[R] odd results from polr vs wilcoxon test

[Thomas Lumley]

On Tue, 30 Dec 2003, Jonathan Williams wrote:

> > summary(fit1)
>
> Call:
> polr(formula = ordered(dat) ~ grp, control = c(maxiter = 10000,
>     trace = 0))
>
> Coefficients:
>         Value Std. Error    t value
> grp -15.82468   169.3329 -0.0934531
>
> Intercepts:
>       Value    Std. Error t value
> 2|3   -18.0223 169.3342    -0.1064
> 3|4   -16.0254 169.3329    -0.0946
> 4|5   -15.4192 169.3327    -0.0911
> 5|6   -13.6274 169.3314    -0.0805
> 6|7    -1.3866   0.5591    -2.4803
> 7|8    -0.2011   0.4495    -0.4474
> 8|9     0.6187   0.4688     1.3198
> 9|10    2.1965   0.7453     2.9472
> 10|12   2.9441   1.0259     2.8698
>
> Residual Deviance: 124.4085
>
> As far as I can see, there is no error message from polr. Could someone let
> me know
> what I am doing wrong?
>

At least five of your parameters have no finite MLE (-13 on a log scale is
very very very small).  This means that the Wald tests are probably
completely useless.  I would look at a likelihood ratio test or a score
test.

It might be nice for polr() [and glm() and coxph() and ...] to diagnose
this phenomenon, but it's a little tricky and hasn't been a high priority.


	-thomas