search

OHDSI / CohortMethod

An R package for performing new-user cohort studies in an observational database in the OMOP Common Data Model.
https://ohdsi.github.io/CohortMethod
82 stars 57 forks source link

Should we add 'assumption test' for Cox Hazard model? #84

Open chandryou opened 5 years ago

chandryou commented 5 years ago

Distriburted Research Network does not gather the whole statistical attributes, so a researcher cannot know if there was important violation of proportional hazard assumption, which may lead to wrong and misleading estimates ref.

Should we add 'assumption test' for Cox Hazard model, whose result a researcher can gather, too?

schuemie commented 4 years ago

Agreed. We should compute the Schoenfeld residuals to test proportionality.

msuchard commented 4 years ago

Schoenfeld $r_ij$ residuals for failure $i$ and covariate $j$ may not be terribly effective as plotting $r_ij$ as a function of covariate $j$ should be centered about 0. However, we only have two discrete values for covariate $j$ (the treatment effect), so this may be hard to see. The alternative is to plot as a function of failure time.

One will need to port the code from cox.zph() into CohortMethod or Cyclops

msuchard commented 4 years ago

Another approach is to plot log(-log(S(t))) where S(t) is the survival function estimate that Cyclops (already? I believe) provides.

S(t) is also approximated into the Kaplan-Meier plots that we have already generated and packaged up across data sources. @chandryou -- maybe this is the way you want to go in your ACS study?

msuchard commented 4 years ago

BTW, the formula for the Schoenfeld residuals is super-simple for a binary covariate and a time-invariant model with Breslow-like ties (which we are using):

https://documentation.sas.com/?cdcId=pgmsascdc&cdcVersion=9.4_3.2&docsetId=statug&docsetTarget=statug_phreg_details52.htm&locale=en