coxHazard {WeightCalibSurvival}

R Documentation

Fit Cox Model from Nested Case-control Samples (NCC) using weight calibration

Description

This function fits the Cox proportional hazards model from a nested case-control study using weight calibration.

Usage

coxHazard(data, ncc.subset, outcome.var, time.vars, covars, 
          anc.covars, risk.time, risk.time0=0, nCntlPerCase=1,
          inclProb.var=NULL, print=1, min.inclProb=1e-6, control=NULL) 

Arguments

data	Data frame containing all the data for analysis.
ncc.subset	Logical vector giving the subjects in the nested case-control sample. The length must be equal to the number of rows in data.
outcome.var	Binary outcome variable in data (1 = case, 0 = non-case).
time.vars	One or two time-to-event variables in data.
covars	Character vector of covariates.
anc.covars	NULL or a character vector of ancillary covariates used to impute missing values of covars for subjects not in the nested case-control sample. If NULL, then random noise will be used for the imputation (see details).
risk.time	Projection time for pure risk.
risk.time0	Initial projection time for pure risk. Only used time.vars contains two time variables. The default is 0.
nCntlPerCase	Number of controls per case. The default is 1.
inclProb.var	NULL or a variable in data giving the probability of being included in the nested case-control sample. This can be computed using the function multipleNCC::KMprob. If NULL, then it is internally computed based on outcome.var, time.vars, and nCntlPerCase. The default is NULL.
print	0 or 1 to print information. The default is 1.
min.inclProb	Positive value <= 1 giving the minimum inclusion probability for being in the nested case-control sample. Any inclusion probability less than min.inclProb will be set to min.inclProb. This cannot be 0, since weights will be formed by taking the reciprocals of the inclusion probabilities. The default is 1e-6.
control	See control.

Details

NOTE: If anc.covars = NULL, then a seed must be set in order to reproduce the results. See Step 1 below.

The algorithm is as follows:

Step 1: Obtain predictions for missing data
For any variable in covars with missing data from phase 1, fit a weighted generalized linear model among the nested case-control subjects (phase 2) with that variable as the outcome adjusting for the other covariates in covars and anc.covars. The weights in the model are the reciprocals of the inclusion probabilities (inclProb.var). If anc.covars = NULL, then runif(nrow(data)) will be used as an ancillary covariate.

Step 2: Create auxiliary statistics (influence functions)

Step 3: Calibrate the design weights

Step 4: Fit the Cox model with and without the calibrated weights.

Value

A list containing two fitted coxph models: with.calibration and without.calibration.

Author(s)

Yei Eun Shin syeeun@gmail.com

References

Shin YE, Pfeiffer RM, Graubard BI, Gail MH. Weight calibration to improve the efficiency of pure risk estimates from case-control samples nested in a cohort. Biometrics. 2019;1-11. https://doi.org/10.1111/biom.13209

See Also

addHazard, diagnosticTests

Examples

  data(sample_data, package="WeightCalibSurvival")

  # Set the input arguments
  ncc.subset       <- sample_data[, "ind.ph2"] 
  outcome.var      <- "ind.fail"
  time.vars        <- "eventime"
  covars           <- c("X1", "X2", "Z1", "Z2")
  anc.covars       <- "U"
  risk.time        <- 8
  inclProb.var     <- "incl.prob"

  coxHazard(sample_data, ncc.subset, outcome.var, time.vars, covars, 
            anc.covars, risk.time, inclProb.var=inclProb.var)

---