Whether we should control for country-level control variables for studying a single country?

[Phuc Nguyen]

In an international setting, when I used the Difference-in-Differences (DiD) approach to examine the impact of laws on firms (the laws were implemented by many countries, I need to control for country-level control variables. However, I am wondering whether I should control for country-level control variables when studying the impact of laws on a firm's productivity in a single country (e.g., panel data for 100 firms with the time period from 2000 to 2020 - just for example) (using (DiD).

For example, I want to study the impact of anti-corruption on firms' productivity in the retailing industry compared to firms in other industries in Zimbabwe. In this case, whether I need to control GDP, Real interest rate change, corporate tax, Unemployment, import%GDP?

[Fei Wang]

Phuc, you may control for year dummies instead of including these country-level indicators. After that, you may still control for regional variables, such as macroeconomic variables in each region of Zimbabwe. But I think a more critical issue is that the DiD framework seems not suitable for your study -- e.g., are there firms not exposed to the anti-corruption law (difficult to find a control group)?

[Phuc Nguyen]

Phuc, you may control for year dummies instead of including these country-level indicators. After that, you may still control for regional variables, such as macroeconomic variables in each region of Zimbabwe. But I think a more critical issue is that the DiD framework seems not suitable for your study -- e.g., are there firms not exposed to the anti-corruption law (difficult to find a control group)?

Dear Fei Wang

Thank you for your kind help, I understand your concern. It is just an example and I acknowledge your concern. For sure that anticorruption will affect all industry. I just want to propose a context similar to what I am studying to ask your advice about whether I need to control for country-level variables.

It seems to me that you suggest that I need to control for year fixed effect. And yes, I will control for both year and firm fixed effect and clustering by firms as well. I am wondering if I understand your point correctly.

Best regards.

[Maarten Buis]

I just want to propose a context similar to what I am studying to ask your advice about whether I need to control for country-level variables.

This is a common problem people have when asking questions on technical sites like these. If you have a question, just ask that question. Don't try to simplify it. That just leads to confusion.

It seems to me that you suggest that I need to control for year fixed effect.

In the oversimplified version of your question, that makes sense: The only source of variation in the macro variables you mentioned is time. So if you control for time you have absorbed the variance in those variables as well. However, if you have a real DID design, then you have groups that are supposed to be similar. The nasty thing about empirical research is that it never corresponds to our theory however hard we try, and there usually are reasons why one group is subject to a law and another group is not and those reasons typically are correlated with other stuff we don't want it to be correlated with, and that is why we need control variables. These may well be regional GDP or regional unemployment rates, or other macro level variables not at the country level but at the level of your group. Since we still don't know your design we have to be more general than is useful. Basically, the true answer is: it depends on your exact situation.

[Jared Greathouse]

Yeah I pretty much agree with these comments. I don't really care what the model looks like (to start with), I'm concerned about the design, since if your design is poor, well, who cares what you adjust for? So my question for you, is how are you constructing your counterfactual? That is, which unexposed units are you using to compare your treated units to?

[Phuc Nguyen]

Yeah I pretty much agree with these comments. I don't really care what the model looks like (to start with), I'm concerned about the design, since if your design is poor, well, who cares what you adjust for? So my question for you, is how are you constructing your counterfactual? That is, which unexposed units are you using to compare your treated units to?

Dear Jared Greathouse , a law when it is implemented in a country, there should be some industries will be exposed to this law less or more than other industries. In this case, the firms in industries being exposed more can be considered as the treatment and firms in other industries are considered as control. The "less or more" should be supported by the extant literature review.

[Phuc Nguyen]

This is a common problem people have when asking questions on technical sites like these. If you have a question, just ask that question. Don't try to simplify it. That just leads to confusion.



In the oversimplified version of your question, that makes sense: The only source of variation in the macro variables you mentioned is time. So if you control for time you have absorbed the variance in those variables as well. However, if you have a real DID design, then you have groups that are supposed to be similar. The nasty thing about empirical research is that it never corresponds to our theory however hard we try, and there usually are reasons why one group is subject to a law and another group is not and those reasons typically are correlated with other stuff we don't want it to be correlated with, and that is why we need control variables. These may well be regional GDP or regional unemployment rates, or other macro level variables not at the country level but at the level of your group. Since we still don't know your design we have to be more general than is useful. Basically, the true answer is: it depends on your exact situation.

Dear Maarten Buis , I got your points and thanks a heap. Best.