- Internal changes to the
`predict()`

methods for flexsurv models, in preparation for the upcoming flexsurv release (#317).

`multi_predict()`

is now available for all prediction types for`proportional_hazards()`

models with the`"glmnet"`

engine, so newly also for`type = "time"`

and`type = "raw"`

(#277, #282).Random forests with the

`"aorsf"`

engine can now predict survival time, i.e.,`predict(type = "time")`

is now available (#308).

- The
`survival_prob_*()`

,`survival_time_*()`

, and`hazard_*()`

helper functions now all take a parsnip`model_fit`

object as the main input, instead of an engine fit as was the case for some of them previously (#302).

`extract_fit_engine()`

now works properly for proportional hazards models fitted with the`"glmnet"`

engine (#266).`multi_predict(type = "survival")`

for`proportional_hazards(engine = "glmnet")`

models: when used with a single`penalty`

value, this value is now included in the results. It was previously omitted (#267, #282).`proportional_hazards(engine = "glmnet")`

models now don’t pretend to be able to deal with sparse matrices when they are not (#291).Fixed a bug for

`proportional_hazards(engine = "glmnet")`

where prediction didn’t work for a`workflow()`

with a formula as the preprocessor (#264).

- The helper functions
`survival_time_coxnet()`

and`survival_prob_coxnet()`

gain a`multi`

argument to allow multiple values for`penalty`

(#278, #279).

The new

`eval_time`

argument replaces the`time`

argument for the time points at which to predict survival probability and hazard. The`time`

argument has been deprecated (#244).The matrix interface for fitting,

`fit_xy()`

, now works for censored regression models (#225, #234, #247, #251).Improved error messages throughout the package (#248).

Added the new

`"aorsf"`

engine for`rand_forest()`

for accelerated oblique random survival forests with the aorsf package (@bcjaeger, #211).Added the new

`flexsurvspline`

engine for`survival_reg()`

(@mattwarkentin, #213).

Predictions of type

`"linear_pred"`

for`survival_reg(engine = "flexsurv")`

are now on the correct scale for distributions where the natural scale and the unrestricted scale of the location parameter are identical, e.g.`dist = "lnorm"`

(#229).Predictions of type

`"linear_pred"`

for`proportional_hazards(engine = "glmnet")`

via`multi_predict()`

now have the same sign as those via`predict()`

(#242).Predictions of survival probability for

`survival_reg(engine = "flexsurv")`

for a single time point are now nested correctly (#254).Predictions of survival probability for

`decision_tree(engine = "rpart")`

for a single observation now work (#256).Predictions of type

`"quantile"`

for`survival_reg(engine = "survival")`

for a single observation now work (#257).Fixed a bug for printing

`coxnet`

models, i.e.,`proportional_hazards()`

models fitted with the`"glmnet"`

engine (#249).

Predictions of survival probabilities are now calculated via

`summary.survfit()`

for`proportional_hazards()`

models with the`"survival"`

and`"glmnet"`

engines,`bag_tree()`

models with the`"rpart"`

engine,`decision_tree()`

models with the`"partykit"`

engines, as well as`rand_forest()`

models with the`"partykit"`

engine (#221, #224).Added internal

`survfit_summary_*()`

helper functions (#216).

For boosted trees with the

`"mboost"`

engine, survival probabilities can now be predicted for`time = -Inf`

. This is always 1. For`time = Inf`

this now predicts a survival probability of 0 (#215).Updated tests on model arguments and

`update()`

methods (#208).Internal re-organisation of code (#206, 209).

Added a

`NEWS.md`

file to track changes to the package.