Country-level differences - fixed effects for panel data?

[Kamilla Plet-Bundgaard]

Hello,

I am currently writing a thesis on the performance of mutual in general. I have collected mutual funds from various countries in Europe, and I have a question relating to one of my sub-analyses. I am hoping to do a regression to see if there are any country-specific factors that affects my dependent variable, which is a HmL variable (the difference in performance) between highly ESG rated mutual funds and low ESG rated mutual funds.

To clarify, my dataset includes a time period of four years, equal to 48 months, and it additionally includes 11 countries, which makes my dataset a paneldata. So in total, I have 48 months of HmL observations for country 1, 48 months of HmL observations for country 2, etc. etc. Furthermore, I have country factors in my dataset such as GDP per capita, AuM, market share, etc.

I am unsure how to perform the best regression on this paneldata-set. I am not interested in any time-effects, but rather country-effects. I was contemplating using a fixed effects model, or else including dummy variables for each country and when running the whole dataset, then excluding one of the 11 dummies to be able to interpret my results.

My supervisor is not very good at guiding me through this, so I was hoping someone in here could help me with this problem.

I look forward to receiving some feedback!

Best,
Kamilla

[Carlo Lazzaro]

Kamilla:
welcome to this forum.
Some issues here:
1) you're seemingly dealing with a T (time-series dimension=48 months)>N(cross-sectional dimension=11 countries), aka long panel dataset. Therefore, the Stata commands that yiou should consider are -xtregar- and -xtgls-;
2) unfortunately, dealing with long panel datasets is much moe demanding than with their N>T counterparts;
3) if you are interested (and your data allow you to do so) in across countries comparisons, fixed effect is not what you're looking for;
4) ignoring time variable may not be a good choice;
5) If I may, I would recommend you to get yourself familiar (at least) with the building blocks of panel data regression;
6) I wouid warmly recommend you to share every single step of your analysis with your supervisor, to avoid going around when the runway is in sight!

[Kamilla Plet-Bundgaard]

Thank you for your reply Carlo. Our goal is to look into the differences between countries, but as our supervisor is not really happy about helping us, it is definitely hard for us to determine what to do about such an analysis. Do you have any recommendations for such cross country comparisons? I see that you mention xtregar or xtgls? Could you elaborate on this?

Me and my fellow study partner are really unsure about this analysis. It was our supervisor suggesting us to look at it. Would it make better sense to run separate regressions for each country? Or maybe run an OLS regression and including dummy variables for all countries (except the one to measure up against)?

[Carlo Lazzaro]

Kamilla:
this is not a good situation for you and your fellow study partner, as, if I may, you seem to have a limited smattering of panel data regression and your supervisor does not seem to be that helpful either.
If you decide to go on with this research:
1) I would recommend you to get yourself familiar (at least) with the building blocks of panel data regression (Microeconometrics Using Stata, Second Edition | Stata Press (stata-press.com) along with any decent textbook on panel data econometrics;
2) take a thorough look at -xtregar- and -xtgls- entries in Stata .pdf manual;
3) a pooled estimator is useful if you do not have evidence of a panel-wise effect;
4) forget running separate regression for each country if the aim of your research is to comnpare different countries;
5) from your description I cannot get if your panel unit is -country- or simething different (say, firms nested within countries).

[Kamilla Plet-Bundgaard]

Thank you again for the quick reply. You are very right with respect to our data and supervisor, however I do think we need to proceed with our analysis. I have read a lot about panel data regression, but I am still unsure what the optimal solution for us is. I will look into the commands you suggest, and see whether this makes any sense. If not, would you believe it to be okay to just do a normal pooled OLS on the dataset I already have?

I am glad that you confirm that it doesn't make any sense to run separate regressions.

To answer 5): I do believe my panel unit is country - we want to look at country-difference. The difference in return (my dependent variable) is also country specific, but it has been calculated based on mutual fund performance in each respective country. Does this make sense?

Kind regards,
Kamilla

[Carlo Lazzaro]

Kamilla:
1) running a pooled OLS if your data show evidence of a panel-wise effect means losing some information. Usually, we go pooled OLS with panel dataset if we do not see any evidence of a panel-wise effect.
2) stick with -country- as the -panelvar-, then;
3) no need to repeat to inform your supervisor of each and every step of your research (and don't care if she/he seems bored or annoyed about that; she/he is paid for that via your academic tuitions), so that I cannot react (or hopefully so) when you'll draft your dissertation.