A Latent Variable Model of Segregation Analysis for Ordinal Traits

Heping Zhang, Rui Feng, Hongtu ZhuJournal of the American Statistical Association 2003 · DOI

TL;DR

This paper introduces a novel latent variable model for analyzing ordinal traits in family studies, addressing a critical gap in existing segregation analysis methods. The model allows for the inclusion of covariates and gene-environment interactions, and its application to alcoholism data reveals a potential major gene component previously undetected.

What Problem Does This Paper Address?

Segregation analysis is a cornerstone of genetic epidemiology, aiming to understand how traits aggregate within families and whether they have a genetic basis, potentially following Mendelian inheritance patterns. While robust methods exist for quantitative and dichotomous traits, a significant limitation has been the lack of suitable models for ordinal traits, which are common in health sciences (e.g., disease stages, severity of disorders). Existing methods often require dichotomizing ordinal data, leading to loss of information and increased heterogeneity. This paper tackles this by proposing a flexible latent variable model that can handle ordinal traits, incorporating environmental factors and gene-environment interactions.

Methods

The authors propose a latent variable model that extends previous work on binary traits. The core idea is to model the observed ordinal trait (e.g., alcoholism categorized into normal, abuse, dependence) using unobserved latent variables that represent:

  • Common unmeasured environmental factors: Capturing shared influences within a family.
  • Genetic susceptibility: Representing the genetic predisposition of family founders.
  • Gene transmission: Modeling how susceptibility genes are passed from parents to offspring.

Key methodological contributions include:

  • Latent Variable Formulation: The model uses Bernoulli and Mendelian transmission assumptions for latent variables, allowing for complex familial correlations.
  • EM Algorithm: The Expectation-Maximization (EM) algorithm is employed for parameter estimation, a standard but computationally intensive approach for models with latent variables.
  • Peeling Algorithm: To address computational challenges, especially with large pedigrees, a "peeling" algorithm is introduced. This recursive process efficiently computes likelihoods by processing family members in a hierarchical manner, starting from the periphery and moving inwards.
  • Asymptotic Theory and Hypothesis Testing: The paper develops asymptotic theory for statistical inference, enabling hypothesis testing for familial aggregation (is the trait heritable?) and the presence of a major gene effect. This involves modified likelihood ratio tests and score tests.

Datasets

The primary dataset used for demonstrating the model's application is the Yale Family Study of Comorbidity of Alcoholism and Anxiety. This study includes data on alcoholism and related factors from 222 probands and 1,416 adult first-degree relatives and spouses. The authors use this dataset to analyze the familial transmission of alcoholism, categorized into three ordinal levels.

Key Insights and Ideas

  • Ordinal Traits Matter: The paper strongly advocates for analyzing ordinal traits directly, rather than dichotomizing them, as this preserves information and can reveal genetic components missed by simpler models.
  • Gene-Environment Interaction is Crucial: The latent variable framework naturally accommodates gene-environment interactions, which are believed to be important in complex diseases.
  • Peeling Algorithm for Efficiency: The development of a peeling algorithm is a significant computational advancement, making the analysis of large pedigrees more feasible.
  • Alcoholism's Major Gene Component: The application to alcoholism data suggests the presence of a major gene effect that was not detected in previous analyses using dichotomous traits, highlighting the power of the new model.

Key Results

  • Familial Aggregation: The analysis of the Yale Family Study data provided overwhelming support for familial transmission of alcoholism (p-value < 10⁻²⁷), confirming a strong hereditary component.
  • Major Gene Effect: A score test and permutation procedure indicated significant evidence for a major gene effect in alcoholism transmission, a finding not apparent in prior dichotomous analyses.
  • Model Performance: Simulation studies demonstrated that the proposed latent variable model is more powerful than traditional methods when analyzing ordinal traits and that the asymptotic theory is adequate for practical applications.
  • Covariate Robustness: Simulations showed that the proposed testing procedures are reasonably robust to the omission of relevant covariates, though large correlated covariates can impact power.

Limitations and Future Work

  • Computational Intensity: While the peeling algorithm improves efficiency, the analysis of very large pedigrees can still be computationally intensive. The authors suggest that exploring more efficient computational algorithms is an important area for future work.
  • Robustness to Misspecification: While preliminary assessments of robustness to covariate omission were conducted, further research is needed to scrutinize the behavior of the tests under broader conditions, especially in misspecified models.
  • Theoretical Validation: The paper notes that while simulation results validate the asymptotic theory, the mathematical proof for the asymptotic distribution of the score test for inheritance is complex and warrants further investigation.
  • Generalizability: The authors emphasize that the inference relies on the assumption that all important covariates are included in the model, and further investigation into the reliability of tests for familial aggregation and genetic inheritance in family studies is needed.

Summary updated

Abstract

AbstractMany health conditions, including cancer and psychiatric disorders, are believed to have a complex genetic basis, and genes and environmental factors are likely to interact in the presence and severity of these conditions. Assessing familial aggregation and inheritability of disease is a classic topic of genetic epidemiology, commonly referred to as segregation analysis. Although today it is routine to conduct such analyses for quantitative and dichotomous traits, methods and software that accommodate ordinal traits do not exist. Tothis end, we propose a latent variable model by extending the work of Zhang and Merikangas, who examined binary traits. The advantage of this latent variable model lies in its flexibility to include environmental factors (usually represented by covariates) and its potential to allow gene–environment interactions. The model building uses the EM algorithm for maximization and a peeling algorithm for computational efficiency. We provide asymptotic theory for statistical inference, and conduct simulation studies to confirm that the asymptotic theory is adequate in practical applications. We also apply our model to examine the familial transmission of alcoholism, which is categorized into three ordinal levels: normal control, alcohol abuse, and alcohol dependence. Our analysis not only confirms that alcoholism is familial, but also suggests that the transmission may have a major gene component not revealed by previous analyses using dichotomous traits.KEY WORDS: AlcoholismEM algorithmFrailtyLatent variableMixture modelPeeling algorithmSegregation analysis
BibTeX
@article{zhang2003latent,
  author = {Heping Zhang and Rui Feng and Hongtu Zhu},
  title = {{A Latent Variable Model of Segregation Analysis for Ordinal Traits}},
  journal = {Journal of the American Statistical Association},
  year = {2003},
  doi = {10.1198/016214503000000981},
}
Figure