Hi all,
I am interested in advice about coming up with the correct regression
standard errors on data that's clustered in three cross-classified
factors.
Here's the basic set-up: I have survey data with (logged) wage (the
depvar) and human capital data for respondents.
This dataset has its own pweights based on the sampling design.
To this dataset, I have merged data on the expected values (per employee)
of certain regulatory costs experienced by employers.
These expected values are calculated for each combination of 50 U.S.
States (plus DC), 39 years, and 15 industry categories.
Within these state/year/industry cells, the expected regulatory costs are
constant.
Years are a major source of variation in regulatory costs, which were
implemented roughly half-way through this period, and then varied greatly
with the enforcement budget of the agency charged with enforcing them.
The are also good reasons (both theoretical and empirical) to believe the
respondents' characteristics and wages are more homogenous (aside from
regulatory costs) within states, industy categories, and years.
The question is what is the best way to correct the SE's for this
clustering. I can use the -cluster- option to -regress-, but what are the
appropriate clusters? I can generate unique identifiers for each
state/year/industry combination and use these.
But then some clusters should be much more similar than others (e.g., same
state and industry, differing years).
I've experimented with using multi-level models, under -xtmixed-, but I
can't get a 3-way cross-classified model to converge.
Any thoughts about how to best correct the SE's for clustering?
And/or how to best test whether anything more complex than considering
each year/state/industry cell to be a cluster is required?
Peter Brownell
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
