I am replicating findings of a paper that uses a panel of US counties. The baseline specification is a regression like this:

$$Y_{zit} = T_{zit} + X_{zt} + a_{t} + b_{i} + u_{zit}$$

Where Y is an outcome, T is the main regressor, i indexes a county, t is time and z indexes the commuting zone, which is an aggregation of counties. X is a vector of control variables that are at the commuting zone level, not at the county level. There are fixed effects of time (a) and county (b). Standard errors are clustered at the commuting zone level.

Now I am trying to run a model with the same data, but collapsing the panel to the commuting zone level. That is to say:

$$Y_{zt} = T_{zt} + X_{zt} + a_{t} + b_{z} + u_{zt}$$

Should I keep clustering at the commuting zone level, now that it is the "individual" dimension of my data? Does it makes sense or clustering just makes sense when you have a higher level (e.g. division, states) in which you would cluster? Is there a way to check whether I should or not cluster?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.