santarabantoosoo
commented
1 year ago I would love to see that addition
Open Generalized opened 2 years ago
santarabantoosoo
commented
1 year ago I would love to see that addition
There are numerous tests for comparing survival curves. You wrote a short vignette about them: https://cran.r-project.org/web/packages/survminer/vignettes/Specifiying_weights_in_log-rank_comparisons.html
But there is a relatively new test, proposed by the ASA Biopharmaceutical Section Regulatory-Industry Statistics Workshop (links below), which gains popularity in the pharmaceutical research. Briefly, it uses the combination of FH(rho, lambda) tests with 3 combination of parameters (F(0,1), F(1,0), F(1,1) and optionally F(0,0)=logrank), able to assess at the same time: early effects, "middle" effects, late effects, diminishing effects, and crossing curves.
The maximum value of the 3 (or 4) mentioned test statistics is taken along with the covariance matrix, and the p-value is calculated based on the multivariate normal distribution. It handles all the cases and doesn't require us to pre-specify the type of change (we may truly not know this, but also want to avoid "data-driven" post-hoc selection).
In R we have 4 ways to do this:
1) the nph package, which has the nicest syntax (the least complex one) 2) the SimTrial (by Merck) project on Github - but it's not a CRAN package 3) the maxcombo package - which has the weirdest syntax (and I didn't test it for this reason yet) 4) the FHtest package, which gives us the FH(r, l) family, so we can calculate the 4, take max() of them, calculate the covariance matrix and pass it to 1-mvtnorm::pmvnorm(...).
I recommend the first one, the nph package, as it has the simplest syntax and the author responds to questions and issues immediately.
The test becomes present in clinical trials ( https://www.google.com/search?q=site%3Aclinicaltrials.gov+%2B+SAP+%2B+"max-combo" ) so it's worth adding it.
Resoures: