I am currently running a logistic regression model in order to analyze my transaction data. Unfortunately I do find contrary recomendations regarding the testing of moderators (btw, some use the term interaction effect, is this really the same?).

My approach looks like this:

1) Generate a new variable (if you can justify this by the literature or by observed confounding) which represents the product of the potential moderator and the respective independent variable

2) Include the new variable into the model - next to all the direct effects

3) If the wald test is significant, the moderating role is proved.

My question: Is this process correct? Other sources recommend to split the sample into two or more groups (with strong and weak/no influence of the moderator.

Thank you very much!


My understanding is that moderation (where the relationship between variable X and Y depends on M) is the same as interaction (here is a great resource on moderation).

I think your approach looks good. As you said, you first create a variable for the product of the potential moderator and the independent variable. Then, you include the new variable in the model (and, as you said, you commonly add the direct effects as well). Then, you test for the significance of the new variable representing the effect of the moderation / the interaction term. One last thing, my understanding is that you commonly plot the moderation / interaction term in order to interpret the effect. See here for a description and handy template.

In terms of splitting the sample into two or more groups, you could on the basis of the potential moderator, which let's say is a continuous variable with a range from 0 to 10. Comparing observations with values on the moderator greater than 5 to those less than or equal to 5, you can examine the nature and strength of the relationship between the independent and dependent variables to see whether there are differences between the groups.

  • 1
    $\begingroup$ Statistical-analysis programs typically can create the interaction term for you; that might be less prone to accidental errors. Breaking up a continuous variable is almost never a good idea. Do check the inherent linearity assumptions in the model. $\endgroup$ – EdM Aug 11 '15 at 14:46
  • $\begingroup$ Thank you very much for your reply. That really helped a lot. Reading through your answers raised two new questions in my mind: As you take the product of two variables, both could - from a statistical standpoint - represent the moderator. The theoretical argumentation defines which one of both is the moderator, right? Can I go through this process for all types of variable combinations (e.g. binary X continious, binary X binary, binary X categorical (which are nothing else than a set of binary variables), ...)? Or are there mathematical restrictions? Thank you very much. $\endgroup$ – Lukas Aug 12 '15 at 15:48
  • $\begingroup$ Yes, that's right about which of the two is moderator. Also, yes, you can create interactions between any of those types (and others) - for example, you can see whether a binary variable moderates the relationship of a continuous variable with an outcome. One thing I thought about and thought to mention, too - with continuous variables, it's common to center them (by subtracting the mean from each value) before creating the interaction. $\endgroup$ – Joshua Rosenberg Aug 12 '15 at 16:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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