Multiple imputation for the analysis of incomplete compound variables

Jiwei Zhao, Richard J. Cook, Changbao WuCanadian 2015 · DOI

TL;DR

This paper introduces novel multiple imputation (MI) strategies for analyzing "compound variables" (like BMI) when their underlying components (like height and weight) are missing. They propose two MI approaches, one imputing the compound variable directly and another imputing the components, and show that the latter is more efficient. A shrinkage estimator is also introduced to mitigate bias when the imputation model is misspecified.

What Problem Does This Paper Address?

In many health and scientific studies, researchers are interested in a "compound variable" that's calculated from several other variables. A common example is Body Mass Index (BMI), which is derived from height and weight. The challenge arises when one or more of these component variables are missing for some individuals. Simply ignoring these individuals (complete case analysis) leads to inefficient use of data and potentially biased results. Existing methods for handling missing data might not be optimal for compound variables, especially when the missingness patterns are complex. This paper aims to develop and evaluate more efficient and robust methods for analyzing such compound variables when their components are incomplete.

Methods

The authors propose two main frequentist multiple imputation (MI) strategies:

  • MI1 (Imputing the Compound Variable Directly): This approach uses an auxiliary model to impute the compound variable (e.g., BMI) directly. This is simpler but might be less efficient if the component variables contain valuable information not captured by the compound variable alone.
  • MI2 (Imputing Component Variables): This strategy imputes the missing component variables (e.g., height and weight) using a multivariate imputation model. This leverages the relationships between the components and is expected to be more efficient.

They also introduce a shrinkage estimator to address potential bias that can arise if the imputation model (especially for MI2) is not perfectly specified. This estimator combines the results from MI1 and MI2 to offer a more robust estimate. The theoretical properties of these estimators are analyzed using large-sample theory, and their performance is evaluated through simulations and a real-world application.

Datasets

The primary dataset used for evaluation is from the Canadian Youth Smoking Survey. This dataset contains information on body mass index (BMI) and its components (height and weight) for adolescents, with missing values for some individuals.

Key Insights and Ideas

  • Leveraging Component Relationships is Key: Imputing the individual components (MI2) of a compound variable is generally more efficient than imputing the compound variable directly (MI1), especially when the component variables have strong interrelationships.
  • Shrinkage for Robustness: A shrinkage estimator can effectively reduce estimation bias when the imputation model for the component variables is misspecified, offering a more reliable result than relying solely on a potentially incorrect model.
  • MI2's Dependence on Model Specification: While MI2 offers higher efficiency, its validity hinges on the correct specification of the multivariate model for the component variables.
  • Practical Application: The methods are demonstrated on a real-world dataset concerning BMI in Canadian youth, highlighting their practical utility in public health research.

Key Results

  • Efficiency Gains: Simulation studies showed that MI1 and MI2 consistently outperformed the naive complete-case analysis (CC) in terms of efficiency (lower variance). MI2 was generally more efficient than MI1.
  • Shrinkage Estimator Performance: The shrinkage estimator (MIS) demonstrated good performance, particularly in reducing bias when the imputation model was misspecified, and often achieved coverage probabilities close to the nominal level.
  • Application Findings: Analysis of the Canadian Youth Smoking Survey data revealed insights into the prevalence of overweight and obesity among youth, with differences observed between genders and age groups, and the imputation methods provided more complete estimates than the CC analysis.

Limitations and Future Work

  • Model Specification for MI2: The performance of MI2 is sensitive to the correct specification of the multivariate imputation model for the component variables. The paper explores a bivariate normal model but acknowledges that identifying the "correct" model can be challenging.
  • Extension to Higher Dimensions: The current work focuses on compound variables with two components. The authors suggest that their developed MI strategies can be extended to compound variables with more than two components, which are common in practice.
  • Nonignorable Missing Data: The current framework assumes missing at random (MAR). The authors note that extending these methods to handle nonignorable missing data scenarios would be a valuable area for future research, though it would likely require incorporating external information or more complex modeling.

Summary updated

Abstract

Abstract In many settings interest lies in modelling a compound variable defined as a function of two or more component variables. When one or more of the components are missing, the compound variable is not observed and a strategy for handling incomplete data is required. Analyses based on individuals with complete data are inefficient and yield potentially inconsistent estimators. We develop a multiple imputation strategy in this setting with an auxiliary model for imputing the compound variable directly, and one based on a multivariate imputation model for the component variables. Asymptotic properties of the imputation‐based estimators are presented for the case in which the imputation model is correctly specified, and a shrinkage estimator is proposed to reduce the bias arising from misspecification of the imputation model. Finite sample properties of the various estimators are examined through simulations. An application to data from the Canadian Youth Smoking Survey involving a study of body mass index illustrates the approach. The Canadian Journal of Statistics 43: 240–264; 2015 © 2015 Statistical Society of Canada
BibTeX
@inproceedings{zhao2015multiple,
  author = {Jiwei Zhao and Richard J. Cook and Changbao Wu},
  title = {{Multiple imputation for the analysis of incomplete compound variables}},
  booktitle = {Canadian},
  year = {2015},
  doi = {10.1002/cjs.11249},
}
Figure