3
$\begingroup$

The partial log-likelihood function in Cox proportional hazards is given with such formula $${}_{p}\ell(\beta) = \sum\limits_{i=1}^{K}X_i'\beta - \sum\limits_{i=1}^{K}\log\Big(\sum\limits_{l\in \mathscr{R}(t_i)}^{}e^{X_l'\beta}\Big),$$ where $K$ is the number of observations for which we have observed an event (where generaly there were $n$ observations, so $K-n$ observations were censored) and $\mathscr{R}(t_i)$ is a risk set for time $t_i$ defined as: $\mathscr{R}(t_i) = : \{X_j: t_j >= t_i, j = 1, \dots, n \}$ (as described here How to compute partial log-likelihood function in Cox proportional hazards model?).

Then I am wondering on how to implement and compute gradient of partial log-likelihood function in Cox proportional hazards model, which is given with formula ($k$-th coordinate for $p$-dimenational vector of partial derivatives, for $k=1,\dots,p$) $$ U_k(\beta)=\dfrac{\partial{}_{p}\ell_k(\beta)}{\partial\beta_k}=\sum\limits_{i=1}^{K}\Big(X_{ik}-A_{ik}\Big), $$ where \begin{equation} A_{ik} = \dfrac{\sum\limits_{l\in \mathscr{R}(t_i)}^{} X_{lk} e^{X_l'\beta}}{\sum\limits_{l\in \mathscr{R}(t_i)}^{} e^{X_l'\beta}} \end{equation} is a mean of $X_{.k}$ ($k$-th variables) over finite population $\mathscr{R}(t_i)$, with the use of \textit{exponentially weightened} form of sampling.

I've tried myself to implement gradient with this code:

library(foreach)
library(dplyr)
coxphGD_step <- function(formula, data, learningRate, beta){
  # collect times, status, variables and reorder samples 
  # to make the algorithm more clear to read and track
  preparedData <- prepareData(formula = formula, data = data)
  # calculate the log-likelihood for this batch sample
  partial_sum <- list()
  foreach(k = 1:nrow(preparedData)) %do% {
    # risk set for current time/observation
    risk_set <- preparedData %>% filter(times >= preparedData$times[k])

    nominator <- apply(risk_set[, -c(1,2)], MARGIN = 1, function(element){
      element * exp(element * beta)
    }) %>% rowSums()

    denominator <- apply(risk_set[, -c(1,2)], MARGIN = 1, function(element){
      exp(element * beta)
    }) %>% rowSums()

    partial_sum[[k]] <- 
      preparedData[k, "event"] * (preparedData[k, -c(1,2)] - nominator/denominator)
  }
  do.call(rbind, partial_sum) %>%
    colSums() -> U_batch

  return(beta + learningRate * U_batch)
}

which transforms data to the input format as in survival::coxph function (and also sorts observations at the end to facilitate the implantation) with below function

prepareData <- function(formula, data) {
  # Parameter identification as in  `survival::coxph()`.
  Call <- match.call()
  indx <- match(c("formula", "data"),
                names(Call), nomatch = 0)
  if (indx[1] == 0) 
    stop("A formula argument is required")
  temp <- Call[c(1, indx)]
  temp[[1]] <- as.name("model.frame")

  mf <- eval(temp, parent.frame())
  Y <- model.extract(mf, "response")

  if (!inherits(Y, "Surv")) 
    stop("Response must be a survival object")
  type <- attr(Y, "type")

  if (type != "right" && type != "counting") 
    stop(paste("Cox model doesn't support \"", type, "\" survival data", 
               sep = ""))

  # collect times, status, variables and reorder samples 
  # to make the algorithm more clear to read and track
  cbind(event = unclass(Y)[,2], # 1 indicates event, 0 indicates cens
        times = unclass(Y)[,1],
        mf[, -1]) %>%
    arrange(times) 
}

which both functions are used in the above function that implements the gradient descent (order I) algorithm to estimate the coefficients in the cox proportional hazards model (this is different method from ?coxph.fit which uses cholesky decomposition of second derivatives to perform Newton-Raphson gradient descent order II algorithm)

library(assertthat)
library(survival)
coxphGD <- function(formula, data, learningRates = function(x){1/x},
                     beta_0 = 0, epsilon = 1e-5, max.iter = 500 ) {
  checkArguments(formula, data, learningRates,
                 beta_0, epsilon) -> beta_start # check arguments
  n <- length(data)
  diff <- epsilon + 1
  i <- 1
  beta_new <- list()     # steps are saved in a list so that they can
  beta_old <- beta_start # be tracked in the future
  # estimate
  while(i <= max.iter & diff > epsilon) {
    beta_new[[i]] <- coxphGD_step(formula = formula, beta = beta_old,
                                    learningRate = learningRates(i),
                                    data = data) %>% 
      unlist # unlist as this might be the result of foreach
    diff <- sqrt(sum((beta_new[[i]] - beta_old)^2))
    beta_old <- beta_new[[i]]
    i <- i + 1  ; cat("\r iteration: ", i, "\r")
  }
  # return results
  list(Call = match.call(), epsilon = epsilon, learningRates = learningRates,
       steps = i, coefficients = c(list(beta_start), beta_new))
}


checkArguments <- function(formula, data, learningRates,
                           beta_0, epsilon) {

  # + check names and types for every variables
  assert_that(is.function(learningRates))
  assert_that(is.numeric(epsilon))
  assert_that(is.numeric(beta_0))

  # check length of the start parameter
  if (length(beta_0) == 1) {
    beta_0 <- rep(beta_0, as.character(formula)[3] %>%
                    strsplit("\\+") %>%
                    unlist %>%
                    length)
  }

  return(beta_0)
}

but I think this implementation is somehow wrong because when I tried to perform gradient descent order I algorithm to estimate coefficients in the Cox proportional model for artificial data (generation based on How to create a toy survival (time to event) data with right censoring) I received results that are far from real coefficients beta = c(2,2) from which I have generated data

set.seed(456)
dataCox <- function(N, lambda, rho, x, beta, censRate){

  # real Weibull times
  u <- runif(N)
  Treal <- (- log(u) / (lambda * exp(x %*% beta)))^(1 / rho)

  # censoring times
  Censoring <- rexp(N, censRate)

  # follow-up times and event indicators
  time <- pmin(Treal, Censoring)
  status <- as.numeric(Treal <= Censoring)

  # data set
  data.frame(id=1:N, time=time, status=status, x=x)
}

x <- matrix(sample(0:1, size = 2000, replace = TRUE), ncol = 2)

dataCox(10^3, lambda = 5, rho = 1.5, x, beta = c(2,2), censRate = 0.2) -> dCox

library(assertthat)
library(magrittr)
library(survival)
library(dplyr)
library(foreach)

coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/x},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results


coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(1000*x)},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results2


coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(1000*sqrt(x))},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results3


coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(10000*x)},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results4

coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(100*x)},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results5

coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(100*sqrt(x))},
        beta_0 = 0, epsilon = 1e-5, max.iter = 50 ) -> results6


coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(100*sqrt(x))},
        beta_0 = 0, epsilon = 1e-4, max.iter = 50 ) -> results7

coxphGD(formula = Surv(time, status)~x.1+x.2,
        data = dCox, learningRates = function(x){1/(50*sqrt(x))},
        beta_0 = 0, epsilon = 1e-5, max.iter = 150 ) -> results8



library(ggplot2)

png('results1.png')
as.data.frame(t(simplify2array(results$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()

png('results2.png')
as.data.frame(t(simplify2array(results2$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()

png('results3.png')
as.data.frame(t(simplify2array(results3$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()


png('results4.png')
as.data.frame(t(simplify2array(results4$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()


png('results5.png')
as.data.frame(t(simplify2array(results5$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()


png('results6.png')
as.data.frame(t(simplify2array(results6$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()


png('results7.png')
as.data.frame(t(simplify2array(results7$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()


png('results8.png')
as.data.frame(t(simplify2array(results8$coefficients))) %>%
  ggplot(aes(V1, V2)) + geom_path() +
  coord_cartesian(xlim=c(0,3), ylim=c(0,3))
dev.off()

enter image description here enter image description here enter image description here enter image description here enter image description here enter image description here enter image description here enter image description here

I mean estimation process stops for $\beta$'s that are not even close to beta=c(2,2). My question is where is a mistake in the philosophy or implementation or is there a simpler way to compute gradient for partial log-likelihood in Cox proportional hazards model?

| cite | improve this question | |
$\endgroup$
-1
+50
$\begingroup$

Try this:

batchData <- preparedData
batchData <- batchData %>% arrange(-times)

  scores <- apply(batchData[, -c(1, 2)], MARGIN = 1, 
                  function(element) {
                    exp(element %*% beta)
                  })

  nominator <- apply(batchData[, -c(1, 2)], 2, function(x) cumsum(scores*x))
  denominator <- cumsum(scores)
  partial_sum <- (batchData[, -c(1, 2)] - nominator/denominator)*batchData[, "event"]

  U_batch <- partial_sum %>% colSums()
| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.