Davron,

I thought that sample size was not a big worry upfront. I am reminded of what Granger wrote in his "Empirical Modeling in Economics" (Cambridge Press, 1999). Essentially, he said "most important macroeconomics series are rather short", and he would do a "post-sample" evaluation for time-series, a "cross-validation technique" for cross-section data, and an mixture of both for panel data. (Ibid., pp. 65-66.Perhaps I am like you, I would like to see a sample size formula like the ones in a statistics book.

Dear Lall, Thank you very much for your answer.

I want to ask one question regarding the STATA commands and methodology

I have panel data for five countries over 20 years period(20 obs per each country). As far as I undestand untill this point is that: First, I can use xtwest stata command for cointegration test for panel data and gcause2 command for granger causality test for panel data and after do some post estimations. or, Second, I can do for each country separete cointegration tests and granger causality test using standart reliable STATA commands. What would you recommend me to do?

Dear Davron,

I would not want to be quoted on this, so I am attaching two pages from Granger I cited. Granger say " it is easier to compare two model..." Take a look and see if it helps. Good Luck.

Lall

Davron,

Since you have 100 observations in total, it is okay to do the Granger Causality test. Remember that you have to try out a number of LAGS to see whether you get good results or not. For instance you may start with the default in many cases it is 2 LAGS. Then you can try 3, 4, 6, 7 LAGS for the test. I think with time series 20 observations is not terribly bad given say, 5 LAGS, at least you have 15 df.

Barnabas Kiiza

Makerere University

Davron,

To the Laal Ramrattan answer for the number of lags in Granger causality test, I woud like to add just that the number of degree of freedom is important for vouching the robustness of the granger tests and consequently, specially one has to take notice of the number of lags in the Granger test in relation to sample size.

In order to enrich this discussion I think reading the paper of Ramos and Macau (2017) worth. I rewrite a part of the paper conclusion: "[...] therefore it is reasonable to conclude that the larger the coupling is, the smaller the sample size required to infer causality."

Ramos, A., Macau, E., 2017. Minimum Sample Size for Reliable Causal Inference Using Transfer Entropy. Entropy 19, 150. https://doi.org/10.3390/e19040150

There isn't a strict minimum sample size for Granger causality tests, but a common guideline is to have at least 50 observations per variable to obtain reliable results.

Thank you Dr Chuck A Arize and Dr Angel Oswaldo Vázquez-Patiño for your valuable contributions