mice插补后用multinom做多项回归的系数命名异常问题求助

I've run into this exact issue before! The problem stems from how nnet::multinom structures its coefficient output versus linear models like lm. For categorical response variables (like age in nhanes2), multinom produces coefficients for each non-reference category—but the mice::pool function doesn't automatically map these to meaningful labels, falling back to generic numbered identifiers instead.

Here are a few solid solutions to fix this:

Solution 1: Manually Correct Coefficient Names #

You can extract the proper coefficient labels from one of the imputed models and overwrite the generic names in the pooled summary:

library(mice)
library(nnet)

# Run imputation (hide progress with printFlag=FALSE)
test <- mice(nhanes2, meth=c('sample','pmm','logreg','norm'), printFlag = FALSE)

# Fit multinomial regression across imputed datasets
m <- with(test, multinom(age ~ bmi, model = TRUE))

# Grab coefficient names from the first imputed model as a template
correct_coef_names <- rownames(coef(m$analyses[[1]]))

# Pool results and fix the row names
pooled_results <- pool(m)
pooled_summary <- summary(pooled_results)
rownames(pooled_summary) <- correct_coef_names

# View the cleaned output
pooled_summary

Solution 2: Use the broom Package for Tidy Model Output #

The broom package simplifies working with model results across different model types. Combine it with dplyr to cleanly pool and summarize your multinomial models:

library(mice)
library(nnet)
library(broom)
library(dplyr)
library(purrr)

test <- mice(nhanes2, meth=c('sample','pmm','logreg','norm'), printFlag = FALSE)

# Extract tidy results from each imputed model
tidy_imputations <- map_dfr(m$analyses, ~tidy(.x, conf.int = TRUE))

# Pool results following MICE's variance pooling rules
pooled_tidy <- tidy_imputations %>%
  group_by(term, response) %>%
  summarise(
    estimate = mean(estimate),
    # Combine within-imputation and between-imputation variance
    std.error = sqrt(var(estimate) + mean(std.error^2)),
    statistic = mean(statistic),
    p.value = mean(p.value),
    conf.low = mean(conf.low),
    conf.high = mean(conf.high)
  )

# View the neatly labeled results
pooled_tidy

Alternative: Use Bayesian Multinomial Regression with brms #

If you're open to Bayesian methods, the brms package handles imputed datasets (mira objects) seamlessly and automatically outputs properly labeled coefficients without extra work:

library(mice)
library(brms)

test <- mice(nhanes2, meth=c('sample','pmm','logreg','norm'), printFlag = FALSE)

# Fit Bayesian multinomial regression across imputations
brm_fit <- with(test, brm(age ~ bmi, family = categorical()))

# Pool and summarize the results
summary(pool(brm_fit))

内容的提问来源于stack exchange，提问作者Branners