Dependent variable higher level than independent variable

[Hanna Lindstrom]

Dear Statalist experts,

I am estimating a panel-data model in which my dependent variable is on county level (organic farmland) and my main explanatory variables are on municipality level (organic food procurement). I understand that Stata's mixed command are only applicable when the dependent variable is on the lowest level. Is there a technique to estimate my model which does not entail aggregating municipality observations into a county mean? I read Griffin (1997) Interaction between individuals and situations: Using HLM procedures to estimate reciprocal relationships. However, I am still struggling with this.

Best regards,
Hanna Lindström

[Phil Bromiley]

You'll increase your chances of a helpful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.
If your dv is at the county level, then you would normally aggregate the lower level variables to the county level.

If you don't then you're treating each municipality as a observation. It seems to me this overstates your sample size. It also forces a correlation among error terms (which is manageable) since you have multiple observations with the same dv which inherently results in ties across observations in the errors.

[Hanna Lindstrom]

Dear Mr Bromiley,
Thank you for your reply. Your points are very clear and intuitive, so I would like to ask you; are there any arguments in favour of maintaining explanatory variables on municipality level? For example, can you argue that the panel with explanatory variables on municipality level contains more information and variation and so would provide more efficient estimates?

Best regards,
Hanna L

[Phil Bromiley]

It depends a bit on your theory. If you have a municipality-level theory, then test it on that level. If you have a higher level theory, test it on that level. The results will often differ.

I have a person bias toward using all the data and lower-level models, but that is not a general prescription.