Great question—handling multiple outcome variables in multiple imputation (MI) requires careful thought to build models that avoid bias and make the most of variable correlations. Below are my top resources, broken down into methodological papers, practical modeling guidance, and software tools tailored to this scenario:

These papers dive into the statistical theory and best practices for multi-outcome MI:

• "Multiple Imputation for Multivariate Missing Data: A Data Analyst’s Perspective" (Schafer, 1997) – A classic foundational text that explores joint modeling approaches for multi-variable (including multi-outcome) scenarios. Focus on chapters covering how to incorporate multiple outcomes into imputation models while preserving inter-variable correlations.
• "Multiple Imputation of Missing Data in Epidemiologic Studies" (White et al., 2011) – Tailored to epidemiological research (which often involves multiple outcomes), this paper specifically addresses adapting imputation models for mixed outcome types (continuous, categorical) and avoiding bias from outcome-induced missingness.
• "Joint Modeling of Multiple Outcomes with Missing Data" (Daniels & Hogan, 2008) – A deep dive into the statistical frameworks for joint imputation of multiple outcomes, including multivariate normal and generalized linear mixed models. Ideal if you want to understand the technical underpinnings.
• "Handling Missing Data in Clinical Trials with Multiple Outcomes" (Mallinckrodt et al., 2013) – Focused on clinical trial settings, this paper provides practical advice on balancing model complexity and stability, plus methods to evaluate imputation quality for multi-outcome datasets.

For hands-on tips to build robust multi-outcome MI models:

• A key best practice from the MI community is to never impute each outcome separately—this breaks inter-variable correlations and leads to biased results. Instead:

  1. Include all variables associated with missingness and all outcomes in your imputation model to leverage their correlations
  2. Match the imputation method to the outcome type (e.g., predictive mean matching for continuous outcomes, logistic regression for binary outcomes)
  3. Avoid excluding outcomes from the imputation model unless you can confirm missingness is completely random (MCAR)

• Many of the papers above include step-by-step case studies for multi-outcome datasets—start with White et al. for a more applied walkthrough.

These tools are designed to handle multi-outcome scenarios smoothly:

• mice (R Package) – The gold standard for fully conditional specification (FCS) MI. It lets you assign different imputation methods to each outcome (e.g., pmm for continuous, logreg for binary) and easily include all outcomes/covariates in the model. Example snippet:

  library(mice)
  # Create predictor matrix including all variables as predictors
  pred_matrix <- make.predictorMatrix(your_multi_outcome_data)
  # Define imputation methods for each variable (adjust based on your data)
  imputation_methods <- c("", "pmm", "logreg", "polr") # Covariate, continuous, binary, ordinal outcome
  # Run multiple imputation
  imputed_datasets <- mice(your_multi_outcome_data, m = 5, method = imputation_methods, predictorMatrix = pred_matrix)
  

• Amelia (R Package) – Great for datasets with mostly continuous outcomes. It uses joint multivariate normal modeling to automatically capture correlations between outcomes, making it simple to set up for large multi-outcome datasets.
• SAS PROC MI – Supports both joint modeling and FCS for multi-outcome data. Use the MODEL statement to specify multiple outcomes and their respective distributions.
• Stata mi Command – Offers a complete workflow for multi-outcome MI, including FCS with support for mixed outcome types. It also has built-in tools to assess imputation convergence and quality.

If you’re new to multi-outcome MI, I’d suggest starting with the White et al. paper and experimenting with mice in R—it’s flexible, well-documented, and widely used in the community. Don’t forget to validate your imputations by checking convergence plots and comparing observed vs. imputed variable distributions!

内容的提问来源于stack exchange，提问作者user166625