Efficient mixed model approach for large-scale genome-wide association studies of ordinal categorical phenotypes

Wenjian Bi, Wei Zhou, Rounak Dey, Bhramar Mukherjee, Joshua N. Sampson, Seunggeun Lee

TL;DR

This paper introduces POLMM, a novel and efficient mixed-model approach for analyzing ordinal categorical phenotypes in large-scale genome-wide association studies (GWAS). POLMM accurately models the complex relationships in such data, controls for inflated Type I error rates, and offers improved statistical power compared to existing methods, making it a valuable tool for uncovering genetic associations with traits like preferences and behaviors.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) are powerful tools for identifying genetic variants associated with complex traits. However, many human traits are not simply binary (yes/no) or continuous (e.g., height) but rather ordinal categorical. Think of survey responses like "strongly disagree," "disagree," "neutral," "agree," "strongly agree," or preference scales from 1 to 9. Analyzing these ordinal phenotypes using methods designed for binary or quantitative traits can lead to inaccurate results, including inflated Type I error rates (false positives) or reduced statistical power (missing true associations). Existing methods struggle to properly account for the ordered nature of these categories and the complex relatedness between individuals in large biobanks.

Methods

The core of this paper is the development of the Proportional Odds Logistic Mixed Model (POLMM). This approach extends the proportional odds logistic model by incorporating a random effect to account for sample relatedness, a crucial factor in biobank data. POLMM uses efficient algorithms (penalized quasi-likelihood and average information restricted maximum likelihood) to fit the mixed model and then employs Saddlepoint Approximation (SPA) to calibrate p-values, which is particularly important for controlling Type I error rates, especially with unbalanced phenotypic distributions.

The authors also present two versions of POLMM:

  • DensePOLMM: Uses a dense genetic relationship matrix (GRM), similar to methods like SAIGE and BOLT-LMM, offering high accuracy.
  • FastPOLMM: Employs a sparse GRM, making it more computationally efficient and memory-friendly, akin to fastGWA.

These methods are designed to be scalable for large datasets with hundreds of thousands of individuals.

Datasets

The primary dataset used for applying and evaluating POLMM is the UK Biobank. This large-scale biobank contains genetic and phenotypic data for hundreds of thousands of individuals. The authors analyzed 258 ordinal categorical phenotypes from 408,961 white British participants. They also used simulated data for extensive validation of their method's performance, including Type I error rates and statistical power under various scenarios.

Key Insights and Ideas

  • Ordinality Matters: Treating ordinal categorical data as quantitative or simply dichotomizing it leads to significant statistical issues. POLMM's proportional odds model directly leverages the ordered nature of the categories.
  • SPA is Crucial for Robustness: Using Saddlepoint Approximation (SPA) for p-value calibration is key to controlling Type I error rates, especially when phenotypic distributions are highly unbalanced (e.g., many individuals in one category and very few in others).
  • FastPOLMM Offers a Practical Trade-off: While DensePOLMM is highly accurate, FastPOLMM provides a substantial improvement in computational efficiency with only a limited loss of power, making it ideal for very large-scale analyses.
  • Rare Variants Benefit from POLMM: The method shows promise in detecting associations with low-frequency and rare variants, which are often missed by other approaches.

Key Results

  • Type I Error Control: POLMM methods (both Dense and Fast) successfully controlled Type I error rates at a stringent significance level (5 × 10⁻⁸) across various phenotypic distributions and MAFs, outperforming other methods that showed inflated error rates, particularly for low-frequency variants and unbalanced data.
  • Statistical Power: POLMM demonstrated superior statistical power compared to other methods, especially when phenotypic distributions were unbalanced.
  • Computational Efficiency: FastPOLMM was significantly faster and used less memory than DensePOLMM and other dense GRM-based methods, fitting a null mixed model for ~400,000 subjects in under an hour with minimal memory.
  • UK Biobank Application: Applying FastPOLMM to 258 ordinal phenotypes in UK Biobank identified 5,885 genome-wide significant variants, with 7.2% being rare variants (MAF < 0.01).

Limitations and Future Work

  • Variance Component Estimation Bias: The paper notes that the variance component estimate (τ) from POLMM is biased and should not be directly used for heritability estimation, a limitation shared with other methods using pseudo quasi-likelihood (PQL). However, they observed more accurate estimates as the number of categories increased.
  • Infinitesimal Architecture Assumption: POLMM, like many mixed models, assumes an infinitesimal genetic architecture (effects are normally distributed). If the genetic architecture is non-infinitesimal, the method might sacrifice power.
  • Future Directions: The authors plan to extend POLMM to gene- or region-based tests for better identification of rare variants. They also aim to evaluate machine learning approaches like REGENIE for ordinal data and adapt POLMM for matched case-control study designs.

Summary updated

Abstract

Abstract In genome-wide association studies (GWAS), ordinal categorical phenotypes are widely used to measure human behaviors, satisfaction, and preferences. However, due to the lack of analysis tools, methods designed for binary and quantitative traits have often been used inappropriately to analyze categorical phenotypes, which produces inflated type I error rates or is less powerful. To accurately model the dependence of an ordinal categorical phenotype on covariates, we propose an efficient mixed model association test, Proportional Odds Logistic Mixed Model (POLMM). POLMM is demonstrated to be computationally efficient to analyze large datasets with hundreds of thousands of genetic related samples, can control type I error rates at a stringent significance level regardless of the phenotypic distribution, and is more powerful than other alternative methods. We applied POLMM to 258 ordinal categorical phenotypes on array-genotypes and imputed samples from 408,961 individuals in UK Biobank. In total, we identified 5,885 genome-wide significant variants, of which 424 variants (7.2%) are rare variants with MAF &lt; 0.01.
BibTeX
@article{bi2021efficient,
  author = {Wenjian Bi and Wei Zhou and Rounak Dey and Bhramar Mukherjee and Joshua N. Sampson and Seunggeun Lee},
  title = {{Efficient mixed model approach for large-scale genome-wide association studies of ordinal categorical phenotypes}},
  journal = {The American Journal of Human Genetics},
  year = {2021},
  doi = {10.1016/j.ajhg.2021.03.019},
}
Figure