Very large t-statistics

[Janik Kaden]

Dear all,

I am performing cluster robust WLS and fixed effects models on a relatively small dataset (N=77) and I observe occasionally very large t-statistics (>100). Many variables are categorical, some are discrete between 0 and 1. Such large t-statistics appear in both. Î usually have arround 12 variables, where the categorical variables have between 2 and 5 specifications. The variables where chosen by a General-to-Specific approach.

Can those large t-values be a sign for any problem with my model? I am a little confused, because mostly I read about that issue it is said that it depends, but I don't really understand on what. I would be really grateful for some assessement.

[Jeph Herrin]

If you have 77 observations and 12 variables, yes, you have a problem: the large t-values are a symptom of over-specification. There is no hard rule, but the usual convention is at lat 10 observations per variable; but even then, you can over-specify - if you have a variable or a combination of variables which predicts the dependent variable (nearly) perfectly, you will see large or missing t-statistics.

hth,
Jeph

[Jeff Wooldridge]

It would help if you follow the guidelines and post your Stata commands and output. I suspect that you might be using the wrong weights and did not compute robust standard errors. It is easy to get the weights flipped, depending on the purpose for weighting. Why are you using WLS? What happens if you use the vce(robust) option after WLS? What if you use OLS with the vce(robust option)?