HRs from a loglog orm do not fall within their corresponding CIs. about rms

[harrelfe]

Sorry to not have gotten to this sooner. The confidence intervals appear to be correct but the hazard ratios are exp(-beta). I'll take a look at the code to try to debug. If you have time to look at it earlier than I get to it that would be great (code for orm and summary.rms).

[yhpua]

Dear Prof. Harrell,

Thank you so much for your reply! I have noticed that the loglog link function in orm.fit.s is given by -log(-log(x )) and not log(-log(x )) as stated in your excellent Regression Modeling Strategies book (Table 15.1, page 362). If the loglog link function is -log(-log(x )), exp(beta) will be given by logP₁/logP₂ while exp(-beta) will be given by logP₂/logP₁. If so, does exp(-beta) give the correct Hazard Ratio?

[harrelfe]

The logic in summary.orm was unnecessarily complex. Thanks for pointing this out. For the PH model (log-log link) I was negating the linear predictor before anti-logging, but was leaving the confidence limits in their original form. So the confidence limits were correct all along, and the point estimate was wrong. I'll have that logic removed for the next submission to CRAN. Changes will be committed to Github momentarily. Let me know if this change does not makes sense.

[yhpua]

Thank you for making those changes.