R: Inference for IDI, continuous NRI, and median improvement

IDI.INF {survIDINRI}

R Documentation

Inference for IDI, continuous NRI, and median improvement

Description

This function performs inference for IDI, continuous NRI, and median improvement. Censoring is adjusted by the inverse probability censoring weight. Proportional hazards models are used as working models.

Usage

IDI.INF(indata, covs0, covs1, t0, npert = 300,
npert.rand = NULL, seed1 = NULL, alpha = 0.05)

Arguments

`indata`	Time-to-event data. The number of columns should be 2. The 1st column should be time-to-event, and the 2nd column is event indicator (1=event, 0=censor).
`covs0`	Covariates/predictors data for a base model (Model 0). Factor variables or character variables are not allowed. If any factor variable is involved in the set of predictors, use `model.matrix()` for dummy coding. `covs0` need to be a design matrix. Also missing value should not be included here.
`covs1`	Covariates/predictors data for a new model (Model 0). Factor variables or character variables are not allowed. If any factor variable is involved in the set of predictors, use `model.matrix()` for dummy coding. `covs1` need to be a design matrix. Also missing value should not be included here.
`t0`	A timepoint to define event=yes/no (case/control). Risk score is calculated as the event probability at `t0` for each model.
`npert`	The number of iterations for the perturbation-resampling. Default is 300.
`npert.rand`	If NULL (default), fresh random numbers will be generated in this routine. If a `(n x m)` matrix is given as `npert.rand`, those numbers are used in the pertubation instead, where `n` is the number of subjects and `m` is the number of iterations of the resampling. The random numbers should be generated from a distribution with mean 1 and variance 1 independently.
`seed1`	A seed for generating random numbers for the perturbation-resampling. Default is NULL.
`alpha`	(1-`alpha`/2) confidence interval will be calcualted. A 0.95 confidence interval will be provided as a default.

Value

`m1`	Result of IDI. Point and corresponding (1-alpha/2) confidence interval are given
`m2`	Result of continuous-NRI. Point and corresponding (1-alpha/2) confidence interval are given. Note that `m2` corresponds to the quantity defined as “1/2 NRI(>0)” in Pencina et al.(2011)
`m3`	Result of median improvement in risk score. Point and corresponding (1-alpha/2) confidence interval are given
`m1.est`	A vector with 3 elements. The 1st element is the point estimate of the IDI and the 2nd element is the average of risk score in “event” group, and the 3rd element is the average of risk score in “non-event” group. The 1st element is equal to the 2nd element minus the 3rd element.
`m2.est`	A vector with 3 elements. The 1st element is the point estimate of the continous-NRI. The 2nd element is the proportion of patients in whom the risk scores with the new model were higher than the risk scores with the old model, among “event” group. The 3rd element is the same proportion but among “non-event” group. The 1st element is equal to the 2nd element minus the 3rd element.
`m3.est`	A vector with 3 elements. The 1st element is the point estimate of the median improvement and the 2nd element is the median of risk score in “event” group, and the 3rd element is the median of risk score in “non-event” group. The 1st element is equal to the 2nd element minus the 3rd element.
`m3.est`	A vector with 3 elements. The 1st element is the point estimate of the median improvement and the 2nd element is the median of risk score in “event” group, and the 3rd element is the median of risk score in “non-event” group. The 1st element is equal to the 2nd element minus the 3rd element.
`point`	An object used in `IDI.INF.GRAPH()`

Note

m2 corresponds to the quantity defined as “1/2 NRI(>0)” in Pencina et al.(2011)

When the base model and the new model are nested, make sure that regression coefficients for the added predictors are significantly different from 0 in the new model, before using this function.

References

Pencina MJ, D'Agostino RB, Steyerberg EW. Extensions of net reclassification improvement calculations to measure usefulness of new biomarkers. Statistics in Medicine 2011. doi:10.1002/sim.5647

Uno H, Tian L, Cai T, Kohane IS, Wei LJ. A unified inference procedure for a class of measures to assess improvement in risk prediction systems with survival data, Statistics in Medicine 2012. doi:10.1002/sim.5647

Examples


  #--- sample data (pbc in survival package) ---
  D=subset(pbc, select=c("time","status","age","albumin","edema","protime","bili"))
  D$status=as.numeric(D$status==2)
  D=D[!is.na(apply(D,1,mean)),] ; dim(D)
  mydata=D[1:100,]

  t0=365*5
  indata1=mydata;
  indata0=mydata[,-7] ; n=nrow(D) ;
  covs1<-as.matrix(indata1[,c(-1,-2)])
  covs0<-as.matrix(indata0[,c(-1,-2)])

  #--- inference ---
  x<-IDI.INF(mydata[,1:2], covs0, covs1, t0, npert=200) ;

  #--- results ---
  IDI.INF.OUT(x) ;

  #--- Graphical presentaion of the estimates ---
  # IDI.INF.GRAPH(x) ;

[Package survIDINRI version 1.1-2 Index]