Interaction terms

[Rick Mould]

Dear Stata experts,

I have some problems understanding the regression outputs with interactions in Stata. I have 4 variables, 1 Dummy and 3 continuous, and want to interact them in a specific way (1 two-way and 1 three-way interaction). I always thought that it is the same to multiply the variables before the regression or to simply use the interaction command (#) in Stata. However, the following commands lead to completely different results:

1) reg var_1 var_2 c.var_2#D c.var_2#D#c.var_3

2) gen int_1=var_2*D
   gen int_2=var_2*D*var_3

   reg var_1 var_2 int_1 int_2

Like I said, var_1, var_2 and var_3 are continous variables and D is a 0/1-dummy.

Does anyone know why this happens or what my mistake is?

Thank you!

[Carlo Lazzaro]

Rick:
please report not only what you typed, but also what Stata gave you back. Thanks.

[Rick Mould]

1)

2)

I hope it's okay that way? Sorry, I know using dataex would be better, but I don't know yet how exactly to use it, so it's faster that way.

[Carlo Lazzaro]

Rick:
when you create an interaction by hand the numbers in the categorical variable (which are meaningless in quantitative terms) do quantitatively contribute to the result of the multiplication, leading you astray.
That's way is wiser to rely on -fvvarlist- wonderful capabilities for creating interactions.

[Clyde Schechter]

Let me also note that in both approaches, you are including interaction terms without including the corresponding "main effects" of D and var3, nor the two-way subinteractions of your three way interaction. Now, that is OK in some rather unusual situations, but it is usually a badly-specified model whose results, even if you can get them consistently, would be uninterpretable. Are you sure that's what you want? Perhaps you really mean

regress var1 c.var2##c.var3##i.D