如何在R中结合PSU、Stratum与多重权重分析？求支持三者的R包

Great questions! Let's break this down step by step to address both your concerns.

1. 结合PSU、Stratum与多重权重的统计分析方法 #

When dealing with complex survey designs that include primary sampling units (PSU), stratification, and multiple weights, the survey package is the gold standard in R—it fully supports all three elements, whether you're working with different types of analytical weights (like sampling weights, post-stratification weights, or calibration weights) or replication weights (e.g., bootstrap or jackknife weights).

Here's a practical example using a sample data frame df with:

• psu: PSU identifier
• stratum: Stratification variable
• sampling_weight: Base sampling weight
• cal_weight: Calibrated secondary weight

First, create a survey design object that incorporates PSU and stratification:

library(survey)

# Build the core survey design with PSU, strata, and base weight
base_design <- svydesign(
  id = ~psu,
  strata = ~stratum,
  weights = ~sampling_weight,
  data = df
)

# If you need to apply a secondary calibration weight, update the design
calibrated_design <- calibrate(base_design, ~some_calibration_var, weights = ~cal_weight)

# Run a simple analysis, like calculating the weighted mean of income
svymean(~income, design = calibrated_design)

# For replication weights (e.g., multiple bootstrap weight columns), use svrepdesign
replication_design <- svrepdesign(
  id = ~psu,
  strata = ~stratum,
  weights = ~sampling_weight,
  repweights = ~rep_wt1 + rep_wt2 + rep_wt3, # Multiple replication weight columns
  type = "bootstrap",
  data = df
)

svymean(~income, design = replication_design)

If you need to analyze multiple weight scenarios side-by-side, you can create separate design objects for each weight or use the mitools package to combine results across weight sets (especially useful if paired with multiple imputation).

2. R Packages That Support PSU, Stratum, and Multiple Weights #

You're right that WeMix is focused more on multilevel modeling with weights and lacks robust support for PSU and stratification. Here are the top packages that cover all three requirements:

• survey: As mentioned above, this is the foundational package for complex survey analysis in R. It handles PSU, stratification, all types of weights (including multiple/replication weights), and supports everything from descriptive stats to regression models.
• srvyr: A tidyverse-friendly wrapper for survey, with syntax that aligns with dplyr. It retains full support for PSU, strata, and multiple weights, making it ideal if you prefer a more intuitive, pipe-based workflow. Example:

  library(srvyr)
  
  # Convert data to a survey design object
  survey_df <- as_survey_design(
    data = df,
    id = psu,
    strata = stratum,
    weights = sampling_weight
  )
  
  # Switch to a secondary weight and calculate a mean
  survey_df %>%
    as_survey(weights = cal_weight) %>%
    summarize(avg_income = survey_mean(income))
  

• mitools: If your multiple weights are paired with multiple imputed datasets, this package works seamlessly with survey to pool results across different weight-imputation combinations.

For most use cases, survey or srvyr will be your go-to choices—they're well-documented, widely used, and fully support all three elements you're working with.

内容的提问来源于stack exchange，提问作者Sorif Hossain