So I have this project to work on with panel data. I am using a dataset featuring most European countries and some of my variables are Macroeconomic such as GDP, GDP per Capita. Thing is, for some cross-sections (countries), I have missing values. In one or two of them, there is like only one or two variables missing, whereas in 3 of them (eg. Latvia's GDP), I have like 5 missing values on the lower end. T goes up to 2018. If I fill the values, I should stand at about (N=19, T=29).

My first thought is to go forecast backwards with ARIMA and Smoothing in those with few observations missing (keep the one that doesn't give me silly negative GDP values) and maybe (linearly) extrapolate for those with more than two missing values?

Couple more questions - Bonus you could say; 1) Some countries, having been part of the USSR because independent in 1991. Should I start my dataset from 1991 simply to include them? Other choice is to include a year more. Still, some Macro varibles like CPI are actually available for those countries pre-1991. 2) How to check/test for heterogeneity?

Few words about the project: It's about causality between energy sources and macro variables. The plan is to go about it using Pesaran's CD for Cross-Dependency, run Unit Root tests (I'll include CIPS in there). See now, to run CIPS on Stata, the panel must be balanced. After that, check for Cointegration. Then, if I do get a long-run relationship (cointegration found), I'll go for FMOLS/DOLS and finally run VECM to check for causality.

Note: The data mentioned is taken from World Bank

  • $\begingroup$ This question needs more details and clarity in order to make it answerable. What is the research about, what method are you using etc. Some partial answers given the state of question now would be: 1. Generally speaking extrapolation of missing data (as opposed to intrapolation) is frown upon outside forecasting. Most panel models do not require balanced panel to begin with, perhaps in cases where you need balanced panel you could convince some readers that it was justified, but I doubt it. 2. Just willy-nilly choosing how to extrapolate missing data is wrong approach. If ARIMA gives bogus $\endgroup$ – 1muflon1 Dec 31 '20 at 11:20
  • $\begingroup$ results in all cases except for 1 why would you expect that in that one case the results are actually appropriate just because they are not obviously nonsensical? Choosing methods based on whether results 'look nice' instead of methods that are case appropriate is literary the antithesis to science. 3. If countries were part of USSR before then you will likely have some structural break there - those can bias results and are hard to control for especially at the start of series. 4. How to check/test heterogeneity for what? Variance of errors? This depends on what model are you using $\endgroup$ – 1muflon1 Dec 31 '20 at 11:25
  • $\begingroup$ @1muflon1 Thanks. I did add some more information about what I'm doing. I hope it helps $\endgroup$ – NeR0 Dec 31 '20 at 11:30
  • $\begingroup$ Maybe this book can help you: amazon.com/dp/1606236393/ref=rdr_ext_tmb with the companion website www.appliedmissingdata.com $\endgroup$ – Jesper Hybel Dec 31 '20 at 12:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.