[Scott Henry]

Good evening Erik-
I thought of a mean and was trying to get a partial data set to run as a test case, just to see that it would print MED as a level 2 variable. However, I wasn't successful yet.

Do you think there a fatal flaw that ml_mediation treats MED as a level 1 variable? Is the code doing any calculation statistically incorrect in the determination of DE, IE and TE that invalidates the analysis?
If not a fatal flaw, any recommendations that would help me address the limitations that might be created (e.g., residuals) or something I need to account for?
Once again, I truly appreciate you taking the time to help me out. Best regards, Scott.

[Erik Ruzek]

If you wanted to calculate a country level mean, you would do the following:

bysort COUNTRYID: egen cmn_MED = mean(MED)

Then, each country would have a single value for MED. But since you have a lot of variation in this variable, I would be hesitant to do that.

Your data structure is more complex than two levels, which is part of the reason why you are having these conceptual issues. You have yearly data on companies which are located within countries. So at the lowest level of your data (level 1), you have a combination of company-year. Each company-year is nested within a company (level 2) and companies are nested within countries (level 3). As a further wrinkle, your MED variable is the same for all companies in the same year. I assume that this MED variable is indeed measured at the country level, however because you have multiple years, it is not a true country-level variable.

Accordingly, to your question, can you run this as a two-level model (company-years within countries) treating MED as a level 1 variable? Yes, you can do this. It has these conceptual issues that muddy the water, but according the model structure you are using, MED is indeed a level 1 variable. Is there a better model? Probably, but it would be much more complex than your current setup. I think you will need to judge what is an acceptable level of complexity for your dissertation and your field. Look for other examples of analyses such as yours - what are people doing? That can be a helpful guide.

Realistically to pull this off, you would have to go back to Clyde's suggestion in post #2 of this thread and use Stata's gsem, which can deal with it. But you will have to do the calculation of the various types of direct and indirect effects using additional postestimation code, and that may be more than you want to take on.

[Scott Henry]

Once again, thank you very much Erik. Yes, MED is as you described (measured at the country level) but I understand the multiple years inhibits it from being a true-country level variable (with a single value). I have some thinking to do and your insights will help me in my committee discussion. With much appreciation, Scott