This page intentionally left blank

Structural Equation Modeling and Natural Systems

This book presents an introduction to the methodology of structural equation modeling,

illustrates its use, and goes on to argue that it has revolutionary implications for the
study of natural systems. A major theme of this book is that we have, up to this point,
attempted to study systems primarily using methods (such as the univariate model)
that were designed only for considering individual processes. Understanding systems
requires the capacity to examine simultaneous influences and responses. Structural
equation modeling (SEM) has such capabilities. It also possesses many other traits that
add strength to its utility as a means of making scientific progress. In light of the
capabilities of SEM, it can be argued that much of ecological theory is currently
locked in an immature state that impairs its relevance. It is further argued that the
principles of SEM are capable of leading to the development and evaluation of
multivariate theories of the sort vitally needed for the conservation of natural systems.
Supplementary information can be found at the author’s website, accessible via
www.cambridge.org/9780521837422.

Ja me s B. Grace obtained his Bachelor of Science from Presbyterian College his

Master of Science from Clemson University and his Ph.D. from Michigan State
University. He served on the faculty at the University of Arkansas and later at
Louisiana State University, where he reached the rank of Professor. He has, for the past
several years, worked at the US Geological Survey’s National Wetlands Research
Center in Lafayette, Louisiana where he is a Senior Research Ecologist. He holds an
Adjunct Professorship at the University of Louisiana.
Structural Equation Modeling and
Natural Systems

JAMES B. GRACE
  
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo

Cambridge University Press

The Edinburgh Building, Cambridge  , UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521837422

© Cambridge University Press 2006

This publication is in copyright. Subject to statutory exception and to the provision of

relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.

First published in print format 2006

- ---- eBook (EBL)

- --- eBook (EBL)

- ---- hardback

- --- hardback

- ---- paperback

- --- paperback

Cambridge University Press has no responsibility for the persistence or accuracy of s
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
 To my wife Peggy,
for her joyous spirit, wisdom, and laughter.

To my mother and my sister, Diane,

for a lifetime of love and support.

and

To Robert Wetzel, my Major Professor,

for his example of dedication to the pursuit of knowledge.
 Contents

Preface page ix
Acknowledgments xi

PART I A BEGINNING
1 Introduction 3
2 An example model with observed variables 22

PART II BASIC PRINCIPLES OF STRUCTURAL

EQUATION MODELING
3 The anatomy of models I: observed variable models 37
4 The anatomy of models II: latent variables 77
5 Principles of estimation and model assessment 115

PART III ADVANCED TOPICS

6 Composite variables and their uses 143
7 Additional techniques for complex situations 181

PART IV APPLICATIONS AND ILLUSTRATIONS

8 Model evaluation in practice 207
9 Multivariate experiments 233
10 The systematic use of SEM: an example 259
11 Cautions and recommendations 275

vii
viii Contents

PART V THE IMPLICATIONS OF STRUCTURAL

EQUATION MODELING FOR THE STUDY OF
NATURAL SYSTEMS
12 How can SEM contribute to scientific advancement? 291
13 Frontiers in the application of SEM 309

Appendix I Example analyses 324

References 350
Index 361
 Preface

This book is about an approach to scientific research that seeks to look at the
system instead of the individual processes. In this book I share with the reader
my perspective on the study of complex relationships. The methodological
framework I use in this enterprise is structural equation modeling. For many
readers, this will be new and unfamiliar. Some of the new ideas relate to statis-
tical methodology and some relate to research philosophy. For others already
familiar with the topic, they will find contained in this volume some new exam-
ples and even some new approaches they might find useful. In my own personal
experience, the approaches and methods described in this book have been very
valuable to me as a scientist. It is my assessment that they have allowed me
to develop deeper insights into the relationships between ecological pattern
and process. Most importantly, they have given me a framework for studying
ecological systems that helps me to avoid getting lost in the detail, without
requiring me to ignore the very real complexities. It is my opinion, after some
years of careful consideration, that potentially they represent the means to a
revolutionary change in scientific inquiry; one that allows us to ask questions
of interacting systems that we have not been able to ask before. These methods
provide many new opportunities for science, I believe, and it is my hope that
others will see their value as well.
It is important for the reader to keep in mind throughout this book the distinc-
tion between statistical procedures and the scientific enterprise. The application
of structural equation modeling (SEM) to research questions embodies both ele-
ments, but the priorities of one do not necessarily match with those of the other.
My approach to this book is from the perspective of a researcher, not a statis-
tician. My treatment is not designed to satisfy the requirements of statisticians
nor those interested in the mathematics. Rather, I strive to keep the focus on
developing models that match the questions being addressed. Many treatments
of statistical methods are prescriptive and based on protocols that have been

ix
x Preface

worked out on the basis of statistical requirements. While I could simply present
SEM protocols for use by the natural scientist, I am of the opinion that pro-
tocols are commonly an impediment to the best use of statistical methods for
research purposes (see also Abelson 1995). For this reason, my emphasis is on
fundamental issues that provide the reader with the material to make their own
decisions about how to apply statistical modeling to their particular research
problems.
The general arena of studying complex relationships and the specifics of
SEM is one where subject matter and statistical analysis intertwine to a greater
degree than is customary for researchers or statisticians. What is distinctively
different about the study of complex, multivariate relationships compared with
univariate hypothesis testing is the degree to which the analyst has to know both
the subtleties of the methods and the particulars of the system being studied.
The goal of this book is to show why it can be worth the effort to develop and
evaluate multivariate models, not just for statistical reasons, but because of the
added scientific insights that can be gained. Those who apply these methods
to their own data may find, as I have, that it is quite enjoyable. Hopefully the
reasons for excitement will be evident as the reader explores the chapters ahead.
 Acknowledgments

I have a great many people to thank for helping me along the way in this major
venture. I thank Alan Crowden, Dominic Lewis, Emma Pearce, and Graham
Bliss of Cambridge University Press for their support with this project. If this
book is at all readable, it is because of the help of a great many individuals who
provided numerous comments. I especially thank Glenn Guntenspergen, who
not only read the whole volume in both first and last drafts, but who provided
sage advice from beginning to end. To him, I owe the greatest debt of gratitude.
I also wish to express special thanks to Sam Scheiner for many insightful
suggestions on both content and presentation, as well as for advising me on
the best way to present an illustration of SEM practice in the Appendix. I am
appreciative of the USGS National Wetlands Research Center for their history
of supporting the application of SEM to natural systems. Several examples in
this book came from their studies.
Two of my comrades in the quest to introduce SEM to the natural sciences,
Bill Shipley and Bruce Pugesek, both provided very helpful comments. It was
Bill who convinced me that the first draft of this work was far too condensed to
be useful to the student. The readers owe him a great debt of gratitude for the
final structure of this book, which attempts to lead one gradually through the
fundamentals of SEM in the beginning, in order to establish a base from which
to jump into more advanced issues later. Bruce is especially to be thanked for
introducing me to SEM and for working patiently with me through the initial
learning process.
I am also thankful for the training and advice I received from some of the leg-
endary figures in the development of structural equation modeling. My biggest
debt of gratitude is to Ken Bollen, who saved me from several fundamental
errors in the early development of many of my ideas. Ken possesses the rare
combination of being brilliant, patient, and kind, which has been enormously
helpful to me as I have struggled to make the statistical methodology fulfill my

xi
xii Acknowledgments

research ambitions. I am also grateful to Karl Jöreskog and Dag Sörbom for their
week-long workshop on SEM early in my learning and for answering my many
pesky questions about all those complications about which instructors hope to
avoid being asked. Bengt and Linda Muthén likewise shared lifetimes of experi-
ence with me in another week-long SEM workshop, again with tolerance for my
questions about fringe methods and thorny problems. Others who helped me
greatly through my involvement in their training classes include Bill Black
(LSU), Adrian Tomer (Shippensburg University), Alex von Eye (Michigan
State University), and most recently David Draper (University of California –
Santa Cruz).
There are many other people who provided helpful comments on all or part
of the book manuscript, including Jon Keeley, Diane Larson, Bruce McCune,
Randy Mitchell, Craig Loehle, Dan Laughlin, Brett Gilbert, Kris Metzger, Gary
Ervin, Evan Weiher, Tim Wootton, Janene Lichtenberg, Michael Johnson, Laura
Gough, Wylie Barrow, Valerie Kelly, Chris Clark, Elsa Cleland, and Ed Rigdon.
I apologize to any who I have left off the list, the process has gone on long enough
to make it hard to keep track. To all who helped, Thank you!
Last and certainly not least are the people who have provided the encourage-
ment and support in all those more personal ways that are essential to a great
and productive life. My deepest gratitude to my loving wife Peggy, who has
enhanced my quality of life in every way and who led me into a better life. To
my Mother and my Sister Diane, I am unspeakably grateful for all the years
of love and support. To Jeremy, Kris, Zach, Abi, Erica, Madison, Sophie, and
Luke, your acceptance and love means more to me than you know. To Bob
Wetzel, I am grateful for his encouragement over all these many years.
 PA RT I

A beginning
 1
Introduction

The purpose and organization of this book

Structural equation modeling (SEM) represents both a different way of ana-
lyzing data, and a different way of doing science. A major theme of this book
is that one of the factors that has limited the advance of ecological science
has been the absence of methods for developing and evaluating multivariate
theories. Understanding systems requires the capacity to examine simultane-
ous influences and responses. Conventional univariate analyses are typically
limited to the examination of a single or at most a few processes at a time.
Further, as will be illustrated in this book, characterizing interacting systems
using univariate methods is commonly misleading and often inadequate. As I
will argue in the final section of the book, conventional univariate hypothesis
testing propagates a reliance on “theories of pieces” where one or two interact-
ing processes are presumed to explain major characteristics of natural systems.
Single-factor hypotheses seldom provide an adequate representation of system
behavior. Worse still, such hypotheses are unable to be placed into a broader
context or to evolve into more complex theories, regardless of the empirical
evidence. Many of the simplistic theories that have occupied ecologists for so
long seem irrelevant when we are faced with the task of predicting the responses
of natural systems to environmental change. I believe ecologists have remained
focused on univariate questions because we have lacked the scientific tools to
ask and answer more complex questions.
Structural equation modeling offers a means of developing and evaluating
ideas about complex (multivariate) relationships. It is this property that makes
SEM of interest to the practitioner of science. As we shall see in this book,
SEM has both tremendous flexibility and significant requirements for proper
use. There are also some serious issues relating to how best to use SEM for
the study of natural systems. Its mode of application in other fields, such as the

3
4 Structural equation modeling and natural systems

social sciences where it has gained a widespread application, may or may not
suit our needs. Further, ways of connecting structural equation models with the
broader scientific process are needed if we are to gain the maximum impact
from our models and analyses. All these issues need to be addressed if SEM is
to be applied properly and is to have utility for advancing the study of natural
systems.
Before I can discuss fully the potential contributions of SEM to the study of
natural systems, we must first have a fairly clear understanding of the principles
and practice of SEM. What is its history? What are the underlying statistical
principles? How are results from SEM applications to be interpreted? What
are the choices to be made and steps to be performed? After such questions
have been addressed and the nature of SEM is clear, I will consider its broader
significance for the study of natural systems.
To start us on common ground, I begin this chapter with a brief discussion
of material that should be familiar to the majority of readers, classic univariate
null hypothesis testing. This material will provide a point of comparison for
explaining SEM. In this chapter, I will present only a brief and simplistic char-
acterization of SEM, primarily from an historic perspective, while in Chapter
2, I will present an example from an application of SEM to give the reader a
tangible illustration.
In Chapter 3, I begin to present some of the fundamental principles of struc-
tural equation models, emphasizing their reliance on the fundamental princi-
ples of regression. This coverage of basic topics continues through Chapters 4
(latent variables) and 5 (estimation and model evaluation). In Chapter 6, I spend
some time presenting a more advanced topic, composite variables, for the dual
purposes of illustrating this important capability and also to help clarify the role
of latent variables. Chapter 7 provides a very superficial overview of some of
the more advanced capabilities of SEM.
Chapters 8 to 11 will emphasize examples of ecological applications to
give the reader more of a sense of how the principles can be applied to nat-
ural systems. Throughout this section of material, I will contrast the kinds of
results obtained from SEM with those obtained from the conventional scien-
tific methods that have guided (and limited) the natural sciences up to this
point. Experience suggests that such comparisons are often the most effective
means of conveying the potential that SEM has to transform the study of natural
systems. This section of chapters will include an illustration of the sustained
application of SEM to an ecological problem, the understanding of patterns of
plant diversity (in Chapter 10). In Chapter 11, I provide a summary of some
cautions as well as a set of recommendations relating to the application of SEM
so as to provide all of this practical advice in one place.
 Introduction 5

In the final section of the book (Chapters 12 and 13), it will be my purpose
to give an overall view of the implications of applying SEM to the natural
sciences. I will discuss from a philosophical point of view some of the things
that I believe have limited the advance of ecological science, and how SEM
can lead to a greater maturation of our theories and investigations. Included in
this last section will be a discussion of how to integrate SEM into the broader
scientific enterprise. Finally, an Appendix provides example applications that
illustrate some of the mechanics and logic associated with SEM. The reader
will be directed to these examples at appropriate places throughout the book.

An historic point of reference – univariate null

hypothesis testing
In the latter half of the nineteenth century, the quantitative sciences began to
shift from a deterministic viewpoint to one that recognized the need to address
explicitly the concept of probability and error. The story is told by Salsburg
(2001) that as scientists began to make more and more precise measurements,
they discovered that deviations between calculations and observations became
more noticeable, rather than less so. Gradually, the view of a “clockwork”
universe in which outcomes are presumed to be deterministic has been replaced
by one based on statistics and probabilities. As a result, for most sciences, the
process of scientific inquiry has evolved to the point that explicit consideration
must be given to issues of error and uncertainty.
Throughout the twentieth century, the acceptance and elaboration of statisti-
cal procedures steadily increased to the point that nowadays most scientists are
taught basic and even advanced statistical methods as part of their core training.
For many areas of inquiry, the very essence of the scientific method has come to
require statistical design and analysis. Conclusions are only deemed acceptable
if based on accepted statistical conventions (though what is deemed acceptable
can vary among disciplines – and among individuals).
In the natural sciences, the fundamental statistical paradigm that is usually
taught can be represented by the generalized univariate statistical formula

y1 = α1 + X + ζ1 (1.1)

Here, y1 refers to an observed set of responses, α 1 represents an intercept for the

population, X refers to some set of independent variables,  represents some
corresponding vector of coefficients (γ s) that empirically link y1 to the elements
in X (a vector of x variables), and ζ 1 represents random errors associated with the
6 Structural equation modeling and natural systems

responses. Individual values of x and their associated γ 1i values (γ 1i is the effect

of xi on y1 ) can represent a suite of factors, such as experimental treatments
and their interactions, restrictions on randomization due to the experimental
or sampling design (e.g., blocking effects), and uncontrolled covariates that
influence responses.
Of course, the scientific method does not depend solely on the estimation of
parameters, such as those implied by the above equation. It also involves a vari-
ety of procedures that are used to draw inferences about the parameter estimates.
Associated with the univariate model there has long been a set of conventions
and principles for null hypothesis testing that seek to determine whether param-
eter values are nonzero. These ideas and procedures can in large part be traced
back to a combination of procedures and ideas developed by Ronald Fisher,
Jerzy Neyman, and Egon Pearson (see Salsburg 2001 for an interesting his-
torical account). The basic approach that is taught typically includes (1) the
establishment of a null hypothesis (usually of no difference between groups)
which is then compared to an alternative hypothesis, (2) the selection of a
significance level (usually 0.05 or 0.01), (3) data collection involving some
degree of random sampling, (4) calculation of estimates of means/intercepts
and variances (as well as other statistical properties of the populations), and (5)
determination of comparative test results and associated probability values. The
values of the test statistics obtained are usually represented to be indicative of
the probability that observed differences between groups may be due to chance
rather than some systematic difference between populations (this is actually
an oversimplification of the true interpretation of p-values). In the conventional
null hypothesis testing procedure, priority is given to the null hypothesis, which
is generally accepted unless there is convincing evidence to the contrary.
Many are aware of long-standing controversies over various aspects of con-
ventional null hypothesis testing. Ronald Fisher, the founding father of much of
the early development of modern statistics, as well as the originator of paramet-
ric p-values, was himself opposed to the use of fixed cutoffs for standardized
hypothesis testing. In spite of this, the codified usage of null hypothesis testing
based on predetermined and fixed p-values became entrenched in standard-
ized statistical protocols (this is sometimes referred to as the Neyman–Pearson
protocol). There has been a steady drumbeat of complaint about overreliance
on null hypothesis tests ever since their introduction (e.g., see Taper and Lele
2004). In most areas of science, this seems to have done little to prevent the
establishment of null hypothesis testing as the dominant framework. Perhaps
one of the more conspicuous exceptions to this has been in the field of SEM,
where priority has been placed on a-priori theory-based models for the past 35
years. The character of this model-based approach will become more apparent
as we work through the material in this book. At this point, I wish only to make
 Introduction 7

the point that SEM practitioners long ago rejected the logical priority of null
hypotheses, though the use of p-values continues to be one of the tools used in
model evaluation.
It is perhaps useful to note that null hypothesis testing has recently been under
attack from several quarters. Biologists have begun to argue more vigorously for
a departure from reliance on null hypothesis testing (Anderson et al. 2000). The
lack of utility of null hypothesis testing has led to recommendations for the use
of model selection procedures as an alternative basis for developing inferences.
A tenacious effort to expose ecologists to this approach (e.g., Burnham and
Anderson 2002) has begun to bring familiarity with these issues to many. At
present, these efforts remain focused on univariate models and have not yet
tapped into the substantial experiential base of SEM practitioners.
Bayesian methods for estimating parameters and probabilities (Congdon
2001, Gelman et al. 2004) also suggest alternatives to null hypothesis testing.
While there are a number of different variants of the Bayesian procedure, the
basic concept is that from a personal standpoint, the concept of probability
is one that is based on the knowledge available to the investigator. In this
framework, empirical evidence is used to update prior probabilities so as to
generate posterior probability estimates. This form of probability assessment is
preferred by many because it corresponds more directly to the intuitive meaning
of probability as a measure of confidence in a result. As will be discussed in
the final chapter, Bayesian approaches are now being considered for use in
estimation and the evaluation of structural equation models.

What are structural equations?

Let us begin at the beginning. Directly stated, a single univariate equation such
as
y1 = α1 + γ11 χ1 + ζ1 (1.2)
representing the effect of a single x on y, can be said to be structural if there
exists sufficient evidence from all available sources to support the interpretation
that x1 has a causal effect on y1 . If information does exist to support a cause
and effect interpretation, the parameter γ 11 provides an estimate of that effect.
There are certain points that should be made about the seemingly straight-
forward definition above. One is that the sort of information required to support
a causal interpretation for a structural equation is basically the same as that
required for a simple regression or an analysis of variance. Thus, we can see
that the family of univariate equations represented by Eq. (1.1) can be classified
as structural equations.
8 Structural equation modeling and natural systems

A difficulty for some may arise from the fact that our definition of structural
includes the word causal. The average person who is neither a scientist nor
a philosopher may be rather surprised to find that scientists and philosophers
have historically had some difficulty with the concept of causation. Because of
the unease some have with discussing causality, the relationships embodied in
structural equations are sometimes referred to as dependencies instead of causes
(e.g., “the values of y depend on the values of x”). Thus, an alternative definition
that is sometimes seen for structural equations is that they represent statistical
dependencies (or statistical associations) that are subject to causal interpreta-
tion. What is ultimately most important to realize is that, while the results of
structural equation analyses are meant to be reflective of causal dependencies,
it is not the statistical results per se that demonstrate causation. Rather, the case
for making a causal interpretation depends primarily on prior experience and
substantive knowledge.
There has existed over the years an ongoing discussion on the nature of
causality and its relationship to structural equations. Perhaps one of the reasons
structural equation modeling has been slow to be adopted by biologists has
been the priority placed by Fisher on the adage that “correlation does not imply
causation”. His emphasis was on manipulative experiments that sought to isolate
causes and this approach remains a strong emphasis in biological research.
One can see this ongoing debate as a recurring process in which from time
to time some feel it wise to caution against overzealous inference of causes.
These cautionary periods are typically followed by a general defense of the
reasonableness of causal thinking. Fisher’s emphasis was on the development
of rigorous experimental protocols designed to isolate individual causes. The
emphasis of those who developed structural equation modeling has not been
on isolating causes, but instead, on studying simultaneous influences. Both of
these scientific goals have merit and, we can think of them as representing the
study of individual processes versus the study of system responses.
Some clearly articulated insights into the nature of causality and the relation-
ship to structural equations can be found in Wright (1921) and Bollen (1989),
and some of these are described below. Recently, Pearl (2000, Causality) has
addressed the issue of causation and structural equations at the level of funda-
mental logic. A distillation of some of these ideas as they relate to biology can
be found in Shipley (2000).
There are a number of arguments that have been made about the tendency
for some scientists and philosophers to steer away from using causal language.
As Pearl (2000) notes, one reason that some mathematicians have steered away
from discussing causation may be the fact that the language of mathematical
equations relies most commonly on the symbol “=” to describe the relation-
ships between y and x. As a mathematical operator, the “=” symbol simply
 Introduction 9

represents equality, not dependence; implying that if A = B then B = A. It is

not clear to me, that the directional neutrality of the equals sign necessarily
causes confusion, because the designation of y variables as dependent and x
variables as independent does enable causal directionality to be explicitly spec-
ified. Thus, the equation y = x does not possess the criterion of reversibility.
Nonetheless, for those who find the language of equations to be insufficiently
precise with regard to causality, graphical methods are often preferred as an
alternative mathematical language (as we will see below).
According to Pearl, another factor contributing to confusion about the con-
cept of causation is that the word “cause” is not included in the language of
probability theory, which is the official mathematical language of statistics.
On page 138 he states, “. . . statisticians, the arbiters of respectability, abhor
assumptions that are not directly testable [via statistical assessment].”
In this same vein, I suspect that there is a degree to which scientists somehow
associate the invocation of causation with a suggestion of “proof”. As Popper
(1959) so persuasively stated, it is very difficult to prove anything to be true.
Rather, science usually proceeds by falsifying hypotheses and accepting only
those that remain.
Alternatively, for some, a statement of causal interpretation may imply “ulti-
mate” causation. This is a reflection of mechanistic thinking that generally
matches poorly with inference based on statistical relationships. Rather, such
thinking reflects the Aristotelean tradition of determinism, which fails to ade-
quately reflect probabilistic causation and the degree of uncertainty that is
associated with measurement.
As we shall see, structural equation modeling does not start with a null
hypothesis that is given priority. Rather, structural equation models are built
upon the complete body of available knowledge. Models are only rejected if
the observed data do not match expectations derived from the model. This
“model-oriented” philosophy facilitates the invocation of causal interpretation
because it automatically incorporates the a-priori knowledge available to the
scientist. This forms the basis for cause and effect interpretations. Stated in a
different way, if a researcher intends to draw interpretations about x having an
effect on y using structural equations, they should be prepared to defend that
supposition with something other than their statistical results.

Criteria for supporting an assumption of causal relationships

in structural equations
A causal interpretation of a structural equation, such as y1 = α1 + γ11 x1 + ζ1
(Eq. 1.2) can be supported if it can be assumed that a sufficient manipulation
10 Structural equation modeling and natural systems

of x would result in a subsequent change in the values of y, independent of

influences from any other factors (Pearl 2000). This criterion might be referred
to as the “intervention outcome assumption”. One obvious set of conditions
that would support such an assumption would be if previous manipulations had
shown such a result. Thus, the assumption that plant growth would be affected
by a manipulation of water supply can, under many circumstances, be supported
by prior experience. How many times do we have to show through manipula-
tions that plant growth can respond to watering before this effect comes to be
expected? This is not to say that a response in plant growth would be expected
for all possible additions of water supply. Rather, the assumption being made
is that if (1) there is an observed correlation between water supply and plant
growth, (2) the supply of additional water occurs prior to the growth response,
(3) other independent influences on growth are not simultaneously changing,
and (4) an interpretation of plant growth responding to water supply is consis-
tent with known mechanisms, then a causal interpretation of the relationship is
reasonable. Of course, when structural equations are used in conjunction with
manipulative experiments, the criteria for establishing causality are automati-
cally established if the manipulations are unambiguous.
We must recognize that absolute certainty (or even absolute confidence)
is not a requirement for reporting scientific results. At times we may have a
stronger or weaker case in support of a causal interpretation. A great deal will
have to do with prior experience with the system under investigation. I will
have more to say about this subject later as it pertains specifically to structural
equation modeling. My position on this is basically consistent with Fisher’s
ideas about hypothesis testing, we should not expect to rely on single studies
to draw general conclusions. Rather, the process of establishing a confident
interpretation comes from the process of investigation.
For now, it must suffice for us to recognize that any claim for a causal
interpretation will, as always, have to be defended. At the risk of being repetitive,
the requirements for supporting causal interpretations are not fundamentally
different from the ones natural scientists already use. A useful discussion of
criteria that can be used to evaluate the plausibility of claims based on statistical
results can be found in Abelson (1995).

What is structural equation modeling?

Structural equation modeling can be defined in its most basic sense as the
use of two or more structural equations to model multivariate relationships.
Multivariate relationships, as the phrase is used here, refers to those that involve
 Introduction 11

g 31
x1 y3

γ11 g 21
b32 z3
b21
y1 y2

z1 z2
Figure 1.1. Graphical representation of a structural equation model (Eqs. (1.2)–
(1.4) involving one independent (x1 ) variable and three dependent (y) variables.

simultaneous influences and responses. Expressed as a series of equations, an

example of a structural equation model might be

y1 = α1 + γ11 x1 + ζ1 (1.2)
y2 = α2 + β21 y1 + γ21 x1 + ζ2 (1.3)
y3 = α3 + β32 y2 + γ31 x1 + ζ3 (1.4)

Using graphical representation (and omitting the intercepts), Figure 1.1 illus-
trates this model. A quick comparison of the equational and graphical forms
of the model (Eqs. (1.2)–(1.4) versus Figure 1.1) illustrates the appeal of the
graphical representation, which has been in use since structural equation mod-
eling (in the form of path analysis) was first introduced. In this illustration,
boxes represent observed variables, arrows between boxes represent the direc-
tional relationships represented by equality signs in the equations, gammas (γ )
represent effects of x variables on y variables while betas (β) represent effects
of ys on other ys, and the zetas (ζ ) represent error terms for response variables.
The idea of combining a series of equations into a multivariate model, which
should be credited to Wright (1921), has proven to be a major step forward in the
advancement of the scientific method. With a series of structural equations, we
can now specify models representative of systems and address many questions
that cannot be answered using univariate models. In Chapter 3, we will describe
the anatomy of structural equation models and illustrate in more detail the vari-
ous ways structural equations can be combined and how they are interpreted. In
later chapters we will include additional elements, such as unobserved (latent)
12 Structural equation modeling and natural systems

and composite variables. We will also see that there are a large number of prob-
lems that can be tackled using SEM, including nonlinear modeling, reciprocal
influences, and many more.
Of prime importance is that structural equation models allow us to address
scientific questions about systems that cannot be addressed using univariate
approaches. As we shall see, these multivariate questions are often precisely
the ones of greatest importance to the scientist. An introduction to SEM brings
into focus the simple fact that up to this point, we have been studying systems
primarily using a methodology (specifically the univariate model) that is not
designed for examining system responses. Rather, the univariate model, which
has come to represent the scientific method over the past 50+ years, is only
designed for the study of individual processes or net effects, not interacting
systems. To make a crude analogy, if the development of the univariate model
is akin to the development of the wheel, the advance represented by combining
structural equations into complex models is akin to the attachment of four
wheels to a frame to make a cart. The change is not just one of degree, it is a
change of a revolutionary sort.

The history of SEM

The simple yet fundamental definitions given above for structural equations and
structural equation modeling belie the complex variety of terms and emphases
that exist in the literature, and the historical development of the subject. If
you examine the body of papers and books relating to this subject, you will
find the term SEM used in several different ways. You will also find a variety
of other terms that are sometimes substituted for SEM or that seem to refer
to something akin to SEM. The most general definition of structural equation
modeling, which is the one espoused in this book, is “modeling hypotheses with
structural equations”. This definition includes both the historic predecessors of
modern SEM as well as recent advances that have developed.
We can trace the modern formulation of SEM back to its roots in the analysis
of path relations by Wright (1918, 1920, 1921). In these early papers, Wright
clearly established a broad intent for understanding multiple causes and multiple
responses, presenting both criteria for causal inference and guidelines for model
building and interpretation. Wright (1932, 1934, 1960, 1968, 1984), along with
others, continued to develop and refine his method of path coefficients over a
substantial number of years. This method came to be known as path analysis.
It is clear in these publications that a new generation of questions was being
addressed through the flexible application of path coefficients, including both
the decomposition of relationships and the estimation of underlying influences.
 Introduction 13

Reviews of the development of SEM (Hägglund 2001, Tomer 2003) empha-

size that modern structural equation modeling has been influenced by a number
of traditions, particularly biometrics, econometrics, psychometrics, and socio-
metrics (these are essentially the quantitative subdisciplines in biology, eco-
nomics, psychology, and sociology). Modern SEM can primarily be attributed
to the remarkable synthesis of path analysis and factor analysis accomplished by
Karl Jöreskog beginning in the early 1970s (Jöreskog 1973). A close association
developed initially between SEM and the particular formulation he introduced,
which is the LISREL model. More will be said about this in a later chapter, but
for now, I simply wish to introduce some of the variety of terms often associ-
ated with SEM. To add one more level of complication, LISREL refers to both
the underlying notational model developed by Jöreskog and others, and to the
computer software he developed (the LISREL program).
Various other terms can be found in the SEM literature; the meaning and
importance of these will become clear later in the book. To a large degree,
the different terms tend to coincide with either methodological distinctions,
emphases of importance, or methodological advances. For example, the mean-
ing of the term “path analysis” has changed from its original meaning. The
techniques originally used by Wright are sometimes called Wrightian path
analysis. This is done to distinguish the fact that modern path analysis is now
based on the LISREL model and incorporates maximum likelihood methods
and tests of overall model fit. As will become clear later, LISREL-based SEM
is sometimes subdivided into path analysis (modern type), factor analysis, and
combined factor–path models, which are referred to in some books as struc-
tural equation models (I suppose this usage could be referred to as “SEM sensu
stricto”).
An illustration of the use of names that represent emphases of importance is
the term causal modeling. This term was popular when the contribution of SEM
to understanding causal relations was being emphasized. Yet another term that
appeared early on and that I will refer to in this book is covariance methods
(sometimes covariance structure analysis). The LISREL model and program
are distinctive because of the fact that they do not deal with the responses of
individuals (at least they did not traditionally). Rather, they are based on the
properties and analyses of covariance (or correlation) matrices. This character-
istic is closely associated with LISREL and, for many, with SEM in general. As
will be discussed later, there are other ways of analyzing structural equations
rather than those that rely on covariances, and the synonymy between SEM and
covariance analysis is becoming a thing of the past.
Another term strongly associated with modern SEM is latent variable mod-
eling. The flexibility of using structural equations to represent relationships
14 Structural equation modeling and natural systems

creates the ability to estimate the influences of unmeasured, underlying causes

or latent variables. The concept of the latent variable has been traced back
to Plato, although its meaning in a statistical context seems to begin with Sir
Francis Galton in the late 1800s. The classic example of a latent property is
the concept of human IQ (the intelligence quotient). It is presumed that the
latent property of general intelligence (G) exists, but there is no direct way
of observing it. Rather, it can only be evaluated indirectly by estimating its
effects. Presenting a simplistic view of this complex topic, if a large number of
individuals are given a battery of ability tests, those individuals who performed
well across the whole range of tasks would be estimated to have a high score
for G, while those who performed poorly would be estimated to have a low
score for G. In addition to the presence of a general intelligence factor, it can be
presumed that there may be more specific intelligence factors, such as mathe-
matical and verbal ability, and analytical reasoning. Thus, one’s capacity for
answering some subsets of questions on a test might be influenced by several
different abilities simultaneously. Problems of the same sort that involve the
estimation of general and specific genetic factors in plants and animals were
also addressed by Wright using his method of path coefficients.
The ability to construct models of latent factors and to estimate their value
from a variety of correlated measures (so-called factor analysis) was first devel-
oped by Charles Spearman. More about the history of the development of factor
analysis and its incorporation into SEM can be found in Hägglund (2001).
Factor analyses can either be exploratory, in that theory is not used to guide
the analysis, or confirmatory, where theory is used to structure the model.
Confirmatory factor analysis has been an integral component of SEM since
the LISREL synthesis, which combined path analysis and factor analysis into
a single procedure.
Of further relevance to this issue of terminology is the fact that there are
new methods related to the LISREL model, as well as to the general enterprise
of SEM, that are of a different sort. For example, there has emerged a suite
of techniques that analyze data at different levels (e.g., within populations and
between populations). These models are variously called multi-level or hier-
archical models and are natural evolutionary developments of SEM. They fit
clearly within the broad definition of SEM used in this book. However, they are
not all based on the covariance procedures of the LISREL model (nevertheless,
they are now being incorporated into the LISREL software program).
It is hoped that this presentation of various terms and their historical associ-
ations will not distract the reader from the fundamental and unified importance
of structural equation modeling. Regardless of exactly how structural equation
models are analyzed and what capabilities are emphasized, the act of using
 Introduction 15

100%
software
80%
model evaluation
60% stat modeling
40% factor analysis
regression
20%
0%
Figure 1.2. Representation of the elements of modern SEM. Stat modeling refers
to statistical modeling.

combinations of equations to specify interconnected causal relationships cre-

ates a capacity for studying hypotheses about systems. As we will examine in
the next section, modern SEM has come to embody many capabilities through
its evolution. This evolution is continuing as more refined and varied capabil-
ities are added to the software packages that implement solutions to structural
equation models.

The elements of modern structural equation modeling

Figure 1.2 provides us with another way of looking at the elements of modern
SEM. This figure is intended to illustrate the various topics that need to be
considered to understand and apply SEM. As indicated in this figure, SEM can
be seen as a combination of regression, factor analysis, statistical modeling,
model evaluation/selection, and software. To clarify, by statistical modeling I
mean the construction of models so as to match the particulars of a situation,
or to allow a particular type of contrast to be evaluated.
Wright’s original formulation of path analysis was an elaboration of regres-
sion, with elements of statistical modeling used to distinguish direct and indirect
effects. In Chapter 3, we will examine some of the regression aspects of SEM
in detail, as an understanding of the role of regression statistics in SEM is vital
for understanding results and interpretations. Factor analysis, another major
element of the modern method, deals with the incorporation of latent or un-
observed variables in a model and is itself an extension of regression tech-
niques, although with particular considerations of its own (to be covered in
Chapter 4).
The synthesis created by the formulation of the LISREL model (Jöreskog
1973) brought together all the elements shown in Figure 1.3. One of the pivotal
things associated with the LISREL synthesis is the utilization of maximum
16 Structural equation modeling and natural systems

Theoretical Content of Model

High Low
Confirmatory Analysis Exploratory Analysis

Structural Equation Modeling

Causal Search Methods

Discriminant Analysis

Regression Trees

PCA, CCA, NMS

Multiple
Regression

Figure 1.3. Classification of some methods used to analyze multivariate data.

PCA refers to principal components analysis, CCA to canonical correspondence
analysis, and NMS to nonmetric multidimensional scaling. Multiple regression,
while not a true multivariate method (i.e., it is based on univariate model with
single response variable), is included because it is often used to analyze data that
are multivariate.

likelihood methods for estimation. It was this advance that permitted (1) a gen-
eralized solution procedure for factor and path models, (2) a proper approach to
nonrecursive models, and (3) an assessment of discrepancies between observed
and expected which permits hypothesis evaluation at the level of the over-
all model. Simultaneous to the development of the LISREL model, Jöreskog
developed computer software (largely because of the iterative requirements of
maximum likelihood methods), and since that time, SEM has come to be defined
in part by the capabilities of the software packages used to implement analyses.
Over the years, numerous capabilities have been added to SEM, both through
theoretical advances and through additions to the various software packages
that have been developed. Such things as multigroup analyses, the inclusion
of means modeling, numerous indices of model fit, modification indices, capa-
bilities for handling categorical variables, and most recently, procedures for
hierarchical models have been added. All of these have expanded the elements
that make up SEM.
 Introduction 17

At present, a number of new developments are taking place which have the
potential to strongly influence the future growth and capabilities of SEM. Novel
approaches to the analysis of network relationships are emerging from fields
as diverse as robotics and philosophy. Their goals range from pattern recog-
nition to data mining, from computer learning to decision support. Some of
these developments are fairly unrelated to the enterprise of structural equation
modeling because of their descriptive nature. However, others have substantial
relevance for the estimation and evaluation of structural equation models and
are already starting to be incorporated into the SEM literature. Collectively,
these new methods, along with SEM itself, fall under the umbrella concept of
graphical models (Pearl 2000, Borgelt and Kruse 2002). Of particular impor-
tance are those graphical models that incorporate Bayesian methods, such as
Bayesian networks (e.g., Neapolitan 2004), which permit what is now being
called Bayesian structural equation modeling (Scheines et al. 1999, Congdon
2003).

Comparing SEM to other multivariate methods

It is legitimate to question how SEM differs from other multivariate methods
that are more familiar to ecologists, such as principal components analysis,
cluster analysis, discriminant analysis, canonical correlation, canonical corre-
spondence analysis, and MANOVA (descriptions of these and other multivariate
methods used for ecological analysis can be found in McCune and Grace 2002).
Fornell (1982) has referred to these methods as “first generation” multivariate
methods, in contrast to the “second generation” methodology of SEM. One
essential distinction between so-called first and second generation methods is
that the first are largely descriptive and more appropriate for exploratory stud-
ies or the development of predictive equations, while the second are capable
of confirmatory tests and are thus more suitable for evaluation of multivariate
hypotheses. A second key difference is that conventional multivariate methods
seek to estimate generic fixed models (e.g., canonical correlation), and lack the
flexibility required to represent the model that best matches a particular situa-
tion. Thirdly, SEM seeks to examine networks of causal relationships, while first
generation multivariate methods are focused on net effects. Below, I elaborate
on the first two of these points.

Confirmatory versus exploratory analyses

First and second generation multivariate methods differ substantially in their
theoretical content. While methods used typically for exploratory analyses do
18 Structural equation modeling and natural systems

sometimes provide tests for the significance of coefficients, their capacity to

represent complex causal relationships is often very limited; stated in another
way, their capability for specification of theoretical content is quite limited.
In contrast, methods such as SEM are capable of possessing a great deal of
theoretical content, although they can be used to fit and test simple models
as well. As we shall see in subsequent chapters, SEM is a very generalized
methodology and many familiar multivariate and univariate procedures can be
seen as special cases of a structural equation model.
Aside from their potential for incorporating a large amount of theoretical con-
tent, first and second generation multivariate methods also differ in the degree
to which they provide feedback from data. This feedback comes in a variety of
forms, ranging from model failure to the detection of deviations between model
and data. Essentially, methods that subject models to “failure tests” allow for
tests of model structure, while those that only estimate parameters do not allow
tests of the model, only tests of individual coefficients.
Figure 1.3 provides a classification of some common multivariate methods in
terms of their capacity for including theoretical content. Those methods capable
of containing high theoretical content are considered to be the most rigorous
for evaluation of complex a-priori hypotheses. Note that some methods have a
broad range of uses. For the purposes of this discussion, low theory methods
are those that initially possess a small theoretical content and then allow a
statistical routine to estimate parameter values that satisfy a set of criteria. In
contrast, highly theoretical models specify relationships based on a detailed
theoretical model and seek to determine if the data collected are consistent
with the expectations of the model. Such analyses also provide estimates of the
relative strengths of relationships in the model in a way that is more suitable
for substantive interpretation; this is a major motivation of SEM that will be
discussed in Chapter 3.
Since methods differ in their characteristics (Table 1.1), they also differ
in the type of application to which they are best suited. For example, SEM
techniques can be applied to problems that range from strictly confirmatory
to highly exploratory. Other methods, such as principal components analysis
(PCA), do not have procedures for directly evaluating a-priori ideas about the
precise nature of underlying factors that explain a set of correlations. At present,
many of the techniques shown in Figure 1.3 are used by ecologists primarily in
a descriptive or exploratory way, even though they can be and sometimes are
used in a confirmatory way.
It is important to keep in mind that exploratory methods can be useful or even
essential in the initial stages of developing multivariate models. Results from a
multiple regression may suggest a multivariate model that can be evaluated later
using SEM. Principal components analysis can suggest the existence of latent
 Introduction 19

Table 1.1. Attributes of multivariate methods related to multivariate hypoth-

esis formulation and evaluation. SEM is structural equation modeling, DA is
discriminant analysis, RT is regression trees, PCA is principal components
analysis, and MR is multiple regression

SEM DA RT PCA MR
Include measures of absolute model fit ✓
User can specify majority of relationships ✓
Capable of including latent variables ✓ ✓
Able to address measurement error ✓
Allows evaluation of alternative models ✓ ✓
Examines networks of relationships ✓
Can be used for model building ✓ ✓ ✓ ✓ ✓

variables which can be evaluated using factor analysis. It is not appropriate

to use SEM in all situations and for all purposes, thus, the other multivariate
methods referenced in this section will continue to play an important role in the
characterization of ecological data.

Philosophy of presentation
As stated previously, the first goal of this book is to introduce the reader to
structural equation modeling. There are many different ways in which this can be
approached. One is to emphasize the fundamental mathematical and statistical
reasoning upon which modern SEM is based. An excellent example of this
approach can be found in Bollen (1989). Another approach is to emphasize the
basic concepts and procedures associated with SEM. A fair number of textbooks
offer this kind of introduction to the subject (e.g., Hair et al. 1995, Schumacker
& Lomax 1996, Loehlin 1998, Maruyama 1998, Raykov & Marcoulides 2000,
Kline 2005). The presentations in these works are meant to apply to any scientific
discipline that chooses to use SEM; in other words, they are descriptions of the
SEM toolbox, along with general instructions. A third kind of presentation
which is especially popular for those who have a background in the basics is
one that emphasizes how individual software packages can be used to apply
SEM (Hayduk 1987, Byrne 1994, Jöreskog and Sörbom 1996, Byrne 1998,
Kelloway 1998, Byrne 2001).
One other approach to the material would be to examine how modern SEM,
which has been largely developed in other fields, might be applied to the study of
natural systems. This is a focus of the first part of this book. Shipley (2000) was
20 Structural equation modeling and natural systems

the first to devote an entire book to present methods associated with modern
SEM specifically to biologists (although classic methods were presented in
Wright 1934 and Li 1975). In this effort, he covered a number of basic and
more advanced topics, with an emphasis on the issue of inferring causation from
correlation. More recently, Pugesek et al. (2003) have presented an overview
of the conventions of modern SEM practice, along with a selection of example
applications to ecological and evolutionary problems. A brief introduction to
the application of SEM to the analysis of ecological communities can also be
found in McCune and Grace (2002, chapter 30).
The need for a presentation of SEM that specifically relates to natural systems
is justified by the way in which disciplinary particulars influence application.
The flexible capabilities of SEM offer a very wide array of possible applications
to research problems. It is clear that the characteristics of the problems them-
selves (e.g., nature of the data, nature of the questions, whether manipulations
are involved, sample size, etc.) have a major influence on the way the SEM
tools are applied. Because many nuances of SEM application depend on the
characteristics of the problem being studied, its application is not so well suited
to a standardized prescription. This is why various presentations of SEM often
differ significantly in their emphasis. For example, researchers in the social
sciences where SEM has been most intensively used thus far, often apply it to
survey data, which tends to have large sample sizes, suites of related items, and
a major concern with measurement (e.g., how do we estimate a person’s ver-
bal aptitude?). When these same methods are applied to experimental studies
where a small number of subjects, for example, forest tracts or elephants, can
be studied intensively, the issues that are most important shift.
I have found that natural systems scientists often find it difficult to grasp
immediately the meaning and value of certain aspects of SEM. If this were not
the case, the books that already exist would be sufficient and biologists would
be applying these methods routinely to their full potential. While there have
been a significant number of applications of SEM to natural systems, it is safe
to say that the capability and appropriate use of these procedures has not been
approached. A decade of effort in applying SEM to ecological problems (as of
this writing) has convinced me that there are substantial challenges associated
with the translation and adaptation of these methods.
In conversations with biologists and ecologists I have consistently heard
requests for examples. Those in the natural sciences often do not see the kinds
of analyses obtained through structural equation modeling in the social sciences
as important or relevant to their problems. Very commonly, biologists still see
the univariate procedures that are applied to their data as sufficient. Why, they
ask, should one go to all the trouble of learning and applying SEM? For these
 Introduction 21

reasons, my approach in this book will be one of illustration and comparison.

The four primary things I hope to convey are:
(1) the basic principles and procedures of structural equation modeling
(2) how the capabilities of structural equation methods can be related to eco-
logical research problems
(3) how the results from these analyses differ from those that would be achieved
if just a univariate approach had been used
(4) how the process of confronting ecological data with structural equation
models can transform the study of natural systems and contribute to the
maturation of our science.
The approach I will use to present the material will be oriented towards
examples. Often, basic issues will not be discussed until after a relevant exam-
ple has been introduced. This is intended to make concepts tangible. This order
will not be followed in all cases, however, for purely practical reasons, neverthe-
less, illustration through example will be applied wherever possible. In order
to accommodate this approach, concepts and procedures will be introduced
gradually. I will generally build from material that is more readily understood
to the more unfamiliar or advanced topics. In the first section of chapters I will
illustrate an SEM application to make the subject more tangible. In the second
section, the basic principles of SEM will be introduced. A brief introduction
to more advanced topics will be the subject of the third section, while example
applications will occupy the fourth section. In the final section, I will discuss
the implications of multivariate model evaluation for the advancement of sci-
ence and consider where this endeavor may take us. With luck, this approach
will provide the natural scientist with a working knowledge of the basics of
structural equation modeling, how it can be applied, and what they stand to
gain in their understanding of natural systems.
 2
An example model with observed variables

In this chapter I present a simple illustration of a structural equation model.

My purpose is to provide a concrete example of structural equation modeling
(SEM) for the reader and to facilitate subsequent discussion. In this chapter,
the emphasis is on illustrating SEM results and comparing those to the results
achieved from univariate analyses. In subsequent chapters I will introduce the
principles and procedures of structural equation models, which will shed further
light on both the origin and interpretation of SEM results.

Background and univariate results

The plant leafy spurge, Euphorbia esula (Figure 2.1), is one of the most
troublesome exotic species in North America. This species was accidentally
introduced through multiple shipments of contaminated crop seed and has
spread throughout the northern and central latitudes, causing substantial eco-
nomic and ecological damage. Considerable effort has been spent testing and
then releasing a number of biocontrol insect species, particularly in the genus
Aphthona (the flea beetles).
The site that this example comes from is the Theodore Roosevelt National
Park, which is a grassland preserve located in North Dakota, USA. From an
initial infestation of 13 hectares in 1970, spurge populations have spread rapidly
in the park. An aggressive biocontrol program was initiated in the 1980s. Since
then, there have been more than 1800 releases and redistributions of populations
of two flea beetle species, A. nigriscutis and A. lacertosa, as control agents for
leafy spurge. During the period 1999–2001, 162 vegetation plots were moni-
tored throughout the park to census vegetation as well as to estimate population
densities of the two flea beetles.

22
 An example model with observed variables 23

Figure 2.1. Euphorbia esula L. From USDA, NRCS. 2005. The PLANTS
database, version 3.5. (http://plants.usda.gov).

Univariate analyses and results

There are a number of questions that can be addressed using the data collected
in this simple survey; some rather obvious and others less so. What is of greatest
interest is the question of whether the biocontrol program is working. In this
section I will present some basic univariate results aimed at shedding light on
this main question. In the next section, which presents the SEM results, we will
revisit that main question and also consider a number of related issues.
Repeated measures analysis of variance was used to examine changes over
time in both the proportion of leafy spurge plants that flowered, and the density
of stems per plot. Examination of leafy spurge populations during the study
(Figure 2.2) showed that average stem density declined from 2000 to 2001,
while the percentage of stems flowering decreased throughout the study. These
results would thus seem to be consistent with the premise that leafy spurge
24 Structural equation modeling and natural systems

50

40

30

20

10

0
1999 2000 2001

% Flowering Stem Density

Figure 2.2. Changes in leafy spurge over time at Theodore Roosevelt National
Park (means ± standard errors). Changes over time for both parameters were
judged to be statistically significant using univariate procedures.

r = −0.20, p = 0.01 r = −0.39, p < 0.001

40 40
20 20
% Change
% Change

0 0
−20 −20
−40 −40
−60 −60
−80 −80
−100 −100
0 2 4 6 0 2 4 6 8
log A. nigriscutis log A. lacertosa

Figure 2.3. Regression of change in spurge between 2000 and 2001 against the
log of the density of Aphthona nigriscutis and Aphthona lacertosa.

populations are stressed (indicated by the declining incidence of flowering) and

declining in the face of the biocontrol effort.
An additional univariate analysis relevant to the main premise being exam-
ined is to see if there are any correlations between the local density of flea
beetles in a plot and the rate of change of spurge in that plot. To determine this,
the relationship between the change in spurge density from year to year was
regressed against the density of each flea beetle to determine whether there was
evidence for an effect of flea beetles on spurge. For this latter determination,
linear regression analysis was used, with the log of the beetle density as the
x variable and the percent change in spurge as the y variable. Changes in spurge
density were found to be negatively correlated with flea beetle populations in
both the 1999–2000 and 2000–2001 time intervals. An illustration of the rela-
tionship for the period 2000 to 2001 is shown in Figure 2.3. As can be seen, there
Table 2.1. Correlations and standard deviations for leafy spurge density, change in density between years, and the population
densities of two flea beetles that are biocontrol agents for spurge in 2000 and 2001. Correlations in bold are statistically significant
at the p < 0.05 level

Spurge Spurge density A. nigris. A. nigris. A. lacertosa A. lacertosa

density 2000 change density 2000 density 2001 density 2000 density 2001
2000–01
n = 162 stems00 change nig00 nig01 lac00 lac01
stems00 1.00
change −0.63 1.00
nig00 0.51 −0.28 1.00
nig01 0.52 −0.20 0.75 1.00
lac00 0.15 −0.24 −0.12 −0.18 1.00
lac01 0.31 −0.39 0.03 0.03 0.70 1.00
std dev 1.1276 18.031 1.7045 1.6647 1.9756 1.8714
26 Structural equation modeling and natural systems

was a negative relationship between change in spurge density and flea beetle
density for both species. These univariate results would seem to further support
the interpretation that the biocontrol agents are contributing to the observed
decline in spurge.
Finally, examination of the bivariate correlations among variables (Table 2.1)
provides additional insights into some of the apparent relationships among the
species. For the sake of simplicity, I will lump bivariate results (i.e., correlations)
with regression coefficients under the general heading of “univariate” results.
Only the correlations for the period 2000 to 2001 are presented here. A more
complete exposition can be found in Larson and Grace (2004). As can be seen
in Table 2.1, 10 of the 15 bivariate correlations are statistically significant at
the p < 0.05 level. Several points are worth noting about these individual
correlations:

(1) The change in spurge density between years is not only negatively correlated
with A. nigriscutis and A. lacertosa in 2001 (as shown in Figure 2.3), but
also with their densities in 2000.
(2) The change in spurge density is additionally correlated (negatively) with
spurge density in 2000.
(3) The spurge density in 2000 correlates with A. nigriscutis density in both
2000 and 2001, while it only correlates with A. lacertosa in 2001.
(4) Species A. nigriscutis and A. lacertosa only correlate with themselves
across times; there are no correlations among species that are apparent.

These results indicate that the full story may be more complicated than described
thus far, primarily because the change in spurge was not only related to flea
beetle densities, but also strongly related to initial spurge density.

Structural equation modeling and results

Description and justification of initial model
Structural equation modeling typically begins with an initial model or selection
of competing models which are formulated based on a-priori information. The
initial model in this case concerns the changes in insects and spurge across
years, and is shown in Figure 2.4. The structure of this model was chosen so as
to disentangle the reciprocal interactions between species, which can sometimes
be difficult to do. Here, the overall reciprocal interaction is addressed by parsing
out the components of interaction (a positive effect of plants on insects and a
negative effect of insects on plants) over time. An extra complication in this
case is the potential for self regulation in spurge. Because of the potential for
 An example model with observed variables 27

A. nigriscutis c A. nigriscutis
2000
0 2001
k e
h
a i
m Change in
Number of
stemss
stems 2000
2000--2001

b g j
A. lacertosa
l f A. lacertosa
a
2000
0 2001
d
Figure 2.4. Initially hypothesized model used to evaluate factors affecting leafy
spurge changes in density over time (from Larson and Grace 2004, by permission
of Elsevier Publishers).

self regulation, change in spurge over time might relate both to flea beetles and
to initial spurge density. These issues were dealt with by (1) separating changes
in stem density over time from instantaneous values of stem density, and (2) by
modeling the interaction with flea beetles over time.
The structure of the initial model (Figure 2.4) was based on a number of
premises, which were derived from general knowledge about these species and
similar systems. These premises are as follows:

(1) In any given year (denoted by year a), the densities of flea beetles can be
expected to covary with the density of spurge because these insects depend
on this species for food. These Aphthona species are presumed to be obligate
feeders on Euphorbia esula, so it would be reasonable to expect that spatial
variations in spurge densities will cause spatial variations in flea beetle
densities (paths a and b).
(2) The densities of insects in a plot would be expected to be related over time.
This premise is based on a number of biological features of the system
relating to dispersal and food requirements. However, it is known that in
some cases high population densities one year can lead to lower ones the
next (e.g., because of interactions with a parasite), so it would not be a
surprise if correlation over time were weak or variable. Thus, here we are
assessing site fidelity for the insects (paths c and d).
(3) There may be time-lag effects in the dependencies of flea beetles on spurge
density. This premise is supported by the fact that the flea beetles feed on
28 Structural equation modeling and natural systems

the plants for some period as larvae before emerging to feed as adults (paths
e and f).
(4) An important question is whether the species of Aphthona may interfere
with one another or otherwise be negatively associated. No a-priori infor-
mation was available upon which to base an expectation, therefore, we
included the pathways (paths g and h) to evaluate this possibility.
(5) The fundamental premise of the biocontrol program is that the flea beetles
will have a negative effect on the change in spurge over time (paths i and j).
Again, because feeding by the insects on plants begins in their larval stage
(note that larvae were not directly observed in this study), it is reasonable to
expect that some of the variation in spurge density changes will be related
to the density of insects in the previous year (paths k and l).
(6) The change in spurge density may be expected to correlate with initial
density due to density dependence. Most likely this would be manifested
as a negative influence of initial density due to autofeedback (a reduced
increase in stems at high densities) or even self thinning (a greater rate of
decline at high densities).

The procedures by which we determine whether the data support our initial
model will be presented later in this book. What the reader needs to know at
this time is that there are processes whereby expectations about the data are
derived from model structure, and the actual data are then compared to these
expectations. This process not only provides estimates for all the parameters
(including the numbered paths, the variances of the variables, and the distur-
bances or error terms), but also provides an assessment of overall model fit.
When the relationships in the data are found to inadequately correspond to the
initial model, the combination of a-priori knowledge and initial results is often
used to develop a revised model, which is then used to obtain final parameter
estimates. The results from revised models are not viewed with as much confi-
dence as results from an initial model that fits well. Rather, they are seen as in
need of further study if confidence is to be improved. That said, it is generally
more appropriate to interpret parameters from a revised model that fits well
than from an initial model that does not.

Assessment of initial structural equation model

Structural equation modeling analyses indicated that the data did not fit the ini-
tial model adequately. This determination was made based primarily on indices
of overall model fit, such as chi-square tests and other measures. The reader
should be aware that these measures of model fit are used in a fundamentally
 An example model with observed variables 29

different way for structural equation models than for null hypotheses. In null
hypothesis testing, priority is given to the hypothesis of no relationship. This is
the case no matter what our a-priori knowledge is about the processes involved.
In contrast, when evaluating overall fit of data to a structural equation model,
priority is given to the model, and test results are used to indicate whether there
are important deviations between model and data. Thus, in SEM, the a-priori
information used to develop the initial model is used as a basis for interpretation.
Stated in other terms, we presume that our other knowledge (aside from the data
at hand) is important and that our data are meant to improve our knowledge.
Based on modification indices (which are provided by some SEM software
programs to indicate more specifically where lack of fit occurs), a relationship
not initially anticipated needed to be added before an adequate match between
model and data could be obtained. This missing relationship indicated by the
data was a negative correlation between A. nigriscutis and A. lacertosa in 2000
(the meaning of which is discussed below).
Now, the procedures in SEM are not completely devoid of null hypothesis
tests. Associated with the estimation of model parameters is the calculation
of standard errors for all path coefficients, and these can be used to estimate
associated p-values. Unlike usual null hypothesis testing procedures, though,
these paths are not always removed from the model simply because we fail to
reject the null hypothesis. Rather, the decision to remove a path is based on the
consequences for overall model fit, the individual p-values, and our substantive
knowledge. In our example, the paths from flea beetles in 2000 to change in stem
counts in year 2001 (paths k and l) were found to be indistinguishable from zero
based on t-tests. Because we were initially uncertain about whether these paths
should be in the model, they were removed. Also, the probability associated with
path h, which represented an effect of A. nigriscutis on A. lacertosa, suggested
little chance that such an effect occurred in this case. Therefore, this path was
also removed. Model chi-square changed little with the removal of these paths,
confirming that these relationships did not represent characteristics of the data.
The path from spurge density to A. lacertosa in 2000 was also found to be
weak and the probability that it differs from zero unlikely. However, this path
was retained in the model because of the potential that such an effect does occur
some of the time. In fact, other data have revealed that the relationship between
spurge density and A. lacertosa density does sometimes show a nonzero value.
One final change to the model was deemed necessary. Results indicated that
path i was nonsignificant (the effect of A. nigriscutis in 2001 on change in spurge
stems). When this path was removed, it was found that there is actually a residual
positive correlation between A. nigriscutis in 2001 and change in spurge stems.
This relationship was represented in the revised model as a correlated error
30 Structural equation modeling and natural systems

R 2 = 0.26 R 2 = 0.61

A. nigriscutis 0.62 A. nigriscutis

2000 2001

0.51 0.23 −0.14 0.09

−0.56 Change in
−0.20 Number of
stems
stems 2000
2000--2001 R 2 = 0.44
ns 0.21
−0.22

A. lacertosa 0.67 A. lacertosa

2000 2001

R 2 = ns R 2 = 0.53

Figure 2.5. Results for revised model. Path coefficients shown are all standardized
values (from Larson and Grace 2004, by permission of Elsevier Publishers).

term, which represents the effects of a joint, unmeasured causal factor (this
topic is explored in more detail in Chapter 3). Following these modifications
of the initial model, a reanalysis was performed.

Results for accepted structural equation model

Upon reanalysis, the data were found to correspond very closely to the revised
model. As mentioned elsewhere, some aspects of this model (particularly the
pathway that was added) run the risk of capitalizing on chance characteristics
of the data because they are based on a revised model. However, subsequent
analyses from the other years and from another area of the park were also
found to be consistent with this revised model (Larson, unpublished data),
which provides us with some additional confidence in our interpretations.
The results from our accepted model, which are summarized in Figure 2.5,
provide a wealth of information about this system. Not only are there 11 path
coefficients shown in the model, but the paths that are omitted provide important
information as well. The main findings are as follows:

(1) A. nigriscutis density appears to track the density of spurge plants to a

substantial degree (path value of 0.51) while A. lacertosa’s association
with spurge is much weaker.
(2) A. nigriscutis and A. lacertosa in 2000 were found to be negatively cor-
related. There are three possible explanations for this. One is that the two
species have different habitat preferences, leading to a partial separation in
space. Another is that the two species were initially released in different
 An example model with observed variables 31

areas and have not fully dispersed yet. The third is that there is a neg-
ative interaction between the two species. There is currently insufficient
information to resolve between these possible causes.
(3) The densities of both A. nigriscutis and A. lacertosa show a fair degree of
consistency over time (path coefficients of 0.62 and 0.67).
(4) Both species of flea beetle in 2001 showed a lag relationship to spurge
density in 2000 (path coefficients of 0.23 and 0.21).
(5) There is evidence for a modest negative effect of A. lacertosa in 2000 on
A. nigriscutis in 2001 (path of −0.14). There is no indication that A.
nigriscutis has a similar effect on A. lacertosa.
(6) Changes in spurge stem densities from 2000 to 2001 were found to be
negatively related to A. lacertosa (path coefficient of −0.22). The most
reasonable interpretation for this relationship is that it represents the effects
of feeding by the insect on the plant. It does not appear that A. nigriscutis is
having a negative effect on spurge, however. Thus, it would seem that the
two biocontrol agents are not equally effective in reducing spurge.
(7) The biggest factor affecting changes in spurge density is negative density
dependence.

A comparison of univariate and SEM results

The results of the analysis revealed a number of findings and implica-
tions that were not apparent from an examination of the univariate results
(Figure 2.3, Table 2.1). Here I make the presumption (supported by a substan-
tial body of theoretical statistics, as well as experience) that path coefficients
are more accurate indicators of true relationships than are univariate regression
results and bivariate correlations. If the reader will accept this presumption for
the time being, a comparison between the univariate and multivariate results can
be made. Here I summarize the SEM findings and some of the key similarities
and differences that were found.

(1) A. nigriscutis tracks spurge strongly while A. lacertosa does not – this was
also found in the bivariate correlations.
(2) A. nigriscutis and A. lacertosa had an initial negative association – this was
not found from examination of bivariate correlations.
(3) Densities of A. nigriscutis and A. lacertosa were fairly consistent over time –
this was found from examination of correlations; however, bivariate results
overestimate the fidelity by a fair bit.
(4) A. nigriscutis and A. lacertosa both show a lag relationship to spurge density
in 2000 – this relationship could not be examined using univariate analyses.
32 Structural equation modeling and natural systems

(5) There is evidence for a negative effect of A. lacertosa in 2000 on A.

nigriscutis in 2001 – there was no evidence for this in the univariate results.
(6) Only A. lacertosa was found to have a significant effect on spurge density –
univariate results suggested that both species had effects on spurge; further,
univariate results substantially overestimate the effects of flea beetles on
spurge.
(7) The biggest factor affecting changes in spurge density is self thinning –
this is consistent with the bivariate correlations; however, the correlations
cannot be used to derive accurate estimates of the relative importances of
factors without the use of a multivariate model.

Conclusions
It is apparent, I hope, that a structural equation model provides a framework for
interpreting relationships that is substantially superior to the piecemeal inspec-
tion of univariate results. When we look at Table 2.1, we see many correlations
that are significant and some that are not. When we look at Figure 2.4, we
see a whole system of interactions with complete accounting of direct and in-
direct interactions and the relative strengths of pathways (in the next chapter
we will discuss the unstandardized path coefficients, which give the absolute
strengths of pathways). Without a substantive model to guide our interpretation,
it would be unclear what relationships are specified by the various correlations in
Table 2.1. As it turns out, some bivariate correlations correspond to path coeffi-
cients and others do not. The reasons for this will also be presented in the next
chapter.
It has also been shown that path coefficients can differ quite a bit from correla-
tion coefficients. The most conspicuous example of that is the apparent negative
relationship between A. nigriscutis and changes in spurge density (Figure 2.2),
which turns out to be a spurious correlation. It comes about because both of
these variables are influenced by a common factor, spurge density in 2000. The
density of A. nigriscutis in 2001 is positively affected by spurge density in 2000
while the change in stems is negatively affected. This structure automatically
creates a negative correlation between A. nigriscutis and the change in spurge
density. It is through the use of our structural equation model that we are able
to ascertain that the correlation is completely spurious.
Finally, the reader should be made aware that obtaining a correct interpreta-
tion depends on having the correct model. We have gone to some effort to ensure
a match between model and data. The fact that they do match is no guarantee
that the model (and our interpretations) are correct. We will have quite a bit
 An example model with observed variables 33

more to say about this issue in later chapters. For now, let us simply say that
SEM, when properly applied, has the potential to provide much greater insight
into interacting systems than do traditional univariate models.

Summary
The results from this example are both ecologically interesting and also very
instructive as to the value of partitioning relationships in a multivariate model.
Bivariate correlations suggest that both flea beetles are contributing to the
observed decline in spurge density over time (Figure 2.3). However, it appears
that in actuality, during the time interval of this study, the biggest impact on
changes in spurge density was from self thinning, with A. lacertosa having a
modest effect on spurge and A. nigriscutis having no impact at all. This illus-
trates how careful one must be in interpreting bivariate correlations between
variables under common influence (in this case, by initial stem densities). Stem
densities appear to be a major driver in this system, both in terms of regulating
flea beetle populations and in terms of self regulation of spurge densities. It
also appears that there may be some negative associations between the two flea
beetles, which could have important implications for biocontrol efforts. Further
work, including carefully targeted experimental manipulations, can help us to
continue to refine our knowledge and our models about this system through an
accumulation of knowledge. I believe that this process of confronting systems
using structural equation models will enhance both our understanding of the
system and our ability to predict its behavior.
 PA RT I I

Basic principles of structural equation

modeling
 3
The anatomy of models I: observed
variable models

Overview of more complex models

The versatility of structural equation modeling is reflected in the wide variety
of models that can be developed using this methodology. To make the material
easier to understand, we will start with the simplest types of model, those that
only involve the use of observed variables. In later chapters we will introduce
abstract variables into our models. The inclusion of these other types of variable
greatly expands the variety of problems that can be addressed. As a preview
of things to come later, here I present an example of a more complex type of
model (Figure 3.1).
The model in Figure 3.1 includes four types of variables. Observed vari-
ables (represented by boxes) represent things that have been directly measured.
Examples of observed variables include the recorded sex of animals in a sample,
or the estimated plant biomass in a plot. Observed variables can also be used to
represent experimental treatments (e.g., predators excluded, yes or no?), spatial
locations (e.g., different sample sites), or interactions among other variables.
Latent variables (represented by circles) represent unmeasured variables, which
are often used to represent underlying causes. In the earliest example of a latent
variable path model, Wright (1918) hypothesized that the relationships amongst
bone dimensions in rabbits could be explained by a number of latent growth
factors. While these latent factors could not be directly measured, their effects
on bone dimensions could be inferred from the pattern of correlations amongst
observed variables. The ability to model latent variables has come to be an inte-
gral capability of SEM. As we shall see later, latent variables can be used for a
variety of purposes, and they do not always represent abstract concepts. A third
variable type also occurs in Figure 3.1, the composite variable (represented by
a latent variable with no indicators and zero error). A composite variable is one
that represents the collective effect of a group of variables. While composite

37
38 Structural equation modeling and natural systems

zm
da a
I M h eh
db b
0

N
dc c

dd d J K L
de e zk zl
f g

ef eg

Figure 3.1. Illustration of a structural model containing four different types of

variables. Boxes represent observed variables a–h, circles represent latent (un-
observed) variables I–M, and the circle with a zero disturbance (N) represents a
composite variable. The ζ term refer to variables that represent unspecified effects
on dependent latent variables (disturbance effects), while δ and ε are variables that
represent unspecified effects on observed variables. The zigzag arrow represents
that L was derived arithmetically from K.

variables have been mentioned in the SEM literature for quite some time, it
is only recently that they have begun to be considered more seriously for rou-
tine inclusion in structural equation models. Finally, error variables (which are
unenclosed, but nevertheless true variables), represent unknown or unspecified
effects on observed variables and latent response variables. Error variables are
usually presumed to represent a collection of effects, including both random
measurement error and unspecified causal influences (i.e., variables that were
not measured). Often, error variables are treated more like summary measures of
the other parameters (e.g., the sum of unexplained variance) than like variables
themselves.
It is not my intention that the model represented in Figure 3.1 will be com-
pletely understood by the reader at this time. Rather, my objective here is simply
to hint at the great versatility embodied in modeling with structural equations
and to point to some of the things that will come later. In the rest of this chapter
 The anatomy of models I: observed variable models 39

we will focus on models that only include observed variables. All of the princi-
ples we discuss related to observed variable models will also apply to the more
complex models that are considered later, when latent and composite variables
are introduced more fully.

Terminology associated with observed variable models

Models that include only observed variables (e.g., Figure 2.5) go by a variety of
names. We could ignore this fact and choose just a single terminology for use in
this book. Such an approach would not be helpful to the reader when trying to
connect the material presented here with the writings of other authors. There-
fore, its seems wise to mention the various terms that have been associated with
these models. This kind of model is variously known as a (1) structural equation
model with observed variables, (2) manifest variable model, or (3) path model.
Analysis of a model that only contains observed variables is frequently referred
to as path analysis. In the context of our historical discussion in Chapter 1,
this would be modern path analysis. I may variously use all of these terms to
refer to this type of model, depending on context.

Historic path analysis versus modern path analysis

While the topic was touched on in Chapter 1, it may be valuable to address
here a question I commonly encounter, Isn’t structural equation modeling with
observed variables the same as the path analysis commonly used by biolo-
gists? This is a bit of a tricky question. The reason it is tricky is because the
capabilities and accepted procedures for analyzing path models have evolved
over the decades. Often these evolutionary advancements are omitted from the
presentations of path analysis in the natural sciences. We will have opportuni-
ties to explain in more detail some of these advances in later sections of the
book. However, for the time being I will mention four things that may help to
distinguish “historic path analysis” from “modern path analysis”.
In historic path analysis, the raw materials for analyses were correlations. In
modern path analysis, the raw materials are covariances. It is now understood
that the analysis of correlations limits the kinds of inferences and comparisons
that can be made. This may seem like a minor point, but it is actually rather
fundamental. Standardized coefficients, such as correlations, are quite valuable
for extracting certain kinds of information. However, they tell somewhat less
than half the story. Comparisons to other data sets depend, in part, on the use
of unstandardized coefficients. The same can be said for the extrapolation of
40 Structural equation modeling and natural systems

findings to other situations or predictive applications. Furthermore, comparisons

among groups (say male and female members of an animal or plant population)
may require the use of unstandardized coefficients. Finally, modern path analy-
sis permits the analysis of differences in means as well as covariance relations.
All these capabilities are lost when only correlations are analyzed.
Another important difference is that in historic path analysis, path coeffi-
cients were only estimated using standard least squares methods of the same
sort as used in multiple regression. In modern path analysis, a number of solu-
tion procedures are available, but maximum likelihood estimation is often the
preferred method. For path models in which all pathways are estimated, the
results derived using standard regression methods are nearly the same as those
obtained using maximum likelihood (when sample sizes are large). However, if
a path model includes reciprocal effects (A affects B and B affects A), regression
gives incorrect estimates. Further, if latent variables are included in a model,
regression no longer suffices as a proper method for estimation. The topic of
estimation methods will be discussed in Chapter 5.
Related to the issue of the accuracy of path coefficients, is the evaluation of
overall model fit. In historic path analysis, fit of the data to the overall model was
not assessed. However, maximum likelihood methods as well as other modern
solution procedures provide for an additional assessment, that of whether the
data are consistent with the specified model overall (typically through a chi-
square or related test of goodness of fit). The inclusion of tests of overall model
fit is a major advance associated with modern path analysis.
In historic path analysis, particularly in the natural sciences, analyses have
often proceeded without regard for the distinction between exploratory versus
confirmatory analyses. There is certainly a place for exploratory path analysis.
This is particularly the case when a new problem is being tackled and there
is insufficient experience and/or theory to provide a solid starting point for
developing a-priori path models. However, within modern path analysis it is
recognized that the strength of inference possible is related to the degree to
which analysis proceeds in a confirmatory fashion. Thus, to the degree possible,
path analysis should be based on an evaluation of a-priori models and not
through a process of looking at all possible models, and then simply picking
the one the author likes best.
This brief discussion greatly oversimplifies the evolution of the meaning
of “path analysis” over time. A much more detailed and accurate portrayal
of the history of modern path analysis can be found in Tomer (2003). The
main point to be made here is that when path analysis is performed using the
modern approach, it proceeds with an awareness of the capabilities of SEM.
Thus, I tell those who ask, if one wants to perform a path analysis, they should
 The anatomy of models I: observed variable models 41

A B
x1 y2 x1 y2

z2 z2
y1 y1
z1 z1

C D
x1 y2 x1 y2

z2 z2

x2 y1 x2 y1

z1 z1
Figure 3.2. Examples of observed variable models. The observed variables are
included in boxes and may be connected by various types of arrows. The unenclosed
symbol ζ represents residual error variables.

read appropriate sections in a contemporary book on SEM (instead of a brief

treatment of the topic in a general statistics book). When approached in this
way, “path analysis” is synonymous with “SEM using observed variables”.

The anatomy of observed variable models

Graphical representation of models is a fundamental part of SEM. This is not
to say that the set of simultaneous equations associated with a model is un-
important. Rather, we should recognize that the visual representation of a model
in graphical form is such a powerful device for expressing structural equation
models that graphical models are nearly always developed by the researcher.
For observed variable models, the elements of graphical representation include,
(1) observed variables (in boxes), (2) double-headed curved arrows (represent-
ing unresolved correlations), (3) single-headed arrows (representing directed
relationships, and (4) residual variances. Typically the symbols representing
residual variances (error terms) are not enclosed by squares (see Figure 3.2).
There are several basic kinds of path model, depending on their structure and
interconnections. Figure 3.2 presents four examples to illustrate some of the
anatomical types. In the terminology of SEM, the independent (x) variables are
42 Structural equation modeling and natural systems

referred to as exogenous, while the dependent (y) variables are called endoge-
nous. Endogenous variables can, in turn, have influences on other endogenous
variables. Commonly, causal order will flow from left to right across a model,
although this will not always be the case.
The ways in which variables in a model are connected are important. Model
A in Figure 3.2 is a simple example where x1 and y2 are only related indirectly.
In other words, the relationship between x1 and y2 can be explained by their
relationships to y1 . Thus, we would say that x1 has an indirect effect on y2 (within
the context of the model). In contrast, model B contains both indirect and direct
paths between x1 and y2 . Direct paths are interpreted as relationships that cannot
be explained through any other relationships in the model. In the case of model
B, the correlation between x1 and y2 is such that it can only be explained by a
direct effect of x1 on y2 , in addition to the indirect effect mediated through y1 .
A further property of model B that should be noted is that it is saturated,
which means that all the possible interconnections are specified. Model A is
unsaturated, which means that some variables are not directly connected (i.e.,
some paths are omitted). Model C contains two exogenous variables, x1 and x2 ,
which are correlated in this case. Such a correlation is generally presumed to be
caused by some common influence that was not measured or is not considered
in this model. Model C also has the property of being unsaturated (the path
between x2 and y2 is omitted). Finally, model D contains reciprocal interactions
between y1 and y2 . Models with reciprocal interactions of this sort are called
nonrecursive. Thus, models without reciprocal interactions (e.g., model C) are
described as being recursive (the term recursive refers to the mathematical
property that each item in a series is directly determined by the preceding
item). In this chapter we will consider models of the first three types, A–C.
Models possessing reciprocal interactions require special care to interpret and
will not be addressed here.
One final property that an observed variable model may possess, and that is
not shown in Figure 3.2, is a correlated error. Such a condition would be rep-
resented by having a two-headed arrow between ζ 1 and ζ 2 , the two error terms,
in any of the models. Correlated errors typically represent some unspecified
cause for association. We will illustrate and discuss the meaning of correlated
errors later in this chapter.

Path coefficients
The path coefficients specify values for the parameters associated with path-
ways between variables. In essence, the path coefficients represent much of the
 The anatomy of models I: observed variable models 43

quantitative meaning of the model (although, as we shall see, there are many
other parameters of importance as well). There are a number of things we will
need to understand in order to properly interpret path coefficients. Some of the
basic concepts will be introduced first. Later we will relate this information to
our example in Chapter 2.

The building blocks of path coefficients: variances, covariances,

correlations, and regression coefficients
Structural equation modeling can include a variety of statistical procedures.
However, for most applications covered in this book, the procedures used are
based on the analysis of variance–covariance matrices. In addition to vari-
ances and covariances, an understanding of correlation coefficients is critical
because of their involvement in the interpretation of standardized path coef-
ficients. The material covered is likely to be familiar to most readers. How-
ever, these basic statistical issues are so critical to understanding SEM that
there is value in reviewing them at this time. Most importantly, this material
in presented so as to alert the reader to the connection between the seem-
ingly sophisticated procedures of SEM, and these most basic elements of
statistics.
We start with a review of the concepts of variance and covariance. The
formula for the best estimate of the population variance for x is

(xi − x̄)2
VARx = (3.1)
n−1
where xi refers to individual sample values, x̄ is the sample mean, and n is the
sample size. The sample size refers to the total number of observations in the
data set. The variance of y is given by the formula

(yi − ȳ)2
VAR y = (3.2)
n−1
The covariance between x and y is related to their joint deviations or cross
products divided by n − 1, or

(xi − x̄) (yi − ȳ)
COVx y = (3.3)
n−1
When x and y are in some raw metric, say density of plant stems/m2 for x
and density of insects/m2 for y, the magnitude of the covariance is related to
the scale used to express the raw values. This can be illustrated using the data
presented in Table 3.1.
44 Structural equation modeling and natural systems

Table 3.1. Illustration of calculations for sums of squares and sums of

cross products

i x y xi − x̄ (xi − x̄)2 yi − ȳ (yi − ȳ)2 (xi − x̄)(yi − ȳ)

1 2 70 −12.6 158.76 −77.3 5 975.3 973.98
2 6 55 −8.6 73.96 −92.3 8 519.3 793.78
3 9 50 −5.6 31.36 −97.3 9 467.3 544.88
4 12 156 −2.6 6.76 8.7 75.7 −22.62
5 15 115 0.4 0.16 −32.3 1 043.3 −12.92
6 15 200 0.4 0.16 52.7 2 777.3 21.08
7 19 155 4.4 19.36 7.7 59.3 33.88
8 20 202 5.4 29.16 54.7 2 992.1 295.38
9 22 295 7.4 54.76 147.7 21 815.3 1 092.98
10 26 175 11.4 129.96 27.7 767.3 315.78

146 1473 504.40 53 492.2 4 036.20

Note: means for x and y are 14.6 and 147.3 respectively.

For this example, the variances of x and y, plus the covariance between x
and y can be calculated using Equations 3.1–3.3 as follows:
504.40
VARx = = 56.04
9
53 492.2
VAR y = = 5 943.58
9
4 036.20
COVx y = = 448.47
9
This yields the following variance–covariance matrix
x y
x 56.04
y 448.47 5943.58
Here, we can see that the variance of y is larger than that of x. The mean of y
(147.3) is also larger than the mean of x (14.6). Because the scales for x and
y are so different, it is difficult to draw intuitive meaning from the covariance
between x and y. In order to put things in terms where we can easily assess
the degree of covariation between x and y, we might like to put x and y on a
common scale. Typically the z transformation is used to adjust the means of
variables to zero and their variances to 1.0. The formula for calculating z scores
for variable x is
xi − x̄
zi = (3.4)
SDx
 The anatomy of models I: observed variable models 45

Table 3.2. Illustration of derivation of z scores and

standardized cross products using data from Table 3.1.
SDx = 7.49 and SDy = 77.09

x y zx zy zx * z y
1 2 70 −1.68 −1.00 1.68
2 6 55 −1.15 −1.20 1.38
3 9 50 −0.75 −1.22 0.94
4 12 156 −0.35 0.11 −0.04
5 15 115 0.05 −0.42 −0.02
6 15 200 0.05 0.68 0.03
7 19 155 0.59 0.10 0.06
8 20 202 0.72 0.71 0.51
9 22 295 0.99 1.92 1.90
10 26 175 1.52 0.36 0.55

146 1473 −0.01 0.00 6.99

SDx , the standard deviation of x, is obtained by taking the square root of VARx .
The end result is that the z scores are expressed in terms of the standard deviation
for that variable.
Using the information in Table 3.2, we can calculate the Pearson product
moment correlation coefficient, rxy , using the formula

(z x × z y )
rx y = (3.5)
n−1
In this case, the formula yields
6.99
rx y = = 0.7767
9
Given that we have the calculated variances and covariance for the data in Table
3.1, we can also calculate the Pearson coefficient directly from the covariance
and standard deviations using the formula
COVx y
rx y = (3.6)
SDx × SD y
which gives the same result,
448.47
rx y = = 0.7767
7.49 × 77.09
As illustrated by this example, data can be standardized either through the use
of z scores, or simply by dividing the covariances by the product of the standard
deviations.
46 Structural equation modeling and natural systems

Finally, there is an important distinction between a correlation coefficient

(r) and a regression coefficient (b). In regression, one variable is considered
to be the response variable (y), and the other variable is considered to be the
predictor (x). The association between the two variables is used to generate a
predictor ( ŷ) based on the formula

ŷ = bx + a (3.7)

where b is the regression coefficient and a is the intercept. The relationship

between the regression coefficient relating x and y, byx , and the correlation
between x and y, rxy , can be represented by the formula

b yx = r x y (SD y /SDx ) (3.8)

It is a simple matter to see from Eq. (3.8) that when variables have been z-
transformed, SD y = SDx = 1, and byx = rxy . Thus, the distinction between
correlation and regression is only noticeable for the unstandardized coeffi-
cients. When dealing with unstandardized coefficients, the relationship between
COVx y , which measures association, and b yx is given by
COVx y
b yx = (3.9)
VARx
It can be seen from Eq. (3.9) that in regression the covariance is standardized
against the variance of the predictor (x), rather than the cross product of SDx
and SDy as in Eq. (3.6).

The rules of path coefficients

If we return to considering path models, as presented in Figure 3.2, we can
recognize two types of path, undirected and directed. Undirected paths, such
as those represented by curved two-headed arrows in models C and D, can be
described as unanalyzed relationships. This leads us to our first rule of path coef-
ficients – the path coefficient for an unanalyzed relationship (i.e., correlation)
between exogenous variables is simply the bivariate correlation (standardized
form) or covariance (unstandardized form) between the two variables. No cal-
culation is required, therefore, to find the coefficient, it can simply be taken
directly from the correlation/covariance table.
Our second rule of path coefficients states that when two variables are con-
nected by a single causal path, the coefficient for a directed path connecting
the variables is the (standardized or unstandardized) regression coefficient. Let
us illustrate this with a three-variable model of the form shown in Figure 3.3.
An associated matrix of correlations is presented to facilitate the illustration.
 The anatomy of models I: observed variable models 47

x1 y1 y2
----------------------------------------
x1 1.0
y1 0.50 1.0
y2 0.30 0.60 1.0

x1 y2
g11 b21
z2
y1
z1
Figure 3.3. Simple directed path model.

Also, for simplicity the discussion of path relations here will be based on stan-
dardized values (correlations). Following the second rule of path coefficients,
the coefficient for the path from x1 to y1 can be represented by the correlation
between x1 and y1 (0.50). Likewise, the coefficient for the path from y1 to y2
can also be represented by the correlation between y1 and y2 (0.60). As with the
first rule of path coefficients, when the second rule of path coefficients applies
and we are only concerned with standardized coefficients, we can simply pick
the coefficient directly out of the table of correlations.
Continuing with our example from Figure 3.3, we can ask, What is the
relationship between x1 and y2 ? This leads us to our third rule of path coef-
ficients, which states that the mathematical product of path coefficients along
a compound path (one that includes multiple links) yields the strength of that
compound path. So, in our example, the compound path from x1 to y2 is 0.50
times 0.60, which equals 0.30.
We can see from the table in Figure 3.3 that in this case, the observed
correlation between x1 and y2 is equal to 0.30, the strength of the compound
path from x1 to y2 . We can illustrate another major principle related to path
coefficients by asking what would it mean if the bivariate correlation between
x1 and y2 was not equal to the product of the paths from x1 to y1 and y1 to y2 ? An
48 Structural equation modeling and natural systems

Table 3.3. Variance/covariance matrix

x1 x2 y
x1 52 900
x2 21.16 0.0529
y 3 967.5 2.645 529

x1 y1 y2
----------------------------------------
x1 1.0
y1 0.55 1.0
y2 0.50 0.60 1.0

g21
x1 y2
g11 b21
z2
y1
z1
Figure 3.4. Model including dual pathways between x1 and y2 .

example of such a case is presented in Figure 3.4. Here we see that the indirect
path between x1 and y2 again equals 0.50 times 0.60, or 0.30. However, the
observed correlation between x1 and y2 is 0.50. This means that the bivariate
correlation between x1 and y2 cannot be explained by the indirect path through
y1 . Rather, a second connection between x1 and y2 (a direct path from x1 to y2 )
is required to explain their overall bivariate correlation.
By having a direct path between x1 and y2 , as shown in Figure 3.4, we now
have a new situation, one where y1 and y2 are connected by two pathways, one
direct from y1 to y2 and another that is indirect through the joint effects of x1
on y1 and y2 . This leads to the fourth rule of path coefficients, which states
 The anatomy of models I: observed variable models 49

that when two variables are connected by more than one causal pathway, the
calculation of partial regression coefficients becomes involved. We must be
clear here what is meant by a “causal pathway”. Certainly a directed path from
one variable to another, such as from y1 to y2 , is an obvious causal path. How
can it be that y1 and y2 are connected by two causal pathways? The answer is
that y1 and y2 are also connected through a joint causal influence exerted by
x1 . This indirect pathway, which can be traced from y1 to x1 and then to y2 (or,
alternatively, from y2 to x1 and then to y1 ) represents a second causal pathway,
one of association by joint causation. Thus, it is important that we should be
able to recognize when two variables are connected by more than one causal
pathway, being aware that some connections (discussed further below) will be
deemed noncausal. To understand this important issue further, we now need to
discuss partial regression coefficients.

Partial regression and its relation to path coefficients

Understanding the related concepts of partial regression and partial correlation
is necessary in order to understand path coefficients. Simply put, a partial
regression or partial correlation between two variables is one that accounts for
the influences of additional variables that affect or correlate with those variables.
As was just stated, partial regression or correlation coefficients become involved
when two variables are connected through more than one causal pathway. Let
us first begin with partial regression, using the data and model in Figure 3.4 for
illustration.
As we know from the second rule of path coefficients, the path coefficient
between x1 and y1 in Figure 3.4 is the same as the bivariate regression coefficient
between these two variables. This is true because x1 and y1 are only connected
through a single causal pathway. Since we are dealing in this example with
z-transformed variables (correlations), the regression coefficient is the same as
the correlation. Thus, without doing any calculations we already know that the
path coefficient between x1 and y1 has a value of 0.55.
It can also be seen in Figure 3.4 that the variables x1 and y2 are connected
by two pathways in this model, one is a simple (single link), direct pathway
and the second is a compound (multi-linkage), indirect pathway through y1 . As
we have said, both of these pathways represent causal connections whereby x1
affects y2 . The coefficient for the direct path from x1 to y2 is designated as γ 21 .
The value of γ 21 can be obtained from the partial regression of y2 on x1 , which
controls for the fact that y2 is also affected by y1 .
Note that I use the customary SEM symbology in this book. According to
this symbology, which will be presented in the context of the LISREL system
50 Structural equation modeling and natural systems

later (Chapter 6, Appendix 6.1), γ represents the effects of exogenous variables

on endogenous variables, and β represents the effects of endogenous variables
on other endogenous variables. Subscripts refer to appropriate elements. Thus,
γ 11 represents the effect on y1 by x1 and β 21 represents the effect on y2 by y1 .
The value of the coefficient γ 21 (effect on y2 by x1 ) can be obtained from the
formula for a partial regression coefficient.
r x y − (r x1 y1 × r y1 y2 )
γ21 = 1 2 (3.10)
1 − r x21 y1
Because Eq. (3.10) is so important to our understanding of path relations, it
may be worth pointing out some of its features. The numerator represents the
total correlation between x1 and y2 minus their indirect connection. This makes
intuitive sense because the total correlation between two variables equals the
sum of all pathways connecting them. The denominator in Eq. (3.10) can be
intuitively understood as a standardization term. In this case, the path coefficient
from x1 to y2 is standardized by the degree to which variance is shared between
x1 and y1 . When variance is not shared between x1 and y1 (i.e., x1 and y1 are
uncorrelated), the effect of x1 on y2 is independent of any effect from y1 on y2 .
However, when variance is shared between x1 and y1 (i.e., they are correlated),
that shared variance is removed from our calculation of the partial regression
coefficient by subtracting it from 1. What is most important about Eq. (3.10) is
that it has the property of being able to yield path coefficients which have the
property that they allow the sum of all paths between two variables (e.g., x1 and
y2 ) to equal the bivariate correlation between x1 and y2 . It is from this property
that Eq. (3.10) is derived (see Appendix 3.1 for the derivation).
In our example from Figure 3.4, we can calculate the partial regression
coefficient for the effect of x1 on y2 as
0.50 − (0.55 × 0.60) 0.17
γ21 = = = 0.24
1 − 0.552 0.6975
We can now say that the standardized effect of x1 on y2 when the effects of y1
on y2 are held constant is 0.24. This means that for every standard deviation
of change in x1 , y2 changes by 0.24 standard deviations. It is worth noting that
we have just obtained statistical control of the relationships. This is in contrast
to experimental control where y1 would be physically held constant and not
allowed to vary. While experimental control is used to force predictors to be
orthogonal, allowing for an evaluation of their separate influences, statistical
control permits us to evaluate the effects while permitting predictors to inter-
correlate. I will have much more to say about the relative merits of statistical
and experimental control in one of the later chapters, as this is a very important
concept in our consideration of multivariate hypotheses.
 The anatomy of models I: observed variable models 51

0.24
x1 y2
0.55 0.47 0.77
z2
y1
0.84
z1
Figure 3.5. Model from Figure 3.4 showing path coefficients.

Now, there is another pair of variables to consider in Figure 3.4, y1 and y2 .

In order to determine the type of path coefficient (simple or partial) that will
represent this path, we again have to ask the question, are these two variables
connected by more than one causal pathway? One causal connection is obvious,
the direct arrow from y1 to y2 . You should now recognize that they are also
causally connected by the fact that x1 affects both of them. Because y1 and
y2 are both affected by a common cause, the effect of x1 , they are causally
connected through a second path, one that spans between them through x1 .
Because the connections between y1 and y2 match the conditions for the
fourth rule of path coefficients, we can calculate the coefficient for the path
from y1 to y2 using the proper modification of Eq. (3.10)
r y1 y2 − (r x1 y1 × r x1 y2 )
β21 = (3.11)
1 − r x21 y1

which gives us
0.60 − (0.55 × 0.50) 0.325
β21 = = = 0.47
1 − 0.55 2 0.6975
Using the path coefficients derived from the correlations in Figure 3.4, we
can now represent the quantitative characteristics of the model as shown in
Figure 3.5. It is sometimes the convention that the width of the arrows is made
proportional to the magnitude of the coefficients to make the representation of
overall relationships more apparent. We should always keep in mind, however,
that the interpretation of the magnitude of a path coefficient requires care, as
will be made clear later in this chapter. As the reader will note, we have two
additional paths in Figure 3.5 that have not been discussed; those from the error
variables to y1 and y2 . More needs to be stated about these.
52 Structural equation modeling and natural systems

Error variables and their path coefficients

In observed variable path models, such as the ones we have been discussing,
typically we ignore the existence of error terms for the observed exogenous
variables. Actually, we are making the claim that there is no error associated
with exogenous (predictor) variables and these variables are measured perfectly.
We will discuss the consequences of this assumption in the next chapter. For
now, we should consider what error terms are, why we describe them as error
variables, and what the path coefficients representing error effects mean.
Stated in simple terms, the coefficients representing the errors are simple
regression coefficients representing the unexplained (residual) variance. If R 2y1
refers to the variance in y1 explained by x1 , then the path coefficient for the
effect of ζ 1 is

ζ1 = 1 − R 2y1 (3.12)

For our example in Figure 3.5, the value of ζ1 equals 0.84. Sometimes the path
coefficients for the error terms are confusing for those first encountering them.
Since the value for R 2y1 = 0.552 = 0.30, we might think the path coefficient
would be the unexplained variance, which is 0.70. However, the path coefficient
from ζ 1 is typically expressed in the same form as for other path coefficients,
the non-squared form (the square root of 0.70 = 0.84).
In a similar fashion, R 2y2 refers to the variance in y2 explained by the combined
effects of y1 and x1 , which in this case is 0.40 (the means of calculating this by
hand is presented below in Eq. (3.16). Given the explained variance of 0.40, we
can calculate the path coefficient as

ζ y2 = 1 − R 2y2 (3.13)

which equals 0.77. Equations (3.12) and (3.13) represent the fifth rule of path
coefficients, that the coefficients associated with paths from error variables are
correlations or covariances relating the effects of error variables. In this case,
the ζ values represent the errors of prediction, though without specifying the
exact causes for that error.

Alternative ways to represent effects of error variables

While the above means of representing the effects of error variables using path
coefficients for error paths is often presented in textbooks, there are alternative
methods that may be preferred by some. Typically we think of error associated
with response variables in terms of variance explained and the related concept
of the R2 . Standardized path coefficients for error effects of the sort shown in
 The anatomy of models I: observed variable models 53

0.24
x1 y2
0.55 0.47
0.59
y1

0.70
Figure 3.6. Alternative representation of effects of error variables (compare to
Figure 3.5).

Figure 3.5, while logical, still require some mental arithmetic to transform the
information into variance explained. Instead of presenting the path coefficients
for error variables, it is possible to present the standardized values of the error
itself, as shown in Figure 3.6. Here, we can directly obtain R2 as 1 minus the
standardized error. Note that in Figure 3.6, the numeric value is located at the
origin end of the arrow, rather than alongside the arrow (the convention for
path coefficients, compare to Figure 3.5). Also, for some audiences, simply
presenting the R2 directly may be preferred over the representation of error
variables.

Partial correlation and its relation to path coefficients

There is one more type of path coefficient I would like to discuss and that
is the partial correlation. Figure 3.7 illustrates a model that contains a partial
correlation (between ζ 1 and ζ 2 ). As we have already stated, according to the first
rule of path coefficients, the coefficients for unanalyzed relationships between
two exogenous variables are simply the correlations or covariances (e.g., x1
and x2 ). So, what about the situation where there are two response variables,
in this case y1 and y2 , that are correlated due to some unspecified reason(s)?
Our sixth rule of path coefficients states that unanalyzed correlations between
endogenous variables represent partial correlations or partial covariances.
The formula for a partial correlation between y1 and y2 in the presence of x1
in Figure 3.7 is
r y y − (r x1 y1 × r x1 y2 )
r y1 y2 •x1 =  1 2   (3.14)
1 − r x21 y1 1 − r x21 y2
54 Structural equation modeling and natural systems

z21 y2 z2

y12
x1
g11 z1
y1
Figure 3.7. Model illustrating correlations between endogenous variables (y1 and
y2 in this case), which are manifested as correlations between the errors.

where r y1 y2 •x1 refers to the partial correlation between y1 and y2 taking into
account their correlations with x1 . The reader should note that technically, a
model with correlated errors between endogenous variables is nonrecursive,
in that there is something akin to a reciprocal interaction between endogenous
variables. In LISREL terminology the correlated error is referred to as ψ 12 (psi
one–two).
Now, for a model such as the one in Figure 3.7, but with no partial correlation
between y1 and y2 , the intercorrelation between the two response variables
would simply be the product of the correlation between x1 and y1 and the
correlation between x1 and y2 . This would be represented by the formula

r y1 y2 = r x1 y1 × r x1 y2 (3.15)

If Eq. (3.15) were substituted into Eq. (3.14) the numerator of Eq. (3.14) will be
zero. Therefore, it is a general rule that under the conditions where Eq. (3.15)
holds, the partial correlation between y1 and y2 will be zero, and y1 and y2 are
said to be conditionally independent. If the relationship in Eq. (3.15) fails to
hold, there will then exist a direct path between y1 and y2 , and its path coefficient
will be the value of the partial correlation given by Eq. (3.14).
Let us consider again the example data in Figure 3.4. If y1 and y2 were
conditionally independent, according to Eq. (3.15), their bivariate correlation
should be 0.55 × 0.50 = 0.275. However, their observed correlation is actually
0.60; this seems to be a rather large difference. According to Eq. (3.14), we can
calculate the partial correlation between y1 and y2 in the presence of x1 as
0.60 − (0.55 × 0.50)
r y1 y2 •x1 =  = 0.45
(1 − 0.552 )(1 − 0.502 )
which represents the value for ψ 12 . So, we can see that there is a rather large
correlation between y1 and y2 that is not explained by their joint dependence on
x1 . As we stated earlier, this correlation between dependent variables suggests
 The anatomy of models I: observed variable models 55

x1 x2 y1 y2
---------------------------------------------------
x1 1.0
x2 0.80 1.0
y1 0.55 0.40 1.0
y2 0.30 0.23 0.35 1.0

0.15
x1 y2

0.80 0.64 0.27 z2

-0.11
x2 y1
z1
Figure 3.8. Model and correlations illustrating concepts of total effect and total
correlation. The matrix represents the bivariate correlations among variables, while
the numbers on the diagram are the path coefficients.

the influence of some other factors that have not been explicitly included in the
model.

Total effects and total correlations

Once we have values for individual paths, we can consider how the different
paths between variables add up. The seventh rule of path coefficients states
that the total effect one variable has on another equals the sum of its direct
and indirect effects. In this case, the total effect of x1 on y2 = 0.24 + (0.55 ×
0.47) = 0.50, which in this case, is the bivariate correlation between x1 and y2 .
Related to the seventh rule is the eighth rule of path coefficients, which states
that the sum of all pathways connecting two variables, including both causal
and noncausal paths, adds up to the value of the bivariate or total correlation
between those two variables. To illustrate this rule, we need to involve a slightly
more complex model, for example of the form of model C in Figure 3.2. Figure
3.8 provides correlations and path coefficients for such a model.
Figure 3.8 offers us a slightly more complex example. Here, we have an
undirected relationship between two exogenous variables (x1 and x2 ), as well as
56 Structural equation modeling and natural systems

multiple directed paths to and among endogenous variables (y1 and y2 ). The first
thing we should notice when comparing the path coefficients to the bivariate
correlations is that all directed paths in this case involve partial regression
coefficients. Only the coefficient for the undirected path between x1 and x2 can
be found in the table of correlations. The reader should be able to calculate all
of the other path coefficients shown using Eq. (3.10).
As for total effects, these can be obtained from the path coefficients. Our
seventh rule of path coefficients allows us to determine that the total effect of
x1 on y1 is simply 0.64. This means that if we were to increase the value of x1
by one standard deviation while holding the value for x2 constant at its mean
value, the value of y1 would increase by 0.64 times its standard deviation. As
we can see, there is only one directed path connecting x1 with y1 . However,
the case for the total effect of x1 on y2 is different. Here, there are two directed
pathways, one simple path from x1 to y2 and one compound path through y1 .
Therefore, the total effect of x1 on y2 is 0.15 + (0.64 × 0.27), or 0.32. Again,
if we were to increase x1 one standard deviation while holding x2 constant, y2
would increase by 0.32 times its standard deviation. In this case, however, y1
would covary in the process because the total effect of x1 on y2 would involve
both paths from x1 to y2 . Finally, the total effect of x2 on y2 can be seen to be
−0.11 × 0.27 = −0.03.
For more complex models, the business of determining which of the paths
connecting two variables are causal can be slightly more tedious than in the
simple example given here. Wright (1960) proposed several simple tracing
rules that can help to identify causal paths. These rules are as follows:

(1) A path can go backwards as many times as necessary before going forward.
(2) Once a path goes forward, it cannot go backwards.
(3) A path cannot go through a variable twice.
(4) A path can include a curved arrow, but only one curved arrow per path.

Reflecting on these tracing rules should reveal that they are perfectly compatible
with our discussion of what constitutes a causal connection. Perhaps the least
transparent of these rules is number 4. To understand rule 4, we should keep in
mind that a curved arrow represents an unresolved shared cause. Thus, a curved
arrow between x1 and x2 actually represents an undescribed variable with arrows
pointing at x1 and x2 . So, when a path proceeds backwards (upstream) through a
curved arrow, it should be viewed as going upstream to the undescribed variable
and then forward (downstream) to the variable on the other end of the curved
arrow. If we were to consider a path that included two curved arrows, we would
be violating rule number 2.
 The anatomy of models I: observed variable models 57

As for the total correlations, the calculation of these involves the eighth rule
of path coefficients. Thus, total correlations involve both the total effects plus
the undirected relations between variables. In our example in Figure 3.8, the
total correlation between x1 and y1 is the sum of the directed and undirected
pathways linking them. In this case, that is 0.64 + (0.80 × −0.11) = 0.55. Note
that this reconstitutes the bivariate correlation between x1 and y1 as seen in
Figure 3.8. Likewise, the total correlation between x2 and y2 can be seen to be
(−0.11 × 0.27) + (0.80 × 0.15) + (0.80 × 0.64 × 0.27) = 0.23.
As a final point of discussion relating to Figure 3.8, the reader may be
surprised to find that the correlation between x2 and y1 , which is 0.40, is sub-
stantially different from the path coefficient between these two variables, which
is −0.11. When such a case is observed, it is often referred to as suppression.
Suppression refers to the fact that the strong intercorrelation between x1 and x2 ,
in combination with the relatively strong effect of x1 on y1 , causes the effect of
x2 on y1 to be fairly unrelated to their net intercorrelation. Keep in mind that the
path coefficient of −0.11 does actually reflect the causal effect that x2 has on
y1 . However, this effect is rather weak, and as a result, the correlation between
x2 and y1 is dominated by the influence of x1 .

The interpretational differences between unstandardized and

standardized path coefficients
The analysis of unstandardized versus standardized data
As is the case with covariances versus correlations, path coefficients and other
model parameter estimates can be expressed in either unstandardized or stan-
dardized form. There are two very important things to mention at the outset.
First, many of the commercial software programs available for conducting SEM
do not provide proper solutions for analyses based on correlation matrices.
Structural equation modeling based on covariance procedures is based on the
presumption that the unstandardized relationships (variances and covariances)
are being analyzed. At the time of writing, only some of the available software
packages (e.g., SEPATH, Mplus) provide the necessary adjustments to provide
correct solutions for correlation matrices. Secondly, when covariance matrices
are analyzed, significance testing of parameters is based on the standard errors
of the unstandardized coefficients. Thus, any standardized path coefficients
that may be presented (for example, by requesting standardized output) have
not been directly tested for significance. Usually, standardized coefficients are
interpreted based on the test results for unstandardized parameters. However,
care must be taken when critical results are marginal and inferences are based
on statements of significance about standardized results.
58 Structural equation modeling and natural systems

A basic issue to consider is the different ways in which these two types of
variable are interpreted and the kind of questions each addresses. Typically,
researchers are fond of standardized parameters because the scale is the same
(standard deviation units) for different relationships. Thus, the primary function
of standardized variables is to allow direct comparisons between relationships
that are measured on different scales. For example, the unstandardized param-
eter relating an animal’s fat deposits to its caloric intake might be measured
as percent body fat per kilocalories. Meanwhile the unstandardized parameter
relating body fat to activity level (e.g., kilometers traveled during migration)
would be on a different scale. In this case, standardized coefficients allow for
a common basis of comparison. As we shall see, however, it is easy to rely too
heavily on standardized path coefficients and an appreciation for unstandardized
coefficients needs to be developed by the user of SEM.
The process of partitioning path coefficients was originally developed using
Pearson product moment correlations, which are standardized parameters. Fur-
ther, many authors choose to introduce the basic concepts about path coeffi-
cients using standardized variables (as I have done above). Sewell Wright, in
his initial work on path analysis, relied exclusively on standardized data and
standardized coefficients to partition direct and indirect relationships among
variables. However, he later came to calculate and report “concrete” (unstan-
dardized) coefficients also, and discussed their use in interpreting path models
(Wright 1984). Often, applications of path analysis in the natural sciences have
relied strictly on the analysis of correlations instead of covariances. In modern
path analysis, as well as in most other forms of structural equation modeling,
we typically perform analyses based on covariances rather than correlations.
Analysis of covariances permits calculation of both unstandardized and stan-
dardized results. It also permits analysis of differences between means as well.
For this reason, it is recommended that SEM be performed using unstandard-
ized data (covariances). It is then possible to calculate both unstandardized and
standardized parameter estimates, and both can be used in drawing interpreta-
tions.

Differences between unstandardized and standardized coefficients

Here I wish to illustrate more explicitly using hypothetical examples why a
complete interpretation of a structural equation model requires the researcher to
examine both unstandardized and standardized coefficients. Later in the chapter
we will compare standardized and unstandardized coefficients for results from
a real example, our study of plants and insects presented in Chapter 2. The
 The anatomy of models I: observed variable models 59

100 100
A B
80 80

60 60
y y
40 40

20 20
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
x x

Figure 3.9. Illustration of two regressions having the same intercept and slope,
but different degrees of scatter around the line. While unstandardized parameters
(intercept and slope) did not differ significantly between the two data sets, the
degree of correlation did, with correlation coefficients of 0.861 in A and 0.495 in B.

reason I emphasize this point rather strongly is because an inappropriate choice

of which coefficients to use can lead to incorrect interpretations (see Grace and
Pugesek (1998) for discussion of a specific ecological example). Here I illustrate
the general issue and use a simple regression example to show what is gained
by having both standardized and unstandardized coefficients (Figure 3.9). At
the end of the discussion, I present some cautionary points about standardized
parameters, and suggest a possible solution that involves an alternative method
of standardization.
The data presented in Figure 3.9 were generated using the equation y =
a + γ x + ζ , with only the value of ζ , the error term, differing between the
two cases, A and B. A random selection of x values were obtained from a nor-
mal distribution and ζ was independent from the other parameters, had a mean
of zero, and a specified standard deviation. Slight, nonsignificant differences
between the unstandardized parameters in A and B are the result of random
sampling (sample size = 100). However, the standardized coefficients (corre-
lations) differ strongly between cases, being 0.861 for case A and 0.495 for
case B.
As this example demonstrates, unstandardized coefficients represent the
effect of x on y in absolute terms. Since the slope here is 2.0 for the population,
we can state that for a change in 10 units in x, there should be a corresponding 20
units change in y on average. This statement holds true for both cases, A and B.
Now, if we took the data shown in Figure 3.9 A and B and z-transformed
the values, the resulting graphs would look virtually identical to each other,
except that the scales for the x and y variables would have mean zero and
be measured in units of the standard deviations of x and y. If we perform a
regression on these z-transformed variables, we now find that standardized
and unstandardized coefficients are the same and they match the correlation
60 Structural equation modeling and natural systems

coefficients given in Figure 3.9. Thus, we have lost any knowledge of the
absolute slope of the relationships using standardized data. The same can be
said when we rely exclusively on standardized coefficients. Thus, a knowledge
of the absolute slope of the relationship is of critical importance in drawing
certain kinds of inferences, such as predictions, comparisons across samples,
and generalizations at large.

The interpretation of standardized coefficients

In the face of the above warnings, we must still recognize that standardized
coefficients are often valuable to us. However, they are frequently misinter-
preted. The biggest cause of misinterpretation may come from a spill over
from our training in analysis of variance; the concept of variance partitioning.
When exogenous variables are uncorrelated with each other, such as is typi-
cally the case in controlled experiments, variance in the response variable can
be apportioned amongst the causal factors. However, when exogenous vari-
ables are intercorrelated, variance partitioning becomes a limiting concept (as
we shall see in a moment). For this reason, standardized path coefficients are
best thought of as the expected change in y as a function of the change in x, in
standard deviation units.
Now, researchers often use standardized coefficients in an attempt to address
a different kind of question. Instead of discussing the expected change in y,
they may try to make statements about how much of the variation in y in a
sample can be associated with the variation in x in that sample. What, they
may ask, causes the variations seen among my sample points in the number of
bird species observed? Are factors x1 and x2 equally important explanations for
those observed variations? The reader should note that this kind of question is
primarily about the data in hand rather than about the phenomenon at large. The
reason for this being about the data in hand is that the question is about observed
variations in the sample. In a different sample, the span of values may be more
or less, and the standardized relationships would not be constant. Furthermore,
this question can only be answered unambiguously in certain types of path
models, those devoid of correlated exogenous variables.
Why are we so readily misled about the interpretation of standardized coef-
ficients? The usual culprit is that we make an analogy with simple regression,
where there are no correlated predictors. For example, in the case of simple
regression, we know that the square of the correlation (which is the same as the
standardized regression coefficient) provides a measure of the variance in y that
is explained by x. For the examples in Figure 3.9, even though the unstandard-
ized effect of x on y is the same in both A and B, the variance in y explained by
 The anatomy of models I: observed variable models 61

Table 3.4. Unstandardized path coefficients

x1 x2 y
x1 –––––
x2 21.16 –––––
y 0.065 23.81 –––––

x differs between the two cases. In A, x is able to explain 74% of the variation
in y in the sample. In B, x is able to explain only 25% of the variation in y.
In multiple regression or a path model, relating the variance explained to
different predictors is more complicated. To consider how multiple predictors
relate to variance explanation in y, we need the formula for calculating total
variance explained in a multiple regression. When there are two correlated
predictors, x1 and x2 , influencing y, the total variance explained in y, R 2y can be
calculated as

R 2y = γ11 (r x1 y ) + γ12 (r x2 y ) (3.16)

In this context, γ11 refers to the standardized effect of x1 on y and γ12 refers
to the standardized effect of x2 on y. Both standardized path coefficients are
partial regression coefficients in this case. Equation (3.16) gives the impression
that we can apportion cleanly variance explanation in y among the xs. This is
not the case.
Another hypothetical case can be used to illustrate the application of Eq.
(3.16). Imagine a situation where y is influenced by two correlated predictors,
x1 and x2 , and the scale of y is in terms of percent change and ranges from 1 to
100. In this example, the scale of x1 is in raw units and ranges from 1 to 1000,
while the scale of x2 is as a proportion and ranges from 0 to 1.0. For simplicity
we let the standard deviation of each be a constant proportion of the range for
each variable. In this example, the covariance matrix has the values given in
Table 3.3.
If we analyze this variance/covariance matrix and present unstandardized
path coefficients, we get the values represented in Table 3.4.
Thus, our equation is

y = 0.065x1 + 23.81x2 + ζ

It is unclear from this equation what the relative importance of x1 and x2 might
be as explanations for the observed variation in y. Now, if we look at the data
in standardized form (Table 3.5),
62 Structural equation modeling and natural systems

Table 3.5. Correlations, standard deviations

x1 x2 y
x1 1.00
x2 0.40 1.00
y 0.75 0.50 1.00
STD 230 0.23 23

we see the correlations among variables, along with the standard deviations
(note, Table 3.3 can be exactly reconstructed from Table 3.5). Looking at the
standardized path coefficients (Table 3.6),

Table 3.6. Standardized path coefficients

x1 x2 y
x1 –––––
x2 0.4 –––––
y 0.655 0.238 –––––

the equational form or representation would be

y = 0.655x1 + 0.238x2 + ζ

We can now use Eq. (3.16) to calculate the total R2 for y.

R 2 = (0.655)(0.75) + (0.238)(0.5)

Solving this in stages, we see that

R 2 = (0.49) + (0.12) = 0.61

Thus, the total proportion of variance in y explained by both predictors is

0.61.
As stated earlier, this calculation of the total variance explained by both
predictors might suggest to the reader that we can cleanly assign variance
explanation to each of the two predictors in this case. However, we must recog-
nize that the explained variance is of three types, (1) variance in y that is only
associated with x1 (x1 ’s unique variance explanation), (2) variance in y that is
only associated with x2 (x2 ’s unique variance explanation), and (3) variance in
 The anatomy of models I: observed variable models 63

y that is explained by both x1 and x2 . The magnitude of the shared variance is

proportional to the magnitude of correlation between x1 and x2 .
It is possible to calculate these three components of the explained variance
using the formula for a semipartial correlation. The semipartial correlation
between x1 and y in the presence of x2 can be found by the equation
r yx1 − (r x1 x2 × r yx2 )
r yx1 (x2 ) =  (3.17)
1 − r x21 x2

The square of the semipartial correlation in Eq. (3.17) gives a measure of the
proportion of the variance in y that is uniquely explained by x1 . The meaning
of a semipartial correlation or regression can be understood to be the additional
variance explanation gained by adding x1 to the collection of predictors (akin
to a stepwise process). Using Eq. (3.17) and the data from Table 3.5, we can
calculate for our example that
0.75 − (0.40 × 0.50) 0.55
r yx1 (x2 ) = √ = = 0.60
1 − 0.40 2 0.92
Thus, the proportion of variance in y uniquely explained by x1 = 0.602 = 0.36.
Because the maximum variance in y that could be explained by x1 if it was the
only predictor is the square of the bivariate correlation, 0.752 = 0.56, and the
variance in y uniquely explained by x1 is 0.36, the shared variance explanation
must be 0.56 − 0.36 = 0.20.
We can further determine using the same procedure (an equation of the
same sort as Eq. (3.17)) that the proportion of variance in y uniquely explained
by x2 = 0.222 = 0.05. As before, the maximum variance in y that could be
explained by x2 if it were the only predictor is 0.502 = 0.25. Since x2 ’s unique
explanatory power is 0.05, we once again arrive at a shared variance explanation
of 0.20. These results are summarized in Table 3.7.
Having gone through the process of showing how variance explanation can
be partitioned, I must again emphasize that this is only relevant when we are
interested in associations in our sample. If we wish to extrapolate to another
situation or compare groups, standardized coefficients will not serve us well
unless the variances are comparable across samples. The reason is very simple.
The standardized coefficients are standardized by the ratios of the standard devi-
ations, and thus, sensitive to the sample variances. Reference back to Eq. (3.8)
reminds us that the standardized effect of x on y can easily be calculated as the
unstandardized effect multiplied by the ratio of the standard deviations (when
this is done, the units cancel out, leaving the coefficients unitless, though we
interpret them as being in standard deviation units). Since we know that further
64 Structural equation modeling and natural systems

Table 3.7. Summary of explained variance components

Proportion of variance
Source of variance explanation in y explained
Unique influence of x1 0.36
Unique influence of x2 0.05
Shared influences of x1 and x2 0.20
Total variance explained 0.61

samples may not always have the same variances, the standardized coefficients
will not automatically extrapolate. To generalize beyond our sample, there-
fore, we should rely on unstandardized coefficients. Pedhazur (1997, page 319)
describes the situation this way, “The size of a [standardized coefficient] reflects
not only the presumed effect of the variable with which it is associated but also
the variances and the covariances of the variables in the model (including the
dependent variable), as well as the variances of the variables not in the model
and subsumed under the error term. In contrast, [the unstandardized coefficient]
remains fairly stable despite differences in the variances and the covariances of
the variables in different settings or populations.” Pedhazur goes on to recom-
mend that authors present both unstandardized and standardized coefficients,
and to be particularly careful to respect the assumptions involved in standard-
ized coefficients.

Guidelines
As a generalization, there are four kinds of conditions in which unstandardized
coefficients are likely to be essential to drawing proper inferences: (1) when
the units of x and y are meaningful and there is interest in describing the abso-
lute effect of x on y, (2) when separate path models for different groups are
being compared (so-called multigroup analysis) and those groups have differ-
ent variances, (3) in repeated measures situations where items are followed
over time, and population variances differ substantially between times, and
(4) when results are compared among datasets or used for prediction and we are
interested in average changes. While some authors will state that under such
circumstances only the unstandardized coefficients can be used, I think that
this is not strictly true. The focus of the research question is what determines
when and how the coefficients should be interpreted. The most general rule
of all about standardized and unstandardized coefficients is that researchers
 The anatomy of models I: observed variable models 65

should use the coefficient type that is appropriate to the inferences they
make.

Summary of the rules of path coefficients

Because I have scattered these important rules through the chapter, I recapit-
ulate them here for the reader. As we will discuss other types of variables in
later chapters (e.g., latent and composite variables), we will introduce some
additional rules that apply to models that include those variable types.

The rules of path coefficients (note, these are all stated in terms
of standardized coefficients, though they also apply to
unstandardized coefficients)
(1) The path coefficients for unanalyzed relationships between
exogenous variables are the bivariate correlations.
(2) When two variables are only connected through a single directed path,
the coefficient for that path corresponds to the bivariate regression
coefficient.
(3) The strength of a compound path (one that involves multiple arrows)
is the product of the coefficients along that path.
(4) When two variables are connected by more than one causal pathway,
the calculation of partial regression coefficients is required.
(5) Coefficients associated with paths from error variables are
correlations representing unexplained effects.
(6) Unanalyzed correlations between endogenous variables are
represented by partial correlations.
(7) The total effect one variable has on another equals the sum of its
direct and indirect effects through directed (causal) pathways.
(8) The sum of all pathways connecting two variables, including both
causal and noncausal paths, adds up to the value of the bivariate or
total correlation between those two variables.

Interpreting the path coefficients of the leafy spurge–flea

beetle example
We now return to the results from our example in Chapter 2 dealing with the
interactions between the plant leafy spurge and the two flea beetles that feed
on it. The diagram (Figure 3.10) now shows both the standardized coefficients,
66 Structural equation modeling and natural systems

R 2 = 0.26 R 2 = 0.61

A. nigriscutis 0.62(0.60) A. nigriscutis

2000 2001
0.23(0.34)
0.51(0.77) -0.14(0.12) 0.09(2.72)

-0.20 -0.56(-8.96) Change in

Number of
(-0.66) stems
stems 2000
2000-2001 R 2 = 0.44
ns 0.21(0.35)
-0.22(-2.2)

A. lacertosa 0.67(0.63) A. lacertosa

2000 2001

R 2 = ns R 2 = 0.53
Figure 3.10. Results for leafy spurge–flea beetle model showing both standard-
ized coefficients (not enclosed by parentheses) and unstandardized coefficients
(enclosed by parentheses).

which were given in Figure 2.5, and the unstandardized path coefficients (in
parentheses). First, we will re-examine the standardized coefficients in this
model while reviewing the rules of path coefficients. Secondly, we will consider
the interpretation of the unstandardized coefficients.

Application of the rules of path coefficients to our example

After considering the rules of path coefficients in this chapter, the reader should
be better able to understand the meaning and origin of the various path coeffi-
cients in Figure 3.10. To help with the process, I reproduce the correlations from
Table 2.1 in Table 3.8. Here are some of the main points that can be illustrated
from Table 3.8 and Figure 3.10 relating to standardized coefficients.

(1) According to the first rule of path coefficients, the path coefficient for an
unspecified relationship (two-head arrow) between exogenous variables
will take on the value of their bivariate correlation. In this case, there is
only one exogenous variable, the number of stems in 2000. Therefore this
rule does not apply to any of the coefficients in Figure 3.10.
(2) According to the second rule of path coefficients, when two variables are
only connected through a single causal pathway, the value of the path
coefficient between them will be the bivariate regression coefficient. For
 The anatomy of models I: observed variable models 67

Table 3.8. Simple bivariate correlations among variables in Figure 3.10

stems00 change nig00 nig01 lac00 lac01

stems00 1.00
change −0.63 1.00
nig00 0.51 −0.28 1.00
nig01 0.52 −0.20 0.75 1.00
lac00 0.15 −0.24 −0.12 −0.18 1.00
lac01 0.31 −0.39 0.03 0.03 0.70 1.00

standardized variables, this will be the same as the correlation coefficient.

In Figure 3.10, there is one path that fits the requirements of the second rule,
the path from number of stems in 2000 to A. nigriscutis (0.51). Reference
to Table 3.8 illustrates that the path coefficient for that path does indeed
match the bivariate coefficient.
(3) According to the third rule of path coefficients, the product of path coeffi-
cients along a complex path represents the strength of that complex path.
This rule applies throughout our model. For example, the effect of the num-
ber of spurge stems in 2000 on A. nigriscutis in 2001 (in standardized terms)
is 0.51 × 0.62 = 0.32.
(4) The fourth rule of path coefficients indicates that there are many cases
where partial regression coefficients are involved; basically anywhere two
variables are connected by two or more causal pathways. The number of
spurge stems in 2000 has three causal pathways to the change in stems from
2000 to 2001. The number of stems also has three causal paths to the density
of A. nigriscutis in 2001, while it has two causal paths to A. lacertosa in
2001.
(5) In this case, I have not presented the path coefficients for the error variable
effects. The fifth rule of path coefficients can be used to determine those
values. Since the path coefficients for error terms are directly related to the
explained variance for endogenous variables, we can determine the path
coefficients for the error variables using Eq. (3.12) if we wish to know
them.
(6) According to the sixth rule of path coefficients, correlations between
endogenous variable errors are partial correlation coefficients. There are
two of these in this model, one involves A. nigriscutis in 2000 and A. lac-
ertosa in 2000. The other partial correlation is between the errors of A.
nigriscutis in 2001 and the change in stems from 2000 to 2001.
68 Structural equation modeling and natural systems

Table 3.9. Standardized total effects of predictors (column headings) on

response variables (row headings) for data in Figure 3.10

stems00 nig00 lac00 nig01 lac01

nig00 0.51 ––––– ––––– ––––– –––––
lac00 ns ––––– ––––– ––––– –––––
nig01 0.52 0.62 −0.14 ––––– –––––
lac01 0.31 0.00 0.67 ––––– –––––
change −0.63 0.00 −0.15 0.00 −0.22

(7) The seventh rule of path coefficients allows us to calculate total effects. In
this case, there are several and they are illustrated in Table 3.9. The reader
should be able to calculate these total effects from the information in Figure
3.10.
(8) In this situation, the eighth rule of path coefficients simply allows us to
see the relationships between bivariate correlations and path coefficients.
One of the more interesting cases is that for A. nigriscutis in 2000 and A.
lacertosa in 2000. The two paths connecting these two variables include
the shared causal effect of stems in 2000 (0.51 × 0.15) plus their direct
correlation (−0.20). Together, these constitute the bivariate correlation,
which is −0.12.

Interpretation of unstandardized coefficients for our example

As we discussed earlier in the chapter, the standardized coefficients only give
part of the picture. These relate primarily to questions about shared variation
in our sample. In other words, the standardized path coefficients reflect the
degree to which observed variation in one variable relates to observed variation
in another. So, looking at Figure 3.10, we can see that the variation we observe
among plots in the change in stems between years is strongly related to variations
among plots of numbers of stems, independent of the effects of other variables.
The exact nature of this direct effect of stems in 2000 on change in stems,
however, is represented by the unstandardized coefficient −8.96. Taking the
units of both variables into account, we would interpret the unstandardized
coefficient as follows: “as the number of stems in a plot goes up by 1 stem, the
change in stems goes down by 8.77 percent”. Therefore we would predict that
if the stems in a plot increased by 10, stems would decline by 87.7%.
Some of the paths in Figure 3.10 do not show a big difference between
the standardized and unstandardized values, in particular, the paths from A.
 The anatomy of models I: observed variable models 69

nigriscutis in 2000 to A. nigriscutis in 2001 and A. lacertosa in 2000 to A.

lacertosa in 2001. This occurs because the units are the same for both variables
involved with each of these paths, their variances are very similar, and variances
are not too far from 1.0.
Finally, we need to be aware that the data for A. nigriscutis and A. lacertosa
densities are the logarithms of the raw scale. Thus, we have to translate unstan-
dardized coefficients if we wish to say, for example, how much A. nigriscutis
would increase in density if we were to increase the number of stems. Basically,
the answer would be 0.51 times the logarithm of the insect density.
While it is difficult to give advice on the use of standardized and unstan-
dardized coefficients that is universally correct, we could say that when you
are interested in inferences about the data in your sample, standardized coeffi-
cients are often what you should use. However, when your goal is to generalize
to other samples or make predictions, the unstandardized coefficients can be
required for proper interpretation.

A possible alternative standardization procedure

In spite of the criticisms of standardization, researchers would generally prefer
a means of expressing coefficients in a way that would permit direct compar-
isons across paths. The debate over this issue goes back to Wright (1921) who
originally developed path analysis using standardized variables. It was Tukey
(1954) and Turner and Stevens (1959) who first criticized the interpretability of
standardized values in regression and path models, and many others have since
joined in that criticism. However, Wright (1960) argued in defense of standard-
ized coefficients saying that they provide an alternative method of interpretation
that can yield a deeper understanding of the phenomena studied. Later, Hargens
(1976) argued that when the theoretical basis for evaluating variables relates
to their relative degrees of variation, standardized coefficients are appropriate
bases for inference. Therefore, we must admit that a form of standardization
would seem desirable from a researcher’s perspective. As Pedhazur’s recent
assessment of this problem concludes, “. . . the ultimate solution lies in the
development of measures that have meaningful units so that the unstandardized
coefficients . . . can be meaningfully interpreted.”
So, how might we standardize using measures that have meaningful units?
We must start by considering what it means to say that if x is varied by one
standard deviation, y will respond by some fraction of a standard deviation. For
normally distributed variables, there is a proportionality between the standard
deviation and the range such that six standard deviations are expected to include
70 Structural equation modeling and natural systems

30 25
25 20

Frequency
Frequency

20
15
15
10
10
5 5

0 0
0--20 21-- 41-- 61-- 81-- 101-- 121-- 141-- 0--8 9--16 17-- 25-- 33-- 41-- 49-- 57--
40 60 80 100 120 140 160 24 32 40 48 56 64
Cover Stand Age, yrs
30
25
30
Frequency

20 25

Frequency
15 20
15
10
10
5 5
0 0
1--2 2--3 3--4 4--5 5--6 6--7 7--8 8--9

5
0
50

00

50

00

50

00

22
05
-1

-3

-4

-6

-7

-9

--1
-1
0-

1-

1-

1-

1-

1-
Fire Severity Index, mm

1-

51
15

30

45

60

75

90

10
Elevation, m

Figure 3.11. Frequency distributions for four variables from a study of wildfire
effects on plant recovery, used to discuss the concept of relevant ranges.

99% of the range of values. As discussed earlier, this may seem reasonable if
(1) we have a large enough sample to estimate a consistent sample variance,
(2) our variables are normally distributed, and (3) variances are equal across any
samples we wish to compare. The reason why many metricians oppose stan-
dardized coefficients is because generally these three necessary conditions are
not likely to hold. Of equal importance, rarely are these requirements explicitly
considered in research publications, so we usually don’t know how great the
violations of these requirements might be.
Figure 3.11 presents frequency distributions for four variables from an SEM
study by Grace and Keeley (2006). In the absence of further sampling, the
repeatability of the sample variance estimate is unknown. This contributes to
some uncertainty about the interpretability of coefficients standardized on the
standard deviations. As for approximating a normal distribution, three of the
four variables are truncated at the lower end of values. Cover can never be less
than 0%, elevation likewise has a lower limit of expression relevant to terrestrial
communities in this landscape, and stand age is also limited to a minimum
value of between 0 and 1 year. None of these deviations is substantial enough to
cause major problems with distributional assumptions (i.e., these variables are
not wildly nonnormal), however, the deviations from idealized normality may
very well impact the relationships between standard deviations, and ranges.
 The anatomy of models I: observed variable models 71

The observed range for cover was from 5% to 153%, while 6 times the standard
deviation yields an estimated range of 190%. The observed range for elevation
was from 60 to 1225 m, while 6 times the standard deviation equals 1550. Stand
age ranged from 3 to 60 years old, with 6 times the standard deviation equaling
75 years. Finally, fire severity index values ranged from 1.2 to 8.2 mm, while 6
times the standard deviation equals 9.9. Thus, observed ranges are consistently
less than would be estimated based on standard deviations, and the degree to
which this is the case is slightly inconsistent (ratios of observed to predicted
ranges for cover, elevation, age, and severity equal 0.78, 0.75, 0.76, 0.71).
It is possible that in some cases information about the range of values likely
to be encountered or of conceptual interest can provide a more meaningful basis
for standardizing coefficients than can the sample standard deviations. We can
refer to such a range as the relevant range (Grace and Bollen 2005). For example,
if we have a variable whose values are constrained to fall between 0 and 100,
it would not seem reasonable for the relevant range chosen by the researcher
to exceed this value, regardless of what 6 times the standard deviation equals.
On the other hand, it may be that the researcher has no basis other than the
observed data for selecting a relevant range. Even in such a case, we can choose
to standardize samples that we wish to compare by some common range so as to
clarify meaning across those samples. Whatever the basis for standardization,
researchers should report both the unstandardized coefficients and the metrics
used for standardization.
In this case, we might choose to specify the relevant range for cover to be
from 0 to 270%. Obviously values cannot fall below 0%, but why choose an
upper limit of 270%? Examination of cover values for all plots across the five
years of the study show that values this high were observed in years 2 and 4 of
the study. By using a relevant range of from 0 to 270, we permit comparisons
across years standardized on a common basis. Of course, this implies that the
slopes measured will extrapolate to that full range, which is an assumption that
should be evaluated closely. For elevation, we might choose the relevant range to
be the observed range, from 60 to 1225 m. This span of 1165 m might be chosen
because we do not wish to extrapolate to lower or higher elevations, in case
relationships to other variables are not robust at those elevations. For stand age,
we could specify the relevant range to be 60 years for basically the same reason.
Finally, the fire index range chosen might also be the observed range, which was
7.0 mm. It is clear that values could be obtained beyond this range in another
fire. It is not known, however, whether the relationship between remaining twig
diameter and herbaceous cover would remain linear outside the observed range.
Based on these determinations, we can generate path coefficients standard-
ized on the relevant ranges for an example model involving these variables.
72 Structural equation modeling and natural systems

elevation
0.037
0.301
-0.022 0.160
-0.450
0.798 -0.450 0.578 0.692

fire plant
stand age
0.085 severity -7.32 cover
0.650 -0.386
0.692 -0.190
Figure 3.12. Example model comparing the unstandardized path coefficients
(upper), coefficients standardized based on their standard deviations (middle), and
coefficients standardized based on their relevant ranges (lower).

Figure 3.12 shows the three types of standardized coefficient. The biggest
numeric differences between values is that when standardized by the relevant
ranges, the values of the coefficients leading to cover are lower because of the
large relevant range selected for this variable. The coefficient for the effect of
age on severity is slightly higher, while that for the effect of elevation on age is
unchanged. Using these coefficients now allows us to describe the importance
of variables using their relevant ranges as the explicit context. These interpreta-
tions are only valid for relative comparisons within the parameter space defined
by the relevant ranges. So, we can say that as fire severity increases across its
relevant range, cover would be expected to decline by 19% of its relevant range.
As elevation increases across its relevant range, the total change in cover from
both direct and indirect causes would be an increase of 21.9% (the total effect).
We now conclude from this analysis that the sensitivity of cover to fire severity
and elevation (19% versus 21.9%) are roughly equivalent in this study, though
of opposing sign. It is possible to test whether these two estimates are reliable
differences, which in this case, they are not.

Conclusions
The sophisticated procedures of modern structural equation modeling start with
the basic principles of regression and the rules of path coefficients. When these
 The anatomy of models I: observed variable models 73

rules and principles are combined with an understanding of model structure, the
anatomy and interpretation of structural equation models should become clear
for a wide variety of model types. For those who wish to understand SEM,
fundamental distinctions between standardized and unstandardized variables
are important. Understanding these issues is especially important for those in
the natural sciences where there has been much use of path analysis based on
correlations. In the next chapter we will continue to build on our knowledge of
the capability of SEM by considering latent variables and issues related to their
use and interpretation.

Appendix 3.1 Derivation of the formula for the partial

regression coefficient
From a mathematical standpoint, the formula for calculating a partial regression
coefficient is quite simple. However, properly interpreting partial path coeffi-
cients is one of the basic elements of mastering SEM. The material in Chapter 3
presents the concepts associated with partial regression using one approach.
Here, I approach the problem from a different direction by deriving the formula
for partial regression, in case this alternative approach is more useful for some
in developing their understanding. For the benefit of the reader, I show all the
steps in the derivation.
There is a fundamental requirement that the correlation between two vari-
ables in a path model must equal the sum of the direct and indirect paths con-
necting them. This is given in Chapter 3 as our eighth rule of path coefficients.
So, for a simple model of multiple regression (Figure 3.1.1), it holds that

r x1 y1 = γ11 + (r x1 x2 × γ12 ) (3.1.1)

and

r x2 y1 = γ12 + (r x1 x2 × γ11 ) (3.1.2)

In this case, our goal is to express the partial path coefficients (which are our
unknown parameters) in terms of the known parameters (the correlations). So,
we must rearrange our equations, yielding

γ11 = r x1 y1 − (r x1 x2 × γ12 ) (3.1.3)

and

γ12 = r x2 y1 − (r x1 x2 × γ11 ) (3.1.4)

74 Structural equation modeling and natural systems

x1 g11
r x 1x 2 y1
x2 g12 z1
Figure 3.1.1. Simple multiple regression model.

Substitution of Eq. (3.1.4) into Eq. (3.1.3) allows us to derive the equation for
γ 11 as follows:
γ11 = r x1 y1 − r x1 x2 × (r x2 y1 − r x1 x2 × γ11 ) (3.1.5)
γ11 = r x1 y1 − r x1 x2 × r x2 y1 + r x21 x2 × γ11 (3.1.6)
γ11 − r x21 x2 × γ11 = r x1 y1 − r x1 x2 × r x2 y1 (3.1.7)
 
γ11 1 − r x21 x2 = r x1 y1 − r x1 x2 × r x2 y1 (3.1.8)
r x y − r x1 x2 × r x2 y1
γ11 = 1 1 (3.1.9)
1 − r x21 x2
Through similar means we can arrive at our equation for γ 12 ,
r x y − r x1 x2 × r x1 y1
γ12 = 2 1 . (3.1.10)
1 − r x21 x2
While we have illustrated the formulae for partial regression coefficients
using a multiple regression model (Figure 3.1.1), it is important to realize how
this extends to a fully directed path model of the form shown in Figure 3.1.2
(similar to Figure 3.4). Let us see why this is the case.
In Figure 3.1.1, the first rule of path coefficients tells us that the relationship
between x1 and x2 can be represented by their bivariate correlation coefficient.
While the model shown in Figure 3.1.2 may be interpreted differently, it is
similar to the previous model in that the standardized path coefficient from x1
to y1 is also represented by their bivariate correlation coefficient, though this
is based on the second rule of path coefficients (which applies to variables
connected by only one causal path).

Table 3.1.1. Correlations associated with example

analyses represented in Figure 3.1.3

variable 1 variable 2 variable 3

variable 1 1.0
variable 2 0.40 1.0
variable 3 0.50 0.60 1.0
 The anatomy of models I: observed variable models 75

g21
x1 y2
g11 b21
z2
y1
z1
Figure 3.1.2. Directed path model.

A x1 0.31
0.40 y1
x2 0.48 0.56

B
0.31
x1 y2
0.40 0.56
0.48
y1
0.20
Figure 3.1.3. Illustration of similarity of path coefficients for multiple regression
model (model A) and directed path model (model B).

To illustrate the similarity between the models in Figure 3.1.1 and 3.1.2, let
us consider data for three variables with the correlations given in Table 3.1.1.
Imagine that in the first case, we use these data to evaluate a multiple regression
model like the one in Figure 3.1.1 and let variable 1 be x1 , variable 2 be x2 , and
variable 3 be y1 . The reader should be able to use the rules of path coefficients
and Eqs. (3.1.9) and (3.1.10) to find the standardized path coefficients for that
model (Figure 3.3).
76 Structural equation modeling and natural systems

Let us now consider applying our data to a different model. Perhaps we decide
that variable 2 actually depends on variable 1 and that variable 3 depends on
both 1 and 2. Such a decision would have to be based on theoretical reasoning,
as there is nothing about the data that will tell us the causal structure of the
model. This decision would lead us to formulate our model as in Figure 3.1.3.
So, now variable 1 is x1 , variable 2 is y1 , and variable 3 is y2 . We can again use
our rules and equations to arrive at the standardized path coefficients. We now
find that the coefficients are the same for both models. The fundamental reason
for this is that in both cases, variables 1 and 3 are connected by two causal
pathways, thus, the equations for partial regression are invoked in both cases.
However, interpretations of the models are somewhat different. In particular, in
model B y1 is interpreted as being influenced by x1 and has an attending error
variable, while the relationship between x1 and x2 in model A is considered
to be the result of an unanalyzed joint cause (a correlation). Also, the other
difference between the two models is that the total effects of x1 and x2 on y1 in
model A are simply 0.31 and 0.48. However, in model B, the total effect of x1
on y2 is 0.31 plus 0.40 times 0.48, which equals 0.50.
 4
The anatomy of models II:
latent variables

Concepts associated with latent variables

What is a latent variable?
As indicated in our overview of structural equation models at the beginning
of Chapter 3, in addition to observed variables in structural equation models,
another type of variable, referred to as a latent variable, is commonly included.
There are a variety of ways latent variables can be defined. Strictly speaking, a
latent variable is one that is hypothesized to exist, but that has not been measured
directly. Bollen (2002) states that we can simply think of a latent variable as
one that is unobserved. In spite of this simple definition, latent variables are
typically used to represent concepts. However, there are many types of concept
and not all of them correspond to latent variables. Further, the algorithms used to
calculate latent variables define their quantitative properties, although how they
are interpreted is a matter of theoretical consideration. Finally, latent variables
are starting to be used in structural equation models for a wider variety of
purposes, such as to represent parameters in latent growth modeling. Because
latent variables are often used to represent abstractions, their meaning is not
always immediately obvious. For this reason, I believe it will be best to rely on
a variety of examples and a discussion of their properties to illustrate the roles
latent variables can play in SEM.
An illustration of a latent variable ξ with a single indicator variable x is
shown in Figure 4.1. There are several things to notice about this representation
and these help to clarify some of the attributes of latent variables. First, a
latent variable is typically (though not always) associated with one or more
observed indicator variables. We use the term indicator to refer to the fact that
the observed variable is not a perfect measure of the concept of interest. Rather,
the observed variable serves to provide us with information about the latent

77
78 Structural equation modeling and natural systems

dx x x
Figure 4.1. Symbolic representation of a latent variable, ξ , a single observed
indicator variable, x, and the error variable δ x associated with x.

variable. Most importantly, we presume that the values of the indicator variable
are reasonably well correlated with the true latent variable values. Through this
correlation, our ability to measure x provides us with information about ξ .
A second thing to observe about Figure 4.1 is that the direction of causality
is represented by an arrow from the latent variable to the indicator. Initially
this can be confusing since the values of the observed variables are used to
estimate the values of the latent variable. However, it is important to realize
that conceptually, the latent variable is the “true” variable of interest, and the
indicator is a manifestation of the effects of that true entity. A classic example
of latent variables is the components of human intelligence. It is presumed that
people have abilities that cannot be directly measured. Rather, performance on
tests of various sorts can be used to provide indirect measures of those abilities.
In a similar fashion, we can presume that many concepts represent properties
that we cannot measure directly but that cause observed manifestations. This is
true even when the concepts are not highly abstract.
A third thing to notice about Figure 4.1 is that there exists an error associated
with the observed indicator. In terms of equations, we can represent this situation
with the formula

x = λξ + δx (4.1)

This equation makes it clear that we are stating that the values of x are the result
of the influence of the latent variable ξ , proportional to λ, its effect on x, plus
error, δ x .
A tangible example of a latent variable and its indicator is represented in
Figure 4.2. Here we declare that what is of interest is a concept referred to as
body size and that we have a single indicator, the mass of individual animals
measured at a single point in time. It is presumed in Figure 4.2 that there is some
measurement error associated with the values of animal mass. By measurement
error, we refer to the degree to which our indicator variable does not correlate
perfectly with our latent variable. We may or may not know something about the
magnitude of that measurement error, but making a distinction between latent
 The anatomy of models II: latent variables 79

body Body
dx
mass Size
Figure 4.2. Example of a latent variable, Body Size, with a single indicator, ani-
mal mass. The term δ x represents the error associated with our values of animal
mass.

and observed variables implies that we believe some degree of measurement

error does exist. The quantification of measurement error will be described
below.

Reason 1 to use latent variables: distinguishing concepts

from observations
Now that a simple example of a latent variable has been presented, we can
start to address the question of why we might include latent variables in struc-
tural equation models. Throughout this chapter we will emphasize a number of
different reasons; here we start with one of the most basic.
One reason an investigator might wish to include latent variables in a model
is to make a distinction between the concepts of interest and the observations
in hand. By including conceptual variables in structural equation models, we
begin to consider more explicitly the theoretical content of variables. Theories
are concerned with concepts and how they interrelate. A major part of the
reason that latent variable models are considered to be more sophisticated than
observed variable models is because of their more theoretical nature. This does
not mean that latent variable models are always superior to observed variable
models, only that latent variable models have the potential to embody a greater
content of information by separating concept from measurement.
It should be said that from a researcher’s standpoint, observed variables may
clearly be embodied with theoretical content. So, I am not advocating that all
SEM models utilize latent variables. However, by including the observed vari-
ables along with latent variables in structural equation models, we make explicit
how we are relating observation to theory. It is often the case with observed
variable models that no discussion of how observations relate to concepts is
given, leaving the reader in the dark on this. This issue has significant implica-
tions for the advancement of science and I will have much more to say about
this subject in the last section of the book. For now, we should recognize certain
immediate consequences of including both observed and latent variables in a
80 Structural equation modeling and natural systems

model. It allows the model to have explicitly linked theoretical and empirical
content. In the natural sciences, we have traditionally related our theories to
the empirical results without an explicit consideration. We theorize about the
effects of body size on survival and then proceed to use some individual and
specific measures of body size and survival to evaluate these theories. We may
spend some time wrestling with the fact that we can measure several aspects
of body size, and they do not all have exactly the same degree of relationship
with our measure(s) of survival. In the end, we are most likely to pick the one
measure of body size, that correlates most highly with our measure of survival
in our data set. The resulting statistical model is often one that feels very specific
because we have picked one predictor from several candidates, and also less
like a proper evaluation of a hypothesis than we would like, because we did not
have an initial theory about how our measurements would relate to our con-
cepts. Latent variable modeling seeks to improve upon this situation. It not only
makes explicit the concepts of interest and how we have estimated them, it also
leads to theories that are empirically meaningful. Ecologists have a long history
of theories with vague and nonoperational concepts. Explicitly linking theory
and observation in the same model has the consequence of promoting theory
with greater empirical validity. Thus, making a distinction between observed
and latent variables would seem to have the potential to lead to better science.

Construct validity and reliability

For a model such as the one in Figure 4.2, the validity of the indicator refers to the
degree to which the observed variable accurately reflects the concept embodied
by the latent variable. Validity defined in this way is a measure of accuracy.
When our model only includes a single indicator variable, as in our example in
Figure 4.2, support for the validity of the indicator must be based on persuasive
argument. In other words, the accuracy of the statement, “body mass is an
accurate indicator of body size” is one of convention. Perhaps many scientists
will be willing to accept this equivalence, but our data provide no support for
this claim. In general, the more conceptually abstract a latent variable is, the
more indicators are required and the more explicit the discussion needed for
the latent variable to be understood.
In contrast to validity is the concept of reliability. Reliability refers to the
degree to which our indicator relates to our underlying true parameter (which
is represented by the latent variable). In mathematical terms, reliability is usu-
ally expressed as the square of the correlation (also called loading) between
indicator and latent variables. Reliability can also be defined as the proportion
of observed variance in common with the true underlying parameter. In turn,
 The anatomy of models II: latent variables 81

Table 4.1. Illustration of the type of data obtained from repeated

measurements of an indicator variable, x (i.e., x – trial 1 refers
to measurements of x in the first of three trials) over n cases. The
average correlation among values from separate trials provides an
estimate of the reliability of our indicator

Animal x – trial 1 x – trial 2 x – trial 3

1 1.272 1.206 1.281
2 1.604 1.577 1.671
3 2.177 2.192 2.104
4 1.983 2.080 1.999
n 2.460 2.266 2.418

the proportion of variance of an indicator that is not in common with the true
underlying parameter is called the proportional error variance (which is 1.0
minus the reliability). As with all other parameters, reliability and error variance
can be expressed either in standardized or unstandardized metric.
Returning to our example in Figure 4.2, assuming we are willing to accept
body mass as a valid single indicator of body size, the reliability of our indi-
cator is determined by its repeatability. If repeated sets of measurements of
body mass are highly correlated with each other, we are said to have a reliable
measure. From a computational perspective, the reliability of our indicator is
equal to the average correlation between repeated measurements of body mass.
Thus, imagine that we obtain the data represented in Table 4.1. Relevant to
our example, we might imagine a calibration process where at the beginning
of our study we took repeated measurements of body mass to determine the
reliability of our measurement process. Data such as those in Table 4.1 can be
used to calculate the correlations between trials, and then we can also deter-
mine the average correlation. Say, for our example, that the average correlation
between measurement trials is 0.75; we would find that the reliability of our
measurement of body mass is 0.75.
Once we have an estimate of reliability for our indicator, there are several
things we can do with that information. If we have multiple measurements of x
for all dates in our study we may simply prefer to average across the multiple
measurements and use those average values. On the other hand, if we have
multiple measures for only the first time period, but single measurements of
body mass were taken for several additional time intervals, we may wish to
use our estimated reliability to specify the relationship between indicator and
82 Structural equation modeling and natural systems

body 0.87 Body

0.25
mass Size
Figure 4.3. Example construct showing standardized path coefficients when the
reliability of the indicator is 0.75 (see text for discussion).

latent variable throughout our model. In general terms, the advantage of such an
approach is that our knowledge about indicator reliability is incorporated into
our model. For single indicator latent variables, the path coefficient between
latent and indicator is defined by the ninth rule of path coefficients. This rule
states that the standardized path coefficient between a latent variable and its
indicator is the square root of the specified reliability of the indicator. The
specification of that reliability takes place outside the model, either through the
default assumption of a perfect correlation (in which case the coefficient has
a value of 1.0), or through the inclusion of some estimated reliability of our
indicator (which is some value less than 1.0).
Now, as stated earlier, there is a direct relationship between reliability and
measurement error. While reliability represents the proportion of variance in
common between indicator and latent variables, the measurement error rep-
resents the quantity of variance in the indicator that is estimated to be error.
Absolute error variance in x can be related to its reliability using the formula:
 
δx = 1 − λ2x × VARx (4.2)

where δ x is the unstandardized error variance in x, λx refers to the loading

between latent and indicator variables (λ2x is the reliability), and VARx is the
variance of x. So, if in our example, the variance of x is 0.22 and the reliability
is 0.75, the error variance in absolute units is 0.055. If we were to express the
results using standardized variables, where the VARx = 1.0, then the standard-
ized error variance would simply be 1.0 minus the reliability (which would be
0.25 in this case). When programming a structural equation model, specify-
ing the error variance (and, therefore the reliability of an indicator) is readily
accomplished with a simple command statement. In Figure 4.3, our example
involving body mass as an indicator of body size is illustrated assuming the
reliability of body mass is 0.75. The standardized path coefficient (also known
as loading) linking indicator to latent variables is the square root of the reli-
ability, 0.87; and the standardized error variance is 1.0 minus the reliability,
which equals 0.25.
 The anatomy of models II: latent variables 83

Types of application involving latent variables

So far, I have illustrated a latent variable construct as an isolated entity, involving
only one latent variable and one indicator. A more complete understanding of
the use of latent variables comes from studying more complex cases, where
there can be multiple latent variables and multiple indicators for individual
latent variables. It may help somewhat in our consideration of latent variables
to introduce certain distinctions. First, I will distinguish between parts of a
latent variable model. Following that, I introduce distinctions about types of
application involving latent variables and their role in regression models. Once
these distinctions are introduced, I will then proceed with models having more
than one latent variable and more than single indicators associated with each.
While already mentioned earlier in the book, here I reintroduce the distinction
between the measurement model and the structural model. The measurement
model is the portion of a structural equation model that relates latent variables
to indicators. The structural model, in contrast to the measurement model, is
the part that relates latent variables to one another. The name for this portion
of the model is selected because it is the part that makes explicit how the
latent variables are being measured. In the observed variable models presented
in Chapter 3, the measurement model is missing. In models that have single
indicators, the measurement model is simplistic. Later in this chapter we will
present models where the measurement model comprises the entire model, as
there are no directed relationships among latent variables.
The other distinction I wish to make has less to do with the anatomy of the
model and more to do with its purpose. Path-focused applications are distin-
guished by the fact that we are primarily interested in the regression relation-
ships among concepts (i.e., the structural model). In contrast, in factor-focused
applications (named for the association with factor analysis), we are concerned
primarily with understanding relationships among a set of correlated response
variables and, in particular, with evaluating ideas about the number and nature
of underlying factors that cause observed variables to correlate. These two types
of analysis evolved from the separate traditions of path analysis and factor anal-
ysis. The synthesis of modern SEM represents a generalized procedure in which
elements from both traditions can be used in a single application. Thus, we now
have path-focused models with latent variables and factor-focused models with
directional effects among factors. It is important to note that there is a third type
of application, the hybrid model, in which both path and factor relationships
are of strong theoretical interest. What I would like to emphasize here is that
in analyses that are primarily concerned with path relations versus those pri-
marily concerned with factor relations, we may have some subtle but important
84 Structural equation modeling and natural systems

Conceptual Size of
Body Size
Model Territory

Observed body singing

Variable Model mass range

indicators
Latent
body Size of singing
Variable mass
Body Size
Territory range
Model
latent
variables
Figure 4.4. Illustration of three types of model. The upper figure illustrates a
simple conceptual model, the middle figure illustrates a related model involving
observed variables, while the lower figure illustrates a latent variable model com-
bining the conceptual and observed variables.

differences in priorities. In the rest of this chapter, I will first elaborate on path-
focused applications and then discuss factor-focused applications. I will end by
extending a factor-focused application into a hybrid model analysis.

Path-focused applications
A survey of the published applications of SEM (including historic path analysis)
in the natural sciences demonstrates that the great majority have been focused
on path relationships. This holds true for models containing latent variables as
well as for the widespread use of observed variable models, which by design
are exclusively path-focused. Only a handful of SEM applications with a strong
emphasis on understanding factors have been published in the natural sciences,
although this may be expected to change as more researchers become familiar
with the technique.
We continue our discussion of latent variables by considering how they can
contribute to path-focused applications. Let us now imagine the situation where
our interest in animal body size is with reference to its role in influencing an
animal’s territory size (Figure 4.4). Our conceptual model of the system at hand,
whether explicitly stated or not, defines the context of our model by representing
the question being addressed. In this case, the question is, how does the size of
 The anatomy of models II: latent variables 85

A B
25 0.64
20

y 15 0.60
10 x y
5 (4.84)
0
0 0.5 1 1.5 2

x
C 0.52

0.87 0.69 1.0

0.25 x xx (6.47)
hy y 0.0

Figure 4.5. Figures used to illustrate important points about measurement error.
Path coefficients in parentheses are unstandardized values. Other values are stan-
dardized. A. Representation of a raw regression of y on x. B. Representation of
the regression in A in diagrammatic form, showing the standardized and unstan-
dardized slopes as well as the standardized error of prediction in y, which is 0.64.
C. Latent variable regression model in which 25% of the variance in x is assumed
to be measurement error.

an individual animal’s territory relate to the size of its body? Also in Figure 4.4,
is the observed variable model which might represent this question if we simply
used our available observed variables, and made no explicit statement about
the concepts they represent. Comparing our observed variable model to our
conceptual model helps to make the points that (1) observed variable models
are typically more specific and less theoretical than conceptual models, and (2)
there is often an unstated presumption that the observed variables used represent
the concepts of interest, although this presumption is not addressed explicitly
in this simple case. In latent variable modeling, we combine the conceptual and
observable into a model that includes both. As a result, our latent variable model
represents a more sophisticated model than either our conceptual or observed
variable models by themselves.

The impact of measurement error on model parameter estimates

in simple regression models
A fundamental justification for the use of latent variables comes from the impact
of measurement error on model parameters. This impact applies to all types of
statistical model, including traditional univariate and bivariate procedures. Fig-
ure 4.5 is presented to help illustrate in a very simple example how measurement
error affects regression relationships, and thus, path-focused applications.
86 Structural equation modeling and natural systems

A 0.64

0.60
x y
(4.84)

B 0.52

1.0 0.69 0.87

0.0 x xx (4.84)
hy y 0.25

Figure 4.6. Illustration of the effects of measurement error in y on regression

between x and y. Path coefficients in parentheses are unstandardized values. Other
values are standardized.

Imagine the case where two variables, x and y, are measured with random
sampling and we believe that x has a causal influence on y. Let us further imagine
that there is some error in the measurement of x, referred to as δ x , such that values
of x obtained are not exactly the same as the true values of x. Such a situation
would seem to be very common in the natural sciences, and thus may be viewed
as a rather normal occurrence. Figure 4.5A illustrates our usual visualization
of the raw relationship between x and y. Simple linear regression reveals a
standardized path coefficient between x and y of 0.60 using least squares, which
is shown diagrammatically as an observed variable model in Figure 4.5B. Now,
let us further imagine that the reliability of our measure of x is equal to 0.75 (as
in the example in Figure 4.3). One implication of measurement error in x is that
some of the error of prediction of y, represented as the residual error variance,
is actually not prediction error, but instead, measurement error. In other words,
measurement error in x gets lumped in with the prediction error of y, leading to
an underestimation of the strength of the true relationship between x and y and
a downward bias of the path coefficient.
Given that we are aware of a less than perfect reliability for x, we can
represent our regression model using latent variables (Figure 4.5C). Here, for
simplicity, we assume no measurement error for y. In the latent variable model,
ξ x represents the true values of x, which are assumed to be error free by defini-
tion, while ηy represents the true values of y. The path from ξ x to x is specified
by the square root of the reliability, 0.87, and the path from ηy to y is set to 1.0.
In this fashion, the path from ξ x to ηy represents the true effect of x on y. Since
we specify a reliability for x of 0.75, the error variance for x can be calculated
using Eq. (4.2), yielding a value of 0.055 (given VARx = 0.22). In Figure 4.5,
the standardized error variance is shown, which is 0.25 (1 – reliability). What
 The anatomy of models II: latent variables 87

difference does all this make to our estimation of the relationship between x
and y? The change in the estimated effect of x on y in standardized terms can
be anticipated by the formula

b =γ ×λ (4.3)

Where b is the standardized regression coefficient when measurement error is

ignored, γ is the true regression coefficient in standardized units, and λ2 is the
reliability. Since in our example λ = 0.87 and b = 0.60, then γ = 0.69. Our esti-
mate of unexplained variance has also changed, from 0.64 to 0.52 (comparing
B to C in Figure 4.5), thus, the respective R2 s are 0.36 and 0.48 (note that R2 = 1
minus the unexplained standardized variance). In this case, unexplained vari-
ance is reduced (and explained variance increased) because we have assigned
part of the variance in x to error. The estimate for the unstandardized path coef-
ficient is also increased when measurement error is accounted for, from 4.84
to 6.47. This example demonstrates an important point about measurement
error – measurement error in predictor variables results in downward bias in
both the unstandardized and standardized coefficients, as well as the estimate
of variance explained.
We can explore the effects of measurement error further by looking at another
simple example, in this case, one where the y variable contains measurement
error but the x variable does not. Imagine that it is our y variable that is influ-
enced by measurement error, with a reliability of 0.75, while our x variable is
measured with enough precision to justify an estimated reliability of 1.0. Com-
paring A and B in Figure 4.6, we can see that in this case, ignoring measurement
error (as is done in A) leads to an underestimate of the standardized effect of ξ x
on ηy , and also an underestimate of the variance explained (or, conversely, an
overestimate of the unexplained variance in ηy ). However, the unstandardized
effect of ξ x on ηy is not influenced by measurement error in y. This demon-
strates another important point about measurement error – measurement error
in response variables does not bias the unstandardized path coefficients, only
the standardized coefficients and our estimate of variance explained. We can
add this to the reasons given in Chapter 3 for evaluating the unstandardized
coefficients when extrapolating from our results.

Reason 2 to use latent variables: adjusting for the effects

of measurement error
The results given above demonstrate that failing to account for measurement
error can lead to biased parameter estimates. The simple examples presented
88 Structural equation modeling and natural systems

involving a single predictor and a single response variable belie the complex
effects that measurement error can have on structural equation models (or other
kinds of statistical model for that matter). In more complex models, a wide
variety of outcomes is possible when measurement error is substantial. The
varieties of consequences are such that no single example can provide a repre-
sentative illustration of the kinds of effects that can be seen. Certain generaliza-
tions do apply, however. First, when there is measurement error in exogenous
variables, there is generally a downward bias in path coefficients from exoge-
nous to endogenous variables. Exceptions to this pattern can be found in models
with multiple, correlated endogenous variables. Here, the effects of measure-
ment error are quite unpredictable, with a wide variety of consequences being
documented. Secondly, when endogenous variables contain measurement error,
the explanatory power of the model is underestimated and standardized coef-
ficients are biased. Thirdly, when both exogenous and endogenous variables
possess measurement error, a wide variety of effects are possible.
Aside from the general consequences of measurement error on model param-
eters, there can be important effects on model fit and the significance of paths.
For a given data set, ignoring or addressing measurement error can shift infer-
ence from one model to another. The greater the amount of measurement error,
the greater the differences between models. Bollen (1989, chapter 5) provides
a detailed consideration of some of the effects measurement error can have in
structural equation models. There can be little doubt that measurement error has
important consequences for all statistical models. Structural equation modeling
and the use of latent variables provide a means of correcting for measurement
error, at least in a fashion and to a degree. This capability provides our second
reason for using latent variables, to account for measurement error so as to
reduce its influence on parameter estimates.

Some realities concerning the specification of measurement error

Specifying measurement error can lead to some discomfort as we begin to realize
how much it influences model parameters and model stability. Do we have a
correct estimate of the true measurement error? Should we ignore measurement
error and settle for the less ambitious goal of data description? If our adjusted
model is different from our unadjusted model, do we really believe that the
results from the adjusted model are closer to the true model? Based on my own
limited experience, I would have to recommend that researchers take the issue
of measurement error seriously and, when possible, use multiple measurements,
multiple indicators, or estimates of reliability to correct for measurement error.
Of course, this will not always be possible or feasible. In the next section, I will
 The anatomy of models II: latent variables 89

elaborate on the use of multiple indicators in latent variable models for cases
where this is possible and feasible.
I offer three pieces of advice for those specifying reliability and measurement
error for single indicators. First, we must keep in mind that reliability is a
sample-specific property. In other words, since reliability is estimated as a cor-
relation between replicate samples, the value of that correlation will be strongly
related to the range of conditions over which calibration has taken place. If, for
example, we are interested in the reliability of body mass values, the range of
animals included in the calibration sample should match the range that applies
to the data being modeled. Taking a limited or nonrepresentative calibration
sample will not provide an appropriate estimate of reliability for our model.
Secondly, there will certainly be plenty of cases in the biological sciences where
we are working with reliable indicators. When the researcher is reasonably con-
fident that repeated measurements would yield an extremely high correlation
between trials, they are fully justified in specifying perfect reliability or ignoring
the need to specify measurement error. Thirdly, we should not be complacent
about specifying high levels of measurement error in our models. The reality
is that when we provide a single indicator and specify that it is not very reli-
able, we will often find that stable solutions for model parameters are difficult
to achieve (for reasons that will be discussed in more detail in Chapter 5).
To suggest some more specific guidance on this matter, Kline (2005, page 59)
offers the following: “. . . reliability coefficients around 0.90 are considered
‘excellent’, values around 0.80 are ‘very good’, and values around 0.70 are
‘adequate’.” Those below 0.50 indicate that at least one-half of the observed
variance may be due to random error, and such unreliable measures should be
avoided.
When reliabilities are low, it would be far better for us to use that information
to try and improve the reliability of our estimate rather than to take the easy
route of “solving” the problem through error specification. Despite the cautions
offered here, we cannot deny that ignoring measurement error is difficult to
justify. Single-indicator latent variable models allow one to explicitly address
the problem of measurement error in path-focused applications.

The use of multiple indicators in path-focused analyses

Rationale
As defined, when the analysis is path-focused, the investigator is principally
concerned with estimating the effect of conceptual variables on one another.
Why then might they wish to complicate things by including latent variables
90 Structural equation modeling and natural systems

with multiple indicators? Make no mistake that incorporating multiple-indicator

latent variables does make the analysis more complex. Given this cost and the
fact that the interest is primarily path-focused, why not simply use an observed
variable path model?
There are two primary and interrelated motivations for using multiple indica-
tors in path-focused analyses. One follows naturally from the above discussion
of measurement error. Using multiple indicators allows for an estimation of
reliability and the apportioning of measurement error, leading to more accurate
parameter estimates. When multiple indicators are included in the model (as
opposed to using specified reliabilities and single indicators), we are assured
that the span of the sample for multiple measures matches the data being mod-
eled. This is for the obvious reason that our multiple measures are part of the
data set being analyzed, rather than taken from some other sample. This leads
us to our third reason for using latent variables – to provide for greater gen-
erality in our models. If we are claiming that body size affects territory size,
that implies that many different indicators of body size will have similar rela-
tionships with territory size. In observed variable models, we typically pick or
measure a single indicator and then claim (or hope) that the result is general.
In multi-indicator latent variable models, we have an explicit test of a gen-
eral effect of body size, rather than a specific effect from one facet of body
size.

Illustration
To further build our understanding of latent variables, let us imagine that we
have in our possession two indicators of body size for individuals of a bird
species, body mass and wing length. What happens if we decide to use these
two measures as multiple indicators of the latent factor, body size? Figure 4.7
represents this situation.
We can represent Figure 4.7 using two equations

x 1 = λ1 ξ + δ x 1 (4.4)
x 2 = λ2 ξ + δ x 2 (4.5)

Now, it is one thing to formulate a set of structural equations, but it is another to

be able to solve them for values of the unknown parameters. Our challenge in
the case of a latent variable with only two indicators is that the only empirical
information we have is the values of x1 and x2 and only one known parameter,
the correlation/covariance between them. To estimate values for ξ in Eqs. (4.4)
and (4.5), we must arrive at values for λ1 , λ2 , δ x1 and δ x2 . Since δ xi = 1 − λ21 , we
 The anatomy of models II: latent variables 91

body
dx1 mass l1
Body
Size
wing l2
dx2 length
Figure 4.7. Representation of a two-indicator model of Body Size.

only have to estimate the λs. We must still make a simplifying assumption to
accomplish this. Resorting to standardized variables for ease of discussion once
again, we have the situation where we have one known piece of information,
r12 , the correlation between x1 and x2 , and two unknowns, λ1 , λ2 . Given that we
have only two indicators for our latent variable, we generally have no reason to
suspect that one indicator is better than the other in the sense that it correlates
more closely with the latent than does the other indicator. Thus, it is logical to
set the standardized values of λ1 and λ2 to be equal. Now, we have reduced the
estimation problem to one known, r12 , and one unknown, λ. All that is left is
to derive λ from r12 .
In Chapter 3 we encountered the statistical principles that relate to two
non-interacting variables under joint causal control. Here we have that same
situation, the only difference being that the causal control is being exerted
by a variable that we have not been able to measure directly, ξ . The model
represented by Eqs. (4.4) and (4.5) (and exemplified in Figure 4.7) is one in
which x1 and x2 are correlated solely because of the joint influences of ξ . In
this model there is an explicit assumption that x1 and x2 do not interact directly.
Therefore, we should be able to calculate the loadings that would result in a
correlation between x1 and x2 of r12 . As we may recall from Chapter 3, the
correlation between two variables is equal to the sum of all paths connecting
them, including both causal and noncausal paths (this is our eighth rule of path
coefficients). In this case, the only path connecting x1 and x2 is the indirect one
through ξ . The strength of that path is the product of the two segments, λ2 ,
which helps us to understand why the correlation between multiple indicators
equals the reliability. Since r12 equals λ2 , it holds that
√
λ= r12 (4.6)
92 Structural equation modeling and natural systems

A 0.52

body 0.86 Body 0.69 Terr. 1.0 singing

0.25 mass range 0
Size Size

B 0.52
body 0.86
0.25 mass
Body 0.69 Terr. 1.0 singing
ing
rangee 0
Size Size
wing 0.86
0.25 length

Figure 4.8. A. Illustration of model having a single indicator of body size with a
reliability of 0.75. B. Model containing two indicators of body size with reliabilities
of 0.75. These models represent two equivalent means of representing measurement
error, either by specification or by the use of multiple indicators. Terr. Size refers
to the latent variable territory size and singing range is its single indicator.

where λi = λ1 = λ2 and refers to the coefficients (loadings) linking indicators

and latent variables.
We can illustrate the application of Eq. (4.6) to our example of multiple
indicators of body size. Imagine that the correlation between x1 and x2 , r12 ,
is 0.75. We would again equate this to the reliability of these two indicators.
The value of λ would be 0.87. Figure 4.8 allows us to compare the situa-
tion where we have a single indicator of body mass with a reliability of 0.75,
to the case where we have two indicators of body mass, each with a relia-
bility of 0.75. The results obtained for loadings, path coefficients, and error
variances are identical. What is different in this case is that our definition of
body size is more general, representing the variance in common between two
indicators.
I would be remiss if I did not digress here to mention that in the analysis of
multi-indicator latent variable models, there always exists the need to provide
some definition of the variance of the latent factor. Since the latent factor is
unmeasured, we must give it some sort of scale by either (1) specifying the
variance of the latent variable to be 1.0, or (2) setting the scale of the latent
variable equal to that of one of the indicators. This latter approach is accom-
plished by specifying that the unstandardized loading for one of the variables
is 1.0.
 The anatomy of models II: latent variables 93

When there are three or more indicators for a latent variable

Let us now take our example of body size and territory size one step further,
by considering indicators of body size more generally. We can theorize that
among a set of individuals (e.g., male red-winged blackbirds, which are highly
territorial during the breeding season) there is a general size factor such that
larger individual birds are simultaneously heavier, have longer bodies, greater
wing spans, larger beaks, and larger dimensions for other morphological char-
acteristics. At the same time, possession of these traits might be expected to
correlate with more impressive displays, louder calls, and a greater capacity to
intimidate potential opponents. If our hypothesis about a general effect of body
size is correct and our several indicators are valid ones, we should expect that
different subsets of the several indicators should form constructs with similar
relationships to territory size.
Reliabilities can also be specified for multi-indicator latent variables. Take,
for example, a case where several indicators exist for a single latent variable.
Factor analysis (also known as common factor analysis) can be used to estimate
the loadings and errors for the indicators, as we shall discuss in the next major
section of this chapter. We may then use those estimates to determine construct
reliability using the formula

 2
λj
ρxi xi =  2  (4.7)
λj + εj

where ρ xi xi is the construct reliability, λ j refers to the standardized loadings

for the j indicators and εj refers to their errors. For the multi-indicator case,
construct reliability can again be used to correct path coefficients using the
formula in Eq. (4.3).
It is important to point out at this juncture, that the solution of multi-indicator
latent variable models moves us into a realm where ordinary least squares
solutions, which can be used in recursive observed variable models, are not
adequate. Rather, iterative methods such as maximum likelihood estimation
are needed for factor models, because of the challenges of estimating unknown
parameters from our known values. These challenges come from the fact that we
have no direct measures of our latent variables, thus, our formulae for calculating
partial regression coefficients do not apply. Because of this, methods such as
maximum likelihood are used to try different parameter values until adequate
ones are found. This subject will be discussed in Chapter 5, where parameter
estimation will be discussed in more detail. Nonetheless, it is still possible for
us to relate the estimated parameters in models containing latent variables to
94 Structural equation modeling and natural systems

body 0.52
0.25 mass 0.86

wing 0.86 Body 0.69 Terr. 1.0 singing

0.25 length range 0
Size Size

beak 0.78
0.39 length

Figure 4.9. Illustration of model of body size influences on territory size having
three indicators of body size.

our simple tracing rules from Chapter 3. Here I illustrate this process using a
model having three indicators.

Illustration of the relationship between path coefficients and bivariate

correlations in multi-indicator models
Figure 4.9 shows hypothetical results for our model of body size with a third
indicator included, beak length. In this case, the example has been constructed
such that beak length does not have the same loading on body size as the
other two indicators; but the relationship between body size and territory size is
unchanged. One way we can understand these results is to relate the standardized
path coefficients to the bivariate correlations among variables, as we did in
Chapter 3. To do this we need to once again rely on the eighth rule of path
coefficients, which says that the bivariate correlation between any two variables
is the sum of all the pathways connecting them. Thus, as discussed above, the
correlation between x1 and x2 can be derived from the single compound path by
which they are linked, which goes from x1 to body size and from body size to x2 .
In this case, the correlation between x1 and x2 is 0.75. Using this approach, we
can determine what all the correlations between variables should be (assuming
the data are consistent with the model), and these are presented in Table 4.2.
Notice that this rule applies to the correlations between our body size indicators
and our indicator for singing range as well (e.g., the correlation between beak
length and singing range would be expected to be 0.78 × 0.69 × 1.0, or 0.54).

Additional types of multi-indicator latent variables

Multiple indicators can be used in different ways to generalize a latent vari-
able. So far, we have emphasized the case where different indicators represent
 The anatomy of models II: latent variables 95

Table 4.2. Bivariate correlations expected, based on path coefficients in

Figure 4.9

Body mass Wing length Beak length Singing range

Body mass 1.0
Wing length 0.75 1.0
Beak length 0.67 0.67 1.0
Singing range 0.60 0.60 0.54 1.0

multi-method multi-property

dx1 dx1 wood

soil C
Soil consumed Fire
Organic Severity
dx2 loss on dx2 fine fuels
ignition consumed

repeated measures multi-sample

dx1 singing range, tt1 dx1 observer 1

Territory Caribou
dx2 singing range, tt2
Size Counts
dx2 observer 2
dx3 singing range, tt3

Figure 4.10. Examples illustrating different ways generality can be enhanced

using multi-indicator latent variables.

different attributes. It may be worth pointing out a few other ways multiple
indicators can be combined. Figure 4.10 shows four different examples. For
simplicity, I only show two or three indicators for each. As described below,
when feasible, more indicators are preferred. As for the examples, all these
represent ways in which the generality of our latent variable is enhanced. We
often select a particular method or technique of measurement, whether it is
for soil analysis, water chemistry, or vegetation assessment, for economy and
simplicity. When we use a single method, however, our results are method
dependent. A solution to this problem is to employ more than one method to
measure the attribute. Here I show two different procedures for measuring soil
96 Structural equation modeling and natural systems

organic content, soil carbon determination (e.g., with a CHN analyzer), and
loss on ignition.
We have already considered one example of a multi-property (or multi-
attribute) model when dealing with body size. Multi-property (also known as
multi-trait) models are often of interest when we wish to test general theories.
The key question we have to address is whether our traits represent a unidimen-
sional or multidimensional concept. Simply put, if they correlate consistently
and strongly, then they behave as two indicators of a single trait.
Frequently, we wish to generalize over time. When we take a single measure-
ment in time, we are often hoping, or presuming, that it generalizes over time
or correlates with conditions at the critical time in which consequences ensued.
Our earlier discussion of territory size implied that it is a static thing. It is rea-
sonable to assume that for many animals, territory size is rather dynamic. One
way to capture this is simply to average the values found in censuses, another is
to use the repeated samplings as multiple indicators (note, sometimes repeated
measures data require the inclusion of correlated error variables because his-
torical conditions cause errors to be nonindependent). The main advantage of
a multi-indicator approach compared to simply averaging the values, is that
it affords us the ability to generalize to other studies using individual census
survey data. If we were to use an average of, say three censuses, our results
would only generalize to other cases where we again average three censuses.
One other example of the use for multi-indicator latent variables is for multi-
sample data that is not explicitly repeated measures, but instead, designed to
permit generalization of a different sort. Perhaps we have the case where two
observer planes fly different survey lines searching for herds of caribou. Aside
from the differences in areas sampled, there are different teams of observers
as well. Because there are different teams of observers, we may not want to
simply add the numbers together, but rather, we can use the data as multiple
indicators. Many variations of such approaches can be used when employing
latent variables with multiple indicators in path-focused models.

Some recommendations and cautions about the use of

multi-indicator latent variables in path-focused models
There are some very good reasons for including multi-indicator latent variables
in path-focused models. As I will argue in the last section of the book, one of
the most important implications of SEM for the study of natural systems is its
ability to allow for the evaluation of multivariate hypotheses. A latent variable
with multiple indicators is, by itself, a multivariate hypothesis. While we have
referred to the indicators of exogenous constructs with “xs”, we cannot lose
 The anatomy of models II: latent variables 97

sight of the fact that the observed variables are response variables and only the
latent variables represent true causes. Ecologists and other natural scientists
frequently propose general hypotheses, such as “body size influences territory
size”. Rarely do we subject these to rigorous evaluation. Latent variable models
hold the promise of permitting such evaluations of general ideas.
Along with the promise held by latent variable models is the need for a
healthy dose of caution, as well as a fairly good fundamental understanding of
the underlying theory and calculations. We have walked through some of the
most basic concepts and will move to another perspective in the next section.
Before moving on to a more explicit consideration of latent variables, there are
a few recommendations and cautions that apply in the context of path-focused
models. The reason I include these thoughts about latent variable models here
rather than later is because, at the moment, the great majority of ecological
applications are path-focused.

Recommendations
It is recommended that when possible, three or more indicators should be used
to represent a latent variable. There are a few reasons why this piece of advice
is frequently given in the SEM literature. First, there are sometimes problems
with achieving unique parameter estimates when latent variables have only
two indicators. This topic was addressed superficially earlier in the chapter,
and will be addressed more thoroughly in Chapter 5. For now, let us just say
that there are some difficulties that can arise when solving models, and having
three indicators provides more information with which to work. I should also
add, however, that when such problems arise, there are simplifications, such as
specifying equal loadings for two-indicator latent variables, which can make
model identification possible. One additional point that can be valuable is to
make sure that when you have two indicators of a latent variable they are scaled
so as to be positively correlated, otherwise it is not possible to set the loadings
as equal.
A second reason to include three or more indicators is to determine more
objectively the degree to which the indicators exhibit commonality. When only
two indicators are used, as long as the correlation between them is decent
(greater than 0.7), they are likely to fit and they will probably contribute equally
to the latent variable. When there are three or more indicators, we can assess not
only the degree of correlation among indicators, but the equality of correlations.
When two of three indicators are highly correlated with each other, but much
less well correlated with the third, that is a warning sign that they may not form
a stable latent variable. The consideration of a number of indicators affords
98 Structural equation modeling and natural systems

the researcher a broader view of the quantitative generality that applies to a

concept.
We should not take the recommendation of using several indicators too far,
however. In the natural sciences there may often be a single or a short list of
measures best suited to measure the property of interest. Because of the nature
of our data, the use of single- or few-indicator latent variables can be expected
to be more common than in the social sciences, where survey data allows for
measurement instruments based on many indicators to estimate highly abstract
constructs.

Cautions
Individual latent variables are said to behave as single causes. From a philo-
sophical standpoint, we must recognize that all single causes can be broken
down into parts, ultimately to the quantum level. For this reason, we should
understand that a latent variable is, to a certain degree, a convenience. Con-
cepts (and latent variables) emphasize the common attributes of a particular set
of items, while de-emphasizing their unique attributes. Ultimately, the value
of this convenience is judged by its utility. Perhaps it is Bollen (1989, page
180) who summed it up best: “Do concepts really exist? Concepts have the
same reality or lack of reality as other ideas. They are created by people who
believe that some phenomena have something in common. The concept iden-
tifies that thing or things held in common. Latent variables are frequently used
to represent unmeasured concepts of interest in measurement models.”
Figure 4.11 attempts to further clarify the unidimensional nature of a latent
variable and how this influences its role in path relations. Simply put, ξ only
represents the information in common among x1 –x3 , it does not represent their
unique information. Stated in another way, the model in Figure 4.11 does not
represent the effects of x1 –x3 on y. It represents the effects of the unmea-
sured factor that makes x1 –x3 correlated with y. Furthermore, even if ξ is the
unmeasured cause of joint correlation in x1 –x3 , if x1 –x3 are not well correlated
(i.e., unless ξ exerts strong influence on the values of x1 –x3 ), ξ will not behave
as a single latent cause, and we may obtain complex model relationships of the
sort discussed later (Figure 4.19). Generally, correlations among a set of indi-
cators need to exceed about 0.7 and be roughly equal, for them to adequately
estimate a latent variable using covariance procedures.
One additional caution to consider when dealing with latent variables has to
do with the names selected. The naming fallacy refers to the fact that putting a
name on a latent variable does not necessarily guarantee that we have measured
the concept embodied by the name. If we claim we are studying effects of body
 The anatomy of models II: latent variables 99

d1 x1 z
x2
d2 x y
d3 x3
Figure 4.11. Venn diagrammatic representation of shared variance among indica-
tors. Each indicator, x1 –x3 , shares some variance with the others (represented by
vertical hatching), and the degree of overlap between them represents their corre-
lations. Only the variance shared by all three variables is linked to the unmeasured
latent variable. Error variances for each variable can be thought of as that part of
the variance not in the hatched portion. This unshared variance includes both the
true error and the unique information contained in the indicator.

size, but we only have a single measure, we have done little to bolster our claim.
The solution to this potential problem is a dedicated effort to demonstrate the
generality of the factor being estimated, which can be accomplished by studying
the relationships among a large set of related indicators. So, if we show that
a latent variable composed of numerous indicators reasonably related to size
does indeed evidence unidimensionality, we have greatly strengthened our case.
It is also worthwhile to be aware that in path-focused models, the predictive
context can subtly influence the meaning of a latent variable. When we have
a model that relates body size to territory size, such as in Figure 4.9, there are
two different ways we can describe what that model is about. First, we might
interpret this model as addressing the question, what is the effect of “body size”
on “territory size”? This is literally what the model states. This statement also
implies that the concept “body size” is independent of the other latent variables
in the model. However, there is a tendency for the researcher to select particular
indicators of body size that most strongly correlate with indicators of territory
size. When we do this, our latent variable “body size” is better described as
“the attributes of body size that influence territory size”. Since a single word or
phrase is always a shorthand statement for a more complex idea, it behooves
us to describe carefully the meaning of our latent variable and the context in
which it is to be interpreted. This is not a problem unique to SEM applications.
Rather, the unique thing about SEM is that we give an explicit and operational
statement of the meaning of our latent variable by the indicators we use to
represent it. As stated earlier, this is one of the many strengths of SEM.
100 Structural equation modeling and natural systems

When we dig into it, we find that there is much to be learned from a dedicated
study of correlated indicators. This aspect of quantitative science has been
somewhat ignored in the natural sciences. There is much to be gained by an
explicit consideration of latent factors and their relationships to indicators. This
is where the topic of factor-focused analyses comes in, and is the subject of our
next major section.

Factor-focused applications
Some concepts related to factor-focused applications
Factor-focused applications have as their background a considerable literature
dealing with factor analysis. In factor analysis there is interest in identifying
some smaller set of forces (unmeasured factors) that can explain the correla-
tions among a set of observed indicators. Through history there have been a
number of approaches proposed for studying factors, including both empirical
methods (e.g., principal components analysis, also known as component fac-
tor analysis), and exploratory methods (such as exploratory factor analysis). In
SEM we consider factor-focused applications that fall within the tradition of a
confirmatory (i.e., theory-driven), common-factor approach.
Our presentation skips over quite a bit of background that may be of interest
to the beginning practitioner of SEM. For those who wish to test hypotheses
involving models with multiple indicators, a working knowledge of exploratory
factor analysis (as well as principal components analysis) is valuable, though
beyond the scope of our presentation. A good reference providing a relevant
and comprehensive treatment of this subject for those working in the natural
sciences is Reyment and Jöreskog (1996).
Figure 4.12 provides us with an overall comparison of principal components
analysis (PCA), exploratory factor analysis (EFA), and confirmatory factor
analysis (CFA) (as is incorporated in SEM). I need to point out that historically
there has been some confusion over the term “factor analysis”, which in the
past included both principal components analysis (also known as component
factor analysis), and what I am referring to here as exploratory factor analysis
(also known as common factor analysis).
The reader is perhaps already familiar with PCA, since it is commonly used
in the natural sciences. In PCA it is assumed that all observed variables are
measured without error. Thus, in our representation of a PCA model, no errors
are shown for the xs. The goal of PCA is to create predictors that represent the
correlations among measured variables, using a reduced set of variables. This
 The anatomy of models II: latent variables 101

PCA EFA CFA

x1 x1 x1
x2 C1 x2 F1 x2 LV1
x3 x3 x3
x4 x4 x4
x5 C2 x5 F2 x5 LV2
x6 x6 x6
Figure 4.12. Graphical comparison of principal components analysis (PCA),
exploratory factor analysis (EFA), and confirmatory factor analysis (CFA).

is accomplished using a rotation procedure that maximizes the variance of the

predictors, while minimizing the deviations. In Figure 4.12, I have shown the
arrows pointing from the observed variables to the components to represent the
fact that components are linear composites of observed variables. Note, how-
ever, that what is usually reported for a PCA are correlations between observed
variables and predictors, rather than the weightings used to create the predic-
tors. Typically, principal components are derived so as to be orthogonal to one
another (i.e., uncorrelated), though oblique (correlated) solutions are possible.
In EFA, it is assumed that some of the variance associated with the observed
variables is error, and factors are estimated so as to explain the common variance
among variables. As stated above, because of the emphasis on common variance,
factor analysis is sometimes referred to as “common factor analysis”. Because
EFA places a higher priority on identifying causes (rather than predictors) than
does PCA, it is more common for an oblique rotation to be used, whereby the
factors may be correlated. Certainly there is no reason why causes should be
uncorrelated in nature. In both PCA and EFA, all factors or components are
allowed to load on all observed variables.
Confirmatory factor analysis differs in important ways from PCA and EFA.
First and foremost, our factors (which are latent variables) and their linkages
to indicators are posited based on theoretical grounds. Again, as with EFA, it
is assumed that the factors are error free while the indicators possess error.
102 Structural equation modeling and natural systems

However, the data are not subjected to an automatic process designed to

maximize common variance among all indicators. Also, rotation is not used
to form factors so as to satisfy some predetermined criterion. Rather, in CFA
we seek to test whether the patterns of relationships in the data are consistent
with those expected from the model. In CFA, the factors are allowed to corre-
late or not, without a requirement for either oblique or orthogonal solutions. In
CFA, as in EFA, ordinary least squares estimation is not capable of deriving
parameter estimates for analyses involving factors; rather, maximum likelihood
or some other iterative method must be used to arrive at weightings that satisfy
the model. What is most distinctive about CFA, though, is that it permits the
evaluation of theory-based expectations. Both EFA and PCA are, in contrast,
simply empirical descriptions of the data. The latent variable scores estimated
in CFA are computed so as to best represent underlying unmeasured causes,
consistent with the objectives of SEM.
Figure 4.13 presents in more detail the structure and notation of a multi-factor
model. One way we can recognize a confirmatory factor model (in contrast to
exploratory models) is that the arrows from both factors do not connect to all
indicators. As has been stated, it is a hallmark condition of confirmatory factor
models that loadings and crossloadings are based on theoretical grounds. Also,
it is the constraint of not specifying all cross loadings that permits models to be
identifiable and for unique solutions to be obtained (in EFA, unique solutions
are not achieved and many different combinations of factor scores can satisfy
the model – a condition called indeterminacy). Structural equation modeling is
about working with confirmatory models, and to put it in the simplest terms,
we evaluate a CFA by determining whether the data match the expectations
imposed by the model. In spite of the limitations, EFA can still be valuable
when pre-existing theory is insufficient to provide adequate guidance and we
are getting started towards a CFA. In contrast, PCA is generally an unreliable
method for leading to interpretable factors.
Now, there is another distinction of importance to mention. The relationships
in the model in Figure 4.13 can be represented by the equation

xi = λi j ξ j + δi (4.8)

where λij represents the loadings and δ i the errors. We should be aware that
the error term δ i is actually composed of two parts. This can be illustrated by
decomposing the variance of xi as follows:

σx2 = σc2 + σδ2 = σc2 + σs2 + σe2 (4.9)

 The anatomy of models II: latent variables 103

d1 x1 l11

l 21
d2 x2 x1
l31
d3 x3
F12
d4 x4 l42

l52
d5 x5 x2
l62
d6 x6
Figure 4.13. Diagrammatic representation of a factor analysis model. The term δ i
represents error for the indicator variables (xi ), λij , represents the loading of factors
εj on indicators, and  represents the covariance between factors.

where
σc2 is the common variance or communality, which is the variance shared
among indicators of a factor ξ ,
σδ2 is the residual variance or uniqueness, which is the variance of x un-
accounted for by the factor,
σs2 is the specific variance or specificity, which represents the variance spe-
cific to a variable that is not true error, and
σe2 is the true error variance, which is solely due to measurement error and
which can be estimated using repeated trials.
In practice, specific variance and true error variance are not distinguished and
these two components are combined in the error variance variables, δ i .
104 Structural equation modeling and natural systems

Illustration of the relationship between path coefficients and bivariate

correlations in multi-factor models
It was instructive in a previous section to relate path coefficients to bivariate
correlations. Here I extend that illustration one step further by considering
models containing multiple factors of the sort shown in Figure 4.13. Once again,
the eighth rule of path coefficients allows us to determine the expectations,
although we have to trace the connections among variables back a little further,
through the correlation between ξ 1 and ξ 2 . So, the expected correlation between
x1 and x4 can be seen to be λ11 × 12 × λ42 .

Types of validity in multi-factor models

We considered a basic definition of the concept of validity earlier. Here we revisit
that concept as it relates to multi-factor measurement models, such as shown in
Figure 4.13. There is actually a rather long list of types of validity that have been
considered in structural equation models. A modern discussion can be found
in Kline (2005). To simplify the discussion somewhat, we can distinguish
between assessments of validity that involve external versus internal compar-
isons. When we compare our factor to some externally established criteria,
we may speak about criterion validity. For example, does our factor correlate
strongly with some key indicator? Two key types of internal validity are conver-
gent validity and discriminant validity, which evaluate the patterns of loadings.
Convergent validity refers to the necessity for indicators of a latent variable to
intercorrelate to a moderate degree. In contrast, discriminant validity refers to
the necessity that indicators of different latent variables do not intercorrelate
too strongly. In other words, for a model such as the one in Figure 4.13, it is
generally necessary that the correlations within the two groups of indicators,
x1 –x3 and x4 –x6 , should be stronger than the correlations between these groups.
Otherwise, the data will not show consistency with a two-factor model.

Examples of factor-focused analyses

There are only a few published examples of ecological applications where
strong emphasis has been placed on discovering the underlying factors causing a
suite of indicators to be correlated. A classic factor-focused biological example
of this was from Wright (1968) who analyzed the skeletal characteristics of
leghorn hens. Aside from the original paper, treatments of this example can be
found in Reyment and Jöreskog (1996) and Shipley (2000). A second example
is the reanalysis of morphological traits of house sparrows by Pugesek and
Tomer (1996). This example is also presented in Shipley (2000) as part of a
 The anatomy of models II: latent variables 105

discussion on latent variables. Here I use data sets dealing with intercorrelated
soil properties to demonstrate how correlations can be understood in terms of
latent soil properties. I should emphasize that the results presented in these
analyses do not represent a thorough treatment of the subject, but serve as an
example.

Intercorrelations among soil properties

A Louisiana coastal grassland
Grace et al. (2000) measured a suite of soil properties as part of a study of spa-
tial variations in plant community structure. The measurement of various soil
properties is common in ecological studies, although it is rare for ecologists to
have evaluated hypotheses about the factors causing soil properties to be inter-
correlated. In McCune and Grace (2002, chapter 31) and Grace (2003b) these
data were subjected to a variety of exploratory analyses as well as confirmatory
factor analysis. The measurement model that resulted from these analyses is
shown in Figure 4.14. Insight into the procedural steps involved in conducting
a factor-focused analysis will be left for a later chapter. Here, I continue with a
focus on the anatomy and interpretation of our model.
According to the results of this analysis, the observed correlations among
soil properties can be explained by two general factors (latent variables) and
three processes that result in correlated errors. Note that it is just as necessary
to provide a justification for correlated errors as for other parameters in the
model. This is an important point to make, because it is usually possible to
make a factor model fit if enough correlated errors are included.
An understanding of the theory behind the model in Figure 4.14 comes,
in part, from the physical setting of this study, and what is known about soil
formation in the grassland being studied. The study area was a coastal post-
pleistocene wet prairie, strongly and conspicuously influenced by the inclusion
of hundreds of small natural mounds per hectare. These mounds generally
possess a coarser soil texture and their presence creates a topology of micro-
elevational variation. In the low areas, hydric soil conditions are evident by the
presence of hydrophilic plants. Sampling was conducted along transects that
passed through both mound and intermound areas, capturing the full span of
variation in this system.
Because of the physical setting and soil formation processes operating at this
site, it was expected that soil properties would be correlated with microelevation.
It was also expected that mineral elements would be related to spatial varia-
tions in soil composition, while soil carbon content, soil nitrogen, and soil pH
106 Structural equation modeling and natural systems

1.0
0.0 elev ELEV

0.51 Ca 0.6
6
0.48
0.43 0.71
0.79 Mg
0.99
0.11 Mn
0.08 MINRL −0.67
0.66
0.27 Zn 0.53
9
0.3
0.29 K 0.
8 −0.33
3
0.8 1
4
0.22 P 0.93

−0.66
0.55 pH HYDR
0.78
0.38 C 7
0.27 0.6
0.53 N
Figure 4.14. Measurement model relating prairie soil variables to latent factors
(modified from McCune and Grace 2002). ELEV refers to elevational influences,
MINRL to soil mineral influences, and HYDR to hydric influences. All coefficients
presented are standardized values.

would be influenced by hydric soil effects. Soil chemical processes were

expected to result in a number of complications to the correlational structure.
Soil minerals, such as calcium phosphate and magnesium phosphate, are com-
mon and expected to contribute to correlations among elements. Finally, in this
study, both carbon and nitrogen content were measured simultaneously using
a CHN analyzer. This could be expected to result in correlated errors based on
methodological effects.
There is sufficient complexity in the correlations among soil parameters to
make the picture a little cloudier than one might like. The model in Figure 4.14
may or may not be definitive for this system. It is likely that other processes
and historical influences not considered in this model are important to explain
some of the discrepancies between data and model. Nevertheless, at least a
major portion of the correlational structure among parameters is consistent
with hydric and mineral influences. Relating these two soil properties to the
 The anatomy of models II: latent variables 107

0.23
Ca 0.8
8
0.47 0.73
Mg
−0.43 MINRL
0.69
Zn 0.71 3
0.23 0.30 0.53
0.19 .65
K 0.30
0.29 −0.97
HYDR
pH
0.91
0.18
org
Figure 4.15. Measurement model of prairie soil conditions (derived from Weiher
et al. 2004). MINRL refers to mineral influences while HYDR refers to hydric
influences.

results of a community ordination (Grace et al. 2000) was quite satisfying as

an explanation for the variations in plant composition across the site.

A Mississippi grassland example

Weiher et al. (2004) conducted a study in a tallgrass prairie in Mississippi which
was designed to examine the factors controlling plant species richness. Also
included in this study was the measurement of a variety of soil parameters.
The parameters measured included calcium, magnesium, zinc, potassium, pH,
and soil organic matter. The relationships among soil factors were examined
for these data, using the model shown in Figure 4.14 as our initial theory, with
missing variables (elevation, manganese, phosphorus, and nitrogen) deleted.
Since the earlier study by Grace and Pugesek (1997) showed that soil carbon
and soil organic content are highly correlated, the substitution of soil organic
for soil carbon was considered to be acceptable. Another difference was that
the pH data used in the model in Figure 4.14 were determined using a base-
saturation method, while those in Weiher et al. (2004) were readings of aqueous
extracts. Analysis of the initial model indicated (1) that the errors of calcium
and magnesium were not correlated, (2) that there was a correlation between
the errors for potassium and zinc, and (3) that pH cross-loaded with both the
MINRL and HYDR factors. The resulting model had a good fit to the data and
is presented in Figure 4.15.
108 Structural equation modeling and natural systems

Table 4.3. Correlations and standard deviations for prairie soil characteristics
(from Grace et al. 2000). N = 105. Correlations shown in bold are significant
at the 0.05 level

Ca Mg Zn K pH C Biomass
Ca 1.0
Mg 0.790 1.0
Zn 0.468 0.450 1.0
K 0.400 0.350 0.687 1.0
pH 0.284 0.309 −0.054 −0.224 1.0
C 0.039 0.145 0.491 0.550 −0.410 1.0
Biomass 0.288 0.137 0.279 0.163 −0.127 0.120 1.0
Std. dev. 1.58 0.447 0.746 1.437 0.271 2.162 1.350

A return to the Louisiana data

Because the two models presented above are not based on the same suite of
indicators, here I return to the data from Grace et al. (2000), and use a subset
of the data that is more equivalent to that in Weiher et al. (2004). To make
the data sets equivalent, pH values based on aqueous extracts (which were
collected but not used in the original analysis) are now used. Of interest in this
more comparable analysis is the question of whether the two grasslands (which
possess quite different soil types) both show a similar factor structure. The data
for this analysis are presented in Table 4.3. Now we will look into the results a
little more deeply.
It is conspicuous when we examine data such as the correlation matrix in
Table 4.3 that most soil characteristics are intercorrelated. Of the 21 possible
correlations, 14 are significant with only a sample size of 107. Only a couple
of correlations (calcium with carbon, zinc with pH) approach 0. Clearly, the
relationships among variables are potentially complex.
It is encouraging to find that when comparable data are used, the factor
loadings and cross loading are qualitatively the same for the Louisiana and
Mississippi data (Figure 4.16). These results tend to support the interpretation
that there are similar soil formation processes at the two sites, even though the
Mississippi site does not have mounds or recent coastal influence. However,
some coefficients clearly suggest additional complications (e.g., MINRL to Zn).
I should point out that, in general, comparisons among data sets should rely
on the unstandardized parameters, rather than the standardized ones as here
(see Chapter 3). Ideally, a multigroup model (Grace 2003a), a common model
type in SEM, would allow us to evaluate directly two samples in an integrative
framework. Such a direct comparison is not possible in this case, however,
 The anatomy of models II: latent variables 109

0.47
Ca 0.7
3
0.28
0.52 0.70
Mg
0.29 0.56 MINRL
Zn 0
0.4 3
5
0. 0 0.17
0.29 .54
K 0.68
0.42 −0.65
HYDR
pH
0.44
0.75
0.44
C
Figure 4.16. Measurement model of coastal prairie soil from analysis of data in
Table 4.3. All values are standardized coefficients.

because several of the variables were not expressed on a common basis, or were
not measured using the same technique, making unstandardized coefficients
incomparable and a rigorous evaluation not possible. Thus, our comparison
between Figures 4.15 and 4.16 can only be rather crude in this case, because of
differences in the data between the studies.
Several of the specifics differ between models (Figure 4.15 versus 4.16).
Based on a superficial examination, (1) the correlation between factors is weaker
in the Mississippi data, (2) the correlated errors differ between models, and (3)
some of the path coefficients are rather different. One’s satisfaction with the
similarities between the two sets of results could easily vary depending on
the objectives of the investigator and the state of their knowledge. This is the
first such comparative analysis of soil intercorrelations that I have seen, and
the similarity between the two sets of results was substantially greater than
expected given the differences between the two ecosystems and the differences
in sampling. The general similarity in results is encouraging for those wishing
to compare the effects of soil conditions on plant communities across systems.
Substantially further work will be required, however, if we are to determine
whether general factor models exist for a range of soils.

Hybrid models
Modern SEM practice and most treatments of the subject are based on the
premise that analyses are fully focused – as much about the measurement model
110 Structural equation modeling and natural systems

as about the path relations. This is partly due to the fact that in the social sciences,
where the dedication to SEM has been the strongest, it has long been recognized
that there is a compelling case for latent variables. In addition, there is also an
abundance of data, such as survey data, that are well suited to factor-focused
investigation. Another reason that the tradition in SEM is for fully focused
analyses using hybrid models is that (1) experience has taught us that latent
variables are valuable, and (2) a thorough understanding of latent variables
requires factor analytic studies.
The situation in the natural sciences is different. The application of SEM to
natural systems is in its infancy. The work has not yet been done that results
in a fundamental merger of path and factor traditions. This is one reason that
the presentation of ideas in this chapter is structured as it is. Most textbooks
on SEM first discuss factor analysis and only then bring latent variables into
path models. Here I have built from strongly path-focused models, starting with
single-indicator latents, to models containing multiple indicators.
The ultimate goal of SEM is to work with models that match both the
particular and general features of the situation. In this section, I will tran-
sition from the results from our example of a soil factor model to a model
that seeks to integrate both path and factor foci. I will then follow that up
with a more general discussion of the complexities that may arise in hybrid
models.

Relating soil properties to plant growth

In the study by Grace et al. (2000), the authors possessed data that included a
suite of soil parameters as well as several response variables relating to the plant
community. The original analysis examined how the full suite of environmental
factors related to community structure (McCune and Grace 2002, chapter 31).
Here, I illustrate a hybrid model by asking how the latent factors in Figure 4.16
relate to community biomass, a question that was not considered in previous
analyses. The proposed model is shown in Figure 4.17.
Note that in the model in Figure 4.17 we are asking how the common vari-
ance among soil indicators relates to biomass. Specific effects of individual soil
variables on biomass are not included in this model. This is a traditional hybrid
model, and our interest in latent variable effects on biomass is justified by our
previous finding that mineral and hydric influences (along with an unnamed
factor causing calcium and magnesium to correlate) provide an adequate expla-
nation for our correlated set of soil variables.
The results from fitting the model in Figure 4.17 to our data in Table 4.3
are not encouraging. Evaluations of model fit (discussed in the next chapter)
 The anatomy of models II: latent variables 111

Ca

Mg
MINRL
Zn
BIOM bio
K
HYDR
pH

C
Figure 4.17. Hypothesized hybrid model relating soil latent factors to plant com-
munity biomass.

indicate that the model has gone from adequate to inadequate when the response
variable biomass is included. Thus, our data do not match the expectations that
can be derived from the model and we must conclude that the model is not the
correct one for the situation. How can this be when the measurement model has
been shown to be adequate in separate analyses and all possible paths among
latent variables are included?
The answer to our dilemma can be found by looking for specific effects of
individual indicators on biomass. Such specific effects are often overlooked,
in part because most SEM software packages are not configured to alert the
researcher to specific effects. In fact, for most software packages (excluding
Mplus) we must specify a different kind of model to even include specific
effects from observed indicators on biomass. In this case, when our model is
appropriately configured, we are able to determine that there is a specific effect
of soil calcium on biomass that is in addition to the general effects of our latent
soil variables. With this reformulated model, we can approximate a satisfactory
solution, which is shown in Figure 4.18.
What is conspicuous from the results of this analysis is that calcium, by
itself, provides an adequate explanation for mineral influences on community
biomass. Thus, while we were successful in estimating latent variables that
explain the correlation structure among indicators, these latent variables do not
necessarily represent the effects of soil conditions on community biomass. I
cannot overemphasize the importance of this result to those wishing to employ
latent variables to understand path relationships. Just because there is evi-
dence for latent factors causing a suite of variables to correlate, that does not
112 Structural equation modeling and natural systems

0.33
0.48 Ca 0.7
2
0.29
0.53 0.69
Mg
0.56 MINRL ns
0.30 Zn 9
0.3 56 1.0
0. 0 0.17 BIOM bio 0.0
.54
0.30 K 0.68 0.22
HYDR R 2 = 0.124
− 0.65
0.38 pH
0.75
0.45 C

Figure 4.18. Model satisfactorily relating soil properties to community biomass,

including both the general effects of latent soil properties and the specific effect of
soil calcium on biomass.

necessarily mean that these latent factors are the causal agents affecting other
variables in the model.

General cautions about the use of hybrid models

A very considerable amount of the effort I have spent in learning about SEM
has been challenged by the need to make certain translations for a clear commu-
nication of its capabilities for use in the natural sciences. Some of the greatest
difficulties have had to do with limitations inherent in the software and its usual
applications. Part of this has been my own learning curve, rather than limitations
of SEM software or literature. I expect that others will also face the minefields
through which I have tiptoed as they apply SEM methods to their research
problems. I do my very best to point out to others in the natural sciences where
some of the mines reside. Some of the most important of these relate to the issue
just addressed about specific versus latent effects in path-focused studies. The
software and most presentations of SEM presume that hybrid models of the sort
shown in Figure 4.17 are the appropriate ones to examine. I will argue that it is
just as likely that models such as the one in Figure 4.19 may apply. There is no
inherent reason to expect that latent factors will capture the variance that corre-
lates with the other variables. For those engaged in analyses, with an interest in
path relations, both general and specific effects are of interest. When properly
applied, structural equation models incorporating multi-indicator latent vari-
ables can examine both general and specific effects, leading to a much greater
understanding of natural systems.
 The anatomy of models II: latent variables 113

A
x1 y1
x2 x h y2
x3 y3

B
x1 x1 h1 y1

x2 x2 x4 h4 h2 y2

x3 x3 h3 y3

Figure 4.19. Models demonstrating a complex mixture of general and specific

effects. In A, specific indicators, such as x1 , have effects on latent variables,
such as η, and latent factors such as ξ can have effects on a specific indicator,
y3 . In B, I have reformulated the model by specifying specific effects as single-
indicator latents, and by using secondary latent variables to represent our general
factors.

Conclusions
Addressing measurement issues can be valuable, both for increasing the accu-
racy of our analyses of relationships among concepts, and in the discovery or
validation of general factors that explain correlated effects. Structural equation
modeling provides us with an opportunity to correct for sampling and other
forms of nonreliability in our measurements. Specification of reliabilities is
generally easy to accomplish and valuable, even if our models only have sin-
gle indicators for estimation of our concepts. The use of multiple indicators to
specify general latent variables helps to add generality to our models. At the
same time, detailed factor-focused analyses enhance our ability to deepen our
understanding of multi-indicator latent variables. Often ecologists have appro-
priate data for the development of latent variable models, but have lacked an
understanding of measurement theory and the characteristics of latent variables.
When we say that an animal’s size influences the size of its territory, what do
we mean? Most of us have assumed all along that general concepts must be
114 Structural equation modeling and natural systems

quantified with specific attributes, unaware that a general concept with robust
statistical properties and theoretical components might be estimated and made
tangible.
For those of us in the biological and ecological sciences, there is a need
to consider carefully the value and procedures involved with using latent vari-
ables. Typically our theoretical knowledge is concentrated in the inner model
and largely absent from our measurement model. Thus, the adoption of latent
variables in our models requires that we become familiar with the implications
of relating latent concepts to one another. Ultimately latent variables represent
a potentially evolutionary step in the maturation of our theories. For this rea-
son, latent variables inspire us to seek more general models that match general
theories.
 5
Principles of estimation and model assessment

Introduction
In previous chapters, I have presented structural equation model parameters,
such as path coefficients, and we have considered their interpretation. Regres-
sion equations have been presented to help the reader understand the meaning
of parameters and how they are commonly expressed. However, we have not
yet considered the important question of how parameter values are estimated.
Since we are dealing with the solution of complex multi equational systems,
this is not a minor matter.
Historically, path models were solved primarily by the use of the familiar
technique of least squares regression (also referred to as ordinary least squares,
OLS). Today, most applications of SEM rely on model fitting programs that
offer a variety of options for estimation methods, as well as many other sup-
porting features. In this chapter our emphasis will be on maximum likelihood
estimation, both because of its central role in the synthetic development of mod-
ern SEM, and because it provides a means of solving nonrecursive and latent
variable models.
Another important issue that is related to the topic of estimation has to do
with the assessment of model fit. One of the most powerful features of SEM
is that techniques exist for comparing the observed relations in data to those
expected based on the structure of the model and the estimated parameters. The
degree to which the data match the model-derived expectations provides us
with the capability of evaluating the overall suitability of a model. This means
we have the capacity to reject models as inadequate and to compare models in
direct tests of alternative hypotheses. While not always a simple matter, model
evaluation is fundamental to the process of SEM and integral to arriving at
suitable parameter estimates. In this chapter I will consider basic principles
associated with estimation and model evaluation. In Chapter 8, illustrations of

115
116 Structural equation modeling and natural systems

model evaluation will be given, along with additional information on some of the
indices of model fit. However, before we can consider the topics of estimation
method and assessment of model fit, we must first contend with a basic issue
that underlies the estimation process, identification.

Identification
We say that a model is identified if it is possible to derive a unique estimate of
each of its parameters. Similarly, we may ask whether an individual parameter
is identifiable in a model, as this specific attribute relates to the overall question
of model identification. The problem in practice is that for some types of models
or for certain data sets, it is not possible to achieve this property of identification
and, as a result, the model cannot be solved for its parameters. Before going
too far into the complexities of this topic, let us begin with what should be a
familiar example, the solution of multi equation systems using algebra.
Consider that we have two algebraic relations,

a+b = 8
2a + b = 14

In this case, we can readily find the solution by solving the top equation for b,
(b = 8 − a), and substituting that expression into the bottom equation, which
leads us to the conclusion that a = 6. Substituting the solution for a from the sec-
ond equation back into the top equation allows us to derive a solution of b = 2.
By extension, it should be clear that when we only know one of the pieces of
information, say a + b = 8, we have a problem with one known value (8) and
two unknowns (a and b). There are any number of solutions that will satisfy this
single equation, such as 1 + 7, or 2 + 6, or 3 + 5, etc. Thus, we cannot achieve
unique solutions and our multi equational system is not identified. This case is
formally described as an underidentified model.

The t-rule
From this example we are able to recognize the most basic requirement for
solving structural equation models, which is also known as the t-rule, where
t, the number of parameters to be estimated, must be less than or equal to
the number of known values. When considering the analysis of covariance
structures, as we are in SEM, we can visualize our known information as the
matrix of variances and covariances. For example, if we have observed data
 Principles of estimation and model assessment 117

Table 5.1. Illustration of variances and covariances

of observed variables, which represent the known
information that can be involved in the estimation
of values for unknown parameters

x1 x2 y1 y2
x1 3.14
x2 0.72 0.26
y1 3.45 0.72 12.55
y2 1.65 0.36 3.85 9.67

g21
x1 y2
g11
r12 b21 z2
g22
x2 g12 y1
z1
Figure 5.1. Example of saturated model involving four observed variables.

for four variables, x1 , x2 , y1 , and y2 , we might obtain data such as shown in

Table 5.1. Here, we can count by hand the number of known values, which
equals the number of variances and covariances in the matrix, 10. We can
also arrive at this number using a simple formula n(n + 1)/2, where n is
the number of observed variables (in this case, 4). To illustrate the problem
further, we can visualize the number of unknowns for a saturated example
model involving x1 , x2 , y1 , and y2 shown in Figure 5.1. Here, the model param-
eters to be estimated include the variances of the four variables, plus the six
structural path coefficients (r12 , γ 11 , γ 12 , γ 21 , γ 22 , β 21 ). Note that the two dis-
turbance terms, ζ 1 and ζ 2 , can be derived from the other parameters using the
principles given in Chapter 3, thus, they do not have to be estimated indepen-
dently. So, there are a total of 10 parameters to be estimated and 10 known
118 Structural equation modeling and natural systems

A
g21
x1 y2
g11
r12 b21 z2

x2 g12 y1
z1
B
g21
x1 y2
g11
r12 b21 z2
g22
x2 y1 y12
g12
z1
Figure 5.2. Illustration of two alternative models involving four observed vari-
ables. In model A, the model is unsaturated and only involves the estimation of 9
parameters; thus, this model is overidentified since our variance/covariance matrix
contains 10 known values. Model B includes a correlation between the errors for
y1 and y2 . In this model there are 11 parameters to be estimated. As a result, the
model is underidentified.

parameters. To clarify, some of the estimation process in this case, specifically

the estimation of variances and the correlation between x1 and x2 , requires
only that we obtain the values directly from the variance/covariance matrix in
Table 5.1.
We can extend our discussion of model identification by recognizing other
kinds of models involving four variables (Figure 5.2). In Figure 5.2A our model
is unsaturated, since there is no direct path from x2 to y2 . This means there is
one less parameter to be estimated for this model (compared to the saturated
model in Figure 5.1). Since the number of known values remains at 10, our
model can be said to be overidentified, which means we have more knowns
 Principles of estimation and model assessment 119

than unknowns. As we shall see later, when a model is overidentified it enables

us to use the surplus information to provide tests of overall model fit (see section
below on assessing model fit).

Other identification requirements

To elaborate on the t-rule, the case in practice is that as a minimum, we must
have as many unique known pieces of information as we have parameters to
estimate. So, the question is, when would known values not represent unique
pieces of information? The short answer to that question is, when elements of
the variance/covariance matrix can be directly derived from one another, the
values are not all unique. Let us illustrate with a simple example. Consider that
we have two equations as follows:

a+b = 8
3a + 3b = 24

When we rearrange the top expression (b = 8 − a) and substitute into the bot-
tom one, we arrive at 24 = 24. While this is a true statement, it is not helpful.
The reason that a drops out of the picture when we rearrange the equations
is because the bottom equation is simply 3 times the top equation. Thus, the
second equation is redundant since it contains no unique information about
the relationship between a and b. So, while we have two equations, they do
not represent two unique pieces of information, but only one. The same thing
can happen in solving certain types of structural equation models. The most
common case is when two predictors (say x1 and x2 ) are very highly correlated
(e.g., r12 = 0.999). In this case, the two predictors actually represent only one
unique piece of information, because they are completely interchangeable. In
the estimation process, it is highly likely (though not guaranteed) that unique
solutions will not be obtained for the parameters. Most SEM programs (as well
as other statistical programs) will report that the matrix is nonpositive defi-
nite, which indicates that there is an apparent mathematical dependency among
elements of the variance/covariance matrix. In cases where it is the characteris-
tics of the data, rather than the structure of the model, that prevent the determi-
nation of unique estimates, this is referred to as empirical underidentification.

Identification in latent variable models

There are certain special identification issues associated with latent variable
models. As we have already mentioned in Chapter 4, when latent variables
120 Structural equation modeling and natural systems

A B
x1 x1 x2
* *
x1 x2 x1 x2 x3 x4

e1 e2 e1 e2 e3 e4
C D
x1 x1 y1 z1

x1 x2 x3 x1 x2

e1 e2 e3 e1 e2
Figure 5.3. Examples of latent variable models used to illustrate certain principles
of model identification.

are introduced into structural equation models, they bring with them additional
parameters to consider, such as their variances. They can also act to increase or
decrease the number of paths in a model. As a practical matter, evaluating the
identification of a structural equation model that includes both latent variables
and directed paths among the latent variables (so-called hybrid models) can be
addressed by separate evaluations of the measurement and structural models.
If the measurement model is identified and the structural model is identified, then
the full model will be identified. While these conditions are sufficient (except in
the case of empirical underidentification), they are not always necessary. Also,
it is important for the potential practitioner to note that there are a variety of
model parameter constraints that can be used to achieve identification, even
when the original model is not identified. Let us look at a few examples to
illustrate some of the possibilities (keeping in mind that a complete treatment
of this topic is beyond our scope).
Our first example in Figure 5.3 (model A) revisits an issue considered in
Chapter 4. Here we see the case of a latent variable associated with two observed
indicators. Since our data matrix only has one covariance (between the two indi-
cators), we lack sufficient information with which to estimate the two loadings
 Principles of estimation and model assessment 121

from the latent to each indicator. The way to solve this problem in practice is
usually to specify that the two loadings are equal, thereby reducing the number
of parameters to estimate. Note that for an equality specification to be reason-
able, the variances of the two indicators should be made equal as well (through
variable coding).
Model B in Figure 5.3 represents a case where, unlike model A, we are
able to satisfy the t-rule. Since there are four observed variables, we have
(4 × 5)/2 = 10 knowns. Including the variances of the xs, we have 9 parameters
to estimate. Therefore, our model is structurally overidentified. However, there
can be a problem associated with empirical underidentification for this model.
If the correlation between the latent variables is indistinguishable from zero,
that implies that the two constructs are independent, 2-indicator models. In this
case, we have the situation where we cannot identify both indicator loadings
for the latent variables, even though we satisfy the t-rule. The solution is the
same as for model A, equality constraints for loadings on each latent variable
will be required in order to arrive at unique solutions. Model C in Figure 5.3
represents a case where, assuming we specify the scale for our latent variable,
we satisfy the t-rule and otherwise should have an identified model.
Finally, model D in Figure 5.3 represents the case where a two-indicator
latent variable model is identified. Since there are three observed variables, we
have six knowns with which to estimate the six unknowns (3 variances and 3
path coefficients). Therefore, involving a two-indicator latent in regression is
another way in which we can estimate models containing latent variables with
less than three indicators.

Solving structural equation models – the parameter

estimation process
In the late 1960s and early 1970s, the search was on for a synthetic approach to
solving the broad range of models related to structural equations. From an histor-
ical vantage point, it can clearly be seen that there were a variety of approaches
being suggested as means to the end of a unified solution to multiequation mod-
els. This is most clearly illustrated in the collection of papers contained in the
volume edited by Goldberger and Duncan (1973). Standing out from the many
heroic efforts contained in that volume is a paper by Karl Jöreskog entitled, “A
general method for estimating a linear structural equation system.” Presented
in this paper in comprehensive, authoritative form is a method for analyzing
covariance structures that has come to be known as the LISREL model. The
LISREL model and its attendant philosophy and procedures have had a rev-
olutionary effect on statistical modeling. There have been many refinements,
122 Structural equation modeling and natural systems

Hypothesized Model Observed Covariance Matrix

x1

y1
y2 +
like
liho
od S = { }
1.3
0.24 0.41
0.01 9.7 12.3

um
xim
ma
compare
Model Fit
Evaluations

Parameter
Estimates Σ =
{ }
s11
s12 s22
s13 s23 s33
Implied Covariance Matrix

Figure 5.4. Simplified overview of the processes associated with parameter esti-
mation and model fit assessment.

and the body of methodology continues to grow and evolve. While we can
(and do) consider the many separate elements that constitute modern SEM, it is
really the synthetic system that evolved from this paper and the influence of the
associated LISREL software (developed with Dag Sörbom) that has become
the core of modern practice. A detailed historic perspective on this can be found
in Cudeck et al. (2001) “Structural Equation Modeling: Present and Future –
A Festschrift in honor of Karl Jöreskog”.
A simplistic representation of the overall process of parameter estimation
and assessment of model fit is presented in Figure 5.4. The elements of the
process represented in this figure are that (1) the hypothesized model implies
general expectations for the observed covariance matrix, (2) when the observed
matrix is combined with the model structure, we can obtain parameter esti-
mates, (3) the parameter estimates in combination with the hypothesized model
yield a model-implied (predicted) covariance matrix, and (4) comparison of the
implied and observed matrices allows for assessments of overall fit of data to
model.
There are several features that characterize modern SEM as represented in
Figure 5.4. First, SEM is based on the analysis of covariances instead of indi-
vidual observations. This is a strong departure from traditional statistical anal-
yses. At present, the continued evolution of SEM is incorporating both covari-
ances and raw observations, although this development is beyond the scope
of the present discussion. Another aspect of modern SEM is the reliance on a
comparison between observed and model-implied covariances. While methods
other than maximum likelihood can be used to perform such comparisons,
 Principles of estimation and model assessment 123

this approach is inspired by the maximum likelihood methodology, where

the emphasis is on simultaneous solutions (in contrast to the earlier path
model method of ordinary least squares regression), and on the differences
between observed and model-predicted covariances. These basic features of
modern SEM have profound implications, not just for the statistical proce-
dures, but also for some of the scientific attributes of multivariate hypothesis
evaluation.
Bollen (1989) has proposed the following general equation to represent
the fundamental consequences of modeling covariances using simultaneous
procedures, such as maximum likelihood:
 =  () (5.1)
Here,  is the population covariance matrix of observed variables and  ()
is the model-predicted covariance matrix expressed in terms of , the matrix
of estimated model parameters. In other words, this equation describes the
fundamental premise of SEM, that the population covariance matrix can be
understood in terms of a statistical model and its parameters. In practice, this
means that we are comparing the observed covariance matrix with the covari-
ance matrix implied by the statistical model. This comparison process involves
an overall test of model fit that plays a major role in multivariate hypothesis
evaluation.

Maximum likelihood estimation

Many of the fundamental properties of modern SEM are derived from the
simultaneous estimation process associated with maximum likelihood estima-
tion. Before proceeding further, however, we must recognize that we do not,
of course, know the population covariance matrix , only the observed covari-
ance matrix S, which serves as our best estimate. The estimation process, thus,
involves comparing the observed covariance matrix S with the covariance matrix
implied by the statistical model, ,
ˆ and the task is to choose values for the struc-
tural coefficients in  such that S is as close to 
ˆ as possible.
The process of maximum likelihood estimation is one that iteratively
searches through progressively refined estimates of parameter values until a
set is found that maximizes the likelihood that differences between S and  ˆ are
only those resulting from normal sampling error. Accomplishing this involves a
model-fitting procedure that minimizes deviations between S and . ˆ The most
commonly used fitting function in maximum likelihood estimation of struc-
tural equation models is based on the log likelihood ratio, which compares the
124 Structural equation modeling and natural systems

likelihood for a given model to the likelihood of a model with perfect fit. This
fitting function is commonly expressed as
 
FML = log 
ˆ  + tr(Sˆ −1 ) − log |S| − ( p + q) (5.2)

where there are p indicators of endogenous latent variables and q indicators of

exogenous latent variables (tr refers to the trace of a matrix). Some of the most
basic properties of Eq. (5.2) can be understood by recognizing that when the
sample matrix and implied matrix are identical, the first and third terms cancel
each other out, as do the second and fourth terms. Thus, our expected value for
FML is 0 when model and data match perfectly.
A number of assumptions are associated with Eq. (5.2) (Bollen 1989,
chapter 4). We must assume that the  ˆ and S matrices are positive definite
(i.e., that they do not have a singular determinant such as might arise from a
negative variance estimate, an implied correlation greater than 1.0, or from one
row of a matrix being a linear function of another), data follow a multinormal
distribution, and the observed covariance matrix follows from a Wishart dis-
tribution (which defines the probability expected when sampling from a multi-
variate normal population). Maximum likelihood estimators, such as FML , also
possess properties such that they are asymptotically unbiased (i.e., they are unbi-
ased in large samples), they are scale invariant, and they lead to best estimators
of structural coefficients. Collectively, the many characteristics of maximum
likelihood estimation have led to its central role in the structural equation
process.

Other estimation methods

There are other methods that exist for performing SEM. Some are simultaneous
“full information” analyses, such as described for the maximum likelihood
method. The characteristic of most importance about such methods is that they,
again, do not involve individual observations, but instead, analyze covariances.
One useful example of such an alternative estimation method is generalized
least squares (GLS). The GLS fitting function can be represented as
1
FGLS = tr(I − S−1 )2 (5.3)
2
where I is the identity matrix. This fitting function can be seen to be based
on squared deviations, thus, it is in the family of least squares methods. The
method is referred to as generalized because the data are weighted so as to
standardize against unequal variances, achieving the property of being scale-
free. The GLS method shares many of the same attributes and assumptions as
 Principles of estimation and model assessment 125

ML estimation. A more complete discussion of the various fitting functions,

including a number of alternatives that are of considerable value in particular
situations, can be found in Bollen (1989, pages 104–123).

Evaluating parameter significance and consideration of the

violation of distributional assumptions
While SEM is indeed a model-oriented approach to hypothesis evaluation,
individual parameter values still possess uncertainty, and can be evaluated
using p-values derived from the attending standard errors. The principles of
the Neyman–Pearson protocol, such as a critical cutoff value for “p” of 0.05,
and the desirability of limiting the number of comparisons, are still widely
practiced. As we shall see below, reliance on individual p-values for param-
eters is generally superseded by evaluations of overall model fit, which are not
only of priority importance, but also favor the hypothesized model over a null
hypothesis (in sharp contrast to the Neyman–Pearson protocol). Nevertheless,
evaluating the significance of individual parameters is an important element
in SEM. As with other applications of individual p-values, there are concerns
about the influence of data distributions on the accuracy of standard errors and
the accompanying p-values.
There are several points worth noting in this overview of the topic of param-
eter evaluation. First, an important distinction needs to be made in SEM prac-
tice between data that contain dichotomous response variables, and those that
are strictly continuous. The analysis of categorical response variables is an
advanced topic that will not be considered in any detail in this introductory
book. Nevertheless, if the reader is familiar with logistic or probit regres-
sion, they will understand the nature of the problem of analyzing dichotomous
responses. Several approaches are possible and much recent work has been
done to permit a thorough inclusion of dichotomous and nominal response vari-
ables in structural equation models (Muthén 1984). Several issues are involved,
only some of which relate to violations of parametric assumptions concerning
normality.
A second point to be made here is that, as in other methods for the analysis of
parametric data, there is no requirement that predictor variables meet paramet-
ric assumptions, such as normality. In any regression, parametric assumptions
relate to the deviations from the prediction line, and therefore only involve
the distributional characteristics of the response variables. This means, among
other things, that predictors can be dichotomous or of any other nonnormal form
(except unordered nominal) without impacting the assumptions of the analysis
process.
126 Structural equation modeling and natural systems

A third point to make is that issues of data distribution (as long as data
are continuous) do not generally impact the parameter estimates themselves,
only their standard errors and our inferences about the probabilities associated
with those parameters. Thus, while we may be concerned with sample-related
problems such as outliers and influential points, we will find that the absolute
values of our parameter estimates are uninfluenced by methods designed to
compensate for data nonnormality.
A fourth point is that in recent years there have evolved a number of methods
for evaluating and compensating for certain deviations from normality that are
widely available in modern SEM software packages. Among these, one that
should generally be avoided (but which still remains because of the historic
role it has played) is the asymptotic distribution-free (ADF) approach. While
the “distribution-free” part sounds appealing, the problem with the use of ADF
methods is the “asymptotic” part, which alludes to the very large sample size
requirements for the validity of such methods. Use of the ADF method has
largely ceased with the development of so-called “robust” methods (Satorra and
Bentler 1988, 1994). In a nutshell, such methods estimate the kurtosis in the data
for the response variables, and provide specific and appropriate adjustments to
the standard errors. Practitioners will find robust methods highly useful for
analyses. Further information on the subject can be found in Kaplan (2000,
chapter 5).
A fifth point is that resampling methods, such as Monte Carlo bootstrapping,
provide another means of dealing with nonparametric data distributions. In
bootstrapping, data are randomly sampled with replacement in order to arrive at
estimates of standard errors that are empirically associated with the distribution
of the data in the sample. The main advantage of this approach is that probability
assessments are not based on an assumption that the data match a particular
theoretical distribution.

Evaluating models by assessing model fit

I have stated earlier in the book that SEM philosophy is fundamentally differ-
ent from the tenets of conventional null hypothesis testing. In SEM, we do not
give priority to null hypotheses. Rather, because SEM takes an approach that
is theory-based, a-priori, theoretically justified models represent the starting
point for model selection. Also, while we are frequently interested in evaluating
alternative scientific hypotheses, there is a recognition that statistical hypothesis
testing is more of a means to an end, than the essence of the scientific endeavor.
To quote an unpublished manuscript by Hayduk, Pazderka-Robinson, and Zed,
 Principles of estimation and model assessment 127

“Researchers using structural equation modeling aspire to learn about the world
by seeking models whose causal specification match the causal forces extant
in the world.” Evaluating models by assessing the fit between expectations and
data, or the comparative fit between alternative models represents the method of
“seeking models”. Characteristic of a theory-based approach, when evaluating
structural equation models we consider substantive information when deciding
the adequacy of a model or when choosing a “best” model. While p-values are
used in evaluating both measures of model fit and individual parameters, a strict
reliance on arbitrary cutoffs such as 0.05 is not adhered to. Ultimately, the goal
is to estimate absolute and relative effect sizes, and to draw inferences about the
processes operating in a system, rather than to reject null hypotheses. Structural
equation modeling philosophy is, thus, fairly compatible with recent criticisms
of null hypothesis testing by biologists (Anderson et al. 2000), although our
approach to model evaluation and selection is somewhat different. In this sec-
tion, I describe some of the elements of model evaluation, with an emphasis on
model fit assessment.

The model-implied covariance matrix and the concept of residuals

As we have already seen, the estimation process uses model-implied covari-
ance matrices in the minimization function used to derive best estimators. This
iterative process arrives at a final “predicted” covariance matrix that is also
used in deriving measures of overall model fit. In addition to being involved in
calculations of model fit, individual predicted covariance matrix elements are
also useful. One of their uses arises from the calculation of residuals that repre-
sent the distance between observed and predicted variances and covariances. A
related use is the derivation of so-called modification indices, which are often
used to help determine if better fitting models exist.
An extremely simple example of the related concepts of model-implied
covariance and residual covariance is illustrated in Figure 5.5. Here standard-
ized data are shown to simplify the presentation. Given the model shown and
the data, we know from the second rule of path coefficients (Chapter 3) that our
best estimate for the parameter value for the path from x1 to y1 is 0.40. Likewise,
our best estimate for the path coefficient from y1 to y2 is 0.50. Based on these
results, the model-implied (also known as “fitted” or “predicted”) correlation
between x1 and y2 is 0.20, the product of the two paths. We can see from the
data matrix that the observed correlation between x1 and y2 is 0.35, thus, there
is a residual (unexplained correlation difference) of 0.15.
We, of course, have the question of whether the residual that is observed
is simply the result of sampling error, or whether it represents a significant
128 Structural equation modeling and natural systems

Hypothesized
Model x1 y2
y1

Data Matrix Model-Implied Matrix

x1 y1 y2 x1 y1 y2
-------- -------- -------- -------- -------- --------
x1 1.00 x1 1.00
y1 0.40 1.00 y1 0.40 1.00
y2 0.35 0.50 1.00 y2 0.20 0.50 1.00

residual = 0.15
Figure 5.5. Illustration of the related concepts of model-implied covariance and
residual covariance. For simplicity, covariances are standardized in this example,
although the principles generalize to unstandardized (true) covariances.

discrepancy between theory (our model) and the real world (the population
being sampled). This issue will be dealt with in the section that follows. For
now, it is important to realize that the very simple example shown in Figure 5.5
generalizes to the universe of structural equation models, including those with
latent variables. Appendix 5.1 provides a more technical and comprehensive
presentation of how structural equation models can be rewritten in terms of
the expected variances and covariances, and how these are represented in the
LISREL system.

Evaluating overall model fit – chi-square tests

It is important to recognize that even incorrect models can be estimated using
maximum likelihood or other simultaneous solution procedures. Under such
circumstances, the estimated structural coefficients may be the best achievable
given the model specified and yet fail to create a  ˆ that approximates S. This
problem is addressed by tests of the overall fit between 
ˆ and S, which determine
if there is a reason to believe that the proposed model might be an adequate repre-
sentation of the data. One of the most commonly used approaches to performing
 Principles of estimation and model assessment 129

such tests (the model χ 2 test) benefits from the fact that the maximum likelihood
fitting function FML follows a χ 2 (chi-square) distribution such that
χ 2 = (n − 1)FML (5.4)
Here, n refers to the sample size. In a model that includes paths between all
pairs of variables, no degrees of freedom are available for tests of overall model
fit. Rather, we say that the model is saturated and has perfect fit. However,
for models that have degrees of freedom, the χ 2 test is the most unambiguous
measure of overall model fit. Further, when comparing two models that differ in
a single constraint, it is possible to apply single-degree-of-freedom χ 2 tests to
assess the consequences of making or relaxing a constraint. Based on parametric
probability theory, a change in model χ 2 that is found to be greater than 3.84
indicates a significant difference between models.
When evaluating a model, the overall χ 2 should always be examined as a first
order of business. The χ 2 is of primary interest because of its fundamental rela-
tionship to the fitting function used in the estimation process (described above).
When χ 2 values have associated p-values greater than 0.05, authors often deem
a model to be acceptable. In my experience, in most situations when the χ 2
p-value is larger than 0.05, there are no pathways that can be added that will
have significant path coefficients based on individual t-tests.1 This condition is
not true 100% of the time, however, so it is still worthwhile examining the indi-
vidual residuals to see if any of them are pronounced (note that the χ 2 evaluation
is based on the sum of all deviations, therefore, many small deviations may yield
a χ 2 equal to that where there is one moderately large deviation). In addition,
models with p-values indicating adequate fit can frequently contain paths with
coefficients that are not significantly different from zero. In other words, the
model χ 2 test is not symmetric in its sensitivity to missing paths versus surplus
paths. Normally missing paths will be detected; surplus paths will not. For this
reason, the p-values for individual path coefficients are typically examined as
well. As will be discussed below, paths with p-values greater than 0.05 are not
always eliminated from the model, despite the fact that a p-value of greater than
0.05 for an individual path indicates that the path may be uncalled for. Again, the
philosophy of SEM is not to be a slave to arbitrary p-value cutoffs, but instead,
to consider both data and theory when evaluating the adequacy of models.
As with the estimation of standard errors for parameters, deviations from
multivariate normality can influence the estimated probabilities associated with

1 Note that when parameter estimates are based on maximum likelihood, the ratio of a parameter
to its standard error is actually a z-value instead of a t-value. T-values and z-values are virtually
identical except when sample sizes are very small.
130 Structural equation modeling and natural systems

the χ 2 . When data appear to have failed to meet the assumption of multivariate
normality by a wide margin, robust standard errors (Satorra and Bentler 1988,
1994) provide an alternative that seems to hold up well to substantial distri-
butional deviations. Resampling (e.g., bootstrapping) methods can be used as
well, in this case it is also possible to correct the evaluation of the overall model
χ 2 for distributional violations (Bollen and Stine 1992).
As just described, a nonsignificant χ 2 value (with probability >0.05) is
generally taken to mean that there is reasonable fit between the model and
the data. However, in practice the p-value associated with the model χ 2 test
is not always a workable gauge of model fit. One reason for problems is that
the p-values associated with a particular χ 2 are affected by the sample size.
Models with large samples sizes will show statistically significant deviations
even when the absolute deviations are small. This unfortunate property of the
χ 2 test in SEM has led to substantial efforts to develop measures of model
fit that are independent of sample size. Probably no other aspect of SEM has
received as much attention as the consideration and development of measures
of overall model fit. At present, two things are generally agreed to: (1) if the
model χ 2 has an associated p-value that is nonsignificant, the model can be
accepted (assuming the model is theoretically justifiable in the first place, and
the data meet distributional assumptions), and (2) if the χ 2 has an associated
p-value that is significant (usually, less than 0.05) and the sample size is large
(say, greater than 150), other fit measures may indicate that the model is never-
theless acceptable. It should be made clear that regardless of sample size, a sig-
nificant chi-square indicates deviations between model and data and this should
never be ignored, although one may decide that the deviations detected are not
of sufficient magnitude to reject the model entirely. The reader should be aware
that the issue of adequacy of fit continues to be debated among SEM specialists.

Other measures of overall model fit

The dozens of measures of model fit that have been proposed can generally
be classified into measures of absolute fit versus those that assess comparative
fit. Measures can also be cross classified as uncorrected versus parsimonious,
where parsimonious measures attempt to adjust for the number of degrees of
freedom. A third category of classification is whether indices are only suitable
for hierarchical (nested) comparisons (i.e., comparisons between models with
the same number and causal order of variables, differing only in the number
of pathways between variables), or can be used for non-nested models (models
that differ in the number of variables or order of pathways). The topic of model
fit indices has probably received more attention than any other aspect of SEM.
 Principles of estimation and model assessment 131

The literature on this subject is deep and not all of it leads to useful advice.
Therefore, I will not take the time here to discuss in detail the great variety
of alternative measures of model adequacy. The interested reader can consult
Bollen and Long (1993), Marsh et al. (1996), Kelloway (1998), or Schermelleh-
Engel et al. (2003). Consensus opinion at the moment suggests that multiple
measures should be considered when evaluating model fit, particularly when
sample size is large (Marsh et al. 1996, Kelloway 1998).
A substantial number of size-corrected indices have been developed as alter-
natives to χ 2 . One such index is the root mean square error of approximation
(RMSEA) proposed by Steiger (1990). This index is certainly not infallible and
there is some degree of controversy about its suitability. Nevertheless, it does
have certain desirable properties including (1) it is one of the few alternative
measures of fit that presents associated 95% confidence intervals and probability
values, (2) it is adjusted for sample size, and (3) it strikes a balance in sensi-
tivity with deviations in the structural model versus the measurement model.
Other indices include measures of goodness of model fit (GFI and AGFI), and
other summations of deviations between observed and predicted covariances
(e.g., the root mean residual, RMR). For measures other than χ 2 and RMSEA,
only rules of thumb are available, and no associated probabilities or error can
be calculated. In Chapter 8, the use of these and other measures of model
fit, including information-theoretic methods (such as the Akaike information
criterion described below), will be illustrated.

Non-hierarchical model comparisons

Comparing models of differing structure requires an alternative type of fit
index. Chi-square differences and related indices such as those already dis-
cussed can only be evaluated for significance when comparing hierarchically
related models. Assessments of model fit for non-hierarchical models can, of
course, be compared in the sense that if one model has inadequate fit and the
other has adequate fit (for a common data set), we may interpret one model
as being a superior representation compared with the other; one of the exam-
ples presented in Chapter 8 illustrates such an application. This approach will
however often not be suitable for the comparisons of interest.
In 1974, Hirotugu Akaike (Akaike 1974) employed information theory to
arrive at an index that has become increasingly used to compare alternative
models (either hierarchical or non-hierarchical). The resulting index, the Akaike
information criterion (AIC), can be viewed to be a modification of the log-
likelihood ratio criterion commonly used to evaluate model fit in SEM. The
modification to the maximum likelihood fitting function introduced by Akaike
132 Structural equation modeling and natural systems

was to correct for the fact that model fit is generally improved by including
additional parameters. To counter this problem and create an index that could
be applied to models of different structure (e.g., with different numbers of
variables included), Akaike sought to take model parsimony as well as model
fit into account. One expression of Akaike’s index is
AIC = c − 2d
where c is the model chi-square and d is the model degrees of freedom. Recall
from the presentation earlier in this chapter that the parameter c is one that
satisfies the criterion
c = sF[S, ()]
where s is the number of samples minus one and the function F is one that min-
imizes differences between the observed covariance matrix S, and the expected
covariance matrix based on the model (). The reader should be aware that
there are actually two different formulae for AIC in the literature, although they
produce equivalent results (see Kline 2005, page 142 for a brief discussion).
Bozdogan (1987) extended the work of Akaike by further correcting for
the effects of sample size on c (which are well documented), arriving at the
following
CAIC = c − (1 + ln S)d
where CAIC refers to the consistent AIC and S is the total number of samples.
Brown and Cudeck (1989), following a somewhat different line of reason-
ing, arrived at another alternative model comparison index, the expected cross
validation index
ECVI = (c/n) + 2(t/n)
where c is S – 1 and t equals the number of estimated parameters. The ECVI
was proposed as a way of assessing for a single sample the likelihood that the
selected model will apply across replicate samples of the same size from a
common population. While devised from a different starting point, the ECVI
shares with the other indices a capacity to compare models of similar or differing
structure.
Finally, indices of Bayesian information criteria (BIC) have been developed
to aid in model selection among structural equation models (Raftery 1993). Such
indices ask the questions which model predicts the data better, or alternatively,
under which model are the data more likely to be true? More will be said about
Bayesian approaches in the final chapter. For now, we can understand BIC to
be another information theoretic index, similar in principle to AIC but based
on Bayes factors.
 Principles of estimation and model assessment 133

The AIC, CAIC, ECVI, BIC, and sample size adjusted BIC are all employed
in the same general fashion to compare alternative models. It is common practice
to select the model with the lowest index value as the “best” model. For AIC, the
model thus selected is the one with the combined lowest chi-square and highest
degrees of freedom. For the CAIC and ECVI, there is an additional weighting
of models to account for differences in sample size. A further discussion of the
principles of model selection based on AIC and CAIC can be found in Burnham
and Anderson (2002).

Model modification and its interpretation

It is to be expected that users of SEM will commonly find indications of the
inadequacy of their models. This will be especially true in the initial stages
of the study of a problem. In addition to indications from the various indices
of overall model fit, residuals representing the specific discrepancies between
S and  ˆ are used by the LISREL computer program to generate modification
indices. Consideration of such modification indices can indicate the reduction in
χ 2 that would result from including a relationship in the model that is currently
absent. It is important for us to understand when modification of a model is
appropriate, and how such changes are to be interpreted.
When a model is found to have an inadequate fit to a set of data, we must
recognize that our statistical model has failed and also that our estimates of
structural coefficients are deemed invalid. At this point, we may either sim-
ply reject the hypothesized model or, as is typically the case, we may seek to
discover a model that is consistent with the data. It is important to be aware
that any subsequent models developed or results achieved are exploratory until
an independent data set is obtained to evaluate the adequacy of the revised
model. This fundamental tenet arises from the fact that structural equation
model evaluation is a confirmatory process. By this we mean that the χ 2 test
is designed to confirm or reject our proposed model, not to inform us of what
model might fit the data if we had been clever enough to guess it in the first place.
When one explores a data set using SEM in order to generate a hypothesis, one
violates certain assumptions of the method. For this reason, model modifica-
tion and the interpretation of results from modified models must be handled
carefully.
While the fitting of structural equation models is a confirmatory evaluation,
the use of SEM in studying hypotheses can range from exploratory to com-
parative to strictly confirmatory. Jöreskog and Sörbom (1996) provide a good
overview of this issue in the following quote:
134 Structural equation modeling and natural systems

We distinguish among the following three situations:

SC – Strictly confirmatory. In this case the researcher has formulated one single
model and has obtained empirical data to evaluate it. The model [and its
parameter estimates] should be accepted or rejected.
AM – Alternative models. Here, the researcher has specified several models and on
the basis of an analysis of a single set of empirical data, one of the models should
be selected.
MG – Model generating. Here, the researcher has specified a tentative initial
model. If the initial model does not fit the given data, the model should be
modified and tested again using the same data. Several models may be tested in
this process. The goal may be to find a model that not only fits the data well
from a statistical point of view, but also has the property that every parameter of
the model can be given a substantively meaningful interpretation. The
re-specification of each model may be theory-driven or data-driven. Although a
model may be evaluated in each round, the whole approach has more to do with
generating than evaluating models.

In spite of the fact that SEM can be used in an exploratory mode, results
thus obtained must be considered tentative. It is particularly important that
substantive interpretations be given to any model modifications. Likewise, it
is not generally recommended that paths judged to be nonsignificant using an
arbitrary p-value of 0.05 not be set to zero and the model re-estimated, unless
one is willing to propose that the deleted path represents a process that one has
now decided is not expected to be important. It is also desirable when making
modifications that only the required minimum changes should be made, in order
to avoid overfitting (defined as “the ill-advised act of changing a model based
on results that are actually due to chance relationships in the data at hand, rather
than based on a revised mechanistic assumption”).
A second element required before the results from a modified model are
considered definitive is subsequent evaluation with a new data set. This require-
ment has been ignored by many. Yet, the requirement for further evaluation of
a modified model in SEM practice is the basis for the philosophy of viewing
repeatability as a primary criterion for theory acceptance. Hopefully, in future
applications there will be an increased awareness of the value of independent
confirmation of SEM model results.

Summary
In this chapter I have discussed the topics of parameter identification, parameter
estimation, and at least to a limited extent, some of the associated assumptions
and alternative procedures. The simplistic consideration of parameter identi-
fication covered in this chapter belies the potential complexity of this topic.
 Principles of estimation and model assessment 135

As will be discussed in Chapter 6, up to the present time, model identification

problems have prevented researchers from having the ability to solve a wide
range of models, particularly those that include composite variables. A further
discussion of these problems and some potential solutions is offered in the
next section. Bayesian approaches to SEM (also discussed in the next section)
can also help in some cases to eliminate problems associated with parameter
identification.
A major emphasis in this chapter has been the issue of model fit assessment.
A key feature of SEM is the ability to allow for the evaluation of models by
comparing observed and predicted covariances. While this kind of evaluation
does not demonstrate that a particular model is true, it can be very helpful in
leading the researcher to reject models that are inconsistent with the data. For
models that are consistent with the data, we still need to be careful to consider
equivalent models, and issues associated with interpretative validity remain.
Finally, Chapter 8 serves as an important companion to the present chapter,
as it provides real-world examples of model evaluation that help to provide a
tangible perspective on the practice of SEM.

Appendix 5.1. Technical presentation of the concept of

model-implied covariances and how they relate to the
LISREL model
In Chapter 5 an elementary treatment of the concept of model-implied covari-
ance is presented. In this appendix, I present a more comprehensive representa-
tion of the matter, and tie it into the mathematical framework of the LISREL sys-
tem. The first subsection dealing with model-implied covariances may provide
some insight into the mathematical underpinnings of modeling covariances.
Of more immediate use for the reader is the material in the second subsection
of this appendix dealing with the LISREL language. In particular, the names of
the various submatrices that are described have practical significance, both for
reading the SEM literature and for reading the output associated with most of
the leading software programs.

Model-implied covariances
As stated in Chapter 5, the following equation summarizes the formal problem
of expressing the relationship between data and model:

 = () (5.1.1)
136 Structural equation modeling and natural systems

 is the population covariance matrix of observed variables and () is the

model-implied covariance matrix expressed in terms of , the matrix of esti-
mated model parameters. The term () is sometimes expressed as .ˆ
To illustrate, take for example a simple regression model
y = γx +ζ (5.1.2)
where y is a dependent variable, x an independent variable, γ is the regression
coefficient, and ζ is the disturbance or error term. In structural equation terms,
both γ and ζ are understood to be structural coefficients, and in this case will
be the two elements of the theta vector in Eq. (5.1.1). Both the variance of y
and its covariance with x can be written in terms of the variances of x and ζ , as
well as the structural coefficients, as shown in Eqs. (5.1.3) and (5.1.4).
VAR(y) = γ 2 VAR(x) + VAR(ζ ) (5.1.3)
COV(x, y) = γ VAR(x) (5.1.4)
Through these equivalencies, a relationship is established between the statistical
model and its implied covariance matrix in terms of Eq. (5.1.1).
We now have an equation (5.1.5) that establishes expected values for the
variance of y and the covariance of y with x, based on the form of the equation
in (5.1.2) and the estimated structural coefficients. These expected values can
then be compared with the observed values to see how well the statistical model
fits the data.
  2
VAR(y) γ VAR(x) + VAR(ζ )
= (5.1.5)
COV(x, y) VAR(x) γ VAR(x) VAR(x)

To better appreciate the flexibility implied by Eq. (5.1.1), we can consider

a different type of example, a simple latent variable model. In this case, the
values of two measured variables, x1 and x2 , are hypothesized to be caused by
a single unmeasured latent variable ξ (xi ) as expressed by the equations (5.1.6)
and (5.1.7).
x 1 = ξ + δ1 (5.1.6)
x 2 = ξ + δ2 (5.1.7)
Each equation includes a term for a random disturbance variable δ. We recognize
that the latent variable ξ has both a mean and variance, with the variance
designated by φ, thus
VAR(x1 ) = VAR(ξ ) + VAR(δ1 ) = φ + VAR(δ1 ) (5.1.8)
VAR(x2 ) = VAR(ξ ) + VAR(δ2 ) = φ + VAR(δ2 ) (5.1.9)
 Principles of estimation and model assessment 137

and the covariance between x1 and x2 is defined by the variance of the common
latent factor φ. Now, Eq. (5.1.1) becomes
 
VAR(x1 ) φ + VAR(δ1 )
= (5.1.10)
COV(x1 , x2 ) VAR(x2 ) φ φ + VAR(δ2 )

As will be shown in the next section, it is possible to express implied covariance

matrices for a wide variety of structural equation models using the language of
LISREL.

The LISREL language

Keesling (1972), Jöreskog (1973), and Wiley (1973) almost simultaneously
developed the LISREL (linear structural relationships) model and its nota-
tion, which has become the standard for expressing structural equation models.
Jöreskog and Sörbom went on to develop the LISREL computer software
for performing SEM, which has played a central role in the application
of these methods, although there are now many other software packages
available.
The LISREL model is based on three matrix equations. The first is the equa-
tion describing the relationships among latent conceptual variables, sometimes
also referred to as the latent variable model or structural model,

η = Bη + Γξ + ζ (5.1.11)

This equation is composed of (1) three vectors of variables, (2) two matrices of
coefficients, and (3) two matrices of covariances. The three vectors of variables
are
η, an m × 1 vector of latent endogenous variables,
ξ , an n × 1 vector of latent exogenous variables, and
ζ , an m × 1 vector of latent errors for the latent endogenous variables,
where m is the number of latent endogenous variables and n is the number
of latent exogenous variables. Note that endogenous variables are dependent
latent variables (i.e., those with an arrow pointing towards them from another
latent variable. Exogenous variables are independent latent variables.
The two matrices of coefficients in Eq. (5.1.11) are
B, an m × m matrix of structural coefficients describing the relationships
among endogenous latent variables, and
Γ, an m × n matrix of structural coefficients describing the effects of exoge-
nous latent variables on endogenous ones.
138 Structural equation modeling and natural systems

The matrices of covariances are

Φ, an n × n matrix of covariances among latent exogenous variables, and
Ψ, an m × m matrix of covariances among the errors of latent variables.
The other two equations that compose the LISREL model represent the
measurement model, the relationships between latent and observed variables.
(Note that observed variables are also referred to as indicators, manifest vari-
ables, and measured variables, and that I switch between these terms regularly.)
These two equations separately describe how the exogenous and endogenous
latent variables are measured.

x = Λx ξ + δ (5.1.12)
y = Λy η + ε (5.1.13)

These equations include four vectors of variables

x, a q × 1 vector of observed indicators of the exogenous latent variables
contained in ξ ,
y, a p × 1 vector of observed indicators of the endogenous latent variables
contained in η,
δ, a q × 1 vector of measurement error terms for the indicators of the latent
exogenous variables in x, and
ε, a p × 1 vector of measurement error terms for the indicators of the latent
endogenous variables in y,
as well as two matrices of coefficients
Λx , a q × n matrix of coefficients relating xs to ξ , the vector of exogenous
latent variables, and
Λy , a p × m matrix of coefficients relating ys to η, the vector of endogenous
latent variables.
While not shown in Eqs. (5.1.12) and (5.1.13), there are two additional matrices
associated with the measurement model,
Θδ , a q × q matrix of covariances among the measurement errors for indi-
cators of the latent exogenous variables in δ, and
Θ , a p × p matrix of covariances among the measurement errors for indi-
cators of the latent endogenous variables in .
Now that we have equations and notation describing the main parts of
the LISREL model, it is possible to use them to express the complete set of
 Principles of estimation and model assessment 139

expectations for a full structural equation model. Equation (5.1.1) now takes
the form
| covariances | covariances |
| among ys | between ys and xs |
 = | - - - - - - - - - - -- -- - - - - - - - - - - - - - - - - - - |. (5.1.14)
| covariances | covariances |
| between xs and ys | among xs |

There is insufficient space in this short chapter to give the derivations of the
four submatrix elements. The interested reader is referred to Hayduk (1987,
chapter 4, section 4) for a lucid account of the process. It is helpful, nonetheless,
to present the equations for the four submatrices individually. First, the equation
for the lower right submatrix is

covariances among xs = Λx ΦΛx + Θδ (5.1.15)

This is essentially Φ, the matrix of covariances among latent exogenous vari-

ables (ξ ), premultiplied by Λx , the coefficients relating x to ξ , plus δ , the
covariances among indicator errors. Secondly, the equation in the upper left
submatrix describing the covariances among latent response variables is neces-
sarily more complex. Here I describe it in two stages, with the first giving part
of the equation in symbols and part in words, as

covariances among ys = Λy [covariances among ηs] Λy + ε (5.1.16)

Now, we fill in the words in Eq. (5.1.16) with

covariances among ηs = A(ΓΦΓ + Ψ)A (5.1.17)

where A = (I – B)−1 , the inverse of the beta matrix subtracted from the identity
matrix and A is the transpose of A. The third submatrix equation to define
(lower left quadrant) is

covariances between xs and ys = Λx ΦΓ A Λy (5.1.18)

which gives the expected cross-products between xs and ys in terms of their

measurement models (Eqs. (5.1.12) and (5.1.13)). Finally, the upper right quad-
rant of our general equation (5.1.14) contains the transpose of the equation in
the lower left (5.1.18)

[covariances between xs and ys] = Λ y AΓΦΛx (5.1.19)

140 Structural equation modeling and natural systems

Plugging Eqs. (5.1.15) through (5.1.19) into (5.1.14) gives the LISREL equiv-
alent of Eq. (5.1.1) in the form given by Jöreskog
|Λ y A(ΓΦΓ + Ψ)A Λ y + Θε Λ y AΓΦΛx |
= | | (5.1.20)
|Λx ΦΓ A Λ y Λx ΦΛx + Θδ |
This is not an equation that the reader needs to memorize. Rather, it is an
expression of the LISREL notation that describes how matrices can be used to
express, in a rather comprehensive way, the covariance expectations from a full
structural equation model that includes both latent variables and measurement
models.
 PA RT I I I

Advanced topics
 6
Composite variables and their uses

Introduction
It would seem that structural equation modeling holds the promise of providing
scientists the capacity to evaluate a wide range of complex questions about
systems. The incorporation of both conceptual and observed variables can be
particularly advantageous by allowing data to interface directly with theory.
Up to this present time, the emphasis in SEM has been on latent variables as
the means of conveying theoretical concepts. It is my view that this is quite
limiting. As we saw in the final section of Chapter 4, causal relationships in
a model may deviate quite a lot from the stereotypic “hybrid” model. In the
current chapter, I discuss the use of an additional variable type, the composite, in
structural equation models. In simple terms, composite variables represent the
influences of collections of other variables. As such, they can be helpful for (1)
representing complex, multifaceted concepts, (2) managing model complexity,
and (3) facilitating our ability to generalize. In my experience, these are all
highly desirable capabilities when representing ecological systems and, as a
result, I frequently find myself including composites in models.
While long recognized as a potentially important element of SEM, composite
variables have received very limited use, in part because of a lack of theoretical
consideration, but also because of difficulties that arise in parameter estimation
when using conventional solution procedures. In this chapter I tackle both the
theoretical and practical issues associated with composites. To accomplish this,
it will be necessary to introduce additional terms, as well as a framework for
deciding when a composite is appropriate to include in a model.
The question of what is appropriate in a structural equation model must be
judged relative to the theoretical concepts, as well as the variables measured,
and the nature of the data. In this chapter I introduce the idea of the “construct
model”, which is the theoretical precursor to the development of a structural

143
144 Structural equation modeling and natural systems

equation model. I will consider the various possibilities for how observed vari-
ables relate to constructs and the resulting model architectures. This transition
from theory to model represents a phase of SEM that is largely absent from
other general treatments. I believe it is the “missing link” in most modeling
exercises, and should be studied carefully by those interested in developing
their own structural equation models.

History
As early as 1964, Blalock pointed out the need to represent some constructs
using composites. In particular, when the indicator variables associated with
a construct are viewed as causes of that construct, rather than effects, a com-
posite variable can be used in a model to represent that association. Heise
(1972) extended the discussion by providing explicit demonstrations of several
examples where composites would be appropriate. In addition, Heise discussed
some of the problems with estimating models containing composites, and pro-
vided a limited set of solutions within a least squares framework. Periodically,
practitioners of SEM have continued to raise the need for careful considera-
tion of the best way to represent constructs given the available data. Bollen
(1984) cautioned that a set of indicators for a latent variable should possess
“internal consistency” (i.e., they should represent a single entity and, therefore,
would be expected to be highly intercorrelated and joint responsive). He went
on to point out that we should differentiate between the case where the con-
cept of interest is the cause of some set of observed responses (i.e., is a latent
factor), and the opposing case where the observed variables have causal influ-
ences on the concept (i.e., composites). A number of authors have continued
to discuss the issue (Bollen and Lennox 1991, MacCallum and Browne 1993),
with an increased interest in the subject emerging in recent years (Bollen and
Ting 2000, Diamantopoulous and Winklhofer 2001, Edwards 2001, Jarvis et al.
2003, Williams et al. 2003). This literature has provided a number of useful
insights, diagnostic procedures, and approaches to the question of how to treat
concepts with only causal indicators. Overall, however, a clear, comprehensive,
and satisfying framework for considering composites has only recently begun
to emerge (Grace and Bollen, unpublished). This chapter presents an overview
of that framework, along with examples, diagnostics, and procedures.

Terminology
A consideration of the subject of composites requires that we have sufficient
terminology to express clearly the relevant relationships. To begin with, I
 Composite variables and their uses 145

z1 0
d1 x1 x1 x1

d2 x2 x1 x2 h1 x2 h1
d3 x3 x3 x3

L→M block M→L block M→C block

Figure 6.1. Representation of blocks specifying relationships between manifest

(observed) indicators (in boxes) and latent or composite variables (in circles). L
refers to a latent variable or unmeasured cause, M refers to manifest variables, and
C refers to a composite, which is a direct product of some set of causes without
error. It is characteristic of my presentation of composites that they have zero error
variance.

distinguish between the underlying theoretical constructs of interest, and the

conceptual variables used to represent how the data relate to the constructs.
I define for this discussion two types of conceptual variable, latent variables,
which are unmeasured causes, and composites, which are variables that rep-
resent collections of causes. Composites can reflect the joint effects of sets of
either manifest or latent variables. A composite can thus be referred to as either
a composite of manifest variables or a composite of latent variables. I also
stress that the issue of how to represent a construct deals with the interrelations
between a latent or composite variable and the associated variables that give
it meaning. A block is the basic unit to consider and several types of block
are illustrated in Figures 6.1 and 6.2. The reader should be aware that in this
chapter I will use the terms “observed” and “manifest” interchangeably to refer
to the directly measured variables. The term “indicator” refers to an observed
or manifest variable that is associated with a conceptual variable.
In Figure 6.1 we see that, given a set of three manifest variables, x1 – x3 , there
are various ways they can be related to conceptual variables, depending on how
the data relate to the theoretical construct. The most common representation is
the L➙M block, where paths go from latent to manifest variables. In this block
type, the latent variable is used to represent an unmeasured cause or factor that
leads to correlations among manifest variables. A great deal has been written
about the properties of L➙M blocks, and they form the backbone of latent
variable modeling (Bollen 1989); this is the subject of Chapter 4. Some of the
expected properties of blocks of this type will be considered below in Example 1,
along with diagnostics that help us to evaluate whether either theory or empirical
relations are consistent with this representation of a construct. An alternative
terminology that has been used for this case refers to a “latent variable with
effect indicators”. There is a long history of the use of the alternative terms
146 Structural equation modeling and natural systems

“effect indicators” (as in the L➙M block in Figure 6.1 and “causal indicators”
(as in the M➙L block in Figure 6.1), and I also use these terms occasionally.
The M➙L block shown in Figure 6.1 represents the situation where mani-
fest variables have causal influences on a latent variable that possesses no effect
indicators (no outward path to a manifest variable). By referring to the concep-
tual variable as a latent variable in this case, we infer that a latent factor does
indeed exist, independent from our data despite the fact that we do not have
effect indicators by which it can be measured. Graphically, this property of
independent existence for the latent variable is represented by the presence of
an error variance ζ 1 , which implies that while the xs have causal effects on η1 ,
they do not completely define it.
In contrast to the M➙L block is the M➙C block, in which the conceptual
variable represented is a composite. Since the error variance is specified to be 0
for a composite, this signifies that it is completely defined by its causes. There
actually exist two kinds of composite – one is the fixed composite. In this type,
the loadings from causes are specified a priori to have particular values by way
of definition. An ecologically relevant example would be the importance value,
which is defined as the sum of the relative density, relative abundance, and
relative frequency (usually for a species within a community). A second type
of composite is the statistical composite, which is the type considered in this
discussion. A statistical composite represents the collective effects of a set of
causes on some response variable(s). Such a composite does not exist separately
from the associated response variables. In simplistic terms, the statistical com-
posite is akin to a multiple regression predictor, some weighted combination of
causal influences that maximizes variance explanation in one or more response
variables. This predictor can have theoretical meaning, and can be quite useful
in developing structural equation models that involve multifaceted concepts.
Composites can represent collections of influences from both latent and
manifest variables. Figure 6.2 presents examples of blocks where relationships
between conceptual variables are considered. The L➙L block represents effects
of latent causes on a latent response, where all latent variables have effect
indicators. This is the standard block type found in the hybrid model discussed in
most introductory SEM treatments. What is represented by this block structure
is that mechanisms of causal interaction are actually among latent factors, and
we can understand the covariances among a set of x and y variables based on
that mechanism. Also implied is that the effect indicators associated with each
latent variable serve as multiple measures of the latent factors. Because the flow
of causation is from the latent factors to the indicators, error terms are used to
represent how well the indicators are predicted by the model; while the latent
factors are presumed to be free from measurement error.
 Composite variables and their uses 147

d1 x1
z1
d2 x2 x1
d3 x3
h1 y1 e1
d4 x4
d5 x5 x2
d6 x6
L→L block

d1 x1 d1 x1
z1 0
d2 x2 x1 d2 x2 x1
d3 x3 d3 x3
h1 h1
d4 x4 d4 x4
d5 x5 x2 d5 x5 x2
d6 x6 d6 x6
L→L(no M) block L→C block
Figure 6.2. Representation of blocks specifying relationships between latent
causes and latent or composite conceptual variables. Block terminology is
described in the caption for Figure 6.1.

The L➙L(no M) block in Figure 6.2 represents causal influence from latent
variables that have effect indicators, to a latent variable for which there is
no effect indicator. This block type resembles the M➙L block in Figure 6.1,
except for the fact that latent causes replace the manifest ones. Similarly, the
L➙C block in Figure 6.2 is analogous to the M➙C block in Figure 6.1. The
relevant situations for these representations will be illustrated in the example
applications below.

Examples
Rather than continuing to elaborate on constructs, blocks, and composites in
the abstract, the remaining development of ideas, as well as the development
of formal notation, will be conducted in the context of specific examples. The
reader should keep in mind that while not representing an exhaustive treatment
of the subject, the examples chosen and procedures discussed in conjunction
with those examples are broadly applicable to the general issues associated with
incorporating conceptual variables into structural equation models.

Example 1: colonization of forest patches by understory herbs

Background
In a recent paper, Verheyen et al. (2003) addressed the influences of a variety
of factors on the ability of herbaceous plant species to recolonize forest stands
148 Structural equation modeling and natural systems

in a fragmented landscape. Their work can be viewed as an effort to evaluate

a dichotomy of pre-existing theories. On the one hand, island biogeographic
theory suggests that successful colonization will be limited by factors such as
the age of an island (in this case, a forest stand is equivalent to an island), and
the distance to a source population. On the other hand, the local conditions in
a forest stand (such as soil conditions and the abundance of competitors) are
frequently invoked as the ultimate limiting factors for successful colonization.
The relative importance of effects is what is of interest in this case.
In this example, the authors examined colonization success and related it
to forest stand attributes in a nature reserve in Belgium that has had a long
history of agricultural use and forest fragmentation. In this landscape, much of
the original forest cover was gradually removed, and in places, discrete stands
of trees were replanted, creating a mosaic of ancient and contemporary stands.
Land use history of the 360-ha preserve was reconstructed using maps dating
back into the 1700s, and aerial photographs dating back to the 1940s. Using the
historical records, the age of each forest stand in the landscape was determined.
Extensive surveys of forest herb species throughout the entire preserve were
conducted in 1999, along with characterizations of forest understory cover
and soil conditions in individual stands. Together, the sources of data allowed
the authors to estimate the number of colonizations by each herb species in
contemporary forest stands, and the distance to the nearest population in the
ancient stands, which was presumed to be a possible source for colonization. A
total of 180 forest stands were in the final data set, with ages from 1 to 195 years
old. All species that were included in the analyses are obligate forest understory
perennials.
The authors’ initial model of construct relations is shown in Figure 6.3.
As represented here, the authors were interested in differentiating between the
effects of two exogenous influences, landscape properties and soil conditions,
on competitor abundance and ultimately colonization success. The authors used
a two-stage analysis to evaluate the model in Figure 6.3, first deriving compos-
ite indices for the conceptual variables, and then using composite scores in
a path analysis of relations among constructed composites. Here I provide a
more formal illustration of how constructs and indicators can be evaluated, and
discuss issues related to the solution of structural equation models containing
composites. A selective subset of the data analyzed by Verheyen et al. (2003) is
examined here, and it is important to point out that the purpose is not to confirm
or contradict the biological conclusions of the original analysis, but rather, to
illustrate the use of composite variables. Data characteristics were considered
in this analysis and corrective actions (e.g., transformations) were taken where
needed; the particulars are not discussed further here.
 Composite variables and their uses 149

Soil Competitor
Conditions Abundance

Landscape Colonization
Properties Success

Figure 6.3. Construct model reflecting the presumed mechanisms causing herb
colonization success to relate to the characteristics of forest stands.

In this case, our starting point is the initial construct model summarizing the
core theoretical questions posed (Figure 6.3). We can start by considering the
question, what are the various ways we might model the relationships shown
in Figure 6.3? To address this question, the available indicators must be con-
sidered and then the linkages between indicators and concepts. The authors
measured eight variables in the process of developing indicators for the four
constructs of interest. Two indicators of landscape conditions were measured,
the estimated age of each forest stand and the distance from each reforested
target stand to the nearest source patch. Three indicators of soil conditions were
measured, including soil texture, soil moisture, and soil pH. Two indicators of
competitor abundance were measured, the cover of herbaceous competitors and
the abundance of understory shrubs. For the herb species whose colonization
will be considered in this example application (Lamium galeobdolon), shrub
abundance was not significantly related to colonization success (although herba-
ceous cover was). So, to simplify the example, I only use a single indicator for
competitor abundance, which will be herbaceous cover. The proportion of sam-
pling points where a species was found in a forest stand serves as the single
indicator of colonization success.
Figure 6.4 shows the presumed linkages between measured indicators and
conceptual variables based on the theoretical reasoning of the authors. In this
initial representation, no attempt is made to express the structure of the blocks,
therefore, concepts are related to observed indicators with nondirectional lines
rather than directional arrows. Disturbance terms are specified for Competitors
and Colonization, while the diagram is ambiguous as to whether Soil and Land-
scape possess disturbances or not, because the directions of influences are not
given.
150 Structural equation modeling and natural systems

pH

moisture Soil Competitors cover

texture

distance
Landscape Colonization col. freq.
stand age

Figure 6.4. Associations between indicator and conceptual variables in Example 1.

pH, x1
Soil Competitors
moisture, x2 x1 h1 cover, y1

texture, x3

distance, x4 Landscape Colonization

col. freq., y2
stand age, x5
x2 h2

Figure 6.5. Model A, all observed (manifest) variables (in boxes) are represented
as “effect indicators” associated with latent variables (in circles). Thus, all blocks
are of the L➙M type.

Possible model structures

There are several possibilities for how the relations in Figure 6.4 might be
modeled. The primary determining factor for deciding which of the possibilities
might be most appropriate is the nature of the causal relations. There are, in
fact, a great many possibilities, although here I will only consider a few key
contrasting types. One of these possibilities is that the most appropriate model is
a classic hybrid model of the sort introduced in Chapter 4. Its specific structure
is shown in Figure 6.5. In this figure we see that the block structure used to
represent relations between indicators and individual latent variables is of the
L➙M type. In this model, which will be referred to as model A, two of the
latent variables, Competitors and Colonization, are associated with one effect
indicator each. As appropriate for such relationships, arrows are directed from
the causes (the latent variables) to the indicators. The two other latent variables
in this model, Soil and Landscape, are associated with multiple effect indicators.
Causal interactions among constructs are presumed to be best represented by
interactions among latent variables in this representation, and direct interactions
among manifest variables are not specified.
 Composite variables and their uses 151

pH, x1
Soil Competitors
moisture, x2 hc1 he1 cover, y1

texture, x3

distance, x4 Landscape Colonization

hc2 he2 col. freq., y2
stand age, x5

Figure 6.6. Model B, soil conditions and landscape properties are represented
as components of M➙L blocks, with the associated manifest variables (x1 –x5 )
representing cause indicators.

A second possibility is that the multiple indicators associated with Soil and
Landscape are causal indicators. This type of structure is illustrated in model B
(Figure 6.6). Here, the arrows pointing from pH, moisture, and texture to Soil
represent the presumption that the measured soil properties jointly determine
the soil influences, rather than vice versa. A similar logic applies to distance
and stand age, which are presumed to contribute to Landscape in this model. In
this case, the multi-indicator blocks are of the M➙L type.
If we assume that the causal indicators in Model B are themselves imper-
fect measures, which is very reasonable in this case, then we might represent
the situation as shown in Figure 6.7 (model C). Here, Soil represents a multi-
dimensional construct related at least in part to True pH, True Moisture, and
True Texture (the use of the term “True” here does not mean that we have esti-
mated the true value of variables using our model, instead, it means that we
recognize that the true value of pH is latent or unmeasured). Similarly, Land-
scape is a multidimensional construct related to True Distance and True Age.
As the names of these latent variables indicate, this model explicitly recognizes
that our manifest variables are not perfect measures of the “true” parameter
values.
Discussion of the range of possible ways to represent the relations in Fig-
ure 6.4 would be incomplete without considering another alternative, which is
referred to as model D (Figure 6.8). In this model, the two multidimensional
constructs, soil conditions and landscape properties, are represented simply as
groups of variables. In this case, the only two block types are L➙M and L➙L.

Formal representations
To discuss the relationships embodied in our models more formally, we can
begin with the three characteristic equations of the LISREL model (the reader
152 Structural equation modeling and natural systems

True pH
pH, x1
x1

True Moist. Soil

moisture, x2 hc
x2
1 Competitors
he cover, y1
True Texture 1
texture, x3
x3

Colonization
True Dist.
he col. freq., y2
distance, x4
x4 2
Landscape
hc
True Age 2
stand age, x5 x5

Figure 6.7. Model C, which represents Soil and Landscape and their associated
causes as multidimensional constructs, allowing for measurement error in all man-
ifest variables. In this case, Soil and Landscape are components of L➙L (no M)
blocks. The line with arrows connecting the latent variables represents all possible
intercorrelations.

Soil Conditions

True pH
pH, x1
x1

True Moist.
moisture, x2
x2
Competitors
cover, y1
he
True Texture 1
texture, x3
x3

Colonization
True Dist.
distance, x4 col. freq., y2
x4 he
2

True Age
stand age, x5
x5

Landscape Properties

Figure 6.8. Model D, partially reduced form of model C omitting explicit con-
sideration of collective soil and landscape effects.

may wish to refer to Appendix 5.1, where the LISREL notation is summarized,
to better appreciate the following presentation). This notation, which applies to
hybrid models such as model A in Figure 6.5 is

x = Λx ξ + δ (6.1)
y = Λy η + ε (6.2)
 Composite variables and their uses 153

and
η = Bη + Γξ + ζ (6.3)
where x and y are vectors of observed indicators of the exogenous and endoge-
nous latent variables, ξ and η are vectors containing the individual exogenous
and endogenous latent variables, Λx and Λy are vectors of coefficients relat-
ing indicators to latent variables, B and Γ are coefficient matrices for effects
of endogenous and exogenous latent variables on endogenous latent variables,
δ and ε are vectors of measurement errors for x and y, and ζ is a vector of
disturbances for the η variables. For exogenous latent variables, ξ, their vari-
ances are represented by the diagonal elements of the matrix Φ, while the
off-diagonal elements of the matrix are the covariances. The disturbance terms
for the endogenous latent variables (ζ) are contained within the diagonal ele-
ments of the Ψ matrix, while the off-diagonal elements of that matrix represent
any covariances among disturbances (normally assumed to be zero). Error vari-
ables (in δ and ε) are expected to be uncorrelated with ξ and η. However, it is
possible to model correlations among errors, typically represented by nonzero
off-diagonal elements in the matrices representing cross products among mea-
surement errors, θ δ and θ ε .
In the case of a model such as model B, which contains causal indicators,
we recognize the need for an additional equation to represent the M➙L block
η c = Λc x + ζ c (6.4)
Further, for ηe variables in general,
η e = Bc η c + Be η e + Γξ + ζ e (6.5)
Here, for clarity, we distinguish two different types of endogenous latent vari-
able. Those with causal indicators, either directly as in the M➙L block, or
indirectly as in the L➙L (no M) block, are represented by the vector η c , while
those with effect indicators are designated by η e . We also distinguish two differ-
ent matrices of structural coefficients, Bc and Be , as well as separate vectors of
disturbance terms, ζ c and ζ e , for the two types of eta variable. In the particular
case of model B, there are no variables in the ξ vector. For model C, Eqs. (6.1)
and (6.2) apply as they did for model A. However, we now have the case where
we need to expand Eq. (6.4) to be more general, so that
η c = Λc x + Γξ + ζ c (6.6)
In the herb colonization example, Soil and Landscape (collectively represented
by the vector η c ) are affected by the exogenous latent variables in the ξ vector.
A summary of the equations that would apply to models A–D is presented in
Table 6.1. As we can see, models A and D can be described unambiguously using
154 Structural equation modeling and natural systems

Table 6.1. Summary of equations relevant for models A–D

Properties and expectations Model A Model B Model C Model D

Equation 1 applies yes no yes yes
Equation 2 applies yes yes yes yes
Equation 3 applies yes no yes yes
Equation 4 applies no yes no no
Equation 5 applies no yes yes no
Equation 6 applies no no yes no

the conventional LISREL notation (Eqs. (6.1)–(6.3)). However, for models

containing blocks of the M➙L and L➙L (no M) type, the additional equations
introduced help us to distinguish among groups of variables with differing
properties.

Composites
So far in our discussion of formal notation I have treated all conceptual variables
as true latent variables – unmeasured variables that exist independent from our
model and data, though we have no direct measures of them. We must now
consider the very practical matter that, for blocks such as M➙L and L➙L (no
M), there is usually insufficient information to estimate the true variance for
latent variables that do not have at least one effect indicator. In the absence of an
ability to estimate the true variance, latent variables without effect indicators can
only represent composites of their causes. For this reason, in the practical case
of estimating models that have latent variables without effect indicators, and
where there are no means to estimate their variances, we can replace M➙L with
M➙C and L➙L (no M) with L➙C. Here, C refers to a composite variable. In the
case of M➙C, the letter C describes a “composite of manifest variables”, while
in the case of L➙C, the letter C describes a “composite of latent variables”.
Composites are, in all cases that I describe in this chapter, considered to be
endogenous (η) variables. However, since they are, in effect, perfect represen-
tations of their cause, they are variables with zero value error variances. While
composites are imperfect representations of any construct they might repre-
sent, they can still be useful approximations of those constructs. Regarding the
example of forest herb colonization, if stand age and distance from source con-
stitute the predominant landscape features of importance for the model, then a
composite formed from their effects will have a very general meaning. If stand
age and distance from source are just two of many unique landscape properties
of importance to herb colonization, our composite will simply represent their
 Composite variables and their uses 155

joint effects, and will be a poor representation of a general landscape effect.

Typically, unless an effort has been made to obtain an exhaustive sampling
of the dimensions of a construct, any composite derived from the measures
of that construct should be understood to represent the collective effects of
its components, regardless of the label placed on that construct. This issue
notwithstanding, composites can be eminently useful as a means of represent-
ing certain types of construct, facilitating general conclusions, and managing
model complexity.

Evaluating the different model possibilities – theoretical considerations

Both theoretical and empirical information contribute to an evaluation of the
suitability of model architecture for a given situation. As stated earlier, there
are a great number of possibilities for how a set of manifest variables may be
connected. In the example under consideration here, involving herb colonization
of forest stands, initial guidance comes from the theory specified in Figure 6.3,
and the associations in Figure 6.4. I begin our evaluation of the applicability of
models A–D in this case by considering the causal relations between constructs
and indicators.
Borrowing, with modification, from Jarvis et al. (2003), certain conceptual
criteria can help in deciding whether a block should be modeled in the traditional
L➙M format (e.g., model A), or whether M➙C or L➙C blocks are more
appropriate (e.g., models B or C). Four kinds of questions can be considered to
gauge the theoretical basis for forming models such as model A versus models
B or C:

(1) What is the direction of causality?

(2) Are the indicators in a single block interchangeable?
(3) Is there an expectation that indicators in a block should covary (e.g., because
of joint causality)?
(4) Do the indicators in a block have a consistent set of causal influences?

For the sake of this discussion, it is recognized that model B is a special case
of model C where the x variables are considered to be without measurement
error. For this reason, in the immediate discussion I will contrast models A and
B, with the understanding that models B and C are equivalent with regard to
this evaluation.
For the first question, we must begin by asking whether the construct “Soil”
has causal influence on pH, moisture, and texture. Stated in a different way,
we are asking whether the variation among stands in soil conditions is such
that it causes common responses in pH, moisture, and texture; in which case
156 Structural equation modeling and natural systems

causation flows from the general construct to the indicators. Alternatively, if

it is the case that pH, moisture, and texture have some substantial degree of
independent effects on Soil, then causation flows from the indicators to the
construct (i.e., the construct comprises its parts). In this study, the authors
felt that causation flows from the indicators to the construct, and argued that
an M➙C block representation was appropriate. A similar determination was
made for the construct “Landscape” which is envisioned by the authors to be
the product of largely independent influences by certain landscape properties.
The second question to be considered is whether the indicators in a block are
interchangeable, if so, they then constitute redundant measures and are likely
to represent effect indicators consistent with an L➙M block. Stated in another
way, we might ask if dropping one of the indicators in a block alters the concep-
tual breadth of the construct. For both Soil and Landscape, the indicators in the
separate blocks would seem to be nonredundant based on the authors’ model
specification. For the construct Landscape, an M➙C block specification would
appear to be clearly indicated. The age of a patch is conceptually quite distinct
from its distance from a source population. If data for the distance to source
populations were to be absent, the construct would reduce solely to an age
dimension. For Soil, there is precedent for common soil factors of the L➙M
type (Grace 2003b, Weiher et al. 2004). However, in this case, the authors
felt that the individual indicators, pH, texture, and moisture, reflect some-
what separate dimensions of Soil and are best represented by an M➙C block
structure.
A third question is whether the indicators are expected to covary. This serves
as an additional way of evaluating the flow of causation in a block. If indicators
are under common causal control by a latent factor, then when that latent factor
varies, the indicators should all vary likewise. Such a situation would imply that
an L➙M block would be an appropriate means of modeling the situation. On
the other hand, if causation flows from the indicators to the latent variable, there
is no basis for having an expectation for correlations among indicators, since
their causes are not specified. We should, therefore, recognize that a correlation
among indicators does not inform us about the direction of causal flow, although
a lack of correlation among indicators would contraindicate the prospect that
a block should be of the L➙M form. For the herb colonization example, the
authors’ presentation did not produce any expectations of whether indicators
in multi-indicator blocks would correlate or not. So, again, by these criteria, it
would seem that Soil and Landscape are constructs best represented as M➙C
blocks given the available indicators.
The fourth question we might consider relates to the first, and asks whether
the indicators have common antecedents. If they do, then causation flows from
 Composite variables and their uses 157

that common antecedent, which can be modeled as a latent variable (of course,
if that latent factor is not the cause affecting other parts of the model, then a
completely different model structure would be implied). On the other hand,
in a model with an M➙C block structure, the indicators do not have common
antecedents, but rather, unique ones. In this example, it does not seem likely
that patch age and distance from source have a determined common antecedent.
The case is less clear for soil properties. Soil formation processes are such that
general soil conditions are capable of having common antecedents, such as
hydrologic and mineral influences. Thus, a potential basis for modeling Soil as
an L➙M block does exist based on this criterion.
Collectively, it would seem that the expectations of the authors, as reflected
by the four questions, would lean towards a specification of the multi-indicator
blocks as M➙C for both Soil and Landscape. This does not guarantee that this
is the correct causal structure, nor does it override the possibility that empirical
characteristics of the data might imply a different block structure. What is most
important to realize is that there are logical tests for developing, justifying, and
interpreting block structure, and that automatically presuming a conventional
hybrid model structure (L➙M) is not justified. Neither is it advisable to rely
simply on data properties to make the determination as to whether a construct
is best represented by a particular structure.

Evaluating the different model possibilities – empirical evaluations

of blocks
The truth is that often it is not known with absolute confidence what the causal
flow is for a set of manifest variables. This is especially true for applications
in the natural sciences, where the concept of latent variables is only beginning
to receive widespread attention. A strength of SEM is that various means of
evaluating model expectations exist (see numerous illustrations in Bollen 1989).
I will address overall model fit later. For now, our considerations relate to the
empirical characteristics of individual blocks. Ultimately these are of great
importance, for no matter what logic one brings to a model, if the indicators in
a block do not behave like effect indicators, then it will be unprofitable to model
them as such. Reviews of SEM studies in certain disciplines (e.g., Jarvis et al.
2003) suggest that researchers commonly misjudge the empirical support for
the causal direction specified in their models. The greatest error is to assume that
blocks are of the L➙M form, without seriously considering the use of M➙L
and L➙L (no M) blocks (or their applied counterparts, M➙C and L➙C). In
the example discussed here, the authors assumed the opposite form (M➙C) for
their multi-indicator blocks. However, it seems that the logical justification for
158 Structural equation modeling and natural systems

Table 6.2. Correlations among variables and the standard deviations

Col. freq. Distance Age Texture Moisture pH Cover

Col. freq. 1.0

Distance −0.5785 1.0
Age 0.6424 −0.5934 1.0
Texture −0.2553 0.1844 −0.3146 1.0
Moisture −0.3369 0.4604 −0.3462 0.5767 1.0
pH −0.0073 −0.0465 −0.0976 0.1324 0.0265 1.0
Cover −0.3423 0.2070 −0.4062 0.3189 0.2952 0.2394 1.0
Std. dev. 0.1867 0.8207 0.3390 0.4588 1.1359 0.7507 0.3017

this assumption is stronger for the Landscape construct and less definitive for
the Soil construct. In this section, I consider some of the empirical properties
to see if either block has properties consistent with the L➙M form.
Table 6.2 presents covariance relations for colonization frequency and other
manifest variables (represented by the correlations and standard deviations) for
the species Lamium galeobdolon. While SEM is based on an analysis of covari-
ance relations (as discussed in Chapter 3), an inspection of correlations can be
instructive. As Bollen and Ting (2000) have described, there is an expectation
that a set of effect indicators associated with a single latent variable in an L➙M
block will be intercorrelated, as implied by Eq. (6.1). For such a set of indica-
tors, their degree of intercorrelation will depend on the strength of the common
influence of the latent cause, relative to their total variances. So, for model A,
we would expect conspicuous correlations among soil pH, moisture, and tex-
ture, because of the joint influence of Soil on those specific factors. We would
have similar expectations for a strong correlation between distance and patch
age. Again, the degree of correlation expected would depend on the relative
importance of the errors, although for a reliable set of indicators, correlations
should at least be moderately high. In contrast, for a set of causal indicators
associated with a single latent variable in a single M➙C block, there is no basis
for expecting any particular correlation pattern. None of the equations that apply
to model B imply common causal influence on sets of causal indicators. So, a
set of causal indicators may or may not intercorrelate in such a case, since our
equations do not describe their interrelationships, except that they are classified
as being of common interest.
Inspection of Table 6.2 reveals a correlation between age and distance of
−0.5934. Thus, a correlation of moderate magnitude is observed for this pair
of indicators. In addition, the correlations between these variables and the other
variables in the model are, very approximately, of the same magnitude. Thus,
based on a crude inspection of correlations in the matrix, we are unable to rule
 Composite variables and their uses 159

out the possibility that either an L➙M or M➙C block structure would be con-
sistent with the empirical characteristics of the indicators related to Landscape.
The correlations between pH, moisture, and texture are 0.0265, 0.1324, and
0.5767. The low magnitude of correlations between pH and the other indicators
in the block suggests that these three soil properties would not be likely to
represent redundant measures of the sort normally found in L➙M blocks. A
method for formally evaluating this has been proposed by Bollen and Ting
(2000) based on vanishing tetrads. Correlations/covariances among a set of
truly redundant indicators in an L➙M block should possess the mathematical
property of vanishing tetrads, with a tetrad being the difference between the
products of pairs of covariances among four random variables. It can be said
in this case that the pattern of correlations among pH, moisture, and texture do
not appear to be consistent with such a block structure.

Comparing results for different models

Results for model A The data presented in Table 6.2 are sufficient to provide
for an evaluation of all models discussed in this paper. The analysis results
presented were obtained using the program Mplus (Muthén and Muthén 2004).
An illustration of a model containing a composite programmed in Mplus can
be found in Appendix I. For the purposes of this illustration involving model
A, we assume a modest, nonzero degree of measurement error for the single-
indicator measures (cover and colonization frequency) of 10%. Fitting of the
data to model A resulted in poor fit based on model chi-square statistics, error
variances for indicators, and examination of the residual covariance matrix. A
chi-square of 45.20 was obtained for the model, with 10 degrees of freedom. The
associated probability of good fit between data and model was found to be less
than 0.00005. Note that since model A is saturated with regard to the structural
model (i.e., all paths between latent variables are estimated), the inflated chi-
square does not result from unspecified relations among latent variables, but
instead, resides in the measurement model. In other words, in this example the
lack of model fit can be attributed to inappropriate relations between latent and
manifest variables. Standardized parameter estimates for model A are shown
in Figure 6.9. Here it can be seen that the standardized error variance for pH is
0.99, indicating that the model provides no explanation of the variance of this
variable. Finally, residuals indicate that, in general, the model does a poor job
of resolving the covariances among indicators.
A modification of model A was evaluated to determine the degree to which
the lack of fit for pH contributed to poor fit for the whole model. In the modified
model, only moisture and texture were used as indicators for Soil, while pH was
 160 Structural equation modeling and natural systems

0.76
0.99 pH, x1 ns
0.83 Soil 0.24 Competitors 0.95
0.37 moisture, x2 cover, y1 0.1
x1 he1
0.69 ns
0.57 texture, x3
ns 0.30
0.57

0.47 0.73 0.31

distance, x4 Landscape Colonization 0.95
−0.87
he2 col. freq., y2 0.1
0.34 stand age, x5 −0.81
x2

Figure 6.9. Standardized parameter estimates for model A. Chi-squared for model
fit was 45.20 with 10 degrees of freedom (sample size = 180) and a p-value
< 0.00005, indicating very poor fit of data to model expectations.

specified to be a single indicator for an additional exogenous latent variable,

True pH. Again, 10% of the variance of observed pH was specified as measure-
ment error. A chi-square of 34.57 with 7 degrees of freedom was obtained for
this model. The associated probability of good fit remained less than 0.00005.
Thus, this result indicates that the observed lack of fit between model and data
is spread throughout the measurement model, and is not solely due to a lack of
correlation between pH and the other soil variables.
As stated earlier, it is not likely that the observed variables measured in this
study were obtained without some degree of measurement error. For this reason,
model B would seem to be a less than ideal representation for our purposes here,
as it assumes the causal indicators associated with soil conditions and landscape
properties are error free. For this reason, we can now ignore model B and focus
our attention on model C, which recognizes that soil and landscape variables
are imperfect measures of the true values. As above, for single indicator latent
variables in this study, an arbitrary measurement error of 10% of the variance
for single indicator blocks is assumed. Before discussing model C, however,
I wish to consider a general issue that relates to models of its type (those that
contain composites). This issue deals with parameter identification, which can
prevent the researcher from obtaining unique and valid estimates.

A digression on the issue of parameter identification in models

containing composites
Model C, because it includes composite variables, has a somewhat different
set of parameters that must be estimated, compared to a similar model without
composites (e.g., model D). As mentioned earlier, since I define composites
as having zero error variances (not all authors do), I can set these parameters
to zero, thus, their identification is not an issue. However, for model C, the
 Composite variables and their uses 161

True pH
pH, x1
x1 0
0.67

0.73
True Moist. ns Soil
0.34
moisture, x2 hc 1
x2
Competitors
ns ns cover, y1
he 1
True Texture
texture, x3
x3
ns 0.43

−0.31
Colonization
True Dist.
−0.37 he 2 col. freq., y2
distance, x4
x4
Landscape 0.72

True Age
0.72 hc 2
stand age, x5
x5 0

Figure 6.10. Select standardized parameter estimates for model C. Here Soil
and Landscape are composites with zero disturbance terms. Model chi-square was
6.005 with 3 degrees of freedom (p = 0.1107).

composite variables introduce four new paths that represent the effects of the
composites (the ηc s) on the ηe variables. In this particular case, the relationships
between the exogenous latent variables (ξ s) and the ηc variables are reduced to
single effects from each ξ . So, for our example, the ten potential paths from the
individual ξ variables to the ηe variables (e.g., in model D) are replaced with
five paths from the ξ to the ηc variables, plus four paths from the ηc variables
to the ηe variables. In spite of the net gain of a degree of freedom from this
substitution, problems remain with parameter identification. Ultimately, there is
a general problem that arises when attempting to identify all paths leading to, as
well as flowing out from, a composite. This problem is similar to the routinely
encountered problem associated with latent variables with effect indicators,
where the scale of the latent needs to be established. Both of these problems
can be solved in the same fashion by specifying a single incoming relationship,
which establishes the scale of measurement.
In spite of the fact that parameter identification issues can be resolved by
specification of select parameters, an issue still remains as to the significance of
the specified parameters. In model C (Figure 6.10), we set the unstandardized
coefficients for the paths from True pH to Soil and from True Distance to
Landscape to 1.0, to establish the scale of measurement for the composites (note,
this is not shown in Figure 6.10 because only the standardized parameters are
presented). This procedure ignores the question of whether there are significant
relationships associated with these paths. More specifically, does True pH have
a significant (nonzero) contribution to Soil, and similarly, does True Distance
162 Structural equation modeling and natural systems

have a nonzero contribution to Landscape? One approach to evaluating fixed

paths from ξ s to ηc s is to use a reduced form model such as model D, in which
the composites are omitted and the direct effects of ξ s on ηc s are tested. For
our example, the reduced form equations are
ηe1 = γe1 ξ1 ξ1 + γe1 ξ2 ξ2 + γe1 ξ3 ξ3 + ζe1 (6.7)
and
ηe2 = γe2 ξ1 ξ1 + γe2 ξ2 ξ2 + γe2 ξ3 ξ3 + βe2 e1 ηe1 + ζe2 (6.8)
For models with more than a single path flowing out from the composites,
such as is the case in model C, an evaluation of model D only provides an
approximate answer to the question of whether the parameters in model C are
significant. The reason for this is that a single estimate for each gamma must
apply to both Eqs. (6.7) and (6.8) in order to represent the situation in model
C. Stated more succinctly, for model C it is assumed that
ηei = γeξ1 ξ1 + γeξ2 ξ2 + γeξ3 ξ3 + ζei (6.9)
holds for all i (i.e., that the condition γe1 ξ1 = γe2 ξ1 and so forth holds for all ξ ).
As we show below, it is possible and even desirable to evaluate Eq. (6.9) as part
of the assessment process.

Results for model C Maximum likelihood estimation yields a chi-square

of 6.005 with 3 degrees of freedom and an associated probability of 0.1107.
Examination of residuals and the chi-square results indicate adequate fit of the
data to model C. Again, to better allow for comparisons among models, for
most comparisons standardized parameter estimates are presented, and it is
assumed that significance tests for unstandardized parameter values apply to
the standardized values as well. In the presentation of results in Figure 6.10,
all numerical estimates shown are associated with significant p-values for the
relevant unstandardized parameters, except for the paths from True pH to Soil
and True Dist. to Landscape, which were set to fixed values to establish the
scales for the composites.
Conspicuous in the results for model C (Figure 6.10) is that neither True
Moisture nor True Texture contributed significantly to the composite variable
Soil. Also, Soil had a significant effect on Competitors, but not on Colonization.
In contrast, the freely estimated relationship between True Age and Landscape
was significant; as well as both paths from Landscape to Competitors and
Colonization. Examination of the residual variances for Competitors and Col-
onization shows that 27% and 57% of the variance in these response variables
were explained.
 Composite variables and their uses 163

Soil Conditions

True pH
pH, x1 0.19
x1
ns
True Moist.. 0.70
ns
moisture, x2
x2
ns Competitors
r, y1
cover,
True Texture
ns he
1
texture, x3
x3
ns ns 0.43

ns
Colonization
ni
True Dist.
he col. freq., y2
distance,, x4
x4 -0.32 2

-0.43
True Age
stand age, x5 0.45
x5

Landscape Properties
Figure 6.11. Standardized parameter estimates for model D. Model D lacks com-
posites and represents a partially reduced form of model C.

Results for model D It is instructive to evaluate model D, which effectively

represents a partially reduced form of model C in which composites are omitted.
In this model, separate effects of the ξ variables on Competitors and Coloniza-
tion are estimated, and the individual ξ variables are associated with constructs
nominally. Since this model is saturated, maximum likelihood estimation yields
a chi-square of 0 due to the fact that there are no degrees of freedom. In a side
analysis, the stability of results was tested by deleting nonsignificant paths,
yielding a chi-square of 9.773 with 6 degrees of freedom (p = 0.1344). This
side analysis indicated that the results for model D are stable.
Results for model D (Figure 6.11) show that only four of the paths from ξ
variables to ηe variables were significant. Variance explained for Competitors
and Colonization in the saturated model was 30% and 57%. Again, for the pur-
poses of comparing among models, the saturated model results are used even
though nonsignificant paths are included. Aside from providing a comparison
to model C, model D also provides a representation of the constructs from
Figure 6.3; except that in this case, the dimensions of soil conditions and land-
scape properties are represented separately. It is possible in such a model to
provide some degree of representation of the effects of the soil and landscape
constructs, even without the use of composites. For example, selective elimina-
tion of categories of variables can be used to derive estimates of their unique and
164 Structural equation modeling and natural systems

Table 6.3. Estimation of unique and shared variance explanation in Com-

petitors and Colonization for soil conditions and landscape properties in
model D

Variance explanation contribution for

response variables (percent of total variance)
Construct(s) Competitors Colonization
Soil only 19.5 22.1
Landscape only 21.1 57.0
Soil and Landscape combined 30.1 57.4
Soil unique contribution 9.0 0.4
(combined – landscape only)
Landscape unique contribution 10.6 35.3
(combined – soil only)
Soil shared contribution 10.5 21.7
(soil only – soil unique)
Landscape shared contribution 10.5 21.7
(landscape only – landscape unique)

shared variance explanation (described in Chapter 3). In this case, if we drop the
variables associated with landscape properties (True Distance and True Age,
as well as their indicators) from the model, the variance explained in Competi-
tors and Colonization is reduced to 19.5% and 22.1% respectively. Also, if we
keep the variables associated with landscape properties in the model, but drop
those associated with soil conditions, the variance explained in Competitors
and Colonization is found to be 21.1% and 57.0% respectively. A summary
of these results can be seen in Table 6.3, along with derivations of unique and
shared variance explanation. Here we can see that soil conditions and land-
scape properties had roughly equal unique variance explanation contributions
for Competitors (9.0% and 10.6%) with 10% shared between them. For Col-
onization, unique variance explanation was very dissimilar (0.4% and 35.3%)
for soil conditions and landscape properties, while 21.7% was shared between
the two groups of predictors. In this case, the path from Competitors to Colo-
nization was nonsignificant, therefore, all effects from predictors are direct and
no indirect effects on Colonization are described.

A further model to consider – model E

Above it was mentioned that model C presumes that Eq. (6.9) holds. This
equation presumes that a single set of gammas associated with the L➙C blocks
can successfully represent the effects of ξ variables on ηe variables. This is akin
to saying that the slopes of the regression relations of ξ s on ηe1 would be the
same as for the effects of ξ s on ηe2 . If this condition holds, then model C is
 Composite variables and their uses 165

0
True pH
pH, x1 0.59
x1 Soil→Com
ns hc 1 0.32
ns 0.70
True Moist.
0
moisture, x2
x2 ns
Competitors
ns Soil→Col
he 1 cover, y1
True Texture hc 2 ns
texture, x3 ns
x3
ns 0.43

−0.35
ns Land→Com Colonization
True Dist.
hc 3 he 2 col. freq., y2
distance, x4
x4
−0.46
1.23 0
True Age Land->Col 0.71
stand age, x5
x5 0.64 hc 4

Figure 6.12. Select standardized parameter estimates for model E. Here four
composites are included, one for each response variable.

our best model for the situation. On the other hand, if Eq. (6.9) does not hold
because regression relations are not similar, our model will underestimate the
variance explanation for Competitors and Colonization. It can easily be seen
that when gammas are very different in Eqs. (6.7) and (6.8), then the gammas
in Eq. (6.9) will be insufficiently relevant to accurately predict the ηe s. To
evaluate this possibility, we consider one additional model, which we refer to
as model E. In this model, separate composites are derived to estimate effects
on Competitors and Colonization.
Results for model E are shown in Figure 6.12. As with model D, when
all paths are included there are no degrees of freedom for testing model fit.
However, deletion of nonsignificant paths allows for model testing and, again,
no unspecified paths are indicated. As before, results presented in the figure are
the standardized parameter estimates for the model, with all paths included.

Comparisons among models

The exercise of comparing different models as contrasting representations of
the construct relations in Figure 6.3 is meant not only to illustrate some of the
variety of modeling possibilities, but also to introduce a number of important
166 Structural equation modeling and natural systems

concepts and procedures. Most of the presentations of SEM in introductory

textbooks are likely to lead new users to consider model A to be the appropriate
representation of the construct relations in Figure 6.3. As we saw, model fit was
found to be completely unacceptable for model A and the resulting parameters,
therefore, unreliable. Several review papers have suggested that researchers fre-
quently fail to consider the full range of possibilities for model development.
Models that contain causal indicators and composites are rather rare compared
to those in which constructs are exclusively represented by effect indicators
in L➙M blocks. In this example, the bulk of the theoretical information and
the great majority of the empirical analyses indicate that model A is not an
appropriate representation of the situation. Conceptually, the constructs were
not envisioned as unidimensional underlying causes, but rather, collections of
influencing factors. Furthermore, the available indicators do not represent a
suite of effect indicators as would be appropriate for an L➙M block. The inad-
equacy of fit of data to model A is clearly indicated by the very low probability
associated with the chi-square test. Thus, we can reject this model as inade-
quate. Consideration of modifications such as splitting out pH from the Soil
block, as well as other attempts at small modifications, also did not lead to the
discovery of an adequate model akin to model A.
Model B was omitted from our analyses because of the strong likelihood
of measurement error existing in the indicators. However, if we had presented
an analysis of model B, it would be found to possess many of the same prop-
erties as were found for model C; although without the modest correction for
measurement error that was built into model C.
The remaining comparisons of most interest are between model C and mod-
els D and E. All three of these models had adequate fit to the data, and the
differences among them can be viewed as tradeoffs between model general-
ity and model accuracy (Levins 1968). Model C is the most general of the
three, representing the collective effects of soil conditions and landscape prop-
erties, with a single composite for each. As described in the above discussion
of Eqs. (6.7)–(6.9), for model C to be appropriate, the estimated composites
would need to be near optimal for representing effects on both Competitors
and Colonization simultaneously. At the other end of the spectrum, model
D is the most accurate and least general of the three models. No attempt is
made to summarize the collective effects of the two multidimensional con-
structs through the use of composites in model D. While a general assess-
ment of unique and shared variance can be used in this case (as demon-
strated in Table 6.3), such calculations are also applicable to models C and
E. Calculations of unique and shared variance explanation can be applied to
any model, including those with composites; and to any collection of variables.
Therefore, this way of dealing with the partitioning of variance explanation
 Composite variables and their uses 167

is not a replacement for the use of composites, but instead, a complementary

procedure.
Compared to model D, which contains no composites, model E represents
an advance in generality, having an individual composite for each effect of the
multidimensional constructs. As would be expected for such a case, the error
variances for Competitors and Colonization are virtually identical for models D
and E. Thus, both models D and E are equally accurate with regard to variance
explanation, but they differ in the way in which effects are represented. Overall,
model D provides a more succinct representation of the individual effects of ξ
variables on Competitors and Colonization. All paths are tested for significance
and coefficients are readily interpretable as individual effects. In comparison,
model E provides a superior representation of collective effects, with single path
coefficients representing the effects of suites of variables, although at some cost
in terms of the interpretation of individual influences (the paths from ξ variables
to ηc variables).
An important comparison to make between models C and E has to do with
the magnitude of the error variances for the ηe variables. It has been recog-
nized previously that when composites with multiple effects on other variables
are not appropriate, there is a conspicuous loss of variance explanation for the
affected variables. In this case, error variances are 0.73 and 0.43 for Competi-
tors and Colonization respectively in model C. In contrast, in model E, the
values are 0.70 and 0.43. A test of the adequacy of model C is whether its error
variances are significantly greater than those for the less general case where
composites have single effects. For such a test, the t-value for a test of differ-
ence between the error variances for Competitors must be conducted and based
on the unstandardized parameters. For model C, the unstandardized residual
variance for Competitors is 0.059 ± 0.007 (std. error), while for model E the
values are 0.057 ± 0.007, yielding a t-value of 0.14, and an associated probabil-
ity of no difference approaching 0.9. For Colonization, residual error variances
were virtually identical (0.013 ± 0.002) for models C and E. Thus, there is no
indication that model E is superior to model C in variance explanation for the ηe
variables. Furthermore, the paths from composites to the ηe variables are also
very similar between models. Collectively, this evidence supports the conclu-
sion that model C is superior to model E overall, because of its greater generality,
and it is our best model (of those examined) for this example and these data.

Example 2: an examination of habitat associations for anurans

Background
Anurans are a group of amphibians, comprising frogs and toads, that are of
high conservation concern due to worldwide population declines. This example,
168 Structural equation modeling and natural systems

Macrohabitat Anuran
Type Diversity

Microhabitat
Conditions

Figure 6.13. Construct model relating macrohabitat type and microhabitat con-
ditions to anuran diversity.

developed from data taken from a study by Lichtenberg et al. (unpublished) in

which 25 wetlands in the lower Mississippi River alluvial valley were examined
for chorusing anuran species and associated habitat characteristics. One major
goal of this study was to understand the habitat characteristics associated with
diversity “hotspots” where a variety of species are abundant. Figure 6.13 pro-
vides an illustration of the construct relations of interest in this example. Stated
in general terms, the goal of their analysis of anuran diversity was to under-
stand the degree to which it depends on the type of habitat (lake, impoundment,
etc.) versus the particular conditions (vegetation, litter, etc.) within that habi-
tat. Specifically, they were interested in the question of whether correlations
between high diversity and particular wetland types can be explained by the
microhabitat conditions within the wetlands. Such information could prove use-
ful in making decisions about types of habitat to protect and the conditions to
maintain.
In this study, several types of wetlands were examined. These could be clas-
sified as being of one of four types – lakes, impoundments, swales, and riverine
areas. Specific microhabitat conditions were assessed at each site, and these
included a number of vegetation and topographic features. Anuran diversity
was assessed using nighttime surveys of chorusing individuals during several
seasons of the year. Chorus surveys during the year were used to arrive at an
estimate of the species using a site.
While macrohabitat types were clearly defined a priori, microhabitat con-
ditions were sampled in a more exploratory fashion. A wide range of charac-
teristics of the vegetation were studied, including herbaceous cover, vegetation
density at different vertical positions, woody cover, tree height, canopy cover,
as well as litter cover and depth (by type). Also measured were hydrologic
 Composite variables and their uses 169

features at the site, including the area of open water and the mean and maxi-
mum water depths.
In conjunction with our presentation of Example 1, we provided a detailed
consideration of how constructs may be represented and both theoretical and
empirical criteria for arriving at decisions about block structure. In this second
example, our emphasis is more on the question of how to model a situation where
an endogenous variable (microhabitat conditions) has multiple indicators and
may involve composites. This question was not addressed in Example 1, where
multiple indicators existed only for the exogenous constructs.
In the current example, we begin our analysis with the construct labeled
Macrohabitat Type. Since our measure is nominal and multi-level (whether a site
is classified as lake, impoundment, swale, or riverine), it immediately suggests
the need to model this construct using a set of dummy variables representing the
possible macrohabitat types. We can assume for the sake of simplicity that
the classification of individual sites as to habitat type was correct. Therefore,
the presumption is that the construct Macrohabitat Type can be modeled using
either an M➙L or M➙C block type. The deciding factors for choosing between
these two block types are whether we believe we have a complete sampling of
all possible macrohabitat types, and whether we can derive an estimate of the
error of prediction for Macrohabitat Type. Since neither of these criteria hold
in this case, the M➙C block structure seems more appropriate.
The construct labeled Microhabitat Conditions is one where the specific
details of how the measured variables would interrelate was not known a priori.
For this reason, Lichtenberg et al. performed an exploratory factor analysis to
see if the correlations among the many measured variables might suggest the
operation of a smaller number of latent factors. I will not go into the details of
that analysis here, but only say that the result was the recognition by the authors
of two factors of importance to anuran diversity, the abundance of herbaceous
vegetation and the abundance of leaf litter. Based on the conceptualization of
Microhabitat Conditions by Lichtenberg et al., it is clear that the indicators
could represent a collection of factors that affect anurans, based on theoretical
grounds. Thus, we begin with the expectation of a block structure of L➙C,
with two latent variables, Herbaceous Vegetation and Litter, contributing to the
construct. A total of seven indicators of the two latent variables were included
in their final model (see below).
Lichtenberg et al. discuss certain issues of measurement regarding the con-
struct Anuran Diversity. It is widely held that there are several causes of
measurement error for wildlife populations and communities. In addition to
the usual matter of sampling, varying detectability can contribute to error.
Lichtenberg et al. addressed the issue of detectability to some degree by using
170 Structural equation modeling and natural systems

0
lake, x1

Macrohabitat Diversity
impound, x2 rich, y8
hc 1 he 3
swale, x3
0

Microhabitat
hc 2

vhit1, y1
wlitr, y5
vhit2, y2
Herbaceous Litter
litrd, y6
he 1 he 2
herbl, y3
litrc, y7
herbc, y4

Figure 6.14. Model F, which shows one of the possible ways that indicators
could be related to the construct model shown in Figure 6.13. Refer to Table
6.3 for definitions of the observed variables. Note that by omission, the riverine
macrohabitat condition represents the baseline against which other macrohabitats
are compared.

the total number of species recorded across samplings, instead of the mean.
Nevertheless, error in assessing the true number of species at each site is likely
to be significant and, while no estimate of this error exists, we again use an
arbitrary estimate of 10% of the total variance.

Possible model structures

In the first example we spent some time discussing and illustrating criteria for
evaluating the cause and effect relations among variables and how these influ-
ence model structure. In this second example, we forego a detailed discussion
of such considerations and, instead, focus on the possible ways to model a case
where composites have directional effects on other composites.
A model that logically follows from the construct relations in Figure 6.13
is shown in Figure 6.14. I refer to this model as model F to avoid confu-
sion with the models discussed in conjunction with the first example. Based
on the information presented above, we can represent Macrohabitat using the
block structure M➙C. The Microhabitat construct is represented in this case as
an L➙C block containing two latent variables, both with multiple indicators.
Diversity is represented by an L➙M block with a single indicator. Consistent
 Composite variables and their uses 171

0
lake, x1

Macrohabitat Diversity
impound, x2 rich, y8
hc 1 he 3
swale, x3
0

Microhabitat
hc 2

vhit1, y1
wlitr, y5
vhit2, y2
Herbaceous Litter
litrd, y6
he 1 he 2
herbl, y3
litrc, y7
herbc, y4

Figure 6.15. Model G, which differs from model F by allowing separate coeffi-
cients (represented by separate paths) to convey the effects of Macrohabitat on the
Herbaceous and Litter dimensions of Microhabitat.

with their declarations as composites, the error variances for Macrohabitat and
Microhabitat are set to zero.
Of the models we consider here, model F is the most abstract. At the same
time, model F is based on the greatest number of assumptions. It is presumed
in this model that the influences of Macrohabitat (ηc1 ) on Microhabitat (ηc2 )
can be summarized by a single coefficient, β c2 c1 , despite the fact that ηc2 is
actually a predictor of Herbaceous and Litter effects on Diversity, rather than
the microhabitat conditions themselves. Stated in other terms, in model F the
covariances between x1 –x3 and y1 –y7 must be resolved by their joint relations
to a linear predictor that depends on ηe3 . These are fairly critical assumptions in
that their failure is likely to mean an underestimate of Macrohabitat effects on
Microhabitat, and possibly unresolved covariances among manifest variables.
An alternative formulation that is perhaps more biologically meaningful,
model G (Figure 6.15), represents the modeling implications of relaxing the
assumptions just discussed for model F. Here we are allowed to consider effects
of Macrohabitat on the two dimensions of Microhabitat, Herbaceous and Lit-
ter. Since these two dimensions are independently estimated in L➙M blocks,
the interpretative meaning of the paths from Macrohabitat to each dimension
is clear, and also independent of other relationships in the model. In contrast,
since the composite Microhabitat is defined as a linear predictor of Diversity,
it depends on the covariances between three latent variables, ηe1 , ηe2 , and ηe3 .
172 Structural equation modeling and natural systems

If any of the covariances among these variables changes, the meaning of ηc2
changes, and thus, the meaning of a direct path from Macrohabitat to Micro-
habitat (as in model F). In model G, the effect of Macrohabitat on Microhabitat
can be summarized by calculating the total effect of ηc1 on ηc2 , which can be
summarized by the equation

βc2 c1 = βe1 c1 × βc2 e1 + βe2 c1 × βc2 e2 (6.10)

where β c2 c1 now refers to a calculated total effect.

Certain restrictive assumptions that warrant discussion remain in model G.
In particular, the assumption is implied that a single combination of coeffi-
cients from macrohabitat types (x1 –x3 ) to the Macrohabitat composite (ηc1 )
simultaneously summarizes the individual effects of x1 –x3 on ηe1 , ηe2 , and
ηe3 . We feel that it is always wise to check such an assumption (as was
done in Example 1). Here, this assumption is explicitly removed in model H
(Figure 6.16), where one composite representing Macrohabitat effects is
replaced by three (Macro➙Herb, Macro➙Lit, Macro➙Div).
Also as was done in Example 1, we offer for comparison a partially reduced
form model, model I (Figure 6.17), in which composites are omitted and con-
structs are represented nominally. While model I offers a less complete mod-
eling of the construct relations (e.g., there are no path coefficients derived
for interactions among the major constructs), it permits a complete evalua-
tion of the significance of all individual effects, thus complementing the other
models.

Comparisons among models

Sample correlations and standard deviations are presented in Table 6.4 for the
manifest variables considered in this example. As stated earlier, we forego
an empirical evaluation of the relations between manifest variables and either
composites or latent variables, as these were addressed by Lichtenberg et al.
We also bypass any discussion of the bivariate relations in Table 6.4 and move
directly into a consideration of model results.
As was brought out in our discussion of Example 1, the solution and evalu-
ation of general models containing composite variables often depend on infor-
mation from more specific models. For this reason, we will begin with the
model containing the least number of assumptions (and thereby, our least gen-
eral model), model I. We will then proceed progressively through models H, G,
and F to determine whether more general models result in a substantial loss of
information.
 Composite variables and their uses 173

Macro→Div
hc1
0
lake, x1 Diversity
rich, y8
Macro→Lit he3
impound, x2
hc2
0
swale, x3 0

Macro→Herb
Microhabitat
hc3
hc4

vhit1, y1
wlitr, y5
vhit2, y2
Herbaceous Litter
litrd, y6
he1 he2
herbl, y3
litrc, y7
herbc, y4

Figure 6.16. Model H, which differs from model G by using separate composites
(ηc1 –ηc3 ) to convey the effects of individual Macrohabitat types on ηe1 , ηe2 , and
ηe3 .

Macrohabitat

lake, x1 Diversity
rich, y8
he3
impound, x2

swale, x3

Microhabitat

vhit1, y1
wlitr, y5
vhit2, y2
Herbaceous Litter
litrd, y6
he1 he2
herbl, y3
litrc, y7
herbc, y4

Figure 6.17. Model I. Partially reduced form model representing effects of Macro-
habitat and Microhabitat nominally through grouping variables.
Table 6.4. Correlations among variables related to anuran richness and their standard deviations. “rich” refers to the number of
anurans at a site, “imp” refers to impoundments, “vhit2” and “vhit1” are measures of vegetation density, “herbl” and “herbc”
are measures of dead and live herbaceous vegetation, “wlitr” is woody litter, “litrd” is litter depth, and “litrc” is litter cover

rich lake imp swale vhit2 vhit1 herbl herbc wlitr litrd litrc
rich 1.0
lake 0.696 1.0
imp −0.167 −0.355 1.0
swale −0.431 −0.659 −0.253 1.0
vhit2 0.372 0.167 0.111 −0.099 1.0
vhit1 0.222 −0.156 0.552 −0.118 0.653 1.0
herbl 0.060 −0.252 0.562 −0.009 0.581 0.825 1.0
herbc 0.091 −0.087 0.419 −0.132 0.437 0.745 0.756 1.0
wlitr 0.509 0.430 −0.284 −0.099 −0.051 −0.290 −0.395 −0.396 1.0
litrd 0.238 0.146 −0.433 0.383 0.027 −0.097 −0.180 −0.281 0.419 1.0
litrc 0.219 0.194 −0.442 0.273 −0.118 −0.414 −0.509 −0.580 0.568 0.762 1.0
sd 2.170 0.510 0.332 0.476 0.512 1.482 0.173 0.122 0.100 0.122 0.148
 Composite variables and their uses 175

Macrohabitat 0.28

lake, x1 0.76
Diversity 0.95
rich, y8
he3
impound, x2
0.73

swale, x3

0.78 0.50 0.37

0.57
Microhabitat

vhit1, y1
0.92 wlitr, y5
0.63 0.60
vhit2, y2
Herbaceous Litter 0.79
0.91 litrd, y6
he1 he2
herbl, y3
0.98
0.82 litrc, y7
−0.30
herbc, y4 0.67 0.61

Figure 6.18. Results obtained for model I, showing standardized values for path
coefficients and the error variances of latent endogenous variables. Correlations
among xs and errors for ys are not shown, for simplicity. Nonsignificant effects
of macrohabitat types were dropped from the final model, which possessed a chi-
square of 50.42 with 39 degrees of freedom and a p of 0.104.

Results for model I The results for model I are given in Figure 6.18. For the
purposes of this evaluation (and in contrast to our practice in Example 1),
nonsignificant effects of macrohabitat types on endogenous variables were
dropped from the final model. In spite of the small sample size, results were
stable and the fit between model expectations and data was acceptable. As these
results show, impoundments had significantly higher levels of herbaceous veg-
etation than did other macrohabitat types. Litter accumulation, in contrast, was
substantially higher in lakes and swales than in impoundments and riverine
habitats (recall, the riverine variable was omitted and, therefore, serves as the
baseline condition). Anuran diversity was found to be higher in lakes than in
all other habitat types.

Results for model H Model H provides for a single path from each compo-
site to replace the multiple paths that would otherwise connect one construct
with another. Aside from that, the models are very similar. Results from the
estimation of model H are presented in Figure 6.19. A comparison of results
from models I and H shows numerous similarities, and a few differences. Model
fit parameters are identical for the two models. Also, variance explanation for
176 Structural equation modeling and natural systems

MacroÆDiv
0.28
1.00 hc1 0.76
0
lake, x1 Diversity
1.17 0.95
he3 rich, y8
MacroÆLit
impound, x2
hc2
1.25 0
swale, x3 0.42
0
1.00 0.62
MacroÆHerb
Microhabitat
hc3
hc4
0.57
vhit1, y1
0.92 1.18 0.87 wlitr, y5
0.63 0.60
vhit2, y2
Herbaceous Litter 0.79
0.91 litrd, y6
he1 he2
herbl, y3
0.98
0.82 -0.30 litrc, y7
herbc, y4 0.67 0.61

Figure 6.19. Results obtained for model H, showing standardized values for path
coefficients and the error variances of latent endogenous variables. Correlations
among xs and errors for ys are not shown, for simplicity. Nonsignificant effects
of macrohabitat types were dropped from the final model, which possessed a chi-
square of 50.42 with 39 degrees of freedom and a p of 0.104; precisely as for
model I.

endogenous latent variables is the same, with R2 s of 0.33, 0.39, and 0.78 for
ηe1 –ηe3 respectively. Loadings in L➙M blocks are the same for both models, as
are outward paths from composites possessing single causes (ηc1 to ηe3 and ηc3
to ηe1 ), in comparison to the equivalent effects in model I (x1 to ηe3 and x2 to ηe1 ).
Composites with multiple causes in model H yielded parameters not found in
model I, such as those associated with ηc4 to ηe3 (Microhabitat to Diversity)
and ηc2 to ηe2 (Macro➙Lit to Litter). The path coefficients associated with
these paths represent standardized collective effects of the composites’ causes
on the response variables involved. Heise (1972) referred to these coefficients
as “sheath” coefficients to designate the fact that they represent a collection of
causes.
The most conspicuous differences between models I and H reside with
the paths from causes to composites. For example, in model I, the effects of
Herbaceous and Litter on Diversity are represented by two path coefficients
(0.50 and 0.37), while in model H, the same effects are represented by three
paths, two from Herbaceous and Litter to Microhabitat (1.18 and 0.87), and
one from Microhabitat to Diversity (0.42). Upon first examination, the paths
 Composite variables and their uses 177

from Herbaceous and Litter to Microhabitat appear unusually inflated (con-

trary to popular notions, standardized coefficients can be greater than 1.0 when
offsetting paths are present), particularly since Herbaceous and Litter are only
modestly correlated (−0.30). However, the equivalency of relationships in mod-
els I and H can be made clear by realizing that the total effects of Herbaceous
and Litter on Diversity are simply compound paths in model H (e.g., the effect
of ηe1 on ηe3 = 1.18 times 0.42 = 0.50). The same holds true for all effects
involving composites (e.g., the effect of x1 and x3 on ηe2 ).

Results for model G Model G, as described above, is a more general state-

ment than model H, and presumes that a single composite is sufficient to rep-
resent the effects of Macrohabitat types on ηe1 , ηe2 , and ηe3 . Results from
maximum likelihood estimation of model G included a chi-squared of model
fit of 66.12 with 38 degrees of freedom, and an associated p-value of 0.0031,
indicating poor fit. It was found that when estimated using a single composite
for Macrohabitat effects, substantial residual correlations between individual
macrohabitat types and the indicators for ηe1 , ηe2 , and ηe3 , were not resolved.
These results suggested that a single composite representing the effects of
macrohabitat types on Herbaceous, Litter, and Diversity was inconsistent with
the data. However, neither impoundment or swale indicators contributed sig-
nificantly to the composite in this case, therefore, it was possible to simplify
the model by dropping these indicators from the model. With these indicators
deleted, model fit improved to 30.03 with 24 degrees of freedom and a p-value
of 0.1837, indicating acceptable fit. The results presented in Figure 6.20 are
based on these results.
The version of model G presented in Figure 6.20, while representing a
model that fits the data, fails to explain significant variation in either ηe1 or ηe2 .
Lichtenberg et al. (unpublished) chose to select this model to represent the direct
and indirect effects of Macrohabitat, because it addressed the central question
of interest in their study. While it succeeds at that more narrow objective, model
G fails to provide an explanation for macrohabitat effects on both diversity and
microhabitat. The clear reason for that failure is that macrohabitat effects on
microhabitat conditions are quite different from those on diversity, and a single
composite cannot adequately represent both.

A consideration of model F The most abstract model, model F, repre-

sents direct effects of Macrohabitat on Microhabitat rather than on the dimen-
sions of Microhabitat (Herbaceous and Litter). Unfortunately, a proper esti-
mation of this model using a simultaneous solution procedure cannot be
achieved. The fundamental difficulty is that a zero error variance is specified for
178 Structural equation modeling and natural systems

0 0.26
lake, x1 1.00
Macrohabitat 0.75 Diversity 0.95
impound, x2 rich, y8
hc he 3
1
swale, x3
0.44
0

ns
Microhabitat
ns hc 2

vhit1, y1
0.92 1.13 0.80 wlitr, y5
0.64 0.59
vhit2, y2
Herbaceous Litter 0.77
0.90 he litrd, y6
1
he 2
herbl, y3
0.98
0.81 −0.47 litrc, y7
herbc, y4 1.0 1.0

Figure 6.20. Results obtained for model G, showing standardized values for path
coefficients and the error variances of latent endogenous variables. Correlations
among xs and errors for ys are not shown, for simplicity. Chi-square for this model
was 30.03 with 24 degrees of freedom and a p of 0.1837.

Microhabitat in order to permit it to represent the effects of Herbaceous and

Litter on Diversity. However, with a zero error variance specified, it is not pos-
sible to estimate a nonzero error variance for Microhabitat based on the effects
of Macrohabitat. Thus, with Microhabitat represented as a composite with zero
error variance, model F is irresolvable. This serves as an additional argument
in favor of modeling approaches in which exogenous influences are measured
against the individual dimensions of endogenous multidimensional constructs
(e.g., models H and G), rather than the composite of their combined effects (in
this case, on Diversity). Using such an approach, effects on the endogenous
composite can be estimated using compound paths. An alternative approach
that could be used is a two-stage solution procedure in which factor scores for
ηc2 are first derived and then used in a simpler model involving only ηc1 , ηc2 ,
and ηe3 . Such a model would, however, fail to provide a broad evaluation of the
relationships embodied in model F, and is not addressed further here.

Conclusions and recommendations

It should be clear from the examples considered in this review that the standard
hybrid model commonly used to describe SEM is an inadequate representation
 Composite variables and their uses 179

of the possibilities. A consideration of composite variables contributes balance

to the modeling process by providing options and encouraging a thorough
evaluation of constructs and the measures at hand. While this problem has been
pointed out by some at various times, the inclusion of composites in structural
equation models remains rare, while the misspecification of models using only
L➙M blocks is common.
At present, much of the literature on composite variables is based on the
premise that composites should have two or more effects on other variables
(outward-directed paths), and that composites with single effects should not
be included in models (MacCallum and Browne 1993, Edwards 2001, Jarvis
et al. 2003, Williams et al. 2003). This recommendation has been offered as a
necessary, though not sufficient, requirement for the identification of composite
error variances. In Example 1, such a model (model C) was indeed found to be
appropriate and satisfactory. However, in many cases models of this sort will
not be adequate, and those with single effects (e.g., model E) will be required.
It can be argued that (1) error variance estimates for latent variables possessing
only causal indicators will seldom be meaningful and, thus, they should not
drive decisions about the appropriate form of composites to include in a model,
(2) including composites having only single outward directed paths can be quite
valuable even though the covariances for a set of data can be represented without
them, and (3) confining the use of composites to cases where there are multiple
outward directed pathways is ill-advised, because it can lead the researcher
into developing and testing models that fail to match the relations in the data.
A key factor that has limited the use of composites has been the aforemen-
tioned problem of parameter identification for models with single effects (as
well as some models with multiple effects). In this example we show that this
can be resolved by reference to a partially reduced form model (of the type
represented by model D), in which specific effects can be evaluated, in con-
junction with the specification of a scale for the composite. Furthermore, we
recommend that for models that incorporate composites, reduced form models
such as model D and single-effect models such as model E should be routinely
examined as part of the evaluation process. When this approach is taken, a
more complete consideration of construct representation is possible for struc-
tural equation models.
As illustrated in Example 2, representing endogenous constructs using com-
posites poses additional theoretical and practical issues. Models of the most
general type, which posit direct effects of exogenous variables on endogenous
composites, while of interest, may be difficult to solve and also to interpret.
Models in which effects on endogenous constructs are modeled through influ-
ences on their dimensions prove to be more tractable, as well as more generally
180 Structural equation modeling and natural systems

interpretable. Again, as a general strategy, comparing results with those from

models with fewer restrictions contributes to the evaluation process.
Modern SEM using latent variables has largely developed through appli-
cations to the social and economic sciences, although it is now widely used
in a broad range of disciplines. Part of the emphasis on common factor-type
models (e.g., the hybrid model) in certain fields is where there is an availability
of redundant measures, such as batteries of questionnaire items or exam ques-
tions, that can be used to estimate underlying causes. Even in these disciplines,
it seems that data and constructs are such that composites could be widely used.
In the natural sciences, redundant measures are much less common, which has
contributed to an emphasis on path analysis using only observed variables,
and little use of latent or composite variables. Certainly multi-indicator latent
variables of the L➙M block type will find wide usage in the natural sciences.
However, we can anticipate that a significant need will develop for considera-
tions of composite variables to represent collections of effects. In the study of
natural systems, it is often the case that the concepts of greatest interest are quite
abstract and heterogeneous. Thus, composites, when used with care, provide
a useful way of summarizing groups of effects that match our theories about
construct relations, and they belong in our SEM tool box.
 7
Additional techniques for complex situations

As mentioned earlier, the array of techniques associated with structural equation

modeling has grown continuously since the LISREL synthesis. Some of this
growth has been associated with refinements and some from new inventions. It
is important for the reader to realize that the development of SEM methodology
is still a work in progress. As described in Chapter 6, the ability to formulate
and solve models containing composite variables, though long discussed, is
only now being achieved. Refinements continue to be made in our ability to
create models that are appropriate for the questions of interest and the data at
hand. Structural equation modeling attempts to do something quite ambitious,
to develop and evaluate multivariate models appropriate to almost any situation.
Not surprisingly, the development of methods to accomplish this goal takes time
and certain statistical limitations must be overcome.
For the most part, the material covered in Chapters 3, 4, and 5 represents
basic principles relating to SEM. Many additional capabilities exist beyond
those covered thus far, and many new developments are emerging. Some of
these additional techniques and new developments are described in this chapter,
although because this is such a vast subject, the treatments presented are only
brief introductions to a select set of topics.
While the material in this chapter is in a section called Advanced topics,
in my experience, the models that are most appropriate to understanding nat-
ural systems often require advanced elements or procedures. Nonlinearities
are commonplace, categorical response variables are often involved, reciprocal
interactions frequently encountered, and data are typically hierarchical in some
fashion. Furthermore, we may wish to create multi-level models that recognize
different processes at different scales, or we may be interested in time-course
models. Thus, creating the most appropriate model for a situation will often
require the use of some advanced techniques beyond the basics covered in the
first section of chapters. This is not to encourage researchers to use models

181
182 Structural equation modeling and natural systems

that are any more complex than necessary. When a simple modeling approach
will suffice, it is to be preferred. Also, when one is starting out with SEM, it is
advisable to begin with simple models to understand the fundamentals involved.
As alluded to in Chapters 4 and 6, there is a long history of misuse of latent
variables and although they are presented in most basic presentations relating
to SEM, to use them properly is an advanced skill.
In this chapter, I will attempt to present in brief and concise form a lit-
tle about a number of more advanced topics. Nearly all of these topics have
one or more books, or many articles written about them. I will not be able to
cover any of them in the depth they deserve in this introductory text. Some
beginning users of SEM may wish to skim this chapter and save the mastery
of these topics for another day. Others may find that their applications require
them to confront some of the topics covered in this chapter from the beginning.
In addition to providing a brief introduction and a limited number of exam-
ples, I will provide references to in-depth treatments of the topics. It can be
argued that SEM is akin to a construction process. As with modern building
construction, tools, techniques, and materials can be used in a variety of cre-
ative ways to solve problems. Here I illustrate a few more of the tools that are
available.

Multigroup analyses
One of the early elaborations of standard SEM practice was the development
of the multigroup analysis. It is not uncommon that one wishes to address
questions about groups. For example, samples may include male and female
members of a population and one may wish to ask whether the multivariate
relations of interest are the same for both groups. Or, we may be interested in
complex interactions involving the responses of different groups. Multigroup
analyses can range from simple (e.g., comparing regressions between groups)
to complex (e.g., evaluating the effects of treatment applications on models
containing conceptual variables). As a fundamental tool in the SEM toolbox,
the multigroup analysis is quite a powerful capability.
The essence of the multigroup analysis is that elements in a data set can
be assigned to one or another group membership, and comparative models that
apply to each group can be developed. Of course, separate models for each group
of samples can always be developed and evaluated separately. However, such
an approach misses the opportunity to arrive at a general answer. What differs
in multigroup analysis is that it is possible to evaluate for all model param-
eters whether they differ between groups and, therefore, whether the groups
 Additional techniques for complex situations 183

z2 y4
d1 x1 l1 l7 e4
l2 g21 l8
d2 x2 x1 h2 y5 e5

x3 l3 g11 b21 l9 y6 e6
d3
z1
h1
l4 l6
l5
y1 y2 y3

e1 e1 e2
Figure 7.1. Example model used to discuss multigroup analysis.

fit a common model. With a multigroup analysis we can determine exactly

what parameters are the same or different across groups, using the data from
all groups simultaneously. As we will see, this framework also permits us to
evaluate differences between groups in mean values, in addition to the correla-
tions/covariances that are the standard stuff of SEM analyses. A slightly more
technical presentation of multigroup analysis, along with ecological examples,
can be found in Grace (2003a) and Pugesek (2003). Example applications can
be found in Grace and Pugesek (1998) and Grace and Jutila (1999). A more
general, nontechnical introduction to the subject can be found in Kline (2005).
Here I give a brief illustration of some of the inferences that can be made by
comparing groups.

Procedures
In multigroup analysis, there are quite a few parameters to compare between
groups (all path coefficients, variances, error terms, and means). Figure 7.1
provides us with an example model to aid in the description of the process.
Additionally, the reader can refer to Appendix 5.1 where a representation of
the LISREL model is given, which may be helpful if one wishes to better
understand the mathematical notation. Essentially, the lambdas (λ) describe
the loadings between latent and observed variables, deltas (δ) are errors for
184 Structural equation modeling and natural systems

indicators of exogenous latent variables, while epsilons (ε) are errors for indi-
cators of endogenous latent variables. Gammas (γ ) describe path coefficients
relating endogenous to exogenous latents, while betas (β) describe path coeffi-
cients relating endogenous to other endogenous latent variables. Zetas (ζ ) are
disturbance terms for endogenous latents.
Because of the large number of comparisons being made, care must be taken
to arrive at a stable solution. Two general strategies have been proposed for
comparing parameters in multigroup analysis. In the first of these, one begins
by performing separate analyses for each group allowing all parameters to
differ between groups. Equality constraints are progressively added and the
appropriateness of such constraints evaluated. Bollen (1989, pages 355–365)
has suggested the following progression as a reasonable way to proceed:

Hform same form

Hλ loadings from latent variables to effect indicators equal
Hλ loadings and correlated errors equal
HλB loadings, correlated errors, and structural path coefficients equal
HλB loadings, correlated errors, structural path coefficients, plus
variances and covariances for endogenous latent variables equal
HλB loadings, correlated errors, structural path coefficients, plus
variances and covariances for both endogenous and exogenous
latent variables equal

where Θ is a matrix containing the variances and covariances for observed vari-
ables, B is a matrix giving the covariances among endogenous latent variables,
Γ is a matrix giving the covariances between exogenous and endogenous latent
variables, Ψ specifies the covariances among errors of exogenous latent vari-
ables, and Φ specifies covariances among errors of endogenous latent variables.
It is always recommended that the first step in a multigroup analysis should be
to consider whether the appropriate model for both groups is of the same form,
for if that analysis fails, it makes little sense to evaluate all the individual param-
eters. The addition of constraints can deviate from the above sequence, how-
ever, depending on the comparisons of greatest theoretical importance. In this
sequence, parameters associated with the measurement model are considered
first, followed by an assessment of similarity in the structural model. Correla-
tions among errors and disturbances are often of least interest, and considered
last. In cases where the questions of most interest have to do with differences
between groups in the structural model (the betas and gammas), they may be
evaluated first. Model evaluation typically employs single-degree-of-freedom
χ 2 tests as well as the use of comparative indices such as AIC.
 Additional techniques for complex situations 185

A second approach to multigroup analysis relies on Lagrange multipliers

(Lee and Bentler 1980). This approach differs from the above-described meth-
ods because the Lagrange multiplier (LM) approach evaluates the validity of
equality constraints multivariately, instead of univariately. In this case, all equal-
ity constraints are evaluated simultaneously, and it is unnecessary to evaluate
a series of increasingly restrictive models in order to find the best final model.
In general, the LM method is both easier to implement and more reliable with
regard to the final solution and is recommended (though not available in all
software packages at the present time).
The reader should be reminded here of the discussion in Chapter 3 relating
to standardized versus unstandardized parameters. As was stated there, when
comparing populations (or groups), unstandardized parameters are often to be
preferred. A key aspect behind this recommendation is the influence that vari-
ances have on standardized parameters. When variances differ between groups
for variables, the basis for standardization of coefficients also differs. Thus, it
is possible for unstandardized coefficients to be the same across groups while
the standardized coefficients appear to differ. The reverse is also possible. Since
the significance tests are based on unstandardized parameters, they are viewed
as a poor basis for comparing groups when variances differ. When variances do
not differ between groups, however, standardized coefficients may be suitable
(as long as they meet the interpretational objectives of the researcher).
In addition to being able to compare all parameters derived from the analysis
of covariances between groups, it is possible to evaluate differences in means
as well. The analysis of means within SEM is based on a regression approach,
where the intercept is used to relate regression results to mean responses. Typ-
ically, the evaluation of means tests to see if the differences between groups is
zero, although interest is often in the magnitude of differences. The analysis of
means within SEM differs from that associated with ANOVA in that it permits
an assessment of the means of latent variables, and it is much more realistic
with regard to the control of covariates, correlated errors, and other types of
nonindependencies.

An illustration
In 1993 and 1994, Heli Jutila conducted a study of vegetation in coastal areas
of Finland. Samples were taken from paired grazed and ungrazed meadows
across several geographic areas. The emphasis in the SEM analysis of these
data (Grace and Jutila 1999) was to understand patterns of species richness and
how they might be influenced by grazing. The effects of geographic region, soil
conditions, and elevation/flooding on community biomass and richness were
186 Structural equation modeling and natural systems

site2 site3
Ungrazed
0
dol
par1 SITE
par2 −0.17
ns 0.10 z
par3 0
par4 0.23
SOIL RICH rich
par5 0.36 0.56
sol1 ns R 2 = 0.56
0 0
sol2 -0.30
FLOD BIOM
sol3
-0.40
R 2 = 0.31
sol4

elev strs mass lmass

site2 site3 Grazed

0
dol

par1 SITE
par2 ns*
0.43* 0.09 z
par3 0
par4 ns*
SOIL RICH rich
par5 0.59 0.56
sol1 0.25* R 2 = 0.47
0 0
sol2 ns*
FLOD BIOM
sol3
-0.31
R 2 = 0.44
sol4

elev strs mass lmass

Figure 7.2. Multigroup results for study of grazing effects. Paths marked with
asterisks in the lower figure are statistically different among groups. From Grace
and Jutila (1999), by permission of Oikos.
 Additional techniques for complex situations 187

believed to apply to both grazed and ungrazed sites, forming the basis for a
multigroup analysis. The form of the model and results for the two groups
are shown in Figure 7.2. The effects of geographic region were modeled as
categorical effects using dummy (yes or no) variables. It is necessary when
modeling with dummy variables to omit at least one variable (in the case of
SITE, sites 1 and 4 were equivalent and were omitted). The interpretation of
the SITE effect is the influence of being at sites 2 and 3 relative to 1 and 4. The
overall influence of sites was modeled as a composite, as described in Chapter
5. SOIL was likewise a composite of the effects of different parent materials
(par 1–6), soil types (sol 1–5), and the depth of the litter layer (dol), with all but
dol being dummy variables. Both flooding (FLOD) and community biomass
(BIOM) had nonlinear relationships with richness, and this was also modeled
using composite effects (this procedure is discussed in more detail later in the
chapter). One set of parameters evaluated in the multigroup analysis but not
shown in Figure 7.2 is the variances, which in this case were not found to differ
significantly. Means were not evaluated in this case as part of the multigroup
analysis.
We can see from the multigroup results (Figure 7.2) that relationships
between SOIL and SITE, as well as SOIL and FLOD, differed between groups.
These differences support the case for suspecting that the soil differences
between grazed and ungrazed meadows might have resulted from grazing itself.
The results further indicate that in the grazed meadows, species richness was
unrelated to site or soil effects, while these relationships were both modestly
important in ungrazed meadows. There is a suggestion from these differences
between groups that grazing effects on richness may be overwhelming site and
soil influences. The other main difference between groups in this study is the
relationship between biomass and richness. We can see that there is evidence
for a strong, predominantly negative effect of biomass on richness in ungrazed
meadows (path coefficient = −0.30). However, in grazed meadows there is no
significant relationship between biomass and richness.
As illustrated in this presentation, multigroup analyses allow for a detailed
inspection of how groups differ. One of the greatest strengths of this type of
analysis is the ability to explicitly consider interactions between the group vari-
able and all within-group relationships. The ability to consider differences in
variances and means between groups also allows for an analysis framework
that is superior to traditional analysis of variance approaches in many regards.
As SEM becomes used more in manipulative experiments, we can expect that
multigroup analyses will become very popular because of the power and flexi-
bility they provide.
188 Structural equation modeling and natural systems

Models containing categorical variables

I should start out by making sure that the reader understands that when categor-
ical variables are exogenous, they pose no real problems for either regression or
structural equation models, as distributional assumptions do not apply. In fact,
categorical exogenous variables are frequently used as alternatives to several
of the procedures discussed in this chapter. Instead of performing multigroup
analyses, such as described in the previous section, ofter analysts will use cate-
gorical variables to represent groups in a single model. Several examples of this
will be presented in Chapter 10, dealing with the analysis of experimental data.
Categorical exogenous variables can also be used to deal with heterogeneous
samples, such as those that involve multi-level sampling (discussed later in the
chapter). In such situations, nominal influences can be modeled as a series of
dummy variables, again relying on categorical exogenous variables.
When categorical variables are endogenous, however, a number of important
issues arise. Some of these issues are sufficiently challenging that many SEM
software programs do not have options for analyzing such models. Both statis-
tical and interpretational issues arise when modeling categorical outcomes. As
we shall see in the next chapter, one of the strengths of Bayesian approaches to
SEM is their inherent fit for modeling categorical responses. For now, however,
we will stay within the framework of conventional SEM and consider both the
difficulties categorical responses pose, and some remedies.
It is quite common for variables to be categorical rather than continuous.
When we count species in a plot or sample, an infinite number of values are not
attainable. Instead, where a maximum of 20 species are obtained in a sample,
that variable has a maximum of 20 levels, and thus, is not a truly continuous
variable. Even when a variable has a central tendency, the very fact that there are
a limited number of levels can result in some deviations from normality. When
the number of levels is high, such deviations are fairly unimportant. The question
we must ask, though, is at what point are the number of levels of a categorical
variable sufficiently low to result in misleading findings? Some effort has been
spent considering this question (Bollen 1989, pages 433–447). Typical answers
fall between 5 and 7 levels as “safe” for treating ordered categorical variables
as if they were continuous. It is also safe to say that when a response variable
possesses only two to three levels, special accommodations should be employed.
The simplest and most extreme case is where only two outcomes are observed,
such as in a dichotomous ordinal variable of the “yes or no” variety.
There are two kinds of problem that arise when dealing with categorical
responses, attenuated correlations and nonnormality. First, correlations involv-
ing ordinal variables, have some undesirable properties. Such correlations are
 Additional techniques for complex situations 189

truncated in comparison to correlations with the underlying continuous response

factor. Imagine that we have two continuous variables, an x and a y, that have a
correlation of 0.5. Further imagine that we score the values of y as being a 0 or 1
if they fall either below or above the mean. If we derive the correlation between
x and the categorical response yc , the correlation will now be substantially less
than 0.5. The reason for this drop is a loss of information. When the underly-
ing continuous response is reduced to a classification, the loss of information
truncates the correlations with other variables. The exact value of the truncated
correlation will depend on the number of responses that fall into the two cat-
egories (0 or 1). Simulation studies show that as the number of categories in
yc goes up, the observed correlation between x and yc will approach the “true”
correlation between x and y. The second issue relating to categorical responses
deals with the nonnormality of residuals. For a dichotomous categorical vari-
able, the precise degree of nonnormality depends on the relative frequency of
0s and 1s. Generally speaking, however, standard errors are biased downward,
leading to inaccurate estimates of probabilities.
In regression models, the approach often taken for dealing with categorical
responses is the logistic model. Oversimplifying a little, the logistic model is
one that treats the responses as approximating an S-shaped response curve,
with the logit transformation being used to linearize that curve. In SEM, the
approach most often taken is based on the probit model work by Muthén (1984).
In the probit model, we assume that underlying the categorical outcomes is
a continuous variable. Stated formally, we assume that behind the observed
response distribution of y is a continuous underlying probability function y*.
Our goal is to understand y* based on our observations of y.
The probit approach is one that involves three steps. In the first step, there
is an estimation of the thresholds. In the case of a dichotomous response, the
assumption is made that if the value of the underlying continuous response is
below the threshold, we will observe a 0. On the other hand, if the value of
the underlying continuous variable is above the threshold, we will observe a 1.
Our estimate of the threshold value depends on the number of 0s and 1s. When
we observe an equal number of 0s and 1s, we would usually assume that the
threshold of y* would be its mean value.
The second step in the probit approach involves the use of polychoric cor-
relations. Stated simply, the polychoric correlation is the correlation between
an x and a y* that would yield the observed correlation between x and y given
the estimated threshold. It is easy to demonstrate the validity of this approach.
If you take some correlated pair of x and y variables and choose a threshold to
use to categorize y into yc , regression using the probit method will allow you
to recover the original correlation between x and y.
190 Structural equation modeling and natural systems

The third step in the probit method involves analysis of the polychoric cor-
relation matrix in place of the observed correlation matrix. We should be aware
that since y is categorical, the mean and variance are not identified parameters.
Therefore, the mean is set to 0 and the variance to 1. As a result, this is one
of the few cases in SEM where we must evaluate the correlation matrix instead
of the covariance matrix.
The remaining problem to address when dealing with categorical response
variables is nonnormality. The earliest approach to addressing this problem
was the asymptotic distribution-free method. Studies have subsequently shown
that extremely large sample sizes are required for this method to yield reliable
results. As an alternative approach, Muthén (1984) has proposed a weighted
least squares method that has been found to behave reasonably well in providing
approximately correct standard errors.

An illustration
When making observations of wildlife species, it is often the case that individ-
uals are observed only occasionally. This alludes to the fact that animal count
data often contain lots of 0s. In a study of the responses of grassland birds
to fire history, Baldwin (2005) found that individual species were encountered
infrequently such that it was best to simply represent the data as either observed
or unobserved. Given the data in this form, the questions posed were about how
different habitat factors relate to the probability of observing a bird species.
In this case, the categorical observations represent y, while the probability of
seeing a bird represents y*. Thus, we envisage the problem as one in which the
probability of seeing a bird varies continuously with habitat conditions, while
some threshold determines whether a bird will be observed or not.
In this study, the primary driving variable of interest was the number of years
since a grassland unit was burned (typically by prescribed burning). Fifteen
management units were selected for sampling, with five having been burned
within one year, five within one to two years, and five within two to three
years. Sampling along transects within the management units produced 173
observations at a scale where bird occurrence could be related to vegetation
characteristics. The vegetation characteristics of presumed importance included
vegetation type, density of herbaceous vegetation, and the density and types of
shrubs. The hierarchical nature of the data was accommodated using cluster
techniques, which are discussed later in the chapter.
Figure 7.3 illustrates the hypothesis evaluated for each of the four most
common species. The two primary questions of interest were (1) whether the
probability of finding a bird in a habitat was affected by the time since burning,
 Additional techniques for complex situations 191

Andro. Spart.
shrublow

shrubmed

brnyr1

bird?

brnyr2
herblow

herbmed

Figure 7.3. Model relating categorical observations of birds of a species to four

types of influence, (1) year since burn (brnyr), (2) density of shrubs (shrublow,
shrubmed, or shrubhigh), (3) density of herbaceous vegetation (herblow, herbmed,
or herbhigh), and (4) vegetation type (Andropogon, Spartina, or mixed).

and (2) whether any relationship between time of fire and the probability of
birds could be explained by associated vegetation characteristics. One may ask
why shrub density and herbaceous density are represented as categorical vari-
ables? In theory, these variables could be represented as continuous influences.
However, examination of the data revealed that the relationships between birds
and vegetation density were nonlinear. The use of ordered categories of veg-
etation density proved effective in this case as a means of dealing with those
nonlinearities.
The results from this analysis for one of the bird species, LaConte’s sparrow,
are given in Figure 7.4. For this species, the probability of observing birds was
least in fields burned three years previously, thus, there was a positive associa-
tion between this species and year 1 fields and year 2 fields. These associations
could be explained as indirect effects, since the associated vegetation conditions
(particularly low shrub density) provide an adequate explanation for the asso-
ciation with burn year. It can be seen that there was an additional influence of
medium levels of herbaceous vegetation on LaConte’s, as well as a preference
for vegetation types that include Andropogon (upland prairie dominated by lit-
tle bluestem). Altogether, these results indicate that this species is most highly
associated with habitat that has been burned between 1 and 2 years previously,
that has a medium level of herbaceous cover, and that is upland prairie. Not
192 Structural equation modeling and natural systems

-0.54

Andro. mixx
0.82
1.12 0.93

0.47 shrublow comm

brnyr1 0.47 0.27
0.38
-0.51 LCSP

brnyr2 0.27 0.65

herbmed
Figure 7.4. Results for LaConte’s sparrows (from Baldwin 2005). A composite
variable, comm, was included to represent the collective effect of vegetation type.

only does the use of categorical modeling accommodate the nature of the data
in this case, the nature of the results would seem to be in a form quite usable
for making decisions about habitat management.

Nonlinear relationships
It has been my experience that nonlinear relationships are very common in eco-
logical problems. There are actually several different ways nonlinear relations
can be addressed, some quite simple, and some more involved. The simplest
means of addressing nonlinear relations is through transformations. In many
cases, transformed variables can be used in place of explicitly modeled non-
linear relations. For monotonic nonlinear relations, simple transformations can
often be quite effective. For transforming unimodal or more complex rela-
tions, however, it must be that the linearization of one relationship does not
create a new nonlinear relation with some other variable. When this can be
avoided, transformations can be effective. An additional way to deal with non-
linear relations was demonstrated in the previous section dealing with bird
populations. By converting vegetation density into an ordered categorical vari-
able, any assumption of a linear relationship involving vegetation was removed.
Multigroup analysis, as discussed in the first section of this chapter, can be used
 Additional techniques for complex situations 193

to address a certain type of nonlinear relation, the interaction. In the example

shown in Figure 7.2, by separating grazed and ungrazed meadows into con-
trasting groups, it was possible to allow the slopes of all relationships to differ
between groups. When a slope does differ between groups, that constitutes a
multiplicative nonlinearity, which we can think of as a significant interaction
between group and relationship.
Continuous nonlinear relationships create some special challenges for struc-
tural equation models, both for observed variable models and especially for
latent variable models having multiple indicators. Challenges posed by non-
linear interactions face both studies of ecological relations and evolutionary
biology (Scheiner et al. 2000). Kline (2005, chapter 13) provides a useful intro-
duction to the topic. Here I will show one example to illustrate how a simple
nonlinearity can be dealt with, and the role that composites can play in sum-
marizing nonlinear effects.

An illustration
Here I give some results from Grace and Keeley (2006) who examined the
response of vegetation to wildfire in California chaparral. Some nonlinearities
could be addressed using transformations. However, the relationship between
species richness and plant cover was one where an explicit modeling of the
nonlinearity was of interest. The results of this analysis can be seen in Figure
7.5, and afford us a means of seeing how this problem was addressed. First,
visual inspection of the relationship between cover and richness revealed a
simple unimodal curve. For this reason, we felt that a polynomial regression
approach could be taken using cover and cover squared. This type of approach
to modeling nonlinearities can often be effective, and is particularly appropriate
where some optimum condition for a response exists. In this case, it appears that
100% cover is optimal for richness, and that at higher levels, richness declines.
The introduction of a higher order term, such as Plant Cover Squared, creates
some potential problems. First, variables and their squares are highly correlated
and this must be modeled. In Figure 7.5, the magnitude of the correlation
between Plant Cover and Plant Cover Squared is not shown, as it is not of
substantive interest. Instead, we followed the recommendation of Bollen (1998)
and represented the relationship using a zigzag arrow that depicts that Plant
Cover Squared is derived from Plant Cover. In some cases, the introduction of
a higher order term results in a host of additional covariances in the model that
must be included, but which are not of substantive interest because the higher
order term is merely included for the purpose of representing the nonlinearity.
To solve some of the potential problems associated with the inclusion of poly-
nomial terms, the higher order terms should be derived by first zero-centering
 194 Structural equation modeling and natural systems

Landscape Conditions Local Conditions

community
heterogen.
distance
from 0.95 optimum
coast abiotic

1.0 Hetero-
0.40 geneity 1.0 R 2 = 0.53
0.38
Landscape
Position 1.0
Richness spp/plot
0.26

0.46 Local 0.25

Abiotic 0.32 0
Optimum
-0.28
Optimum
ns Cover

-0.43
Stand 0.45 Fire Plant
Plant
Cover
Age Severity Cover
Squared

1.0 1.0 1.0 1.0

fire total total
age cover
index #1 cover
squared

Figure 7.5. Relationship between herbaceous richness and various factors fol-
lowing fire in southern California chaparral (from Grace and Keeley 2006). This
model includes an explicit nonlinearity between Plant Cover and Richness, along
with a composite variable summarizing the relationship called Optimum Cover.

the original variable before raising to a higher power. This reduces the correla-
tion between first and second order terms. Second, one should always specify a
correlation between polynomial terms to represent their intercorrelations. High
correlations between terms may still exist. However, the use of maximum like-
lihood estimation reduces problems potentially created by high multicollinear-
ity, such as variance inflation, nonpositive definite matrices (ones that contain
apparent contradictions), and empirical underidentification.
One further problem associated with polynomial regression is the interpre-
tation of path coefficients. Typically, we would interpret a path coefficient, say
from Plant Cover to Richness, as the expected change in Richness if we were
to vary Plant Cover while holding the other variables in the model constant.
However, it is not reasonable to expect to be able to hold Plant Cover Squared
constant, since it is directly calculated from Plant Cover. Thus, the independent
coefficients from Plant Cover and Plant Cover Squared do not have meaning,
but instead, it is their collective effects that are of interest. To address this, we
 Additional techniques for complex situations 195

included a composite of latent effects, which is designated Optimum Cover. As

discussed in Chapter 5, composites have their own set of issues to address, thus,
this approach to modeling a nonlinear relationship involves several steps. That
said, the resulting model would seem to effectively represent the relations of
interest.

Latent growth models

A relatively recent innovation in SEM is the development of latent growth
models. These are growth models in the sense that they deal with progressive
temporal dynamics. Some, though not all, formulations of latent growth models
require the analysis of the raw data, because they are based on a repeated mea-
sures concept. Latent variables are used in such models to represent population
parameters, such as means, intercepts, and slopes, as well as to represent the
influences of covariates. Covariates here refer to treatment variables, nonlinear
terms, and both time-varying and time-invariant factors. To use such models,
the researcher needs to be working with longitudinal data measured on at least
three different occasions. Here I will introduce a simple example that illustrates
a few of the many capabilities of latent growth modeling. Additional informa-
tion about latent growth models can be found in Duncan et al. (1999) and Kline
(2005, chapter 10).

An illustration
In the previous section, the illustration involved a study of vegetation responses
following wildfire (Grace and Keeley 2006). Nonlinear relations were illus-
trated as part of a static model representing spatial variations in diversity across
a heterogeneous landscape. In the same study, the authors were also interested
in certain hypotheses regarding changes in diversity over time. Previous studies
suggested that following fire in these systems, there is an initial flush in herba-
ceous plant diversity associated with fire-induced germination. Over time, it is
expected that diversity will decline as time since disturbance increases. Exam-
ination of the temporal dynamics of richness (Figure 7.6A) shows that there
was only a hint of a declining trend. The authors were aware, however, of sub-
stantial year to year variations in rainfall. Furthermore, there is strong evidence
that rainfall in this system is a substantial limiting factor for plant richness.
Patterns of year to year rainfall (Figure 7.6B) suggest the high levels of rich-
ness observed in 1995 and 1998 can be explained by the very high amounts
of rainfall in those years. To evaluate this hypothesis, a latent growth model
was constructed, with annual rainfall included as a time-varying covariate. The
results are shown in Figure 7.7.
196 Structural equation modeling and natural systems

A 60

Richness, spp/plot
50
40
30
20
10
0
94 95 96 97 98
B
Precipitation, mm

300
250
200
150
100
50
0
94 95 96 97 98
Figure 7.6. A. Changes in herbaceous richness over a five-year period following
fire (from Grace and Keeley 2006). B. Variations in annual precipitation accom-
panying changes in richness.

The results from the latent growth analysis of richness showed that the data
are consistent with the hypothesis of a general decline over time. When the
influence of rainfall was controlled for in the model, an average decline of 4.35
species per year was estimated for the five-year period, although the decline
function was not strictly linear. For the overall model, a robust chi-square of
14.2 was found with 9 degrees of freedom and an associated p-value of 0.115.
The variance in richness explained by the model for the five years were 34, 62,
68, 76, and 60%. Overall, these results support the contention that behind the
somewhat noisy observed dynamics are two major influences, post-disturbance
decline and responsiveness to precipitation.

Hierarchical data and models

It is very common that data are hierarchically structured. Samples may be
associated with sample units that are located within larger units or clusters.
This is almost universally the case in studies of sessile organisms, and also
 Additional techniques for complex situations 197

INTERCEPT SLOPE

rich94 rich95 rich96 rich97 rich98

ppt94 ppt95 ppt96 ppt97 ppt98

Figure 7.7. Latent growth model representing temporal dynamics in herbaceous

richness based on the influence of a common intercept and a common (though
complex) slope, as well as the influence of year to year variations in precipitation.

typically true for sampling efforts involving mobile animals. When the hier-
archical structure of data is ignored, it can lead to downward biased standard
errors; when hierarchical structure is incorporated into the analysis, it can lead
to an increased understanding of the system.
There are several different ways that hierarchical structure can be modeled,
ranging from the simplistic to the sophisticated. We have examined a couple of
the possibilities in previous sections. Multigroup analysis is an explicit means
of dealing with hierarchical structure. This approach is most useful when there
are primary hypotheses of interest about the highest level of structure. If one
is interested in comparing survival of males and females in a population, a
multigroup analysis is a good way to look at main effects and interactions.
Sampling structure can also be dealt with using exogenous categorical variables
representing that structure. In Figure 7.2, we can see that the influences of three
geographic areas (sites 1–3) along the Finnish coast were accommodated using
two dummy variables. Such an approach permits the hierarchical structure to be
included in the model, and also allows for general evaluation of the differences
attributable to each site. Using variables to represent sites, as with certain other
approaches, does not, however, allow for interactions between the site and the
other processes to be considered.
198 Structural equation modeling and natural systems

When the units of hierarchical structure are not of interest, for example, one
uses cluster sampling and is not particularly interested in the differences among
clusters; the data can be adjusted to make the clustering transparent. This kind
of correction can be done by hand or is implemented in some software packages.
The basic approach is to estimate standard errors and chi-square tests of model
fit, taking into account nonindependence of the observations.
A somewhat recent development in the field of SEM are software packages
that permit multi-level modeling (also known as hierarchical linear modeling).
Multi-level modeling recognizes that there are relationships of interest within
and between sample levels. For example, if one samples a large number of
populations collecting data on individuals in each population, it is of interest to
ask questions about the processing operating within and between populations.
Multi-level modeling permits the development and evaluation of population and
individual models simultaneously. Given the interest in the effects of spatial
scale in many ecological problems, multi-level models would seem to hold
promise for addressing a new set of analyses. The interested reader is referred
to Little et al. (2000) and Hox (2002) for a detailed discussion of this topic
and to Shipley (2000, chapter 7) for an illustration of its application to natural
systems.

Feedback loops: nonrecursive models

While recursive models, which we have examined throughout the book, repre-
sent only unidirectional flow of effects, there are many cases where interactions
can be reciprocal, or nonrecursive. In general, nonrecursive models represent
some sort of feedback between system components. Thus, there is an implicit
temporal element in these models. The inclusion of reciprocal interactions can
be viewed as an act of time compression, where the temporal dynamics are
simplified to a static representation of feedback. Often there is more than one
way to model feedback relations. The simplest form of nonrecursive model
involves reciprocal effects (e.g., Figure 7.8A). Feedback can also involve more
than two variables, such as in a case where y1 ➙ y2 ➙ y3 ➙ y1 .
Figure 7.8A provides a representation of a nonrecursive feedback model.
In this model we posit that correlations between two variables, y1 and y2 can
be explained, in part, through their interactions. In Chapter 2, we looked at the
interactions between a plant, leafy spurge, and two species of insects that feed on
it. In that case, since timecourse data were available, the reciprocal interactions
between species were represented in a modified panel design. A simple panel
 Additional techniques for complex situations 199

z1 z1
A B
b
11
y1 y11 y12
b
b r12 21
b 12
21
b
12
y2 y21 y22
b
22
z2 z2

Figure 7.8. A. A simple, nonrecursive model with reciprocal influences. B. A

panel model in which reciprocal interaction is played out over time; yij is entity i
at time j.

model is shown in Figure 7.8B for comparison. Here it is represented that

variables y1 and y2 influence each other over time (from time 1 to time 2). An
influence of each variable on itself over time is also represented in the model
(by the paths from yi1 to yi2 ).
As was shown in Chapter 2, a temporal or dynamic approach to studying
reciprocal interactions can partition out many different elements of the inter-
action. This might lead one to think that reciprocal interactions should always
be represented using timecourse data. The detailed nature of panel models
can become limiting when representing entities changing over time along with
many factors that influence those entities. One other complication is the need
for the precise timecourse data that captures the various steps in the interaction.
Collectively, these challenges mean that in some cases we may wish to use
nonrecursive models to represent reciprocal feedback relations. In such a case,
we are focusing on the equilibrium net effects of the interaction instead of the
dynamics and components.
If we stop to think about it, we can anticipate that nonrecursive models
may pose some special challenges. Feedback loops imply an infinite series of
back and forth effects between variables. This means the errors of the vari-
ables possessing reciprocal interactions can be expected to be correlated. So,
nonrecursive models will typically possess (1) extra arrows (e.g., two arrows
between two variables), and (2) correlated errors. With so many parameters to
estimate, such models can be expected to have special problems with model
identification. Below I provide an illustration of how limiting this problem can
be for the inclusion of feedback in structural equation models.
200 Structural equation modeling and natural systems

There are a couple of other aspects of nonrecursive models worth mentioning.

The first is the basis issue of estimation. As mentioned in Chapter 5, recursive
path models with observed variables can be estimated with either OLS (ordi-
nary least squares) techniques or using ML (maximum likelihood estimation)
methods. However, nonrecursive models cannot be correctly estimated using
OLS. The nonindependence of errors of the sort found in nonrecursive mod-
els leads to improper solutions. For this reason, typically maximum likelihood
estimation is used. It is worth pointing out that there is actually a type of least
squares approach called two-stage least squares (2SLS) that is not only suit-
able for estimating nonrecursive models, but has some advantages in dealing
with the fact that in these models there are correlations between residuals and
causes, which violates the assumptions of other estimation methods. The reader
interested in this topic should refer to Bollen (1996).
Nonrecursive models have special issues regarding interpretation. Referring
back to Figure 7.8, not only does y1 have an effect on y2 , it also has an indirect
effect on itself. In the simple case where β12 = β21 = 0.5, the indirect effect
on itself through one feedback loop is 0.5 × 0.5 = 0.25. Also, y1 has an indi-
rect effect on y2 . The magnitude of that effect is 0.125. Such indirect effects
diminish asymptotically, nevertheless, they must be accounted for in calcula-
tions of indirect effects and R 2 s. Fortunately for the researcher, many of the
modern SEM software programs perform these calculations for us. What is of
greater importance is that these results depend on an assumption of equilib-
rium for the temporal dynamic that is being represented. As has been shown
by Kaplan et al. (2001), violations of equilibrium cause significant bias in
estimated effects. This places a burden on the researcher to argue convinc-
ingly that the interaction being modeled can be interpreted as an equilibrium
result.

Illustration
There have been very few applications of nonrecursive models to ecological
problems. One exception is the analysis of limnological interactions by Johnson
et al. (1991), who examined interactions between various components of aquatic
foodwebs, including reciprocal interactions between phytoplankton and zoo-
plankton. Their work is discussed further in Chapter 9, as an example of the
use of SEM with experimental studies. To illustrate the potential utility of non-
recursive models in the current, brief discussion, I will rely on a hypothetical
example. In this example, we imagine two potentially competing species that
occur together across some range of environmental conditions. Let us further
 Additional techniques for complex situations 201

z2
x1 g21
y2
b12
g31
g11 b32 b23 y11

y1 y3
b13
z1 z3
Figure 7.9. Nonrecursive model with two reciprocally interacting entities (y2 and
y3 ) and two variables having joint control over them (x1 and y1 ).

imagine that we have an estimate of the degree of correlation between these

species sampled across some area of interest, along with estimates of their stan-
dard errors, which permits us to examine their covariance. By simply referring
to these as species 1 and species 2, we can use Figure 7.8A as a representation
of their association. The first question we might ask is, can we solve for their
individual effects on one another (β12 , β12 ) given the available information? If
we take the variances for y1 and y2 as known from their sample variances, we
have a model with two unknown parameters (the path coefficients), but with
only one known piece of information (the covariance between species). We
know from our discussion in Chapter 5 that this model is underidentified and
cannot be solved for unique values of the parameters. Stated in a different way,
we can derive any number of values for the two path coefficients that will yield
any given covariance or correlation between the two species. So, our model is
inestimable.
Based on earlier discussions, we should expect that any association between
species 1 and species 2 might be spurious unless there are no common fac-
tors that influence their abundances (other than their competitive interaction).
Consider again the two flea beetle species that feed on leafy spurge, from our
example in Chapter 2. Any association between these two could be influenced
by a joint dependence on spurge for their food. Their association might also
be affected by some habitat quality variable that influences their survival or
reproduction. Based on this thinking, we might have a model such as the one
in Figure 7.9. The question is, can we solve such a model?
202 Structural equation modeling and natural systems

A consideration of model identification in nonrecursive models

The first question we must ask for any model, but particularly for nonrecursive
ones, is whether the model is identified. The basic requirements for identification
were discussed in Chapter 5. Those basic requirements include the t-rule, which
states that there must be as many unique pieces of information available as there
are parameters to estimate. Other requirements include an adequate sample size
and sufficient independence for each variable such that it constitutes unique
information (i.e., empirical identification). For a recursive path model, this will
usually be achievable as long as the model does not possess the simultaneous
properties of being saturated plus having correlated errors. For nonrecursive
models, there are additional challenges.
Examination of the model in Figure 7.9 shows that there are 8 path coeffi-
cients to be estimated. As there are 4 variables, the total number of knowns is
6. This is a case of a saturated model that also contains a reciprocal interaction
and a correlated error. Typically, such a model cannot be identified unless two
of the paths can be omitted. So, what alternatives are there?
One approach to satisfying the t-rule would be to obtain multiple indicators
for the variables in the model and to construct a latent variable model. Having
even a few multiple indicators would allow one to have more knowns than
unknowns. However, while the t-rule is a necessary condition for identification,
it is not sufficient. Kline (2005, pages 242–247) summarizes two additional
rules that must apply for model identification. It is beyond the scope of this
overview chapter to describe this somewhat intricate set of rules in detail. In
essence, these rules state that one cannot achieve identification of a nonrecursive
model that has correlated errors, unless there are particular patterns of effects
omitted from the model. For our example in Figure 7.9, we cannot estimate
both γ21 and γ31 , nor can we estimate both β12 and β13 . So, only a model with
independent effects from x1 and y1 on each of the species (represented by y2
and y3 in this model) can be estimated.

Summary
Reciprocal effects and feedback loops can be important components of systems.
Representing these temporal dynamics in static models using nonrecursive rela-
tionships has its limitations. There are a number of things that can be done to
arrive at estimable models. However, the restrictions are rather severe and many
models will not be estimable. For this reason, it is perhaps more important in
this case than in any other for the researcher to consider carefully potential
 Additional techniques for complex situations 203

identification problems when developing their initial models, and when select-
ing parameters to measure. It is also very important to consider whether the
assumption of equilibrium is approximated for the system, otherwise, estimates
of reciprocal effects will not be unbiased. For all these reasons, one should be
careful when incorporating feedback in structural equation models, and should
treat such models as complex cases requiring advanced techniques.
 PA RT I V

Applications and illustrations

 8
Model evaluation in practice

The contribution of model evaluation

to scientific advancement
In Chapter 5 I described how the estimation process allows for evaluations of
model fit. Such evaluations form the core of SEM. In the last section of this
book I will go on to discuss how this contributes to the importance of SEM
as a research tool. Briefly, I believe that because attention has been focused
on univariate models, our theories have remained simplistic, emphasizing indi-
vidual mechanisms rather than an understanding of systems. The evaluation of
general ecological theories has also suffered from being very informal, with
few rules agreed upon for deciding how evidence should be weighed in favor
of one theory or another. This informality has often resulted in prolonged and
unresolved debates (e.g., Grace 1991, Grimm 1994, Stamp 2003); signs of an
immature scientific process in operation.
In this chapter I will present a few examples of the evaluation of struc-
tural equation models that have been applied to natural systems, relying on the
methods presented in Chapter 5. As mentioned in the earlier discussion of SEM
principles, we must always be cognizant of the fact that results and conclusions
from SEM are dependent on the appropriateness of the model specified. As
was illustrated earlier and as will be shown in some of the chapters that follow,
inappropriate models can produce results that are quite misleading.
The current chapter is important because it describes some of the procedural
steps involved in arriving at our best approximation of the correct model. What
is often (though not always) omitted from papers dealing with SEM applications
in the natural sciences is a description of the degree to which investigators gain
a new level of understanding of their system through the model evaluation
process. Our first empirical example, which is presented a few pages below, is a
unique illustration of this learning process. Here we will see laid out before us

207
208 Structural equation modeling and natural systems

the process whereby multivariate models are evaluated and conclusions drawn,
including (1) the evaluation of adequacy for individual models, (2) comparisons
among alternative models, and (3) model selection.

Types of model evaluations and comparisons

We should start by recognizing that there are several categories of model eval-
uation types. A simple scheme is to describe them as (1) strictly confirmatory,
(2) involving a nested series of models, or (3) purely exploratory. This clas-
sification recognizes the importance of the strength of a-priori support for a
causal structure. It is always valuable to keep in mind the words of Sewall
Wright (1968) who stated, “The method itself depends on the combination of
knowledge of degrees of correlation among the variables in a system with such
knowledge as may be possessed of the causal relations. In cases in which the
causal relations are uncertain, the method can be used to find the logical conse-
quences of any particular hypothesis in regard to them.” What is clearly implied
here is that the researcher must judge the strength of support for any causal struc-
ture and mechanistic interpretation, and involve this degree of support in the
analysis process. The data set currently in hand may or may not represent our
strongest source of information about the system being studied. As with the
Bayesian philosophy, the relative strengths of prior and current information are
situational and influenced by many things. This cannot be ignored when eval-
uating models and attempting to draw conclusions that have general validity.
There are many other types of model comparisons that can be recognized
besides the ones just mentioned. For example, an alternative approach to model
evaluation involves the use of sequential examinations, where we determine if
additional variables contribute explanatory or predictive power to a base model.
Such model evaluations represent somewhat different questions that are, them-
selves, sequential. Another kind of evaluation involves non-nested model com-
parisons, where models have different structures, either different numbers or
types of variables, or different causal relations. Non-nested model comparisons
such as these have not been a traditional part of SEM practice. As described in
Chapter 5, there is currently considerable interest in information theoretic ap-
proaches to model selection in regression models (e.g., Burnham and Anderson
2002). While information theoretic approaches to model evaluation are directly
applicable to SEM, their use to select among non-nested models without
strong guidance from theoretical considerations is not to be recommended.
The exception to this is when one is performing exploratory analyses designed
to narrow the range of causal possibilities. Even here, some knowledge of causal
 Model evaluation in practice 209

x1 y2

y1
Figure 8.1. Illustration of a simple model that might be subjected to a strictly
confirmatory analysis.

mechanisms is required in order to make very much progress, if the goal is to

arrive at interpretable results.

Strictly confirmatory evaluations

On one end of the spectrum of possibilities is the strictly confirmatory evalu-
ation. In a strictly confirmatory evaluation, data are compared to only a single
model. Such a case might occur when a researcher has little doubt about the
causal structure in a model. Thus, the focus of the study is usually on the spe-
cific values of the parameters, rather than the choice among models or different
structures. In the case of a strictly confirmatory study, if an evaluation of over-
all model fit shows major discrepancies, the model fails and any subsequent
analysis must be viewed as exploratory.
On first examination, strictly confirmatory evaluations may seem to be rather
dull; this is not the case at all. Often we are primarily interested in the absolute
and relative strengths of various pathways in the model. This kind of interest
was clearly articulated by Sewall Wright in his early development of path
analysis, who stated that path analysis gives quantitative meaning to a qualitative
hypothesis. In a strictly confirmatory evaluation, even if a particular pathway
turns out to have a nonsignificant path coefficient, we may not wish to interpret
that result as an indication that our model is incorrect. Remember, our support
for a causal structure should be supported by alternative sources of information
(e.g., previous experience or known mechanisms), and what is in question in
the analysis is the magnitude of the path coefficients. We can use Figure 8.1 to
illustrate this point. Let us assume that we are interested in three parameters
with causal order as shown. Let us further assume that we have good reason to
believe, based on prior knowledge, that all three pathways among variables can
be important in the system we are studying. The model shown in Figure 8.1 is a
210 Structural equation modeling and natural systems

saturated model, thus, the fit of data to model will be assessed as being perfect,
and no degrees of freedom exist to provide a chi-square test of model fit. In
this case, even if the results show that one of the path coefficients cannot be
distinguished from a value of zero, we should still retain all paths in the model.
It would be a mistake in such a situation to constrain the nonsignificant path to
be zero and re-estimate the model. Rather, we conclude that all pathways are
valid, though in this particular sample, one of the paths cannot be distinguished
from a zero value (perhaps a larger or more extensive sample would have led to a
path coefficient deemed to be statistically significant at p = 0.05, for example).
It is anticipated, thus, that future studies of this system will, at times, find all
the paths in Figure 8.1 to be of significant importance. Commonly in such a
case, we will be interested in the strength of the paths to inform us about the
influence of different processes controlling the behavior of y2 (assuming we
continue to feel we have justification for a causal interpretation).

Nested comparisons among models

We generally recognize two types of model comparisons, nested and non-nested.
Nested model comparisons involve models that are similar in causal structure
(in that they possess the same variables in the same causal order), and differ
only in terms of which paths are deemed significant. For any given model, there
is typically a suite of nested models that differ only through the addition or
subtraction of paths. Figure 8.2 illustrates the nested set of models that exists
for three variables with sequential dependence as in Figure 8.1. We can see
from this figure that for a simple, three-variable model, there are eight models
in the nested series.
We might imagine a couple of diametrically opposed approaches to selecting
a model from a nested set. The strategy of model trimming starts from a model
that includes all the paths that could theoretically occur, and eliminates paths that
are not supported by the data. Assuming some paths are eliminated through this
process, there will be created some degrees of freedom that permit an estimate of
the degree of model fit. Most experts in SEM advise against model trimming as
a routine approach to model evaluation. When there is doubt about the inclusion
of paths in a model, it is more often recommended that the researcher take a
model building approach in which pathways with weak theoretical support are
omitted from the initial model. In such a strategy, only when the model fails
are other pathways considered for inclusion.
A competing models strategy is often possible, and when a limited set of
nested models can be identified as being of prime interest, strength of inference
associated with model comparisons can be protected. The reasons we may prefer
 Model evaluation in practice 211

x1 y2 x1 y2
(A) (B)
y1 y1

x1 y2 x1 y2
(C) (D)
y1 y1
x1 y2 x1 y2
(E) (F)

y1 y1

x1 y2 x1 y2
(G) (H)
y1 y1
Figure 8.2. Nested set of models (A–H) with common causal structure. Solid
arrows represent significant paths and dashed arrows indicate nonsignificant paths.
In this case, models with correlated errors are omitted for simplicity.

a competing models approach are because (1) a more specific evaluation is being
performed, and (2) our comparison-wise error rate is substantially reduced. For
unlimited model comparisons, we should be aware that even simple models
have a substantial number of alternatives (as suggested in Figure 8.2). Models
of moderate complexity can have hundreds of alternatives (especially if we
consider the possibility of correlations among errors).
In the evaluation of any given model, we can simultaneously evaluate
(1) overall model fit, (2) the significance level of included pathways, (3) the
consequences for model fit of including additional paths, and (4) comparative
model fit. Such evaluations are accomplished both by assessments of model
fit (e.g., the chi-square test), through the examination of residuals or modifica-
tion indices, and by using information theoretic measures such as AIC or BIC
(see Chapter 5). Residuals between predicted and observed covariances and any
associated modification indices can indicate that a missing path would signifi-
cantly improve model fit. If such a path is subsequently added, further evaluation
can confirm that the path coefficient is indeed significant. Thus, model evalua-
tion typically involves a comparison among nested models as an inherent part
of the assessment.
212 Structural equation modeling and natural systems

It is important to keep in mind that each data set only permits a single
untainted assessment of a model or limited set of competing models. Frequently,
analyses will suggest either the addition or deletion of paths based on the fit to
data. The fundamental principle of hypothesis evaluation in SEM is that only
the minimum number of changes to a model should be made, and the results
based on a modified model must be considered provisional. Thus, it is not
proper to present results of an analysis based on a modified model and claim
that the resulting model has been adequately evaluated. When a single data set
is utilized, only the initial model is subjected to a confirmatory assessment, the
modified model is actually “fitted” to the data. Where sufficient samples are
available, multiple evaluations can be conducted by splitting a data set, with
each half being examined separately. Even greater confidence can be achieved
if there exist independent data sets for model evaluation.
Every sample is likely to have some chance properties due to random sam-
pling events. Adding excess paths to a model in order to achieve perfect fit
is referred to as overfitting. Overfitting can be thought of as fitting a model
exactly to a set of data, representing both the general features of the data and
its idiosyncrasies. When overfitting takes place, the model no longer represents
an approximation to the underlying population parameters.
As stated above, additional power and interest can be achieved through the
use of a limited set of a-priori competing models. When this is possible based
on theoretical grounds, there is generally a greater opportunity for substantively
based decisions about the appropriate model and a reduced likelihood of laps-
ing into empirically driven changes. We should not, of course, be unwilling
to use empirical findings to suggest a modified model. This kind of empirical
feedback to our models is precisely what we hope to gain from SEM. How-
ever, it is very important to always remember that any change to a model,
including the addition of correlated errors, must be theoretically justified and
interpreted. This sort of explicit process has not been the norm in the natural
sciences up to this point. Next I present an illustration that is an exception to that
pattern.

Example of nested model evaluation: hummingbird pollination

A nice ecological example of hierarchical model evaluation deals with hum-
mingbird pollination of flowers. In 1992, Randy Mitchell (Mitchell 1992) used
data from a study by Campbell et al. (1991) dealing with plant traits and pol-
linator behavior to illustrate certain principles of SEM. Table 8.1 presents the
correlations among floral traits and hummingbird pollination behavior, as well
as the variable means and standard deviations from that analysis.
 Model evaluation in practice 213

Table 8.1. Correlation matrix and standard deviations for plant traits and pol-
linator behavior from Mitchell (1992) based on data from Campbell (1991).
Minimum sample size was 82

Corolla Corolla Nectar No. of Approaches Probes Fruit

length width production flowers per hour per flower set

Corolla length 1.0

Corolla width 0.420 1.0
Nectar 0.367 0.297 1.0
production
No. of flowers 0.149 0.117 0.024 1.0
Approaches 0.390 0.144 0.378 0.231 1.0
per hour
Probes per 0.284 0.079 0.326 0.074 0.743 1.0
flower
Fruit set 0.414 0.338 0.294 −0.140 0.230 0.233 1.0
std. dev. 3.01 0.44 0.26 0.71 0.12 0.11 0.43

Background
In this example, the plant involved was Ipomopsis aggregata, scarlet gilia,
which was pollinated primarily by two species of hummingbird, broad-tailed
and rufous (Selasphorus platycercus and Selasphorus rufus). The goal of this
study was to estimate the effects of plant floral traits on hummingbird visitation
(both in the form of approaches and probes), and the combined effects of plant
traits and hummingbird visitation on fruit set, measured as the proportion of
marked flowers that developed fruits. The plant floral traits examined included
the average length and width of the corolla for flowers open at that time, esti-
mated floral nectar production, and the number of open flowers. Hummingbird
approaches were measured as the number per hour, as were the number of probes
by a hummingbird per flower. Some of the variables were transformed prior to
analysis, including nectar production (as the square root), number of flowers
(as the natural log), and fruit set (as the arcsine square root of the proportion).
The goal of the analysis by Mitchell was to evaluate the two competing
models shown in Figure 8.3. For model A, it is presumed that the floral traits
(corolla length, corolla width, nectar production, and number of flowers) can
affect both approaches to plants and the number of probes per flower. This is the
more general hypothesis of the two as it allows for more biological processes,
such as repeated probing of flowers possessing high nectar content. This more
liberal hypothesis was compared to a more restrictive one, model B. In the
second model, it was reasoned a-priori that hummingbirds might base their
choice of plants primarily on visual cues. This second model also presumed
that fruit set would be affected by probes, not simply by approaches.
214 Structural equation modeling and natural systems

Model A NUMBER OF FLOWERS APPROACHES

COROLLA LENGTH

FRUIT SET
COROLLA WIDTH

NECTAR PRODUCTION PROBES/FLOWER

Model B NUMBER OF FLOWERS

NU APPROACHES

COROLLA LENGTH

FRUIT SET
COROLLA WIDTH

NECTAR PRODUCTION PROBES/FLOWER

Figure 8.3. Competing models relating plant traits to hummingbird behavior

(approaches to flowers and probes per flower), and ultimately to fruit set (1992, by
permission of Blackwell Publishers).

Results
Mitchell found that both models failed to fit the data adequately. For
model A, a chi-square of 23.49 was obtained with 4 degrees of freedom
(p < 0.001). For model B, a chi-square of 26.46 with 9 degrees of freedom
(p = 0.003) was found. Perhaps of equal or even greater importance in this case
was that only 6% of the variance in fruit set was explained by both models.
To make a strong statement about the proper procedures for using SEM,
Mitchell concluded that his two models were not adequate and he did not
continue to explore further configurations that had no theoretical support.

Overview
If one performs an analysis of model A, one finds that substantial residuals
exist for relationships between plant traits and fruit set. If paths from corolla
length and number of flowers to fruit set are allowed, the chi-square drops to
6.08 with 2 degrees of freedom (p = 0.05), and the R2 for fruit set increases
 Model evaluation in practice 215

from 6% to 23%. Mitchell was aware of this possibility, but chose not to take
the analysis in this direction. Instead, he reasoned, “Although these paths [from
floral traits to fruit set] have no clear biological meaning themselves, this may
indicate a common factor (such as energy reserves or photosynthetic rate) that
influences both fruit set and the floral characters.” I applaud Mitchell for taking
this opportunity to make an important point about SEM methodology – that
sometimes the results indicate that key system properties were not measured
and an adequate model cannot be constructed. I also applaud the reviewers and
editor of his manuscript for allowing this “nonresult” to be published, providing
us with a clear demonstration of this kind of SEM application.

Hummingbird pollination revisited

In 1994, Mitchell returned to the question of how plant traits relate to pollinator
behavior and fruit set (Mitchell 1994), this time armed with data to estimate
a more complete model. These data were collected from two, somewhat con-
trasting areas in central Colorado, referred to as the Almont and Avery sites.
Five distinct subpopulations of scarlet gilia plants were examined at each site,
with 24–30 plants assessed in each subpopulation (yielding a total of 139 plants
at Almont and 145 at Avery). Hummingbird approaches and probes per flower
were again assessed for each plant, along with number of flowers per plant,
corolla length, corolla width, and nectar production. Additionally, total flower
production over the season was measured. Nectar production was estimated
using flowers that were covered to prevent consumption and then sampled using
a micropipette. Fruit set was again measured, this time by dividing the num-
ber of developed fruits by the total number of flowers. Additional plant traits
measured in this study (that were not measured in the previous study, Mitchell
1992) included height of the tallest inflorescence, the number of flowers open
at time of observation, and the total dry mass (above and below ground) of each
plant collected approximately 1 month after the period of observation. Plants
for which all measures were not taken were deleted from the data to ensure a
complete data set, yielding 130 plants for Almont and 127 for Avery. The corre-
lations, means, and standard deviations for both Almont and Avery populations
are reproduced in Table 8.2.
Mitchell devised a set of theoretically supported models to evaluate using
this more complete data set. Figure 8.4 presents a graphical summary of most
of these models. The figure shows, using letters B–E, the paths that were either
added (for model B) or subtracted (models C–E) in comparison to model A (the
model specified by all solid arrows). Two additional models not represented
in Figure 8.4, F and G, were also evaluated. Since all models had the same
Table 8.2. Correlations, means, and standard deviations for plant traits, hummingbird behavior, and fruit production (from Mitchell 1994,
Table 1)

Open flowers Corolla length Corolla width Nectar prod. Height Dry mass Total flowers Approaches Probes Fruit set Total fruits
(ln#) (mm) (mm) (µL)−2 (ln cm) (ln g) (ln#) (#/hr) (#/fl/hr) (prop.) (ln#)
Almont
Open flowers 1.0
Corolla length 0.084 1.0
Corolla width 0.190 0.093 1.0
Nectar 0.156 0.350 0.441 1.0
Height 0.368 0.091 0.093 0.107 1.0
Biomass 0.809 0.148 0.206 0.164 0.523 1.0
Total flowers 0.855 0.096 0.195 0.167 0.477 0.935 1.0
Approaches 0.248 0.174 0.157 0.271 0.155 0.226 0.277 1.0
Probes/flower −0.145 0.099 −0.028 0.141 −0.008 −0.115 −0.073 0.676 1.0
Fruit set −0.102 0.127 −0.037 0.213 0.308 −0.012 −0.072 0.182 0.247 1.0
Total fruits 0.748 0.069 0.156 0.152 0.501 0.833 0.871 0.236 −0.050 0.296 1.0
Mean 2.56 27.8 3.22 1.80 3.73 1.12 4.88 0.151 0.056 0.99 4.32
Std. dev. 0.755 2.473 0.399 0.492 0.251 0.717 0.718 0.171 0.069 0.177 0.859
Avery
Open flowers 1.0
Corolla length 0.166 1.0
Corolla width 0.002 0.248 1.0
Nectar 0.154 0.062 0.149 1.0
Height 0.325 0.305 0.291 0.015 1.0
Biomass 0.650 0.340 0.261 0.001 0.402 1.0
Total flowers 0.754 0.178 0.093 −0.036 0.390 0.748 1.0
Approaches 0.233 −0.093 −0.003 0.203 0.080 0.240 0.220 1.0
Probes/flower 0.060 −0.199 −0.063 0.199 0.019 0.167 0.123 0.841 1.0
Fruit set −0.006 −0.106 0.114 0.068 0.254 0.002 −0.108 0.180 0.166 1.0
Total fruits 0.683 0.110 0.027 −0.073 0.417 0.637 0.881 0.197 0.119 0.165 1.0
Mean 2.18 26.8 3.92 1.85 3.62 0.92 4.35 0.156 0.081 0.94 3.52
Std. dev. 0.550 2.483 0.325 0.365 0.285 0.558 0.572 0.208 0.114 0.142 0.643
 Model evaluation in practice 217

E
COROLLA LENGTH E APPROACHES
E
E
COROLLA WIDTH E

E
E
NECTAR
PRODUCTION E

DRY MASS B PROBES/FLOWER

B E D D
D
E
PROPORTION
HEIGHT C FRUIT SET
C
OPEN FLOWERS C D

C TOTAL
TOTAL FLOWERS
FRUITS

Figure 8.4. Representation of the competing models evaluated by Mitchell (1994).

Model A is represented by all solid arrows. Model B involved the addition of two
paths, indicated by the letter B. All other models depicted here (C–E) involved the
removal of paths. By permission of Blackwell Publishers.

structure, in terms of the causal order, this evaluation is considered to be a

nested model comparison.
In brief terms, the models examined in this analysis are as follows:
Model A – This model proposes that fruit set is influenced by both polli-
nator behavior and plant internal resources, and that pollinator behavior
is influenced by certain plant traits. It is specifically hypothesized that
any correlations between floral traits (corolla length, corolla width, nec-
tar production, or number of open flowers) and fruit set will be mediated
through their effects on pollinator behavior (i.e., there will be no direct
paths from floral traits to fruit set or production). In model A, the effects of
dry mass on fruit set and total fruits produced is completely predicted by
plant height and the total number of flowers produced, which themselves
depend on dry mass.
Model B – This model proposes that plant height and total flowers will not
adequately represent the effects of dry mass on fruit set and total fruits.
218 Structural equation modeling and natural systems

Because of this, there will be additional effects of dry mass on fruit set
and total fruits, indicated by the paths labeled B.
Model C – This model evaluates the possibility that fruit set and production
are not limited by plant resources; thus, all paths from height and total
flowers to fruit set and total fruits are omitted.
Model D – This model evaluates the possibility that pollinator behavior has
no effect on fruit set and total fruit production.
Model E – This model evaluates the possibility that nectar production and
visible plant characters have no effect on pollinator behavior.
Model F – This model evaluates the possibility that plant mass is the under-
lying cause of any correlations among corolla length, corolla width, and
nectar production. This model is not causally nested within the general
framework shown in Figure 8.4, because it involves directional arrows
from dry mass to corolla length, corolla width, and nectar production.
Model G – This model considers the possibility that dry mass is merely
correlated with plant height and total flower production, instead of their
causal determinant. This model also involves a change in the basic causal
order.

The statistical fit of these 7 models was evaluated in two ways. First, chi-
squares permitted a consideration of the degree to which the data deviated from
the models. This considers the absolute fit of the model. Second, chi-square
difference tests were used to evaluate whether some models possessed better fit
when paths were either added or removed. As described in Chapter 5, this latter
analysis is based on the fact that differences between chi-squares, themselves,
follow a chi-square distribution.
The results of the evaluation of competing models are given in Table 8.3.
For the Almont site, model A had an adequate absolute fit to the data based on
a chi-square of 28.9 with 26 degrees of freedom (p = 0.315). Adding the paths
from dry mass to fruit set and total fruits (model B) resulted in a decrease in
chi-square of 1.59 with two degrees of freedom. A single degree of freedom chi-
square test requires a change in value of 3.841, and a two-degree-of-freedom
value of 5.991 is required before a significant difference between models is
achieved. Thus, model B was not found to be a significantly better fitting model
than model A. Given the logical priority given to model A in this study, its
greater parsimony, and the fact that paths from dry mass to fruit set and total
fruits are not significant at the 0.05 level, model B was rejected in favor of model
A. Deletion of paths associated with models C–E led to significant increases in
chi-square, also indicating that these models were inferior to model A. Since
model A fitted the data well for the Almont site and because models F and G
 Model evaluation in practice 219

Table 8.3. Measures of fit for the models evaluated in Mitchell (1994)

Goodness of fit Nested comparison with model A

X2 df p X2 df p
Almont
Model A 28.91 26 0.315 ----- ----- -----
Model B 27.32 24 0.290 1.59 2 0.45
Model C 313.1 30 0.000 284.2 4 0.0001
Model D 47.80 30 0.021 18.89 4 0.0008
Model E 75.99 36 0.000 47.08 10 0.0001
Model F 70.41 29 0.000 ----- ----- -----
Model G 23.46 14 0.053 ----- ----- -----
Avery
Model A 59.01 26 0.000 ----- ----- -----
Model B 55.33 24 0.000 3.68 2 0.16
Model C 306.5 30 0.000 247.5 4 0.00001
Model D 68.11 30 0.000 9.10 4 0.059
Model E 90.88 36 0.000 31.87 10 0.0004
Model F 67.16 29 0.000 ----- ----- -----
Model G 25.78 14 0.028 ----- ----- -----

Note – Model A is the baseline model

both fitted the data less well than model A, Mitchell deemed model A to be the
best representation of the data of any of the models.
For the Avery site, model A did not have an adequate absolute fit to the data
based on a chi-square of 59.01 with 26 degrees of freedom (p < 0.001). While
not reported by Mitchell, other indices of model fit agree with this assessment.
The root mean square error of approximation (RMSEA) gives a value of 0.090
and a 90% confidence interval for the RMSEA from 0.054 to 0.125. In order
for the RMSEA to indicate a nonsignificant difference between model and data,
the minimum RMSEA would need to include a value of 0.0 within its range,
which in this case, it did not. Models B–F did not have adequate absolute fits
to the data either. Only model G, which represents the case where all plant
traits freely intercorrelate, showed a model fit that was close to adequate. Based
on this information, Mitchell concluded that the model fitting data at the two
sites was not constant and that the best model for the Avery site was not fully
determined.
There are a fair number of interpretations to be drawn from the results from
this study. For brevity, I will focus only on the results for the Almont site
(Figure 8.5). Mitchell’s original model, which he deemed most biologically
220 Structural equation modeling and natural systems

R 2 = 0.12
COROLLA LENGTH APPROACHES

0.20 0.18 0.76

COROLLA WIDTH

NECTAR
PRODUCTION
R 2 = 0.57
DRY MASS PROBES/FLOWER

0.52 R 2 = 0.27
−0.32 PROPORTION
HEIGHT 0.44 FRUIT SET
0.94 R 2 = 0.21
OPEN FLOWERS 0.41
0.86 R2 = 0.73 −0.28
0.95
TOTAL FLOWERS TOTAL
FRUITS
R 2 = 0.87 R 2 = 0.90
Figure 8.5. Results of analysis for Almont site. Directional paths with dashed lines
were nonsignificant. From Mitchell (1994) by permission of Blackwell Publishers.

reasonable from the beginning, matched the covariance structure of the data.
This does not prove that the model represents the true forces controlling the
system. Instead, the adequacy of model fit serves to “fail to disprove” the model
structure. This is no small achievement for a model of this complexity.
There was substantial variation in the degree to which variation in endoge-
nous variables was explained in the model. Looking first at plant traits, plant
height was only modestly related to dry mass. However, both total flowers and
open flowers were related quite strongly to dry mass. Considering pollinator
behavior, approaches by hummingbirds to plants were significantly related to
both nectar production and the number of open flowers. However, only 12% of
the variation in approaches between plants was explained by these two factors.
Probes per flower were strongly determined by the number of approaches and
the number of open flowers (R 2 = 0.57). As for reproductive success, the pro-
portion of fruit set was significantly related to height, which is a function of
dry mass. Proportion of fruit set was also found to be negatively related to the
total number of flowers, suggesting that the plant was indeed resource limited in
 Model evaluation in practice 221

filling fruits. Variation among plants in total fruits per plant was well predicted
by the total number of flowers (R 2 = 0.90).
Looking at overall path relations, we see that total fruit set was largely
determined by the total number of flowers produced, and thus, ultimately deter-
mined by dry mass of the plant. Hummingbird behavior did not explain any
of the observed variation in total fruits in this model, even though approaches
to plants and fruit set were significantly correlated (r = 0.236). Based on this
result, we must conclude that the correlation between approaches and total
fruits was spurious and not causal. This does NOT mean that approaches do
not affect fruit production in some broader range of circumstances. Rather, it
simply means that the variation in fruit production observed in this data set was
related almost entirely to plant size and not to variation in pollination.
This example does an excellent job of illustrating how critical the model is
to interpretation. Mitchell was painstaking in his evaluation of model adequacy.
This gives substantially more credence to the interpretations that derive from
the results than if he had simply searched for the model that best fitted the
data with little regard to a-priori theoretical justification. It is easy to see from
the results of this analysis, that Mitchell was justified in being cautious about
interpreting the results from the earlier study (Mitchell 1992), where plant size
was not measured.

Example of a step-wise hierarchical analysis

Sometimes the questions we wish to ask about a system are a little different
from the ones illustrated thus far. This can lead to alternative approaches to the
evaluation of multivariate models. An example comes from a study I conducted,
in collaboration with Glenn Guntenspergen (Grace and Guntenspergen 1999),
that sought to use SEM to evaluate whether residual spatial variation in species
richness might be explained by past storm events. In this case, model fit was
only one of the criteria used to compare two competing models.

Background
A number of previous studies, including SEM studies using both experimental
and nonexperimental data (Grace and Pugesek 1997, Gough and Grace 1999),
provided a significant amount of experience upon which to base our initial mod-
els and their interpretations. The question addressed in this study was whether
landscape position could explain unique variation in species richness, that could
not be explained by contemporary environmental conditions. Thus, our ques-
tion was of the step-wise sort. To achieve this objective we compared the two
models shown in Figure 8.6. These models differ in that model B includes two
222 Structural equation modeling and natural systems

km elev1 elev2 elev3 elev4 Model A

1.0 0.95 0.92 0.90 0.98
DIST
FLOOD
SEA 2
−0.67 R = 0.45

−0.76 −0.36

1.0 −0.55 1.
1.0
salt SALT RICH rich
2
R = 0.63 R 2 = 0.54
0.21
ns −0.29
0.31
−0.14
DIST ns −0.62
DIST PABUN
RIVR R 2 = 0.10 2
R = 0.44
1.0
1.0 1.0
m dist ltcap

km elev1 elev2 elev3 elev4 Model B

1.0 0.95 0.92 0.90 0.98
DIST FLOOD
SEA 2
R = 0.45

−0.67
0.56 −0.36
−0.76

1.0 −0.14 1.0

salt SALT RICH rich
R 2 = 0.63 R 2 = 0.66
ns
0.21
0.31 ns −0.29
−0.14
DIST −0.62
ns DIST PABUN
RIVR R 2 = 0.10 R 2 = 0.44
1.0
1.0 1.0
m dist ltcap

Figure 8.6. Competing models used to evaluate whether landscape position vari-
ables, distance from the sea (DIST SEA) and distance from the river’s edge
(DIST RIVR) contributed to the explanation of unique variation in species richness
(RICH). From Grace and Guntenspergen (1999) by permission of Ecoscience.
 Model evaluation in practice 223

paths not found in model A, from DIST SEA (distance from the sea) to RICH
(plant species richness), and from DIST RIVR (distance from river’s edge) to
RICH. The results from this analysis are mentioned again in Chapter 10, where
they are put into the context of a systematic use of SEM to understand plant
diversity patterns.

Results and Conclusions

The process of model selection was based on two criteria. The first criterion
was whether including direct effects of position variables on richness would
lead to a better fit between model and data. The chi-square for model A was
87.69, with 32 degrees of freedom, a p-value less than 0.0001, and a fit index
of 0.937. In contrast, the chi-square for model B was 43.54 with 31 degrees of
freedom, with a p-value = 0.068. Thus, tests of overall model fit indicated that
the data did not fit model A, but they did fit model B.
A second, and in this case critical, criterion was whether more variance in
species richness would be explained by the second model. Just because model
B is a better model than model A, it does not necessarily mean that variance
explanation for species richness is higher. It is entirely possible that model B
could essentially explain the same total amount of variance in richness, but
through different relationships. For this reason, a t-test was used to determine
whether there was a significant reduction in error variance for richness in model
B compared to model A. Results showed that distance from the mouth of the
river did explain an additional 12% of the observed variance in richness, while
distance from the river’s edge did not contribute new information. Based on
these findings, we concluded that some of the spatial variation in species rich-
ness in this system can only be explained by past events, such as previous
tropical storms.

An example of model selection based on chi-square evaluations

Meziane and Shipley (2001) conducted a comparative study of morphological
and physiological traits for 22 plant species. Among their objectives was to
determine the best model to describe the causal relationships among the set
of intercorrelated traits. In addition to considering how their own data fit
various models, they also evaluated two other previous data sets collected for
other species, including one data set that was from an extensive field study. The
traits that were included in their study are given in Table 8.4. Figure 8.7 shows
the five alternative models considered in the study.
224 Structural equation modeling and natural systems

Table 8.4. Traits examined in the study by Meziane and Shipley

Trait Definition
Specific leaf area (SLA) Leaf surface area per gram leaf tissue mass
Leaf nitrogen concentration ([N]) mg nitrogen per gram tissue mass
Net leaf photosynthetic rate (A) nmol CO2 uptake per gram leaf tissue per
second
Stomatal conductance (G) mmol water loss per gram leaf tissue per
second

MODEL A MODEL B

SLA [N] A G SLA [N] G A

ε1 ε2 ε3 ε1 ε2 ε3

MODEL C MODEL D

SLA [N] A G SLA [N] G A

ε1 ε2 ε3 ε1 ε2 ε3

ε1
MODEL E [N]
SLA A G

ε2 ε3

Figure 8.7. Alternative models considered in Meziane and Shipley (2001).

Reprinted by permission of Oxford University Press.

In this study, the authors performed experimental manipulations of light lev-

els and nutrient supply, so as to generate a broad range of responses and to gain
insights into causal relationships. Mean responses to experimental manipula-
tions were examined using ANOVA. Samples were combined across all treat-
ment combinations to evaluate the path models shown in Figure 8.7. Because
 Model evaluation in practice 225

Table 8.5. Chi-squares and probabilities (in parentheses) comparing data on

plant traits to the five models shown in Figure 8.7 (modified from Meziane and
Shipley 2001). ** indicates probabilities < 0.05

Data set Model A Model B Model C Model D Model E

Model df 6 6 4 4 6
trt 1 3.7 (0.72) 12.6 (**) 2.8 (0.60) 7.0 (0.13) 6.4 (0.38)
trt 2 7.4 (0.29) 7.0 (0.32) 5.0 (0.28) 4.3 (0.37) 7.4 (0.28)
trt 3 12.0 (0.06) 10.2 (0.12) 9.9 (**) 4.6 (0.33) 12.9 (**)
trt 4 11.8 (0.07) 13.7 (**) 0.5 (0.97) 3.2 (0.52) 2.6 (0.86)
all trts 14.1 (0.08) 19.4 (**) 10.0 (0.27) 9.5 (0.30) 11.0 (0.20)
(df = 8)
data set 2 29.8 (**) 10.7 (0.10) 22.8 (**) 1.4 (0.85) 29.6 (**)
data set 3 18.9 (**) 13.2 (**) 11.7 (**) 2.1 (0.72) 27.0 (**)
data set 4 74.5 (88) 133.0 (**) 13.0 (**) 93.3 (**) 54.3 (**)

Note that treatments 1–4 were modifications of light and nutrient supply conducted in
this study. Data set 2 was from all plants studied in Shipley and Lechowicz (2000).
Data set 3 was for only the C3-type plants in Shipley and Lechowicz (2000). Data set
4 was from Reich et al. (1999).

MODEL F

SLA [N] A G

e1 e2 e3
Figure 8.8. Final, general model accepted by Meziane and Shipley (2001).
Reprinted by permission of Oxford University Press.

the sample size was low (n = 22), an alternative test statistic was employed that
is exact for small samples and relatively robust to distributional violations (see
Shipley 2000, chapter 3). This test statistic was evaluated using chi-square tests
and results are shown in Table 8.5.
None of the models was found to fit all data sets, although model D did fit all
except those from the comparative field study by Reich et al. (1999). Because
this study sought to provide a unified solution that would hold across all data
sets, the authors fitted all data to a sixth model, which is shown in Figure 8.8.
Chi-square tests indicate that model F fits adequately for all data sets, and thus
represents a general model for the system studied.
226 Structural equation modeling and natural systems

An example of model selection based on consistent Akaike’s

information criterion (CAIC)
Background
Studies of natural selection often rely on an examination of the relationships
between a set of life history characteristics and a variety of environmental
conditions. An example of such an analysis for intraspecific variation in fish
populations comes from the work of Johnson (2002).
Brachyrhaphis rhabdophora is a fish species endemic to Costa Rica, occur-
ring in freshwater streams. Populations can be found living under a wide range of
conditions that vary in fish density, stream productivity, physical habitat charac-
teristics, and predation risk. Substantial life history variation among populations
of B. rhabdophora has previously been documented. Traits of importance to
fitness known to vary include offspring size, fecundity, reproductive allocation,
size at maturity for males, and size at maturity for females.
In this study, populations of B. rhabdophora were sampled in four successive
years from a set of 27 different sites throughout northwestern Costa Rica, with 12
of the sites represented in all samplings. Approximately 200 fish were collected
from each site during each sampling.
Johnson established a general model, shown in Figure 8.9, that represented
a comprehensive hypothesis of the habitat factors that might affect life history
variation in this species. Model solution was accomplished in two stages. First,
principal components analysis was used to calculate estimates of life history
scores. Those scores were then used in an analysis using the SEM software
AMOS. Johnson developed a set of 17 a-priori candidate models based on the
global model in Figure 8.9. These models were evaluated for each of the four
years of data collection using the CAIC index. Only values of CAIC differing
by more than 2 were considered to be substantially different.

Results and Conclusions

Model selection results showed that CAIC values were lowest for single-
selection factor models for all years (Table 8.6). Furthermore, CAIC values
were unable to distinguish among alternative hypotheses which selective factor
was most important. These findings led the author to suggest that the selective
environment may not be easily decomposed into individual components. It is
important to keep in mind, however, that the sample size in this study was
sufficiently low that these conclusions may not be consistent with tests of abso-
lute fit. Typically, I would always recommend that models selected be further
examined for adequate fit using measures of absolute fit, such as the chi-square
test.
 Model evaluation in practice 227

Canopy
Size at Maturity
Resource Males
Availability

1
Density 5 Size at Maturity
Females
Density
Effects 2
Life Reproductive
6 History Allotment

7 Extrinsic
3
Mortality 8 Number of
Offspring
Predation Unknown
4 Effects

Habitat Offspring
Stability Size

Width
Flow
Gradient

Figure 8.9. Global model evaluated by Johnson 2002 for the effects of envi-
ronmental conditions on life history characteristics of the fish B. rhabdophora.
Hypotheses tested relative to this diagram and their results are given in Table 8.6.
Reproduced by permission of Oikos.

Equivalent models
A careful consideration of alternative models often demonstrates that there are
multiple models existing that are indistinguishable (Figure 8.10). This may be
the result of statistical equivalency (because there is insufficient power to dis-
criminate among alternatives) or because of pure equivalency (the existence
of alternative causal structures with identical statistical expectations). Further-
more, for saturated models (those specifying all possible interactions among
variables), all variations of the causal order of the model are also saturated and,
therefore, equivalent. There has been a significant discussion of this issue for
over a decade among SEM practitioners (Glymour et al. 1987; Hayduk 1996,
chapters 3 and 4; Raykov and Penev 1999; Spirtes et al. 2000). Recently,
Shipley (2000) has discussed this important topic in detail using ecologically
relevant examples.
A variety of rules for identifying equivalent models have been devel-
oped, the historically most important being the development of the TETRAD
228 Structural equation modeling and natural systems

Table 8.6. A-priori set of 17 candidate models evaluated by Johnson (2002)

using CAIC. For the structure of the model, see Figure 8.9. Selective agents are
as follows: R = resource availability, D = density effects, M = extrinsic mor-
tality, and H = habitat stability. Sample sizes show the number of populations
sampled each year (n)

Selective Paths in CAIC 1996 CAIC 1997 CAIC 1998 CAIC 1999
agents model n = 21 n = 14 n = 16 n = 19
R 1,8 20.2 18.5 18.9 19.9
D 2,8 20.4 18.4 19.1 19.7
M 3,8 20.9 18.3 18.9 22.0
H 4,8 20.7 18.2 19.6 20.2

R+D 1,2,8 33.1 29.6 31.5 32.5

R+D 1,5,8 32.6 32.6 32.3 29.8
R+D 1,2,5,8 36.5 32.8 34.2 35.6

D+M 2,3,8 36.1 32.8 30.8 34.4

D+M 2,6,8 38.7 32.4 34.0 46.6
D+M 2,3,6,8 37.4 32.8 34.3 38.4

D+H 2,4,8 34.1 31.3 42.2 32.7

D+H 2,7,8 39.0 29.0 29.6 38.4
D+H 2,4,7,8 37.7 32.9 34.4 35.6

R+D+M 1,2,3,8 48.6 53.3 43.3 50.0

R+D+M 1,2,3,6,8 50.0 53.4 46.9 53.9
R+D+M 1,2,3,5,6,8 53.3 56.5 49.6 57.0
R+D+M+H all (1–8) 78.0 84.1 71.8 84.3

software program (Glymour et al. 1987), which automates the identification of

equivalent models for a given set of data. Shipley (2000) reduces much of the
elaborate discussion of this issue to some simple rules for recognizing which
model structures indicate the presence of equivalent models that lead to iden-
tical covariance expectations. I will not go into detail on this important topic
here, but instead refer the reader to recent treatments of this subject cited above.
I will, however, comment on the importance of examining alternative models
as a means of checking one’s causal logic.
Hayduk (1996) makes a compelling case for the value of considering equiv-
alent models as a way of advancing one’s thinking about their system. For
systems where the directionality of causal effect is clear, equivalent models can
 Model evaluation in practice 229

ec ec

C C
ed ed
A A
eb
D D
ee ee
B B
E E

ec

C C
ed
A A
eb eb
D D
ee ee
B B
E E
Figure 8.10. Example of a few equivalent models that possess identical covariance
expectations.

easily be dismissed and robustness established. For systems with a high degree
of integration or feedback, equivalent models can challenge one’s assumptions
and interpretations. To some degree, considering equivalent models provides
one with a test of where our understanding lies relative to this problem.

Discovery triggered by lack of model fit

An important contention of this chapter and even of this book is that the process
of model evaluation can lead to new discoveries about the system under study.
Here, one can make an analogy to celestial mechanics, emphasizing the numer-
ous discoveries about our solar system and universe that have resulted from
the discovery of discrepancies between the predicted and observed planetary
orbits. Examples of discoveries resulting from findings of inadequate model fit
and residual relationships are accumulating for ecological studies as well. Here
I present one brief example from Grace and Pugesek (1997) to illustrate the
point.
230 Structural equation modeling and natural systems

RMSEA = 0.085
DIST BIOM LIGHT
GHT 90% CI = 0.059; 0.112
p-value = 0.016
Df = 28
dist massm2 lighthi lightlo

RMSEA = 0.065
DIST BIOM LIGHT 90% CI = 0.034; 0.094
p-value = 0.188
Df = 2
dist massm2 lighthi lightlo

Figure 8.11. Illustration of lack of model fit (upper figure) resolved by addition
of a pathway (lower figure). This figure only shows part of the full model (refer to
Chapter 10, Figure 10.10, for the full model). DIST refers to disturbance, BIOM
refers to above-ground biomass (living and dead), LIGHT refers to percent of
full sunlight reaching the ground surface. RMSEA is the root mean square error of
approximation (refer to Chapter 5). Results indicate that data deviated significantly
from the upper model (p = 0.016), but fit to the lower model was substantially
better (p = 0.188).

In the study by Grace and Pugesek (1997), we proposed initially that dis-
turbance affects plant biomass, which in turn affects light penetration (Figure
8.11). For simplicity, only part of the whole model is shown here, although it
is important to realize that the results are only valid for the full model (Chapter
10, Figure 10.10).
The finding that there was unexplained residual variance between DIST
and LIGHT was initially a surprise. How is it that disturbance, which in this
system was caused by grazers (nutria, rabbits, wild hogs) and by wrack deposits
or scouring by water, could affect light, independent of effects on biomass?
More detailed examination of the data suggested an interpretation. It seems
that vegetation that has recently been disturbed has a more upright architecture
than less disturbed vegetation. The result is that disturbed vegetation allows
more light to penetrate through the canopy per gram biomass than undisturbed
vegetation (Figure 8.12).
This finding turns out to have led to a line of thought which, I believe, may
be of quite general importance. These and other subsequent results suggest
that (1) light penetration is a key predictor of the dynamics of diversity in plant
 Model evaluation in practice 231

Figure 8.12. Illustration of variations in plant architecture, the importance of

which was discovered through an examination of residual effects. Plants on the left
and right are presumed to have equal above-ground biomass values. Those on the
left represent the architecture associated with recently disturbed sites, while those
on the right represent the undisturbed vegetation. Through effects on architecture,
disturbance leads to a greater penetration of light through the canopy, which allows
other species to establish and persist in disturbed sites.

communities (Grace 1999), and (2) plant morphology and litter dynamics should
be far more important than total biomass in explaining community diversity
dynamics (Grace 2001). Empirical studies to evaluate this proposition are lead-
ing to a much more mechanistic understanding of richness regulation that may
permit us to predict, among other things, the effects of dominant (including
invasive) species on native diversity patterns (Jutila and Grace 2002).

Summary
A serious analysis of model fit is important both for the evaluation of scientific
hypotheses, as well as the results and their interpretation being dependent on
the model. When theory is well understood, samples are adequately large and
representative, and measurements are reliable, then evaluations of model fit
provide clear indications of the match between hypotheses and interpretations.
One finds that most evaluations of model fit in those fields with a substantial
history of using SEM (e.g., social and economic sciences) are of the nested type,
where it is generally recognized that there is a limit to the ability of statistical
testing to resolve issues of causal structure. The lesson for ecologists is that
care must be exercised in drawing conclusions too firmly based on a single data
set, regardless of how well the data fit a model.
Biologists and other scientists in many different fields are just beginning to
explore the utility of SEM. This means that modeling efforts are often quite
exploratory, with many more unknowns than knowns. The example by Mitchell
presented early in the chapter represents an outstanding illustration of how
232 Structural equation modeling and natural systems

exploratory analyses can proceed with care, and lead to convincing conclusions
when the conventions of SEM model evaluation are applied. Some questions,
of course, are more difficult to address than others. This is notably the case for
interacting components of individuals or populations, where the high degree of
integration and feedback makes causation sometimes difficult to disentangle.
Here, we will require both persistence and patience to make progress, since the
observations we require are often novel and without precedence in ecological
study.
I believe the process of evaluating multivariate models can be a pivotal
step forward in the maturation of the ecological sciences. This process can be
unambiguously strong in its rejection of pre-existing theory. However, through
the examination of residual effects, one can often discover new relationships
that are of importance, which motivates one to modify one’s theory so that a
better match with nature is obtained. The net results of such model evaluations
will be a reduction in entrenched theory, a higher empirical content to new
theories, and will allow us to have a greater capacity to understand and predict
system behavior.
 9
Multivariate experiments

Basic issues
The gold standard for studying causal relations is experimentation. As Fisher
(1956) labored so hard to demonstrate, experimental manipulations have the
ability to disentangle factors in a way that is usually not possible with non-
experimental data. By creating independence among causes, experimentation
can lead to a great reduction in ambiguity about effects. There is little doubt
for most scientists that well designed and properly analyzed experiments pro-
vide the most powerful way of assessing the importance of processes, when
appropriate and relevant experiments are possible.
In this chapter I address a topic that generally receives little attention in dis-
cussions of SEM, its applicability to experimental studies. I hope to deal with
two common misconceptions in this chapter, (1) that multivariate analysis is
only for use on nonexperimental data, and (2) that when experiments are possi-
ble, there is no need for SEM. In fact, I would go one step further and say that
the value of studying systems using SEM applies equally well to experimental
and nonexperimental investigations.
There are several reasons why one might want to combine the techniques
of SEM with experimentation. First, using experiments to evaluate multivariate
relationships provides inherently more information about the responses of a
system to manipulation. It is often difficult and sometimes impossible to exert
independent control over all the variables of interest in a system. Examination of
how the various pathways among variables respond to experimental treatment
can yield important insights into system function and regulation. It can also
isolate effects of interest in the presence of covarying factors.
A second reason to combine SEM with experimental studies is the fact
that “replicates” are often quite dissimilar in important ways. One clear case
of such a situation is with manipulations of ecosystems. Often experiments

233
234 Structural equation modeling and natural systems

involve units of some size in order to incorporate all the processes of interest,
and to achieve some degree of realism. The various “replicates” being studied
may differ amongst themselves in characteristics that cause their individual
responses to vary widely. The use of SEM automatically leads the researcher to
consider the broader suite of variables that may influence results. These covari-
ate variables can often be of overwhelming importance, and understanding their
role in the operation of the system can greatly improve a researcher’s chances
of obtaining general and coherent findings.
Yet a third reason to use SEM when designing and analyzing an experiment
is a desire to accommodate cases where ecosystem manipulations are not simple
or precise. For example, in the case of prescribed burning in grasslands, a fire
started in the morning may have quite different characteristics from a fire ignited
the afternoon of that same day. Even within an individual fire, spatial variations
in vegetation, topography, and fire behavior may cause substantial variation in
fire temperatures, residency times, and soil heating. Recognizing that treatments
such as burning may not be precise, automatically encourages the researcher
to measure covariates that might impact the responses and to incorporate these
covariates into the model to be evaluated.
As a fourth motivation, experimental multivariate studies can help evaluate
assumptions and predictions that arise from nonexperimental manipulations. By
combining intensive multivariate experiments with extensive nonexperimental
studies, the strengths of both approaches can be integrated into a broader under-
standing of system behavior.
At present, SEM is less commonly combined with experimental studies than
it might be. At the risk of being repetitive, it is difficult to study systems effec-
tively using methods that are designed for the study of individual factors and
effects. Methods such as ANOVA, MANOVA, and ANCOVA are not really
designed to look at system behavior, but instead, to summarize net effects. As
the following examples illustrate, such net effects usually hide a rich comple-
ment of individual processes that are only revealed when multiple pathways
are considered. Furthermore, a multivariate mindset automatically motivates
the researcher to incorporate a broader range of variables that control system
behavior, allowing for the effects of manipulated variables to be put into system
context.
I should emphasize here that I do not mean that we should abandon the use of
conventional univariate analysis when performing experiments. Rather, what I
recommend is that we include both kinds of analyses as complementary sources
of information. Most of the examples that follow are selected from studies in
which conventional analysis of variance methods were also applied, and it is
important to keep this in mind.
 Multivariate experiments 235

Table 9.1. Illustration of a response-surface design for an exper-

iment that involves different levels of fertilizer additions and
different frequencies of clipping. The code F0C0 refers to the
unfertilized and unclipped treatment, etc.

Fert0 Fert1 Fert2 Fert3 Fert4

Clipped0 F0C0 F1C0 F2C0 F3C0 F4C0
Clipped1 F0C1 F1C1 F2C1 F3C1 F4C1
Clipped2 F0C2 F1C2 F2C2 F3C2 F4C2
Clipped3 F0C3 F1C3 F2C3 F3C3 F4C3
Clipped4 F0C4 F1C4 F2C4 F3C4 F4C4
Clipped5 F0C5 F1C5 F2C5 F3C5 F4C5

Dealing with treatment variables in multivariate experiments

Experiments frequently involve subjecting a system to some limited set of treat-
ments. Our experiments may also involve treatments being applied to different
groupings, such as vegetation types. Thus, we can expect to encounter categor-
ical variables in many experimental analyses.
One way to deal with this issue is to avoid using categorical levels of
the manipulated variables, but instead, to use a response-surface design (see
Cottingham et al. 2005 for some of the rationale). In response-surface designs,
experimental units are subjected to a large number of treatment combinations,
often with only one replicate per combination. Such a design relies on a regres-
sion approach to analysis, and is most informative when many levels of the
experimental factors are used. An example of this approach is given in Table
9.1, for a hypothetical experiment involving different levels of fertilization and
clipping in a plant community.
In Chapter 7 we addressed a number of different ways categorical variables
can be handled. Here, some of that discussion is revisited in the context of treat-
ment variables in experimental studies. When nominal variables are encoun-
tered in an analysis, there is little choice but to treat them using dummy variable
modeling. However, when dealing with ordinal variables where the levels can
be ranked, there are more options as to how they can be viewed. Overall, the
options are (1) reliance on standard correlations/covariances, (2) adjustment
using polychoric or tetrachoric correlations/covariances, (3) performance of a
multigroup analysis, and (4) use of latent class modeling. If the reader has not
read about these topics in Chapter 7, they should now do so, to better follow
the discussion that follows.
236 Structural equation modeling and natural systems

Reliance on standard correlations/covariances

In so-called “fixed-effect” studies, the universe of interest is defined by the
treatment levels applied in the study. We might compare male and female indi-
viduals in a population, and treat sex as a fixed-effect exogenous variable in
the model. Alternatively, we might compare two burning treatments and wish
only to draw inferences for those two specific treatments. In this case, standard
Pearson correlations or covariances are sufficient to define the relationship
between treatment and continuous response variables. Furthermore, because
distributional assumptions in the analysis only relate to the residuals, the sta-
tistical distribution of a categorical predictor variable does not necessarily lead
to a violation of parametric assumptions. Thus, one approach to dealing with
a categorical treatment variable is to treat it as you would any other variable.
This has been the most common approach used in ecological studies.
It should be added here that the fixed-effect approach can be applied not
only to manipulated variables, but also to hierarchical levels in the sampling
scheme. In recent years, so called hierarchical linear modeling (Hox 2002)
has evolved to allow structural equation models to be analyzed in a multi-
level sampling framework. We might, thus, consider both within-population and
between-population multivariate effects. In cases where the investigator wishes
to account for sampling level in a single model, the choices are data adjustment
(Shipley 2000, chapter 7) or incorporation of a sampling level variable in the
model. Examples of this latter approach will be shown below.

Adjustment of correlations/covariances
In “random-effect” studies, inference is drawn to a larger universe of possi-
bilities using a limited subset of treatment levels. For example, one may wish
to consider the effect of nutrient loading on a lake ecosystem. The experi-
mental treatment levels are thus viewed as finite samples from an underlying
continuum. As discussed in Chapter 7, when standard correlations are used
to represent an underlying continuous variable, correlations are attenuated. In
practice, this means that values of the correlation fall in a narrower range of
values than the true values. Through a somewhat involved set of procedures
that define thresholds and make assumptions about the relationship between
categories and the underlying continuum, it is possible to adjust correlations
involving categorical variables. In the case of a categorical treatment variable
and continuous response variable, a polyserial correlation can be substituted
for the Pearson product-moment correlation and we can obtain an estimate of
the correlation between the response variable and the underlying continuous
treatment variable. The discrepancy between Pearson and polyserial correla-
tions is greatest when a continuous variable is measured as a dichotomous one.
 Multivariate experiments 237

As the number of categories in an ordinal categorical variable increase, the

Pearson and polychoric correlations converge. As a general rule, if an ordi-
nal variable is measured with 5–7 levels, practitioners often perform analyses
without adjustment.
As mentioned above, sometimes investigators wish only to consider the
effects of the particular treatments applied in their study. For example, an
experiment with two levels of fertilization may rely on unadjusted correla-
tions/covariances, and draw inferences only to their particular study. On the
other hand, it is common in the social sciences to treat many categorical vari-
ables as expressions of underlying continuous variables. Thus, there is some
latitude afforded to the investigator as to the use of adjustments, depending
on how one wishes to define the underlying latent variable, and the sphere of
inferences they wish to draw from the analysis.

Multigroup analyses
It is always possible to subdivide data into groups (say, treatment types) and
perform separate analyses on each group. Within SEM, this can be performed
in a comprehensive way using multigroup analysis as described in Chapter 7.
Multigroup analysis allows the investigator to determine whether parameters in
a group are constant across groups, or whether they vary. This applies not only
to path coefficients, but to all model parameters (e.g., factor loadings, latent
variable variances), and can be extended to a comparison of parameter means
across groups.
Among the many advantages of multigroup analysis is the investigation of
interactions among variables. When one wishes to study the interaction between
a treatment variable and a covariate, for example, one approach would be to
create an interaction variable and include it in the model. This can be cum-
bersome and sometimes difficult to implement. Within a multigroup analysis,
slopes of relationships are allowed to vary across groups, permitting a detailed
examination of interactions. An example of a multigroup analysis can be found
in Chapter 7.

Examples of multivariate experiments

The effects of contaminants on pond ecosystems
Background
In 1991, Johnson et al. published the results of a study in which SEM was
used to examine the effects of atrazine additions on pond mesocosms. A pri-
mary motivation in this analysis was to determine the indirect as well as direct
238 Structural equation modeling and natural systems

consequences of adding this potential toxicant to freshwater ecosystems. This

study involved a 3-year investigation of 450 000-liter mesocosms (experimen-
tal ponds). Mesocosms were exposed to 0, 20, 100, or 500 µg/L of atrazine,
a widely used herbicide. Mesocosms were periodically dosed to maintain the
intended concentration level, with ecosystem data collected periodically. The
data used in this analysis were from the third year of the study, during which
variables were measured eight times from May through October. Data from the
eight mesocosms were combined across the eight time periods to provide 64
measurements. These data were treated as independent estimates, which may
not be justified. Nevertheless, this example provides an interesting illustration
of the kind of insight that can be gained using SEM to analyze experimental
data on ecosystems.
A variety of ecosystem properties were measured in this study, in addition to
the atrazine levels, including phytoplankton concentration, zooplankton abun-
dance, aquatic vegetation, and the presence or absence of grass carp in the
mesocosm. Sampling of the planktonic components of the system, which were
the response variables of primary interest, was performed at a variety of depths
in each mesocosm, and the variable “pond depth” was included as an additional
factor that might affect other parts of the system.

Results
The analyses of these data using SEM were accompanied by a detailed discus-
sion of the process of model development and refinement. The reader is encour-
aged to read the original paper (Johnson et al. 1991) to see the steps the authors
took in reaching the final model. Here I present a figure showing the final
results (Figure 9.1). One major focus of this study was to determine the effects
of atrazine on aquatic vegetation, phytoplankton and zooplankton. As can be
seen by the number of significant relationships among the variables examined in
this system, there was a complex interplay of parts that influenced how response
variables were affected by atrazine. These interactions included the following:

(1) Grass carp played an important role in these mesocosms by reducing the
amount of aquatic vegetation. Through this mechanism, grass carp had an
indirect positive effect on both phytoplankton and zooplankton by reducing
aquatic vegetation; which had a very strong negative effect on phytoplank-
ton.
(2) The depth of the pond also had important effects, both on aquatic vegetation
and on phytoplankton. In general, aquatic vegetation and phytoplankton
were reduced in deeper ponds. Since aquatic vegetation had a negative effect
on phytoplankton, we can see that the effect of pond depth on phytoplankton
 Multivariate experiments 239

grass
atrazine pond
carp
concent. depth
presence
0.43 2.19 0.34

GRASS POND
ATRAZINE
ATRA −0.21
CARP DEPTH
−0.71
−0.82
−0.38 0.88 −0.50

AQUATIC −1.08 PHYTO-

VEG. PLANKTON
0.41
−0.25
emergent 0.30 0.50 0.64
vegetation ZOO-
submersed PLANKTON phytopl.
vegetation abund.
1.22

zooplank.
abund.

Figure 9.1. Final model from analyses by Johnson et al. (1991) on the effects
of atrazine on aquatic mesocosms. Loadings from latent variables (in circles) to
measured variables (in boxes) are influenced by the effects of standardization
on units using a prior version of LISREL software which did not scale latent
and indicator variables equally. Thus, loadings cannot be interpreted as if they
are standardized values, although path coefficients (directional arrows) can be.
Reproduced by permission of the Ecological Society of America.

consisted of a direct negative effect (likely to be caused by reduced light

penetration at greater depths) and a positive indirect effect (through the
adverse effects of depth on macrophyte growth). The total effect of depth
can be seen to be −0.50 + (−0.21 × −1.08) = −0.27.
(3) Atrazine had strong negative effects on both aquatic vegetation and phyto-
plankton (being a herbicide, this is to be expected). However, the net effect
on phytoplankton in this system was (−0.82 × −1.08) – 0.71 = 0.18,
which is a positive relationship that is nonsignificant at the sample size of
this study. This offsetting pair of direct and indirect effects means that the
net effects of atrazine on phytoplankton will vary widely depending on the
abundance of aquatic vegetation in the system studied.
(4) In order to understand the responses of zooplankton and phytoplankton
in this system it was necessary to model their relationship as a reciprocal
240 Structural equation modeling and natural systems

one, with arrows going in both directions. This is basically a predator–

prey system in which phytoplankton enhance zooplankton and zooplankton
reduce phytoplankton. Zooplankton also responded to pond depth, perhaps
because there was a greater refuge from predation from other predators,
including other fish species not modeled here, in deeper ponds. The total
correlation between phytoplankton and zooplankton is partitioned in quite
a complex way in this model and consists of several pathways.
(a) phytoplankton → zooplankton
(b) zooplankton → phytoplankton
(c) phytoplankton → zooplankton → phytoplankton
(d) zooplankton → phytoplankton → zooplankton
(e) phytoplankton ← pond depth → zooplankton.
As Johnson et al. (1991) point out, the reciprocal (nonrecursive in the
language of SEM) relationship means that phytoplankton affect themselves
indirectly and zooplankton do likewise. As Hayduk (1987) illustrates in
detail, this “effect on self” involves a loop in which a unit change, e.g., in
phytoplankton, causes a change in zooplankton, which in turn causes an
additional change in phytoplankton, and so on. The total effect of these loop
interactions on the relationship between phytoplankton and zooplankton is
composed of the basic effect (which is the usual direct effect) and the loop-
enhanced effect, the effect through the loop. The indirect effect through the
loop is 1/(1 − L), where L = β pz × β zp . The total effect through a loop is
the product of the direct and indirect effects (as opposed to the sum, which
applies for non-loop relationships).

Conclusions
This example illustrates that there is substantial value in studying a system
response to manipulation, in order to get a more realistic and predictive under-
standing of the potential effects of contaminants. Imagine, for a moment, that
the web of relationships studied in this case was ignored and that only the
response of phytoplankton to atrazine was examined. Depending on the sample
size and statistical power, we would find for a case like the one studied by
Johnson et al. either a nonsignificant or weak positive effect (despite the fact
that the herbicide atrazine is known to have a negative effect on algae in isola-
tion!). How would we interpret such a result? We might also imagine that when
the effects of atrazine are studied in ponds, mesocosms, or whole lakes, the
responses of phytoplankton vary from strong negative effects to clear positive
effects, depending on the abundance and role of macrophytes in these stud-
ies. How would our chances of publication and convincing interpretation be
 Multivariate experiments 241

affected by a widely varying response, as might be expected for this case? The
example by Johnson et al. (1991) clearly illustrates many of the advantages of
using SEM in the study of ecosystem responses.

Response of a prairie plant community to burning and fertilization

Background
The “coastal prairie” ecosystem in North America is the southern-most lobe
of the tallgrass prairie biome, and stretches along the western Gulf of Mexico
coastal plain from central Louisiana to south Texas. Approximately 99% of this
endangered ecosystem has been radically altered or badly degraded by agri-
culture, industry, or urban development. For conservationists, a high priority
is currently placed on the protection and proper management of the remaining
native fragments. Our next example deals with data from a study of the effects
of prescribed burning and soil fertility on the vegetative community on one
such native fragment (Grace et al., unpublished). In order to keep things sim-
ple, this presentation will focus on the joint responses of plant species richness
and community biomass to experimental treatments. In this case, the multivari-
ate model for analysis was expanded because of effects of spatial gradients,
temporal variation, and experimental manipulations that were not anticipated
when the study was designed.
The site involved in this study has been managed by annual mowing for the
past few decades. In this experiment, 10 × 10 m plots were established and ran-
domly subjected to combinations of burning and fertilization treatments using a
random block design (note block effect was not found to be important and was
removed from the analyses). Burning treatments included being (1) unburned,
(2) annually burned, and (3) once burned. Fertilization treatments were (1) with,
or (2) without annual applications of NPK fertilizer. A pretreatment examina-
tion of vegetation was conducted in early spring of 2000, prior to treatment
applications. Vegetation, including species richness and above-ground biomass
(live and dead), was measured at the end of the growing season in 2000, 2001,
and 2002.

Results
Repeated measures analysis of covariance was used to examine mean responses
of richness and biomass to experimental treatments (Figure 9.2). For richness,
results showed significant effects of time, row (as a spatial covariate), and the
fertilizer treatment. For biomass, significant effects of time and its interactions
with burning and fertilizer treatments were found.
242 Structural equation modeling and natural systems

45

Richness per plot

40
35 * U-U
30 U-F
25 A-U
20 * * A-F
15 O-U
10 O-F
5
0
2000 2001 2002
Year
2500
Biomass, g/m2

2000 U-U
U-F
1500 ** ** A-U
1000
* A-F
* * O-U
500
+
O-F

0
2000 2001 2002
Year
Figure 9.2. Responses of means for richness and biomass to experimental treat-
ments over three years of study. U–U = unburned + unfertilized, U–F = unburned
+ fertilized, A–U = annually burned + unfertilized, A–F = annually burned
and fertilized, O–U = once burned + unfertilized, O–F = once burned + fertil-
ized. Asterisks represent means that differ significantly from the U–U treatment at
p < 0.05, while the “+” symbol indicates differences at the p < 0.01 level.

In order to better understand how richness and biomass are jointly regulated
in this system, a multivariate analysis was conducted to evaluate the general
model shown in Figure 9.3. This model supposes that richness and biomass
might influence each other, and these interactions could contribute to responses
of both to treatments. Reciprocal arrows between richness and biomass were
meant to include both negative effects of biomass on richness through competi-
tive effects (which has been documented for this site by Jutila and Grace 2002),
as well as postulated positive effects of richness on biomass.
The first step in evaluating the model in Figure 9.3 was to determine the
shape of the relationship between biomass and richness. A positive relationship
between these two variables would indicate the possibility of an effect from
richness to biomass; a negative relationship would indicate the possibility of a
 Multivariate experiments 243

UU rich
UF
AU trt
AF
OU
OF biom

Figure 9.3. General multivariate model proposed to evaluate how interactions

between richness and biomass might influence responses to treatments. Hexagon
represents composite variable.

50 r = 0.47
Richness/plot

40

30

20

10
0 500 1000 1500 2000 2500
Biomass, g/m2
Figure 9.4. Observed relationship between biomass and richness in study of
coastal prairie. Sample size in this study is 72 and includes values pooled across
treatments and times.

negative effect of biomass on richness; and a hump-shaped (unimodal) relation-

ship would indicate the possibility of both types of effect. Examination of the
relationship between biomass and richness revealed a negative linear correla-
tion (Figure 9.4). Therefore, the path from richness to biomass was eliminated
for theoretical reasons, and the model shown in Figure 9.5 was used to represent
the most plausible causal order of relationships. Analysis was performed using
a composite variable (Chapter 6) to present a consolidated representation of
treatment effects.
The effects of individual treatments are omitted from this presentation, for
simplicity. The results in Figure 9.5 suggest (1) that biomass responded more
strongly to treatment manipulations than did richness, and (2) treatments had
244 Structural equation modeling and natural systems

UU 0.28 rich R2 = 0.304

UF
AU
trt −0.50
AF
OU
OF −0.54 biom R2 = 0.295

Figure 9.5. Results from analysis of initial model (Figure 9.3). Path from rich to
biom was not theoretically justified in this analysis, because the observed correla-
tion between these two variables (Figure 9.4) was negative and linear.

both direct effects on richness as well as indirect effects through impacts on

biomass.
We know that the model in Figure 9.5 is an inappropriate oversimplifica-
tion of the factors affecting richness. The impact of variations among years is
omitted, as is the effect of spatial variation in site conditions. As to this latter
point, examination of the data, as well as general field observations, indicated
the presence of a subtle gradient in site wetness running along the long axis of
the experimental plot array. To evaluate the role of spatial and temporal vari-
ations in wetness, indices of the known affinities of these species for wetland
conditions were used to calculate a wetness index for each plot. Subsequently,
the model in Figure 9.6 was evaluated, with results as shown below.
The model in Figure 9.6, with accompanying results, supports several impor-
tant conclusions:

(1) The correlation between biomass and richness observed in Figure 9.4 does
not represent a causal interaction between these two variables, but instead,
joint control by the other variables.
(2) Row effects on richness can be explained entirely by variations in wetness
across rows.
(3) Year effects were of three sorts, one being an increase in biomass over time,
a second being an increase in community wetness during the experiment,
and a third being a decrease in richness independent of wetness or biomass.
(4) Upon considering the year effects, the authors concluded that the path
from year to rich most likely represented a system-wide decline in richness
resulting from a cessation of mowing during the course of the study. This
effect was further deemed to be a form of “experimental drift” that was an
unintended consequence of the experiment.
 Multivariate experiments 245

R 2 = 0.55
0.35
row wet −0.36

UU ns rich R 2 = 0.77
UF 0.65
AU trt −0.53 ns
AF ns
−0.54
OU
OF biom R 2 = 0.63

year 0.58

Y1 Y2 Y3
Figure 9.6. Model and results when the effects of year, spatial variation (the row
variable), and vegetation wetness tolerances were incorporated.

To provide a result of more general utility (because it is not contaminated

with treatment artifacts), the effects of the path from year to rich, which is
presented to represent experimental drift, were removed from the data and the
model rerun. By calculating the residual differences between richness and the
predicted scores using the individual path coefficient, it was possible to remove
this one component of the year effect. The remaining effect of year on richness
mediated through wetness was retained.
The results of this modified model are shown in Figure 9.7. It can be seen that
there are only two changes of significance from Figure 9.6. First, the amount
of variance in richness is reduced and the percent explained drops from 77%
to 43%. Secondly, the path from wet to rich increases in absolute strength
from −0.36 to −0.66.

Conclusions
Conventional univariate analyses, such as the repeated measures ANCOVA are
often important in a complete evaluation of experimental results. Such analy-
ses permit a detailed examination of net treatment effects on mean responses.
However, such univariate analyses are unable to evaluate the multivariate rela-
tionships among variables. Thus, a combination of univariate and multivariate
techniques can be most effective.
246 Structural equation modeling and natural systems

R 2 = 0.55
0.35 −0.66
row wet
UU rich R 2 = 0.43
ns
UF 0.65
AU
trt ns
AF ns
−0.54
OU
OF R 2 = 0.63
biom
year 0.58

Y1 Y2 Y3
Figure 9.7. Model with effects of “experimental drift” (the path from year to rich)
removed.

There has been and remains a great interest in the relationship between
biomass and richness. It is instructive in this case to see how a multivariate
analysis sheds light on the interpretation of this relationship. The observed
correlation between richness and biomass found in this study (Figure 9.4) is
suggestive of a competitive effect on richness. This correlation is supportive of
an interpretation in which treatment, row, and year effects might influence rich-
ness through their impacts on biomass. However, when a multivariate analysis
is performed, we can see that the correlation between richness and biomass,
though it is fairly strong, results from common control rather than biological
interactions. Both richness and biomass are under strong control from com-
mon causes and as emphasized in the previous chapter, relationships between
variables possessing common causes cannot be accurately evaluated using a
univariate or bivariate approach.
This example also illustrates another important capability of a multivariate
experiment. Univariate results showed that vegetation wetness increased over
time while biomass increased and richness decreased. On the face of it, it might
seem that the decrease in richness over time resulted from increased rainfall and
increased competitive exclusion. Multivariate analysis isolated a direct effect of
year on rich that was interpreted as an undesired side effect of the experimental
treatments due to the cessation of mowing during the study. It was then possible
to remove this side effect (dubbed by the authors “experimental drift”) from the
data while retaining the other components of the year effect.
 Multivariate experiments 247

Manipulation of an intertidal marine community

Background
In 1994, Wootton published the results of an intensive analysis of a marine
intertidal community using a combination of manipulative experiments and
path analysis. Marine intertidal communities have been shown through previous
experiments to be highly interactive, both within and between trophic levels.
Wootton made several arguments in favor of multivariate experimentation. First,
he argued that the possibility of indirect effects makes the analysis of species
pairs an inadequate tool for predicting community responses to environmental
change. Secondly, he argued that interactions were typically quite asymmetric,
with some species and certain interactions being more important to overall
community dynamics. Thus, he built the case for the need to determine the
relative importance of pathways in the system. A third motivation he presented
was that a reductionist approach that tried to understand community interactions
by assembly of all the pairwise interactions was hopelessly complex for all
but the most simple systems. Finally, he noted that manipulation of all the
elements in a community was, at times, not possible due to logistic or other (e.g.,
endangered species) reasons. As Wootton clearly saw, the combination of SEM
and experimentation has the potential of overcoming a variety of difficulties in
studying interacting systems. These issues were reiterated in a later review of
the widespread importance of indirect effects in ecological systems (Wootton
2002).
Wootton (1994) began his study with an initial experiment that excluded
birds, which serve as top predators, from a middle intertidal community on
an island off the coast of Washington state in North America. Five 1.5-year-
old gaps in the mussel bed were selected for study and pairs of plots were
established in each, with one of each pair randomly assigned to be caged or
not. Snails (Nucella), small starfish (Leptasterias), mussels (Mytilus), goose
barnacles (Pollicipes), and acorn barnacles (Semibalanus) were measured in
each plot over two years. Comparisons of treatment means found that when
birds were excluded (1) snails, mussels, and acorn barnacles were reduced, and
(2) goose barnacles were enhanced.
Wootton recognized that at least three different mechanistic scenarios were
possible given the pattern of results. To evaluate these possibilities, he pooled
data across treatments and years, and performed a path analysis that compared
the fit of the data to five contrasting models. Predicted and observed correlations
were compared using linear regression to assess model fit. Chi-square tests were
also performed. As Wootton himself noted, “. . . the conclusions [from this
initial analysis] should be treated as predictions that point to the most important
248 Structural equation modeling and natural systems

Table 9.2. Correlations and standard deviations from initial bird exclusion
experiment conducted by Wootton (1994). GooseB refers to goose barnacles.
AcornB refers to acorn barnacles

Birds GooseB Mussels AcornB Snails ln Tide Ht.

Birds 1.0
GooseB −0.955 1.0
Mussels 0.399 −0.468 1.0
AcornB 0.802 −0.809 −0.103 1.0
Snails ln 0.403 −0.349 −0.303 0.532 1.0
Tide Ht. 0 0.039 −0.444 0.335 −0.210 1.0
Std. dev. 0.47 26.76 16.97 15.48 0.90 0.16

experiments to be conducted next, not as conclusions to be set in stone.” This

is true not only because of the exploratory (though theory-driven) nature of the
analysis, but also because of the low sample size and the fact that data from the
two years were pooled and thus treated as independent observations.

Results
The results from the initial bird exclusion experiment are presented in Table
9.2, and the model selected to best represent the data is shown in Figure 9.8.
Examination of the correlations indicate that exposure to bird foraging leads to
a strong reduction in goose barnacles and increases in acorn barnacles, snails,
and mussels. These interpretations from the correlation matrix are matched
with results from the analysis of mean responses as well. The correlation matrix
further indicates that both goose barnacles and mussels are negatively associated
with all community members, while acorn barnacles and snails are positively
associated with each other.
Comparing the correlation matrix to the path model results indicates the
following:

(1) The effects of birds on snails, acorn barnacles, and mussels are all indi-
rect, being mediated by direct effects on goose barnacles. Thus, birds are
primarily feeding on goose barnacles.
(2) Goose barnacles have strong negative effects on acorn barnacles that are
only partially offset by indirect beneficial effects caused by reductions in
mussels.
 Multivariate experiments 249

Birds −0.96
R 2 = 0.91

GooseB 0.92 R 2 = 0.65

−0.45
−1.1 Snails
ns
−0.54
−0.43 Mussels AcornB 1.5
R2 = 0.40 0.14 R 2 = 0.97
−0.76
Tide Ht

Figure 9.8. Model selected to represent initial bird exclusion experiment in

Wootton (1994). Path coefficients presented are standardized. GooseB refers to
goose barnacles. AcornB refers to acorn barnacles. Reproduced by permission of
the Ecological Society of America.

(3) Acorn barnacles have a strong positive effect on snails, and there is no
indication that there is a reciprocal effect of snails on acorn barnacles.
(4) Goose barnacles have a complex effect on snails. This effect has positive
direct and indirect components that are offset by a negative effect on acorn
barnacles, which stimulate snails.
(5) Tide height plays an important part in this system. Its actual influence is
poorly represented by the correlation results because of the importance of
the indirect effects that tide height has on acorn barnacles and snails.

Conclusions
It can be seen from this example, as with previous examples, that a multivariate
examination of the data yields a very different interpretation of the underlying
mechanisms. In this case, the author possessed several plausible alternative
interpretations about underlying causal mechanisms. However, two of the three
models evaluated were found not to fit the covariance patterns in the data. Thus,
SEM has the ability to provide guidance as to which proposed mechanisms
do not appear consistent with the results. In this study, Wootton went on to
make predictions from this initial experiment that were subsequently tested
with further experiments. Such experimental tests are considered in the final
chapter relating SEM to ecological forecasting.
250 Structural equation modeling and natural systems

The effects of fire on an exotic tree that invades coastal

tallgrass prairie
Background
Ecosystem manipulations, such as those involving prescribed fire, often pose
special challenges for experimental study. This example comes from a study of
the potential for prescribed burning to control an invading exotic tree, Chinese
tallow (Triadeca sebifera) (Grace et al., unpublished). The system being invaded
is coastal tallgrass prairie, one of the most endangered ecosystems in North
America. Chinese tallow poses a special threat because of its ability to turn a
diverse and fire-dependent native grassland into a nonflammable monoculture
thicket in only a few years. Because the coastal prairie is typically managed
using prescribed fire to mimic the natural fire regime, the question of interest
is whether fire can control this exotic tree in the same way that fire naturally
keeps the prairie from succeeding to woodland.
A few characteristics of the biological interactions in this system are of
prime importance. First, as with most native grasslands, the coastal tallgrass
prairie is a fire-dependent ecosystem. Secondly, Chinese tallow is a fire sup-
pressing species. It both suppresses fuel species and itself rarely carries a fire.
Thirdly, it is known that once Chinese tallow establishes to a certain point, it
becomes immune to fire management. The objective of this study was to deter-
mine whether there exists some minimum tree size below which fire control
is effective. If such a critical minimum size exists and can be identified, it can
be used to develop protective habitat management plans that are effective and
efficient. Thus, of prime interest was the relationship of tree size to its response
to burning.
This study was conducted at the Brazoria National Wildlife Refuge in coastal
Texas which is one of the most important pieces of coastal prairie under federal
protection. Approximately 5000 hectares of virgin prairie remain at this site. In
addition, thousands of acres of abandoned agricultural lands are in the process
of recovering prairie vegetation. As with many habitats of conservation impor-
tance, conditions were not favorable at this site for generating a large number
of replicate habitat units for experimental manipulation. Several factors had to
be considered. First, it was important that the destructive process of creating
fire breaks in this system be kept to a minimum. Secondly, numerous gradients
and historical differences contribute to a mosaic landscape in which pre-existing
management units (which can be burned independently) each have unique prop-
erties. Thirdly, the trees themselves contribute to the development of a mosaic
in fuel and other conditions, creating a heterogeneous fire that can burn one
tree and completely miss the tree next to it. Finally, stands of tallow contain a
 Multivariate experiments 251

season
burned

soil

study
fuel fate ht_chg
area

elev

damage
init_ht

Figure 9.9. Initial model of how height and other factors might influence the
effects of burning on the Chinese tallow tree.

continuum of sizes from seedling to reproducing adult. While seedlings often

establish beneath adult trees, they are dispersed by birds and also establish in
the open or in gaps in the vegetation caused by animal activities.
In this experiment, 20 trees in each of five size classes (0–0.1 m, 0.1–1 m,
1–2 m, 2–3 m, and >3 m) were randomly selected in each of four habitat
management units, for a total of 400 trees. Some seedlings did not survive to
the time of initial burning and tags were lost from a few other trees, resulting in a
total sample size of 378. For each tree, a 1 meter diameter plot was established
with the tree in the center. Variables measured before burning included tree
height, basal diameter, estimated fuel load, elevation, and soil texture. One
management unit at each of the two areas was selected for summer burning and
the other selected for winter burning. Following each burn, the percentage of
the plot that burned was recorded, along with the degree of damage the tree
incurred (specifically, how high the tree was scorched and what fraction of the
tree was damaged), and whether it ultimately died, retained its original height,
or was topkilled. The heights of all trees were tracked over a two-year period,
with the end of the second growing season designated for the final comparative
measurements.
The model proposed initially, shown in Figure 9.9, incorporates both con-
trolled variables, such as (1) initial tree height, (2) area, and (3) season of burn,
along with uncontrolled variables, such as (4) location along elevation gradi-
ents, (5) soil conditions, and (6) fuel under or around each tree. The response
variables measured included (7) the percentage of a plot that burned, (8) the
252 Structural equation modeling and natural systems

Table 9.3. Correlations among variables remaining in model of fire effects on

Chinese tallow. Htchg refers to tree height change during the study. Initht refers
to the initial height of the tree prior to burning. Numbers in bold are significant
at p < 0.05

Fuel Burn Damage Fate Htchg Season Area Initht

Fuel 1.0
Burn 0.21 1.0
Damage 0.20 0.88 1.0
Fate −0.07 −0.42 −0.57 1.0
Htchg −0.18 −0.32 −0.45 0.60 1.0
Season −0.12 −0.39 −0.27 0.01 −0.25 1.0
Area 0.16 0.59 0.54 −0.39 −0.26 0.0 1.0
Initht −0.26 −0.03 −0.21 0.48 0.38 0.0 0.0 1.0

percentage of a tree damaged, (9) its fate (whether it died, was topkilled, or
survived intact), and (10) the height change of the tree during the study. Height
change basically quantified the degree to which resprouting might have allowed
the tree to recover from the burn. The logic of this initial model was that some
factors of importance, such as soil conditions and fuel would potentially depend
on study area and elevation gradients within sites. It was also hypothesized that
the amount of fuel beneath trees would be affected by tree size.
The amount of fuel was expected to influence both the completeness of burn,
as well as the amount of damage that would be done to trees. However, fires are
also affected by immediate weather and other influences, such as soil moisture,
that vary from site to site and time to time. In this model, such generalized
effects were confounded with study area, which was postulated to include such
effects.
Finally, the fate of a tree and ultimately its change in height during the study
were expected to depend on how complete the burn was around a tree, and how
much damage it sustained. In woodland burns, heterogeneity in fire behavior
can be substantial at the level of the individual tree, and obviously only trees that
actually sustained significant fire damage are expected to respond to burning.

Results
Initial model results found that both soil conditions and elevation influenced
fuel, but did not affect plant responses in any major way. For this reason, it
was possible to simplify the model by removing these two factors. Correlations
among the remaining factors are shown in Table 9.3 and the final model is
shown in Figure 9.10.
 Multivariate experiments 253

−0.30
season
−0.11 −0.38 burned
R 2 = 0.51
0.46
0.08 −0.12
0.86

−0.17
0.12
0.17
study 0.11 fuel fate 0.56 ht_chg
area R2= 0.11 0.12 R 2 = 0.55 R 2 = 0.54
0.10

0.43
−0.39
−0.26 −0.26
damage
−0.21 R 2 = 0.82
init_h
init_ht
0.05

Figure 9.10. Final model for fire effects on Chinese tallow tree.

Figure 9.10 reveals 20 significant direct relationships among variables, sug-

gesting that the factors affecting tree responses to fire are quite complex. To a
certain degree, this complexity of outcome results from the large sample size
associated with this study. We can see that paths whose standardized coefficients
were as low as 0.05 were deemed statistically significant in this analysis. Eight
paths had coefficients less than 0.15. The actual complexity of this model was
even greater than apparent in Figure 9.10, because the relationships between
initial tree height and fuel, fate, and height change were nonlinear, each hav-
ing a unique shape. Appendix I at the end of this book presents the program
statements for the full model, as well as an explanation for the procedures for
estimating the strength of the nonlinear paths.
Several main points can be drawn from the results of this study.

(1) Initial tree height was well correlated with the responses of trees to burning.
While a number of processes appear to be involved, the biggest fraction of
this relationship was through a direct effect on the fate of the tree that
was unrelated to fuel suppression, or to the fact that larger trees suffered
proportionally less damage.
(2) The study areas also had a substantial impact on the results. This effect
could result from a number of factors relating to the nature of the fires,
the overall fuel conditions, or the conditions affecting tree growth. The
authors concluded that the study area effect was most likely to be related
254 Structural equation modeling and natural systems

to differences in fuel conditions (both general and at the time of the burn),
which had strong effects on the completeness and intensity of the burns. It
would seem that in this study, the fuel load immediately around the tree did
not represent the general effects of fuel on fire effects, although there were
some modest effects detected.
(3) The season of burn was related to several pathways in the model. Some of
these appear to result from a confoundment of factors caused by the small
number of burn units. Effects of season on initial fuel conditions and the
percentage of plots that burned were deemed coincidental. The main effect
of season that was important in this study was the direct effect on height
change. This effect represents the growth responses of resprouts to season.
It was apparent that trees burned during the dormant season produced
resprouts that fared much better than those produced by trees burned in
summer. There was a similar but weaker effect of season on tree fate.

Conclusions
This example represents a case where an analysis of multivariate relationships
among trees seeks to overcome the limits of field experimentation. Since the
ultimate objective of the research was to develop a predictive model that can be
tested against repeated trials, the limits of the results will ultimately be judged
by the success of prediction over a range of conditions. Because of the goal of
predicting effects in other locations, this experiment contributes to a program
of study that seeks to be generally consistent over space and time, rather than
intensively precise for a particular space and time.

Linking experimental and nonexperimental studies of coastal

plant diversity
While this chapter emphasizes the value of an experimental approach, ecol-
ogists, perhaps more than many other natural scientists, are keenly aware of
the limits imposed by experimentation. Typically, we do not strive simply to
understand the pieces of natural systems that fit within the confines of aquaria,
incubators, test tubes, greenhouses, mesocosms, or even individual study sites.
Rather, we wish to know how whole ecological systems function, and how
they will respond to perturbations and environmental changes. Nonexperimen-
tal hypothesis testing, a strong suit of the SEM tradition, can play an invaluable
role in allowing the researcher a more extensive view of how a system works.
Our interpretations from nonexperimentally evaluated models are neverthe-
less limited in the strength of their inference in many cases (see Chapter 4).
Experimentation plays a vital role in evaluating our assumptions and
 Multivariate experiments 255

Table 9.4. Correlations. Those in bold are significantly different from zero
at the p < 0.05 level for a sample size of 254

Richness Fenced Fertilized Abiotic Disturb. Biomass

Richness 1.0
Fenced 0.1076 1.0
Fertilized −0.2683 0.0041 1.0
Abiotic 0.0249 −0.0180 0.0014 1.0
Disturb. −0.4400 −0.3114 −0.1341 0.1730 1.0
Biomass 0.5343 0.2119 0.3172 −0.2401 −0.5525 1.0

hypotheses, as well as in determining the limits of our causal interpretations. For

this next example, we return to an experimental study conducted by Gough and
Grace (1999). In this study, several variables were experimentally manipulated
in order to test the assumptions of prior nonexperimental SEM.

Background
In part because of a desire to test experimentally assumptions about the relation-
ships among environmental and community properties, a study was undertaken
to manipulate as many of the factors thought to be important in controlling
species richness as feasible. This study was conducted at the Pearl River (see
Chapter 10 for a more complete description of this system) in oligohaline and
mesohaline marshes. The manipulations included (1) the addition of NPK fer-
tilizer to alter community biomass, (2) changing the levels of flooding stress by
raising or lowering sods of plants, (3) exposing sods to different salinity regimes
by transplanting them between oligohaline and mesohaline locations, and (4)
the use of fences to exclude herbivores and thereby reduce disturbance levels.
Treatments were applied in a full factorial study. While species composition
was measured in this study, the focus of the model was on species richness
as the primary response variable of interest. Eight replicates of each of the 32
treatment combinations were used in this study, which included a total of 256
experimental units. Problems resulted in the loss of data for two experimental
units, leaving 254 samples in the analysis.

Results
Table 9.4 gives correlations for the variables analyzed by Gough and Grace
(1999). Prior to analysis using LISREL, compositing techniques were used to
provide estimates of total abiotic effects, disturbance, and biomass for subse-
quent analysis. In the preanalysis, first- and second-order polynomial effects
256 Structural equation modeling and natural systems

salt
FENCE FERT

flood ABIOTIC 0.22 −0.19 −0.50

0.25

soilcarb −0.16

−0.24 RICH 1.0

0.17 −0.13 R 2 = 0.60
rich

−0.31

0.65(+/-)
1.0 DIST BIOM
dist
R 2 = 0.11 R2 = 0.39
−0.49

mass masssr

Figure 9.11. SEM results for experimental studies of the factors affecting species
richness in coastal wetland plant communities (from Gough and Grace 1999).
Reproduced by permission of the Ecological Society of America.

of salinity, flooding, and soil organic variations were formative indicators for
the statistical composite variable “abiotic”. First- and second-order polynomial
terms were also incorporated into both “disturbance” and “biomass” in order
to achieve linear relations among variables. In the process of performing the
preanalysis, all variables were z-transformed to means of zero and variances
of 1.0.
Comparison of the SEM results to those apparent from only the raw cor-
relations is revealing. To reduce complexity, only the effects of fencing and
fertilization on richness will be emphasized here. Additional interpretations
can be found in Gough and Grace (1999).
Variations in species richness in this study were unrelated to whether or not
the plots were fenced (r = 0.108), even though fencing reduced disturbance
levels and disturbance, in turn, was related to richness. As the SEM results
show (Figure 9.11), the experimental manipulation of fencing actually had
three significant effects.

(1) One effect of fencing was an indirect enhancement of richness caused by a

reduction in direct disturbance effects. This is represented by the path from
Fence to Disturbance to Richness. We interpreted this indirect effect as a
suppression of the direct loss of disturbance-sensitive species.
 Multivariate experiments 257

(2) The other indirect enhancement was through a reduction in disturbance,

subsequently leading to a recovery in biomass and an associated recovery
in richness (the path from Fence to Disturbance to Biomass to Richness).
(3) The third significant effect of fencing observed was a direct negative effect
of fencing on richness. This path appears to represent increased competitive
exclusion as litter built up within the fenced areas. The reason the loss of
species through this path was not related to biomass accumulation appears to
be because the accumulation of litter suppressed both richness and biomass
accumulation.

Conclusions
The value of this multivariate experimental approach and analysis should be
clear. A univariate experiment would only have revealed that fencing had no net
effect on richness. What would have been hidden from view were the effects
of three significant processes, two positive and one negative, which offset one
another.
As for the effects of fertilization, behind the net negative effect on richness
there appear to lie four significant processes.

(1) Two rather weak effects operated through a reduction in disturbance asso-
ciated with fertilization. We hypothesize that, in this case, the more rapid
growth rate of plants in fertilized plots showed a more rapid recovery from
disturbance and, thus, evidence of disturbance was slightly lowered by fertil-
ization. As described above for the effects of fencing, reducing disturbance
can promote richness through two somewhat different paths.
(2) The third indirect effect of fertilization was through an enhancement of
biomass and an associated enhancement of richness, presumably primarily
in plots recovering from disturbance.
(3) The fourth process, implied by Figure 9.11, is a direct negative effect of
fertilization on richness. This strong negative effect occurred independent
of increases in biomass, as was also seen for the fencing treatment. Analysis
of the timecourse of events in this study would indicate that fertilization led
to an early increase in biomass that ultimately inhibited later plant growth.
Ultimately, much of the loss of species associated with fertilization was not
accompanied by an increased biomass at the end of the study.

The story for fertilization would seem to be similar to that for fencing in that
the net effect only tells a small part of the total story. The actual direct negative
effect, independent of the other variables, is actually quite a bit stronger (path
coefficient = −0.50) than the net negative effect (r = −0.27) representing all
258 Structural equation modeling and natural systems

the processes. It is possible that our interpretations of all the paths are not
complete. However, the multivariate analysis clearly indicates that fertilization
is influencing richness through a variety of different processes, and it gives an
estimate of the relative importance of direct and indirect influences. It is worth
pointing out that the rather large number of total experimental units in this study
was important in revealing many of the indirect effects that were observed.

Conclusions
It is hopefully clear through the examples presented in this chapter that exper-
imental studies of systems can and often should be conducted using SEM. The
basic test an investigator can apply to determine their needs is to ask themselves
the following question: am I interested in how the various parts of the system
interact to produce the results? If the answer is yes, conventional univariate
methods will not be adequate, although they can complement the analyses.
Once an investigator decides that they wish to use SEM in an experimental
situation, they are automatically encouraged to include the important unmanip-
ulated factors into their models. This moves the experiment from being simply
an investigation of a single process, to the study of how their system is regu-
lated. I expect that often ecologists will find that the process of initial interest
to them is not really the one most important in regulating their system, at least
that has been my experience.
Society asks us to provide answers that are relevant to real world concerns
and that can be applied to conservation solutions. Understanding the effects
of individual processes will not, in my opinion, generate the information we
need. The conduct of multivariate experiments is a largely neglected topic
that I believe will become the standard for understanding ecological systems.
There exist great opportunities for developing methods and applications for
experimentally evaluating multivariate hypotheses using SEM.
 10
The systematic use of SEM: an example

This chapter illustrates the systematic application of a multivariate perspec-

tive using SEM to explore a topic. In this presentation, the statistical details
of the analyses will be ignored; these have been presented in earlier chapters,
or can be found in the various publications referenced throughout. Here, the
emphasis is on illustrating the broad enterprise of developing, evaluating, refin-
ing, and expanding multivariate models in order to understand system behavior
and regulation. Throughout, the focus will be on the research enterprise rather
than the analytical details. Thus, the philosophy and practice of SEM will be
in the forefront, while the analysis of covariances, maximum likelihood, and
mathematical details will be de-emphasized.

Background studies and findings

In 1992, Laura Gough and I conducted a study designed to examine the rela-
tionship between plant community biomass and species richness. This work
was conducted in coastal marsh communities. The purpose of this study was
to first characterize the relationship between biomass and richness. Then we
planned to determine the role of competition in controlling the relationship.
We expected that we would find a unimodal relationship between biomass and
richness, primarily because of several key papers that had been published pre-
viously (Al-Mufti et al. 1977, Huston 1980, Wheeler and Giller 1982, Moore
and Keddy 1989, Wisheu and Keddy 1989, Shipley et al. 1991, Wheeler and
Shaw 1991). We also expected this relationship because there were several com-
peting theories attempting to explain this phenomenon (Grime 1979, Huston
1979, Tilman 1982, Taylor et al. 1990, Keddy 1990). Figure 10.1 presents
one of the early examples that inspired much of the subsequent work on this
topic. A further influence on us at that time was work by Paul Keddy and his

259
260 Structural equation modeling and natural systems

30

20
SD

10

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 2600
Maximum standing crop + litter (gm−2)
Figure 10.1. Example of the relationship between biomass (measured as the
maximum standing crop including litter), and species richness (SD, or species
density), presented by Al-Mufti et al. in 1977. Solid circles represent woodland
herbs, open circles represent grasslands, and triangles represent tall herbs of open
areas. Reproduced by permission of Blackwell Publishers.

colleagues (see above references), who were attempting to establish general,

quantitative, predictive relationships between community biomass and richness.
Combining our data from two coastal riverine systems with the data of one
of our colleagues, Kathy Taylor, we found a pattern rather unlike the one we
expected (Figure 10.2). Instead of a unimodal curve, we found a unimodal
envelope (see additional discussion of this relationship in Marrs et al. 1996).
Also, we found biomass to be rather unimportant as a predictor, but found
that species richness was strongly correlated with microelevation and sediment
salinity and, to a lesser degree with soil organic matter. Table 10.1 shows the
results of the multiple regression that we performed using all of the measured
predictor variables. The fit of this multiple regression model to the data is shown
in Figure 10.3.
Based on the results we obtained and our familiarity with wetland systems,
we formulated a multivariate conceptual model (Figure 10.4). We further sup-
ported this model with additional data about the existing species pools (we
used the term “potential richness”) and how they varied with salinity. To quote
Gough et al. (1994),

In conclusion, in this study we found that biomass was not an adequate predictor of
species richness. One reason for this inadequacy appears to be that while stresses
such as flooding and salinity may greatly reduce the pool of potential species that
 The systematic use of SEM: an example 261

Table 10.1. Multiple regression results for species richness as a function of

environmental variables (modified from Gough et al. 1994)

Cumulative
Predictor variables Coefficient Std error R-square p<
Constant −3.90 1.084 0.001
Biomass −0.0011 0.0003 0.02 0.001
Elevation 3.10 0.377 0.57 0.001
Salinity 0.51 0.137 0.69 0.001
Soil organic 0.052 0.011 0.82 0.001

12
10
Richness

8
6
4

2
0
0 1000 2000 3000 4000
Biomass, g m−2
Figure 10.2. Relationship between total above-ground community biomass
(live + dead) per m2 and number of species per m2 , found by Gough et al. (1994)
in coastal marsh systems. Reproduced by permission of Oikos.

can occur at a site, those species that have evolved adaptations to these factors may
not have substantially reduced biomass. Thus, we recommend that models
developed to predict species richness should incorporate direct measurements of
environmental factors as well as community attributes such as biomass in order to
increase their applicability.

Multivariate hypothesis formulation and evaluation

Initial evaluations of a multivariate model at the Pearl River
In 1993, Bruce Pugesek and I initiated a more extensive study designed to
evaluate the ideas presented in Figure 10.4 using SEM. The system to which
we wished to apply this model was a coastal riverine marsh landscape located
262 Structural equation modeling and natural systems

12

10

Observed Richness 8

0
0 2 4 6 8 10 12
Predicted Richness
Figure 10.3. Fit of data to the multiple regression model in Table 10.1 (from
Gough et al. 1994). Reproduced by permission of Oikos.

Potential
Richness

Environmental Competitive
Biomass
Conditions Exclusion

Realized
Richness
Figure 10.4. Hypothesized conceptual model of factors controlling species rich-
ness in plant communities.

on the shore of the Gulf of Mexico (Figure 10.5). This landscape contains a
number of conspicuous environmental gradients, including gradients in salinity
(salt marsh to fresh marsh), microelevation (from plants in deep water to those
growing on raised levees), and soil organic content (from sandy soil deposits
to organic muck sediments). We also knew from previous studies that natural
disturbances were common and resulted from the activities of wild mammal
populations, as well as from flooding and storm effects. Thus, this system
contained a variety of stresses and disturbances of the sort commonly invoked
in theories about diversity regulation.
 The systematic use of SEM: an example 263

Figure 10.5. Aerial photograph of the Pearl River marsh complex, which was the
site of these investigations.

The next step in the process was to convert the conceptual model
(Figure 10.4) into a construct model, as represented in Figure 10.6. The pro-
cess of specifying a construct model is an important step forward in theory
maturation. The conceptual model presents important ideas, however, it is very
general and the meanings of the terms are somewhat vague. This is not to say
that the construct model in Figure 10.6 is without ambiguity. Models always
have a context, some set of physical circumstances under which the model
makes sense. I have argued previously (Grace 1991) that the context for mod-
els is often not clearly specified in ecological theories, leading to irresolvable
debate. The application of SEM seeks to make concepts and context tangible in
stages, which preserves both general and specific perspectives on the problem.
This topic will be discussed in more detail in Chapter 12 as it relates to the
concept of theory maturation.
One thing that happens when we specify our construct model is that there is
an immediate expectation that the concepts represented will have to be made
operational. This means that it is soon going to be necessary to specify the
meaning of the concepts by describing exactly how they will be measured.
This reality suggested to us a distinction that we wanted to make that was not
specified in the conceptual model; that there are two distinctly different kinds
264 Structural equation modeling and natural systems

ABIOTIC

RICH

DIST BIOM

Figure 10.6. Construct model showing major factors presumed to be controlling

species richness (taken from Figure 1 in Grace and Pugesek 1997). ABIOTIC refers
to abiotic environmental variables, DIST refers to disturbances of vegetation by
animals and flooding, BIOM refers to community biomass, and RICH refers to
species richness. Reproduced by permission of The University of Chicago Press.

of environmental variables, abiotic conditions and disturbances. Making this

distinction was actually not a mandatory requirement, as the use of composites
(Chapter 6) does allow us a method for dealing with highly heterogeneous con-
cepts. However, since much attention has been paid to the effects of disturbance
versus the effects of stress on richness, the distinction between abiotic factors
and disturbance was one we thought would be valuable.
The next step in the process was to develop the structural equation model
(Figure 10.7). As indicated by the model structure, we specified that abiotic
conditions would be measured by combining specific abiotic measurements into
indices. We also proposed to characterize the disturbance regime by quantifying
the observable signs of disturbance in the vegetation. To measure community
biomass, we decided to measure the standing crop of plant material, which
was the parameter measured by Al-Mufti et al. (1977), and also to measure the
amount of light passing through the vegetation, as an additional measure of the
quantity of vegetation. Richness was specified as being measured simply by
the number of species we found in one m2 plots.
An important concept associated with SEM is that the observed variables are
only required to be indicators of the processes and parameters of interest rather
than perfect measures. In order for a measured variable to serve as an adequate
indicator, it is only necessary for the relative values of that variable to be
correlated with the underlying process of mechanistic importance. For example,
when we specify that the conceptual variable DIST would have an effect on
 The systematic use of SEM: an example 265

abiotic
index 1
ABIOTIC
abiotic
index 1

RICH rich

dist DIST BIOM

massm 2 lighthi lightlo

Figure 10.7. Initial structural equation model with latent variables representing
concepts and observed variables representing indicators. Reproduced by permis-
sion of The University of Chicago Press.

BIOM, what is meant is that the disturbance regime has an effect on community
biomass. When a measure of recent disturbance is used as an indicator of
DIST, we are proposing that recent indications of disturbance correlate with
the disturbance regime over a period of time, reflected by the vegetation.
Once the initial structural equation model shown in Figure 10.7 was formu-
lated, we designed a sampling scheme and schedule. Data were then collected
over a two-year period to test this model. Only the results from the first year
were used in Grace and Pugesek (1997), with the second year’s data saved for a
subsequent test of whether the model results would hold up over time, and how
the system changes between years. Data collected included sediment salinity,
site microelevation, soil organic and mineral content (components of the abiotic
indices), recent disturbance (dist), measured as percentage of the plot disturbed,
above-ground biomass per m2 (massm2 ), percentage of full sunlight reaching
the ground surface in the densest part of the plot (lightlo), percentage of full
sunlight reaching the ground surface in the sparsest part of the plot (lighthi),
and the number of species in a plot (rich).
As is often the case when evaluating multivariate models, the fit between our
data and the initial model indicated that it was not adequate. Simply put, our data
were not consistent with the expectations implied by the initial model. The part
of the hypothesis in Figure 10.7 that failed was the proposition that biomass and
light readings can both be used to represent a single conceptual variable. The
symptoms of this failure were that lighthi and lightlo were highly correlated,
266 Structural equation modeling and natural systems

abiotic
index 1
ABIOTIC
abiotic
index 1

RICH rich

dist DIST BIOM LIGHT

massm2 lighthi lightlo

Figure 10.8. Modified full model defining BIOM and LIGHT as separate con-
ceptual variables.

but that massm2 was not well correlated with either light variable. Because of
this, we reformulated our model, as shown in Figure 10.8. An important lesson
was learned here, regardless of your conceptualization of the problem, if two
variables are not consistently and equally well correlated, they will not function
as multiple indicators of a single conceptual variable. Of course, we could have
used the heterogeneous indicators to represent a composite. However, that did
not fit with our objectives in this analysis.
Once the model was reformulated and reanalyzed, another inconsistency
between model and data emerged. A large residual correlation between DIST
and LIGHT was found to exist. As was discussed in Chapter 8 (see Figure 8.12),
this particular residual represents an effect of disturbance on plant morphology
that moderates the effect of disturbance on light. In order to proceed further,
it was necessary to include a pathway from DIST to LIGHT. Only then was
there a consistent match between the relationships in the model and those in the
data.
The results shown in Figure 10.9 represent the partitioning of covariances
in the data as specified by the relationships in the model. Since these results
were obtained using a maximum likelihood statistical procedure, they satisfy
the criterion of being a simultaneous solution for all relationships. It is not my
purpose here to describe all the ecological interpretations of these results. The
interested reader can consult the paper by Grace and Pugesek (1997). What I
would like to point out, however, is that while this model fits the data, we must
conclude that our original formulated model (Figure 10.7) did not. Thus, further
 The systematic use of SEM: an example 267

abiotic 0.99
index 1
0.99 ABIOTIC −0.52
abiotic
index 1

−0.25 RICH 1.0

rich
0.20 R 2 = 0.42
−0.39
−0.57(+/-)
−0.68 −0.51
1.0 BIOM LIGHT
dist DIST
R 2 = 0.61 R 2 = 0.59

0.30
0.86 0.98 0.97

massm2 lighthi lightlo

Figure 10.9. Results for accepted model. Standardized partial regression coeffi-
cients are given. The (+/−) next to the path from LIGHT to RICH signifies that
this path was unimodal. Reproduced by permission of The University of Chicago
Press.

evaluation is still needed using an independent data set, before we can conclude
that our accepted model is valid for the system sampled. As this rather weak
conclusion reveals, the demands that SEM places on empirical validation are
quite stringent. Stated in another way, the model evaluation philosophy pushes
the scientist rather hard to demonstrate that their results have consistent validity
(i.e., consistent applicability in different places and over time), not just local
application.

A more detailed examination of the data

As mentioned earlier, the construct model presented in Figure 10.6 actually
represents a family of models that can be represented using a single data set.
To illustrate this point, Figure 10.10 shows the results of a more specific and
detailed model in which the abiotic factors were represented individually. The
results are based on the exact same data set used to arrive at the more general
results in Figure 10.9, except for the fact that the more general model combined
the individual abiotic data into indices. Again, no discussion of the ecological
interpretations will be presented here. Later in the chapter, I will present an
example that used a derivative of this model for the purpose of exploring the
role of historical factors in controlling species richness.
268 Structural equation modeling and natural systems

1.0
salt SALT −0.23
0.17
0.15
1.0
floodhi
−0.35
0.97 FLOOD
floodlo
0.24
−0.30
0.99
soilorg RICH 1.0
rich
0.96 INFERT R 2 = 0.45
−0.21 −0.43
soilcarb
0.31 −0.63(+/-)
−0.68 −0.59
1.0 BIOM LIGHT
dist DIST
R 2 = 0.70 R 2 = 0.65

0.26
0.86 0.98 0.97
massm2 lighthi lightlo

Figure 10.10. Results for a more specific version of the model. Reproduced by
permission of The University of Chicago Press.

An experimental test of the multivariate model

A certain amount of experimental evidence exists that supports the dependency
assumptions in the above models (for a review of some of this literature, see
Grace 1999). This is not to say that there are no feedback processes that have
been omitted; for example, a reciprocal effect of species richness on community
biomass. What has been assumed is that the relationship between biomass and
richness is asymmetric, with the predominant influence being in the direction
specified. Regardless of this, the principle of theory maturation pushes us to ask
whether the results from the nonexperimental studies so far described have any
predictive power. Testing predictions based on our accepted models (Figures
10.9 and 10.10) not only addresses the question of whether there is consistent
validity, but also represents an opportunity to further refine our model and to
determine the limits of its applicability.
Between 1993 and 1995, Laura Gough devised and conducted an extensive
experimental study designed to manipulate a number of key variables believed
to control species richness in the Pearl River system. The results from these
experiments were then used in two ways to evaluate multivariate hypotheses
(Gough and Grace 1999). First, her results were compared with predictions from
a model like the one in Figure 10.10, except that it was based on the second year’s
 The systematic use of SEM: an example 269

10 10
Observed Species Richness

A B
8 8

6 6

4 4

2 2

0 0
0 2 4 6 8 0 2 4 6 8
Predicted Species Richness
Figure 10.11. Comparisons between predicted and observed species richness
taken from Gough and Grace (1999). A. All plots. B. Excluding fertilized and
fenced plots. Reproduced by permission of the Ecological Society of America.

data collected by Grace and Pugesek. Secondly, the results were represented as
a structural equation model of their own. This model was presented in Chapter 9
and can be seen in Figure 9.11.
Figure 10.11 presents two graphs showing how observed values of species
richness in Gough’s experiments compared to those predicted from non-
experimental data in the broader landscape. When all plots were included
(Figure 10.11A), there was considerable scatter and only 35% of the variance
was explained. Further analyses showed that this was due to the fact that fencing
and fertilizing caused effects that were not quantitatively predicted. With plots
subjected to either of these treatments removed (Figure 10.11B), the remaining
treatments, which included those subjected to changes in salinity and flooding
as well as the controls, demonstrated a stronger correlation between predicted
and observed (R2 = 0.63).
Details of the interpretations of the experimental study can be found in
Gough and Grace (1999). What should be pointed out here, however, is that
many aspects of the model of this system based on nonexperimental data were
supported by the results of the experimental treatments. It was our interpretation
that where the model based on nonexperimental data failed to predict accurately
was for conditions that were experimentally created, but that did not exist
naturally in the field. In other words, it appears that the nonexperimental data
can be used to reveal how factors relate in the unmanipulated community, while
experimental treatments permit one to ask the question, what will happen if we
change the conditions in the system? Thus, when it comes to comparing the
270 Structural equation modeling and natural systems

strengths and limitations of experimental versus nonexperimental studies, each

has a role to contribute.

A search for evidence of historical effects on species richness

in the Pearl River system
In 1994, Glenn Guntenspergen and I set out to see if we could discover the exis-
tence of additional factors that might explain variation in plant species richness
at the Pearl River study area. Of particular interest to us were factors that could
be related to landscape position and that might reflect past events. Prior work
(Brewer and Grace 1990) had suggested that periodic tropical storms, which are
common in this region, leave a long-lasting effect on plant community zona-
tions. To examine the possibility that this might be important in understanding
diversity patterns at the Pearl River, we established a sampling scheme that
placed plots relative to the mouth of the Middle Pearl River (downstream to
upstream), and relative to the river channel (streamside to interior). The assump-
tions that we wished to test were (1) that distance from the river’s mouth would
reflect the effects of past saltwater intrusions from tropical storm events, and
(2) that distance from the stream channel would reflect past overbank flooding
events. It is important to point out that the specific question we were asking was
not if these events happened, but whether they had lingering effects of richness
that were not reflected in current environmental conditions.
To address this question, we first developed a multivariate model that
included a minimum set of the best predictors of richness (based on the previ-
ous experience of Grace and Pugesek 1997). This model included soil salinity,
microelevation, disturbance, and light readings (as a measure of plant abun-
dance). Then we asked, if the inclusion of landscape position variables might
explain additional variance in richness. This model and its results were presented
in Chapter 8 as an illustration of sequential hypothesis testing (see Figure 8.6)
and will not be reproduced here. What is important to the current discussion is
that we found that distance from the mouth of the river did explain an additional
12% of the observed variance in richness, while distance from the river’s edge
did not contribute new information. Based on these findings, we concluded
that landscape position could reveal effects of past events that influence current
diversity patterns.

A further examination of spatial effects

At a later time, Glenn Guntenspergen and I returned to the question of whether
there were hidden controls of richness that could be detected from landscape
 The systematic use of SEM: an example 271

position. This time, we were joined in the search by an eager group of graduate
students at the University of Louisiana, who were involved in a multicampus
course in Biocomplexity offered by the National Center for Ecological Analysis
and Synthesis. Together, we re-examined the earlier data Guntenspergen and I
collected to see if the grids of plots at each site along the river held additional
clues. This time, the question we wished to address was whether small-scale
historical effects might show up as positive correlations in richness among
adjacent plots that were not related to known environmental gradients (Mancera
et al. 2005).
Starting with an examination of the data, we determined that there was
spatial autocorrelation among plots. In other words, we found that plots that
were spatially close were similar to one another in richness, more often than
would be expected by chance. Such spatial autocorrelation has been reported
before (Legendre 1993) and is probably very common. It is entirely possi-
ble, of course, that such spatial autocorrelation in richness simply reflects
spatial autocorrelation in controlling environmental conditions. As seen in
Figure 10.12, the relationship between spatial patterns in richness and spa-
tial patterns in environmental variables represents an important problem for
interpretation. Do the spatial patterns represent a tight mapping to spatial vari-
ations in environmental conditions, or do they represent historic effects such
as dispersal? To address this problem we first factored out the variation in
species richness that could be ascribed to known environmental factors, and
then tested for whether residual richness still showed spatial autocorrelation.
This sequential hypothesis testing is represented in Figure 10.13, with the
test of neighbor richness represented by a “ghost” variable, indicating that
its effect was determined after the effects of the other variables had been
considered.
The analyses showed that once environmental factors were considered, spa-
tial autocorrelation in species richness disappeared. This means that we were
unable to find any evidence of small-scale historical effects, or other unmea-
sured causes of spatially controlled variations in richness in this system.

Applicability of findings to other systems

More recently, the questions that have interested me are (1) how plant diver-
sity is regulated in a wide array of systems, and (2) whether there are general
features that apply broadly across systems. In a study of meadows in Finland,
Grace and Jutila (1999) examined the relationships of plant richness to both
grazing and environmental gradients (see Figure 7.2). These results generally
indicate support for a common construct model such as shown in Figure 10.6.
272 Structural equation modeling and natural systems

Figure 10.12. Topographic plots of spatial variation in species richness and other
variables at one of the five sample sites along the Pearl River, based on sampling
in a 5 × 7 grid of 1 m2 plots. In these figures, rows are at different distances from
the river’s edge (in meters) and columns are at different locations along the river
at a site (also in meters). From Mancera et al. 2005. Reproduced by permission of
Springer Publishers.

General support for such a construct model was also found by Grace et al.
(2000) in coastal tallgrass prairie, although the influence of recent disturbances
was minor in this system. In a study of woodlands in Mississippi, Weiher
et al. (2004) found that the presence of trees in a prairie grassland moderates
soil and biomass effects on herbaceous richness; requiring an alteration of the
general construct model for such situations. Investigations of diversity regu-
lation in California chaparral (Grace and Keeley 2006) revealed both similar
construct relations, along with the importance of landscape features, and the
increasing importance of spatial heterogeneity with increasing plot sizes. In
other studies, the overwhelming importance of abiotic constraints on diver-
sity patterns of serpentine endemics (Harrison et al. 2006) suggests a dif-
ferent balance of forces at play for plants occupying extreme environmental
 The systematic use of SEM: an example 273

Neighbor
salt SALT Richness

elev1

elev2
FLOOD RICH rich
elev3

elev4

dist DIST PABUN

light

Figure 10.13. Model used to evaluate relationship between species richness and
contemporary variables at each of the five grid sample sites at the Pearl River.
PABUN represents plant abundance. The “ghost” variable, Neighbor Richness,
was evaluated for a relationship to RICH after the effects of all other variables
were removed; representing a sequential variance explanation test.

conditions. Altogether, these studies suggest support for certain general fea-
tures across systems, and numerous specific factors of importance in particular
situations or contexts. We will return to the question of generality in our final
chapter, where we consider how SEM methods may evolve to permit the evalu-
ation of very general models that can apply across systems diverging in specific
properties.

Summary
This chapter has sought to give the reader an insight into an example of the
ecological perspective that can be created through a committed approach to
multivariate model development, evaluation, and refinement. The point of pre-
senting such an extensive example is not to imply that all of these steps are
required. Instead, what I hope to have accomplished is to show the reader how
the adoption of a multivariate perspective opens up a new way of learning
about ecological systems. It has been my personal experience that the pursuit
of a multivariate understanding has greatly enhanced the ecological insights
274 Structural equation modeling and natural systems

I have gained. I am eager to extend these studies to include additional variables

and pathways, such as reciprocal interactions between biomass and richness,
exploration of the role of habitat variability, and interactions with other trophic
levels. In Chapter 12, I follow up on this extended example with a more philo-
sophical discussion of how multivariate theories can contribute to the maturation
of ecological science.
 11
Cautions and recommendations

In this chapter I consider some of the pitfalls that can be encountered, and present
a few recommendations that may help to avoid problems in implementing and
interpreting results from SEM. Once the reader is finished with this chapter,
they may wish to visit Appendix I to see how some of the recommendations
from this chapter are applied to example SEM applications.

General limits of statistical analysis

Robert Abelson, a statistics professor at Yale for many years, published a book
a few years ago entitled, Statistics as Principled Argument (1995). The the-
sis of this book is that the proper use for statistical analyses and results is as
an aid to interpretation rather than as a set of rigid protocols for establish-
ing facts. A related principle is that statistical analysis of data is a matter of
approximation. We seek to use models that are approximately correct for our
systems. We rely on samples of data that are hoped to approximate some larger
population. We rely on distributional assumptions that are approximately cor-
rect. And, we hope that our next effort will be an even better approximation.
Those of us in fields that are just beginning to use SEM are in the luxurious
position of being able to benefit from decades of intense work that has gone on
in other disciplines. Structural equation modeling developed to its current state
through the combined efforts of practitioners working in a variety of fields, but
especially the humanities. There now exists a vast body of literature on SEM.
Importantly, several software packages are available for SEM analyses, and
many have gone through numerous improvements over the years. There exist
dozens of books summarizing both basic and advanced issues related to SEM.
Finally, there are thousands of trained practitioners and many opportunities
for training for the student wishing to get started. Thus, the beginner can take

275
276 Structural equation modeling and natural systems

some comfort that this “new” method has substantial theoretical and practical
backing.
It would be a mistake, however, to think that the proper application of SEM
methods is straightforward, or that all the major issues with multivariate mod-
eling have been worked out. Many of the models we might wish to solve are
currently not solvable. Sometimes the solutions offered by authors are not appro-
priate. A few of these issues have been discussed in previous chapters. Aside
from these important technical issues, there is a significant need to rely on
fundamental SEM principles rather than prepackaged protocols, if we are to
apply the methodologies to environmental problems where the context of data,
theory, and research objectives may be different. There are many opportuni-
ties for misapplication of SEM methods, some with minor consequences and
some with major. In the first two sections of this book, I tried to be rather care-
ful to emphasize proper applications and interpretations. In the third section
(Chapters 8, 9, and 10), however, I have been quite liberal in my use of pub-
lished applications of path analysis and SEM in the ecological sciences. This
was done deliberately so that I could emphasize the merits of using multivariate
models. It must be disclosed, however, that many of the examples presented do
not provide good illustrations of the best way to conduct analyses. Therefore,
in this chapter I provide a brief description of some of the problems faced in
the application of SEM. This is followed by a brief set of recommendations to
provide additional guidance for the newly beginning SEM practitioner.

Cautions
Kline (2005) offers a list of 44 common errors made using SEM. Here I borrow
heavily from his description of such problems, with the caveat that I qualify
some of his comments. These problems fall under the four headings of errors
of model specification, problematic data, errors of analysis, and errors of inter-
pretation. Table 11.1 contains a paraphrased list of the problems as he presents
them, and I provide some limited discussion of them by group afterwards. In
this discussion I do not address all of the listed issues, some are not discussed
because they are self-evident, and some because they are explicitly covered
elsewhere in the book and cannot be encapsulated briefly.

Specification errors
Specification errors can indeed cause some of the greatest problems with SEM
results. Basing analyses on an inappropriate model can lead to results that
are not close to the true values, because the model structure does not match the
 Cautions and recommendations 277

Table 11.1. List of commonly made errors (adapted from Kline 2005)

Errors of specification
1. Failure to specify model(s) prior to data collection.
2. Omission of important factors from model.
3. Failure to have sufficient number of indicators for latent variables.
4. Using unsound indicators for latent variables.
5. Failure to carefully consider directionality of arrows.
6. Specification of feedback effects to mask uncertainty about relations.
7. Overfitting of model.
8. Including error correlations without sufficient theoretical reasons.
9. Cross-loading indicators on multiple factors without justification.
Improper treatment of data
10. Failure to root out errors in the data.
11. Ignoring whether pattern of missing data is systematic (versus random).
12. Failure to examine data distributional characteristics.
13. Failure to screen for outliers.
14. Assuming relations are linear without checking.
15. Ignoring lack of independence among observations.
Errors of analysis
16. Failure to rely on theory to guide decisions of model acceptance.
17. Failure to check accuracy of computer syntax.
18. Failure to check for admissibility of solutions.
19. Reporting only the standardized parameter estimates.
20. Analyzing a correlation matrix when inappropriate.
21. Improperly performing covariance analysis using correlation matrix.
22. Failure to check for constraint interactions.
23. Failure to properly address collinear relations.
24. Estimation of complex model using small sample size.
25. Setting inappropriate scales for latent variables.
26. Ignoring problems resulting from improper starting values.
27. Failure to check model identification when solving models.
28. Failure to recognize empirical underidentification.
29. Failure to separately analyze measurement and structural models.
30. Failure in multigroup analyses to establish common measurement model.
31. Analysis of categorical items as continuous indicators.
Errors of interpretation
32. Relying solely on indices of overall model fit.
33. Interpreting good model fit as proving validity of model.
34. Interpreting good fit as suggesting model is good at explaining variance.
35. Relying solely on statistical criteria for model evaluation.
36. Relying too heavily on p-values.
37. Inappropriate interpretation of standardized parameters.
38. Failure to consider equivalent models.
39. Failure to consider alternative nonequivalent models.
40. Reification of latent variables.
41. Equating a latent variable with its label.
42. Believing SEM can compensate for either poor data or weak theory.
43. Failure to report sufficient information to permit proper scrutiny.
44. Interpreting significant results as proof of causality.
278 Structural equation modeling and natural systems

situation. Kline begins this set of errors by stating that the failure to specify your
model prior to collecting data is an error to avoid. While it is certainly true that
it is desirable to have your model prior to collecting your data, I do not believe
(1) it is reasonable to expect beginning practitioners to be sufficiently experi-
enced to always specify a correct or near-correct model prior to data collection,
(2) forming a model based on pre-existing data always leads to problems, or
(3) forming a model prior to collecting data necessarily leads to a correct model.
We must first recognize that initial applications of SEM to a problem take place
in the face of insufficient information. Often we do not have sufficient experi-
ence to anticipate the appropriateness of the data and models used. While I do
recommend specification of models before data collection, this still does not
solve the problem of insufficient experience. With experience in modeling a
system, initial specification of models provides important opportunities.
Omitting important factors from models can be a major limitation to our
success. When omitted factors are uncorrelated with other predictors, their
omission leads to reduced predictive power. When omitted factors are correlated
strongly with included factors, their omission can bias coefficients. Only the
relentless pursuit of a problem can uncover either the adequacy of a model, or
the missing factors and their degree of correlation with included factors.
In the social sciences general recommendations exist for the number of indi-
cators to include when using a latent variable model. The “magic number” often
recommended is 3 indicators per latent. Such prescriptions must be understood
within context, however. In Chapter 4, I describe the fundamental principles
associated with latent variables and their measurement. I show that the speci-
fication of single-indicator latents can be quite valuable. I also argue that the
criterion of indicator replaceability should guide decisions about the adequacy
of a set of indicators. Furthermore, whether the objective of an analysis is
inward- or outward-focused strongly influences questions about the number
of indicators per latent. Finally, the nature of the data relative to the concept
embodied by the latent variable is particularly important. Keep in mind that
a latent variable is simply an unmeasured variable. It may represent a narrow
concept, such as an individual’s body weight, or it may represent something
more abstract, such as the concept of body size. When latents represent narrow
concepts, it may be that only one or two indicators are needed to assess that
concept adequately. The more general the concept, the more important it will
be to assess a wide array of properties that relate indicators to its meaning.
Errors 4–6 in Table 11.1 all relate to the general problem of having an
inappropriate model. Sometimes this results from insufficient experience with
the subject being modeled, and sometimes it results from carelessness by the
researcher or analyst. The point here is to urge the practitioner to avoid being
careless by developing an appreciation of the possibilities of alternative model
 Cautions and recommendations 279

structures, and the consequences of making a poor choice of model. It is usually

when an author makes incautious claims from a weakly justified model that
harsh criticism is due. Again, the solution (hopefully) is the sustained testing
and further application of a model to ascertain its ability to withstand repeated
examination and analysis.
Errors 7–9 in Table 11.1 all relate to questionable decisions about details of
modeling. We have previously discussed the pitfalls of overfitting (Chapters 5
and 8). In the process of evaluating models, we will find that modification indices
produced when programs judge the fit to be poor, often indicate that incorpora-
tion of error correlations will improve model fit. Careful scrutiny of output will
also reveal that often the same problem can be “solved” by including directed
paths. A correlation, whether between errors or not, represents either an unana-
lyzed relation or the influences of an unmeasured factor. Either way, we should
avoid the temptation of inserting error correlations in models as a convenient
way of getting them to fit. Rather, they should be indications for a careful search
for the cause of the offending discrepancy between data and model. Finally, the
existence of cross loadings of indicators on multiple factors can sometimes
represent a poorly formed model. Again, reference to theory is required to pro-
vide an assessment of whether cross loadings make theoretical sense or not.

Improper treatment of data

There are generally two types of error that relate to our data. The first set
relates to things that can only be found by looking at the data, such as entry
errors, distributional characteristics, outliers, or nonlinearities. The obvious
solution to these problems is to inspect the data carefully and be aware of the
consequences of such errors. The second set of errors relate to not thinking
enough about the data. Are data missing by design? For example, if values
are missing when organisms die in a study, the data are not randomly missing
from the sample. Fortunately there exist solutions for such problems, but only
if one thinks deeply enough about the issue to correctly diagnose the situation.
Also, are the observations independent or are there obvious dependencies (e.g.,
repeated measures or nested sampling)? Remedies for these issues are discussed
in Chapter 7, although again, the problem must be recognized.

Errors of analysis
Most of the errors of analysis listed in Table 11.1 are technical problems. Basi-
cally, we find that the results achieved from SEM analyses are dependent on
properly coded models that achieve admissible solutions providing unique esti-
mates based on the proper algorithms for the data used and model specified.
280 Structural equation modeling and natural systems

There are also certain requirements for proper results that involve recognizing
when certain conditions are not met (e.g., common measurement model across
groups), or being aware that the test of overall model fit integrates lack of fit in
both the measurement model and structural (inner) model.
Only a few of the errors listed require additional, specific mention. It is a
mistake to rely exclusively on statistical criteria when judging model accep-
tance. The good news is that for observed variable models, judging adequacy
of fit is usually fairly clear cut. Latent variable models with lots of indicators,
on the other hand, pose the greatest difficulties. As we discussed in Chapters 4
and 6, such models tend to accumulate a lack of fit load, and when combined
with large sample sizes, discrepancies between model and data become very
noticeable. The point is, even well-fitting models can be way off the mark if
theory is either insufficient or ignored. At the same time, lack of fit may rep-
resent an isolated misspecification in a model that does not apply to the model
generally, although it must still be addressed.
The problem of sample size adequacy is one that has no easy answers. It is true
that certain sample sizes are so low that they do not provide much opportunity
to detect relationships. Bootstrapping methods, as well as other approaches
(Bayesian estimation), can allow us to avoid the large-sample assumptions of
maximum likelihood. On the other hand, some models are not easily resolvable
even with large samples, and others can provide stable, reasonable solutions
even with very small samples. It is true that model complexity plays a role.
However, the clarity of patterns in the data also plays an overwhelming role in
sample size adequacy. Thus, it is not possible to make reliable pronouncements
in the absence of knowing something about the data. More discussion of this
topic can be found in Shipley (2000, chapter 6).
The management of collinear relations in models is indeed one that requires
care. When two predictors are very highly correlated (r > 0.85), they begin to
become somewhat redundant. If appropriate, modeling them as multiple indica-
tors of a common latent factor is an effective solution that removes problems of
variance inflation and parameter bias. When it is not appropriate to model them
as multiple indicators, some other means of reducing the model to something
equivalent to a single predictor may be needed. Options range from dropping
a variable, to the use of composites, as is needed in the challenging case of
nonlinear polynomial relations, where first- and second-order terms are highly
correlated (Chapter 7).

Errors of interpretation
A few of the errors of interpretation listed in Table 11.1 result from a failure to
understand technical issues. If one relies solely on indices of overall model fit,
 Cautions and recommendations 281

they fail to recognize that lack of fit can either be spread uniformly throughout a
model, or be concentrated in one spot. When all but a few expected covariances
match observed ones exactly, the average discrepancy may not be too high,
even though one or more of the specific discrepancies may indicate a problem.
Also, interpreting measures of fit as indicating a model is “adequate” for all
purposes is a mistake. While we judge a model first by adequacy of fit, poor
predictive power (indicated by low explained variances) can indicate that the
utility of the model may be low. This could be an inherent limitation of the
phenomenon being studied, or it could be a data or model problem. Further,
equating the magnitude of p-values with effect sizes represents a major mis-
understanding of what p-values mean. It is the parameter value that is ultimately
important. Finally, inappropriate use of results, such as an incorrect usage
of standardized path coefficients, usually results from a misunderstanding of
subtle technical issues influencing these parameters. This issue is discussed in
detail at the end of Chapter 3. The improper use of standardized parameters is a
very common problem that needs attention, so care does need to be paid to this
issue.
Most problems in this category, however, represent a misunderstanding or
misuse of statistical analysis. The most egregious error is to assume or imply
that adequacy of fit equates to a demonstration of causation. Many pieces of
evidence are needed, including both an absence of equivalent models with equal
theoretical support, and the existence of independent information supporting
a causal interpretation. The consideration of alternative (either equivalent or
nonequivalent) models (Chapter 8) is a very good way to take a reality check. If
you can convince a skeptic that the other possibilities are not supportable, you
will have achieved a strong “fail to reject” result for your model. Similarly, even
the most sophisticated SEM application cannot change the underlying data and
availability of theory. More likely, SEM will uncover the inadequacies. It is my
experience that adding an SEM analysis to a manuscript does not change the
journal for which it is appropriate. That is determined by the data. Structural
equation modeling can, however, greatly improve our extraction of information
from that data, assuming analyses are correct and meaningful.

Summary recommendations
Given that there are so many particulars that need to be considered in an SEM
application, as implied from the above list of possible errors, my advice to the
beginning practitioner is rather general. In many ways, the following account
is redundant with previous advice, though I believe it may be appreciated by
some readers that it should be presented in one place in simple form.
282 Structural equation modeling and natural systems

Getting started
The transition to SEM is often a bit of a bumpy ride for a number of reasons. It is
hoped that the other chapters of this book will be of some help to the beginning
SEM user by providing a basic understanding of concepts and procedures.
The examples presented (as well as the analyses presented in Appendix I) also
provide opportunities for the reader to become familiar with some of the variety
of possible applications. The suggestions presented here are not meant to be a
rigid protocol, but instead, suggest one approach to take so as to proceed in an
orderly fashion.
The starting point for an SEM analysis will often be the selection of a
system property, or set of system properties, whose variations one wishes to
understand. Note that this is a different emphasis from univariate analyses,
which typically place the focus on a relationship between two variables (e.g.,
the effect of fertilizer addition on plant growth), or perhaps a pair of relation-
ships (e.g., the interactive effects of fertilizer and herbivore exclusion on plant
growth). Certainly a key relationship between two variables can be central to
a multivariate model. However, for a multivariate analysis to be successful,
we need to include the main variables that control the variations in our focal
system properties. Since our objective is to understand some response vari-
able within a system context, we need to approach the problem with a system
perspective.
It is often useful to select or develop a conceptual or other model before
developing the structural equation model. A conceptual model used as a starting
point should include the processes and properties that are known or suspected
to control the focal system properties, regardless of whether they will all be
measured in this study or not. This can help with both model development and
the interpretation of results.
Translation of a conceptual or other nonstatistical model into either a con-
struct model or directly into a structural equation model is the next step. Devel-
oping the initial structural equation model can begin either with a consideration
of the inner model among concepts, or with a measurement model that seeks to
understand factor structure. Regardless of where model development begins, a
model that considers the conceptual variables to be studied as well as the way
they will be measured is the objective at this stage. Sometimes, the develop-
ment of the structural equation model will occur without the development of an
associated conceptual or other model. Ultimately, this depends on the amount
of prior knowledge.
What is often unappreciated is the variety of ways an analysis can be for-
mulated. The flexibility of SEM permits an array of possibilities for how to
 Cautions and recommendations 283

formulate a hypothesis. For example, tests of categorical variable effects on a

system can be evaluated using either the inclusion of a dummy variable in a
general model, or the use of multigroup analysis. Also, longitudinal data can
be analyzed using either growth or time series modeling. Multi-stage analyses,
such as the evaluation of covariates on residuals, can accomplish things not
generally achieved in conventional analysis. Being aware of this flexibility may
permit a researcher to recognize how models different from the examples they
have seen can be formulated to answer the questions that interest them.

Sampling considerations
Once an initial model has been developed, sampling proceeds and the decisions
made at this step can have a substantial influence on subsequent analyses,
results, and interpretations. It should be clear that the degree to which results
can be said to represent something beyond the sample itself depend on how
well the sample represents the sphere of generalization.

Sample size
A first point of consideration is sample size. Again, earlier chapters showed that
even very small samples can be analyzed by resorting to small-sample methods.
However, we are rarely able to have a great deal of confidence in the generality
of studies based on small samples. In many cases in ecological studies, the
ecological units of interest may be large and replication may be challenging. It
is realistic to realize that not all ecological units will make ideal study subjects
for multivariate models, where many parameters are to be estimated. This is
not to say that multivariate models cannot be evaluated for, say, replicate large
ecosystems; only that it cannot be done very precisely using a limited number
of such systems.
Some guidelines are available regarding sample size requirements. Monte
Carlo simulation studies can be implemented in several of the current SEM
software packages, and this is the most efficient way to determine the sample size
needed to obtain particular levels of power. In the absence of such information,
many authors make recommendations as to the number of samples needed to be
estimated per parameter, a common number being 10. Other authors have more
general recommendations such as, 200 samples is a satisfactory number, 100
samples as minimal, and 50 samples as a bare minimum. When working with
smaller sample sizes, bootstrapping procedures are especially recommended
because they do not rely on large-sample assumptions. Additional discussion
of this subject can be found in Hoyle (1999).
284 Structural equation modeling and natural systems

Sample distribution
Multigroup analysis has long permitted analysis of samples that fall into dis-
tinct categories of interest. However, problems arise when data are clustered or
subsampled, and that subsampling is ignored in the analysis. Typically, ignoring
subsampling leads to an increase in the chi-square of the model, because there
is a source of variation in the data that is not reflected in the model. Also, R-
squares and standard errors are overestimated when data structuring is ignored.
A simple and common procedure for handling data hierarchy is the use of
so-called background variables. Several approaches to this problem have been
discussed in both Chapters 7 and 9. However, until fairly recently, more sophis-
ticated methods explicitly designed for handling data structure in multivariate
modeling were not available in SEM software packages. Now, both LISREL
and Mplus, as well as a program called HLM (Hierarchical Linear Modeling,
www.ssisoftware.com) have capabilities for handling hierarchical data. These
more complete methods for hierarchical data handling are based on analysis of
raw data and cannot be accomplished through the analysis of covariances. As a
consequence, results akin to those obtained by repeated measures ANOVA can
be obtained, allowing individuals in the population to possess different response
patterns. Recent advances in Mplus permit an amazing array of analyses to be
performed at different levels in hierarchical data.
One other issue related to sample distribution has to do with distribution in
parameter space. The span of variation in variables captured by sampling, as
well as the distribution of those samples, can have a substantial effect on results
and interpretations. If the goal of a study is to obtain a representative random
sample, random or stratified random sampling is required. Often, however, sam-
pling is stratified or restricted, for example, sampling across gradients. While
various sampling approaches should arrive at similar estimates for the unstan-
dardized coefficients (slope of relationship), they will differ in the standardized
coefficients, as well as the R-squares. This is basically a signal to noise prob-
lem. Sampling at two ends of a gradient will be expected to generate a strong
regression relationship. This can actually be helpful if the goal is to estimate
a linear slope and to ensure that endpoints are well measured. Samples spread
evenly along a gradient will provide an even better representation of a gradi-
ent, although there will be more unexplained variance. Systematic sampling,
regardless of its merits for efficiency, does not provide an accurate measure of
variance explained for the population at large. Traditionally, most hypothesis
testing enterprises are based on a presumption of random sampling. Sequential
methodologies, such as SEM or Bayesian learning, seek repeatability and can
thus be less restrictive about sampling so as to adequately estimate population
variances.
 Cautions and recommendations 285

Analysis
Initial evaluation
When working with pre-existing data, one should be aware that the strength of
inference that can be achieved is dependent on the degree to which analyses
are exploratory versus confirmatory. Beginning users often have the impression
that confirmatory evaluation of models requires that they should not look at
the data prior to fitting the proposed model. This is both a practical and a
theoretical mistake. The appropriate criterion for a confirmatory analysis is
that the researcher should not have used exploratory SEM analyses of the data
to arrive at the final model, unless it is to be acknowledged that the results are
exploratory and provisional. This applies to both the analysis of pre-existing data
as well as cases where initial models were developed prior to data collection.
The place to start with an analysis is a thorough screening and examination
of the data, point by point in most cases. Paying inadequate attention to this step
is a common mistake made by new analysts. Results of all sorts, particularly
regression results (and therefore, SEM results as well) are extremely sensitive
to outliers. These outliers cannot always be detected in univariate space or even
in bivariate space (an outlier in bivariate space is a combination of x and y that is
very different from any other combination). The researcher needs to know that
all the points are representative. Often, apparent outliers can be found in random
samples because random samples do not capture the edges of parameter space
easily unless there is a very large sample. In that case, such samples are not true
outliers. Nevertheless, a decision must be made whether to remove an outlying
sample value that is useful in defining a slope in order to avoid overestimating
variance explained by a relationship. Again, the choice may depend on whether
the goal of the study is to estimate population values, or to develop general
equations.

Preparing for model evaluation

It is important to provide summaries of the data that describe overall charac-
teristics, in addition to those that relate to model evaluation. It is also useful to
perform more conventional analyses such as ANOVA to provide overall assess-
ments of mean differences, although means can also be evaluated in SEM
(Pugesek 2003).
Statistical assumptions play a major role in all statistical analyses. Therefore,
it is necessary to examine the characteristics of variables and their residuals to
determine what kind of data are being analyzed. When it does not appear that
residuals will meet the assumption of multivariate normality (which is admit-
tedly difficult to evaluate), options that do not require this assumption should
286 Structural equation modeling and natural systems

be used. Fortunately for the modern user of SEM, such methods are available.
For large samples, the Satorra–Bentler Robust method provides standard errors
that are accurate for a variety of residual distributions. For small or large sam-
ples, bootstrapping methods can accomplish this. It is important to point out
that parameter estimates, such as path coefficients, are unaffected by resid-
ual distributions, because their estimation does not depend on the normality
assumption.
While normality problems are easily addressed, extreme skew can affect
covariance, and therefore, path coefficients. For this reason, transformations of
the individual variables may be required. Categorical variable modeling options
in programs like Mplus provide a variety of ways to address the analysis of
extremely skewed or otherwise nonnormal variables.
Nonlinearities can lead to serious problems in analyses. Generally it is possi-
ble to detect nonlinearities through the examination of bivariate plots. The sim-
plest approach to correcting for nonlinear relations is through transformations
of the raw variables. As illustrated in Chapter 7, nonlinear modeling for cases
where multi-indicator latent variables are not involved can be accomplished.
Nonlinear relationships (including interactions) among latent variables have
generally been quite difficult to implement due to the complexity of specifica-
tion. This is no longer true for the latest version (version 3) of Mplus, which
now automates the process.
A number of strategies can be used for missing data. While listwise deletion
(deletion of any cases where values for a variable are missing) is often the
simplest approach, there are other options. Imputation of missing values has
received considerable attention and can often be used successfully if there are
not very many missing values for a variable, and they are considered to be
missing at random. Some programs have rather elaborate and quite clever ways
of dealing with missing values, particularly if values are missing for a reason
(for example, through mortality of individuals). In such situations, deletion of
cases with missing values can lead to bias in the sample, and is to be avoided.

Evaluation of models
Once data have been prepared for analysis, it may be necessary to modify the
initially hypothesized model(s). The processes of data collection, data examina-
tion, and data screening can often lead to the conclusion that the resulting data
set differs in important ways from what was initially intended. One should make
as few changes as possible to theoretically based models. However, it should
be clear that in the initial stages of evaluating models relating to a system, there
will be substantial learning associated with all parts of the process.
 Cautions and recommendations 287

The formulation of competing models is highly useful as a means of bringing

focus to an analysis. When causal structure of a model is clear, alternative
models will usually represent evaluations of various processes. When causal
structure is not clear, alternative models may be used to eliminate some of the
logical possibilities.
Analyses of models can often proceed in parts. It is typical for models with
multi-indicator latent variables for analysis of the measurement model to pro-
ceed prior to evaluation of the full model. At all stages in this process, evaluation
of model fit will need to be made. As discussed in Chapter 5, the current con-
vention for covariance modeling is to use a variety of fit indices in making
decisions about models. In spite of their present popularity, indices of relative
model fit, such as the AIC, are not sufficient to determine model adequacy,
although they can support model comparisons. Absolute measures, such as the
model chi-square are very informative, and if the chi-square indicates data fit to
the model, that is usually a reliable indicator of model adequacy. Unfortunately,
chi-square values are sensitive to sample size and for large samples, chi-squares
usually indicates significant deviations from the model, though these may be
trivial differences in parameter values. For this reason, a number of other indices
have been developed that are adjusted for sample size. An examination of the
residuals should also be standard practice. The analyst is referred to Fan et al.
(1999) and Marsh et al. (2004) for recent discussions of this topic.

Interpretation of findings
Virtually everything we have covered so far in this book can come into play
when we draw interpretations. Results are certainly highly dependent on the
correctness of the model structure. They are also dependent on the way data are
sampled. For these reasons, in manuscripts it is best to describe SEM results
in terms that make the fewest assumptions, saving causal inferences for the
discussion section of the paper. Few things raise reviewers’ concerns as read-
ily as what sounds like pre-interpreted results. Structural equation modeling
lingo contains plenty of terms that can encourage us to state results as if their
interpretations are pre-judged, such as discussions of direct and indirect effects.
Sensitivity to this issue can save the author from harsh reactions.
Often, results and associated modification indices will indicate that models
should be changed. When deciding on both model modification and the inter-
pretation of results the first priority should be given to reliance on substantive
(theoretical) logic. An additional caveat is to avoid overfitting to the data and
thereby capitalizing on chance characteristics of the data set. It is important to
always report how the analysis was conducted and never present a final result
288 Structural equation modeling and natural systems

that was formulated through a very exploratory process, as if it was the initial
model and the study was confirmatory. When data sets are large, consider split-
ting the data set into an exploratory sample and a confirmatory sample. Finally,
plan to do followup studies in which modified models receive further testing
with new data sets. Ultimately, the two ingredients for success are a sustained,
multifaceted approach to a problem (Chapter 10 may provide the reader with
some ideas on this subject) combined with a strong reliance on theory. It is
unlikely that we will determine the general validity of our solutions from a
single study, but with repeated efforts conducted using comparable methods,
strong inference can develop.

Conclusions
While the evaluation of multivariate models holds great promise for ecological
studies, the current level of rigor for ecological applications is less than it needs
to be. Proper application of methods such as SEM, requires an appreciation
of the need to collect appropriate samples, handle data properly, formulate
models correctly for the sample being analyzed, perform analyses carefully,
and interpret results appropriately. Because of the complexity of the issues
involved, a commitment of some time and effort is required, compared with
many of the univariate procedures ecologists are accustomed to learning. I
believe the effort is well worth it and hope that the material in this book will
make the process easier for others.
 PA RT V

The implications of structural equation

modeling for the study of natural systems
 12
How can SEM contribute to scientific
advancement?

Motivation for studying complexity

Ecological science has changed in character over the past few decades. Of
course it changed before that time, but the past 30 years is the part to which
I have been a direct witness. I like to say that during that period, ecologists
have gone from being the harbingers of doom to the advisors of the future.
Increasingly, we are asked, what will happen if . . . ? While general opinions
abound, it is rare that we have solid answers.
A rising awareness that the ecological science of the past will not meet
tomorrow’s needs can be seen in several recent major science initiatives from
many of the world’s leading science agencies. Among these initiatives are ones
that promote greater focus on long-term studies, those that enhance efforts to
promote ecological synthesis, those that foster interdisciplinary studies, and
those that encourage a broader range of scientists to apply their knowledge to
applied problems.
Recently, the call for studies designed to address “biocomplexity” has been
taken up in several parts of the world. One of the leading proponents for the
study of biocomplexity has been Rita Colwell, of the US National Science
Foundation, who states, “The fundamental and underlying principle [behind the
NSF Biocomplexity Initiative] is that we must move from strictly reductionist
research to research that synthesizes information and work toward a holistic
approach to understanding and wisely managing the environment.” She goes
on to say, “After investing in biocomplexity research, we’ll be able to make
[better] predictions concerning environmental phenomena as a consequence
of human actions taken.” Similar research initiatives in other nations are also
under way at this time.
Implied by the recent emphasis on biocomplexity is that many of our past
efforts have emphasized what I will call “biosimplicity”, which I define as

291
292 Structural equation modeling and natural systems

the search for simple general explanations for complex ecological problems.
Simple generalizations will continue to fascinate ecologists. They will not, I
contend, provide adequate answers to environmental problems. We have been
doing the best we can with the resources and tools available to us. This book
presents new tools and approaches for the biologist to consider, ones that are
specifically designed to confront complexity. There is a special emphasis in this
book on ways to analyze complex data and achieve answers that are specific to
a particular system, and yet have generality across systems as well.
Methodologies that have been developed for the study of multivariate rela-
tions provide us with a means to advance ecological understanding in a number
of new ways. Their capacity to sort out simultaneous effects can allow us to
move beyond dichotomous debates of this factor versus that factor. They can
also provide us with the means to accommodate both the specific and general
features of data. Perhaps of greatest importance, they push us to collect data in
a different way, one that will, I believe, clarify our ideas and refine our under-
standing. Ultimately, the premise of this book is that the study of multivariate
relationships using methods like SEM can advance our understanding of natural
systems in ways not achieved through other methods.

A critical look at the current state of ecological theory

When Peters (1991) published his book, A Critique for Ecology, I must admit
that I agreed with many of his criticisms, although not with his solution. What
is most relevant here, though, is that sometimes we gain an appreciation for the
value of a new approach through the eyes of a critical discussion of existing
practice. It has been my experience that the merits of studying multivariate
models are often not fully appreciated unless we first consider some of the
limitations of more familiar approaches. This chapter will discuss a number
of examples of existing ecological models and hypotheses that, I believe, can
be advanced using a multivariate approach. These criticisms must be tempered
by the recognition that virtually all current ecological models will presumably
be replaced by more mature and refined models as further work proceeds. The
switch from univariate to multivariate theory, however, represents a change
from looking at individual effects to system responses. Thus, it can potentially
transform the scientific process. The criticisms offered in this chapter may
offend some. It is not my intent to be demeaning. Instead, what I do wish to
do is to challenge the reader to think of how we may take ecological models
and hypotheses to a different level of refinement. If the reader is moved to
consider how we might progress our understanding, even if it is not along the
lines suggested in this book, a valuable purpose will have been served.
 How can SEM contribute to scientific advancement? 293

description
and
comparison

hypothesis
prediction
formation

hypothesis increasingly formulation

refinement clear, explicit, of tests
and predictive

hypothesis
evaluation
Figure 12.1. Representation of the progress of the scientific method.

In this chapter, I will set the stage for a critical look at current ecological
theory by first discussing concepts relating to the development and maturation
of scientific theories. I contend that a univariate mindset has caused ecologists
and others studying natural systems to develop and cling to simplistic theories
that fail to mature. There are many conspicuous properties of immature theories,
and it is readily apparent that such properties are commonly associated with the
most discussed ecological theories of the day. I go on to discuss properties of
SEM that can contribute to theory maturation, and within that discussion point
out topics in our field where current limitations are apparent. In the last part
of the section, I discuss and illustrate an associated set of scientific behaviors,
such as theory tenacity, that contribute to the accumulation of a large set of
immature theories that contradict each other but that do not evolve over time
and for which there is no resolution of merit. My thesis is that the ability to
develop and evaluate multivariate theories has the capacity to greatly enhance
the maturation of theory and lead to scientific progress.

The concept of theory maturation

The study of multivariate hypotheses can be viewed as following the same
logic as the study of any other hypothesis, going from formulation to testing and
refinement. Figure 12.1 provides a very simplistic representation of the scientific
method. One reason I review the steps in the scientific method here is because I
believe that, for the most part, studies of ecological systems have often not gone
beyond the initial stages of the process. Furthermore, I believe this is primarily
because the mindset created by conventional univariate hypothesis testing has
294 Structural equation modeling and natural systems

Theory Maturation

Clear, Convincing,
Novel Idea or
Mechanistic, Predictive
Bold Proposition
Explanation

Figure 12.2. The starting and ending points in theory maturation.

limited the completeness of the theories we have developed. Considering where

we are in the process of theory maturation and the contribution that multivariate
models can make, may contribute to our understanding of what is gained by
using SEM to study natural systems.
The concept of theory maturation parallels our advancement through the
scientific process. An excellent introduction to this important subject was pre-
sented by Loehle (1987), who discussed where the study of hypotheses should
lead us. Figure 12.2 provides the dominant characteristics of the end points
of the process. Theory development often begins with an exciting idea that
suggests a new perspective. Some scientists prefer to emphasize this stage of
the process, offering a continuous progression of novel or bold propositions.
Usually controversy surrounds these ideas specifically because the theories are
in need of clarification and evaluation. Other scientists can spend a career try-
ing to verify and support a new theory they have proposed. Such a tendency is
described by Loehle (1987) as theory tenacity, which can be important in the
early stage of theory assessment, although it eventually becomes an impediment
to progress.
The process of theory maturation is one that continues to refine predictions
and evaluate hypothesized models to the point where our models possess what
can be called the “traits of mature theory”. Such traits include clarity, operational
meaning, high empirical content, quantitative expression, and predictive power.
According to the concept of theory maturation, the study of many scientific
problems is initially characterized by a crudeness of understanding that com-
plicates the process of evaluation. This has been a factor contributing to many
disputes in ecology. Often new hypotheses are presented in terms that are not
clearly defined, and such hypotheses must go through a period of refinement
before it is even clear how to develop appropriate tests (see, for example, the
treatment of such a debate in Grace 1991). Furthermore, it is often the case that
the meanings of key terms in models are ambiguous and “unbounded”, lacking
a description of the realm of space, time, and system components to which they
best apply. Analysis of many of the most visible debates in ecology over the
past 30 years shows that theories possessing immature characteristics are often
associated with these discussions.
 How can SEM contribute to scientific advancement? 295

In general, it is helpful for the researcher to assess the maturity level of

a theory before deciding how to study it. To some degree, such assessments
are usually made, although often not in an explicit fashion. An awareness
of the processes involved in theory maturation can facilitate the scientist’s
approach to a problem. Scientists may, of course, differ from one another in
their own objectives – some wishing to focus on developing new ideas, and
others more interested in working on a problem until its associated theory is
mature enough to allow for predictive applications to ecological management.
What is important is to be aware of the elements in the progress of theory
maturation, and to conduct studies that will advance the development of our
understanding.

The properties of structural equation modeling that can

contribute to theory maturation
Here I think it may be useful to describe in greater detail how a multivariate
perspective on scientific theory, as exemplified by the procedures associated
with SEM, can contribute to theory maturation. To accomplish this, we need
to consider some of the general properties that are associated with SEM. The
reader should be aware that these properties are not always incorporated into
each SEM use. As with the concept of theory maturation, there is a process
of methodological maturation that can be expected when new methods are
introduced. Thus, the discussion that follows really represents the mature use
of SEM in a sustained application.

Property 1: encourages a careful consideration of theory

and measurement
It is to be expected that in the initial stages of exploring a topic, models and
ideas will be vague, unbounded, and frequently nonoperational. It can also be
argued that some ecological ideas appear to have remained in such a state for
decades, without resolution of exactly what mechanisms are being evaluated,
how they are to be measured appropriately, and where the associated models
apply. One factor that can help keep theories in an abstract state is the desire
by some to emphasize the generality of a principle over its applicability. Often
generality is emphasized by making terms more general and, in the process,
even less well defined. There can be merit to such an approach, for a time.
Eventually, I will contend, there comes a time when the merits of a theory must
296 Structural equation modeling and natural systems

be demonstrated in unambiguous terms. For this to happen, a clear system for

evaluation is needed.
Structural equation modeling begins with a consideration of the construct
model, a general statement about the kinds of processes contributing to rela-
tionships and ultimately to observed patterns. Not all applications of SEM may
begin in this way, however. Rather, one may start with an abstract theory or an
empirical result from an exploratory study. Both these starting points can lead
to the formulation of an explicit structural equation model.
The researcher interested in developing a multivariate theory about a prob-
lem might begin with a path model involving only relations among observed
variables. This represents the stage of development of the great majority of
applications of SEM/path analysis to ecological problems. Often when a model
is in this form, we may be unclear as to whether the researcher has considered
the distinctions between the underlying concepts and the observed indicators,
or the issue of measurement error.
A more explicit way to express a multivariate model is as a latent variable
model. This may be a model that has only single indicators for each concep-
tual variable. The value in showing the full form, even though the statistical
meaning of the conceptual variables is exactly represented by the indicator
variables, is clarity. Expressed in the form of a full structural equation model,
a clearer distinction is made between the variables of conceptual interest and
the data at hand, which can often be a rather poor substitute for perfect mea-
sures of the factors that are interacting. Stated in a different way, a path model
with unspecified variable type incorporates one level of theory, one that con-
siders how components interrelate. However, it ignores another level of the-
ory, measurement theory, the science of considering how observations relate
to underlying causes. A full structural equation model with single indicators
may or may not resolve how concepts are to be measured, but at least it rec-
ognizes that these issues exist and deserve future consideration. Sometimes
there is existing information about the reliability of a given indicator variable
as a measure of a conceptual variable that can be used to remove bias from
parameters.
A full structural equation model with multiple indicators typically represents
a multivariate model in which issues of measurement have received substantial
thought. An example of such a model can be found in Pugesek and Tomer
(1996). As indicated in Chapter 4, models with multi-indicator latent variables
can become quite complex. The main point is that SEM encourages a careful
consideration of how conceptual variables are measured and how they inter-
relate, contributing to theory refinement and maturation.
 How can SEM contribute to scientific advancement? 297

Property 2: models are open to inclusion of additional factors

Multivariate models are well suited to inclusion of additional factors. In the
case of the Pearl River example in Chapter 10, results suggested that plant
biomass and light penetration were not adequate indicators of a single latent
conceptual variable. Inclusion of two latent variables not only resulted in a
greater consistency between theory and data, but also led to a greater explanation
of variance, and an isolation of two different kinds of effects of biomass on
richness. A further example was the work by Grace and Guntenspergen (1999)
in which two spatial variables were tested to see if they provided additional
insight into patterns of richness; one variable did and the other did not. If we
had relied only on logic, as we might have in a purely theoretical modeling
effort, we certainly would have included both terms. Model evaluation usually
provides clear guidance as to when variables contribute new information and
when they do not.
Examination of some ecological topics shows a preponderance of “closed”
models. By closed models, I refer to those that emphasize some small number
of factors and that are not able to incorporate additional factors, even if they
are known to be important. One topic that has historically shown a tendency
towards closed models is life history theory. In the early 1970s there was sub-
stantial interest in the theory of r- and K-selection (MacArthur and Wilson 1967,
Pianka 1970, Gadgil and Solbrig 1972), which emphasizes the importance of
the tradeoff between reproduction and survival (Figure 12.3). Once sufficient
examples had been examined, it became clear that other factors also influenced
reproductive effort and survival, and also that many different mechanisms could
play a role (Stearns 1977). Shortly after this discovery, confidence in the empir-
ical utility of distinguishing between r- and K-selected species waned, and the
importance of this theory in explaining life histories received little additional
discussion.
Another example of a one-dimensional model is the R* model of Tilman
(1982), which emphasizes the effects of a tradeoff between tolerance of low
resource levels and the ability to grow under high resource levels. Tilman has
subsequently incorporated a number of other assumptions into his models (e.g.,
Tilman 1988). However, these models remain set with a fixed and limited num-
ber of processes, unable to easily detect or accommodate additional factors
based on deviations between predictions and observations.
Following criticisms of the theory of r- and K-selection, Grime (1977)
proposed a two-dimensional model of plant life histories, involving tradeoffs
among growth rate, tolerance, and reproduction (Figure 12.3). There has been
298 Structural equation modeling and natural systems

One-dimensional Models
r- selected K- selected

growth rate resource tolerance

specialists specialists

Two-dimensional Modell
C

R S

Multidimensional Model
C1

R2 C2

R1
S1

S3
S2

Figure 12.3. Representation of one-dimensional and two-dimensional closed

models, in comparison with a multidimensional model that can accommodate as
many dimensions as the data indicate are important.

a long history of resistance to the acceptance of this model by many (Tilman

1987, Loehle 1988, Grubb 1998, Wilson and Lee 2000) in spite of the fact that
it makes some important and previously ignored distinctions. I believe that the
reason for this resistance is the inherently closed nature of the model. Is there
a reason why the theory cannot become refined to accommodate additional
dimensions, as Grubb (1998) has proposed? Perhaps the very structure of the
model, as well as the nonstatistical nature of the theoretical modeling pro-
cess contribute to this inflexibility. The lowermost model in Figure 12.3 shows
the form that a more complete, multidimensional model might take. Here, the
model shows how multiple components of competitive ability and stress toler-
ance could be incorporated into a multidimensional model of open architecture.
Through the use of SEM, it would be possible to allow the data to tell us how
many dimensions are important, and also exactly how important each dimension
 How can SEM contribute to scientific advancement? 299

is under different circumstances. I believe that the adoption of such a model

would represent an increase in theory maturation for a topic that has remained
in gridlock for 25 years.

Property 3: allows for strong and constructive feedback from

empirical results
In the years of Sir Francis Bacon (circa 1561–1626), there existed in European
university education and scholarly circles a kind of scientific orthodoxy in which
ideas were deemed to reign supreme over observations. Theories, both about
nature and mankind, were hotly debated and resolutions were obtained through
logic and reasoning. Whenever nature failed to behave as expected, based on
the carefully crafted conclusions of philosophers, most notably followers of
the Aristotelean tradition, it was concluded that observations were “messy”,
not a clear reflection of truth. It was Bacon’s contribution that he advocated an
empiricist’s philosophy that sought to liberate us from logical orthodoxy. At the
same time, Bacon was also skeptical about an overreliance on pure description.
Rather, he espoused an approach to the development of theory that relied heavily
on empiricism and experiment. Bacon represented his philosophy using several
classic drawings that served as frontispieces for his books. Figure 12.4 shows
one such drawing from his book Novum Organum, which represents the ship
of scientific exploration passing beyond the pillars of established ideas.
I think that to a significant degree, the empirically directed theory building
nature of the SEM process is in line with the scientific philosophy espoused by
Sir Francis Bacon. Users of SEM will attest to the fact that these techniques
place strong demands on their models. Theory, no matter how logical in the
mind of the scientist, must be able to reconcile deviations that are found between
model and data. At the same time, empirical findings typically result in model
evolution and thereby cause our models to grow. The process of rigorous model
evaluation ultimately leads to a high degree of empirical content of the resulting
model, encouraging the scientist to sail beyond his or her established notions
about what is important, and incorporate those factors demonstrated to have
importance.

Property 4: flexible and useful under a wide range of conditions

One of the attributes of SEM is that it imposes a certain logical procedure in
which ideas about systems are compared with observable properties. As we have
seen throughout this book, metricians have worked very hard to find creative
300 Structural equation modeling and natural systems

Figure 12.4. Engraved title page in Sir Francis Bacon’s book, Novum Organum,
published in 1620 (engraving by Simon Pass). The latin motto written beneath the
ship translates into, “Many shall pass through and learning shall be increased”.
Image reproduced by permission of Richard Fadem.

ways to devise appropriate representations for a variety of multivariate models.

Examples of this creativity can be seen in the selection of ecological topics
covered in Pugesek et al. (2003), as well as in examples shown in Marcoulides
and Schumacker (1996), Schumacker and Marcoulides (1998), and Cudeck
et al. (2001). In spite of all these efforts, structural equation models still have
a limit to their flexibility. Much of my efforts in working with SEM have
been aimed at adapting and combining techniques in order to evaluate models
 How can SEM contribute to scientific advancement? 301

that fall outside of the mainstream of application. Often this process involves
tradeoffs and compromises that give up one model attribute (e.g., estimation of
measurement error) for another (e.g., across-system generality). Examples of
these efforts are found throughout this book, though particularly in Chapters 6,
7, 10, and 13. Two of my several objectives in this work have been to demonstrate
the ability of SEM to be flexible, and also to strive for continued developments
that will permit us to address a broader range of models. There remain, at the
present time, a number of important difficulties in dealing with certain kinds of
problems. Given the methodological advances that are being made, I am hopeful
that further progress can permit SEM to achieve the degree of flexibility needed
for use by researchers studying ecological systems.

Property 5: contributes to development of a predictive

level of understanding
From a strictly statistical perspective, SEM is not always an optimal vehicle
for prediction. The philosophy of covariance-based SEM is geared towards
general theoretical explanation, instead of variance explanation. In particular,
latent variables seek generality over precision. Nevertheless, when the goal is
prediction, structural equation models can be formulated in a way that is well
suited to the task. I believe that the simplest way this will be accomplished
will be by linking structural equation models to forecast and system dynamic
models (Grace 2001), combining the strengths of each. Illustrations of this will
be given in the next chapter.

Examples of how multivariate theory might contribute

to theory maturation
Consideration of the progress of theory maturation in the case
of the intermediate stress hypothesis
It may be useful to consider the progress of theory maturation for a particular
ecological topic, to illustrate some of the preceding points. Here I will focus on
clarity of theoretical concepts and their measurement as it relates to the inter-
mediate stress hypothesis, also known as the hump-backed model of diversity
and productivity.
In 1973, Grime (1973) found a number of interesting patterns in a compar-
ative study of plant species richness in central England. Initially, he empha-
sized two primary and separate hypotheses about the factors controlling these
Table 12.1. Concepts and variables associated with the Intermediate Stress Hypothesis

Author Name of theoretical structure Scope of application Causal or (predictor) variables

Grime 1973a Intermediate stress hypothesis Plants Intensity of environmental stress
Grime 1979 Hump-backed model Plants (1) dominance, (2) stress, (3) disturbance, (4) niche
differentiation, (5) ingress of suitable species, (6)
species pool (standing crop and litter)
Huston 1979 Dynamic equilibrium model All organisms, Rate of competitive displacement, population growth
latitudinal gradients, rates (soil resource availability, productivity,
regional variations population densities, environmental fluctuations)
Tilman 1982 Resource ratio model All organisms Resource supply rates, environmental heterogeneity
(productivity)
Taylor et al. 1990 Species pool hypothesis Plants Pool size of adapted species (soil fertility)
Rosenzweig & Unimodal model All organisms Productivity (rainfall, ocean depth, evapotranspiration,
Abramsky 1993 soil fertility)
Grace 1999 Multivariate model Herbaceous plants Environmental conditions, species pool, total biomass,
(environmental indicators, above-ground biomass +
litter, light reaching ground surface)
 How can SEM contribute to scientific advancement? 303

patterns, the intermediate stress hypothesis and the intermediate disturbance

hypothesis. The intermediate stress hypothesis stated that species density
(= species richness) is highest at intermediate levels of environmental stress,
because of offsetting effects of abiotic limitation and competitive exclusion.
In 1979, Grime presented the humped-back model, which proposed that low
above-ground standing crop could serve as an indicator of high environmen-
tal stress. A number of authors have since proposed that there exists a gen-
eral relationship between species diversity and environmental stress or habitat
productivity. Most commonly they have proposed that diversity is highest at
intermediate levels of stress or productivity.
Table 12.1 summarizes information on some of the conceptual and predictor
variables that have been proposed to control diversity in conjunction with the
intermediate stress hypothesis. It can be safely stated that, at the present time,
despite a great deal of theoretical and empirical work on this topic, there is little
consensus about the mechanisms that are most important, and even whether
there is a consistent humped-back relationship that applies generally. A major
effort to evaluate and synthesize our understanding of this relationship and its
empirical support has taken place in the past few years (Palmer 1994, Abrams
1995, Grace 1999, Waide et al. 1999, Gross et al. 2000, Mittelbach et al. 2001).
Overall, this literature suggests:
(1) A weak but discernible unimodal (= humped-back) relationship is common,
but not consistent.
(2) Empirical supporting evidence is strongest for herbaceous plants, and gen-
erally weak for consumer species.
(3) Even in the best cases, measures of productivity explain only a modest
portion of the observed variance in species richness.
(4) Control of species richness in natural systems is strongly multivariate and,
where data are sufficient, multivariate models do quite well in explaining
richness variations (average R2 for herbaceous communities is 57%).
(5) The theoretical basis whereby environmental stress may affect animal diver-
sity is strongly disputed, with a large number of competing hypotheses.
(6) The mechanisms whereby herbaceous plant richness is regulated are more
generally agreed upon.
(7) It is argued by most who have reviewed this topic that more theoretical and
empirical work is needed to clarify the issues.
My own assessment is that this topic exhibits a low degree of theory mat-
uration. Most models are based on vague and even nonoperational theoretical
constructs. Also, most deal with only one or two of the many processes that con-
trol diversity. A few, such as Grime’s humped-back model have a high degree
304 Structural equation modeling and natural systems

of empirical content and a well-bounded realm of applicability. However, most

have no clearly defined relationship between constructs and indicators, and
have made little effort to provide rigorous empirical evaluations. What is per-
haps most distressing is that there is little indication of theory maturation for
most models. The model presented by Grime in 1979 is identical to the one
presented in his recent update of that book (Grime 2001). The model presented
by Huston in 1979 is unchanged in his 1994 assessment of diversity, although
he does discuss species pools in his 1999 paper (Huston 1999). Tilman’s model
has likewise remained unchanged between its first appearance in 1982 and in its
most recent treatment (1997). The same can be said for the species pool model,
which is unchanged from Taylor et al. (1990) to Aarssen and Schamp (2002).
Examining the body of work dealing with this topic, it would appear that,
following the initial publication of their ideas, most authors have focused on
a tenacious defense of their original formulation. When changes are presented
they are usually one of two types, (1) extending the generality of their ideas
by using even more abstract terms, or (2) providing a broader range of sup-
porting examples that rely on a greater array of indicators. The result is that,
generally speaking, there is little movement towards a consensus on even the
most basic points of concept and measurement. Typifying this situation is the
fact that virtually the same handful of various and poorly defined examples are
presented in support of the three contrasting models of Grime (1979), Tilman
(1986 – see his Figure 2.4) and Huston (1994 – see his Figure 5.4). Rosenzweig
and Abramsky (1993) noted that, of all the studies they examined linking pro-
ductivity to diversity, none actually measured productivity, but instead, used
indicators ranging from annual precipitation to depth below the ocean surface.
As Palmer (1994) has noted, individuals continue to propose more and more
competing explanations for diversity patterns (there are now over 100) without
any resolution. Abrams (1995) arrived at a similar conclusion.

Consideration of the progress of theory maturation in the case

of the effects of species diversity on community productivity
Current disagreement about the relationship between species diversity and com-
munity productivity (Tilman et al. 1996, Huston 1997, Wardle 1999, Hodson
et al. 1998, Lawton et al. 1998, Loreau et al. 2002) reminds us that controversy
is a common part of our science. In order to put this debate in its proper per-
spective, we can view it in terms of theory maturation. I believe it is accurate
to say that the hypothesis that species diversity enhances habitat productivity
has remained in a fairly immature state until very recently. I would further
 How can SEM contribute to scientific advancement? 305

argue that this debate has been complicated by ambiguous definitions, unclear
predictions, equivocal tests, and considerable controversy – all signs that the
hypothesis being addressed needs clarification and refinement. I believe that
adoption of a multivariate approach to theory construction would faciliate the
theory maturation process in this case as well.
We have already seen that certain refinements in both definitions and exper-
imental methods have had to be made in order to facilitate progress. First,
the consequences of drawing a limited number of species from a larger pool of
species are not really so simple (Huston 1997, Wardle 1999, Mikola et al. 2002).
Separating the number of species selected from the productivities of the particu-
lar species that wind up in a sample has turned out to be more difficult than most
investigators initially expected. The recognition of this “sampling effect” has
forced researchers to make a distinction between the effects of having a variety
of species in a sample, and the effects of winding up with a particularly produc-
tive species in a sample just because many were selected. As the importance of
this distinction has started to sink in, it has been necessary to recognize that a
key process that should be isolated is “overyielding”, the capacity for a mixture
of species to be more productive than any of the individual species in the mix-
ture. Thus, we now see that evaluating the hypothesis that diversity enhances
productivity requires us to partition at least two processes: the sampling effect
and the overyielding effect.
A second area in which definitions have had to be refined deals with the
term diversity. Initially it was assumed that species diversity automatically led
to functional diversity. However, it has now been found quite useful to estimate
the effect of growth form or functional group diversity on productivity because,
on average, species of different functional groups are more reliably different
than species of the same functional group. Again, here it seems that separating
the effects of the diversity of specific plant attributes, such as the ability to fix
nitrogen, from other types of plant variety may help us to further refine our
hypotheses and experiments.
Many of the experiments that have been conducted to address the potential
effects of diversity on productivity have been extremely impressive from the
standpoint of the effort involved. In spite of this, there has been much con-
troversy about the interpretation of the results. Some problems have resulted
from the fact that the above-described distinctions had not been made at the
time the experiments were initiated. Other problems arose because of the diffi-
culties associated with developing an operational test of an abstract hypoth-
esis. The particular set of species considered, the growing conditions, and
the duration of the study can all have specific effects that might confound
any attempt to understand the general effect of diversity on productivity from
306 Structural equation modeling and natural systems

an individual study. The process of theory maturation requires that we use the
lessons learned from trying to implement experimental tests in order to further
refine hypotheses. One such refinement that is often valuable is to “bound” the
hypothesis by describing the conditions under which it is more or less likely to
occur.
Failure to reconcile the hypothesis that diversity enhances productivity with
other ideas about how these two variables relate represents another area in which
refinements need to be made. Some of the initial resistance to the hypothesis
that diversity enhances productivity comes from evidence that high produc-
tivity inhibits diversity for herbaceous plants. So, in experiments involving
herbaceous plants that have examined the effect of diversity on productivity
(the majority of studies thus far), how is it that the negative feedback from
enhanced productivity on diversity is ignored? Reconciliation of the opposing
processes relating diversity and productivity will ultimately need to be consid-
ered if we are to draw conclusions about the effects of diversity on productivity
that are meaningful in the real world.
Another area where further refinement is needed deals with the spatial
scale at which the relationship between diversity and productivity is examined.
We may imagine (and even predict) that for a natural landscape, having a large
pool of species in that landscape will enhance the ability of plants to colonize
disturbed sites rapidly, occupy stressful areas more completely, and respond
to environmental changes more quickly. However, this is not the question that
has been addressed up to this point, at least not at the scale of experimental
evaluation (Symstad et al. 2003). Whenever there is a disconnect between the
idea being tested, that the number of species existing in a small plot affects
the productivity of that plot, and the broad appeal of the idea that diversity is
important in maintaining a productive landscape, further theory maturation is
needed.
Finally, another factor that is playing a role in the debate over the relation-
ship between diversity and ecosystem function is the relevance of this idea to
conservation efforts (Grime 2002). There has been a great deal of emphasis
placed by some ecologists on the prospect that an enhancement of productivity
by diversity represents a justification for the preservation of diversity (Naeem
2002). Many ecologists, myself included, think we should be careful in this
arena. Does this mean that if a species is rare and has no measurable effect on
ecosystem function, as certainly will be true for many rare species, that it has no
value? Traditionally ecologists have argued that diversity is important because
it represents successful evolutionary experiments, and the loss of diversity is a
loss not only of variety itself, but also of genetic information. It is important
 How can SEM contribute to scientific advancement? 307

that we should not allow a desire to justify the importance of diversity to society
to be the supporting basis for keeping a scientific hypothesis in an immature
form.
Overall, the current controversy about how species diversity influences habi-
tat productivity represents one stage in the development of an important ques-
tion. Compared to most previous ecological debates, I am encouraged by the
degree to which ideas, definitions, and experimental tests have been refined.
Much of this has been caused by the intense scrutiny the work on this topic has
faced. However, there is much to do. Ecological questions of this magnitude
require a long-term sustained effort to evaluate. Furthermore, I believe that an
explicitly multivariate approach could help in a number of ways. There is still
little effort aimed at integrating any positive effects of diversity on productiv-
ity with models that explain how productivity and other factors can control
diversity.

Multivariate models offer opportunities for maturing

our ideas and predictions
It should be pointed out that SEM by itself is not an ultimate solution for
understanding the regulation of species diversity, or for arriving at some general
synthesis of the topic. However, I believe that the methods and philosophy
associated with SEM can greatly add to theory maturation on this subject.
Admittedly, this conclusion is based on my perceptions of what I have learned
studying herbaceous plant diversity using a multivariate approach. As Kaplan
(2000) has recently emphasized, the interface between SEM and more complete
theoretical understanding is in need of further work; a theme expanded upon in
the chapter that follows.
Generally, what are needed are approaches that facilitate a progressive refine-
ment of ideas and explanatory power, while permitting broad general compar-
isons. This is a tall order and will not be achieved easily. Nevertheless, I believe
the organizational framework of SEM, its capacity for representing multifactor
explanations, its openness to modification, and its uncompromising feedback
from empirical tests will help. I further believe that this framework will allow a
continued refinement of models and tests that will lead to the inclusion of more
factors of importance and the explanation of a greater array of field conditions.
What will not be accomplished by using a multivariate framework alone is a
resolution of the exact mechanisms behind all the relationships. A combination
of approaches will be required to achieve this lofty goal. It is also unclear at
308 Structural equation modeling and natural systems

present whether the widespread usage of a multivariate approach will contribute

to consensus, or simply a proliferation of various multivariate models. I believe
that regardless of what happens, the new models that will evolve will be more
explicit, more realistic and more predictive, all of which will contribute to
theory maturation. Towards the end of the next chapter, I present an example
of a multivariate theory developed from the sustained use of SEM.
 13
Frontiers in the application of SEM

The “symphonic” nature of natural systems

A system can be defined as an interacting set of parts. There are, of course, many
kinds of system. I sometimes like to draw a parallel between natural systems and
symphonic music (another kind of system), as a means of conveying certain
ideas about what we gain from a multivariate approach to ecology. In sym-
phonic music, a composition relies on the participation and interplay of many
instruments. As we listen to an orchestra (preferably live for greatest effect) we
gain an entirely different experience compared with the effects of listening to
the instruments individually. The same is true with the study of natural systems.
The preceding chapters have sought to convey through examples what can be
gained from achieving a multivariate perspective. While descriptive multivari-
ate methods have been increasingly employed by natural scientists, the study
of hypotheses has largely relied on univariate methods. The result has been
an emphasis on individual processes, with little opportunity to consider their
role in the system. My own earlier experience with the study of competitive
interactions convinced me of the value of building and evaluating multivariate
hypotheses, where the importance and impact of individual processes can be
seen in some context. It is, of course, true that the simultaneous study of all
parts of a system is not feasible. However, it would seem that our understand-
ing of the observed variation in a system property of key interest can often be
improved without resorting to extremely complex models.

Frontiers in statistical methodology

Recent and future advances in statistical methodology have the potential to add
in important ways to the questions that can be posed and addressed. Already

309
310 Structural equation modeling and natural systems

we have seen how the introduction of a statistical concept such as a latent vari-
able can imply new elements of our research paradigm, leading to new lines
of inquiry. A significant number of statistical inventions have been described
in this book and their role in permitting different types of models to be esti-
mated has been illustrated. In Chapter 7, a number of more advanced statistical
techniques were referenced, including multigroup analysis, categorical variable
modeling, the treatment of nonlinear relationships, latent growth models, hier-
archical methods such as multi-level modeling, and the modeling of reciprocal
interactions. All of these open up new questions that can be addressed by the
researcher.
Currently, progress is being made on many fronts related to conventional
SEM. Recent compilations of advancements can be found in Cudeck et al.
(2001), Marcoulides and Schumacker (2001), and Pugesek et al. (2003). These
address a wide range of topics from nonlinear modeling to alternative estima-
tion techniques, to categorical variable modeling. The reader should be aware,
however, of parallel methodological advances taking place outside of the tradi-
tional domain of SEM that have the potential to greatly alter the way we work
with multivariate hypotheses. The broad enterprise of modeling multivariate
relations is currently undergoing a staggering rate of growth, with a wide vari-
ety of approaches being developed and promoted under the umbrella concept
of “graphical models” (e.g., Borgelt and Kruse 2002). These efforts have been
driven by a wide variety of objectives, from the desire to develop artificial
intelligence to the wish to mine vast quantities of data. These methods are also
highly dependent upon the capacity for high-speed computers as well as new
algorithms for solving multiequational systems and searching iteratively for
solutions. Borgelt and Kruse (2002) provide the following list of approaches to
the search for relationships in multivariate data:

r classical statistics
r decision/classification and regression trees
r naive Bayes classifiers
r probabilistic networks (including Bayesian and Markov networks)
r artificial neural networks
r neuro-fuzzy rule induction
r k-nearest neighbor/case-based reasoning
r inductive logic programming
r association rules
r hierarchical and probabilistic cluster analysis
r fuzzy cluster analysis
r conceptual clustering
 Frontiers in the application of SEM 311

Consideration of many of these approaches is beyond the scope of this book. The
reader can find out more about these and other methods in Dasarathy (1990),
Bezdek and Pal (1992), Langley et al. (1992), Muggleton (1992), Pearl (1992),
Agrawal et al. (1993), and Anderson (1995). Shipley (2000) also provides an
overview of some of these methods in chapter 8 of his book.
It would seem that one important external influence impinging on classical,
maximum likelihood SEM is Bayesian reasoning. Bayesian thinking is lead-
ing to distinctively different methods for estimating and interpreting statistical
models. It is also leading to novel approaches to evaluating networks of rela-
tionships. Emerging from the growth in interest in, and applications of Bayesian
statistics is the recognition of a new future for SEM, Bayesian structural equa-
tion modeling (Jedidi and Ansari 2001). Thus, it is valuable, I think, to describe
briefly what these new ideas might mean for SEM.

Bayesian inference and structural equation modeling

Bayesian ideas
There has long been debate and dissatisfaction within the field of statistics on
how to express probabilities in a meaningful way. While there have been a num-
ber of different systems proposed for describing probabilities (Oakes 1990), the
two dominant approaches today are frequentist (von Mises 1919) and Bayesian
(Bayes 1763). While conventional SEM practice shares some characteristics
with the Bayesian philosophy, such as an emphasis on sequential learning, a
disregard for null hypotheses, and a less rigid approach to hypothesis testing,
likelihood methods, as emphasized in Chapter 5, are frequentist in nature. At
the heart of the frequentist view, there is an assumption of the existence of
population parameters and general probabilities, as well as a belief that the
long run frequency of repeatable events provides an estimate of those popula-
tion probabilities. In this system, inferences are drawn concerning populations,
and the p-value refers to the probability that if a study were conducted 100
times, the true value would fall within the 95% confidence intervals 95% of the
time. Characteristic of the implementation of the Neyman–Pearson protocol
(Chapter 1), either–or conclusions are typically drawn about the truth of some
proposition relative to an alternative hypothesis.
Long before the frequentist view became the backbone of modern statistics,
the alternative ideas known as Bayesian inference were elaborated, first by
Bayes himself, and then shortly after that by Laplace (1774). There are many
different ways we could describe the Bayesian view and, in fact, a great many
312 Structural equation modeling and natural systems

“flavors” of Bayesian inference have been proposed (Good 1983). I think it is

most helpful to define Bayesian inference as the quantification of uncertainty.
This definition makes clear from the outset that the emphasis is on what is known
(our current knowledge) about the probability of observing some event, and the
associated range of uncertainty. This is quite a contrasting emphasis from the
frequentist goal, which is estimating the value of a population parameter, which
is presumed to exist, but is unknown to us.
What is of great practical importance at the present time, is that the ways
of solving a wide range of models using Bayesian methods has very recently
become possible, thanks to the development of efficient ways of sampling data
and very fast computing speeds. This means that a discussion of the merits of
Bayesian methods is suddenly much more relevant. This is particularly true for
structural equation models, which are typically much more complex than the
average univariate model, and therefore, even more challenging for Bayesian
estimation.
Because the Bayesian view emphasizes uncertainty, it is inherently refer-
enced to the person’s perspective. In other words, uncertainty is a property of
the quality of a person’s knowledge, rather than a statement about a population.
This has profound implications for the way that inference is drawn, the nature of
the conclusions, and the reactions that different individuals have to its validity.
From an abstract perspective, it would seem that how uncertain we are about
something is more of a psychological issue than a scientific matter. However,
from a practical perspective, we interface the world and make decisions based
on our estimated certainty about observable events. Applied science is typi-
cally driven by uncertain information (as well as perceived importance). For
this reason, a Bayesian approach has, at the very least, a substantial niche in
the conduct of statistical inference.
Among the many differences between frequentist and Bayesian views is that
Bayesian’s are not simply interested in estimating the first and second moments
of the distribution (mean and variance), they are interested in describing the
entire probability distribution. Once we know the probability distribution of
some occurrence or co-occurrence, we can estimate directly and precisely the
likelihood of any particular range of events of interest. This preoccupation with
the probability distribution is precisely why Bayesian inference gives more
detailed and potentially more accurate estimates, and also why it is more compu-
tationally difficult. Frequentists can invoke many different assumed probability
distributions. Typically, however, the normal distribution is assumed for contin-
uous functions, and a few different distributions such as binomial or poisson are
assumed for discrete events. A presumption that one’s variables approximate
the normal or multinormal distribution provides frequentists with a “quick and
dirty” calculation of confidence intervals for the population parameters. This
 Frontiers in the application of SEM 313

has been (and continues to be) an enormously valuable compromise in drawing

statistical inference.
Because Bayesian inference is focused on uncertainty, it is inherently linked
to sequential learning. Here I define sequential learning as the progressive
refinement of one’s knowledge based on accumulated experience. Sequential
learning implies that our estimate of the probability of observing some out-
come is updated based on additional data. This concept is embodied in the
Bayes Theorem, which stated in words is, our estimated odds of observing
some outcome are equal to the observed frequency of outcomes from data plus
our prior estimate of the odds before we had the data.
Now, when one hears this definition, I expect we may have two, nearly simul-
taneous reactions (I certainly do). First, it seems unreasonable on the face of
it to combine one’s uninformed prior estimate of odds with those informed by
data. At the same time, if one possessed really good prior information about the
odds of an event, then combining the odds derived from pre-existing experience
(and potentially a lot of previously obtained data) with those derived from the
current sample would seem like a superior idea (if one’s purpose is prediction).
I think this pair of reactions summarizes nicely the view that Bayesian estimates
can either be worse than or better than those derived from frequentist estimation
methods, depending on the quality of prior information. Furthermore, the rele-
vance of combining prior information with current data to derive an averaged
estimate will depend on both (1) the quality and magnitude of the data, and (2)
the relevance of prior information to the current case. In the study of natural
systems, it is not always clear that we are sampling repeatable phenomena. Nat-
ural systems, particularly communities and ecosystems, can possess a great deal
of spatial and temporal heterogeneity. This fact of life adversely impacts both
frequentist and Bayesian inference, the latter because it affects the relevance of
prior information.
Now, some claim that the issue of combining prior odds with results from
a data set may not pose as large a problem for Bayesian estimation as it might
seem. The use of uninformative priors (vague initial estimates) presumably
reduces their influence on the estimated posterior odds. Of course, there is a
tendency for uninformative priors to contribute to a broad probability distri-
bution, even if it does not have a strong influence on the estimation of the
mean response. Thus, uninformative priors may adversely impact parameter
estimates.
To some degree, the use of priors in Bayesian estimation can initially distract
us from a number of potential advantages that arise from the mechanics of
the estimation procedure. Those mechanics now involve various resampling
methods, such as the Metropolis–Hasting method (Hastings 1970) or the Gibbs
method (Casella 1992) which are used to estimate the precise shape of the
314 Structural equation modeling and natural systems

probability (or joint probability) distributions. Having the precise shape of the
probability distribution yields inherently superior information, at least in theory.
In practice, there are a number of things that can interfere with reaping the
potential benefits of Bayesian estimation, especially problems in estimation.
Nevertheless, available information based on comparisons and considerations
of Bayesian versus frequentist (typically maximum likelihood) estimates yield
the following points (all of which should be considered preliminary at the
present time).
(1) With large samples that conform well to distributional assumptions and
in the absence of informative priors, Bayesian and frequentist (maximum
likelihood) estimates produce similar results (Scheines et al. 1999).
(2) Bayesian estimates of centrality and precision measures are substantially
superior when maximum likelihood is used on data that deviate a great
deal from the associated distributional assumptions.
(3) Bayesian estimation is more suited to small sample data because maximum
likelihood depends on asymptotic behavior.
(4) The choice of priors has a greater impact on small samples in Bayesian
estimation.
(5) Bayesian methods provide more intuitively useful statements about prob-
abilities, particularly in the context of decision making.
(6) Bayesian estimation, when good priors exist, can overcome some problems
of model underidentification, enhancing the range of models that can be
solved.
(7) Bayesian modeling holds promise for facilitating the modeling of non-
linear and complex model structures.
(8) Bayesian estimation may be inherently better suited for modeling cate-
gorical outcomes, because it is based on conditional probabilities.
(9) Bayesian estimation for structural equation models with multi-indicator
latent variables is potentially problematic, because the lack of a simultane-
ous solution procedure can yield inconsistent estimates (Congdon 2003,
chapter 8).
(10) It is unclear whether methods for evaluating model fit under Bayesian
estimation are as reliable as, or superior to, those achieved under maximum
likelihood.
Overall, it would seem that Bayesian approaches hold promise for general
application to SEM (Rupp et al. 2004). That said, maximum likelihood may be
comparable for a significant range of applications. Current SEM software has
now automated a very wide range of model types, including latent class anal-
ysis, mixture growth modeling, incorporation of interactions in latent variable
 Frontiers in the application of SEM 315

models, and hierarchical models. Finally, the potential pitfalls for Bayesian
estimation in complex models have not been adequately researched as of yet.
Initial indications are that particular problems arise in latent variable models
involving multiple indicators per latent, when using Bayesian estimation. Since
this is a particular strength of ML SEM procedures and represent a very com-
mon model type, further work is needed to ascertain whether Bayesian methods
will be applicable to the full range of model types.

Bayesian networks
What we might refer to as “Bayesian SEM” goes beyond considerations of esti-
mation and can be viewed as falling under the umbrella of “Bayesian networks”
(Pearl 2000, Jensen 2001, Neapolitan 2004). This is a topic currently experi-
encing explosive growth and interest. Thus, any attempt to generalize about it is
subject to significant misrepresentation. That said, it would seem that the field
of Bayesian networks includes interest in both causal as well as descriptive and
predictive modeling, and is thus not synonymous with Bayesian SEM. Never-
theless, there is a great deal of applicability of Bayesian network (BN) methods
to SEM problems. This applicability relates not just to Bayesian estimation,
as described above, but also to a Bayesian perspective on conditional proba-
bility that can be used to evaluate network structure. I am reluctant to delve
too deeply into the specifics of BN methods in this book, because the termi-
nology associated with networks is both completely different from that used in
conventional SEM, and somewhat involved. Bayesian network methods make
reference to “directed acyclic graphs” (DAGs) instead of “recursive models”,
“nodes” instead of “observed variables”, “edges” instead of “paths”, and so on.
Therefore, I refer the reader to Pearl (2000), Shipley (2000), and Neapolitan
(2004) for a description of this terminology, as well as the topic of “d-separation”
which deals with how models can be evaluated for conditional independence.

Conclusions about Bayesian methods and SEM

In this discussion of Bayesian inference and its relation to SEM, I have tried to
give a very general description of a highly complex topic. A thorough treatment
of Bayesian SEM would require a book of its own, if for no other reason than
the fact that it involves completely different solution procedures, software, and
terminology. In spite of all that, the scientific goal of developing and testing
multivariate hypotheses so as to understand systems goes beyond whether we
use frequentist or Bayesian methods. The questions addressed by SEM, many of
the interpretations, and the insights produced also apply regardless of the method
316 Structural equation modeling and natural systems

of estimation and inference. Whether maximum likelihood methods remain

widely used, or whether Bayesian procedures and ideas come to predominate
in SEM applications, I believe the study of multivariate models will lead us to a
better understanding of natural systems and a greater maturation of our theories.
Ultimately, Bayesian methods are focused on providing the best prediction of
the next event, while ML emphasizes an evaluation of the data.

Frontiers in research applications

While advances in analytical methods can allow us to address new questions,
new ways of applying methods can also advance the scientific enterprise. Up to
the present, structural equation modeling has been largely confined to playing a
strict role as a means of statistical modeling. While the inclusion of latent vari-
ables can allow us to more directly represent concepts in our models, in most
cases, the latent variable effects in our models must represent single effects.
This limitation is discussed in the chapter on latent variables (Chapter 4) and
in Chapter 5, composite variables are introduced as an aid to modeling multi-
dimensional constructs. I believe that composite variables have the potential to
be applied in ways that will permit some entirely new kinds of research appli-
cations. In particular, the use of composites allow us to begin to formulate more
general statistical models that better match our theories. In this book, the use of
composite variables, either to represent general constructs, or to represent non-
linear effects, has been prominent. I have included this material despite the fact
that many of the statistical procedures for analyzing models containing com-
posites have not been established in the literature. Thus their inclusion is one
of necessity rather than because they represent an accepted element of SEM.
There are other necessities that drive our application of SEM. One is that
we need to have greater flexibility in relating general theoretical models to
structural equation models. We also need to be able to better link mechanistic
theory to statistical implications. This implies a desirability of linking SEM with
statistical forecast models and with system simulation models. Such linkages
between model types are limited only by example applications. Thus, much
of the potential for multi-modeling approaches can be revealed simply by dis-
cussing ways different techniques can be combined. Let us begin by considering
some of the strengths of each major model type.
Conceptual models have long served as the primary means of summarizing
general features of ecological systems. In the past few decades, mathematical
modeling, either of particular processes (commonly accomplished using ana-
lytical mathematics), or of system dynamics (commonly accomplished using
system simulations) has played a major role in ecological investigations. To
 Frontiers in the application of SEM 317

some degree, each type of modeling has a particular kind of utility. Conceptual
models are capable of great generality and flexibility. They can, in some cases,
imply a great deal of content without requiring specification of explicit details.
They may also serve as logical starting points for exploring topics or organiz-
ing our knowledge. Analytical mathematical models, in contrast, seek to make
mechanistic assumptions and their consequences explicit. They are often used
to explore the interactions between a small number of variables. Such models
can have very general implications, although the business of determining their
relevance to real ecosystems is often left to the field biologist. System simulation
models have come to be the most common type of mechanistic model. Motives
for their development and usage range from organizational to exploratory to
predictive. Their size and complexity also vary greatly. System simulations can
grade from those exploring the dynamics of a single process to those involving
hundreds of equations. Traditional statistical models also have a long history of
application in ecology. Experimental designs and analyses, sampling programs,
and hypothesis evaluations of all sorts rely on statistical models. Such models
can, and often are, used for static or dynamic predictive modeling.

Relating conceptual models to structural equation models

The first step in multivariate hypothesis testing is to translate the important
elements of an initial conceptual model into a construct model. This step repre-
sents an opportunity to evaluate our general conceptual models for tangibility.
The researcher may find, for example, that the way that a concept is described
or represented in a general conceptual model may not make sense in a statistical
modeling context. Concepts represented may prove to be too abstract to trans-
late in a general way, or may need to be subdivided or even renamed before their
theoretical meaning becomes clear. It may also be the case that conceptual mod-
els include processes or components that are virtually impossible to measure.
As described in Chapter 12, this translation process has the potential to lead
to theory maturation and clarification of our general conceptual models. Thus,
it is expected that the interplay between conceptual modeling and multivariate
hypothesis testing will lead to refinement of conceptual models, in addition to
hypothesis testing.
Because of the capacity for latent variable models to represent both link-
ages between conceptual variables and how concepts are to be measured, such
models can play an expanded role compared to our traditional use of statistics.
Certainly the construct model can be viewed as a particular type of concep-
tual model. Perhaps the biggest differences between structural equation models
and general conceptual models are (1) there is an agreed upon convention for
expressing and interpreting structural equation models, (2) structural equation
318 Structural equation modeling and natural systems

heterogeneity

species
richness
species pool
recruitment extinction

environmental disturbance
conditions

plant
abundance

growth biomass
loss

Figure 13.1. A multivariate conceptual model of primary relationships controlling

species richness (from Grace 1999). The model is presented in dynamics systems
format and follows the conventions of Hannon and Ruth (1997). State variables are
shown in boxes. Clouds indicate unspecified sources and sinks. Circles represent
control variables and those with linkages to thick arrows regulate changes in state
variables. Reproduced by permission of Elsevier Publishers.

models only include system variables that can be measured or estimated from
data, and (3) the measurement portion of structural equation models explicitly
defines how the concepts will be quantified, and therefore, their explicit mean-
ing. For these reasons, it is expected that structural equation models will come to
play a role in the conveyance of ecological theory that extends beyond their role
as a specific statistical hypothesis. The construct model, which represents the
statistical expectations in a more general way, without concern for how
the constructs are to be measured, can play a special role in transitioning from
the general to the specific. An illustration of this transitioning was presented in
Chapter 10 (compare Figures 10.4, 10.6, and 10.7).

Incorporating statistical results into conceptual models

By combining the results of our site-specific studies of species richness with
a thorough review of the published literature, I have proposed a more explicit
conceptual model, informed by structural equation model results (Grace 1999).
This multivariate conceptual model (Figure 13.1) seeks to subsume the individ-
ual mechanisms for richness control proposed by Grime (1979), Huston (1979),
Tilman (1982), and Gough et al. (1994) into a single multivariate framework
 Frontiers in the application of SEM 319

that is consistent with field and experimental data. According to this conceptual
model, species richness is controlled by the size of the species pool, species
recruitment into a site, and local extinction, all of which can be affected by
habitat heterogeneity. At the same time, community biomass is controlled by
growth and loss processes and, in turn, can suppress recruitment and increase
extinction at elevated levels. Finally, this model presumes that environmental
conditions can have differential effects on species pools and plant growth. Thus,
an environmental condition can be evolutionarily stressful (i.e., few species can
live there) and yet not appear to be ecologically stressful (i.e., the species that
can tolerate those conditions are quite productive). Collectively, this conceptual
model differs from any of the pre-existing models of this subject by being both
multivariate and relatable to the statistical expectations of field data. I believe
it also has many of the desirable properties described in Chapter 12, such as
being able to accommodate new information as it is found, operational, reason-
ably general, and predictive. Current work is seeking to incorporate additional
processes that are now believed to be important, and to include more explicit
information about the relative importance of different pathways.

Statistical forecasting
The linkage of statistical models to forecasting is simple and direct. Regres-
sion relationships have historically been used to estimate coefficients that can be
applied to forecast future outcomes (Pankratz 1991). Here I use the term forecast
to emphasize that projected outcomes are conditional on a number of assump-
tions that may or may not hold. It is reasonable to expect that estimation of
predictive relationships can often be improved upon using modern multivariate
methods. There are two basic characteristics of structural equation and related
models that make this so. First, multivariate models partition net effects into
separate pathways, allowing for greater accuracy in conditional predictions.
Secondly, the use of latent variables allows for the removal of measurement
error and the production of more accurate path coefficients.
The value of isolating effects using SEM can be illustrated by referring
back to the example of atrazine effects on ponds by Johnson et al. (1991;
Figure 9.1). When atrazine, a phytotoxin, was added to replicate mesocosm
ponds, the responses of algal populations were variable and did not consistently
differ from control ponds overall. Use of this net outcome as the basis for pre-
dicting risk would be of limited generality or utility, as it was found that other
factors, including aquatic vegetation, grass carp, pond depth, and zooplank-
ton densities all affected algal populations. Once the effects of other factors
were considered using SEM, the direct negative effect of atrazine on algae was
found to be quite strong. Furthermore, the dissection of effects using a model
320 Structural equation modeling and natural systems

RICH

pool
effects species sploss
sploss
rate
species
recruit
pool
recruit germ
exclude
richness
rate effect
submodel
total splossmax
environ. mass
factors recruitmax
seasons

LIVE
MASS

growth
biomass
bioloss
effects growth
submodel
gromax persistence
environmental biomassmax

submodel

LITTER litter
submodel
addition decomposition
decomprate

Figure 13.2. Representation of a system model used to simulate dynamics of

richness, live biomass, litter, and environmental effects. Model symbolism repre-
sents the STELLA programming language (High Performance Systems®). Refer
to Grace (2001) for details. Reproduced by permission of Oikos.

containing causal dependence gives more accurate prediction coefficients than

would be obtained from a multiple regression approach. A direct translation of
unstandardized path coefficients into a forecast model would allow forecasting
under a broad range of scenarios. In general, results from multivariate models
should permit more versatile and accurate forecasting in most cases.

System simulation
I believe there can be an important complementarity between system simula-
tion and SEM. System simulation has tremendous flexibility and the capacity
to incorporate a nearly unlimited amount of theoretical detail. A considerable
variety of kinds of mechanism can be included, and these can readily be parti-
tioned to incorporate spatial heterogeneity and scalability. The primary limita-
tion with system simulation models is that they generally project or extrapolate
from what is already known. Thus, like other theoretical mathematical models,
the potential for informative feedback is limited. Discrepancies between pre-
dicted and observed can certainly be detected and analyzed. However, for many
models, the number of estimated parameters that might be involved can mask
 Frontiers in the application of SEM 321

50
A
40
Richness
30

20

10

0
0 300 600 900 1200
Biomass, g m−2

50

40
B
Richness

30

20

10

0
0 300 600 900 1200
Biomass, g m−2
Figure 13.3. A. Relationship between total biomass and richness expected along
a resource gradient based on simulation results (Grace 2001). B. Relationships
expected along a nonresource gradient. Field results are generally more consistent
with case B than with case A, and emphasize the importance of nonresource stresses
on communities.

the reasons for observed discrepancies. Because of their capacity for directed
discovery, structural equation models can not only reveal that data do not fit a
model overall, but can also show precisely where discrepancies occur. Through
this capability, SEM has the ability to assist in the construction and evaluation
of simulation models.
On the other side of the process, it is possible to use the results of system
simulations to generate multivariate theoretical expectations that can be evalu-
ated using subsequent SEM analyses. Such an approach was used to explore in
more detail the explicit interworkings of the various processes in Figure 13.1
(Grace 2001). To accomplish this, I created a system simulation model (shown
in Figure 13.2) designed to capture some of the processes implied by the
322 Structural equation modeling and natural systems

ENV RICH

BIOM

B
ENV RICH

BIOM

Figure 13.4. Qualitative representation of multivariate expectations based on

simulations whose results are shown in Figure 13.3. A. Expected pattern for change
in richness along a resource gradient where environmental effects on richness
parallel those on biomass. B. Expected pattern for change in richness along a
nonresource gradient where environmental effects on richness are independent of
effects on biomass.

conceptual model. To provide a greater degree of realism, I separated biomass

into live and dead components and confined plant growth, species recruitment,
and species loss to occur during the growing season. Another distinction I
made was to distinguish two kinds of environmental gradient, resource gra-
dients (such as a soil fertility gradient) and nonresource gradients (such as a
salinity or flooding stress gradient).
Among the explorations made using the dynamic system model was the
consideration of how biomass and richness would respond to resource gradi-
ents versus nonresource gradients. The presumption made was that resource
gradients would affect both community biomass and species pools, while non-
resource gradients would also affect species pools, but not necessarily impact
biomass. As shown in Figure 13.3A, a unimodal curve between biomass and
richness would be expected along a purely resource gradient. Along a nonre-
source gradient, however, the most likely pattern would be a unimodal envelope
of values (Figure 13.3B). Within a multivariate statistical context, Figure 13.4A
shows the expected pattern of relationships associated with a resource gradient
and Figure 13.4B shows that expected for a nonresource gradient. If these theo-
retical explorations are valid, based on a review of the field data, it appears that
 Frontiers in the application of SEM 323

environmental gradients in the field typically behave as nonresource gradients

in that environmental effects on richness are often unrelated to environmental
effects on biomass.

Conclusions
Structural equation modeling will not always be appropriate or useful in every
study. Furthermore, it will not solve all our problems and can yield results no
better than the data upon which analyses are performed. Yet, it would seem that
it has the ability to allow us to tune in to nature’s symphony to a greater degree
than ever before. Only time and experience will tell whether these methods will
revolutionize science. It would seem from the current vantage point in history
that they have great potential. It is my hope that biologists will be open to the
process of exploring their utility in helping us to understand natural systems
and their dynamics.
 Appendix I
Example analyses

Purpose
The purpose of this appendix is to give the reader a little experience with the process
of structural equation modeling. I should make the disclaimer that my treatment of the
illustrations in this appendix is admittedly superficial. I expect that most readers will
forgive this transgression, although some methodological experts may wish that I went
into far more detail here rather than assuming that the reader understood all the more
subtle nuances from the main chapters of the book. It is beyond the scope of an appendix
such as this to give a complete exposition on model evaluation. It is also beyond our scope
to give even a superficial overview of the variety of models that can be analyzed using
SEM. Our task is further complicated by the significant number of competing software
packages, their various features, and the significant issue of how they arrive at the results
they yield. Somehow, in spite of all this, it seems useful to provide some illustrations
of example analyses, allowing the reader to see some of the steps in the process, some
sample output, some of the pitfalls and how they may be avoided or resolved.
It seems that the wisest thing to do in this appendix is to first give an overview of
some of the various resources that are available to help the reader towards becoming a
competent SEM analyst. I will assume, rightly or wrongly, that the material and advice
given in the text will be sufficient for the individual to understand at least the basic
analyses. Chapter 11 provides a brief summary of some of the main things to avoid and
to attempt to accomplish when performing SEM. So, if we assume that one knows what
one wants to do, this appendix serves to provide some examples of analyses, including
how to translate questions into models and how to interpret output and create revised
models. Additional supporting material can be found at www.jbgrace.com.

About SEM software

In this appendix I emphasize model construction and interpretation. It is not my intent to
train the user in the use of particular software packages. Each software program comes
with its own user manual, and all have a great number of features that go far beyond what
can be covered here. That said, it is useful to at least comment on the various software
programs designed to perform SEM analyses.

324
 Example analyses 325

There are a number of commercial software programs currently available for per-
forming SEM. The first widely used program of this sort was LISREL, which as of this
writing is in its eighth generation. Over the years, the authors (Karl Jöreskog and Dag
Sörbom) have added a number of modules and features to LISREL, as they continue to
do. The original programming language for LISREL involved a rather cumbersome spec-
ification of matrix elements. Since that time, a more intuitive programming language,
SIMPLIS, has been added and most recently, a graphical programming capability has
been introduced. A number of other programs are also available for covariance analyses,
including EQS, AMOS, CALIS (a module of SAS), SEPATH (a module of Statistica),
and Mplus. New users of SEM inevitably wish to know which of these programs is best.
One point in this regard is that most of these programs are updated fairly frequently,
making any description I might give of the limitations of a program potentially out of
date by the time this material is read. For this reason, I suggest readers perform internet
searches for the programs that interest them and go to the software developer’s web
site to find out the latest features and associated costs. Also, free student editions of the
software are available for download for many of these programs and many of the exam-
ples that follow can be run using the free versions of the software. Finally, published
reviews can often be found in the journal Structural Equation Modeling, as well as in
other outlets.
Given the above caveats, I will go on to say that as of this writing, AMOS, EQS, and
LISREL all possess graphical programming capabilities. It is generally true that most
new users would prefer to begin with a graphical programming environment, so these
all make good choices. On the other hand, I think when one is aware of how simple the
command-based programming is for programs such as Mplus, CALIS, and the SIMPLIS
module of LISREL, there is no need to automatically pick a graphical-based package.
More important to the selection process in my opinion is the availability of features.
When it comes to features, most of the main programs being used today possess some
unique features, which frustrates our ability to choose one program that is “the best”
at all things. Again, this can change at any time, so the reader may wish to compare
features at the time they read this. At the moment, I will note my enthusiasm for Mplus,
which is the program I use most often, for having the most features. Other programs
such as LISREL and EQS are also feature rich, while the program AMOS has some very
friendly bootstrap features. Again, let the buyer be aware that all these comments will
soon be out of date. Also, some programs are considerably more expensive than others,
which can influence choices.

Example applications
In this appendix I offer just a few simple examples to illustrate some of the kinds
of models that can be run and the type of output produced. First I set the stage and
then present a graphical representation of the model or models of interest. One or
more hypothetical data sets are presented for the purpose of seeing if the data fit the
expectations of any of the models. Note that the data will typically be summarized as the
correlations plus the standard deviations. This information is sufficient for calculating
the covariances, which will be the raw material for our analyses in most cases. Next,
output from analyses are presented, along with a brief discussion. Those interested can
download the student version of Mplus from their web site to run these and other models
326 Structural equation modeling and natural systems

?
x1 y2 e2

x2 y1 e1
Figure A.1. Simple path model representing a set of possible relations between
exogenous and response variables.

if they like. The Mplus program code that accompanies the example analyses presented
in this appendix can be downloaded from www.JamesBGrace.com.
One additional comment about the examples to follow is warranted. In these examples
I tend to discuss the estimation of the “effects” of one variable on another. It is important
to always keep in mind that the SEM analysis does not establish causality, but that the
ability to interpret relationships as effects or influences depends on outside information.
Of course, when such outside information is lacking, the whole enterprise of SEM is
somewhat compromised. For that reason, I make the assumption that there is reason
to justify the structure of the models presented. For those working with real examples,
sometimes thought experiments can help to establish or justify model structure. For
example, for an x → y relationship, we can ask, if we were to manipulate x, would we
expect y to respond? We may also ask, if we were to manipulate y would x respond? If
the reasonable answers to those two questions are yes and no, respectively, a basis for
establishing directional influence is established, although the degree of confidence in
that conclusion is influenced by many things.

Example 1: a simple path model

The situation
Imagine we are interested in the influences of some system drivers on a system response,
in the presence of an intervening variable. In such a case, our system drivers might be
represented as exogenous observed variables x1 and x2 , our intervening variable as y1 ,
and the ultimate system response of interest as y2 (Figure A.1). There are a number of
questions we might ask about this system. It is often the case that we wish to know
if the relationships between the x variables and y2 are mediated by y1 . The model in
Figure A.1, omitting the paths with question marks, might be designated the “indirect
effects” model. In this model we ask if any effects of x1 or x2 on y2 can be explained
by y1 . Models that contain the paths with the question marks then become alternatives
to the indirect effects model. The direct paths from x1 or x2 to y2 presumably represent
separate mechanisms whereby the exogenous variables influence y2 (separate from those
mechanisms that involve mediation by y1 ).

Some possible applications

One example that would fit this situation is where x1 and x2 are soil conditions, y2 is the
proportion of plants in a plot that are flowering, and y1 is the average size of the plants
 Example analyses 327

Table A.1. Hypothetical data for simple

path model example. N = 100

x1 x2 y1 y2
x1 1.0
x2 0.200 1.0
y1 0.387 0.576 1.0
y2 0.679 0.356 0.766 1.0
Std. dev. 1121.6 1.79 278.9 0.0856

Table A.2. Variance/covariance matrix

derived from Table A.1. N = 100

x1 x2 y1 y2
x1 1 257 986
x2 401.53 3.204
y1 121 059 287.56 77,785
y2 65.19 0.054 55 18.287 0.007 327

in the plot. In a situation such as this, one might naturally ask whether variations in soil
conditions relate to the frequency with which plants flower, and if so, whether this can
be explained by the fact that plants are larger under certain soil conditions and larger
plants are more likely to flower.
Another example that would fit this situation is where we are interested in variations
in population densities of an herbivorous insect as our y2 variable, and we wish to relate
population densities to the abundance of its preferred food plant, y1 , when there are two
habitat factors x1 and x2 thought to influence plant growing conditions. We might wonder
if there is something about either of the habitat factors that might have some additional
influences on insect populations, independent of influences mediated by influences on
its food plant. These additional influences would show up as significant direct paths
from the xs to y2 .

Data
Illustration 1.1
Here I wish to convey a simple but critical lesson about running models. To accomplish
this, we will first convert the data in Table A.1 into its raw form, a variance and covariance
matrix. Typically when one analyzes data using any of the SEM software packages, the
raw data will be examined and for many (though not all) analyses, a variance–covariance
matrix will be used to perform the analyses. Let us see what our raw data will look like.
We can make the calculations directly. The variances that will go into the diagonal will
simply be the squares of the standard deviations from Table A.1. The covariances can
be calculated by multiplying the correlations by the product of the standard deviations
of the two variables involved. The logic for this can be seen by revising Eq. (6) in
Chapter 3. The results are presented in Table A.2.
328 Structural equation modeling and natural systems

Table A.3. Recoded data from Table A.1.

N = 100

x1 x2 y1 y2
x1 1.0
x2 0.200 1.0
y1 0.387 0.576 1.0
y2 0.679 0.356 0.766 1.0
Std. dev. 1.1216 1.79 2.789 0.856

Our goal now will be to use the data in Table A.2 to evaluate the model in Figure A.1
that omits the paths with question marks. The output from that analysis is as follows:

Results and discussion of initial analysis for Illustration 1.1

NO CONVERGENCE. SERIOUS PROBLEMS IN ITERATIONS CHECK YOUR
DATA, STARTING VALUES AND MODEL
What you have just witnessed is the result of out-of-bound calculations resulting
from having numbers in the variance/covariance matrix that are too large. This is a
very common problem for first time analysts. Fortunately, the solution is very simple, to
recode the data. I would recommend that the data be coded so that the standard deviations
are all in the range between 0.1 and 10. With that in mind, we now adjust our data from
Table A.1, with the resulting data presented in Table A.3.

Illustration 1.2
The goal now is to use our recoded data to evaluate the model in Figure A.1. We again
start with the “indirect effects” model, which omits the paths with question marks. The
results obtained from that analysis are presented below. Note that the form of the results,
including the types of indices presented and the form of the output vary depending
on software package. The Mplus package presents a fairly abbreviated form of output
compared with some others, such as LISREL.

Results for Illustration 1.2

THE MODEL ESTIMATION TERMINATED NORMALLY
TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 57.425
Degrees of Freedom 2
P-Value 0.0000
RMSEA (Root Mean Square Error of Approximation)
Estimate 0.526
90 Percent C.I. 0.414, 0.648
Probability RMSEA <=0.05 0.000
 Example analyses 329

Discussion of Illustration 1.2

So, we now have a successful run and have achieved output. If the reader has read
Chapter 5, they will be able to recognize that the measures of fit given, the chi-square test
and the RMSEA, both indicate some important deviations between model expectations
and data. We would generally prefer the p-value associated with the chi-square to be
one that gives us confidence that model and data are reasonably consistent. As was
discussed earlier, chi-square increases with increasing sample size, and this measure of
fit is sometimes too sensitive for practical purposes. However, a sample size of 100,
which is what is used in this case, it not excessively large, and as a result, we would
reasonably expect a p-value greater than something like 0.05 if data supported this model.
The RMSEA adjusts for sample size, so we do not attribute its indication of significant
deviations to a sample size problem, regardless of how many samples we have. Since
both indices indicate we have a poorly fitting model, we reject this model.
Since our data did not match the initial model, we conclude that this is the wrong
model for these data. Therefore, we do not bother to look at the parameter estimates or to
see whether any of the paths appear to be nonsignificant. Since we have concluded that
we ran the wrong model, we cannot trust the results. What we would do next depends
on our a-priori theoretical position. Our choices are (1) to run the models of initial
theoretical interest – a “competing models” strategy, (2) to look at model residuals or
modification indices in order to educate our guess about which path(s) need to be added
to achieve good model fit, and which path(s) should be eliminated because they are not
significant, (3) to make changes until we stumble across a model that shows good fit to
the data, or (4) to include all paths (which would give us perfect fit) and then eliminate
those that are nonsignificant.
In this case, we started with an initial set of alternatives that we wished to compare.
Our interest was equal between the possibility that effects of x1 and x2 on y2 are com-
pletely mediated by y1 , and the possibility that a direct path from either x1 or x2 to y2
might exist. Our most powerful move at this point is to run our next model without
looking at residuals or modification indices. This means we are not letting an interim
look at the data influence our testing process.

Illustration 1.3
Our next move will be to run an alternative model, one that includes a path from x1 to
y2 (refer back to Figure A.1).

Results for Illustration 1.3 – model fit

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 3.548
Degrees of Freedom 1
P-Value 0.0596
RMSEA (Root Mean Square Error of Approximation)
Estimate 0.160
90 Percent C.I. 0.000, 0.353
Probability RMSEA <=0.05 0.092
330 Structural equation modeling and natural systems

Discussion of Illustration 1.3 – model fit

The addition of the path from x1 to y2 resulted in a reduction in the model chi-square
from 57.425 to 3.548. Thus, our single-degree-of-freedom change was 53.877. Since
the change in chi-square also possesses a chi-square distribution, we can treat the dif-
ference as a test criterion. Typically we would judge a decrease in chi-square of 3.84 or
more to indicate a significant improvement in fit when relaxing a single constraint on
the model. Clearly, the data fit the new model substantially better that they did the original
model.
Now, we have judged the second model to be better than the first, but is the fit of
the data to this model adequate or do additional paths need to be added? The absolute
values of the fit indices suggest a reasonably good fit between data and model. The
p-value for the chi-square is 0.0596, while the p-value of the RMSEA is 0.092. We
can also note that the 90% confidence interval for the RMSEA includes the value of
0.0. This means that perfect fit between data and model cannot be ruled out with these
data.
Since our measures of absolute model fit are adequate based on usual criteria, we
conclude that our second model may be acceptable. We might feel sufficiently confident
to simply accept the model. However, if we were of a suspicious mind about this model,
we might worry that a p-value for our chi-square of 0.0596 suggests that with a larger
sample size, we would be able to detect deviations between data and model. There are
other things we can and should examine to evaluate our model, particularly the model
residuals.

Results for Illustration 1.3 – expected covariances, residuals,

and modification indices
Observed Covariances

Y1 Y2 X1 X2
Y1 7.779
Y2 1.829 0.733
X1 1.211 0.652 1.258
X2 2.876 0.545 0.402 3.204

Model Estimated Covariances

Y1 Y2 X1 X2
Y1 7.701
Y2 1.810 0.725
X1 1.198 0.645 1.245
X2 2.847 0.654 0.398 3.172

Residuals for Covariances (observed-estimated)

Y1 Y2 X1 X2
Y1 0.078
Y2 0.019 0.008
X1 0.013 0.007 0.013
X2 0.029 −0.109 0.004 0.032
 Example analyses 331

Standardized Residuals

Y1 Y2 X1 X2
Y1 0.000
Y2 0.000 0.000
X1 0.000 0.000 0.000
X2 0.000 −0.076 0.000 0.000
MODIFICATION INDEX FOR Y2 ON X2 = 3.486

Discussion of Illustration 1.3 – residuals and modification indices

An examination of the estimated variance–covariance matrix points out that an unsat-
urated model imposes a set of expectations on the data. These expectations guide the
process of parameter estimation. However, our best estimates of parameter values typi-
cally still leave us with deviations between observed and estimated/expected values. The
residuals are the deviations between observed and expected. Generally, the standardized
deviations are easier to interpret as these are in correlation metric. It is possible to ask
most programs to calculate the expected decrease in chi-square that should occur if we
were to include a path in the model (these are the so-called modification indices). In this
case, the chi-square change expected for adding a path from x2 to y2 is 3.486. This value
is below the single-degree-of-freedom chi-square change value of 3.84, indicating that
if this change was made the model would not be significantly improved.
The examination of residuals is informative. In this case, it is fairly easy to know
where to look for possible improvements in the model, since there is only one pathway
that is not included in our model. However, for more complex models, residuals and
modification indices can provide valuable clues about where data are failing to match the
expectations of a model. When we use these residuals and indices to find where changes
need to be made for models that are clearly inadequate, we are now applying SEM in an
exploratory mode. In this example, we managed to avoid lapsing into exploratory mode
by focusing on models of initial interest, and basing our evaluation on indices of model
fit. That said, I would rather wind up with a model that matches the data well using an
exploratory process than base my interpretations on a model that obviously does not
match the data. It is really only through repeated applications or very large samples that
we will find out whether our model has general validity.

Results for Illustration 1.3 – parameter estimates

Now, we are finally ready to examine parameter estimates. Some output relating to path
coefficients and other parameters are presented here.
MODEL RESULTS
Estimates S.E. Est./S.E. StdYX
Y1 ON
X1 0.704 0.195 3.608 0.283
X2 0.809 0.122 6.618 0.519
Y2 ON
Y1 0.182 0.016 11.114 0.592
X1 0.343 0.041 8.449 0.450
332 Structural equation modeling and natural systems

0.34
x1 y2 0.175

0.70
0.40 0.18
0.81
x2 y1 4.553
Figure A.2. Unstandardized parameter estimates for accepted model.

Residual Variances
Y1 4.553 0.644 7.071 0.591
Y2 0.175 0.025 7.071 0.241

Discussion of Illustration 1.3 – parameter estimates

The “estimates” presented in this case are the unstandardized parameters. Associated
with each is the normal-theory standard error estimate. If we had reason to believe
that our data failed to meet parametric assumptions, we could choose either ROBUST
estimation, which corrects, at least partly, for nonnormality in the calculation of the
standard errors. Alternatively we could use BOOTSTRAP procedures, which estimate
standard errors using resampling methods. Determinations of whether parameters are
significantly different from zero can be made using the “Est./S.E.” ratios, which approx-
imate a t-distribution. The p-values can be obtained elsewhere for deciding whether
the investigator is confident that the parameter is indeed a nonzero value. For a sample
size of 100, a p-value great than 0.05 is obtained when the t-value is less than 1.984
(two-tailed test). Finally, since we requested them, the standardized parameter values
are presented.
We should note that the tests of parameter significance are based on the unstan-
dardized parameters, and that the standardized parameters are simply summaries of
the model results. We should also note that the R2 values can be calculated simply as
1 minus the standardized error variances. Thus, the R2 for y1 is 0.409 and the R2 for y2
is 0.759.
It is customary and convenient to summarize our results graphically, which is done in
Figure A.2. Here I present the unstandardized parameters to get the reader accustomed
to their use. Since the path coefficients are in absolute units, we can relate results to the
population at large. Considerable discussion in Chapter 3 deals with the interpretation
of coefficients, and the interested reader should refer to that section if needed.
It is often the case that we also examine the standardized results, which are summa-
rized in Figure A.3. Here, the path coefficients are summarized in standard deviation
units. As discussed in Chapter 3, for normally distributed variables, standardized path
coefficients can be approximately interpreted as representing the variation across the
range expected for a response variable, as you vary a predictor across its range. We
must be careful to realize, however, that coefficients standardized based on the sample
standard deviations suffer from some loss of precision compared with unstandardized
 Example analyses 333

0.45
x1 y2 0.24

0.28
0.20 0.59
0.52
x2 y1 0.59
Figure A.3. Standardized parameter estimates for accepted model.

coefficients. This can be remedied by selecting relevant ranges for the variables and
using those to standardize the raw parameters, providing a means of standardization
that does not suffer from reliance on sample variances. Again, see the latter part of
Chapter 3 for a discussion of this option.
The ultimate interpretations the researcher makes concerning the results depend
on many things, including the system being studied, the associated theoretical knowl-
edge and prior experience, the questions being asked, and characteristics of the sample.
Hopefully the main chapters in this book, along with illustrative publications using these
methods, provide sufficient instruction on how to draw careful and meaningful interpre-
tations of results such as these. Basically, my advice is to rely on fundamental statistical
principles rather than someone else’s protocol.

Example 2: a multigroup comparison

The situation
There are many reasons to compare groups. This can be an excellent approach to use
when analyzing experimental data. It can also be used to compare different samples
or populations. Multigroup analyses can allow for the detection of interactions as well
as main effects. For continuity, we will continue with our previous example and now
assume we have a second sample we wish to compare to the original. In such a case, we
wish to know whether the relationships in the two samples are the same, or whether the
relationships in the two samples differ in important ways. We can represent the situation
graphically with a two-group model as shown in Figure A.4.

Some possible applications

Imagine that we performed an experiment in which we randomly assigned fertilizer to
half our plots, and again looked to see how soil factors and plant size relate to frequency
of flowering. In this case, we might wish to ask the basic question of whether the system
responded in a fundamentally different way when it was fertilized. Perhaps we might
think that there should be no important differences in model parameters, although we
might shift the mean plant size to a larger value and likewise shift the incidence of
flowering to a higher proportion. Alternatively, we might think that direct addition of
fertilizer would eliminate the importance of one of our soil factors by removing nutrient
limitation. Or, perhaps we wonder whether fertilization might enhance flowering for
334 Structural equation modeling and natural systems

Sample A

x1a y2a e 2a

x2a y1a e 1a
Sample B

x1b y2b e 2b

x2b y1b e 1b
Figure A.4. Multigroup model of the form obtained in Example 1.

plants of any given size. All these questions can be addressed using the results from a
multigroup analysis.
It is not necessary for our comparison to involve an experimental treatment, of course.
We might again consider a case where we are interested in how herbivorous insects
associate with a particular food plant in the face of covarying habitat factors. This time
we ask whether the species initially studied has the same responses and relationships
to habitat as a second, closely related species also found in the same sample. Here, our
two groups would be independent estimates of each insect species and their relations
to plant density and habitat conditions. In such a situation, our questions may have to
do with differential efficiency of converting food into individuals, differential habitat
preferences, or mortality rates for the two species.

Data
Let us imagine that we have two samples now, our original sample (from Table A.3)
and a second sample. These are shown together in Table A.4. Keep in mind, that while
I show the correlations, the models will be assessed using the covariances.

Results for Illustration 2.1 – model fit and equality constraints

The overall concept in a multigroup model is to determine whether a single model applies
to all data sets. In this case, we have two data sets, each representing a separate group. We
are assuming independence among samples in this example, although that assumption
could be relaxed if need be. Our goal in this analysis is to determine which if any
parameters differ between groups. Parameters in this case include the path coefficients,
 Example analyses 335

Table A.4. Correlations and standard devia-

tions for the two insect species. N = 100 sam-
ples for each group

x1 x2 y1 y2

Group 1
x1 1.0
x2 0.200 1.0
y1 0.387 0.576 1.0
y2 0.679 0.356 0.766 1.0
Mean 6.22 0.756 100.75 10.25
Std. dev. 1.1216 1.79 2.789 0.856
Group 2
x1 1.0
x2 0.212 1.0
y1 0.501 0.550 1.0
y2 0.399 0.285 0.389 1.0
Mean 6.01 0.998 95.56 21.55
Std. dev. 1.32 1.66 2.505 1.701

the variances of the variables, and the means for the variables. There is a particular
sequence of evaluations usually recommended for determining whether parameters are
equal across groups. Stated briefly, our strategy is to begin by allowing all parameters to
be uniquely estimated for each group. It is a basic requirement that the models be of the
same form, which is something that is tested by allowing all parameters to be unique,
and then determining whether model fit is adequate for each model separately. In this
case, our data do fit a common model.

Discussion of Illustration 2.1 – model fit and equality constraints

Once we have determined that both data sets fit a common model, we then begin a series
of single-degree-of-freedom chi-square tests by setting individual parameters to be equal
across groups. Some programs (EQS) have the ability to perform one multivariate test
for all equalities using Lagrange multipliers. In this case, we illustrate the method of
multigroup analysis using the more commonly used univariate stepwise procedure.
Table A.5 gives the results for our series of equality tests. Our baseline model allowing
all parameters to be unequal across groups had a chi-square of 4.981 with two degrees of
freedom. The two degrees of freedom come from the fact that the paths from x2 to y2 for
the two groups were not specified in the base model. When we forced path coefficients to
be equal across groups one at a time, there was a slight, but nonsignificant increase in
the model chi-square with each additional degree of freedom. Therefore, we conclude
that all pathways have the same quantitative slopes across the two groups. When we
forced variances to be equal across groups, only the variances of y2 were found to be
different among the groups. An indication of this difference can be seen in Table A.4,
where the standard deviation of y2 in group 1 can be seen to be 0.856, while the standard
336 Structural equation modeling and natural systems

Table A.5. Summary of tests for equality of parameters across groups. The base
model is one where all parameters are unique for each group. As parameters
are constrained to be equal across groups, we determine whether model chi-
square increases significantly, if so, the parameter is not equal across groups.
If chi-square does not increase significantly, parameters are judged to be indis-
tinguishable across groups

Chi-square Model df Equal across groups?

Base model 4.981 2
Pathways: x1 → y1 5.042 3 equal
x2 → y1 5.413 4 equal
x1 → y2 5.416 5 equal
x2 → y2 5.433 6 equal
Variances: x1 8.079 7 equal
x2 8.752 8 equal
y1 10.995 9 equal
y2 143.606 10 not equal

Means: x1 12.474 10 equal

x2 14.076 11 equal
y1 211.882 12 not equal
y2 695.039 12 not equal

deviation of y2 in group 2 is 1.701. Finally, two of the means were found to be equal
across groups, those for x1 and x2 . When the means for y1 and y2 were constrained to
be equal across groups, model chi-square was greatly inflated and overall model fit was
poor.

Results for Illustration 2.1 – parameter estimates

In Figure A.5, I show both the unstandardized and standardized path coefficients. Both
types of results are presented here to make a very important point about interpreting
multigroup analyses. First, we should note that all unstandardized path coefficients
were found to be equal across the two groups (Table A.5). Thus, we may judge that
the structural parameters defining the effects in this model are constant. Upon close
inspection of Figure A.5, you will notice that some of the standardized path coefficients
are rather different between groups, despite the fact that the unstandardized coefficients
are judged identical. In particular, this can be seen for the paths from x1 to y2 , and from
y1 to y2 . Why is this the case? The answer lies in the fact that the variances differ among
groups for y2 . More specifically, the variance of y2 in group 2 is roughly double that
in group 1. Since standardized coefficients are standardized by the standard deviations,
their values are “contaminated” by the unequal variances among groups. There is also an
effect of unequal variances on the error variance of y2 ; in this case both the unstandardized
and standardized error variances are affected.
These results are designed to bring home the points mentioned earlier about the
challenges to interpreting results based on standardized coefficients. This does not mean
 Example analyses 337

Sample A
0.34/0.48
x1a y2a 0.18/0.23

0.74/0.34
0.42/0.20
0.18/0.55
0.76/0.50

x2a y1a 3.97/0.57

Sample B
0.34/0.25
x1b y2b 2.27/0.80

0.74/0.34
0.42/0.20
0.18/0.28
0.76/0.50

x2b y1b 3.97/0.57

Figure A.5. Results from multigroup analysis. Coefficients presented are the
unstandardized values followed by the standardized values. The parameters that
were found to be significantly different among groups are summarized in Table
A.5 and only included the variance of y2 and the means of y1 and y2 .

that all comparisons are automatically incorrect. It is only when variances differ among
groups that we will have a problem. In fact, since the declaration of significant differences
among groups is based on tests of the unstandardized parameters, we know that the
standardized path coefficients in Figure A.5, even those that appear to be different, do
not represent significant differences. If those paths did differ significantly and there were
also differences in variances among groups, then the standardized coefficients could lead
to some incorrect inferences. Again, the solution mentioned at the end of Chapter 3 to
standardize based on relevant ranges would solve the problem, and if applied to this
example, we would see that all such standardized path coefficients would be equal
across groups.
One other point that is made by this example to illustrate why statisticians are not
enamored with judging a model by the R2 values. The R2 for y2 in group 1 is 0.769,
while the R2 for y2 in group 2 is only 0.204. Would we judge the model for group 2 to
be inferior to the one for group 1? We should not, because the path coefficient estimates
are the same for both. It may be that we actually obtained the correct path coefficients in
group 2 in the face of more noise, which could be construed as a major success. There
is a natural attraction for researchers to focus on explained variance in the response
variables. As we see here, it is not one that should be accepted too readily, as it focuses
on the wrong parameters.
338 Structural equation modeling and natural systems

Table A.6. Data borrowed from Table A.4, with the

added component that a dichotomous cluster vari-
able is now included. N = 100

c x1 x2 y1 y2
c 1.0
x1 0.25 1.0
x2 0.41 0.200 1.0
y1 0.33 0.387 0.576 1.0
y2 0.37 0.679 0.356 0.766 1.0
Mean 0.5 6.22 0.756 100.75 10.25
Std. dev. 0.20 1.1216 1.79 2.789 0.856

Example 3: hierarchical data

The situation
It is extremely common for data to be sampled in some fashion that restricts random-
ization. A multigroup analysis can be used when we are explicitly interested in different
subgroups of our sample. However, in many cases, we are not interested in the sam-
ple clusters, but instead, simply wish to control for their effects. There are two basic
approaches we can use, one is to include cluster variables and the other is to perform
behind-the-scenes adjustments to the data. Both can be effective.

Some possible applications

For experimental data, use of a random block design creates a structure in the data that
needs to be taken into account in the analysis. Likewise, stratified random sampling, in
which coverage of a range of conditions is enhanced by defining strata and randomly
sampling within those strata, is a commonly used device. Even when not used as a
strategy per se, the subsampling of disjunct locations or populations creates a kind of
structure in the data that needs to be accounted for. Sometimes the researcher is interested
in the magnitude of this effect, and sometimes it is the across-strata results that are of
greatest interest.

Data
For this example, we continue using the data from Example 1, which was utilized as
the data for group 1 in Table A.4. In this case, we imagine that these data were actually
taken from two disjunct sampling locations and combined, and we now wish to control
for this cluster sampling in the analyses. It would be our general assumption that if the
clusters differed consistently in some fashion, ignoring them should create bias in the
parameter estimates observed in Figure A.2.

Results
Here we will not worry about the steps involved in the process of arriving at the final
results, as the reader should now be familiar with the process. Instead, we focus on
 Example analyses 339

A
0.34
x1 y2 0.175

0.70
0.40 0.18
0.81
x2 y1 4.55

B
0.33
x1 y2 0.155
0.62
0.40 0.19
−0.07

x2 0.70
y1 4.35

Figure A.6. Analysis results either A. ignoring clustering in the data, or B. adjust-
ing for clustering.

the results obtained. For comparison, we will refer to the results in Figure A.2, which
represent what one would obtain here if we ignored the clustering. When our model was
modified to include the cluster variable, the first thing that was found was that model
fit was substantially reduced. Instead of a chi-square of 3.548 with 1 degree of freedom
(p = 0.0596), we now have a chi-square of 7.803 with 2 degrees of freedom (p = 0.0197).
Inclusion of a path from x2 to y2 was indicated by these deviations, which could lead us
to accept a different model from these same data once clustering is taken into account.
The model with the path from x2 to y2 included is found to have a chi-square of 1.226
with 1 degree of freedom (p = 0.2681), indicating good fit.
Results from the analysis including a cluster variable can be seen in Figure A.6.
For simplicity, the paths involving the cluster variable are omitted from the presenta-
tion. The principal change in results compared to when the effects of clustering were
ignored (Figure A.2), is the inclusion of the path from x2 to y2 . Since these are the
unstandardized results, which seems like the appropriate metric to use when comparing
non-nested models as we are here, it is a little more challenging to judge the magnitude
of the differences in parameters. We can improve our perspective a little by express-
ing the differences in path coefficients as percentage changes, and they range from
0% up to a maximum of 17% (for the path from x2 to y1 ). The take home message
here is that it is both possible and desirable to control for nonindependencies in the
data.
340 Structural equation modeling and natural systems

x1
z2 y4

x2 x1 h2 y5

x3 y6
h1
z1

y1 y2 y3
Figure A.7. Latent variable model with multiple indicators. Error terms for indi-
cators are omitted for simplicity of presentation.

Example 4: evaluating latent variable models

The situation
Latent variables can be incorporated into models in a great many different ways. Here
I will offer a single illustrative example designed to cover some of the more common
issues, as well as some more specialized points. Figure A.7 shows a classic “hybrid”
model that considers directed relations among latent factors possessing multiple indi-
cators. In this model we assume that the entities of theoretical importance are the latent
factors and it is these that interact (i.e., no direct interactions among indicators is spec-
ified). We also assume that we have appropriate indicators for each latent variable and
that the indicators possess the properties of validity and reliability. Chapter 4 provides
an extended discussion of the properties of latent variables. Additional references giving
more detail on this subject include Bollen (1984) and Bollen and Ting (2000).

Some possible applications

It would seem certain subject areas, such as population biology, might be fertile ground
for latent variables. Individual organisms, be they whales, mites, or macroalgae, have
correlated sets of attributes that are presumably reflective of underlying traits we cannot
measure directly. Perhaps the model in Figure A.7 might apply to a case where ξ 1 is
body size, η1 is a male bird’s territory range, and η2 is that bird’s reproductive success.
The indicators in such a case would be those that we believe adequately reflect the
latent factors of interest. In truth, any situation where we believe (1) that unmeasured
variables lie at the heart of the interactions of interest, and (2) we have various ways of
approximating the properties of those unmeasured variables, will potentially be suitable
for the use of multi-indicator latent variable models.
 Example analyses 341

Table A.7. Correlations among indicators associated with latent variable

example. s.d. refers to standard deviations. N = 100

x1 x2 x3 y1 y2 y3 y4 y5 y6
x1 1.0
x2 −0.86 1.0
x3 −0.64 0.70 1.0
y1 −0.23 0.25 0.24 1.0
y2 −0.19 0.20 0.19 0.64 1.0
y3 −0.22 0.24 0.23 0.77 0.62 1.0
y4 −0.19 0.20 0.20 0.38 0.31 0.71 1.0
y5 −0.16 0.18 0.17 0.33 0.27 0.62 0.72 1.0
y6 −0.17 0.19 0.18 0.35 0.28 0.66 0.76 0.66 1.0
s.d. 2.9 1.1 0.06 0.9 1.0 2.3 1.6 0.8 1.9

Data
I begin by throwing another curve at the reader. One that represents an easy programming
mistake to make which has fatal consequences. We will then correct the data and proceed
through the model analyses.

Results – initial analyses of data in Table A.7

It is convention that the analysis of hybrid models begins with a test of the measurement
model. This is accomplished by allowing the latent variables to intercorrelate freely,
thereby saturating the structural model (i.e., specifying all possible paths among latents).
When we do this with the data in Table A.7, we obtain the following:
NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED.
We can get some idea why our model failed by examining the initial estimates, which tell
us about the direction the iterations were heading when the program gave up. One part of
the printout looks immediately suspicious. The notation here refers to the measurement
of the latent variable XI-1 by the indicators X1, X2, and X3.
MODEL RESULTS

Estimates
XI1 BY
X1 1.000
X2 ********
X3 8697.837

These results describe the initial estimates of the loadings for the three indicators on
the latent variable ξ 1 (“XI-one”). The loading for x1 is set at 1.0 automatically by the
program in order to set the scale for the latent variable (the scale for latent variables must
always be specified in some fashion). The results presented indicate that the program
was unable to estimate a loading for the path from ξ 1 to x2 , and that the loading for x3
was very unrealistic. To see what a more normal pattern would look like, here we see
the interim results for η1 .
342 Structural equation modeling and natural systems

ETA1 BY
Y1 1.000
Y2 0.894
Y3 2.474

There are a number of things we might try at this point to get started. Since SEM
programs use iterative algorithms, we might specify starting values. We can also increase
the number of iterations attempted. None of these changes will help in this case, however.
Another approach is to set the scale for the latent variable using a different indicator.
When we do this, we achieve the following:
THE MODEL ESTIMATION TERMINATED NORMALLY
If we now examine the estimates for loadings for ξ 1 , we find

XI1 BY
X1 −2.418
X2 1.000
X3 0.041

So, we obtained seemingly reasonable estimates. However, we see that one has a negative
loading (X1) and the others are positive. Programs vary in how well they are able to
handle this situation where indicators are not coded to correlate positively. In general,
this is to be avoided. So, to eliminate this as a potential source of problems, we recode our
original data in Table A.7 so that the indicators X2, X2, and X3 are positively correlated.
Our new matrix of correlations is given in Table A.8.
This may be a good time to comment on another problem that frequently arises in
analyses of latent variable models, the so-called Heywood case. A Heywood case refers
to a solution that is inadmissible, in that some parameter estimates have illogical values.
The most common such situation is when one obtains negative error variance estimates.
Sometimes this can be solved by recoding the data or by changing the indicator used to fix
the scale of the latent variable. Another approach is to constrain the estimate to be greater
than or equal to zero (using programming constraints). However, in recalcitrant situations
where this results in unacceptable solutions, one further alternative is to specify a small,
nonzero, positive value for the error variance that keeps coming up with a negative
estimate. This approach invalidates our estimates of loadings and residual variances
for the indicators on that latent variable. However, it at least allows us to have an
admissible solution with which to proceed. It is hoped that the researcher will eventually
identify the true source of the problem and resolve it through a reformulation of the
model.

New data
Further results and discussion – evaluation of the measurement model
As stated above, our evaluation of a hybrid model is usually performed as a two-stage
process, with the first stage being a test of the measurement model. As stated above,
this is accomplished by modifying the model in Figure A.7 so that the latent variables
intercorrelate freely, thus allowing us to focus on the adequacy of our measurement
model in a separate analysis. When we run such a model using the data in Table A.8,
we get the following message:
 Example analyses 343

Table A.8. Data were recoded from Table A.7 so that indicators for all latent
variables are positively correlated. N = 100

x1 x2 x3 y1 y2 y3 y4 y5 y6
x1 1.0
x2 0.86 1.0
x3 0.64 0.70 1.0
y1 0.23 0.25 0.24 1.0
y2 0.19 0.20 0.19 0.64 1.0
y3 0.22 0.24 0.23 0.77 0.62 1.0
y4 0.19 0.20 0.20 0.38 0.31 0.71 1.0
y5 0.16 0.18 0.17 0.33 0.27 0.62 0.72 1.0
y6 0.17 0.19 0.18 0.35 0.28 0.66 0.76 0.66 1.0
s.d. 2.9 1.1 0.06 0.9 1.0 2.3 1.6 0.8 1.9

THE MODEL ESTIMATION TERMINATED NORMALLY WARNING:

THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE.
PROBLEM INVOLVING VARIABLE Y3
A nonpositive definite matrix means that something does not add up; in this case,
it relates specifically to the theta matrix, which contains information about the residual
covariances among indicators. This is a serious problem that means our results cannot
be trusted. We can obtain more information from the printout to help us spot the source
of the problem. If we examine the residuals for the indicator y3 , we find the highest
values for its covariances with y4 and y6 . In and of themselves, these do not provide
us with enough of a clue as to the possible misspecification we are dealing with. It is
important at this point to consider some of the reasons we might have a problem in a
measurement model. Some of the possibilities might include (1) inadequacies in the data,
(2) unmeasured influences affecting indicators, and (3) incorrect causal specification. We
will ignore the first of these here. Our data have been selected to represent an underlying
structure and we accept for the moment their adequacy.
The second of the above possibilities (unmeasured influences) would suggest a cor-
related error that represents the effects of some unmeasured factor on two or more of our
indicators. In this case, the residuals represent linkages between indicators loading on
different latent variables. It is certainly possible that some unmeasured factor is causing
us to have correlations among the indicators of different latent variables. We can look
further for clues to see if such error correlations could cause our lack of fit. If we request
modification indices, they suggest the following as ways of improving model fit.
MODEL MODIFICATION INDICES
BY Statements M.I.
ETA2 BY Y1 8.887
ETA2 BY Y3 23.864

WITH Statements M.I.

Y2 WITH Y1 21.814
Y3 WITH Y2 8.153
Note that the lingo in which these results are expressed is such that “BY” is shorthand for
“measured by” and “WITH” is shorthand for “correlated with”. The modification indices
344 Structural equation modeling and natural systems

suggest that if we allow the errors of Y2 and Y1 to correlate (which is indistinguishable

from allowing an additional correlation between Y2 and Y1), we would reduce our
chi-square by 21.814. This is an interesting possibility to consider. We see elevated
residuals between Y3 and Y4 and between Y3 and Y6, which does not fit logically
with the suggestion of allowing Y1 and Y2 to correlate. Furthermore, if we examine the
correlations between Y1, Y2, and Y3 in Table A.8, we do not see anything imbalanced in
the triad of correlations. We might still be tempted to see if we can solve our estimation
problem by including a correlated error between Y1 and Y2. Our evidence does not
indicate that Y1 and Y2 are unusually correlated, however, suggesting that we should
look elsewhere for our clues about the problem. If, despite the warning signs, one
includes a correlation between the error of Y1 and Y2, just to see if this would indeed fix
our estimation problem, the answer is “no”. We still have a nonpositive definite matrix
indicative of unresolved relationships involving Y3.
Returning to our modification indices, these statements indicate that the largest reduc-
tion in model chi-square could be obtained by allowing Y3 to load on ETA2. We might be
sufficiently frustrated at this point to simply make that change. That could be a fatal error
unless theory suggests that as a reasonable way to reformulate our model. So, instead,
we ask ourselves, does it make substantive sense that Y3 could be an indicator for both
ETA1 and ETA2? Typically, the answer to this question will be “no”. Crossloading indi-
cators can sometimes make sense, but in a model such as this one where the indicators
have been selected for the purpose of representing our separate constructs, we would
hope we had a sufficiently clear rationale that we could obtain good indicators in the first
place. For example, if Y3 was an indicator of territory size, like singing frequency and
ETA2 is reproductive success, it would not make sense to equate singing frequency with
indicators of reproductive success, such as number of eggs produced by the bird’s mate.
Now, I have made this example tricky for a reason, we need to take to heart the
message at the end of Chapter 4 that the causal logic in the standard hybrid model (as
shown in Figure A.7) does not necessarily represent the true situation. In Figures 4.18 and
4.19 I describe some other possible types of misspecification that involve direct causal
effects by individual indicators. This represents our third category of possibilities. Can
we find clues as to a specific effect of an indicator in the output? Since the modification
indices suggest a linkage between ETA2 and Y3 and the residuals suggest unspecified
relationships between Y3 and Y4 and between Y3 and Y6, perhaps there is a specific
effect of Y3 on ETA2. There are other possibilities, but what is being raised here is
that perhaps the specific attribute associated with Y3 has, for example, a specific benefit
on reproductive success. If that were the case, this influence should be included in the
model. If we include this path, we now obtain a proper solution, as well as a good fit of
data to model,
TESTS OF MODEL FIT
Chi-Square Test of Model Fit
Value 23.315
Degrees of Freedom 25
P-Value 0.5592
Subsequent examination of parameter estimates and residuals reveal no signs of any
problems. We have now obtained an adequate measurement model. While the results
are not presented here, the form of this model is as shown in Figure A.8.
 Example analyses 345

x1
z2 y4

x2 x1 h2 y5

x3 y6
h1
z1

y1 y2 y3
Figure A.8. Modified measurement model evaluated in stage one of the analysis.
In measurement model evaluations, all latent variables are allowed to intercorre-
late freely, so that the measurement relations are the focus of the analysis. The
modification made was to specify an effect of y3 on η2 .

Evaluation of the full model

We now move on to consider the full model, which includes the hypothesis of whether
the data are consistent with the structural relations (directed relations among latents)
specified in Figure A.7. Omitting the path from ξ 1 to η2 , we obtain a model chi-square
of 3.517 with 24 degrees of freedom, which is a nearly perfect fit. We can be pretty
confident now that our data fit our model. The results are shown in Figure A.9.
We will not take the time to interpret the results obtained. Hopefully the reader
has sufficient grasp of the issues to realize both the subtleties involved in interpreting
SEM results and the role of theory in guiding interpretations, which will depend on the
particular system under study.

Example 5: models including composite variables

The situation
Chapter 6 goes into great detail discussing the considerations involved in deciding when
to use composite variables. My purpose here is simply to walk through the stages of
an analysis so the reader can see the mechanics of the operation. As the reader may
be aware from Chapter 6, there are problems in solving models containing composites,
because the composites introduce new parameters to be estimated. We generally set the
error variances for composites to zero and also set the scale for the composites by fixing
a path from one of the component causes. While these specifications allow the model to
be solved, there is a loss of information associated with fixing the path from one of the
component causes to the composite. Basically, the problem is that because the path is
specified at a fixed value by the investigator, it cannot be tested. The resolution of this is
346 Structural equation modeling and natural systems

0.23
x1 0.89 0.91 y4

0.97 0.79
x2 x1 h2 y5
0.72 0.84
0.27 −0.77
x3 y6
0.93
h1 1.45
0.90 0.72 0.86

y1 y2 y3
Figure A.9. Results obtained for modified latent variable model. Errors for indi-
cators are omitted to simplify presentation. Coefficients shown are standardized
values.

to use a two-stage analysis with the composite omitted from the model in the first stage.
This process will be illustrated below.

Some possible applications

I find that there are a great variety of situations when I might wish to incorporate a
composite variable, usually so as to achieve a more general perspective on findings. For
example, in the immediately preceding illustration (Figure A.9), we find two influences
of ξ 1 on η1 , one general and one specific. The path coefficients from ξ 1 to η1 and from
y3 to η1 are challenging to interpret because ξ 1 and y3 are highly intercorrelated (λ =
0.86). We might wish to know the combined effects of ξ 1 and y3 on η1 . This could
be accomplished using a composite variable representing their combined influences. In
an alternative situation, we might have a whole suite of specific effects that we wish
to compare to some other set of effects, for example the combined influences of soil
factors compared to the combined influences of disturbance factors on plant growth.
Composites could be used here as well. Finally, I typically find it necessary to use
composites when modeling nonlinear relationships. The individual path coefficients for
first and second order terms in a polynomial regression are rather uninterpretable. How
are we to examine the sensitivity of y to variations in x when holding x2 constant? Since
x2 is a direct function of x, it is not possible to hold it constant while varying x, except in
a very abstract sense. What we wish to know is the combined influences of x and x2 on
y. A composite can allow us to estimate such an effect. Here I present a simple example
involving a set of exogenous predictors and a single response variable (Figure A.10) to
illustrate the two-stage nature of the modeling process.
 Example analyses 347

Table A.9. Correlations among variables in

composite model. N = 100

x1 x2 x3 x4 y1
x1 1.0
x2 0.25 1.0
x3 0.02 0.26 1.0
x4 0.12 0.31 0.25 1.0
y1 0.55 0.47 0.42 0.41 1.0
Std. dev. 1.23 0.66 2.30 1.56 2.09

x1
0

x2
c y1 e1
x3

x4
Figure A.10. Model in which a composite variable is used to represent the col-
lective effects of x1 –x3 on y1 .

Data
Results and discussion of first stage of analysis of composite model
As stated above, the first stage in the analysis is to omit the composite from the model
and evaluate the component processes whose effects will be composited. In this par-
ticular case, we simply have a saturated model equivalent to a multiple regression
(Figure A.11). Since the model is saturated, the fit is perfect, thus, our evaluation is
quantitative – determining the magnitudes of the path coefficients, their significance,
and the variance explained by the model.

Results and discussion from second stage of analysis

The second stage of the analysis includes a so-called “sheaf” coefficient (Heise 1972)
from the composite to y1 representing the collective effect of x1 –x3 . It is most common
that we present standardized coefficients in models involving composites because of
the difficulties associated with interpreting unstandardized parameters for composites
(Figure A.12). We omit the coefficients from the components to the composite in this
348 Structural equation modeling and natural systems

x1 0.46

x2 0.21

0.30 y1 0.43
x3
0.21

x4
Figure A.11. Results from first stage of analysis of composite model.

x1
0

x2 0.66
c y1 0.43
x3
0.21
x4
Figure A.12. Results from second stage of analysis of composite model. Coeffi-
cients shown are standardized values.

presentation because we wish to focus on the net effect. One can see that aside from
compositing the effects of x1 –x3 , the other results are the same as those from stage 1 of
the analysis. Now we can consider the collective effect of x1 –x3 , giving us flexibility to
provide results that match with our constructs of theoretical interest.

Conclusions
A facile use of SEM comes from a reliance on fundamental statistical and scientific
principles, not from blindly following conventions. One way (actually the best way) to
develop confidence in one’s analysis is to generate simulated data where you control
what relationships are built into the data. Then you can analyze those data in various
ways to see what you get out. It is extremely instructive if you use conventional univariate
 Example analyses 349

analyses in comparison with multivariate ones to gain a respect for how far off the mark
univariate results can be. You will also find out that ignoring certain features of the
situation in multivariate models (e.g., data clustering) can have either subtle or not so
subtle effects. Several of the SEM programs have the capability to generate simulated
data. It is also possible to simulate data sets using any number of other means. You
will gain both an appreciation for the assumptions that lie behind the analyses and an
appreciation for SEM. Try it, you will be glad you did.
 References

Aarssen, L. W. & Schamp, B. S. (2002). Predicting distributions of species richness and

species size in regional floras: Applying the species pool hypothesis to the habitat
templet model. Perspectives in Plant Ecology, Evolution and Systematics, 5, 3–12.
Abelson, R. P. (1995). Statistics as Principled Argument. Hillsdale, NJ: Lawrence
Erlbaum Publishers.
Abrams, P. A. (1995). Monotonic or unimodal diversity – productivity gradients: what
does competition theory predict? Ecology, 76, 2019–2027.
Agrawal, R., Imielienski, T., & Swami, A. (1993). In: Mining association rules between
sets of items in large databases. Proceedings of Conference on Management of
Data. New York: ACM Press.
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions
on Automatic Control AC, 19, 716–723.
Al-Mufti, M. M., Sydes, C. L., Furness, S. B., Grime, J. P., & Band, S. R. (1977). A
quantitative analysis of shoot phenology and dominance in herbaceous vegetation.
Journal of Ecology, 65, 759–791.
Andersen, J. A. (1995). An Introduction to Neural Networks. Cambridge, MA: MIT
Press.
Anderson, D. R., Burnham, K. P., & Thompson, W. L. (2000). Null hypothesis testing:
problems, prevalence, and an alternative. Journal of Wildlife Management, 64,
912–923.
Bacon, F. (1620). Novum Organum. London: Bonham Norton and John Bill.
Baldwin, H. Q. (2005). Effects of fire on home range size, site fidelity, and habitat asso-
ciations of grassland birds overwintering in southeast Texas. M.S. thesis, Louisiana
State University, Baton Rouge.
Bayes, T. (1763). An essay towards solving a problem in the doctrine of chances. Philo-
sophical Transactions of the Royal Society of London, 53, 370–418.
Bezdek, J. C. & Pal, N. (1992). Fuzzy Models for Pattern Recognition. New York: IEEE
Press.
Blalock, H. M. (1964). Causal Inferences in Nonexperimental Research. Chapel Hill,
NC: University of North Carolina Press.
Bollen, K. A. (1984). Multiple indicators: internal consistency or no necessary relation-
ship. Quality and Quantity, 18, 377–385.

350
 References 351

Bollen, K. A. (1989). Structural Equations with Latent Variables. New York: John Wiley
& Sons.
Bollen, K. A. (1996). An alternative 2SLS estimator for latent variable models. Psy-
chometrika, 61, 109–121.
Bollen, K. A. (1998). Path analysis. pp. 3280–3284. In: Encyclopedia of Biostatistics.
P. Armitage and T. Colton (eds.). New York: John Wiley & Sons.
Bollen, K. A. (2002). Latent variables in psychology and the social sciences. Annual
Review of Psychology, 53, 605–634.
Bollen, K. A. & Lennox, R. (1991). Conventional wisdom on measurement: a structural
equation perspective. Psychological Bulletin, 110, 305–314.
Bollen, K. A. & Long, J. S. (eds.) (1993). Testing Structural Equation Models. Newbury
Park, CA: Sage Publications.
Bollen, K. A. & Stine, R. (1992). Bootstrapping goodness of fit measures in structural
equation models. Sociological Methods and Research, 21, 205–229.
Bollen, K. A. & Ting, K. (2000). A tetrad test for causal indicators. Psychological
Methods, 5, 3–22.
Borgelt, C. & Kruse, R. (2002). Graphical Models. New York: John Wiley & Sons.
Bozdogan, H. (1987). Model selection and Akaike’s Information Criterion (AIC).
Psychometrika, 52, 345–370.
Brewer, J. S. & Grace, J. B. (1990). Vegetation structure of an oligohaline tidal marsh.
Vegetatio, 90, 93–107.
Browne, M. W. & Cudeck, R. (1989). Single sample cross-validation indices for covari-
ance structures. Multivariate Behavioral Research, 24, 445–455.
Burnham, K. P. & Anderson, D. R. (2002). Model Selection and Multimodel Inference.
Second Edition. New York: Springer Verlag.
Byrne, B. M. (1994). Structural Equation Modeling EQS and EQS/Windows. Thousand
Oaks, CA: Sage Publications.
Byrne, B. M. (1998). Structural Equation Modeling with LISREL, PRELIS, and
SIMPLIS. Mahway, NJ: Lawrence Erlbaum Associates.
Byrne, B. M. (2001). Structural Equation Modeling with AMOS. Mahway, NJ: Lawrence
Erlbaum Associates.
Campbell, D. R., Waser, N. M., Price, M. V., Lynch, E. A., & Mitchell, R. J. (1991).
A mechanistic analysis of phenotypic selection: pollen export and corolla width in
Ipomopsis aggregata. Evolution, 43, 1444–1455.
Casella, B. (1992). Explaining the Gibbs sampler. The American Statistician, 46, 167–
174.
Congdon, P. (2001). Bayesian Statistical Modeling. Chichester: Wiley Publishers.
Congdon, P. (2003). Applied Bayesian Modeling. Chichester: Wiley Publishers.
Cottingham, K. L., Lennon, J. T., & Brown, B. L. (2005). Knowing when to draw the
line: designing more informative ecological experiments. Frontiers in Ecology, 3,
145–152.
Cudeck, R., Du Toit, S. H. C., & Sörbom, D. (eds.) (2001). Structural Equation
Modeling: Present and Future. Lincolnwood, IL: SSI Scientific Software Inter-
national.
Dasarathy, B. V. (1990). Nearest Neighbor (NN) Norms: NN Pattern Classification
Techniques. Los Alamitos, CA: IEEE Computer Science Press.
352 References

Diamantopoulous, A. & Winklhofer, H. M. (2001). Index construction with formative

indicators: an alternative to scale development. Journal of Marketing Research, 38,
269–277.
Duncan, T. E., Duncan, S. C., Strycker, L. A., Li, F., & Alpert, A. (1999). An Introduction
to Latent Variable Growth Curve Modeling: Concepts, Issues, and Applications.
Mahwah, NJ: Lawrence Erlbaum Associates.
Edwards, J. R. (2001). Multidimensional constructs in organizational behavior research:
an integrative analytical framework. Organizational Research Methods, 4, 144–
192.
Fan, X., Thompson, B., & Wang, L. (1999). Effects of sample size, estimation methods
and model specification on structural equation modeling fit indexes. Structural
Equation Modeling, 6, 56–83.
Fisher, R. A. (1956). Statistical Methods and Scientific Inference. Edinburgh, UK: Oliver
and Boyd.
Fornell, C., ed. (1982). A Second Generation of Multivariate Analyses: Volumes 1 and
II. New York: Praeger Publishers.
Gadgil, M. & Solbrig, O. T. (1972). The concept of r- and K-selection: evidence
from wild flowers and some theoretical considerations. American Naturalist, 106,
14–31.
Gelman, A., Carlin, J. B., Stern, H. S., & Rubin, D. B. (2004). Bayesian Data Analysis.
Boca Raton: Chapman & Hall.
Glymour, B., Scheines, R., Spirtes, R., & Kelly, K. (1987). Discovering Causal
Structure: Artificial Intelligence, Philosophy of Science, and Statistical Modeling.
Orlando, FL: Academic Press.
Goldberger, A. S. & Duncan, O. D. (1973). Structural Equation Models in the Social
Sciences. New York: Seminar Press.
Good, I. J. (1983). Good Thinking. Minneapolis: University of Minnesota Press.
Gough, L. & Grace, J. B. (1999). Predicting effects of environmental change on plant
species density: experimental evaluations in a coastal wetland. Ecology, 80, 882–
890.
Gough, L., Grace, J. B., & Taylor, K. L. (1994). The relationship between species richness
and community biomass: the importance of environmental variables. Oikos, 70,
271–279.
Grace, J. B. (1991). A clarification of the debate between Grime and Tilman. Functional
Ecology, 5, 503–507.
Grace, J. B. (1999). The factors controlling species density in herbaceous plant
communities: an assessment. Perspectives in Plant Ecology, Evolution and Sys-
tematics, 2, 1–28.
Grace, J. B. (2001). The roles of community biomass and species pools in the regulation
of plant diversity. Oikos, 92, 191–207.
Grace, J. B. (2003a). Comparing groups using structural equations. chapter 11, pp. 281–
296. In: B. H. Pugesek, A. Tomer, & A. von Eye (eds.). Structural Equation
Modeling. Cambridge: Cambridge University Press.
Grace, J. B. (2003b). Examining the relationship between environmental variables and
ordination axes using latent variables and structural equation modeling. chapter
7, pp. 171–193. In: B. H. Pugesek, A. Tomer, & A. von Eye (eds.). Structural
Equation Modeling. Cambridge: Cambridge University Press.
 References 353

Grace, J. B. & Bollen, K. A. (2005). Interpreting the results from multiple regression
and structural equation models. Bulletin of the Ecological Society of America, 86,
283–295.
Grace, J. B. & Guntenspergen, G. R. (1999). The effects of landscape position on
plant species density: evidence of past environmental effects in a coastal wetland.
Ecoscience, 6, 381–391.
Grace, J. B. & Jutila, H. (1999). The relationship between species density and community
biomass in grazed and ungrazed coastal meadows. Oikos, 85, 398–408.
Grace, J. B. & Keeley, J. E. (2006). A structural equation model analysis of post-
fire plant diversity in California shrublands. Ecological Applications, 16, 503–
514.
Grace, J. B. & Pugesek, B. (1997). A structural equation model of plant species richness
and its application to a coastal wetland. American Naturalist, 149, 436–460.
Grace, J. B. & Pugesek, B. H. (1998). On the use of path analysis and related procedures
for the investigation of ecological problems. American Naturalist 152, 151–159.
Grace, J. B., Allain, L., & Allen, C. (2000). Factors associated with plant species richness
in a coastal tall-grass prairie. Journal of Vegetation Science, 11, 443–452.
Grime, J. P. (1973). Competitive exclusion in herbaceous vegetation. Nature, 242, 344–
347.
Grime, J. P. (1977). Evidence for the existence of three primary strategies in plants
and its relevance to ecological and evolutionary theory. American Naturalist, 111,
1169–1194.
Grime, J. P. (1979). Plant Strategies and Vegetation Processes. London: John Wiley &
Sons.
Grime, J. P. (2001). Plant Strategies, Vegetation Processes, and Ecosystem Properties.
London: John Wiley & Sons.
Grime, J. P. (2002). Declining plant diversity: empty niches or functional shifts? Journal
of Vegetation Science, 13, 457–460.
Grimm, V. (1994). Mathematical models and understanding in ecology. Ecological Mod-
elling, 74, 641–651.
Gross, K. L., Willig, M. R., & Gough, L. (2000). Patterns of species density and pro-
ductivity at different spatial scales in herbaceous plant communities. Oikos, 89,
417–427.
Grubb, P. J. (1998). A reassessment of the strategies of plants which cope with shortages
of resources. Perspectives in Plant Ecology, Evolution and Systematics, 1, 3–31.
Hägglund, G. (2001). Milestones in the history of factor analysis. pp. 11–38. In: R.
Cudeck, S. H. C. Du Toit, & D. Sörbom (eds.). Structural Equation Modeling:
Present and Future. Lincolnwood, IL: SSI Scientific Software International.
Hair, J. F., Jr., Anderson, R. E., Tatham, R. L., & Black, W. C. (1995). Multivariate Data
Analysis. Fourth Edition. Englewood Cliffs, NJ: Prentice Hall.
Hannon, B. & Ruth, M. (1997). Modeling Dynamic Biological Systems. New York:
Springer.
Hargens, L. L. (1976). A note on standardized coefficients as structural parameters.
Sociological Methods & Research, 5, 247–256.
Harrison, S., Safford, H. D., Grace, J. B., Viers, J. H., & Davies, K. F. (2006). Regional
and local species richness in an insular environment: serpentine plants in California.
Ecological Monographs, 76, 41–56.
354 References

Hastings, W. (1970). Monte Carlo sampling methods using Markov chains and their
applications. Biometrika, 57, 97–106.
Hayduk, L. A. (1987). Structural Equation Modeling with LISREL. Baltimore, MD:
Johns Hopkins University Press.
Hayduk, L. A. (1996). LISREL Issues, Debates, and Strategies. Baltimore, MD: Johns
Hopkins University Press.
Heise, D. R. (1972). Employing nominal variables, induced variables, and block variables
in path analyses. Sociological Methods & Research, 1, 147–173.
Hodson, J. G., Thompson, K., Wilson, P. J., & Bogaard, A. (1998). Does biodiversity
determine ecosystem function? The ecotron experiment reconsidered. Functional
Ecology, 12, 843–848.
Hox, J. (2002). Multilevel Analysis. Mahway, NJ: Lawrence Erlbaum Associates.
Hoyle, R. H. (ed.) (1999). Statistical Strategies for Small Sample Research. Thousand
Oaks, CA: Sage Publications.
Huston, M. A. (1979). A general hypothesis of species diversity. American Naturalist,
113, 81–101.
Huston, M. A. (1980). Soil nutrients and tree species richness in Costa Rican forests.
Journal of Biogeography, 7, 147–157.
Huston, M. A. (1994). Biological Diversity. Cambridge: Cambridge University Press.
Huston, M. A. (1997). Hidden treatments in ecological experiments: Re-evaluating the
ecosystem function of biodiversity. Oecologia, 110, 449–460.
Huston, M. A. (1999). Local processes and regional patterns: appropriate scales for
understanding variation in the diversity of plants and animals. Oikos, 86, 393–
401.
Jarvis, C. B., MacKenzie, S. B., & Podsakoff, P. M. (2003). A critical review of construct
indicators and measurement model misspecification in marketing and consumer
research. Journal of Consumer Research, 30, 199–218.
Jedidi, K. & Ansari, A. (2001). Bayesian structural equation models for multilevel data.
pp. 129–158. In: Marcoulides, B. A. & Schumacker, R. E. (eds.), New Developments
and Techniques in Structural Equation Modeling. Mahway, NJ: Lawrence Erlbaum
Associates.
Jensen, F. V. (2001). Bayesian Networks and Decision Graphs. New York: Springer
Verlag.
Johnson, M. L., Huggins, D. G., & deNoyelles, F., Jr. (1991). Ecosystem modeling
with LISREL: a new approach for measuring direct and indirect effects. Ecological
Applications, 1, 383–398.
Johnson, J. B. (2002). Divergent life histories among populations of the fish
Brachyrhaphis rhabdophora: detecting putative agents of selection by candidate
model analysis. Oikos, 96, 82–91.
Jöreskog, K. G. (1973). A general method for estimating a linear structural equation sys-
tem. pp. 85–112. In: A. S. Goldberger & O. D. Duncan (eds.). Structural Equation
Models in the Social Sciences. New York: Seminar Press.
Jöreskog, K. G. & Sörbom, D. (1996). LISREL 8: User’s Reference Guide. Chicago:
Scientific Software International.
Jutila, H. & Grace, J. B. (2002). Effects of disturbance and competitive release on
germination and seedling establishment in a coastal prairie grassland. Journal of
Ecology, 90, 291–302.
 References 355

Kaplan, D. (2000). Structural Equation Modeling: Foundations and Extensions. Thou-

sand Oaks, CA: Sage Publishers.
Kaplan, D., Harik, P., & Hotchkiss, L. (2001). Cross-sectional estimation of dynamic
structural equation models in disequilibrium. pp. 315–339. In: R. Cudeck, S. H. C.
Du Toit, & D. Sörbom (eds.). Structural Equation Modeling: Present and Future.
Lincolnwood, IL: SSI Scientific Software International.
Keddy, P. A. (1990). Competitive hierarchies and centrifugal organization in plant com-
munities. pp. 265–289. In: J. B. Grace & D. Tilman (eds.). Perspectives on Plant
Competition, New York: Academic Press.
Keesling, J. W. (1972). Maximum Likelihood Approaches to Causal Flow Analysis.
Ph.D. Dissertation, Department of Education, University of Chicago.
Kelloway, E. K. (1998). Using LISREL for Structural Equation Modeling. Thousand
Oaks, CA: Sage Publications.
Kline, R. B. (2005). Principles and Practice of Structural Equation Modeling. 2nd
Edition. New York: The Guilford Press.
Langley, P., Iba, W., & Thompson, K. (1992). An analysis of Bayesian classifiers. In:
Proceedings of the 10th National Conference on Artificial Intelligence. pp. 223–
228. Cambridge, MA: MIT Press.
Laplace, P. S. (1774). Mémoire sur la probabilité des causes par les événements.
Mémoires de l’Academie de Science de Paris, 6, 621–656.
Larson, D. L. & Grace J. B. (2004). Temporal dynamics of leafy spurge (Euphorbia
esula) and two species of flea beetles (Aphthona spp.) used as biological control
agents. Biological Control, 29, 207–214.
Lawton, J. H., Naeem, S., Thompson, L. J., Hector, A., & Crawley, J. J. (1998). Biodiver-
sity and ecosystem function: getting the ecotron experiment in its correct context.
Functional Ecology, 12, 848–852.
Lee, S. Y., & Bentler, P. M. (1980). Some asymptotic properties of constrained gen-
eralized least squares estimation in covariance structure models. South African
Statistical Journal, 14, 121–136.
Legendre, P. (1993). Spatial autocorrelation: Trouble or a new paradigm. Ecology, 74,
659–673.
Levins, R. (1968). Evolution in Changing Environments. Princeton, NJ: Princeton
University Press.
Li, C. C. (1975). Path Analysis – A primer. Pacific Grove, CA: Boxwood Press.
Little, T. D., Schnabel, K. U., & Baumert, J. (eds.) (2000). Modeling Longitudinal and
Multilevel Data. Mahway, NJ: Lawrence Erlbaum Associates.
Loehle, C. (1987). Hypothesis testing in ecology: psychological aspects and the
importance of theory maturation. The Quarterly Review of Biology, 62, 397–
409.
Loehle, C. (1988). Problems with the triangular model for representing plant strategies.
Ecology, 69, 284–286.
Loehlin, J. C. (1998). Latent Variable Models. Third Edition. Mahway, NJ: Lawrence
Erlbaum Associates.
Loreau, M., Naeem, S., & Inchausti, P. (2002). Biodiversity and Ecosystem Functioning.
Oxford: Oxford University Press.
MacArthur, R. H. & Wilson, E. O. (1967). The Theory of Island Biogeography. Princeton,
NJ: Princeton University Press.
356 References

MacCallum, R. C. & Browne, M. W. (1993). The use of causal indicators in covariance

structure models: some practical issues. Psychological Bulletin, 114, 533–541.
Mancera, J. E., Meche, G. C., Cardona-Olarte, P. P. et al. (2005). Fine-scale environmen-
tal control of spatial variation in species richness in a wetland community. Plant
Ecology, 178, 39–50.
Marcoulides, G. A. & Schumacker, R. E. (eds.) (1996). Advanced Structural Equation
Modeling: Issues and Techniques. Mahway, NJ: Lawrence Erlbaum Associates.
Marcoulides, G. A. & Schumacker, R. E. (eds.) (2001). New Developments and
Techniques in Structural Equation Modeling. Mahway, NJ: Lawrence Erlbaum
Associates.
Marrs, R., Grace, J. B., & Gough, L. (1996). On the relationship between plant species
diversity and biomass: a comment on a paper by Gough, Grace, and Taylor. Oikos,
75, 323–326.
Marsh, H. W., Balla, J. R., & Hau, K.-T. (1996). An evaluation of incremental fit indices:
a clarification of mathematical and empirical properties. pp. 315–353. In: B. A.
Marcoulides & R. E. Schumacker (eds.). Advanced Structural Equation Modeling.
Mahway, NJ: Lawrence Erlbaum Associates.
Marsh, H. W., Hau, K. T., & Wen, Z. (2004). In search of golden rules: comment on
hypothesis testing approaches to setting cutoff values for fit indexes and dangers
in overgeneralizing Hu and Bentler’s findings. Structural Equation Modeling, 11,
320–341.
Maruyama, G. M. (1998). Basics of Structural Equation Modeling. Thousand Oaks, CA:
Sage Publications.
McCune, B. & Grace, J. B. (2002). Analysis of Ecological Communities. Gleneden
Beach, Oregon: MJM.
Meziane, D. & Shipley, B. (2001). Direct and indirect relationships between specific
leaf area, leaf nitrogen and leaf gas exchange. Effects of irradiance and nutrient
supply. Annals of Botany, 88, 915–927.
Mikola, J., Salonen, V., & Setälä, H (2002). Studying the effects of plant species rich-
ness on ecosystem functioning: does the choice of experimental design matter?
Oecologia, 133, 594–598.
Mitchell, R. J. (1992). Testing evolutionary and ecological hypotheses using path anal-
ysis and structural equation modelling. Functional Ecology, 6, 123–129.
Mitchell, R. J. (1994). Effects of floral traits, pollinator visitation, and plant
size on Ipomopsis aggregata fruit production. The American Naturalist, 143,
870–889.
Mittelbach, G. G., Steiner, C. F., Scheiner, S. M. et al. (2001). Ecology, 82, 2381–2396.
Moore, D. R. J. & Keddy, P. A. (1989). The relationship between species richness and
standing crop in wetlands: the importance of scale. Vegetatio, 79, 99–106.
Muggleton, S. (ed.) (1992). Inductive Logic Programming. San Diego, CA: Academic
Press.
Muthén, B. (1984). A general structural equation model with dichotomous, ordered
categorical, and continuous latent variable indicators. Psychometrika, 49, 115–132.
Muthén, L. K. & Muthén, B. O. (2004). Mplus User’s Guide. Third Edition. Los Angeles,
CA: Muthén and Muthén.
Naeem, S. (2002). Ecosystem consequences of biodiversity loss: the evolution of a
paradigm. Ecology, 83, 1537–1552.
 References 357

Neapolitan, R. E. (2004). Learning Bayesian Networks. Upper Saddle River, NJ: Prentice
Hall Publishers.
Oakes, M. (1990). Statistical Inference. Chestnut Hill, MA: Epidemiology Resources
Inc.
Palmer, M. W. (1994). Variation in species richness: towards a unification of hypotheses.
Folia Geobot. Phytotax. Praha, 29, 511–530.
Pankratz, A. (1991). Forecasting with Dynamic Regression Models. New York: John
Wiley & Sons.
Pearl, J. (1992). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. San Mateo, CA: Morgan Kaufmann.
Pearl, J. (2000). Causality. Cambridge: Cambridge University Press.
Pedhazur, E. J. (1997). Multiple Regression in Behavioral Research, 3rd edition. Toronto:
Wadsworth Press.
Peters, R. H. (1991). A Critique for Ecology. Cambridge: Cambridge University Press.
Pianka, E. R. (1970). On r- and K-selection. American Naturalist, 104, 592–
597.
Popper, K. R. (1959). The Logic of Scientific Discovery. London: Hutchinson.
Pugesek, B. H. (2003). Modeling means in latent variable models of natural selection.
pp. 297–311. In: B. H. Pugesek, A. Tomer, & A. von Eye (eds.). Structural Equation
Modeling. Cambridge: Cambridge University Press.
Pugesek, B. H. & Tomer, A. (1996). The Bumpus house sparrow data: a reanalysis using
structural equation models. Evolutionary Ecology, 10, 387–404.
Pugesek, B. H., Tomer, A., & von Eye, A. (2003). Structural Equation Modeling.
Cambridge: Cambridge University Press.
Raftery, A. E. (1993). Bayesian model selection in structural equation models. pp. 163–
180. In: K. A. Bollen & J. S. Long (eds.). Testing Structural Equation Models.
Newbury Park, CA: Sage Publishers.
Raykov, T. & Marcoulides, G. A. (2000). A First Course in Structural Equation Model-
ing. Mahway, NJ: Lawrence Erlbaum Associates.
Raykov, T. & Penev, S. (1999). On structural equation model equivalence. Multivariate
Behavioral Research, 34, 199–244.
Reich, P. B., Ellsworth, D. S., Walters, M. B. et al. (1999). Generality of leaf trait
relationships: a test across six biomes. Ecology, 80, 1955–1969.
Reyment, R. A. & Jöreskog, K. G. (1996). Applied Factor Analysis in the Natural
Sciences. Cambridge: Cambridge University Press.
Rosenzweig, M. L. & Abramsky, Z. (1993). How are diversity and productivity related?
pp. 52–64. In: R. E. Ricklefs & D. Schluter (eds.). Species Diversity in Ecological
Communities. Chicago: University of Chicago Press.
Rupp, A. A., Dey, D. K., & Zumbo, B. D. (2004). To Bayes or not to Bayes, from
whether to when: Applications of Bayesian methodology to modeling. Structural
Equation Modeling, 11, 424–451.
Salsburg, D. (2001). The Lady Tasting Tea. New York: Henry Holt & Company.
Satorra, A. & Bentler, P. M. (1988). Scaling corrections for chi-square statistics in covari-
ance structure analysis. pp. 308–313. In: Proceedings of the American Statistical
Association.
Satorra, A. & Bentler, P. M. (1994). Corrections to test statistics and standard errors in
covariance structure analysis. pp. 399–419. In: A. von Eye & C. C. Clogg (eds.).
358 References

Latent Variables Analysis: Applications for Developmental Research. Thousand

Oaks, CA: Sage Publishers.
Scheiner, S. M., Mitchell, R. J., & Callahan, H. S. (2000). Using path analysis to measure
natural selection. Journal of Evolutionary Biology, 13, 423–433.
Scheines, R., Hoijtink, R., & Boomsma, A. (1999). Bayesian estimation and testing of
structural equation models. Psychometrika, 64, 37–52.
Schermelleh-Engel, K., Moosbrugger, H., & Müller, H. (2003). Evaluating the fit of
structural equation models: Test of significance and descriptive goodness-of-fit
measures. Methods of Psychological Research – Online, 8, 23–74.
Schumacker, R. E. & Lomax, R. G. (eds.) (1996). A Beginner’s Guide to Structural
Equation Modeling. Mahway, NJ: Lawrence Erlbaum Associates.
Schumacker, R. E. & Marcoulides, G. A. (eds.) (1998). Interaction and Nonlinear Effects
in Structural Equation Modeling. Mahway, NJ: Lawrence Erlbaum Associates.
Shipley, B. (2000). Cause and Correlation in Biology. Cambridge: Cambridge Univer-
sity Press.
Shipley, B. & Lechowicz, M. J. (2000). The functional coordination of leaf morphology
and gas exchange in 40 wetland plant species. Ecoscience, 7, 183–194.
Shipley, B., Keddy, P. A., Gaudet, C., & Moore, D. R. J. (1991). A model of species
density in shoreline vegetation. Ecology, 72, 1658–1667.
Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, Prediction, and Search.
Cambridge: MIT Press.
Stamp, N. (2003). Theory of plant defense level: example of process and pitfalls in
development of ecological theory. Oikos, 102, 672–678.
Stearns, S. C. (1977). The evolution of life history traits: a critique of the theory and a
review of the data. Annual Review of Ecology and Systematics, 8, 145–171.
Steiger, J. H. (1990). Structural model evaluation and modification: an interval estimation
approach. Multivariate Behavioral Research, 25, 173–180.
Symstad, J. J, Chapin, F. W., Wall, D. H. et al. (2003). Long-term and large-scale
perspectives on the relationship between biodiversity and ecosystem functioning.
Bioscience, 53, 89–98.
Taper, M. L. & Lele, S. R. (2004). The Nature of Scientific Evidence. Chicago, Illinois:
University of Chicago Press.
Taylor, D. R., Aarssen, L. W., & Loehle, C. (1990). On the relationship between r/K
selection and environmental carry capacity: a new habitat templet for plant life
history strategies. Oikos, 58, 239–250.
Tilman, D. (1982). Resource competition and community structure. Princeton, NJ:
Princeton University Press.
Tilman, D. (1986). Resources, competition and the dynamics of plant communities.
pp. 51–75. In: M. J. Crawley (ed.). Plant Ecology. London: Blackwell Scientific
Publications.
Tilman, D. (1987). On the meaning of competition and the mechanisms of competitive
superiority. Functional Ecology, 1, 304–315.
Tilman, D. (1988). Plant Strategies and the Dynamics and Structure of Plant Commu-
nities. Princeton, New Jersey: Princeton University Press.
Tilman, D. (1997). Mechanisms of plant competition. chapter 8. In: M. J. Crawley (ed.).
Plant Ecology, 2nd edn. Malden, MA: Blackwell Scientific Publications.
Tilman, D., Wedin, D., & Knops, J. (1996). Productivity and sustainability influenced
by biodiversity in grassland ecosystems. Nature, 379, 718–720.
 References 359

Tomer, A. (2003). A short history of structural equation models. pp. 85–124. In:
B. H. Pugesek, A. Tomer, & A. von Eye (eds.). Structural Equation Modeling.
Cambridge: Cambridge University Press.
Tukey, J. W. (1954). Causation, regression, and path analysis. pp. 35–66. In: O.
Kempthorne, T. A. Bancroft, J. W. Gowen, & J. D. Lush (eds.). Statistics and
Mathematics in Biology. Ames, IA: Iowa State College Press.
Turner, M. E. & Stevens, C. D. (1959). The regression analysis of causal paths. Biomet-
rics, 15, 236–258.
Verheyen, K., Guntenspergen, G. R., Biesbrouck, B., & Hermy, M. (2003). An integrated
analysis of the effects of past land use on forest herb colonization at the landscape
scale. Journal of Ecology, 91, 731–742.
von Mises, R. (1919). Grundlagen der Wahrscheinlichkeitsrechnung. Mathematische
Zeitschrift, Vol. 5. (referenced in Neapolitan (2004).)
Waide, R. B., Willig, M. R., Steiner, C. F. et al. (1999). The relationship between
productivity and species richness. Annual Reviews in Ecology and Systematics, 30,
257–300.
Wardle, D. A. (1999). Is “sampling effect” a problem for experiments investigating
biodiversity–ecosystem function relationships? Oikos, 87, 403–407.
Weiher, E., Forbes, S., Schauwecker, T., & Grace, J. B. (2004). Multivariate control of
plant species richness in a blackland prairie. Oikos, 106, 151–157.
Wheeler, B. D. & Giller, K. E. (1982). Species richness of herbaceous fen vegetation
in Broadland, Norfolk in relation to the quantity of above-ground plant material.
Journal of Ecology, 70, 179–200.
Wheeler, B. D. & Shaw, S. C. (1991). Above-ground crop mass and species richness
of the principal types of herbaceous rich-fen vegetation of lowland England and
Wales. Journal of Ecology, 79, 285–302.
Wiley, D. E. (1973). The identification problem for structural equation models
with unmeasured variables. In: A. S. Goldberger & O. D. Duncan (eds.).
Structural Equation Models in the Social Sciences. New York: Seminar Press
A. S.
Williams, L. J., Edwards, J. R., & Vandenberg, R. J. (2003). Recent advances in
causal modeling methods for organizational and management research. Journal of
Management, 29, 903–936.
Wilson, J. B. & Lee, W. G. (2000). C-S-R triangle theory: community-level predictions,
tests, evaluation of criticisms, and relation to other theories. Oikos, 91, 77–96.
Wisheu, I. C. & Keddy, P. A. (1989). Species richness-standing crop relationships along
four lakeshore gradients: constraints on the general model. Canadian Journal of
Botany, 67, 1609–1617.
Wootton, J. T. (1994). Predicting direct and indirect effects: an integrated approach using
experiments and path analysis. Ecology, 75, 151–165.
Wootton, J. T. (2002). Indirect effects in complex ecosystems: recent progress and future
challenges. Journal of Sea Research, 48, 157–172.
Wright, S. (1918). On the nature of size factors. Genetics, 3, 367–374.
Wright, S. (1920). The relative importance of heredity and environment in determining
the piebald pattern of guinea pigs. Proceedings of the National Academy of Sciences,
6, 320–332.
Wright, S. (1921). Correlation and causation. Journal of Agricultural Research, 10,
557–585.
360 References

Wright, S. (1932). General, group, and special size factors. Genetics, 17, 603–619.
Wright, S. (1934). The method of path coefficients. Annals of Mathematical Statistics,
5, 161–215.
Wright, S. (1960). Path coefficients and path regressions: alternative or complementary
concepts? Biometrics, 16, 189–202.
Wright, S. (1968). Evolution and the Genetics of Populations, Vol. 1: Genetic and
Biometric Foundations. Chicago: University of Chicago Press.
Wright, S. (1984). Diverse uses of path analysis. pp. 1–34. In: A. Chakravarti (ed.).
Human Population Genetics. New York: Van Nostrand Reinhold.
 Index

Akaike Information Criterion (AIC), canonical correlation, 17

131–132, 184–185, 226, 287 canonical correspondence analysis, 17
alpha, defined, 5 categorical variables (see variable,
amphibians, 167 categorical)
Andropogon, 191 causal, defined, 8
anurans (see amphibians) causal indicator, 146
Aphthona causal modeling, 13
lacertosa, 22 causal pathway, 49
nigriscutis, 22 causality, 8–10, 281
Aristotle, 299 cautions, 96–100, 112, 275–288
association by joint causation, 49 chi-square tests, 128–130, 287
asymptotic distribution free, 126, 190 single degree of freedom test, 129, 184
asymptotically unbiased, 124 cluster analysis, 17
atrazine, 237 clockwork universe, 5
attenuated correlations, 188 cluster sampling, 198
collinearity, 280
background variables (see variable, communality, 103
background) competing models, 210–221
Bacon, Sir Francis, 299 composite variables (see variables, composite)
basic effect, 240 composite
Bayes theorem, 313 fixed, 146
Bayesian analysis, 7, 311–316 statistical, 146
Bayesian Information Criteria, 132 compound path, 47, 49
Bayesian networks, 315 conceptual model (see model, conceptual)
best estimator, 124 conditionally independent, 54
beta, defined, 11, 50, 137, 184 confirmatory analysis, 17–19, 133
biased parameter estimates, 87 confirmatory model evaluation (see model
biocomplexity, 291 evaluation, confirmatory)
biosimplicity, 291 consistent validity (see validity, consistent)
birds, 190–192 construct reliability, 93
biocontrol, 22 construct model (see model, construct)
block, 145 convergent validity (see validity, convergent)
bluestem, 191 correlated error, 29, 42
body size, 78, 90–92, 93, 94 correlation, polychoric (see polychoric
bootstrapping, 126, 130, 241–246, 286, correlation)
332 correlation coefficient, 45

361
362 Index

covariance, 43 evaluating parameters, 125–126

covariance matrix example programs, 326–349
model-implied (predicted), 122, 127–128, exotic species, 250–254
135–137 expected cross validation index, 132
observed, 122 experiments, 227, 233–258
covariance structure analysis, 13 exploratory analysis, 17–19
criterion validity (see validity, criterion) exploratory model evaluation (see model
evaluation, exploratory)
d-separation, 315 external validity (see validity, external)
data,
adjustment, 236 factor analysis, 14, 100
missing, 286 common, 100
screening, 285 component, 100
transformation, 192, 286 confirmatory, 14
defining the variance of a latent factor, 92 factor-focused applications, 83
delta, defined, 136, 138, 183 factor loading (see loading)
dichotomous variables (see variable, feedback loops, 198
dichotomous) fertilization, 241
direct path (aka direct effect), 42 fire, 190, 250–254
directed acyclic graph (DAG), 315 fish, 226
discriminant analysis, 17, 19 Fisher, Ronald, 6
discriminant validity (see validity, fitting function, 124, 128
discriminant) fixed effects, 236
distribution flea beetles, 22
multinormal, 124 forecasting, 318, 319–320
Wishart, 124 forest, 147
diversity, 254, 259–274 frequentist analysis, 311
dummy variables (see variable, dummy) full information estimation, 124
functional diversity, 305
edges, 315
effect indicator, 146 gamma, defined, 5, 11, 50, 136, 137, 184
empirical underidentification (see model, Gibbs method, 313
empirical underidentification) goodness of fit, 131
epsilon, defined, 138, 184 generalized least squares, 124
equality constraint, 184 graphical models, 17, 310
equivalent models (see models, equivalent) grassland (see also meadows), 105–109,
equivalency 190
pure, 227, 233–258 grazing, 185
statistical, 227, 230 growth models (see latent growth models)
error, correlated (see correlated error)
error of measurement (see measurement error) herbs, 147
error variable (see variable, error) Heywood case, 342
error variance, 82 hierarchical models (see model, hierarchical)
errors, historical effects, 270
in model specification, 276–279 hummingbirds, 212
of analysis, 279–280 hump-backed model of diversity, 301
of data treatment, 279 hybrid model (see model, hybrid)
of interpretation, 280–281
variable omission, 278 identification of models or parameters (see
estimation (see parameter estimation) model, identified)
eta, defined, 137 imputation, 286
Euphorbia esula, 22 inadmissible solutions, 342
 Index 363

indeterminacy, 102 measurement model (see model, measurement)

indicator variable (see variable, indicator) measurement theory, 296
indirect effect, 42 measures of model fit (see also chi-square
information theoretic methods, 208 tests)
interaction effects, 237 absolute, 130
intermediate disturbance hypothesis, comparative, 130
303 parsimonious, 130
intermediate stress hypothesis, 301 uncorrected, 130
internal consistency, 144 Metropolis–Hasting method, 313
internal validity (see validity, internal) missing data (see data, missing)
interpretation, 287–288 model,
intervention outcome assumption, 10 assessing fit, 126–131
island biogeographic theory, 148 building, 210
conceptual, 317–319
Jöreskög, Karl, 13, 121 construct, 143, 263, 282, 318, 319–320
empirical underidentification, 119
kurtosis, 126 fit measures (see measures of model fit)
hierarchical, 196, 198, 236
LaConte’s sparrow, 191 hybrid, 83, 109–112
Lagrange multipliers, 185 identification, 202
lambda, defined, 78, 183 identified, 116–121
lambda-x, defined, 138 latent variable, 137
lambda-y, defined, 138 logistic, 189
landscape position (position in the landscape), manifest variable, 39
270 measurement, 83, 138
latent growth models, 195–196 modification, 133
latent variable (see variable, latent) nested (hierarchical), 130
latent variable modeling, 13 non-nested (nonhierarchical), 130, 131–133
cautions, 112 nonrecursive, 42, 198–203, 240
leafy spurge, 22 observed variable, 37, 41–42
least squares regression, 115 overidentified, 118
leghorn hens, 104 panel, 198
linearity (see nonlinear relationships) path, 39
LISREL probit, 189–190
the model, 13, 15, 121, 137 recursive, 42
the software (see software, LISREL) reduced form, 162
loading, 80, 82, 102 simulation, 320
log likelihood ratio, 123 structural, 83, 137
logistic model or function (see model, logistic) underidentified, 116
logit transformation, 189 model evaluation, 286–287
longitudinal data, 195, 283 confirmatory, 208, 209–210
loop-enhanced effect, 240 exploratory, 208
nested series, 208, 210–221
manifest variable (see variable, manifest) non-nested, 208
marine intertidal community, 247–249 step-wise, 221–223
maximum likelihood estimation, 123–124, model-implied covariance matrix (see
128 covariance matrix, model-implied)
MANOVA, 17 model selection, 223–226
meadows, 185 model trimming, 210
means comparisons, 185, 237 models,
measurement error, 78, 85–89 (see also alternative, 227
reliability) equivalent, 227–229
364 Index

modification indices, 127, 331, 343 path-focused applications, 83, 84–85

Monte Carlo methods, 126, 283 Pearson, Egon, 6
Mplus software (see software, Mplus) phi, defined, 136, 138
multigroup analysis, 182–187, 192–198, 237 pollination, 212–221
multi-level modeling, 198 polychoric correlation, 189, 235
multinormal distribution (see distribution, ponds, 237–241
multinormal) positive definite, 124
multi-stage analysis, 283 prairie, 105–109, 191, 241–246, 250–254
multivariate normality, 129 prediction, 268–269, 301
multiple indicators, 89–100 principal components analysis, 17, 18, 19, 100
multiple regression, 19 prior odds, 313
probit model (see model, probit)
naming fallacy, 98 proportional error variance, 81
negative variance, 124 psi, defined, 54, 138
nested model evaluation (see model
evaluation, nested) R2 , 53, 61
Neyman, Jerzy, 6 R* model, 297
Neyman–Pearson protocol, 6, 125 r- and K-selection, 297
nodes, 315 random effects, 236
nonindependency, 185 reciprocal interactions, 42, 239–241
nonlinear relationships, 191, 192–195, 286 recommendations, 275–288
non-nested model evaluation (see model recursive model (see model, recursive)
evaluation, nonnested) regression coefficient, 46
nonnormality (see normality) regression trees, 19
nonpositive definite matrix, 119, 343 relevant range, 71
nonrecursive model (see model, nonrecursive) reliability, 80–82, 89 (see also construct
normality, 125, 188, 190 reliability)
null hypothesis testing, 5–7, 126 repeated measures, 195
resampling methods, 126
observed variable (see variable, observed) residuals, 127–128, 227, 230, 287
oblique, 101 response surface design (of experiments),
ordinal variables (see variable, ordinal) 235
ordinary least squares, 115 robust methods, 126, 130, 285, 286, 332
orthogonal, 101 root mean residual, 131
outliers, 285, 286 root mean square error of approximation, 131
overfitting, 134, 212
overidentified (see model, overidentified) sample distribution (sampling distribution),
overyielding, 305 284
sample size, 283
panel model (see model, panel) sampling, 283–284
parameter estimation, 121–126 sampling effect, 305
partial correlation, 49, 53–55 Satorra–Bentler robust method (see robust
partial regression, 49, 50, 73–76 methods)
path analysis, 12, 39 saturated model, 42
historic, 39–41 sequential hypothesis testing, 271
modern, 13, 39–41 scale invariant, 124
Wrightian, 13 scientific method, 5, 293
path coefficients, 42–65 semipartial correlation, 63
alternative form of standardization, 69–72 sequential learning, 313
rules summarized, 65 serpentine endemic plants, 272
standardized vs. unstandardized, 57–65 sheath coefficient, 176
 Index 365

sigma, 123, 136 total correlation, 55, 57

simple pathway, 49 total effect, 55
simulation models (see model, simulation) tracing rules, 56
single degree of freedom chi-square test (see two-stage least squares, 200
chi-square test, single degree of freedom)
singular determinant, 124 uncertainty, 312
software, 324 underidentified model (see model,
AMOS, 325 underidentified)
CALIS, 325 uninformative prior, 313
EQS, 325 uniqueness, 103
HLM, 284 univariate analyses, 3, 5, 85–89
LISREL, 13, 284, 325 unsaturated model, 42
Mplus, 284, 325
SEPATH, 325 validity, 80, 104
SIMPLIS, 325 consistent, 267
TETRAD, 227 convergent, 104
soil, 105, 107, 110–112 criterion, 104
sparrows, 104 discriminant, 104
spatial autocorrelation, 271 external, 104
spatial effects, 270–271 internal, 104
species pool model of diversity, 304 vanishing tetrads, 159
specificity, 103 variable,
spurious correlation, 32 background, 284
standard deviation, 45 categorical, 187, 188–192, 197, 235,
statistical control (vs. experimental control), 283
50 composite, 37, 143–180, 195, 243, 345
statistical equivalency (see equivalency, dichotomous, 188
statistical) dummy, 187
step-wise model evaluation (see model endogenous, 41
evaluation, step-wise) error, 38, 52–53
structural equations, 7–9 exogenous, 41
defined, 7, 8 indicator, 77, 264
structural equation modeling, latent, 14, 37, 77–114
compared to other multivariate techniques, observed, 37
19 ordinal, 188
defined, 10 variance, 43
history, 12–15 variance explanation
structural model (see model, structural) shared, 63
suppression, 57 unique, 62
variance partitioning, 60
tetrachoric correlation, 235
TETRAD software (see software, TETRAD) weighted least squares, 190
theory tenacity, 294 wetlands, 255, 259
theta, defined, 136 Wishart distribution (see distribution, Wishart)
theta-delta, defined, 138 Wright, Sewell, 11, 12, 58
theta-epsilon, defined, 138
theta matrix, 343 xi, defined, 136, 137
threshold, 189, 236
transformations (see data transformation) z transformation, 44
t-rule, 116, 202 z-values, 129
t-test, 129 zeta, defined, 5, 136, 137, 184