如何在gtsummary的tbl_uvregression中结合coxph指定聚类效应？

Hey there! Let's fix your clustering issue with tbl_uvregression and coxph—I've been in this spot before, so I know exactly where to adjust.

First, Let's Clear Up the Missteps in Your Original Code #

Your current approach has a few key issues that are preventing the clustering effect from being applied:

• coxph doesn't recognize the lme4-style (1|grade) syntax for clustering. Instead, we use the cluster() function built into the survival package to adjust standard errors for clustered data.
• You set y = response, but coxph requires a Surv() object (combining time and status) as the outcome for survival analysis.
• The family = binomial argument is for generalized linear models, not Cox proportional hazards models—this shouldn't be included here.

Correct Code Implementation #

Here's the revised code that properly applies clustering by grade in your single-variable Cox regression table:

library(coxph)
library(gtsummary)
library(survival)

# Load trial dataset
data(trial)

# Create the clustered single-variable Cox regression table
trial %>% 
  tbl_uvregression(
    method = coxph,
    # Define the survival outcome correctly with Surv()
    y = Surv(time, status),
    exponentiate = TRUE,
    pvalue_fun = function(x) style_pvalue(x, digits = 2),
    # Add cluster(grade) to the formula to adjust for clustering
    formula = "{y} ~ {x} + cluster(grade)"
  )

Why This Works #

• The cluster(grade) term tells coxph to compute robust standard errors that account for within-cluster correlation (i.e., observations in the same grade group are not independent).
• We're using the correct survival outcome (Surv(time, status)) which is required for Cox models.
• Removed the unnecessary family = binomial argument since it's irrelevant for survival analysis.

Verify the Results #

To confirm the clustering is working, you can compare the standard errors/confidence intervals from this table to a standalone coxph model with clustering:

# Example standalone clustered Cox model for one variable
coxph(Surv(time, status) ~ age + cluster(grade), data = trial) %>% 
  summary()

The standard errors and CIs from this should match what's in your gtsummary table—this confirms the clustering adjustment is applied correctly.

内容的提问来源于stack exchange，提问作者gideon1321