I am testing an interaction effect where $X$ and $Y$ are continuous variable and $M$ (Moderator) is a categorical variable (effects coding $+1$, $-1$).

I have no clue about how to do a post-hoc probing of slopes in this case.

For a continuous moderator we calculate $Z_{\rm above}$ and $Z_{\rm low}$, calculate their crossproduct with $X$ (e.g., $Z_{\rm Above}*X$ and $Z_{\rm low}*X$) and then run the regression with their combinations of ${\rm High}$ and ${\rm Low}$ with $Y$.

I do not see that the same method above could be applied for a categorical moderator. How can I test that?


1 Answer 1


I'm not sure I follow your strategy with continuous moderators, but the proper approach with categorical moderators is essentially the same as with continuous ones. Effect coding is a perfectly reasonable strategy for representing categorical covariates, and it makes no difference in terms of how you test the interaction. To test an interaction, we add a product term to the model, and test the beta associated with that term. Assume $Y$ is normally distributed, so that we're talking about a standard linear regression model, and that $X$ and $Z$ are continuous. Let's say that you are primarily interested in the relationship between $X$ and $Y$, but think the specific nature of that relationship may depend on the value of $Z$, so that we will call $Z$ a moderator here. Then, the model is: $$ \hat{Y}=\hat{\beta}_0+\hat{\beta}_1X+\hat{\beta}_2Z+\hat{\beta}_3XZ $$ The test of $\hat{\beta}_3$ will assess the existence of the interaction. Note that you have to include $Z$ in the model (see here and here for discussions of that issue). Now, let's consider a situation like yours where you wonder if a categorical covariate, $Z$, with two levels moderates the relationship between $X$ and $Y$, then your model would be: $$ \hat{Y}=\hat{\beta}_0+\hat{\beta}_1X+\hat{\beta}_2Z+\hat{\beta}_3XZ $$ exactly the same! Note that the coding strategy for $Z$ is not represented--it's irrelevant, reference cell coding or any other valid coding scheme would be employed and tested the same way. Again, you examine the test of $\hat{\beta}_3$ to see if the moderation is 'significant'.

The situation is a little more complicated if your categorical covariate has more than two levels. As you probably know, categorical covariates with $k$ levels are represented by $k-1$ 'dummy' variables. Thus, for example, if $k=3$, you need two new variables. Let's assume the situation is as above, but you are wondering if the relationship is moderated by $Z$, a categorical covariate with an arbitrarily large number of levels. Then the model would be: $$ \hat{Y}=\hat{\beta}_0+\hat{\beta}_1X+\hat{\beta}_2Z_1+\hat{\beta}_3Z_2+\cdots+\hat{\beta}_kZ_{k-1}+\hat{\beta}_{k+1}XZ_1+\hat{\beta}_{k+2}XZ_2+\cdots+\hat{\beta}_{2k-1}XZ_{k-1} $$ That's an ugly formula, but it's the way it's done. The important part is this: because you now have more than one $Z$ variable, to test if the moderation is 'significant', you drop all $k-1$ interaction terms, fit the reduced model, and perform a nested model test. In a standard linear regression context like the situation we're assuming here, that can be the F change test: $$ F_{change}=\frac{\left(\frac{SSE_{reduced}-SSE_{full}}{k-1}\right)}{\left(\frac{SSE_{full}}{df_{full}}\right)} $$ where $SSE$ is the sum of squared errors from the ANOVA table, and $p$ is the number of parameters (betas) you are estimating for that model. This $F_{change}$ value is assessed against the $F$ distribution with $(k-1,df\ {\rm error}_{full})$ degrees of freedom. If you were working with the generalized linear model, you would use the likelihood ratio test instead. (NB, most software can do these tests for you; e.g., in R the anova() command can perform nested model tests.)

To understand the effect of the moderation, I think it's best to make a scatterplot of the data and superimpose several regression lines over the points, one for each level of the moderator (i.e., $k$ lines, not $k-1$). In addition, it's typically best to plot the points associated with the different levels of $Z$ with different symbols and colors. If your moderator is continuous, it's often convenient to plot lines at the mean of $Z$, 1 SD above the mean and 1 SD below, which is what I think you are referring to in the question.

  • $\begingroup$ Thanks gung for ur detailed answer.I didn't get notified of this answer.Sorry for late reply. I am actually interested to know how to conduct a post-hoc probing once βˆ3 is confirmed to be significant. Please have a look at [this tutorial][1] where how to conduct a post-hoc analysis is explained for continuous moderators (p - 45-53). I want to do it for categorical one. You mention that - "it's often convenient to plot lines at the mean of Z, 1 SD above the mean and 1 SD below". Yes, it is. But can't I do it for categorical moderators? That is my question. [1]: goo.gl/JN26d $\endgroup$
    – Rahul
    Aug 12, 2012 at 23:00
  • $\begingroup$ @Rahul, only the last sentence is about how you would adapt the procedure for continuous moderators, the bulk of the last paragraph explains how to do this for continuous moderators. $\endgroup$ Aug 13, 2012 at 1:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.