Is there some convention in sample selection for time series analysis? And what is the most important factor in this regard?
I mean, I have seen Kónya, I. (2018) use annual data from 1995–2016 to estimate a DSGE model.
I have seen Uribe, M., & Schmitt-Grohé, S. (2017) use annual data from 1980-2011 to estimate a VAR model.
I have seen Drechsel, T., & Tenreyro, S. (2018) use annual data from 1900-2015 to estimate a VAR and DSGE model.
Shousha (2016) uses a quarterly sample from 1994–2013.
What appears to be the pattern in these papers for me is the following;
- First, choose a sample period you want to analyze, maybe, based on a structural break in the data or maybe before or after some event.
- Then choose which frequency you want to analyze...annually, monthly, quarterly.
- Then use all available data you can get based on 1 and 2.
- Then check the parameters of the model if they are stable after estimation.
With all that said, there is a ton of information on the internet saying, for example, that sample size should be at least 50 or 80 or some threshold. Meanwhile, I do not see these thresholds in the literature. Instead, what I see is that researchers focus on some period they want to analyze at some desired frequency, not caring so much about the number of observations (which as I have shown above can be even 20).
By the way, let's say the parameters of the model are stationary for the sample 1980-2018. Then quantitative results would change if say the sample 1985-2018 is used, although both models are valid. I guess one should place more weight on qualitative results (like signs) than on quantitative results which are relatively more sensitive to sample size.