Ugh! I did wonder, which is why i faaffed around trying to retrofit the Cox model. And it would be interesting to do a comparison of distributions on the two models.

That said, it struck me that the biggest single flaw in trying to generalise out from what is very weak data is...

...dis-continuity. Cox defo does not like dis-continuity and it seems perfectly resonable to question whether trans crims and trans non-crims are part of a single homogenous population or effectively a totally different population taking a very different value on some covariant.