predict.flexsurvreg: Predictions from flexible survival models

Description

Predict outcomes from flexible survival models at the covariate values specified in newdata.

Usage

# S3 method for flexsurvreg
predict(
  object,
  newdata,
  type = "response",
  times,
  start = 0,
  conf.int = FALSE,
  conf.level = 0.95,
  se.fit = FALSE,
  p = c(0.1, 0.9),
  ...
)

Value

A tibble with same number of rows as newdata and in the same order. If multiple predictions are requested, a tibble containing a single list-column of data frames.

For the list-column of data frames - the dimensions of each data frame will be identical. Rows are added for each value of times or p requested.

This function is a wrapper around summary.flexsurvreg, designed to help flexsurv to integrate with the "tidymodels" ecosystem, in particular the censored package. summary.flexsurvreg returns the same results but in a more conventional format.

Arguments

object

Output from flexsurvreg or flexsurvspline, representing a fitted survival model object.

newdata

Data frame containing covariate values at which to produce fitted values. There must be a column for every covariate in the model formula used to fit object, and one row for every combination of covariate values at which to obtain the fitted predictions.

If newdata is omitted, then the original data used to fit the model are used, as extracted by model.frame(object). However this will currently not work if the model formula contains functions, e.g. ~ factor(x). The names of the model frame must correspond to variables in the original data.

type

Character vector for the type of predictions desired.

"response" for mean survival time (the default). "mean" is an acceptable synonym
"quantile" for quantiles of the survival distribution as specified by p
"rmst" for restricted mean survival time
"survival" for survival probabilities
"cumhaz" for cumulative hazards
"hazard" for hazards
"link" for fitted values of the location parameter, analogous to the linear predictor in generalized linear models (type = "lp" and type = "linear" are acceptable synonyms). This is on the natural scale of the parameter, not the log scale.

times

Vector of time horizons at which to compute fitted values. Only applies when type is "survival", "cumhaz", "hazard", or "rmst". Will be silently ignored for all other types.

If not specified, predictions for "survival", "cumhaz", and "hazard" will be made at each observed event time in model.frame(object).

For "rmst", when times is not specified predictions will be made at the maximum observed event time from the data used to fit object. Specifying times = Inf is valid, and will return mean survival (equal to type = "response").

start

Optional left-truncation time or times. The returned survival, hazard, or cumulative hazard will be conditioned on survival up to this time. start must be length 1 or the same length as times. Predicted times returned with type "rmst" or "quantile" will be times since time zero, not times since the start time.

conf.int

Logical. Should confidence intervals be returned? Default is FALSE.

conf.level

Width of symmetric confidence intervals, relative to 1.

se.fit

Logical. Should standard errors of fitted values be returned? Default is FALSE.

p

Vector of quantiles at which to return fitted values when type = "quantile". Default is c(0.1, 0.9).

...

Not currently used.

Author

Matthew T. Warkentin (https://github.com/mattwarkentin)

Examples

Run this code


fitg <- flexsurvreg(formula = Surv(futime, fustat) ~ age, data = ovarian, dist = "gengamma")

## Simplest prediction: mean or median, for covariates defined by original dataset
predict(fitg)
predict(fitg, type = "quantile", p = 0.5)

## Simple prediction for user-defined covariate values
predict(fitg, newdata = data.frame(age = c(40, 50, 60)))
predict(fitg, type = "quantile", p = 0.5, newdata = data.frame(age = c(40,50,60)))

## Predict multiple quantiles and unnest
require(tidyr)
pr <- predict(fitg, type = "survival", times = c(600, 800))
tidyr::unnest(pr, .pred)