what kind of model is suitable for me?

[Vincent Wey]

Hello you guys!

I really am devastated about my choice to choose a quantitative analysis for my scientific work at the Uni.

What I want to do is to research the influence of Government Ideology on the economic support (Economic-suppport-Index) in times of Covid-19. First I was gonna use a regular linear OLS model when I learned that I have to consider the panel-structure of my data (yes I really am a rookie). Then I was going to use a FE model, until i read that it only consideres changes within groups (in my case countries) and not between the groups (what I actually want). Then I read about random effects models, that can only be used if some requirements are met that can be tested with the Hausman-Test - which I did and failed.

What type of model can I use?
Thanks in advance for your answers

Best regard
Vincent

PS: my dependent variable is actually scaled ordinally. Yes, i know, you should use logistic regressions with ordinal dependent variables normally. But since it can be understood quasi-metric, I thought I could use the "simple" regression model. Was I wrong? I hope, anyone can help me.

[Clyde Schechter]

There are a couple of different approaches you can take.

You could escape from the panel data constraints by reducing our data set to one observation per country covering the entire time period. If the ideology predictor variable is not constant within country, then I wouldn't recommend this overall approach, but if it doesn't change within country, or changes only minimally in a small number of observations, then this could be viable. You can reduce to one observation per country in several ways. You could pick a calendar year at random and just analyze data from that year. Or you could pick a different year at random for each country and analyze just that ensemble. You could aggregate up all the observations by taking means or medians of the variables (or other appropriate summary statistics depending on the nature of the variable).

That said, that approach discards information and is really only a method of last resort. Better is to go in the opposite direction. Supplement your panel data with more data. You are failing the Hausman test, so the remedy for that is to gather data on other relevant variables. If you can introduce sufficiently many good covariates into your model, they may be able to absorb the variance that the error terms share with the predictors so that the fuller model can pass the Hausman test. The draw back of this approach, of course, is that it is only feasible if it is possible to find a sufficiently rich set of covariates with data you can get your hands on.

[Vincent Wey]

Hello Clyde, thanks for your really helpful and detailed answer.

Right now I think I will do several approaches in my model: first I will reduce my model to only countries without a regime change and proceed like you told in the first sentence (with arithmetic mean). Then I will take several cross-section observations and do a simple regression, that should include the countries with a regime change. At last I will run the fixed effects model that reflects the changes of ideology in my model.

Maybe this is not how science works, but I think I should only demonstrate that I have a somewhat functioning understanding of regression models.

Thanks so much Clyde!!