vce(cluster county) or vce(cluster state) or vce(unconditional)?

[Lars Pete]

The data is xtset county year.
The dep var is number of nurses in a COUNTY in a given year.
Treatment is STATE-LEVEL binary and staggered but interacted with a standardized continuous variable.

Should I use vce(cluster county) or vce(cluster state) or vce(unconditional)?

Thanks much.

[Andrew Musau]

Normally, with panel data, you would cluster at the panel-identifier level (i.e., county in this case). However, if the treatment is administered at the state level and you have a sufficient number of states (30+), there is a strong argument for clustering at the state level.

[Lars Pete]

Andrew, thank you very much for your response.
There are 40 states in data. 10 states are never treated.
What if the treatment indicators are also interacted with standardized continuous county-level variable (technological penetration) and there are 5 types of treatment (a state could adopt any one type of treatment)? So, the treatment group is group of counties at technology level 1 sd above mean in treated state being compared to another group of counties in untreated states with technology level at mean=0. Would there be a case for clustering at the county level in this case? Thanks.

[Carlo Lazzaro]

Lars:
what if, as per FAQ, you shared what you typed and what Stata gave you back? Thanks

[Andrew Musau]

The argument for clustering at the county level may stem from having county-level panel data. However, once information on the treatment assignment is available, it becomes difficult to justify clustering at the county level while arguing against clustering at the state level.

[Lars Pete]

Hi Andrew, I agree the treatment assignment is at state-level. When triple interaction is used (when treated * post is also interacted with county-level standardized tech variable, i.e., \sum_{i=1}^{k} treat_{ki(j)}* post_i(j) * std_tech_i), i is county, j is state, k is treatment type, it basically becomes a county level continuous treatment variable. Moreover, if there are 5 types of treatment (k=1to 5), there is further cross county variation in the triple interaction treatment. If I don't do triple interaction, that is just: i.e., treat_{ki}* post_i, then the case for clustering at the state-level becomes straightforward. However, in this case: \sum_{i=1}^{k} treat_{ki(j)}* post_i(j)* std_tech_i. ------ it doesn't look that straightforward. Would clustering at state-level appropriate in this case too, knowing that depvar is count data? Thanks.