Further to John Sorkin's post on the cox.zph: You get test(s) of whether there is an interaction between a variable, say, sex, and time.
Suppose it is significant. You will have no clue whether the M/W hazard ratio is increasing or decreasing by time. Suppose it is not significant. You will have no clue whether the (non-significant) M/W hazrad ratio exhibits a pattern that is worth looking further into or not. In this sense the cox.zph is a perfect tool to allow you to write 'we checked for non proportionality' instead of 'we have no clue of how the M/W ratio varies by time'. If you label it what it is, namely a test of interaction, you might realize that you should ESTIMATE the shape and size of the interaction before deriving a test, either ad-hoc by the Shoenfeld residuals or by proper modeling. See for example pp 202 ff. in 'Epidemiology with R' by (surprise, surprise) me, published by OUP a few months ago. b.r. Bendix Carstensen Senior Statistician Steno Diabetes Center Copenhagen Clinical Epidemiology Niels Steensens Vej 2-4 DK-2820 Gentofte Denmark tel: +45 30 91 29 61 b...@bxc.dk bendix.carsten...@regionh.dk http://BendixCarstensen.com ________________________________ Region Hovedstaden anvender de personoplysninger, du giver os i forbindelse med din henvendelse. Du kan læse mere om formålet med anvendelsen samt dine rettigheder på vores hjemmeside: www.regionh.dk/persondatapolitik ______________________________________________ R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
