icenReg (version 1.3.5)

icenReg_cv: Get Cross-Validation Statistics from icenReg Regression model

Description

Computes the cross validation error icenReg regression models by using multiple imputations to get estimates of the censored values and taking the average loss across all imputations. Allows user to provide their own loss function.

Usage

icenReg_cv(fit, loss_fun = abs_inv, folds = 10, 
             numImputes = 100, useMCore = F)

Arguments

fit

model fit with ic_par or ic_sp. See details

loss_fun

Loss function. See details

folds

Number of cross validation folds

numImputes

number of imputations of missing responses

useMCore

Should multiple cores be used? See details

Details

For interval censored data, the loss function cannot be directly calculated. This is because the response variable is only observed up to an interval and so standard loss functions are not well defined (i.e. (10 - [7, 13])^2 is not a single number. Here the predicted value is 10, and the response is known to be between 7 and 13). To handle this, icenReg_cv uses multiple imputations; it first takes a posterior sample of the regression coefficients and conditional on these coefficients and the censored interval for each value, it draws a sample for each of the censored values. This process is repeated numImputes times to account for the uncertainty in the imputation. The cross validation error will be computed using the user provided loss_fun. This function must take two vectors of numeric values; the first being the predicted values and the second being the true values. The default function is abs_inv, which we believe is well suited in most survival analysis settings. It should be noted that the predicted values used in icenReg_cv are the estimated median for a given set of covariates. In many cases, this may not necessarily be the prediction that minimizes the loss function, but it is often still a reasonable estimate. The object fit must be an icenReg regression model, i.e. either a parametric model (via ic_par) or a semiparametric model (via ic_sp). We caution that the semi-parametric models appear to have heavier bias in estimating the out of sample bias, at least with the abs_inv loss function. The reason for this is that the semi-parametric models only assign probability mass to to Turnbull intervals, which in most cases will be strictly greater than 0, while most parametric models have possible probability on the real line. This biases both the imputations and the estimates of the semi-parametric model to not be close to 0, which is heavily pelanized by the abs_inv function. By both imputing no values close to 0 and predicting no values close to 0, the ic_sp model is being overly optimistic. In short, caution should be used when using the CV criteria to compare parametric and semi-parametric models. Multiple cores can be used to speed up the cross validation process. This must be done with the doParallel package.

Examples

##Not run: somewhat computationally intense
  # should not take more than a few minutes in worst case scenario
  
  #library(splines)
  #library(doParallel)
  #simData <- simIC_weib(n = 1000)
  
  # fit1 <- ic_par(cbind(l,u) ~ x1 + x2, data = simData)  
       # True model
  
  # fit2 <- ic_par(cbind(l,u) ~ x1 + x2, model = 'po', data = simData) 
       # Model with wrong link function
  
  # fit3 <- ic_par(cbind(l,u) ~ bs(x1, df = 6) + x2, data = simData)  
       # Overfit model
  
  # myCluster <- makeCluster(5)
  # registerDoParellel(myCluster)
        #Setting up to use multiple cores
  
  # icenReg_cv(fit1, useMCore = T)
  # icenReg_cv(fit2, useMCore = T)
  # icenReg_cv(fit3, useMCore = T)
  
  # stopCluster(myCluster)