求R语言中计算部分截尾变量与连续变量相关系数及p-value的方法

Got it, let's figure out how to calculate a correlation coefficient (and its corresponding p-value) between a partially censored variable and a continuous variable in R—since you mentioned that the SurvCorr package only gives the coefficient without the p-value, here are two solid approaches you can use right away:

Since SurvCorr doesn't compute p-values out of the box, a permutation test is a great non-parametric way to estimate the significance of your observed correlation. Here's how to implement it:

# Load required packages
library(SurvCorr)
library(survival)

# Simulate example data (adjust to match your dataset)
set.seed(123)
n <- 100
censored_time <- rexp(n, rate = 0.5)
censoring_status <- rbinom(n, 1, 0.7)  # 30% of observations are censored
continuous_var <- rnorm(n, mean = 0.2 * censored_time, sd = 1)

# Calculate the observed correlation coefficient
observed_corr <- SurvCorr(Surv(censored_time, censoring_status), continuous_var, method = "pearson")$correlation

# Run permutation test to get p-value
num_permutations <- 1000
permuted_corrs <- numeric(num_permutations)

for (i in 1:num_permutations) {
  # Shuffle the continuous variable to break the relationship
  shuffled_continuous <- sample(continuous_var)
  permuted_corrs[i] <- SurvCorr(Surv(censored_time, censoring_status), shuffled_continuous, method = "pearson")$correlation
}

# Compute two-sided p-value
p_value <- mean(abs(permuted_corrs) >= abs(observed_corr))

# Print results
cat("Observed Correlation Coefficient:", round(observed_corr, 4), "\n")
cat("Permutation Test p-value:", round(p_value, 4), "\n")

The core idea here is: if there's no true correlation between the censored and continuous variable, shuffling the continuous variable should produce correlation coefficients clustered near zero. The p-value tells you how often a correlation as extreme as your observed one would occur by random chance.

The corcens package is built specifically for calculating correlations between a censored variable and a continuous variable, and it directly returns both the coefficient and p-value. Here's how to use it:

# Install the package first (if you haven't already)
# install.packages("corcens")
library(corcens)

# Use the corcens function with your data
corr_result <- corcens(x = Surv(censored_time, censoring_status), y = continuous_var, method = "pearson")

# View the full results (includes coefficient, standard error, and p-value)
print(corr_result)

This package supports Pearson, Spearman, and Kendall correlation methods, and uses either asymptotic normality or bootstrapping to compute the p-value—you can adjust the method argument to match your analysis needs.

内容的提问来源于stack exchange，提问作者user5054