[R] Fleming-Harrington weighted log rank test

[David Winsemius]

> On Feb 13, 2018, at 4:02 PM, array chip via R-help <r-help at r-project.org> wrote:
> 
> Hi all, 
> 
> The survdiff() from survival package has an argument "rho" that implements Fleming-Harrington weighted long rank test. 
> 
> But according to several sources including "survminer" package (https://cran.r-project.org/web/packages/survminer/vignettes/Specifiying_weights_in_log-rank_comparisons.html), Fleming-Harrington weighted log-rank test should have 2 parameters "p" and "q" to control the weighting for earlier vs later times in the follow-up.
> 
> For example, setting rho=1 in survdiff() uses the Peto-Peto modification of Gehan-Wilcox weights, which I can confirm by setting p=1 & 1=0 in comp() from survminer package. similarly rho=0 is equivalent to p=0 & q=0
> 
> I am interested in putting more weights on survival difference in later follow-up time. According to comp() from survminer package, that would set p=0 & q=1 for Fleming-Harrington weights. 
> 
> My question is how I can do the same by setting certain values for "rho" in the regular survival() function?

I think that survdiff uses a different version than what you have found. The G-rho family weights are:

w_j = [Sˆ(tj)]^ρ

So rather than two parameters on S(t) and (1-S(t)) as in the p,q version, you only have one parameter applied to S(t). This class handout says that the G-rho,gamma weighting scheme is not available in survdiff.

-- 
David Winsemius
Alameda, CA, USA

'Any technology distinguishable from magic is insufficiently advanced.'   -Gehm's Corollary to Clarke's Third Law