Issue 129

[adibender]

• additions:
  • predict.pamm returns predicted hazard/cumulative hazard/survival probability for each subjects survival time in newdata. Takes into account subject specific offset. newdata can be data.frame, is transformed to PED using same settings as for data used during fit.
  • predictSurvProb.pamm S3 extension for pamm objects so that PAMMs can be used with pec package. As for predict.pamm, newdata can be provided as untransformed data

[adibender]

Once this is merged, I will push the updates to CRAN.

[adibender]

ping @fabian-s
Btw, wenn du keine Zeit hast, brauchen wir auch nicht unbedingt eine Review. Denke es sollte schon passen. Die Hauptarbeit wird eh vom pec package gemacht

[fabian-s]

1. why not return the entire dataframe? are there any downstream uses that rely on this being a dataframe with as many rows as observational units instead of as many as "intervals spent under risk"? The final value is pretty useless I think (unless type = "surv_prob", ok....)
2. suggested change (but see above re "pred"):

if (type != "hazard") {
  newdata <- newdata %>%
    group_by(.data$id) %>%
    mutate(pred = cumsum(.data[["pred"]] * exp(.data[["offset"]])))
}
if (type == "surv_prob") {
  newdata <- newdata %>%
    mutate(pred = exp(-.data[["pred"]]))
}

newdata %>%
  pull(.data[["pred"]]) 

[adibender]

Yeah, I just wanted to provide some S3 method for prediction for consistency. So decided to return a vector of the same length as number of rows in input data (as is the case for most predict methods, e.g. lm, glm, gam). Thought about returning data frame with extra column, but this would replicate the functionality of add_* functions.
Maybe return the E(survival time|covariates)? 🤷‍♂️

[fabian-s]

ah ok, sorry, obvs you want a vector return not a data.frame....
expected survival time makes a lot of sense for type = "response".
for the others, maybe a list of length n of (step)functions or a list of dataframes cbind(tstart, tstop, estimate) .

[adibender]

The only Q is how to obtain the expected survival times 😄

1. calculate hazard (per interval) -> simulate survival times using rpexp -> use mean or median
2. Calculate survival prob over follow up -> use median survival time (could be NA if S(t|x)>.5 for all t )
3. Is it possible to calculate E(t|x) analytically, similar to parametric models?

[fabian-s]

analytically should be possible I think? the density of the pexp isn't to terrible, after all... if not, expected survival should still be ok to compute since expectation of a positive random variable is simply the integral of S(t), which we know how to compute from the fit. practical problem there is what the upper limit of the integral should be (certainly not just maximal follow-up Tmax, but if the hazard estimate in the final interval is small and S(Tmax) is still >>0, extending that last interval for a long time will pull the expectation way up....)

[fabian-s]

sorry, pack ich grade nicht das wirklich im detail anzusehen.
das was ich geshen habe ist kleinkram, sieht sonst super aus und ist sicher ne super wichtige ergänzung mit dem pec-interface... 👍

[fabian-s]

should this have a se arg as well (and se.type = c("bla", "delta", "sim"))?

[adibender]

will try to implement. Is there a technical term for "bla". I usually call it direct transformation, but maybe there's a term for that?

[fabian-s]

Is there a technical term for "bla"

..comedy gold right here 🤣
direct would seem to be the more professional sounding choice.

[adibender]

Only fixing #128 and #129 for now. predict.pamm will be added in separate PR (see #136)