JMP > Proportional Hazard with Continuous parameter > Cutoff Optimization

[#1]

Hi JMP Community,

JMP 16.1 Standard

Windows

This might be a naive question. I am exploring the impact of a continuous baseline variable on the survival of subjects using the Proportional Hazard platform. I identified some factors with strong associations (see one example below), but I am still determining how to visualize this relationship. 

For example, I can create a survival plot if I create arbitrary sub-groups using the median as cutoff point (see below). However, this approach does not take into account the actual optimal cutoff point of the continuous variable.

Hence, I am looking for an approach to estimate this cutoff point for all identified parameters. Beside a brute-force bootstrap approach (cycling through likely cutoffs), I could not find a more elegant approach to this apparent simple problem (Yes, I read the documentation for the platform).

Any input would be greatly appreciated.

Please, feel free to let me know if my question is misguided.

Thank you.

Best,

TS

 

[#2]

Have you considered a profiler on the fitted model? I see that the Proportional Hazard platform cannot save a prediction formula, and it does not provide a profiler either. Can you use the Parametric Survival platform? If so, it provides both features.

[#3]

I support the proposal by Mark as I use this proceding very often by myself. At this occasion I would appreciate if the Parametric Survival Platform of JMP would provide risk ratios in addition as the majority of parametric survival functions implemented fulfil the proportional hazards assumption.

Winfried

[#4]

Hi JMP Team,

I decided to adopt the brute-force approach by developing a short script that iterates in 50 steps through a range of cutoffs between the 25th and 75th percentiles of the variable on interest. For each cutoff, I run the Proportional Hazard platform on the dichotomized variable of interest, recording the test performance. The optimized cutoff is then set to the value that maximizes the test performance.

Best,

TS