My question involves Stata's capability to estimate design-based truncation
regression models. I am using complex survey data, incorporating weights,
strata, and psu. I also have a truncated dependent variable in which I have
no observations for cases when y is below a certain cutoff and, therefore,
want to estimate a truncated regression model. While Stata's survey
estimation capabilities include models for other limited dependent
variables (e.g., svyintreg), it does not explicitly implement this
capability for truncated regression. Thus, as I see it, I have several
options, which I would appreciate any comment on.
1) Use -truncreg- options to estimate truncated regression models that
partially adjust for design elements. Specifically, that would mean
ignoring strata (or disaggregating analyses by strata) while accounting for
just the pweights and clusters. This is unsatisfactory because (i) possible
problems in variance estimates for analyses that are not fully
disaggregated by strata and (ii) problems with having incorrect df in any
postestimation tests.
2) Estimate totally model-based -truncreg- and then use -suest- to survey
adjust the estimates. Unfortunately, this won't work unless -truncreg- is
somehow reprogrammed to accept iweights, which are needed for use with
-suest-. Even if this capability were implemented, I am not sure
postestimation tests would be performed with correct df.
3) This is probably the best option, but I'm not sure I have the
statistical/programming ability to implement it. It is to use Stata's -ml-
capabilities to create an estimation command that accounts for both the
truncation and survey design. I know -ml- has survey estimation
capabilities, but am not sure if truncated regression estimation has any
inherent or theoretical limitations for such implementation.
In any event, I would greatly appreciate any technical, methodological,
programming, and/or theoretical comments anyone could provide on this issue.
