LRTEST error

[Joe Tuckles]

Hi, trying to run an LR test but getting an error. Any way to run this with the type of models I'm using? Or is there a different method to compare models like these?

. glm cv dep if include==1, fam(poisson) link(log) nolog vce(robust) eform

Generalized linear models                         No. of obs      =      1,246
Optimization     : ML                             Residual df     =      1,244
                                                  Scale parameter =          1
Deviance         =  711.1348832                   (1/df) Deviance =   .5716518
Pearson          =  969.8694409                   (1/df) Pearson  =   .7796378

Variance function: V(u) = u                       [Poisson]
Link function    : g(u) = ln(u)                   [Log]

                                                  AIC             =   .9543619
Log pseudolikelihood = -592.5674416               BIC             =  -8155.716

------------------------------------------------------------------------------
             |               Robust
          cv |        IRR   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      dep    |   1.326143   .0272457    13.74   0.000     1.273803    1.380633
       _cons |   .1137594   .0091845   -26.92   0.000       .09711    .1332632
------------------------------------------------------------------------------
Note: _cons estimates baseline incidence rate.

. estimates store m1

. glm cv i.binary dep if include==1, fam(poisson) link(log) nolog vce(robust) eform

Generalized linear models                         No. of obs      =      1,246
Optimization     : ML                             Residual df     =      1,243
                                                  Scale parameter =          1
Deviance         =  707.9852305                   (1/df) Deviance =   .5695778
Pearson          =   969.663448                   (1/df) Pearson  =   .7800993

Variance function: V(u) = u                       [Poisson]
Link function    : g(u) = ln(u)                   [Log]

                                                  AIC             =   .9534392
Log pseudolikelihood = -590.9926153               BIC             =  -8151.738

-----------------------------------------------------------------------------------
                  |               Robust
               cv |        IRR   Std. Err.      z    P>|z|     [95% Conf. Interval]
------------------+----------------------------------------------------------------
1.binary          |   1.369277   .2226052     1.93   0.053      .995658    1.883095
           dep    |   1.320781   .0273566    13.43   0.000     1.268237    1.375502
            _cons |   .0884578   .0140422   -15.28   0.000     .0648056    .1207424
-----------------------------------------------------------------------------------
Note: _cons estimates baseline incidence rate.

. estimates store m2

. lrtest m1 m2
LR test likely invalid for models with robust vce
r(498);

[Marcos Almeida]

Yes, please try without robust SEs.

[Marcos Almeida]

Also, you can rely on AIC or BIC.

[Joe Tuckles]

Thanks for your help. I get different model results if I remove the robust SE

[Jeff Wooldridge]

Just use the z statistic on i.binary. And I’d drop eform. You’re simply testing for a zero effect of i.binary, right?

[Joe Tuckles]

I’m sorry I don’t understand. Use the z statistic instead of the lr test?

[Jeff Wooldridge]

Yes. There’s no robust version of the LR test. It doesn’t exist. So use the Wald test. With one parameter to test, this is just the asymptotic t test.