SurvGPR: Fit a Gaussian process regression model to right-censored...

In ajmolstad/SurvGPR: Survival time prediction using Gaussian process regression

Description

A function for fitting a Gaussian process regression model to right-censored survival time data. See github.com/ajmolstad/SurvGPR for examples.

Usage

SurvGPR(time, status, Z, K, tol = 1e-7, max.iter = 100, max.iter.MM = 100, quiet = FALSE, 
        max.samples = 1e5, kern.type = c("multi-K"), initializer = 0)

Arguments

time	An n-variate or containing the failure/censoring times on the original scale – not log-transformed.
status	An n-variate binary vector of same length as time: 0 indicates censored, 1 indicates failure.
Z	An n \times q design matrix for the linear mean function. Note that the first column should contain all ones to account for the intercept. We recommend construction using model.matrix.
K	Candidate kernel matrices in the form of an array of dimension n \times n \times M. The algorithm will work best if these kernels have diagonal entries on similar scales.
tol	The convergence tolerance. Default is 1e-7.
max.iter.MM	The maximum number of iterations for the inner M-step algorithm.
max.iter	The maximum number of total EM-iterations.
kern.type	A character argument that can be set to either K+I or multi-K, indicating which algorithm to use. We highly recommend using multi-K in all cases, although if M=1, K+I can be faster, but less stable.
quiet	TRUE/FALSE – print algorithm progress?
initializer	Only used when kern.type = "multi-K": either 0,1 indicating the type of initialization used for the Monte-Carlo EM. The default is 0, sets all variance components equal and gets an initial estimate of \hat{β} using the updating equation from Algorithm 2. The code 1 which fits a variance components model to the IPW-mean-imputed dataset.
max.samples	An upper bound on s_k, the Monte-Carlo sample size for the kth iteration. Note that the final imputed values of log-survival for censored subjects will be the average of max.samples Monte-Carlo draws.

Value

beta	\hat{β}: The estimated regression coefficient vector corresponding to the columns of Z.
sigma2	\hat{σ}^2: The estimated variance components: a vector of length M+1, with the final element corresponding to the variance of ε.
Tout	The log-failure and imputed log-failure times obtained from our MCEM algorithm. These are primarily to be used in the prediction function.
Yimpute	The mean-imputed values of the training time-to-failures based on the method of Datta et al (2005).

ajmolstad/SurvGPR documentation built on Jan. 8, 2022, 2:38 p.m.