Moderated multiple regression using Process macro version 3.

5
Using ‘Multicategorical’ option with categorical moderator

Mike Crowson, Ph.D.

The University of Oklahoma
September, 2020

Download data from: https://drive.google.com/file/d/1wZYs_ua3SSuOdydVdRcL7giHcyqlOfsg/view?usp=sharing

Youtube video demo on using Multicategorical option (Sept 2020): https://youtu.be/IG-FnhvTud4
Video demo with same data with continuous focal and moderator variables (Sept 2020): https://youtu.be/4gGJCX6rrAo
Video and Powerpoint that covers concepts of moderated regression and Process in more depth (2019):
https://www.youtube.com/watch?v=JZ2fl8exNTE
 In this analysis, we are modeling predictors of
depressive personality. Affective lability is our
Affective focal predictor and anxiety is treated as a
lability moderator of the effect of affective lability on
Depress
depression. For this analysis, we will use ‘anxLev’
as a categorical moderator of the relationship
between affective lability (continuous) and
anxLev depressive personality (continuous). Specifically,
this variable is treated as ordered categorical and
contains three levels, coded 1=low anxiety,
2=medium anxiety, 3=high anxiety.
We keep our Model number set to 1.
Here, we go under the Multicategorical tab and indicate that Variable W is categorical. There are various coding systems
allowed, but I will stick with ‘indicator’ since this represents traditional dummy coding of the categorical variable.
Selection of this option will
generate syntax for you to plot
simple slopes

Selecting this option will help

us to better interpret the
intercept in the model.
The first thing to note is how the categorical variable is being
recoded into dummy variables (W1 and W2). Group 1 on
‘anxLev’ is coded with 0’s across W1 and W2, making this
the reference group. Group 2 is coded W1=1 and W2=0.
Group 3 is coded W1=0 and W2=1.
Here, we see the predictors in the model
account for significant variation in
depressive personality characteristics, R-
square=.4063, F(5,460) = 62.9731, p<.001.
The intercept (constant) in the model is the
conditional mean for individuals with values of
0 on the set of predictors in the model. The
only group where that occurs is the reference
group (coded W1=0 and W2=0 on the dummy
variables). Moreover, because I selected ‘Only
continuous variables that define products’, that
means that the affective lability variable has
been mean centered as well [The mean on the
centered variable, therefore, is 0]. Thus, I can
interpret the intercept as the conditional mean
for a person in the reference group. That is, it is
the mean score on depressive traits for persons
in the low anxiety group scoring at the mean on
affective lability.
Since the reference group is coded 0 on both
W1 and W2, the slopes for W1 and W2
represent group comparisons (where we are
holding affective lability constant). Specifically,
the slope for W1 is the difference in conditional
means between persons in the medium anxiety
group and the low anxiety (reference) group
involving persons scoring at the mean on
affective lability. The difference between
groups is .3534, which is significant at p<.001.
The slope for W2 is the difference in means
between persons in the high anxiety group and
the low anxiety (reference group) involving
persons scoring at the mean on affective
lability. The difference is .6835, which is
significant at p<.001.
The conditional mean for persons in the
medium anxiety group who fall at the mean of
affective lability is: 1.4453 + .3534 = 1.7964.
The conditional mean for persons in the high
group is computed as: 1.4453 + .6835 = 2.1288.

So the conditional means for our anxiety

groups are:

Low anxiety: 1.4453

Medium anxiety: 1.7964
High anxiety: 2.1288
The regression slope for affective lability
represents the predictive relationship when W1
and W2 are coded 0 – i.e., the slope for
individuals in the reference group.

We see this slope is positive and significant

(b=.3025, s.e.=.1010, p=.0029).
Finally, we have our interaction terms in the
model.

Int_1 reflects the interaction between affective

lability and the W1 dummy variable.

Int_1 reflects the interaction between affective

lability and the W2 dummy variable.

The slopes for these terms represent the

difference in the slopes for the middle and high
anxiety groups and that of the low anxiety
(reference group). For example, the slope for
the middle anxiety group (where W1 = 1)
is: .2866 + .3025 = .5891. The slope for the high
anxiety group (where W2 = 1) is: .3025 + .0057
= .3082
Given the slope for Int_1 is statistically
significant (p=.0264), we infer that there is a
significant difference in the slope for
affective lability between the low anxiety
(reference) group and the medium anxiety
group (i.e., group 2). The non-significant
slope for Int_2 means that there is no
difference in regression slopes for affective
lability between persons in the high anxiety
group (i.e., group 3) and the low anxiety
(reference) group.
You can think of this result as an omnibus test of the interaction between affective lability and anxiety level in the model.
Specifically, the R-square change is can be thought of as the change in R-square between a model that does not include
the interaction terms and the current model that does.

The R-square change is statistically significant, which indicates that the slopes vary significantly across groups. This means
that we have evidence of a significant interaction between group membership and affective lability.
Here, we have the simple slopes reflecting the predictive relationship between affective lability and depression in
each of our three groups. Recall that the slope for affective lability in our regression table is the conditional effect of
affective lability on depression in group 1 (i.e., the reference group).
If we sum the slope (see below) for affective lability (for group 1) and the difference in slopes with group 2 (see
slope for Int_1), we get: .3025 + .2866 = .5891 (which you find in the table of conditional effects above).
If we sum the slope (see below) for affective lability (for group 1) and the difference in slopes with group 3 (see
slope for Int_2), we get: .3025 + .0057 = .3082 (which you find in the table of conditional effects above).
As you can see, the simple slopes are all positive and statistically significant. Interestingly the slope for the middle
anxiety group was stronger than it was for the low and high anxiety groups.
We can copy and paste the syntax from
our output file into a syntax file and run
to plot out the simple slopes.
Appendix A: Generating moderated multiple regression results through General
linear model menu in SPSS
We can generate regression results that contains the interaction terms in SPSS by using the General Linear Model
menu. When you select ‘Parameter estimates’ the regression output will be generated. However, the intercept will be
for the group on the factor variable coded highest. So with the current coding of the moderator variable (1=low anxiety,
2=medium anxiety, 3=high anxiety), the intercept would reflect the mean of the high anxiety group. To have our output
align with what we generated through process, we need to recode our variable so that 3=low anxiety, 2=medium
anxiety, 1=high anxiety).
Original variable coding New variable Value codes for
new variable
Our factor variable is entered as a fixed factor. Our centered focal predictor (affective lability) is entered as a covariate. Next,
we click on the Model tab…
Under the Model tab, select Build Terms. Enter each of the predictors as separately (Click Type drop-down and select ‘Main
effects’). To add the interaction term, select ‘Interaction’ and highlight the two predictors at the same time and move to box
on the right.
You can select the same things you typically would if
running an ANOVA. But to get the model in
regression format, click on ‘Parameter estimates’.
We can obtain marginal means via this
route.
The corrected Model F-test and
partial eta-squared in the GLM
output will be the same as that
shown in our Process output for
the overall model fit.
Here we have our model
presented via the GLM route as a
regression model.

Here are the results presented in

Process. The values are the same.
Intercept, which is the mean
depression score among
persons in the low anxiety group
falling at the mean on affective
lability.
Slope for the centered affective
lability predictor, reflecting its
relationship to depressive traits in
the low anxiety group.
Difference in conditional
means between the high
anxiety group and the low
anxiety group.
Difference in conditional means
between the medium and low
anxiety groups.
Difference in regression slope
for affective lability between the
high anxiety group and the low
anxiety group.
Difference in regression slope
for affective lability between the
medium anxiety group and the
low anxiety group.
Marginal means on anxiety
Appendix B: ‘What if my moderator is continuous and my focal IV is categorical?’
This is the model we will generate output for…
Here, I’ve placed the categorical anxiety variable in the X box and selected Variable X under multicategorical. I’m keeping
the coding as ‘indicator.’ The moderator Variable W (which is continuous) is not selected under the M’ulticategorical’ tab.
The categorical re-coding scheme is applied to the X variable (in this case, anxLev).
The dummy coding for the X variable is the same as that applied to the previous
analysis with the moderator W.
Current results: The regression results here
are the same since the both models are
mathematically equivalent. The only
difference is that the X variables (above)
were previously W variables (below), and
there’s some reordering of the variables in
the table.

Previous model results

 Current results

Recall that X1 is the difference between the medium anxiety group (i.e., 2) and the low anxiety group (i.e., 1), and X2 is
the difference between the high anxiety (i.e., 3) and low anxiety groups, for persons scoring at the mean of the
moderator (since we requested that our continuous variables in any interaction term be mean centered prior to forming
product terms).

The conditional mean for low anxiety group is 1.4453.

The conditional mean for the medium anxiety group is 1.4453 + .3534 = 1.7987
The conditional mean for the high anxiety group is 1.4453 + .6835 = 2.1288
The simple effects/slopes tests look a little bit different from what we saw before. They are broken up into tables
reflecting differences in marginal group means at different levels of the moderator.

Variable X1 is the difference between the medium group (group 2) and the low group (group 1), and variable X2 is the
mean difference between the high group (group 3) and the low group. For persons scoring 1 sd below the mean on the
continuous moderator (affective lability), the difference between the medium and high group means is .1516 (p=.0780).
The difference between the high group and the low group is .6975 (p<.001).

-1sd

Group means
mean
Simple effects/slopes tests at the
mean of the moderator contain
same values that show up in
original regression table.
 +1sd
Simple effects/slopes tests at +1 sd
of the moderator.

For persons scoring 1 sd above the mean on the continuous moderator (affective lability), the difference between the
medium and high group means is .5552 (p=.0002). The difference between the high group and the low group is .6876
(p<.001).
Visualization of group means at three levels of
the moderator.

The anxiety groups differ most for persons falling

at the mean on affective lability.
Appendix C: ‘What if my moderator and focal predictor are categorical?’
In this short demo, we have a categorical affective lability variable (coded 1=low, 2=medium, 3=high) moderating the
effect of anxiety level (coded 1=low, 2=medium, 3=high) on depressive traits. Under the multicategorical tab you will
select both Variable X and Variable Y.
Both variables are recoded into dummy
variables. The ‘low’ group for each
variable serves as the reference group
(since it is coded 0 across dummies).
The intercept is the mean depressive score for persons who are low on BOTH anxiety and affective lability. The mean for these
individuals is 1.200. The slope X1 is the mean difference (group 2 – group 1 = .2524, p=.0188) between persons medium and
low on depressive traits for persons who are low in affective lability. The slope for X2 is the mean difference (group 3 – group 1
= .7400, p<.001) between persons high and low on depressive traits for persons low on affective lability. The slope for W1 is
the difference in means (group 2 – group 1 = .2190, p=.0412) between persons medium on affective lability and low on
affective lability among those who are low in anxiety. The slope for W2 is the difference in means (group 3 – group 1 = .5333,
p=.0029) between persons who are high and low on affective lability who are in the low anxiety group.
The interaction terms various parts of the total interaction between the two variables.

The test of R-square change here tests the overall

interaction effect. Notice it is not significant.
Plot of marginal means based on syntax
provided in output
If we re-run the model as a factorial ANOVA through General Linear Model route (after recoding our categorical
affective lability variable in the manner shown in Appendix A)…
If we re-run the model as a factorial ANOVA through General Linear Model route (after recoding our categorical
affective lability variable in the manner shown in Appendix A)…
…our results in the Parameter estimates table will be the same as those from the Process output.