Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

From |
Stas Kolenikov <skolenik@gmail.com> |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: st: GLLAMM problem in comprob3 |

Date |
Sat, 5 May 2012 11:18:07 -0500 |

I imagine that -gllamm- tries different approximations to the likelihood (e.g., when switching from adaptive to regular quadrature), but since your data are so sparse, none of these approximations is accurate enough, so they both just break down. Your model is most likely empirically underidentified: if an individual has all zeroes, then there is no way to pull out any reasonable estimate for that person's random effect. You simply need to have more ones to get any meaningful results; no sensible method will help you with that. For plain data, there's Firth's logistic regression (-findit firthlogit-) that is similar to ridge regression in that it shrinks the parameters towards zero by imposing a likelihood penalty that is equivalent to Jeffrey's prior. I am not sure if there are any extensions of that for multilevel data. On Sat, May 5, 2012 at 6:55 AM, P.T.Dijkstra <p.t.dijkstra@rug.nl> wrote: > Dear all, > While using the GLLAMM package, I sometimes obtain the following error message: >> can't get correct log-likelihood: -31.09976 should be -31.123589 >> something went wrong in comprob3 >> > > > I am under the impression that GLLAMM cannot estimate the model correctly because I have almost no "ones" in the dataset. > There are two groups: group 1 has just 20 times a one out of 970 observations, group 2 has only 2 ones out of 1000 observations. > > > I try to estimate a group effect as follows: > gllamm success groupeffect if group <= 2, i(personID group) family(binom) link(logit) robust adapt search(25) iterate(250) > > > The output: > Running adaptive quadrature > Iteration 0: log likelihood = -37.679599 > Iteration 1: log likelihood = -33.411411 > (...) > Iteration 249: log likelihood = -31.095211 > Iteration 250: log likelihood = -31.195684 > > Adaptive quadrature has converged, running Newton-Raphson > Iteration 0: log likelihood = -31.195684 (not concave) > Iteration 1: log likelihood = -31.135656 > (...) > Iteration 4: log likelihood = -31.123589 > can't get correct log-likelihood: -31.09976 should be -31.123589 > something went wrong in comprob3 > r(198); > > > > > Does anyone recognize this problem with the same data structure or is there anything else I can do? > I know that this kind of problem has been posted before, but any help is highly appreciated. > > > Best, > Peter Dijkstra > University of Groningen > > > > * > * 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/ -- ---- Stas Kolenikov -- http://stas.kolenikov.name ---- Senior Survey Statistician, Abt SRBI -- Opinions stated in this email are mine only, and do not reflect the position of my employer * * 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/

**References**:**st: GLLAMM problem in comprob3***From:*"P.T.Dijkstra" <p.t.dijkstra@rug.nl>

- Prev by Date:
**Re: st: Testing for differences in skewness and kurtosis?** - Next by Date:
**st: How to run a chow test for endogenous variables** - Previous by thread:
**st: GLLAMM problem in comprob3** - Next by thread:
**st: SSC Archive activity, April 2012** - Index(es):