SC504

Week 6
Logic of multivariate analysis:
the elaboration approach
The Elaboration Model

Babbie uses the term, Elaboration Model,

to refer to this protocol for investigating
cause and effect among variables.
Other names:
Lazarsfeld method.
Columbia school.
Origin of the Model
The American Soldier
Stouffer et al. sought to understand how to
increase the motivation and morale of
soldiers in the U.S. Army.
The researchers sought ways to explain
seemingly non-rational behavior among U.S.
soldiers during their training for duty during
WWII.
Origin of the Model
The American Soldier (Continued)
Consider this intuitively plausible proposition:
1. Soldiers with more education are more likely to resent being
drafted than soldiers with low education.
Because the opportunity cost of being drafted to highly
educated men is greater (lost income, career progression
etc)
Origin of the Model
The American Soldier (Continued)
Although this hypothesis seemed intuitively reasonable, it was
not supported. Instead, the opposite was observed. Well
educated soldiers resented being drafted less than poorly
educated soldiers.
The reason turned out to be that poorly educated soldiers had
more friends whose draft was deferred - because these friends
were working in strategic industries etc. College-educated
soldiers had college friends who did not work and were less
likely to have reason for deferral.
Reference group comparisons drove resentment, not
individual costs and benefits.
How was this discovered? By introducing a test factor (friends
deferred) into an observed relationship (high education ->
more resentment) = elaborating the relationship between x
and y.
 Original observation

Source: Rosenberg 1968

Broken down by test factor

Source: Rosenberg 1968

Main Rationale for the Elaboration
Model
Experimental Control
In the physical and life sciences, it is
possible to physically control most
variables.
In the social sciences, one can rarely
physically control variables (e.g., one
cannot take out religious beliefs from a
person to isolate the effects of education
on income, controlling for religious
beliefs).
Therefore, the social sciences rely more
Steps in the Elaboration Model
1. A relationship is observed between two variables x
and y.
The researcher begins by examining frequencies
and bivariate relationships (e.g., the greater
education, the greater the income).
2. A third variable z is held constant by subdividing
the cases according to the attributes of this third
variable.
The cases are divided into the groupings of the
third variable (e.g., education and income are
divided by males and females and the original
relationship observed separately within each
group). Or in regression, this means we add the
variable and control for it. we are now
Steps in the Elaboration Model
1. In the context of regression analysis, this
control is achieved by entering the third
(test) variable into the model and examining
what happens to the coefficient of the
original variable of interest.

Model 1 : Education predicting income

Model 2: Education and gender predicting
income
[look to see what happens to education
coeff when gender included in the model]
 Types of test variable
Elaboration of a bivariate relationship with a
third variable can take two (or sometimes
three) forms. The variable is assumed to be
either:
Antecedent
Intervening
(Interacting)
What does this mean?
Lets think about what causality actually is and
what assumptions we need to infer a causal
relationship
 Some assumptions necessary for
inferring causality from correlation XY
X is temporally prior to Y
E.g. compulsory education is prior to
adult criminality, so could be a cause.
Attitudes towards police competence
and adult criminality could arise in any
temporal order so causality cannot
easily be assumed
 Some assumptions necessary for
inferring causality from correlation XY
X and Y are not jointly caused by Z (ie
another variable not in our model)
E.g. if education levels and criminal behaviour are
both related to (or caused by) parental social
class, then we cannot conclude education is a
cause of crime (also known as spuriousness)
Necessity of this assumption underlies logic
of randomised experiments. Guarantees that
treatment and control groups are exactly
same in relation to all possible vectors of Zs
except that one group receives treatment
 Education and crime

Educati
on

Crime

Does education cause crime (i.e.

does having more education make
people less likely to commit crime?)
 Antecedent test variable: parental
social class

Educati
on

Parent
al
social Crime
class
Relationship is explained by joint dependence of crime and
education on class (ie within high and low social class groups
considered separately, there is no correlation between educ and
crime
No causal relationship.
(Regression coefficient would go to zero when class added to
model)
Or: intervening test variable: adult
income

Educati
on Crime
Intervening test variable (also called
mediation or interpretion)

Adult
income

Educati
on Crime

Relationship is interpreted (in Lazarsfeldian nomenclature) by

dependence of income on education and dependence of crime on
income
Intervening variables also called mediators the effect of X on Y is
through its effect on Z
Regression coefficient would go to zero when income added to model
 Intervening or antecedent?
There is no empirical way to distinguish
between what is antecedent and what is
intervening. This has to be an assumption
If controlling for a test variable Z makes
original XY relationship disappear or
reduce, the true data generating
mechanism may be either:
Z X or: X Z Y
Y
Combining antecedent and intervening
effects: direct and indirect effects

Educati
on
Incom
e
Social
class Crime

Indirect effect of social class through education

and income
Direct effect of social class on crime (presumably
through some unobserved mechanisms like area,
attitudes to authority, friendship networks)
Interaction (moderator) effects (also
called specification)

Female

+ve
Incom
e Crime
-ve

Effect of income moderated by gender

e.g. higher income men lower crime
higher income womenno effect on
crime
 Suppressor relationships
This complicates things
We are investigating relationship between education
and holding left wing views

Educati
on Left
wing
political
We find no relationship (no correlation, a
=zero non-sig
views
regression coefficient close to zero, no association in a
cross tabulation
Can we conclude there is no causal link? Unfortunately
no!
 Suppressor relationships
Lets say we test for income as an intervening variable (more education
enables =more earning potential)

Educati +ve Incom -ve

on e Left
wing
political
+ve views
We find that education is positively correlated with income and income is
negatively correlated with being left wing (rich people are more
conservative). But we also find, once controlling for income that the direct
effect of education on being left wing is now positive!
Suppressor effect whereby intervening variable Z is related inversely to the
outcome Y in relation to X. Direct effect of education on political views is
cancelled out in the observation but is still real if the model is correct.
 Hierarchical regression
modeling
The trick to doing this kind of analysis is to think of the hypothesised temporal
order and estimate one model at each stage, adding variables as you go along.
Here we have education->income->political view
Step 1: fit a model predicting political view with education
Step 2: add our test variable, income to the model and watch what happens to the
coefficient for education

This is consistent with the suppressor relationship in the previous slides.

Question: What would we see in these models if, in the world (ie the mechanism
that generated the data), more income made people more left wing?
 Thinking about what elaboration
means for regression analysis
The world:

Educati +ve Incom +ve

on e Left
wing
political
zero views

Our regression model results:

Model 1: Education predicting political views. Coefficient?
Model 2: Education and Income predicting political views.
Coefficients?