Coefficients of divided sample both smaller than coefficients in full sample

[Ian Kingsbury]

I'm running a simple regression and comparing results with certain exclusion criteria. For example, I want estimates for

Y=X1+B2+XB3 if Year==2010 | Year==2011

as well as Y=X1+B2+XB3 if Year==2012 | Year==2013

The coefficients that I am getting for XB3 (an interaction term) are both smaller in these restricted samples than if I just run an estimate for the full sample (ie no "if" statement), which includes (exhaustively) 2010-2013 data. Is that possible?

[Carlo Lazzaro]

Ian:
I find difficult to reply positively without looking at what you typed and what Stata gave you back (as per FAQ).
Moreover, I may want to share an example/excerpt of your data via -dataex-. Thanks.

[Ian Kingsbury]

Hi Carlo,

Thanks for your response. Allow me to hopefully clarify a bit.

For the full sample I have the following, where y is an indicator variable, Var1 is an indicator variable, var2 ranges from 0 to 33 and var3 is the interaction of var1 and var2.

reg y var1 var2 var3, robust

However, when I restrict the sample to certain years, the coefficient of var3 (my variable of interest) shrinks in both cases to below the one estimated for the full sample

[FernandoRios]

Hi Ian
First of all, if you are going to use interactions, it is preferable that they are introduced in your model using factor notation. Something like:

reg xFirstdecision  var1 var2 c.var1#c.var2, robust

Second, since you are using interactions, it is perfectly possible to find what you are finding right now. Your data is an example of it. My guess is that the distribution of var1 and var2 are different for both samples, which explains why you observe the unusual behavior of the coefficients.
That being said, perhaps you can add an interaction, instead of looking at two separate models

gen ydummy=inlist(year,2014,2015,2016)
reg xFirstdecision  i.ydummy var1 var2 c.var1#c.var2 i.ydummy#(c.var1 c.var2 c.var1#c.var2), robust

This will not Solve the problem, because it is not a problem at all, but will put all the numbers together for you to analyze.
HTH
Fernando

[Ian Kingsbury]

Interesting. Thanks!

Could you kindly help me to clarify these coefficients though? var1 is the effect of variable1 in 2014-2016, var2 is the effect of variable2 in 2014-2016, c.var1#c.var2 is the effect of the interaction in 2014-2016? Is that correct? And what are the remaining coefficients?