search

chjackson / flexsurv

The flexsurv R package for flexible parametric survival and multi-state modelling
http://chjackson.github.io/flexsurv/
54 stars 28 forks source link

Add "quantile" as an option for the type of summary.flexsurvreg #49

Closed leonardommarques closed 6 years ago

leonardommarques commented 6 years ago

Fix issue #48

I have included the type="quantile" and the quantiles parameter for summary.flexsurvreg.

I have also written two test functions:

1) compare the results of type="quantile", quantiles=0.5 and type="median". 2) expect an error if quantiles is not in between 0 and 1.

I was asked to reimplement the "median" option as a special case of your "quantile". I think that this reimplementation will not be backwards compatible, since type="median" returns only three columns and type="quantile" returns four. I am afraid some users are implementing something like summary(..., type="median")[1,1]. But I can do it if you think it is fot the best.

chjackson commented 6 years ago

Thanks Leonardo - looks great. Just merged it into the master branch with some minor edits.

chjackson commented 5 years ago

@leonardommarques Can I just clarify why you do

med_from_start = start_p * (1-t)

in summary.fns? This seems backwards to me - e.g. the 10% quantile of the survival time distribution is returned for quantiles = 0.9. See issue #57. I know in survival analysis we usually work in terms of survivor functions rather than CDFs, but a quantile is usually defined from P(X < q) not P(X > q).

leonardommarques commented 5 years ago

@chjackson , I agree that quantiles are usually defined from P(X < q). I implemented this function thinking about an inverse Survival function. If I understood correctly, in case we change it to match P(X < q) we might run into some inconsistencies. Check the issue #57:

In summary, I think it is good that the function keeps the current behavior, especially if the amount of downloads after this implementation is high. But I understand it is good to agree with the standard definitions and with the comunity. Plus, it is your package and I trust you know what is the best to do.

jrdnmdhl commented 5 years ago

Perhaps an optional argument to reverse consistent with the way the p-functions work?

From: Leonardo notifications@github.com Sent: Tuesday, October 23, 2018 3:10 PM To: chjackson/flexsurv-dev Cc: Subscribed Subject: Re: [chjackson/flexsurv-dev] Add "quantile" as an option for the type of summary.flexsurvreg (#49)

@chjacksonhttps://github.com/chjackson , I agree that quantiles are usually defined from P(X < q). I implemented this function thinking about an inverse Survival function. If I understood correctly, in case we change it to match P(X < q) we might run into some inconsistencies. Check the issue#57https://github.com/chjackson/flexsurv-dev/issues/57:

In summary, I think it is good that the function keeps the current behavior, especially if the amount of downloads after this implementation is high. But I understand it is good to agree with the standard definitions and with the comunity. Plus, it is your package and I trust you know what is the best to do.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/chjackson/flexsurv-dev/pull/49#issuecomment-432380597, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AKbbr_LhNnyGQBQkUvGMlk_78SxFBNfTks5un2m_gaJpZM4U9trJ.

chjackson commented 5 years ago

I think it's currently inconsistent - the function returns the mean and median of the distribution of survival times. So it's fairly clear to me that type="quantile" with quantiles=0.1 should return the 10% quantile of the distribution of survival times X, that is the value q for which P(X<q) = 0.1. Also this feature is not released on CRAN yet so it's fine to change I think.

chjackson commented 5 years ago

Changed in [8e6cc7b]