High-dimensional, outcome-dependent missing data problems: Models for the human loci

Area of Science:

• Genetics
• Biostatistics
• Computational Biology

Background:

• Missing data is prevalent in high-dimensional biological datasets.
• Imputation and expectation-maximization (EM) algorithms are used for data reconstruction.
• Integrating regression models into imputation may reduce bias in regression coefficients.

Purpose of the Study:

• To evaluate outcome-based EM algorithms for reconstructing KIR diplotypes with missing data.
• To compare strategies incorporating high-dimensional regression models against a baseline EM algorithm.

Main Methods:

• Extended a previously proposed EM algorithm to include a high-dimensional regression model.
• Evaluated three strategies: allelic predictors only, allelic predictors with haplotype selection, and penalized regression.
• Compared these strategies with a baseline EM algorithm without an outcome model via simulation.

Main Results:

• Outcome-based EM algorithms outperformed the baseline in extreme scenarios of effect sizes and missingness.
• In most cases, the baseline EM algorithm performed superiorly or comparably.
• The inclusion of an outcome model can potentially introduce harmful effects and bias.

Conclusions:

• Outcome-based missing data models in high-dimensional settings require careful application.
• These models may lead to biased results, particularly when reconstructing KIR diplotypes.
• A baseline EM algorithm without outcome modeling is often a more robust approach.