如何用tidyr批量拟合多个Cox-PH模型？生存对象处理存疑

Got it, let's tackle this problem efficiently. When you're dealing with thousands of variables, looping through each one can be slow and messy—using the tidyverse (especially purrr and dplyr) is the way to go here. Here's a step-by-step solution that's both faster and cleaner:

Step 1: Load Required Packages #

First, make sure you have these packages installed and loaded—they'll handle data wrangling, model fitting, and result tidying:

library(survival)
library(tidyverse)
library(broom)

Step 2: Prepare Your Example Data (with Reproducibility) #

Let's start with your sample data, plus a seed to ensure results are consistent:

set.seed(123) # Fix randomness for testing
s <- Surv(sample(100:150, 5), sample(c(T, F), 5, replace = T))
df <- data.frame(var1 = rnorm(5), var2 = rnorm(5), var3 = rnorm(5))

Step 3: Batch Fit Models with purrr (No For Loop!) #

We have two efficient approaches here—pick the one that fits your use case:

Approach 1: Map Over Variable Names (Memory-Efficient) #

This method avoids reshaping your entire dataset, which is great when you have thousands of variables and want to save memory:

# Create a tibble of variable names, then fit models and extract results
cox_results <- tibble(variable = colnames(df)) %>%
  mutate(
    # Fit a Cox-PH model for each variable
    model = map(variable, ~ coxph(s ~ df[[.x]])),
    # Tidy model results (exponentiate=TRUE gives hazard ratios instead of log-HR)
    tidy_results = map(model, tidy, exponentiate = TRUE)
  ) %>%
  # Unnest the tidy results into a flat table
  unnest(tidy_results) %>%
  # Keep only the columns you care about
  select(variable, term, estimate, std.error, statistic, p.value)

Approach 2: Reshape Data to Long Format (More Intuitive) #

If you prefer a more visual approach (or need to inspect per-variable data later), reshape your data into long format first, then group and fit:

# Combine survival data with your predictor variables
surv_data <- tibble(time = s[,1], status = s[,2]) %>%
  bind_cols(df)

# Reshape to long format: one row per variable-sample pair
long_surv_data <- surv_data %>%
  pivot_longer(cols = starts_with("var"), names_to = "variable", values_to = "value")

# Fit models per variable and extract results
cox_results_long <- long_surv_data %>%
  group_by(variable) %>%
  nest() %>%
  mutate(
    model = map(data, ~ coxph(Surv(time, status) ~ value, data = .x)),
    tidy_results = map(model, tidy, exponentiate = TRUE)
  ) %>%
  unnest(tidy_results) %>%
  select(variable, term, estimate, std.error, statistic, p.value)

Step 4: Supercharge with Parallel Processing (For Thousands of Variables) #

If you have thousands of variables, you can use parallel computing to cut down runtime drastically. Use furrr (a parallelized version of purrr):

library(future)
library(furrr)

# Enable multi-session parallelism (uses all your CPU cores)
plan(multisession)

# Run the same variable-name approach in parallel
cox_results_parallel <- tibble(variable = colnames(df)) %>%
  mutate(
    model = future_map(variable, ~ coxph(s ~ df[[.x]])),
    tidy_results = future_map(model, tidy, exponentiate = TRUE)
  ) %>%
  unnest(tidy_results) %>%
  select(variable, term, estimate, std.error, statistic, p.value)

# Reset to sequential processing when done
plan(sequential)

Why This Is Better Than a For Loop #

• Faster: purrr functions are optimized for vectorized operations, avoiding the overhead of repeated subsetting in a for loop.
• Cleaner: Results come out in a tidy data frame, making it easy to filter, visualize, or export.
• Scalable: Parallel processing lets you leverage multiple CPU cores for large datasets.

内容的提问来源于stack exchange，提问作者Ronny Efronny