predict.mexhaz: Predictions based on a mexhaz model

Description

Function for predicting the (excess) hazard and the corresponding (net) survival from a model fitted with the mexhaz function for a particular vector of covariates. If the survival model was fitted with an expected hazard (excess hazard model), the estimates obtained are excess hazard and net survival estimates. Corresponding variance estimates are based on the Delta Method or Monte Carlo simulation (based on the assumption of multivariate normality of the model parameter estimates). This function allows the computation of the hazard and the survival at one time point for several vectors of covariates or for one vector of covariates at several time points. When the model includes a random effect, three types of predictions can be made: (i) marginal predictions (obtained by integration over the random effect distribution), (ii) cluster-specific posterior predictions for an existing cluster, or (iii) conditional predictions for a given quantile of the random effect distribution (by default, for the median value, that is, 0).

Usage

# S3 method for mexhaz
predict(object, time.pts, data.val = data.frame(.NotUsed = NA),
marginal = FALSE, quant.rdm = 0.5, cluster = NULL, conf.int = c("delta", "simul", "none"),
level = 0.95, delta.type.h = c("log", "plain"), delta.type.s = c("log-log",
"log", "plain"), nb.sim = 10000, keep.sim = FALSE, include.gradient =
FALSE, dataset = NULL, ...)

Value

An object of class predMexhaz that can be used by the function plot.predMexhaz to produce graphics of the (excess) hazard and the (net) survival. It contains the following elements:

call: the mexhaz function call on which the model is based.
results: a data.frame consisting of: the time points at which the (excess) hazard and the (net) survival have been calculated; the values of the covariates used to estimate the (excess) hazard and the (net) survival; the name of the cluster (when cluster-specific posterior predictions are requested), or the quantile and its value (when conditional predictions are requested); the (excess) hazard values with their confidence limits; and the (net) survival values with their confidence limits.
variances: a data.frame consisting of two columns: the variance of the logarithm of the (excess) hazard and the variance of the (excess) cumulative hazard for each time points or each vector of covariates. The object variances is produced only when conf.int is set to "delta".
grad.loghaz: a data.frame consisting of the components of the gradient of the logarithm of the (excess) hazard for each time points or each vector of covariates. The number of columns corresponds to the number of model parameters. This object is produced only when conf.int is set to "delta" and include.gradient to TRUE.
grad.logcum: a data.frame consisting of the components of the gradient of the logarithm of the (excess) cumulative hazard for each time points or each vector of covariates. The number of columns corresponds to the number of model parameters. This object is produced only when conf.int is set to "delta" and include.gradient to TRUE.
vcov: a matrix corresponding to the covariance matrix used to compute the confidence intervals.
type: this item can take the value multitime (computation of the hazard and the survival at at several time points for one vector of covariates) or multiobs (computation of the hazard and the survival at at one time point for several vectors of covariates). It is used by plot.predMexhaz and points.predMexhaz.
type.me: the type of predictions produced in case of a model including a random intercept. Can take the values marginal (marginal predictions), posterior (cluster-specific posterior predictions), or conditional (conditional prediction for a given quantile of the distribution of the random intercept).
ci.method: the method used to compute confidence limits.
level: level of confidence used to compute confidence limits.
delta.type: type of confidence limits for the hazard and the survival when using the Delta Method.
nb.sim: number of simulations used to estimate the confidence limits when ci.method is set to "simul".
sim.haz: matrix containing the simulated hazards (each column representing a simulated vector of values); only returned when keep.sim is set to TRUE.
sim.surv: matrix containing the simulated survival probabilities (each column representing a simulated vector of values); only returned when keep.sim is set to TRUE.

Arguments

object: an object of class mexhaz, corresponding to a hazard-based regression model fitted with the mexhaz function.
time.pts: a vector of numerical values representing the time points at which predictions are requested. Time values greater than the maximum follow-up time on which the model estimation was based are discarded.
data.val: a data.frame containing the values of the covariates at which predictions should be calculated.
marginal: logical value controlling the type of predictions returned by the function when the model includes a random intercept. When TRUE, marginal predictions are computed. The marginal survival is obtained by integrating the predicted survival over the distribution of the random effect. The marginal hazard rate is obtained as the opposite of the marginal time derivative of the survival divided by the marginal survival. When FALSE (default value), cluster-specific posterior predictions or conditional predictions are calculated depending on the value of the cluster argument.
quant.rdm: numerical value (between 0 and 1) specifying the quantile of the random effect distribution that should be used when requesting conditional predictions. The default value is set to 0.5 (corresponding to the median, that is a value of the random effect of 0). This argument is ignored if the model is a fixed effect model, when the marginal argument is set to TRUE, or the cluster argument is not NULL.
cluster: a single value corresponding to the name of the cluster for which posterior predictions should be calculated. These predictions are obtained by integrating over the cluster-specific posterior distribution of the random effect and thus require the original dataset. The dataset can either be provided as part of the mexhaz object given as argument or by specifying the name of the dataset in the dataset argument (see below). The cluster argument is not used if the model is a fixed effect model. The default value is NULL: this corresponds to marginal predictions (if marginal is set to TRUE, the preferred option), or to conditional predictions for a given quantile (by default, the median) of the distribution of the random effect (if marginal is set to FALSE).
conf.int: method to be used to compute confidence limits. Selection can be made between the following options: "delta" for the Delta Method (default value); "simul" for Monte Carlo simulations (can be time-consuming, especially for models using B-splines for the logarithm of the baseline hazard); "none" indicates absence of confidence limits estimation.
level: a number in (0,1) specifying the level of confidence for computing the confidence intervals of the hazard and the survival. The default value is set to 0.95.
delta.type.h: type of confidence limits for the hazard when using the Delta Method. With the default value ("log"), the confidence limits are based on a Wald-type confidence interval for the logarithm of the hazard, otherwise they are based directly on a Wald-type CI for the hazard.
delta.type.s: type of confidence limits for the survival when using the Delta Method. With the default value ("log-log"), the confidence limits are based on a Wald-type confidence interval for the logarithm of the cumulative hazard; when the argument is set to "log", they are based on a Wald-type CI for the cumulative hazard; otherwise they are based directly on a Wald-type CI for the survival.
nb.sim: integer value representing the number of simulations used to estimate the confidence limits for the (excess) hazard and the (net) survival. This argument is used only if conf.int is set to "simul".
keep.sim: logical value determining if the simulated hazard and survival values should be returned (only used when conf.int is set to "simul"). These simulated values can be used by the riskfunc function to compute simulation-based confidence intervals for hazard ratios / risk ratios. The default value is set to FALSE.
include.gradient: logical value allowing the function to return the components of the gradient of the logarithm of the hazard and of the logarithm of the cumulative hazard for each prediction. This argument is used only if conf.int is set to "delta". The default value is FALSE.
dataset: original dataset used to fit the mexhaz object given as argument to the function. This argument is only necessary if cluster-specific posterior predictions are requested (and if the dataset is not already provided in the mexhaz object). The default value is set to NULL.
...: for potential additional parameters.

Author

Hadrien Charvat, Aurelien Belot

References

Charvat H, Remontet L, Bossard N, Roche L, Dejardin O, Rachet B, Launoy G, Belot A; CENSUR Working Survival Group. A multilevel excess hazard model to estimate net survival on hierarchical data allowing for non-linear and non-proportional effects of covariates. Stat Med 2016;35:3066-3084 (doi: 10.1002/sim.6881).

Skrondal A, Rabe-Hesketh S. Prediction in multilevel generalized linear models. J R Stat Soc A Stat Soc 2009;172(3):659-687 (doi: 10.1111/j.1467-985X.2009.00587.x).

Examples

Run this code


data(simdatn1)

## Fit of a fixed-effect hazard model, with the baseline hazard
## described by a linear B-spline with two knots at 1 and 5 year and
## with effects of age (agecr), deprivation index (depindex) and sex
## (IsexH)

Mod_bs1_2 <- mexhaz(formula=Surv(time=timesurv,
event=vstat)~agecr+depindex+IsexH, data=simdatn1, base="exp.bs",
degree=1, knots=c(1,5), verbose=0)

## Prediction at several time points for one vector of covariates
Pred_Modbs1_2A <- predict(Mod_bs1_2, time.pts=seq(0.1,10,by=0.1),
data.val=data.frame(agecr=0,depindex=0.5,IsexH=1))

## Prediction for several vectors of covariates at one time point
Pred_Modbs1_2B <- predict(Mod_bs1_2, time.pts=10,
data.val=data.frame(agecr=c(-0.2,-0.1,0), depindex=c(0.5,0.5,0.5),
IsexH=c(1,1,1)))

## Prediction for all individuals of the study population
## at one time point

Pred_Modbs1_2C <- predict(Mod_bs1_2, time.pts=10,
data.val=simdatn1)


# Example of cluster-specific posterior prediction (not run)

## Fit of a mixed-effect excess hazard model, with the baseline hazard
## described by a cubic B-spline with two knots at 1 and 5 year

# Mod_bs3_2mix <- mexhaz(formula=Surv(time=timesurv,
# event=vstat)~agecr+IsexH, data=simdatn1, base="exp.bs", degree=3,
# knots=c(1,5), expected="popmrate", random="clust", verbose=1000)

## Posterior predictions at several time points for an individual
## in cluster 15 with a specific vector of covariates
# Pred_Modbs3_2A <- predict(Mod_bs3_2mix,
# time.pts=seq(0.1,10,by=0.1), data.val=data.frame(agecr=0.2, IsexH=1),
# cluster=15, dataset=simdatn1)

Run the code above in your browser using DataLab