Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Dalhia <ggs_da@yahoo.com> |
To | statalist@hsphsun2.harvard.edu |
Subject | Re: st: xtnbreg, nbreg, and tests of assumptions |
Date | Wed, 15 Dec 2010 09:12:34 -0800 (PST) |
oops sorry. don't know what I was thinking. Thanks Mary for the correction. Here are the results for xtnbreg that don't make sense. Basically, I have panel data on hospitals (private, public, and associates), and looking at the averages of the number of training days for each hospital type, I can see that private hospitals have lower number of training days compared to public hospitals. Associate hospitals fall in the mid-range. However, when I run this model using xtnbreg (with random effects), I get a funny result. It looks like public and associates have lower rate of training days in a year compared to private. Am I interpreting the coefficients wrong or is there something else going on? (output attached below). When I run it using nbreg I get the opposite result (the result I was expecting - public and associates are have greater rate of training per year compared to private). Thanks for your help. Dalhia . xtnbreg train asso pub if train<12000, re irr note: you are responsible for interpretation of non-count dep. variable Fitting negative binomial (constant dispersion) model: Iteration 0: log likelihood = -1341968.9 Iteration 1: log likelihood = -1341967.5 Iteration 2: log likelihood = -1341967.5 Iteration 0: log likelihood = -504693.72 Iteration 1: log likelihood = -35614.007 Iteration 2: log likelihood = -35604.55 Iteration 3: log likelihood = -35604.545 Iteration 4: log likelihood = -35604.545 Iteration 0: log likelihood = -35604.545 Iteration 1: log likelihood = -35595.175 Iteration 2: log likelihood = -35595.145 Iteration 3: log likelihood = -35595.145 Fitting full model: Iteration 0: log likelihood = -81145.913 Iteration 1: log likelihood = -49940.372 (not concave) Iteration 2: log likelihood = -42786.562 (not concave) Iteration 3: log likelihood = -35793.307 Iteration 4: log likelihood = -33256.88 Iteration 5: log likelihood = -33190.785 Iteration 6: log likelihood = -33150.666 Iteration 7: log likelihood = -33150.622 Iteration 8: log likelihood = -33150.622 Random-effects negative binomial regression Number of obs = 7522 Group variable: fi Number of groups = 1873 Random effects u_i ~ Beta Obs per group: min = 1 avg = 4.0 max = 5 Wald chi2(2) = 7.29 Log likelihood = -33150.622 Prob > chi2 = 0.0261 ------------------------------------------------------------------------------ train | IRR Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- asso | .8803461 .0551126 -2.04 0.042 .7786914 .9952712 pub | .9029852 .0380889 -2.42 0.016 .8313349 .9808108 -------------+---------------------------------------------------------------- /ln_r | -.8268984 .0334362 -.8924322 -.7613647 /ln_s | .7346747 .0714634 .5946091 .8747404 -------------+---------------------------------------------------------------- r | .4374038 .0146251 .4096582 .4670286 s | 2.084804 .1489872 1.812322 2.398253 ------------------------------------------------------------------------------ Likelihood-ratio test vs. pooled: chibar2(01) = 4889.04 Prob>=chibar2 = 0.000 . --- On Wed, 12/15/10, Mary E. Mackesy-Amiti <mmamiti@uic.edu> wrote: > From: Mary E. Mackesy-Amiti <mmamiti@uic.edu> > Subject: Re: st: xtnbreg, nbreg, and tests of assumptions > To: statalist@hsphsun2.harvard.edu > Date: Wednesday, December 15, 2010, 7:03 PM > I think you meant to say that your > *dependent* variable is count with > overdispersion, and your *independent* variable is time > invariant. > Independent variables predict dependent variables. > > Please post the -xtnbreg- command you used and the results > you find > questionable. > > > On 12/15/2010 7:30 AM, Dalhia wrote: > > Hi, > > > > I am trying to figure out whether I should use nbreg > (with > > correction for autocorrelation and heteroskedasticity) > or xtnbreg (with > > random effects)? My independent variable is count with > significant overdispersion, and I have panel data (cross > sectional time series). One of my main dependent variables > is time invariant, and therefore I cannot use xtnbreg fixed > effects. xtnbreg random effects is giving me some funny > results that are hard to believe, but how should I decide > which one I should be using (xtnbreg or nbreg)? Also, are > there tests to check if the assumptions of these models are > satisfied in my data? > > > > Finally, I have two independent variables, > predicted by the same dependent variables. But I can't find > a version of SUR appropriate for negative binomial. I am not > really interested in cross-equation testing. If I don't do a > seemingly unrelated regression, does that bias the > coefficients or does it produce inefficient > > coefficients. > > > > Thanks so much. I really appreciate your help. > > Dalhia > > > > --- On Wed, 12/15/10, Maarten buis<maartenbuis@yahoo.co.uk> > wrote: > > > > From: Maarten buis<maartenbuis@yahoo.co.uk> > > Subject: Re: st: Difference between xtlogit, > xtmelogit, gllamm > > To: statalist@hsphsun2.harvard.edu > > Date: Wednesday, December 15, 2010, 10:28 AM > > > > --- On Wed, 15/12/10, Rajaram Subramanian Potty > wrote: > >> I have event history data and this data has been > converted > >> into discrete time to fit discrete time hazard > model. > >> Now, I want to fit a multilevel model. But there > are: three > >> different proceudres such as xtlogit, xtmelogit > and gllamm. > >> I want to know which procedure is more appropriate > for > >> analysing the discrete time data. > > All three will do for a basic multilevel model for the > odds > > (not the hazard) of survival. If you want to model a > > multilevel model for the hazard of survival you can > use > > -gllamm- with the cll link function. The difference > between > > -xtlogit- and -xtmelogit- is that the latter can > accomodate > > more complex multilevel structures. > > > > Hope this helps, > > Maarten > > > > -------------------------- > > Maarten L. Buis > > Institut fuer Soziologie > > Universitaet Tuebingen > > Wilhelmstrasse 36 > > 72074 Tuebingen > > Germany > > > > http://www.maartenbuis.nl > > -------------------------- > > > > > > > > > > * > > * 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/ > > > > > > > > > > > > * > > * 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/ > > > > > > -- > Mary Ellen Mackesy-Amiti, Ph.D. > Research Assistant Professor > Community Outreach Intervention Projects (COIP) > School of Public Health m/c 923 > Division of Epidemiology and Biostatistics > University of Illinois at Chicago > ph. 312-355-4892 > fax: 312-996-1450 > > * > * 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/ > * * 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/