Adj. R2

[Barbara Heerkens]

I ran multiple panel data regressions with
-xtreg, re vce(cluster id)-. (in the regression also industry and year fixed effects are used)

In order to obtain the adjusted R2 I read on Statalist I had to run the OLS -reg, vce (cluster id)-.

However, in the bottom-up approach of adding variables to the regression, my adjusted R2 drops even when I add the first variable, from 0.1569 to 0.1312.
This first variable was Size for explaining Leverage, so it should not have dropped in my opinion?
I do note that the amount of observations drops from 458 to 303, could that be a reason?
Or does the adjusted r squared not really fit with the random effects model with fixed effects for industry and year?

[Carlo Lazzaro]

Barbara:
I would start off with investigating the simplest possible reason about what you're experiencing, that is that you have missing values for -size- predictor and, due to listwise deletion, your sample drops from 485 to 303 observations.
I would also check whether (as expected) the R2 changed.
As an aside, exception made for the likelihood based regression models that fail to converge, the approach of adding one variable at time and/or increrasing the number of predictors in order to find out "the best" (whatever that means) model, should not encouraged. I would rather suggest to consider the model specification that gives the fairest and truest view of the data generating process.