Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Panel Data - Missing Overall Model Significance Statistics

    Dear all,

    I have estimated a fixed effect panel regression model for work performance (dependent variable y) in my data set for about N = 100 workers for T = 15 years.
    Since my data suffers from heteroskedasticity i used the , fe robust specification.

    All was fine and my overall modell and predictor variable are highly significant.

    Recently, i found out i need to include another explaining variable (salary):

    If i include it in the model two unpleasent things happen:

    1) Stata dont show a p- and F- statistic for the overall model (only a dot)
    Code:
                                                     F(20,93)          =          .
    corr(u_i, Xb)  = -0.9496                        Prob > F          =          .
    Suprisingly, the p- and F- values are NOT lost if i dont use the "robust" but only a simple , fe

    2) All explanatory variables become insignificant (including salary)

    I tried this exercise with salary as well as ln(salary) with the same result.


    While for 2) i could imagine that just my model has not a good explanatory power i am completely clueless what made the overall model statistics vanish in the , fe robust case.

    Anybody facing a similar problem or has an idea what the problem here is?

    Many thanks in advance,
    Alexander-Florian

  • #2
    This is a common situation and it is not a problem. In -fe robust- you are actually getting the cluster robust VCE (unless you are using a Stata version earlier than 13), clustering on the panel variable declared in your -xtset- command. For the cluster robust VCE, the number of degrees of freedom is the number of clusters. When you added the additional predictor to your model, the number of predictors was pushed higher than the number of degrees of freedom in the model, so it is not possible to calculate the overall model F statistic.

    This is, in general, not a problem. Unless the test of the omnibus null hypothesis for the model is an explicit part of your research goals, these overall model fit statistics are usually of no importance anyway. Most models consist of variables you are interested in and other variables included to adjust for their nuisance effects or reduce residual variance. The statistical significance of the overall model is then an uninteresting question: only the significance of the variables you are interested in matters. Moreover, you will note that each of your predictors has its own test, and these are not missing. You can still rely on those tests.

    In the unusual situation where the test of the omnibus null hypothesis is an explicit part of your research goals, then you have a problem, because it is mathematically impossible to do this with cluster robust fixed-effects estimation. In this case you would either have to remove some variable(s) from your model, or give up either cluster robust variance estimation or fixed effects estimation.

    By the way, it would have been more helpful had you shown the actual command you used and the complete output from that command, rather than just the small part that exhibits the phenomenon you are concerned with. In this case, it is such a common issue that I felt comfortable giving my answer without seeing enough information to be completely sure, but typically one needs to see the complete command and output as there are often clues in the rest of the information.

    Correction: the number of degrees of freedom is the number of clusters minus one. The conclusion remains the same.
    Last edited by Clyde Schechter; 29 Jun 2018, 09:28.

    Comment


    • #3
      Clyde,

      many thanks for your helpful reply and for your advise how to use the forum more effectively!

      Best A.

      Comment

      Working...
      X