Interaction variables

[Chiraz KARAMTI]

Hi everyone,
I have a quite bizarre question! I'm running a model with reform (competition, regulation and privatization) variables (continuous) and I want to analyze the reforms interaction. To do so I created an interaction variable with dummy variables (I have both continuous and dummy measures for reforms). The variable introduced into the model are continuous (for example the number of competitors to reflect competition and number of years since the establishment of an independant regulator to reflect regulation). Now I created a new inetraction variable equal to 1 if when competion and regulation exist and 0 if not. Can we say that is an interaction variable? Can we inetrpret it as one??
Unfortunately, when I created a continuous interaction variable, the model results turn bad and I loose significance of many variables.
I'm not sure if my question is clear for you:-(

Help!!

[Clyde Schechter]

Yes it is an interaction variable, but it cannot be used with the continuous variables. The actual variables that are constituents of the interaction variable must (nearly) always be included in the model along with the interaction. Otherwise the model is mis-specified.

Continuous interaction variables are always tricky. The assumption that the effects are actually multiplicative is a very strong one and is frequently wrong. I usually don't advocate turning continuous variables into categories, but the complications associated with a three way interaction are probably one of the situations where I would. So I would generate dichotomous variables for the presence or absence of computation and regulation. Call those computation2 and regulation2. I gather privatization is already a dichotomy. So then I would use i.computation2##i.regulation2##privatization in my model (without the original computation and regulation continuous variables) to express all three "main" effects, all three 2-way interactions between them, and the one three-way interaction. You will definitely want to use -margins- to understand the results!

[Chiraz KARAMTI]

Hi Clyde,
I totally agree with your opinion concerning the interaction between continuous variables. However, how can I use interactions without the original variables in the model? is it logical? I already tried including interactions 2 by 2 and than a trinity variable.
But still my question is, if the model includes the original continuous variables and I introduce interactions using binary indicators, are we then still talking about interactions?

[Clyde Schechter]

But still my question is, if the model includes the original continuous variables and I introduce interactions using binary indicators, are we then still talking about interactions?

And my answer is still no. That is not a valid interaction model.

You can't use interactions without the original variables in the model. (Exception: if one of the original variables is colinear with a fixed effect.)

What I suggest is substituting the discrete variables for the original continuous variables throughout.

[Chiraz KARAMTI]

Ok, understood. Thank you very much.