Running model diagnostic tests using gologit2

[Pip Roddis]

Hi,

I'm using the gologit2 program in Stata to run generalised ordered logit models with partial proportional odds (using autofit). The pseudo R2 that is provided by default is the McFadden, however I would prefer to have theNagelkerke pseudo R2 (so I can compare it to some other results that I already have from a proportional odds model). Does anyone know how I can change the default setting?

Additionally, I am also trying to run a likelihood ratio test and goodness-of-fit test (either Pearson or Deviance) on the model. Does anyone know how I can do this on a gologit2 autofit model? The fitstat command and lrtest command aren't working.

Many thanks in advance,
Pip

[Mike Lacy]

-gologit2- and -fitstat- are user-written commands, not part of official Stata. Without modifying their programs, there isn't a way to get the Nagelkerke R2 after -gologit2-.

Regarding the use of -lrtest-, take a look at Example 1 in the help for -gologit2- and see if this is relevant. (It shows an example of comparing models that differ in their imposition of the proportional odds assumption.)

Regarding the Nagelkerke R2: [set selfpromotion on]
I'd recommend you look at the user-written (by me) command -r2o-, which implements a different and arguably better R2 measure for ordinal response models, as shown in the article cited in the help for that command. The -r2o- command has an option that will allow it to be run after commands such as -gologit2- that are not part of official Stata.
Comparison of proportional and nonproportional odds models is one of the useful purposes of the -r2o- measure.
[set selfpromotion off]

[Richard Williams]

Do you have the most current versions of gologit2 (from SSC, not Stata Journal) and spost13_ado (which has the latest version of fitstat) ? Both work fine for me.

. webuse nhanes2f, clear

. gologit2 health weight height

Generalized Ordered Logit Estimates             Number of obs     =     10,335
                                                LR chi2(8)        =     447.41
                                                Prob > chi2       =     0.0000
Log likelihood = -15540.691                     Pseudo R2         =     0.0142

------------------------------------------------------------------------------
      health |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
poor         |
      weight |  -.0125667   .0026198    -4.80   0.000    -.0177014    -.007432
      height |   .0252413   .0043858     5.76   0.000     .0166454    .0338372
       _cons |  -.7262079   .6763911    -1.07   0.283     -2.05191    .5994942
-------------+----------------------------------------------------------------
fair         |
      weight |  -.0172167   .0016788   -10.26   0.000    -.0205071   -.0139264
      height |   .0396816   .0027561    14.40   0.000     .0342797    .0450835
       _cons |  -4.187096   .4181421   -10.01   0.000    -5.006639   -3.367552
-------------+----------------------------------------------------------------
average      |
      weight |   -.018172   .0015315   -11.87   0.000    -.0211736   -.0151704
      height |   .0442024   .0024304    18.19   0.000     .0394388     .048966
       _cons |  -6.173409   .3648046   -16.92   0.000    -6.888413   -5.458405
-------------+----------------------------------------------------------------
good         |
      weight |  -.0203431   .0019184   -10.60   0.000    -.0241031    -.016583
      height |   .0468921   .0029476    15.91   0.000      .041115    .0526692
       _cons |  -7.633728   .4391588   -17.38   0.000    -8.494463   -6.772992
------------------------------------------------------------------------------

. fitstat

                         |    gologit2 
-------------------------+-------------
Log-likelihood           |             
                   Model |  -15540.691 
          Intercept-only |  -15764.397 
-------------------------+-------------
Chi-square               |             
      Deviance(df=10323) |   31081.382 
                LR(df=8) |     447.413 
                 p-value |       0.000 
-------------------------+-------------
R2                       |             
                McFadden |       0.014 
      McFadden(adjusted) |       0.013 
            Cox-Snell/ML |       0.042 
  Cragg-Uhler/Nagelkerke |       0.044 
                   Count |       0.306 
         Count(adjusted) |       0.030 
-------------------------+-------------
IC                       |             
                     AIC |   31105.382 
        AIC divided by N |       3.010 
              BIC(df=12) |   31192.301

[Pip Roddis]

Dear Richard and Mike,

Many thanks to you both for your help. Installing the most up-to-date versions of gologit2 and spost13_ado seems to have done the trick, and fitstat is now working perfectly.

Thank you again for your help with this matter!

All the best,
Pip

[Aysha Karim]

Do you have the most current versions of gologit2 (from SSC, not Stata Journal) and spost13_ado (which has the latest version of fitstat) ? Both work fine for me.

. webuse nhanes2f, clear

. gologit2 health weight height

Generalized Ordered Logit Estimates Number of obs = 10,335
LR chi2(8) = 447.41
Prob > chi2 = 0.0000
Log likelihood = -15540.691 Pseudo R2 = 0.0142

------------------------------------------------------------------------------
health | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
poor |
weight | -.0125667 .0026198 -4.80 0.000 -.0177014 -.007432
height | .0252413 .0043858 5.76 0.000 .0166454 .0338372
_cons | -.7262079 .6763911 -1.07 0.283 -2.05191 .5994942
-------------+----------------------------------------------------------------
fair |
weight | -.0172167 .0016788 -10.26 0.000 -.0205071 -.0139264
height | .0396816 .0027561 14.40 0.000 .0342797 .0450835
_cons | -4.187096 .4181421 -10.01 0.000 -5.006639 -3.367552
-------------+----------------------------------------------------------------
average |
weight | -.018172 .0015315 -11.87 0.000 -.0211736 -.0151704
height | .0442024 .0024304 18.19 0.000 .0394388 .048966
_cons | -6.173409 .3648046 -16.92 0.000 -6.888413 -5.458405
-------------+----------------------------------------------------------------
good |
weight | -.0203431 .0019184 -10.60 0.000 -.0241031 -.016583
height | .0468921 .0029476 15.91 0.000 .041115 .0526692
_cons | -7.633728 .4391588 -17.38 0.000 -8.494463 -6.772992
------------------------------------------------------------------------------

. fitstat

| gologit2
-------------------------+-------------
Log-likelihood |
Model | -15540.691
Intercept-only | -15764.397
-------------------------+-------------
Chi-square |
Deviance(df=10323) | 31081.382
LR(df=8) | 447.413
p-value | 0.000
-------------------------+-------------
R2 |
McFadden | 0.014
McFadden(adjusted) | 0.013
Cox-Snell/ML | 0.042
Cragg-Uhler/Nagelkerke | 0.044
Count | 0.306
Count(adjusted) | 0.030
-------------------------+-------------
IC |
AIC | 31105.382
AIC divided by N | 3.010
BIC(df=12) | 31192.301

can this fitstat command be used when we use svy or gsvy while fitting gologit model?

[Richard Williams]

The easiest way to answer a question like this is just to try it and see if it works. But no, as far as I can tell, fitstat does not work after any command when svy: has been used. I think most or all of the statistics it produces are inappropriate and cannot be calculated when svy: has been used. I talk about things you can and cannot do with svy: at

https://www3.nd.edu/~rwilliam/xsoc73...yCautionsX.pdf