Panel models: compare coefficients (suest & xtreg alternatives)

[Dmitry Ilin]

Dear Statalist,

I have a panel data of a large number of companies (2481) for several years (unbalanced panel). I would like to estimate a multivariate model to investigate the impact of intellectual capital (IC) on company's performace, keeping two controls also in the model. Thus, three regressors. Now, I need to distinguish between capital intensive companies and non intensive companies. I introduced a dummy for this purpose. Generally, I could do it in two ways:

1) run two separate regressions and test whether the coefficient in front of IC is statistically different from each other;
2) interact the dummy with IC and other two control models and run one model and test whether interacted terms are significantly different from zero.

The second option is less preferable for me as the interacted terms cause quite a lot of multicollinearity (VIFs around 100) and therefore my inferences suffer. The first option seems okay in this respect. However, here is another pitfall. I first run pooled OLS on two subsamples, check the coefficients with "suest" command, and then estimate "xtreg, fe" model for these subsamples. As you might know, "suest" does not allow xtreg command. So the question is - how to get this comparison of coefficients in "xtreg" models' case?

Thanks for help,
Dmitry

[Aspen Chen]

Indeed -suest- won't work with -xtreg-. This is not a "smart" suggestion. But since you are using the fe option, can't you just create dummy variables for the fixed effects and run them with as OLS models?

[Carlo Lazzaro]

Dmitry,
as an aside to Aspen's sound advice, if you dedide for pooled OLS, cluster standard error is warranted.
Beside, provided that it is not my research field, I would be curious to investigate if a statistical significant interaction exists between capital intensity and intellectual capital, neglecting the interactions between intellectual capital and the two other controls.
Eventually, if your final choice were -xtreg-, did you contrast -fe- vs. -re- specification via -hausman- test?

Kind regards,
Carlo.

[Dmitry Ilin]

Indeed -suest- won't work with -xtreg-. This is not a "smart" suggestion. But since you are using the fe option, can't you just create dummy variables for the fixed effects and run them with as OLS models?

Thanks, Aspen, for the suggestion. You are right, however, my concern is that since I have 2481 companies, I need to have 2480 dummies. Including such an amount of dummies in the model will eat up half of the degrees of freedom. I also suspect potential multicollinearity issues. And finally, I am not sure if Stata could handle a regression with 2483 explanatory variables. All in all, I suppose this way will kill the model statistically.

[Dmitry Ilin]

Dmitry,
as an aside to Aspen's sound advice, if you dedide for pooled OLS, cluster standard error is warranted.
Beside, provided that it is not my research field, I would be curious to investigate if a statistical significant interaction exists between capital intensity and intellectual capital, neglecting the interactions between intellectual capital and the two other controls.
Eventually, if your final choice were -xtreg-, did you contrast -fe- vs. -re- specification via -hausman- test?

Kind regards,
Carlo.

Hello, Carlo,
Yes, clustered st. errors are already in
Hausman test is conducted and pointed towards fe.
Concerning the point about interactions - it is interesting. I was also wondering whether I should actually interact dummy for capital intensive companies with the control variables (firm size and firm leverage), leaving only one interaction with intellectual capital. On the one hand, If I would go with the first approach and estimate the same model on two separate samples, then size and leverage can be different, and they are different in fact. On the other hand, if I stay with one model just adding to it interacted terms, then this is a question of your belief, I guess - do you expect size and leverage to be different in intellectual intensive companies as compared to non intensive companies? The subtlety is that interacting dummy with control variables adds further multicollinearity..

[Carlo Lazzaro]

Dmitry:
my very gut-feeling is that intellectual intensive companies are, in general, smaller in size that non intensive ones. I do not know about the leverage.
However, please take this as my prior belief, as I know no data to support this statement.
The challenge, as I can see it, is to build a convincing model (as far as the theoretical building-blocks of your research field are concerned) avoiding collinearity.
For instance, does a single interaction in your model between capital intensity and intellectual capital make sense when contrasted against the theoretical mainstream? Does it make sense statistically? Does it cause sky-rocketing VIF values?
Kind regards,
Carlo

[Aspen Chen]

Dmitry, those are valid concerns. I tested the idea out using the nlswork dataset (webuse nlswork), which has more than 5,000 cases spread across 15 years in unbalanced patterns. Two OLS models would actually converge within 10 minutes or so. It is the task of stacking the estimates that killed the idea--probably because the dummy variables differ too much between the two models.

This is an interesting question. I'll keep digging. Meanwhile, I think Carlo has a point about the interaction effect route. That seems to be what the standard literature on this subject suggests.

[sarry]

Hi, I´m trying make a mediated moderated regresion, I have dependent variable is porcentage and the mediador variable is dummy variable. I do know if I use ols or gsem. I have a unbalanced data.

I appreciate your helps

[Stephen Jenkins]

"Sarry": please do not use this thread to post your question. Start a new one that is directly related to your own question. But before posting, please do two important things. (1) Read the Forum FAQ (hit the black bar at the top of the page) and digest the advice about how to compose questions in a manner most likely to lead to a helpful response. (2) Re-read the FAQ and note the strong recommendation to be registered using your real name (firstname lastname). It's easy to re-register: hit the Contact Us link at the bottom right of page, and make your request. Thank you.