Bayesian Factor Analysis for Binary and Ordinal Phenotypes with Missingness

Area of Science:

• Computational biology
• Statistical genetics
• Machine learning

Background:

• Many factor analysis methods require quantitative data, limiting their use with common binary/ordinal phenotypes.
• Clinical screening and questionnaires often yield non-continuous data, posing analytical challenges.

Purpose of the Study:

• To introduce FABOr (Factor Analysis of Binary and Ordinal data), a novel Bayesian framework for matrix factorization.
• To extend FABOr to handle missing not at random (MNAR) data.
• To evaluate FABOr's performance against existing methods for binary and ordinal data imputation.

Main Methods:

• FABOr utilizes a Bayesian framework with continuous priors for latent variables and appropriate binary/ordinal likelihoods for observed phenotypes.
• The framework incorporates extensions for analyzing data with structured missingness (MNAR).
• Performance was assessed using simulated data and the Simons Foundation SPARK dataset for autism spectrum disorder (ASD).

Main Results:

• FABOr demonstrated comparable performance to benchmark methods for binary data imputation.
• FABOr significantly outperformed all tested benchmark methods on ordinal data imputation.
• Application to the SPARK dataset showed FABOr improved imputation accuracy by up to 5% for binary and 23% for ordinal phenotypes.

Conclusions:

• FABOr provides an effective Bayesian approach for matrix factorization of binary and ordinal data.
• The method shows particular strength in handling ordinal phenotypes and structured missingness.
• FABOr enhances data imputation accuracy in real-world datasets, such as those for autism spectrum disorder research.