I made my own risk-variable and I have to analyse the effects of my risk-variable on NPL with cross-country panel data. My professor tells me I have to perform all kinds of regressions (fixed effects, pooled ols, random effects) but I have two questions.
First of all, I don't understand why you would need to do all of them? I would perform some tests and then select the best one.
Secondly, I don't understand why he wants me to put only year-dummies into the fixed effects regression model, why no country-dummies?.
Hello ask your professor these questions. Without knowing your research question, your variables, your methods and so on it is impossible to to answer the question.. Get a good book on panel data analysis and consult it frequently.. References are available at the z-library google instructions for it's use.. May the force be be with your professor. Best wishes, David Booth
Doing multiple analyses on the same data set and then choosing the one that produces the results you like is ethically questionable. You should explain more about what you professor asked you to do.
In practical terms, we usually stick to FE. Nevertheless, if you are in a learning process, it is common that you have to choose between FE and RE, using, for instance, some sort of of Hausman test. In this respect, I recommend a regression-based procedure suggested by Wooldridge in his advanced book, as it is robust to heteroskedasticity (in contrast to the original Hausman test).
Secondly, you do not need to introduce, in principle, country dummies, because this would be a way of implementing fixed effects. FE can be implemented through dummies, demeaning, first differences, etc. Nevertheless, the software will do this for you. You do not need to do it manually.
I hope it helps.
These are my results, here you can see which variables I use, the research question is: what effect does the environmental risk score have on the NPL and z-score
I agree that I gave to little information, I hope this helps
Has your professor said which to use and why? As José-Ignacio Antón says, if doing multiple procedures is intended as a learning activity for you that is fine, if it is fishing that is different.
I have a different take on this in relation to the first part of your question:
All the models are special cases of the within-between random effects model:
Pooled OLS is when the between variance is constrained to be zero;
Fixed effects is when the variance is effectively infinite;
Random effects is when the the between variance is not constrained but estimated.
This case is made in: Article Fixed and Random effects models: making an informed choice
and also see https://statmodeling.stat.columbia.edu/2010/12/17/so-called_fixed/
In the random effects model you can have both between year and between county variances.
Florence Peers
The regression-based test (à la Hausman) proposed by Wooldridge. He describes it in "Econometrics of cross-section and panel data" and the Stata code can be found in Cameron and Trivedi "Microeconometrics using Stata".
https://stats.stackexchange.com/questions/139466/procedure-for-the-cluster-robust-hausman-test
I have myself applied it here:
https://www.cambridge.org/core/journals/journal-of-pension-economics-and-finance/article/supplementary-private-pensions-and-saving-evidence-from-spain/91248880AAF307F27CDBE448EDA830A8
Then, although you will probably be using FE, you can just use clustered standard errors at the individual/country/whatever-unit-you-consider level, and you will be dealing at the same time with serial correlation and heteroskedasticity.