So I am running a pooled OLS. My dependent variable is house prices, and my main explanatory variables are job creation and distance to job creation, plus a raft of control variables.

The housing data is a cross section, and the firm data is panel.

The firm level data is job creation, sector, and country of origin. I calculate distance to job creation using Arc.

How do I decide whether it is better to cluster at the house level, or the firm level?

  • $\begingroup$ Which variables are "firm data"? I would guess you mean "job creation" and "distance to job creation". Please edit to clarify. It might also help if you were more explicit about what the two models you're proposing look like (I'm particularly curious what you'd be doing with the response if you did this at the firm level). $\endgroup$ – Glen_b -Reinstate Monica Oct 8 '15 at 22:16

It's hard to answer your question precisely since it is not at all clear what you are doing. However, there are some general principles courtesy of Cameron and Miller's JHR paper.

There's no formal test that will tell you at which level to cluster. If you think that the regressors or the errors are likely to be uncorrelated within a potential group, then there is no need to cluster within that group. Larger and fewer clusters have less bias, but they have more variability, so there's a kind of a trade-off there. To be conservative and avoid bias, use bigger and more aggregate clusters when possible, up to and including the point at which there is concern about having too few clusters. Unfortunately, there's no clear definition of "too few". In practice, people will often cluster at progressively higher levels and stop clustering when there is relatively little change in the standard errors.


Depending on the structure of your dataset, it might even be possible to cluster in two dimensions, i.e. house and firm level. It depends on whether the house and firm level are nested or not. If they are, ignore what I say and go to the very good answer of Dimitriy.

Petersen (2008) gives the theoretical justification for clustering on both time and firm level. His programming advice can be found here.

  • $\begingroup$ Thanks for the link- much appreciated! I'll look into it $\endgroup$ – Kelly Oct 14 '15 at 12:15

One common way to compare models is to use the sum squared errors (or sum squared distances).

Another option is using set entropy. You build yourself an entropy function and determine which split is better at describing your data.

  • 1
    $\begingroup$ I'm not sure this really deals with the particular issues that would be raised by this study. $\endgroup$ – Glen_b -Reinstate Monica Oct 8 '15 at 22:12
  • $\begingroup$ Thanks! I will look into, it might not deal with the issue, but it has been helpful for me to know this suggestion. Atleast I can look into this method, and see what it doesn't work etc, that's important too. $\endgroup$ – Kelly Oct 14 '15 at 12:16

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.