Kit has just posted my new package -boottest- on SSC. It offers:
* Extremely fast execution of the wild bootstrap (with null imposed), as recommended by Cameron, Gelbach, and Miller (2008) for inference from estimators with clustered standard errors and few clusters. With few clusters, it is often practical (if overkill!) to do as many as 100,000 replications.
* The "score bootstrap" of Kline and Santos (2012), which is an adaptation of the wild bootstrap to Maximum Likelihood estimators like logit, probit, cmp, and gsem.
* Easy access to the score/Lagrange Multiplier test, which I prefer to call the Rao test in honor of its living-legend creator. It appears to me that the Rao test has generally been inconvenient in Stata. A wrapper program, -scoretest-, makes the Rao test easier to access.
In all cases it can test one or more (joint) linear hypotheses.
boottest works as a post-estimation command, sort of like -test-. It works after OLS/2SLS in regress, ivreg, ivregress, ivreg2, restricted OLS in cnsreg, and after most ML-based estimation commands. (Exceptions: tobit and sem. But gsem works.)
When testing a single null-hypothesis constraint after OLS/2SLS, it can also provide bootstrap-based confidence intervals, whose boundaries are found through an iterative search.
-cgmwildboot- also does the wild bootstrap. It is slower, but has the advantage of supporting multi-way clustered standard errors.
To install, type "ssc install boottest". Comments welcome.
--David
* Extremely fast execution of the wild bootstrap (with null imposed), as recommended by Cameron, Gelbach, and Miller (2008) for inference from estimators with clustered standard errors and few clusters. With few clusters, it is often practical (if overkill!) to do as many as 100,000 replications.
* The "score bootstrap" of Kline and Santos (2012), which is an adaptation of the wild bootstrap to Maximum Likelihood estimators like logit, probit, cmp, and gsem.
* Easy access to the score/Lagrange Multiplier test, which I prefer to call the Rao test in honor of its living-legend creator. It appears to me that the Rao test has generally been inconvenient in Stata. A wrapper program, -scoretest-, makes the Rao test easier to access.
In all cases it can test one or more (joint) linear hypotheses.
boottest works as a post-estimation command, sort of like -test-. It works after OLS/2SLS in regress, ivreg, ivregress, ivreg2, restricted OLS in cnsreg, and after most ML-based estimation commands. (Exceptions: tobit and sem. But gsem works.)
When testing a single null-hypothesis constraint after OLS/2SLS, it can also provide bootstrap-based confidence intervals, whose boundaries are found through an iterative search.
-cgmwildboot- also does the wild bootstrap. It is slower, but has the advantage of supporting multi-way clustered standard errors.
To install, type "ssc install boottest". Comments welcome.
--David
Comment