2

I am using dynamic model with panel quarter data using Stata. And my sample contain 16 nations from 2000 to 2010. Is there an approximated number of observations in the panel data to be considered as a time persistent process?

mpiktas
  • 35,099
user31458
  • 21
  • 1
    This may simply be my own ignorance, but I am unfamiliar with the definition of "time persistent process". Googling didn't turn up anything obvious. Could you define your term please? – Glen_b Oct 14 '13 at 01:28
  • 2
    @Glen_b time persistence is simply a fancy term describing non- stationary data. – mpiktas Oct 14 '13 at 07:23
  • The number of observations doesn't change whether or not something is stationary, though the number of observations might affect your ability to detect it. – Glen_b Oct 14 '13 at 07:38

1 Answers1

1

Time series issues, such as unit roots, etc in panel data can be accounted for when there is enough time series dimension for single unit regression estimation. This means at least 30 observations. If you have less, you can only use ideas from time series regressions, such as doing regression on growth rates instead of levels, etc.

In fact J. Wooldridge in his book "Econometric Analysis of Cross Section and Panel Data" recommends to treat all the time series issues as a question of covariance matrix of the unit error term. Translated to Stata parlance, use cluster-robust standard errors for your analysis and you should be ok, with the usual caveat that there are no magical fixes in modelling, i.e. if your model is not sound, no fancy estimation method is going to help you.

mpiktas
  • 35,099
  • I disagree with the "This means at least 30 observations" prescription, particularly if panels are nested within, for example, states, cities, people, etc. The ability to discern such effects depends greatly on the size of the error attached to the integrating process. – Alexis Jun 29 '14 at 00:49
  • How much would you recommend? You can employ super-consistency, but then too the number of observations should be around 20. All the proofs associated with detecting unit roots I've seen rely on the asymptotics on $T$, unless you are ready to make a strong assumptions on the errors. But to make such assumptions you need to test them, and for that usually more data points are needed, not less. – mpiktas Jun 30 '14 at 07:52
  • I would not recommend any "rule of thumb". – Alexis Jun 30 '14 at 16:16