Multiple Imputation of Missing Data in Practice: Basic Theory and Analysis Strategies

Cover
Half Title
Title Page
Copyright Page
Dedication
Contents
Foreword
Preface
1. Introduction
	1.1. A Motivating Example
	1.2. Definition of Missing Data
	1.3. Missing Data Patterns
	1.4. Missing Data Mechanisms
	1.5. Structure of the Book
2. Statistical Background
	2.1. Introduction
	2.2. Frequentist Theory
		2.2.1. Sampling Experiment
		2.2.2. Model, Parameter, and Estimation
		2.2.3. Hypothesis Testing
		2.2.4. Resampling Methods: The Bootstrap Approach
	2.3. Bayesian Analysis
		2.3.1. Rudiments
		2.3.2. Prior Distribution
		2.3.3. Bayesian Computation
		2.3.4. Asymptotic Equivalence between Frequentist and Bayesian Estimates
	2.4. Likelihood-based Approaches to Missing Data Analysis
	2.5. Ad Hoc Missing Data Methods
	2.6. Monte Carlo Simulation Study
	2.7. Summary
3. Multiple Imputation Analysis: Basics
	3.1. Introduction
	3.2. Basic Idea
		3.2.1. Bayesian Motivation
		3.2.2. Basic Combining Rules and Their Justifications
		3.2.3. Why Does Multiple Imputation Work?
	3.3. Statistical Inference on Multiply Imputed Data
		3.3.1. Scalar Inference
		3.3.2. Multi-parameter Inference
		3.3.3. How to Choose the Number of Imputations
	3.4. How to Create Multiple Imputations
		3.4.1. Bayesian Imputation Algorithm
		3.4.2. Proper Multiple Imputation
		3.4.3. Alternative Strategies
	3.5. Practical Implementation
	3.6. Summary
4. Multiple Imputation for Univariate Missing Data: Parametric Methods
	4.1. Overview
	4.2. Imputation for Continuous Data Based on Normal Linear Models
	4.3. Imputation for Noncontinuous Data Based on Generalized Linear Models
		4.3.1. Generalized Linear Models
		4.3.2. Imputation for Binary Data
			4.3.2.1. Logistic Regression Model Imputation
			4.3.2.2. Discriminant Analysis Imputation
			4.3.2.3. Rounding
			4.3.2.4. Data Separation
		4.3.3. Imputation for Nonbinary Categorical Data
		4.3.4. Imputation for Other Types of Data
	4.4. Imputation for a Missing Covariate in a Regression Analysis
	4.5. Summary
5. Multiple Imputation for Univariate Missing Data: Robust Methods
	5.1. Overview
	5.2. Data Transformation
		5.2.1. Transforming or Not?
		5.2.2. How to Apply Transformation in Multiple Imputation
	5.3. Imputation Based on Smoothing Methods
		5.3.1. Basic Idea
		5.3.2. Practical Use
	5.4. Adjustments for Continuous Data with Range Restrictions
	5.5. Predictive Mean Matching
		5.5.1. Hot-Deck Imputation
		5.5.2. Basic Idea and Procedure
		5.5.3. Predictive Mean Matching for Noncontinuous Data
		5.5.4. Additional Discussion
	5.6. Inclusive Imputation Strategy
		5.6.1. Basic Idea
		5.6.2. Dual Modeling Strategy
			5.6.2.1. Propensity Score
			5.6.2.2. Calibration Estimation and Doubly Robust Estimation
			5.6.2.3. Imputation Methods
	5.7. Summary
6. Multiple Imputation for Multivariate Missing Data: The Joint Modeling Approach
	6.1. Introduction
	6.2. Imputation for Monotone Missing Data
	6.3. Multivariate Continuous Data
		6.3.1. Multivariate Normal Models
		6.3.2. Models for Nonnormal Continuous Data
	6.4. Multivariate Categorical Data
		6.4.1. Log-Linear Models
		6.4.2. Latent Variable Models
	6.5. Mixed Categorical and Continuous Variables
		6.5.1. One Continuous Variable and One Binary Variable
		6.5.2. General Location Models
		6.5.3. Latent Variable Models
	6.6. Missing Outcome and Covariates in a Regression Analysis
		6.6.1. General Strategy
		6.6.2. Conditional Modeling Framework
		6.6.3. Using WinBUGS
			6.6.3.1. Background
			6.6.3.2. Missing Interactions and Squared Terms of Covariates in
			6.6.3.3. Imputation Using Flexible Distributions
	6.7. Summary
7. Multiple Imputation for Multivariate Missing Data: The Fully Conditional Specification Approach
	7.1. Introduction
	7.2. Basic Idea
	7.3. Specification of Conditional Models
	7.4. Handling Complex Data Features
		7.4.1. Data Subject to Bounds or Restricted Ranges
		7.4.2. Data Subject to Skips
	7.5. Implementation
		7.5.1. General Algorithm
		7.5.2. Software
			7.5.2.1. Using WinBUGS
	7.6. Subtle Issues
		7.6.1. Compatibility
		7.6.2. Performance under Model Misspecifications
	7.7. A Practical Example
	7.8. Summary
8. Multiple Imputation in Survival Data Analysis
	8.1. Introduction
	8.2. Imputation for Censored Event Times
		8.2.1. Theoretical Basis
		8.2.2. Parametric Imputation
		8.2.3. Semiparametric Imputation
		8.2.4. Merits
	8.3. Survival Analysis with Missing Covariates
		8.3.1. Overview
		8.3.2. Joint Modeling
		8.3.3. Fully Conditional Specification
		8.3.4. Semiparametric Methods
	8.4. Summary
9. Multiple Imputation for Longitudinal Data
	9.1. Introduction
	9.2. Mixed Models for Longitudinal Data
	9.3. Imputation Based on Mixed Models
		9.3.1. Why Use Mixed Models?
		9.3.2. General Imputation Algorithm
		9.3.3. Examples
	9.4. Wide Format Imputation
	9.5. Multilevel Data
	9.6. Summary
10. Multiple Imputation Analysis for Complex Survey Data
	10.1. Introduction
	10.2. Design-Based Inference for Survey Data
	10.3. Imputation Strategies for Complex Survey Data
		10.3.1. General Principles
			10.3.1.1. Incorporating the Survey Sampling Design
			10.3.1.2. Assuming Missing at Random
			10.3.1.3. Using Fully Conditional Specification
		10.3.2. Modeling Options
	10.4. Some Examples from the Literature
	10.5. Database Construction and Release
		10.5.1. Data Editing
		10.5.2. Documentation and Release
	10.6. Summary
11. Multiple Imputation for Data Subject to Measurement Error
	11.1. Introduction
	11.2. Rationale
	11.3. Imputation Strategies
		11.3.1. True Values Partially Observed
			11.3.1.1. Basic Setup
			11.3.1.2. Direct Imputation
			11.3.1.3. Accommodating a Specific Analysis
			11.3.1.4. Using Fully Conditional Specification
			11.3.1.5. Predictors under Detection Limits
		11.3.2. True Values Fully Unobserved
	11.4. Data Harmonization Using Bridge Studies
	11.5. Combining Information fromMultiple Data Sources
	11.6. Imputation for a Composite Variable
	11.7. Summary
12. Multiple Imputation Diagnostics
	12.1. Overview
	12.2. Imputation Model Development
		12.2.1. Inclusion of Variables
		12.2.2. Specifying Imputation Models
	12.3. Comparison between Observed and Imputed Values
		12.3.1. Comparison on Marginal Distributions
		12.3.2. Comparison on Conditional Distributions
			12.3.2.1. Basic Idea
			12.3.2.2. Using Propensity Score
	12.4. Checking Completed Data
		12.4.1. Posterior Predictive Checking
		12.4.2. Comparing Completed Data with Their Replicates
	12.5. Assessing the Fraction of Missing Information
		12.5.1. Relating the Fraction of Missing Information with Model Predictability
	12.6. Prediction Accuracy
	12.7. Comparison among Different Missing Data Methods
	12.8. Summary
13. Multiple Imputation Analysis for Nonignorable Missing Data
	13.1. Introduction
	13.2. The Implication of Missing Not at Random
	13.3. Using Inclusive Imputation Strategy to Rescue
	13.4. Missing Not at Random Models
		13.4.1. Selection Models
		13.4.2. Pattern Mixture Models
		13.4.3. Shared Parameter Models
	13.5. Analysis Strategies
		13.5.1. Direct Imputation
		13.5.2. Sensitivity Analysis
	13.6. Summary
14. Some Advanced Topics
	14.1. Overview
	14.2. Uncongeniality in Multiple Imputation Analysis
	14.3. Combining Analysis Results from Multiply Imputed Datasets: Further Considerations
		14.3.1. Normality Assumption in Question
		14.3.2. Beyond Sufficient Statistics
		14.3.3. Complicated Completed-Data Analyses: Variable Selection
	14.4. High-Dimensional Data
	14.5. Final Thoughts
Bibliography
Authors Index
Subject Index