Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Stas Kolenikov <skolenik@gmail.com> |
To | statalist@hsphsun2.harvard.edu |
Subject | Re: st: Wald test in Random Coefficient Model |
Date | Fri, 25 Feb 2011 10:26:10 -0500 |
On Fri, Feb 25, 2011 at 10:05 AM, Joerg Luedicke <joerg.luedicke@gmail.com> wrote: >> The behavior of this test is not quite what we would expect. >> The problem is that the null hypothesis is on the boundary of >> the parameter space, i.e. the null hypothesis is that the >> country level variance equals 0, and a variance can only be >>>= 0. There is a discussion of that in (Gutierrez et al. 2001). > That's right. I guess the rule of thumb in this case would be to > divide the p-value by 2? Nope. That's the rule of thumb that works when you have only one restriction. Here you have two, on the random intercept and on the random slope, so the correct asymptotic distribution is a mixture of chi^2(0), chi^2(1) and chi^2(2). The latter one is the most conservative, but even that distribution gives you a p-value of zero, as shown in the line: > LR test vs. logistic regression: chi2(3) = 2132.55 Prob > chi2 = 0.0000 If you insist on Wald test (on the difference between tests in the asymptotic trio, see http://www.citeulike.org/user/ctacmo/article/890474), then you would need figure out what the parameterization of the variance components is (type -matrix list e(b)-) and then use -test- or -testnl- to create the test of interest. If the variance components are parameterized using logs, then you cannot achieve a value of zero, and would have to use the LR test. -- Stas Kolenikov, also found at http://stas.kolenikov.name Small print: I use this email account for mailing lists only. * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/