Hi,
I am doing regression analysis with a dataset where the regressor is clustered. I am using coarsened exact matching (Stata's cem package) to address this problem, but I am not quite sure how credible my solution is.
So my unit of observation is a postal code area. I have 2500-3000 postal codes for which I observe the number of manufacturing jobs. I am trying to see if this is affected by electricity transfer price. I observe electricity transfer price at electricity distribution network level, and there's only one transfer price within each distribution network. Each distribution network contains numerous postal codes, so the postal codes are clustered in transfer price following the distribution network areas.
To overcome the problem of clustering, my idea is to use coarsened exact matching (cem) to find similar postal codes in terms of different controls, and then see what is the effect when similar postal codes in different clusters are compared. I am simply first running the cem algorithm without weights, and then in the regression absorbing the effect of cem strata. This approach gives me results that look reasonable and very much what I would expect to see. However, I am not experienced enough to say whether my strategy actually accounts for the problem or if it's just a coincidence that the results seem expected.
- Atte
I am doing regression analysis with a dataset where the regressor is clustered. I am using coarsened exact matching (Stata's cem package) to address this problem, but I am not quite sure how credible my solution is.
So my unit of observation is a postal code area. I have 2500-3000 postal codes for which I observe the number of manufacturing jobs. I am trying to see if this is affected by electricity transfer price. I observe electricity transfer price at electricity distribution network level, and there's only one transfer price within each distribution network. Each distribution network contains numerous postal codes, so the postal codes are clustered in transfer price following the distribution network areas.
To overcome the problem of clustering, my idea is to use coarsened exact matching (cem) to find similar postal codes in terms of different controls, and then see what is the effect when similar postal codes in different clusters are compared. I am simply first running the cem algorithm without weights, and then in the regression absorbing the effect of cem strata. This approach gives me results that look reasonable and very much what I would expect to see. However, I am not experienced enough to say whether my strategy actually accounts for the problem or if it's just a coincidence that the results seem expected.
- Atte
Comment