Skip to content
Prev 76776 / 398502 Next

convergence for proportional odds model

liu abc <liu2074 at yahoo.com> wrote:
It seems to be the starting values.  Using lrm() from the Design package gave
Obs  Max Deriv Model L.R.       d.f.          P          C        Dxy
        10      6e-06      11.43          1      7e-04      0.909      0.818
     Gamma      Tau-a         R2      Brier
     0.931        0.6      0.768      0.014

     Coef    S.E.  Wald Z P
y>=2 -10.904 5.137 -2.12  0.0338
y>=3 -14.336 6.287 -2.28  0.0226
dose   3.160 1.399  2.26  0.0239

and giving polr starting values:
Call:
polr(formula = resp ~ dose, data = dose.resp, start = c(-1, -4,
    3))

Coefficients:
    dose
3.158911

Intercepts:
     1|2      2|3
10.90172 14.33296

Residual Deviance: 10.34367
AIC: 16.34367

Even then, summary(m1) gives the same problem (as it refits).  There is
separation in the data, of course, but I presume the ordinality gives
some extra information.

David Duffy.