Mediation, Moderation, and Interaction:

Definitions, Discrimination, and (some) means of testing

Quantitative SIG
February 25th 2013

Department of Education
University of Oxford

[email protected]

1 Note, The content within this presentation is expanded upon in a book chapter which can be found here:
DOI: 10.1007/978-94-6209-404-8_13
 Abstract
 In 1986 Baron and Kenny set out to clarify the differences between the terms “Moderation”
and “Mediation” as used in the social sciences.

 Twenty seven years later, the seminal paper that this collaboration resulted in (published in
the Journal of Personality and Social Psychology) has been cited over 35,000 times
(Google Scholar on 09/01/2013).
– This is approximately 1300 times year – roughly equall to once every 7 hours of every day of every year for over a
quarter of a century.

 However and despite this citation record, the uncertainty surrounding these terms has not
gone away.
– Academics still struggle to define, distinguish and utilise these terms while related under- and post-graduate teaching
is still the exception.

 This presentation sets out simple, clear definitions that distinguish “Mediation” from
“Moderation”, and both from “Interaction” as well as all three from a number of other
commonly-used terms.

2
 Contents
Part 1 – Theoretical Issues :
– Definitions & Discrimination
 (Including 4 ways of testing hypotheses of “mediation”)
Part 2 – Practical Issues (theory in practice):
– Including (some) means of testing hypotheses of “moderation”:
1. Sub-group Comparison
2. Statistical Interaction Term
3. Random Slope Effect
Part 3 – An example of how to test an hypothesis of moderation:
– Triple-analysis giving comparisons between:
1. A Sub-group Comparison
2. A Statistical Interaction Term
3 3. A Random Slope Effect
 Part 1 – Theoretical Issues:
Unambiguous Definitions

4
 Definitions (I)

Mediation
 A tri-variate unidirectional (thus causal) hypothesis:
– What are the mechanisms of effect underlying a pre-established causal statistical
relationship?
 A pre-established causal relationship between two variables (X, Y) is theorised to exist due to
an intermediate third variable (Me)

 The additional (third) variable that is hypothesised to have this effect

is known as a “mediator” though is also sometimes referred to as an:
– “intermediate variable” (Kraemer et al., 2001)
– “explanatory link” (Rose, Holmbeck, Coakley, & Franks, 2004)

 Mediators have “mediating effects” that are otherwise labelled:

– “indirect effects”
– “surrogate effects”
– “intermediate effects”
5 – “intervening effects” (MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002; *Wu & Zumbo, 2007)
 Definitions (II)

Moderation
 Another tri-variate unidirectional (thus causal) hypothesis:

– Under what conditions/for whom/when is a pre-established

statistical relationship evident?
– Completely different from mediation

 The presence of a third “moderator” variable

(Mo) is also termed:
– a “causal interaction effect” (Wu & Zumbo, 2007)
– an “effect-modifier” (Hinshaw, 2002)

*Wu & Zumbo, (2007)

6
 Definitions (III)

(Statistical) Interaction
 A bidirectional (thus non-causal) multi-variate hypothesis

 Implies that two or more concepts represented by measured

variables, “work together” or, “have a combined effect” in eliciting
a third (e.g. Kraemer et al., 2001; Talamini et al., 2002)
≠ behavioural, psychological, gene-environment “interaction”
(obviously these are non-statistical!)

 Commonly tested with the specification of a “Statistical

Interaction Term”
7
 Aside: “Indirect effects”

 An “Indirect effect” is another term that can be encountered in

discussions of, “mediation”, “(statistical) interaction”, and
“statistical interaction terms”

 Both mediation and statistical interaction feature effects that

can be thought of as “indirect” but these are distinct and should
not be confused.

 Within Structural Equation Modelling (SEM) software and texts

it is common to find the tern “indirect effect” reserved for just
those indirect effects that are associated with mediation

8
 Definitions (IV)

 Understanding Moderation and Mediation to be tri-variate one-way

hypotheses gives the quantitative researcher an intellectual framework to
postulate complex but still testable causal hypotheses

 Variation between disciplines in the acceptance and use of these

hypotheses however

 For example, the famous Baron and Kenny (1986) was published in a
psychology journal: The Journal of Personality and Social Psychology
– What about areas of academia though? More specifically, those that concern themselves
with quantitative approaches to educational research? For example,
 Economists & Econometricians?
 Sociologists?
 Social Geographers?

9
 Political Scientists?
– & what about “educational researchers” themselves?
 Part 1 – Theoretical Issues:
Discrimination

10
 Mediation ≠ Moderation

The classical diagrams:

Mediation: Moderation:
Answers hypotheses of, Answers hypotheses of,
“How” and, “Why”* “When” and, “For whom”*

X Y X Y

Me Mo
Me = Mediator Direct effects
Mo = Moderator Moderated effect
11 * Wu & Zumbo (2007)
 Why the confusion?
 The simple similarity of the two words

 Their changing definitions over time

 The similar purposes for which both are used in research. Moderation and
Mediation are both:
1. …“theories for refining and understanding a causal relationship” (Wu &
Zumbo, 2007)

2. ...tri-variate hypotheses
3. ...unidirectional (thus causal) hypotheses

 They may even be combined to provide a powerful intellectual framework for

specifying and testing multi-variate hypotheses
– E.g. “How” does X manage to impact Y - is it via intermediate effects
on Z [mediation]? Further, under what conditions of W does Z express
12 this role [moderation]?
*Often only in the context
of experimental research
though
Kraemer et al. (2001) Hinshaw (2002) Nicholson et al. (2005) Essex et al. (2006) Wu & Zumbo (2007)
+-+-+Moderation+-+-+
Moderator has temporal Moderator has Moderator has Moderator has Moderator has
precedence precedence precedence precedence precedence
Moderator and that Moderator and that Moderator and that Moderator and that Moderator and that
Some Past Guides*

moderated are moderated are moderated are moderated are moderated are
uncorrelated uncorrelated uncorrelated uncorrelated uncorrelated
Co-domination of
moderated and Answers, "for whom"
moderator and "when"
Moderator is a trait
Moderator is observed
+-+-+Mediation+-+-+
That being mediated That being mediated That being mediated That being mediated has That being mediated has
has precedence has precedence has precedence precedence precedence
Mediator and that Mediator and that Mediator and that Mediator and that Mediator and that
mediated are mediated are mediated are mediated are mediated are
correlated correlated correlated correlated correlated
Either co-domination of Either co-domination of
mediated and mediated and
mediator (partial) mediator (partial) Answers, "how" and
OR OR "why"
Mediator dominates
that mediated Mediator dominates that
(total) mediated (total) Mediator is a state

13 Mediator is observed or
manipulated
 Real-world ambiguities between
mediation and moderation (I)
 Even with unambiguous definitions, both hypotheses can be
appropriate for the same set of measures
– In developmental science, this often depends upon the timeframe
– For example, Masten (2007) describes the relationship between
adversity (X), stress-response (Y) and stress-regulators (Z):

Period A, Period B,
stress-regulators developing - mediation stress-regulators developed - moderation

Adversity Stress- Adversity Stress-

Regulators (t 2) Response

Stress- Stress-
Regulators (t1) Direct effect Regulators
Moderated effect
14
 Real-world ambiguities between
mediation and moderation (II)

 Masten (2001) also gives the developmental science

example of...
– “...a risk-activated moderator analogous to an automobile airbag or
immune system response.”

15 – Aside: This is also an excellent description of the ideal functioning

of social interventions such as Head Start and Sure Start
 Combining moderation and mediation

 A, “risk-activated moderator” (Masten, 2001) is

just one example of how moderation and
mediation may be integrated.
 The more (in)famous examples are
– “moderated-mediation”
– “mediated-moderation”
 Both are 4 variable hypotheses [W, X, Y, Z] that postulate
conditional (moderated) mechanisms of effect (mediation)
whereby X impacts Y

16 (Aside: Wu and Zumbo (2007) provide further excellent descriptions and discussion)
 *SPSS macro available from: http://www.afhayes.com/spss-sas-and-mplus-macros-and-code.html
Moderated-mediation*

X Y

Me

Mo
Hypothesising the ‘when’ and ‘for whom’ of
an initially mediated relationship

Me = Mediator Direct effect

Mo = Moderator Moderated effect
17 Additional effects
 Mediated-moderation

X Y

Mo Me

Hypothesising the ‘how’ and ‘why’ of an

initially moderated relationship

Me = Mediator Direct effect

Mo = Moderator Moderated effect
18 Additional effect
 Moderation ≠ (Statistical) Interaction (I)

 Moderation is unidirectional & thus a causal hypothesis

– “The relationship between X and Y varies by Z”

 (Statistical) interaction is bidirectional & thus a non-causal

hypothesis,
– “There is a combined effect of X and Z on Y”

 Moderation can be viewed as a more restricted version of Statistical

Interaction as evidenced by the alternative name for Moderation
given by Wu and Zumbo (2007): “Causal Interaction”

19
 Moderation ≠ (Statistical) Interaction (II)

 Although the statistical methods that are used to test hypotheses

of (Statistical) Interaction can also be applied to hypotheses of
Moderation, to conclude Moderation from these methods
necessitates relying heavily upon pre-existing knowledge, be this
from:
– past research findings
– broader substantive theories
– other top-down sources of knowledge (e.g. Nicholson, Hursey, & Nash, 2005).

 This is broadly similar to the difference between correlation

coefficients and regression coefficients
– And their varying suitability in answering research questions

20
 “(Statistical) Interaction” ≠
“Statistical Interaction Terms”

 Again, Statistical Interaction is a bidirectional hypothesis that two

or more concepts represented by measured variables, “work
together”/“have a combined effect” upon a third

 A “Statistical Interaction Term” is a statistical artefact where:

– 2+ variables are zero-centred (typically by z-scoring), multiplied together,
and commonly entered into a regression equation (though also ANOVA) of
the form (here: tri-variate):
 Y=X+Z+XZ [+e]
– Note, Statistical Interaction Terms include a “main effect” [here: Z] as well as
a combined effect [XZ] .
– This main effect is necessary when specifying a “Statistical Interaction Term
– ...But it is completely unnecessary for establishing moderation

21
 Moderation, (Statistical) Interaction,
and Statistical Interaction Terms

 The relationship between Moderation,

Statistical Interaction, and Statistical
Interaction Terms takes the following form:
– Moderation is a more restricted unidirectional
alternative to the bidirectional hypotheses of
Statistical Interaction although both are often
tested through the specification of Statistical
Interaction Terms.

22
 Operationalisation (I)

Mediation
 MacKinnon and colleagues (2002) discuss fourteen
statistical methods to test hypotheses of Mediation(!)
(see Hayes, 2009)

 Four particularly common methods are:

1. The ’Causal Steps Approach’ of Baron and Kenny

(1986)
 Strongly criticized as having the least statistical power
to accurately detect Mediation Effects (Fritz & MacKinnon, 2007; Hayes,
2009)
23
 Operationalisation (II)

2. The Sobel Test (Sobel, 1982)

– A more formal test of mediation compared to the
Causal Steps Approach
– Multiple regression analyses are conducted and
the results of each are combined (see Preacher & Hayes, 2004)
– Various macros and online calculators are
available for this additional step (e.g.: http://www.danielsoper.com)

24
 Operationalisation (III)

3. Statistical Bootstrapping
– One of the problems with the Sobel Test is that it
assumes normality in the distribution of variables
which limits its appropriate application
 Statistical bootstrapping does not make this assumption
– Not only is this non-parametric technique applicable with
non-normally distributed variables, but it also retains its
reliability with lower sample sizes as compared to the
Sobel Test (see Preacher & Hayes, 2004).

25
 Operationalisation (IV)

4. Statistical Path Analysis (often within

“Structural Equation Modelling”, SEM)
– Commonly incorporates the above Bootstrapping
approach within a broader statistical modelling
framework noted to be an especially suitable
technique for, “theory driven tests of hypotheses
of causal mediation” Reynolds and Ou (2003)
– A good overview is provided by Tatsuoka (1973)
who documents both the historical origins of path
analysis and provides an account of its initial
26 take-up by educational researchers.
 Note: “Full” vs. “Partial” Mediation

 “Full” AKA ”Complete” Mediation:

– When the (a priori established) impact of X on Y
falls to insignificance when the intervening
mediator Z is considered
 “Partial” Mediation:
– When the (a priori established) impact of X on Y
remains statistically significant when the
intervening mediator Z is considered
 For much more detail, see:
27 – http://davidakenny.net/cm/mediate.htm
 Operationalisation (V)

Moderation (a unidirectional hypothesis remember)

 Explicit (unidirectional; uni-directional):
1. Random Slope Effects
 (special conditions necessary in data though)

 Implicit:
2. Sub-group comparisons [e.g. a girl sample and a boy sample]
3. “Mixture Modelling”
– “(Statistical) Interaction” (bi-directional as it also tests 2+ variables as, “working
together” or, “have a combined effect”):
4. “Statistical Interaction Terms”
– Y=X+Z+XZ [+e]
5. (v.arguably) Partial-Correlations [compare to Bivariate equivalents]
6. (v.arguably) Configural Frequency Analysis
– Others?

28
 Note: “Variable-based” versus
“Person-based” Approaches

 Mediation is concerned solely with how variables relate to one

another – It is tested with “Variable-based” statistical techniques

 Moderation is concerned both with the relationships between

variables as well as distinctions between groups of people
– This is reflected in a distinction between the means of testing
hypotheses of moderation:

29
 Aside speculation: “Self-Moderation”?

 If specifying an interaction term [Y=X+Z+XZ] allows us appropriately test

hypotheses of moderation
 Then does specifying a quadratic term [Y=X+XX] equally allow us to test a
hypothesis where a measured concept [X] implicitly limits its own relationship to
something else [Y]?
– e.g. Psychoactive drug use on neurological effects?
– Evident in Quantile Regression? Any effect from moving averages?

 A strong underlying causal mechanism would seem to be needed in order to make

this leap in interpretation
 This idea would make Masten’s (2001) “risk-activated moderators” equivalent to
“indirect self-moderation”
– The risk moderates it’s own effect by triggering an intermediate moderator
 [note: mediation here]
(Note: [XZ] also captures non-causal “working together” just as [XX] captures a
30 quadratic effect)
 Part 2 – Practical Issues
(theory in practice)

31
 (some) means of testing hypotheses of “moderation”

…Back in 1986….
 Baron & Kenny laid down detailed guidelines concerning how researchers should
attempt to actually test their hypothesis of moderation:
– The best means of testing a hypothesis of moderation depends upon how the
moderated and moderating variables are measured:
 Ordinal/Continuous
 Dichotomous/Categorical

 In response to practice at the time, Baron & Kenny were very specific over when it
was acceptable to dichotomize a hypothesized moderator originally measured on a
continuous scale:
1. When the moderator effect is hypothesized to be a step function
2. When the moderator effect is hypothesized to be quadratic and the
independent/predictor variable is categorical

32
 …Fast-forward fifteen years (to 2001)….
 Kraemer et al. refer to a “struggle” between two approaches that are used for
testing hypotheses of moderation:
1) Sub-group comparisons (that commonly dichotomise samples)
2) Statistical Interaction Terms
3) Compare and Contrast these to Random Slope Effects

33
 1. Sub-group comparisons

 Only provide an indirect test of moderation

– Where the moderation effect is captured by a bi-variate
relationship differing between 2+ groups of people
 (Thus making it a person-based approach to moderation)

 If the hypothesised moderator was originally measured on an ordinal

or continuous scale then an intermediate step in the analysis is
needed whereby sub-groups are created by
categorising/dichotomising the ordinal/continuous moderator
– This intermediate step is controversial though as it discards data, can warp
distributions, and ultimately, can produce highly erroneous findings
 Examples:
– An effect separately tested within boys and girls
– An effect that varies between ‘high’, ‘medium’, and ‘low’ levels of family income
34
 2. Statistical Interaction Terms

 A multi-stage procedure:
1. Mean-centre your predictor [X] and moderator [Z] variables
2. Construct a new ‘interaction’ variable of the form:
predictor x moderator [XZ]
3. Use this variable as a predictor of your outcome(s) [Y] along
with the original variables [X, Z]
 Y=X+Z+XZ [+e]
4. Do not interpret the standardised regression co-efficients of the
[XZ] statistical interaction term – interpret only those
unstandardised
1. If z-scoring is used to mean-centre, then the metric of these unstandardised
coefficients will be in standard deviations

35
 (Example using the ITALASSI software*)

36  *http://www.provalisresearch.com/ITALASSI/ITALASSI.php
 ...A reminder...

 Kraemer et al. (2001) differentiate conclusions from

the statistics that underlie these
– “In short, statistical interaction is a property of which linear model the
researcher selects, not a property of the population… or outcome.”

 Wu & Zumbo (2007) present a good argument for why neither

“statistical interaction terms” and “(statistical) interaction”
constitute a test of moderation on their own
 Remember, moderation is a hypothesis whereby a causal
relationship is specified, not something one can conclude from
mere statistics
– Statistics alone cannot answer a hypotheses of moderation – a
sound conceptual understanding is just as important
37
 3. Random slope effects

 The most explicit test of moderation.

 Literally tests the graphical representation of
moderation that was earlier presented
S*
X Y
Direct effect
Mo Moderated effect

* Denotes the slope of the regression co-efficient

38 that represents the effect of X on Y
 Description

 Allows the regression slope between two

variables [X, Y] to vary between individuals
 This new variation [s] may then be used in
regression equations of its own:
y y
From: To:

x x
s
39 X Y X Y
 Complications:
 Random Slope Effects require nested data and therefore multi-level
modelling
– As well as a very specific set of relationships to be specified between
variables:
 Moderator [Z] must be at the between level (level 2)
 Moderated relationship [Y = X] must be at the within level (level 1)

 Speculation – there is limited availability of this method within common

statistical packages and thus also to researchers?

 Also, standardised beta coefficients are unavailable with this method

– Though again, z-scoring apriori is an option

Aside: Is the whole of multi-level modelling itself representing sets of

moderating relationships?
40
 Summary

1. Sub-group comparisons
– Simple but an imprecise specification of moderation?
– Different sample sizes = lower power?
– The intermediate step of categorisation/dichotomisation is
heavily criticized

2. Statistical Interaction Terms

– Still only an implicit test of moderation – reliant entirely upon
theory for moderation to be concluded

3. Perhaps Random Slope Effects then?

– Uses ALOT of processing power… Therefore limited in the
number you can simultaneously specify and needs nested data.
41 – Highly limited conditions when these may be specified
 Aside speculation (I): “Exogenous” versus
“Endogenous” Moderators?

 Terms borrowed from Econometrics

– These terms have different meanings within SEM though!
 More simply, “Within-system” moderators versus “between-system
moderators” then
– Suitability of the means of testing a moderation varies between
these:
 An Endogenous/Within-system moderation:

A B C

(Statistical
Interaction
Term)
D F
42 E
 Aside speculation (II): “Exogenous” versus
“Endogenous” Moderators?

 An Exogenous/Between-system moderation:

(Sub-group comparisons)
Group A: Group B:

A B C A B C

D E D E F
F

43
 Part 3 – An example of how to test
an hypothesis of moderation

44
 A contrived example (I)

Theoretical Background
1. A mother’s age at the birth of her child is known to significantly impact her
child’s cognitive development
– Children of younger mothers demonstrate are likely to demonstrate
slower cognitive development

2. High quality pre-school is theorised to partially protect children from such

outcomes
– Such that attending pre-schools that are of high quality should moderate
the relationship between a mother’s age at child-birth and her child’s
subsequent cognitive development

 Let us therefore apply the last 3 alternatives were have covered for testing
this hypothesis of moderation.
45
 A contrived example (II)

Outline of measures, data format, and analysis strategy

Multi-level Structural Equation Modelling (SEM)
– (unfortunately a necessary complication to permit this contrived example)

Outcome variable [Y]:

– General Cognitive Ability (GCA) at mean age 58 months
 As measured by the British Ability Scales
 (at entry to reception class)
Predictors:
– General Cognitive Ability at mean age 36 months
 (at entry to pre-school)
– Mother-age at child-birth [X]
– Overall Pre-school ‘Quality’ as assessed by the ECERS-R [Z]
 (the hypothesised moderator)
– (Aside: Uncorrelated with mother’s age)

46 Y=X+Z+XZ [+e]
 1 .Sub-group Comparison
 A “multi-level mixture model” via the Mplus software package
– In which a [mean  1 standard deviation] dichotomisation
strategy was used to form groups of n=538 and n=393
children attending ‘low’ and ‘high’ quality pre-schools

 In which the following effects of mother-age [X] on age 58

month cognitive ability [Y] were found:
• ‘Low’ quality: 0.11 (p<0.001) (standardised beta)
• ‘High’ quality: -0.01 (p=n.s.) (standardised beta)
• Significantly different: (t929=2.7174, p=0.0067)

Conclusion:
47 “A differential impact of mother’s age upon child cognitive
development was found in low vs. high quality pre-schools”
 2. Statistical Interaction Term
 A “multi-level path model” via the Mplus software package
– Y=X+Z+XZ
 Significant impacts on age 58 month cognitive ability [Y] from:
1) ability at 36 months, p<0.001
2) mother-age at child-birth, p<0.001 [X]
3) pre-school quality x mother-age, p<0.05 [XZ]
1) [No significant main effect of quality, Z]

Conclusion:
“Mother’s age has a smaller effect on children’s general cognitive
ability the higher the quality of the pre-school these children
attended”
48 – the same conclusion as from sub-group comparisons
 3. Random Slope Effects
 A “multi-level path model with random effects” via the Mplus
software package

 Significant impacts on reception entry cognitive ability from:

1) earlier abilities and maternal age [X] (both p<0.001)
2) Pre-school quality negatively (p<0.05) impacted the
regression slope between mother-age and age 56 month
cognitive ability [XZ]

Conclusion:
“Mother’s age has a smaller effect on children’s general cognitive
ability the higher the quality of the pre-school these children
attended”
– the same conclusion as from sub-group comparisons
49
 Conclusions from this example

 …All means of testing addressed the hypothesis of moderation,

but in quite different numerical ways…

 The precise information one gets from each approach varied:

– We only got direct protector (moderator) effects [Y=x+Z+xz]
with Statistical Interaction Terms and Random Slopes Effects
– We only got interpretable standardised regression coefficients
from the Sub-Group Comparison procedure

 Final caveat:
– This was only a very small example using only 3 variables and
1 hypothesised moderator!
50
 ...The Final Slide
 If you are interested in the history of the terms, “moderation” and
“mediation” as they have been developed and disseminated through the
social sciences, then do read this paper:
– Kenny, D.A. (2008). Reflections on Mediation. Organizational Research Methods, 11(2),
353-358
 http://davidakenny.net/doc/orm.pdf

 Today’s content is more fully explored in an up-coming book chapter:

– Hall, J. & Sammons, P. (forthcoming). Mediation, Moderation, & Interaction: Definitions,
Discrimination & (Some) Means of Testing. In Teo, T. (Ed.) Handbook of Quantitative
Methods for Educational Research. Rotterdam: Sense Publishers

 Any Questions?
– Thank you for coming, your patience, and your attention!

51
 References
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social
psychological research: conceptual, strategic, and statistical considerations. Journal of
Personality and Social Psychology, 51, 1173-1182.
Essex, M. J., Kraemer, H. C., Armstrong, J. M., Boyce, W. T., Goldsmith, H. H., Klein, M. H., et al.
(2006). Exploring Risk Factors for the Emergence of Children’s Mental Health Problems.
Archives of General Psychiatry, 63, 1246-1256.
Farrington, D. P., & Loeber, R. (2000). Some benefits of dichotomization in psychiatric and
criminological research. Criminal Behaviour and Mental Health, 10, 100-122.
Hinshaw, S. P. (2002). Intervention research, theoretical mechanisms, and causal processes
related to externalizing behavior patterns. Development and Psychopathology, 14, 798-818.
Kraemer, H. C., Stice, E., Kazdin, A., Offord, D., & Kupfer, D. (2001). How Do Risk Factors Work
Together? Mediators, Moderators, and Independent, Overlapping, and Proxy Risk Factors.
American Journal of Psychiatry, 158, 848-856.
MacCallum, R. C., Zhang, S., Preacher, K. J., & Rucker, D. D. (2002). On the Practice of
Dichotomization of Quantitative Variables. Psychological Methods, 7(1), 19-40.
Masten, A. S. (2001). Ordinary magic: Resilience processes in development. American
Psychologist, 56(3), 227-238
Masten, A. S. (2007). Resilience in developing systems: Progress and promise as the fourth wave
rises. Development and Psychopathology, 19(3), 921-930.
Nicholson, R. A., Hursey, K. G., & Nash, J. M. (2005). Moderators and Mediators of Behavioral
Treatment for Headache. Headache, 45, 513-519.
Wu, A. D., & Zumbo, B. D. (2007). Understanding and Using Mediators and Moderators. Social
52 Indicators Research: An International Interdisciplinary Journal for Quality of Life Measurement.
 Technical Appendices
MPLUS SUBGROUP COMPARISON (VIA MODEL:
DICHOTOMISATION) SYNTAX: %WITHIN%
MISSING ARE ALL (-999999); %OVERALL%
idvariable = childid; rgcam on bgcam;
CLUSTER = centreid; rgcam on q53am;
WITHIN = bgcam q53am;
CENTERING = GRANDMEAN (rgcam); %group#1%
CLASSES = group (2); rgcam on bgcam;
KNOWNCLASS = group (group1=0 rgcam on q53am;
group1=1);
DEFINE:
%group#2%
IF (ecers_r LE -1.163) THEN group1=0;
rgcam on bgcam;
IF (ecers_r GE 0.846) THEN group1=1;
rgcam on q53am;
ANALYSIS:
TYPE = MIXTURE TWOLEVEL;
53 ALGORITHM = INTEGRATION;
 MPLUS INTERACTION TERM SYNTAX:
MISSING ARE ALL (-999999);
idvariable = childid;
CLUSTER = centreid;
BETWEEN = ecers_r;
WITHIN = q53am bgcam ;
CENTERING = GRANDMEAN (rgcam);
MODEL:
DEFINE: qualage = ecers_r*q53am; %WITHIN%
rgcam on bgcam q53am;
ANALYSIS:
%BETWEEN%
TYPE = RANDOM TWOLEVEL;
rgcam on ecers_r ;
54 ALGORITHM = INTEGRATION; rgcam with qualage;
 MPLUS RANDOM SLOPES SYNTAX:
MISSING ARE ALL (-999999);
idvariable = childid;
CLUSTER = centreid;
BETWEEN = ecers_r;
WITHIN = q53am bgcam; MODEL:
CENTERING = GRANDMEAN %WITHIN%
(rgcam); rgcam on bgcam;
ANALYSIS: s | rgcam on q53am;
TYPE = RANDOM TWOLEVEL;
%BETWEEN%
ALGORITHM = INTEGRATION;
s on ecers_r;
rgcam on ecers_r;
55
 Previous Version of this
Presentation – pre-2013

56
 Interaction, Moderation, and Mediation:
Definitions, Discrimination, and (some) means of testing

Quantitative SIG
February 6th 2012
[email protected]
Department of Education
University of Oxford

57
 Preliminary note:
 This presentation has been written for a mixed audience of undergraduates-through-
professors and from those starting-out in statistics, through those who routinely carry
out quantitative analyses, all the way to those who routinely teach/tutor on this subject.
 As such, the content of this presentation is mixed:
1. Part of it is purely pedagogical and this is for the benefit of those who come to
this with little prior knowledge
2. Part of it is for those who feel that they know a little and who come to this seeking
clarification and/or guidelines
3. The final part of this is for those who feel confident on these topics
 For them, there are novel speculations designed to test their preconceived
ideas
 However, although this means that there is content for everyone within this
presentation, this also means that there is content which will either seem too simple or
too advanced – I’m afraid that this is unavoidable....
– ...As such, please try not to feel too frustrated at such content!

58
 Abstract
 In 1986 Baron and Kenny set out to clarify the differences between the terms
“Moderation” and “Mediation” as used in the social sciences.
 Twenty six years later, the seminal paper that this collaboration resulted in
(published in the Journal of Personality and Social Psychology) has been cited
around 30,347 times (Google Scholar on 12/01/2012).
– This is an average of 1,167 each year - the equivalent of more than once every 8 hours for over a quarter of
a century.
 However and despite this citation record, the uncertainty surrounding these terms
has not gone away.
– Academics still struggle to define, distinguish and utilise these terms while related undergraduate teaching
is still an exception.
 This presentation sets out simple, clear definitions that distinguish “Interaction”,
“Moderation”, and “Mediation” as well as a number of other commonly-used terms.
– An introduction is given on how to use these concepts in ‘real-life research’ with worked-through examples
provided.
 The presentation slides themselves also serve as a short primer for future reference.

59
 Contents
Part 1 – Theoretical Issues :
– Definitions and discrimination
 (Including 2 ways of testing hypotheses of “mediation”)

Part 2 – Practical Issues (theory in practice):

– (some) means of testing hypotheses of “moderation:”
1. Sub-group Comparison
2. Statistical Interaction Term
3. Random Slope Effect
– A contrived example of how to test a hypothesis of moderation.
Triple-analysis giving comparisons between:
1. A Sub-group Comparison
2. A Statistical Interaction Term
60 3. A Random Slope Effect
 Part 1 – Theoretical Issues

61
 Definitions I

 (Statistical) Interaction
– A phrase implying 2+ variables “work together” to impact a third
– ≠ behavioural/psychological interaction!
 Moderation
– A tri-variate hypothesis:
 Under what conditions/for whom/when is a pre-established
statistical relationship evident?
– (Often addressed through ‘Statistical Interactions’)
 Mediation
– A distinctive & different tri-variate hypothesis:
 What are the mechanisms of effect underlying a pre-
established causal statistical relationship?
62
 Definitions II
 Understanding Moderation and Mediation to be tri-variate hypotheses gives the quantitative
researcher an intellectual framework to postulate complex but still testable hypotheses
 Compare and contrast to common undergraduate teaching where:
– The link between formulating hypotheses and matching quantitative analyses can be
missed due to different lecturers leading different classes
– Even when the link between hypothesis formulation and testing is not missed, it can be
limited to:
1. bi-variate tests of “differences” (t-test, [M]AN[C]OVA & non.par.eq.)
2. bi-variate tests of “associations” (correlations, regressions)
3. It is not unusual to find undergraduate social science statistics courses commonly
ended in nebulous descriptions of “multi-variate” statistics which aren’t explicitly linked
to equivalent “multi-variate” hypotheses
 Ultimately then: How are undergraduate students even supposed to know how to formulate
hypotheses concerning 3+ concepts without an underlying intellectual framework?
– Moderation and Mediation provide this and link to statistical analyses

63
 Statistical Interaction ≠ Moderation

 Moderation is a uni-directional hypothesis

– “The relationship between X and Y varies by Z”

 Statistical interaction is a bi-directional statistical phrase,

– “There is a combined effect of X and Z on Y”

 Thus, to conclude moderation from a statistical interaction involves

relying heavily upon pre-existing knowledge
– Be this: research findings, a priori theories, on-the-ground
knowledge etc
– Moderation from interaction is broadly analogous to testing

64 bivariate uni-directional hypotheses with correlation coefficients

 “(Statistical) Interaction” ≠ “Statistical Interaction Terms”

 Again, “(Statistical) Interaction” is a phrase used to imply that 2+

variables “work together”/“have a combined effect” upon a third

 A “Statistical Interaction Term” is a statistical artefact whereby 2+

variables are zero-centred (typically by z-scoring), multiplied
together, and commonly entered into a regression equation (though
also ANOVA) of the form (here: tri-variate):
– Y=X+Z+XZ [+e]
 Note, Statistical Interaction Terms include a “main effect” of the moderator
[here: Z] as well as a combined effect [XZ] .
 Though the inclusion and significance of this main effect is completely
unnecessary in establishing moderation, it is necessary when specifying a
“Statistical Interaction Term”.

65
 Mediation ≠ Moderation

The classical diagrams:

Mediation: Moderation:
Answers hypotheses of, Answers hypotheses of,
“How” and, “Why”* “When” and, “For whom”*

X Y X Y

Me Mo
Me = Mediator Direct effects
Mo = Moderator Moderated effect
66 * Wu & Zumbo (2007)
 Why the confusion?
 Well, Moderation and Mediation are both…
1. …“theories for refining and understanding a causal relationship” (Wu &
Zumbo, 2007)
2. ...uni-directional hypotheses
3. ...tri-variate hypotheses

 They may even be combined to provide a powerful intellectual framework for

specifying and testing multi-variate hypotheses
– E.g. “How” does X manage to impact Y - is it via intermediate effects
on Z [mediation]? Further, under what conditions of W does Z express
this role [moderation]?

67
 Aside: “Indirect effects”

 “Indirect effect” is another term that can be encountered in

discussions of mediation, statistical interaction, and statistical
interaction terms.
 Both mediation and statistical interaction feature effects that
can be thought of as “indirect” but these are quite distinct and
shouldn’t be confused.
 Within Structural Equation Modelling (SEM) software and texts
it is usually common to find the tern “indirect effect” strictly
limited to those indirect effects that are associated with
mediation (only)

68
*Often only in the context
of experimental research
though
Kraemer et al. (2001) Hinshaw (2002) Nicholson et al. (2005) Essex et al. (2006) Wu & Zumbo (2007)
+-+-+Moderation+-+-+
Moderator has temporal Moderator has Moderator has Moderator has Moderator has
precedence precedence precedence precedence precedence
Moderator and that Moderator and that Moderator and that Moderator and that Moderator and that
moderated are moderated are moderated are moderated are moderated are
uncorrelated uncorrelated uncorrelated uncorrelated uncorrelated
Some Past Guides*

Co-domination of
moderated and Answers, "for whom"
moderator and "when"
Moderator is a trait
Moderator is observed
+-+-+Mediation+-+-+
That being mediated That being mediated That being mediated That being mediated has That being mediated has
has precedence has precedence has precedence precedence precedence
Mediator and that Mediator and that Mediator and that Mediator and that Mediator and that
mediated are mediated are mediated are mediated are mediated are
correlated correlated correlated correlated correlated
Either co-domination of Either co-domination of
mediated and mediated and
mediator (partial) mediator (partial) Answers, "how" and
OR OR "why"
Mediator dominates
that mediated Mediator dominates that
(total) mediated (total) Mediator is a state

69 Mediator is observed or
manipulated
 ...Real-world ambiguities (I)...
 Even with unambiguous definitions, both hypotheses can be
appropriate for the same set of measures
– In developmental science, this often depends upon the timeframe
– For example, Masten (2007) describes the relationship between
adversity (X), stress-response (Y) and stress-regulators (Z):

Period A, Period B,
stress-regulators developing - mediation stress-regulators developed - moderation

Adversity Stress- Adversity Stress-

Regulators (t 2) Response

Stress- Stress-
Regulators (t1) Direct effect Regulators
Moderated effect
70
 ...Real-world ambiguities (II)...

 Masten (2001) also gives the developmental science

example of...
– “...a risk-activated moderator analogous to an automobile airbag or
immune system response.”

71 – Aside: This is also an excellent description of the ideal functioning

of social interventions such as Head Start and Sure Start
 Combining moderation and mediation

 A, “risk-activated moderator” (Masten, 2001) is

just one example of how moderation and
mediation may be integrated.
 The more (in)famous examples are
– “moderated-mediation”
– “mediated-moderation”
 Both are 4 variable hypotheses [W, X, Y, Z] that postulate
conditional (moderated) mechanisms of effect (mediation)
whereby X impacts Y

72 (Aside: Wu and Zumbo (2007) provide further excellent descriptions and discussion)
 *SPSS macro available from: http://www.afhayes.com/spss-sas-and-mplus-macros-and-code.html
Moderated-mediation*

X Y

Me

Mo
Hypothesising the ‘when’ and ‘for whom’ of
an initially mediated relationship

Me = Mediator Direct effect

Mo = Moderator Moderated effect
73 Additional effects
 Mediated-moderation

X Y

Mo Me

Hypothesising the ‘how’ and ‘why’ of an

initially moderated relationship

Me = Mediator Direct effect

Mo = Moderator Moderated effect
74 Additional effect
 Operationalisation (I)

Mediation
 Two common means of testing are:
1. The Sobel Test
 E.g. http://www.danielsoper.com/statcalc3/calc.aspx?id=31

2. Path Analysis
 “StructuralEquation Modelling (SEM)” if latent variables
are being specified

75
 Note: “Full” vs. “Partial” Mediation

 “Full” AKA ”Complete” Mediation:

– When the (a priori established) impact of X on Y
falls to insignificance when the intervening
mediator Z is considered
 “Partial” Mediation:
– When the (a priori established) impact of X on Y
remains statistically significant when the
intervening mediator Z is considered
 For much more detail, see:
76 – http://davidakenny.net/cm/mediate.htm
 Operationalisation (II)

Moderation (a uni-directional hypothesis remember)

 Explicit (uni-directional):
1. Random Slope Effects
 (special conditions necessary in data though)

 Implicit:
2. Sub-group comparisons [e.g. a girl sample and a boy sample]
3. “Mixture Modelling”
– “(Statistical) Interaction” (bi-directional as it also tests 2+ variables as, “working
together” or, “have a combined effect”):
4. “Statistical Interaction Terms”
– Y=X+Z+XZ [+e]
5. (v.arguably) Partial-Correlations [compare to Bivariate equivalents]
6. (v.arguably) Configural Frequency Analysis
– Others?

77
 Note: Variable-based vs. Person-based

 Mediation is concerned solely with how variables relate to one

another

 Moderation is concerned both with the relationships between

variables as well as distinctions between groups of people
– This is reflected in the means of testing hypotheses of
moderation:

78
 Aside speculation - “Self-Moderation”?

 If specifying an interaction term [Y=X+Z+XZ] allows us appropriately test

hypotheses of moderation
 Then does specifying a quadratic term [Y=X+XX] equally allow us to test a
hypothesis where a measured concept [X] implicitly limits its own relationship to
something else [Y]?
– e.g. Psychoactive drug use on neurological effects?
– Evident in Quantile Regression? Any effect from moving averages?

 A strong underlying causal mechanism would seem to be needed in order to make

this leap in interpretation
 This idea would make Masten’s (2001) “risk-activated moderators” equivalent to
“indirect self-moderation”
– The risk moderates it’s own effect by triggering an intermediate moderator
 [note: mediation here]
(Note: [XZ] also captures non-causal “working together” just as [XX] captures a
79 quadratic effect)
 Part 2 – Practical Issues
(theory in practice)

80
 (some) means of testing hypotheses of “moderation”

…Back in 1986….
 Baron & Kenny laid down detailed guidelines concerning how researchers should
attempt to actually test their hypothesis of moderation:
– The best means of testing a hypothesis of moderation depends upon how the
moderated and moderating variables are measured:
 Ordinal/Continuous
 Dichotomous/Categorical

 In response to practice at the time, Baron & Kenny were very specific over when it
was acceptable to dichotomize a hypothesized moderator originally measured on a
continuous scale:
1. When the moderator effect is hypothesized to be a step function
2. When the moderator effect is hypothesized to be quadratic and the
independent/predictor variable is categorical

81
 …Fast-forward fifteen years (to 2001)….
 Kraemer et al. refer to a “struggle” between two approaches that are used for
testing hypotheses of moderation:
1) Sub-group comparisons (that commonly dichotomise samples)
2) Statistical Interaction Terms
3) Compare and Contrast these to Random Slope Effects

82
 1. Sub-group comparisons

 Only provide an indirect test of moderation

– Where the moderation effect is captured by a bi-variate
relationship differing between 2+ groups of people
 (Thus making it a person-based approach to moderation)

 If the hypothesised moderator was originally measured on an ordinal

or continuous scale then an intermediate step in the analysis is
needed whereby sub-groups are created by
categorising/dichotomising the ordinal/continuous moderator
– This intermediate step is controversial though as it discards data, can warp
distributions, and ultimately, can produce highly erroneous findings
 Examples:
– An effect separately tested within boys and girls
– An effect that varies between ‘high’, ‘medium’, and ‘low’ levels of family income
83
 2. Statistical Interaction Terms

 A multi-stage procedure:
1. Mean-centre your predictor [X] and moderator [Z] variables
2. Construct a new ‘interaction’ variable of the form:
predictor x moderator [XZ]
3. Use this variable as a predictor of your outcome(s) [Y] along
with the original variables [X, Z]
 Y=X+Z+XZ [+e]
4. Do not interpret the standardised regression co-efficients of the
[XZ] statistical interaction term – interpret only those
unstandardised
1. If z-scoring is used to mean-centre, then the metric of these unstandardised
coefficients will be in standard deviations

84
 (Example using the ITALASSI software*)

85  *http://www.provalisresearch.com/ITALASSI/ITALASSI.php
 ...A reminder...

 Kraemer et al. (2001) differentiate conclusions from

the statistics that underlie these
– “In short, statistical interaction is a property of which linear model the
researcher selects, not a property of the population… or outcome.”

 Wu & Zumbo (2007) present a good argument for why neither

“statistical interaction terms” and “(statistical) interaction”
constitute a test of moderation on their own
 Remember, moderation is a hypothesis whereby a causal
relationship is specified, not something one can conclude from
mere statistics
– Statistics alone cannot answer a hypotheses of moderation – a
sound conceptual understanding is just as important
86
 3. Random slope effects

 The most explicit test of moderation.

 Literally tests the graphical representation of
moderation that was earlier presented
S*
X Y
Direct effect
Mo Moderated effect

* Denotes the slope of the regression co-efficient

87 that represents the effect of X on Y
 Description

 Allows the regression slope between two

variables [X, Y] to vary between individuals
 This new variation [s] may then be used in
regression equations of its own:
y y
From: To:

x x
s
88 X Y X Y
 Complications:
 Random Slope Effects require nested data and therefore multi-level
modelling
– As well as a very specific set of relationships to be specified between
variables:
 Moderator [Z] must be at the between level (level 2)
 Moderated relationship [Y = X] must be at the within level (level 1)

 Speculation – there is limited availability of this method within common

statistical packages and thus also to researchers?

 Also, standardised beta coefficients are unavailable with this method

– Though again, z-scoring apriori is an option

Aside: Is the whole of multi-level modelling itself representing sets of

moderating relationships?
89
 Summary

1. Sub-group comparisons
– Simple but an imprecise specification of moderation?
– Different sample sizes = lower power?
– The intermediate step of categorisation/dichotomisation is
heavily criticized

2. Statistical Interaction Terms

– Still only an implicit test of moderation – reliant entirely upon
theory for moderation to be concluded

3. Perhaps Random Slope Effects then?

– Uses ALOT of processing power… Therefore limited in the
number you can simultaneously specify and needs nested data.
90 – Highly limited conditions when these may be specified
 A contrived example (I)

Theoretical Background
1. A mother’s age at the birth of her child is known to significantly impact her
child’s cognitive development
– Children of younger mothers demonstrate are likely to demonstrate
slower cognitive development

2. High quality pre-school is theorised to partially protect children from such

outcomes
– Such that attending pre-schools that are of high quality should moderate
the relationship between a mother’s age at child-birth and her child’s
subsequent cognitive development

 Lets therefore apply the last 3 alternatives were have covered for testing this
hypothesis of moderation.
91
 A contrived example (II)

Outline of measures, data format, and analysis strategy

 Multi-level Structural Equation Modelling (SEM)
– (unfortunately a necessary complication to permit this contrived example)
 Outcome variable [Y]:
– General Cognitive Ability at mean age 58 months
 As measured by the British Ability Scales
 (at entry to reception class)
 Predictors:
– General Cognitive Ability at mean age 36 months
 (at entry to pre-school)
– Mother-age at child-birth [X]
– Overall Pre-school ‘Quality’ as assessed by the ECERS-R [Z]
 (the hypothesised moderator)
– (Aside: Uncorrelated with mother’s age)

92 Y=X+Z+XZ [+e]
 1 .Sub-group Comparison
 A “multi-level mixture model” via the Mplus software package
– In which a [mean  1 standard deviation] dichotomisation
strategy was used to form groups of n=538 and n=393
children attending ‘low’ and ‘high’ quality pre-schools

 In which the following effects of mother-age on age 58 month

cognitive ability were found:
• ‘Low’ quality: 0.11 (p<0.001) (standardised beta)
• ‘High’ quality: -0.01 (p=n.s.) (standardised beta)
• Significantly different: (t929=2.7174, p=0.0067)

Conclusion:
93 “A differential impact of mother’s age upon child cognitive
development was found in low vs. high quality pre-schools”
 2. Statistical Interaction Term
 A “multi-level path model” via the Mplus software package
– Y=X+Z+XZ
 Significant impacts on age 58 month cognitive ability [Y] from:
1) ability at 36 months, p<0.001
2) mother-age at child-birth, p<0.001 [X]
3) pre-school quality x mother-age, p<0.05 [XZ]
1) [No significant main effect of quality, Z]

Conclusion:
“Mother’s age has a smaller effect on children’s general cognitive
ability the higher the quality of the pre-school these children
attended”
94 – the same conclusion as from sub-group comparisons
 3. Random Slope Effects
 A “multi-level path model with random effects” via the Mplus
software package

 Significant impacts on reception entry cognitive ability from:

1) earlier abilities and maternal age (both p<0.001)
2) Pre-school quality negatively (p<0.05) impacted the
regression slope between mother-age and age 56 month
cognitive ability

Conclusion:
“Mother’s age has a smaller effect on children’s general cognitive
ability the higher the quality of the pre-school these children
attended”
– the same conclusion as from sub-group comparisons
95
 Conclusions from this example

 …All means of testing addressed the hypothesis of moderation,

but in quite different numerical ways…

 The precise information one gets from each approach varied:

– We only got direct protector (moderator) effects [Y=x+Z+xz]
with Statistical Interaction Terms and Random Slopes Effects
– We only got interpretable standardised regression coefficients
from the Sub-Group Comparison procedure

 Final caveat:
– This was only a very small example using only 3 variables and
1 hypothesised moderator!
96
 ...The Final Slide
 If you are interested in the history of the terms, “moderation” and
“mediation” as they have been developed and disseminated through
the social sciences, then do read this paper:
– Kenny, D.A. (2008). Reflections on Mediation. Organizational
Research Methods, 11(2), 353-358
 http://davidakenny.net/doc/orm.pdf

 Any Questions?
– Thank you for your patience and attention!

97
 References
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social
psychological research: conceptual, strategic, and statistical considerations. Journal of
Personality and Social Psychology, 51, 1173-1182.
Essex, M. J., Kraemer, H. C., Armstrong, J. M., Boyce, W. T., Goldsmith, H. H., Klein, M. H., et al.
(2006). Exploring Risk Factors for the Emergence of Children’s Mental Health Problems.
Archives of General Psychiatry, 63, 1246-1256.
Farrington, D. P., & Loeber, R. (2000). Some benefits of dichotomization in psychiatric and
criminological research. Criminal Behaviour and Mental Health, 10, 100-122.
Hinshaw, S. P. (2002). Intervention research, theoretical mechanisms, and causal processes
related to externalizing behavior patterns. Development and Psychopathology, 14, 798-818.
Kraemer, H. C., Stice, E., Kazdin, A., Offord, D., & Kupfer, D. (2001). How Do Risk Factors Work
Together? Mediators, Moderators, and Independent, Overlapping, and Proxy Risk Factors.
American Journal of Psychiatry, 158, 848-856.
MacCallum, R. C., Zhang, S., Preacher, K. J., & Rucker, D. D. (2002). On the Practice of
Dichotomization of Quantitative Variables. Psychological Methods, 7(1), 19-40.
Masten, A. S. (2001). Ordinary magic: Resilience processes in development. American
Psychologist, 56(3), 227-238
Masten, A. S. (2007). Resilience in developing systems: Progress and promise as the fourth wave
rises. Development and Psychopathology, 19(3), 921-930.
Nicholson, R. A., Hursey, K. G., & Nash, J. M. (2005). Moderators and Mediators of Behavioral
Treatment for Headache. Headache, 45, 513-519.
Wu, A. D., & Zumbo, B. D. (2007). Understanding and Using Mediators and Moderators. Social
98 Indicators Research: An International Interdisciplinary Journal for Quality of Life Measurement.
 Technical Appendices
MPLUS SUBGROUP COMPARISON (VIA MODEL:
DICHOTOMISATION) SYNTAX: %WITHIN%
MISSING ARE ALL (-999999); %OVERALL%
idvariable = childid; rgcam on bgcam;
CLUSTER = centreid; rgcam on q53am;
WITHIN = bgcam q53am;
CENTERING = GRANDMEAN (rgcam); %group#1%
CLASSES = group (2); rgcam on bgcam;
KNOWNCLASS = group (group1=0 rgcam on q53am;
group1=1);
DEFINE:
%group#2%
IF (ecers_r LE -1.163) THEN group1=0;
rgcam on bgcam;
IF (ecers_r GE 0.846) THEN group1=1;
rgcam on q53am;
ANALYSIS:
TYPE = MIXTURE TWOLEVEL;
99 ALGORITHM = INTEGRATION;
 MPLUS INTERACTION TERM SYNTAX:
MISSING ARE ALL (-999999);
idvariable = childid;
CLUSTER = centreid;
BETWEEN = ecers_r;
WITHIN = q53am bgcam ;
CENTERING = GRANDMEAN (rgcam);
MODEL:
DEFINE: qualage = ecers_r*q53am; %WITHIN%
rgcam on bgcam q53am;
ANALYSIS:
%BETWEEN%
TYPE = RANDOM TWOLEVEL;
rgcam on ecers_r ;
10 ALGORITHM = INTEGRATION; rgcam with qualage;
 MPLUS RANDOM SLOPES SYNTAX:
MISSING ARE ALL (-999999);
idvariable = childid;
CLUSTER = centreid;
BETWEEN = ecers_r;
WITHIN = q53am bgcam; MODEL:
CENTERING = GRANDMEAN %WITHIN%
(rgcam); rgcam on bgcam;
ANALYSIS: s | rgcam on q53am;
TYPE = RANDOM TWOLEVEL;
%BETWEEN%
ALGORITHM = INTEGRATION;
s on ecers_r;
rgcam on ecers_r;
10