Missing Model F statistic

[andrew rich]

Hello all,


I am performing linear regression on panel data using a LSDV estimator. The regression command takes the form of something like: reg y x1-x12 i.firm i.year i.country, robust ... Everything is progressing as hoped, but I wanted to run an additional robustness check on the data and I include country by year dummies (i.e., i.country##i.year). The results come back as hoped in terms of direction of certain coefficients and their corresponding t values, but the overall model F stat is blank. From reading, I understand this is likely because there are so many predictors in the model. My question is ... can I safely interpret this auxiliary model as supporting my other primary results and correspondingly say in my paper that the results of this additional test further supports my hypotheses? or... is the model invalid since the model F statistic is blank?


Thanks!

[Carlo Lazzaro]

Andrew:
welcome to this forum.
See -help j_robustsingular-
I think you can reasonably interpret the results of your second model as informative: the role of F-test is to inform about the joint statistical significance of the coefficients included in the right-hand side of your regression equation. If the F-test fails statistical significance, your OLS is, basically, no more informative than the mean of the regressand.
Conversely, whether ort not your second model supports the robustness of the first one, I cannot say.

[andrew rich]

Andrew:
welcome to this forum.
See -help j_robustsingular-
I think you can reasonably interpret the results of your second model as informative: the role of F-test is to inform about the joint statistical significance of the coefficients included in the right-hand side of your regression equation. If the F-test fails statistical significance, your OLS is, basically, no more informative than the mean of the regressand.
Conversely, whether ort not your second model supports the robustness of the first one, I cannot say.

Carlo, thank you for your reply. I think I am a bit caught up in your language. Are you saying that if I run: reg y x1-x12 i.firm i.year##i.country, robust and get a missing F test, then the results could still be informative as a second model (or robustness check), but I probably wouldnt want to rely on a model with a missing F test by itself?

More directly, if the coefficients in this second model support my primary findings found in the first model, can I say in my paper that I found additional support / robustness when adding the country by year dummies or would that be incorrect given that the second model had a missing F test?


Thanks again.

[Carlo Lazzaro]

Andrew:
sorry for being unclear.
As far as my previous statement is concerned, the missing F-test is not of relevant concern.
Conersely, I would shy away from phrasing such as "the second model confirms the first one" as they are, in fact, two different regression models.
As an aside, whenever we talk about robustness, we should first answer to the following question: robustness with respect to what? Misspecification, heteroskedasticity, else....

[andrew rich]

Andrew:
sorry for being unclear.
As far as my previous statement is concerned, the missing F-test is not of relevant concern.
Conersely, I would shy away from phrasing such as "the second model confirms the first one" as they are, in fact, two different regression models.
As an aside, whenever we talk about robustness, we should first answer to the following question: robustness with respect to what? Misspecification, heteroskedasticity, else....

Thank you, Carlo. It seems that you are nudging me to be more clear in my wording, which I understand and appreciate. You are right and such nuanced language is often overlooked even in top academic journals. I will temper my language according as to say something like: the effects of interest are robust to the inclusion of country by year dummies (under the shared understand that this may further help mitigate concern over omitted variable bias).

[Carlo Lazzaro]

Andrew:
the main point is, as usual, to give the fairest and truest view of the data generating process.
If the interaction that you included in your second case is part of the data generating process (see for instance what others did in the past in your research field when presented with the same research topic) your second model should replace your first one.
Omitted variable bias (even though this fortunate and everlasting label does not main what -estat ovtest- actually does) is, if detected, a serious problem, in that it can be a proxy for endogeneity: if that were (unfortunately) the case, all the coefficients would be biased.