TL;DR
This paper tackles the challenge of analyzing complex biological data, like genetic studies, where individuals are sampled non-randomly (case-control studies) and exhibit intricate dependencies. The authors introduce novel methods, AEP and APL, that accurately estimate model parameters by combining Gaussian Processes with case-control sampling adjustments, offering a more robust and principled approach than existing techniques.
What Problem Does This Paper Address?
Analyzing complex datasets often involves two major challenges: ascertainment and complex dependencies. Ascertainment occurs when the probability of sampling an individual depends on their outcome (e.g., having a disease), which is common in case-control studies for rare diseases. Complex dependencies arise when individuals are related or share similar genetic backgrounds, meaning their data points are not independent. Traditional statistical models struggle with these issues, leading to biased results. Specifically, existing methods for Gaussian Processes (GPs) and Generalized Linear Mixed Models (GLMMs) often ignore the ascertainment bias or struggle with the computational demands of full-rank, non-sparse covariance structures common in modern high-dimensional data like genome-wide association studies (GWAS). This paper aims to provide a statistically sound and computationally tractable likelihood-based solution for these scenarios.
Methods
The core contribution of this paper lies in developing approximate likelihood methods for Gaussian Processes under case-control sampling. They propose two main approaches:
- Ascertained Pairwise Likelihood (APL): This method extends pairwise likelihood (PL) to account for ascertainment. PL approximates the joint distribution of all variables by a product of marginal distributions of pairs of variables. APL modifies this by conditioning on the sampling indicators, allowing for accurate ratio computation.
- Ascertained Expectation Propagation (AEP): This is a generalization of the Expectation Propagation (EP) algorithm, a popular technique for approximating complex distributions. AEP adapts EP to handle ascertainment by modifying the iterative update step to account for the sampling probabilities. This method is more computationally intensive but is shown to be more robust.
Both methods aim to estimate the kernel hyperparameters of GPs, which are crucial for modeling the complex dependencies. They are compared against a moment-based method (PCGC) and plain EP (which ignores ascertainment).
Datasets
The paper uses simulated data that closely mimics real Genome-Wide Association Studies (GWAS) with a case-control sampling design. They also analyze real data from the Wellcome Trust Case Control Consortium, focusing on seven complex disorders with population prevalences ranging from 0.1% to 6%, using approximately 3,700 individuals and 280,000 genetic variants.
Key Insights and Ideas
- Ascertainment is Crucial: Ignoring the case-control sampling scheme in GP models can lead to nonsensical probabilistic interpretations and severely biased parameter estimates, as demonstrated by the poor performance of plain EP.
- AEP's Robustness: The AEP method proves to be remarkably robust to modeling violations, such as errors in feature standardization, making it a reliable choice for real-world applications where data preprocessing might not be perfect.
- Non-linear Kernels are Tricky: Standard methods like PCGC and APL, which rely on first-order Taylor expansions, struggle with non-linear kernels. AEP, however, shows promise in handling these more complex dependency structures.
- Likelihood-Based Inference is Valuable: While moment-based methods like PCGC are computationally efficient, the likelihood-based AEP provides a more principled framework for model comparison, inference, and prediction, which is essential for a comprehensive understanding of the data.
Key Results
- Accurate Hyperparameter Estimation: AEP and APL (and PCGC) accurately estimate hyperparameters when the true trait prevalence is known and the model is correctly specified. Plain EP is severely biased unless the prevalence is 50% (no ascertainment).
- Robustness to Sample Size and Dimensionality: All methods improve with larger sample sizes and higher dimensionality. AEP is particularly accurate with a small number of features.
- Sensitivity to Misspecification: AEP is highly robust to feature standardization misspecification, while APL is highly sensitive, and PCGC is moderately sensitive.
- Real-World Performance: In real data analysis of complex genetic disorders, AEP and PCGC provide concordant heritability estimates, while APL estimates are significantly down-biased, aligning with simulation findings.
- Improved Latent Variable Inference: AEP provides a clearer separation between cases and controls in the posterior distribution of latent variables, which is clinically useful for assessing disease severity.
Limitations and Future Work
- Computational Complexity: AEP scales cubically with sample size, limiting its direct application to very large datasets. The authors suggest that GP approximation techniques like mixture-of-experts or random feature expansions could be used to scale up AEP.
- Theoretical Guarantees: While AEP performs well empirically, formal theoretical guarantees for its performance are still an active area of research, especially for standard EP.
- Advanced Kernels and Inference Tasks: The paper suggests future work could explore GPs with more advanced kernels, as well as extending the approach to outcome prediction and testing of fixed effects, which currently rely on heuristic methods.
- Statistical Theory for Non-Sparse Dependencies: The paper highlights a gap in statistical theory for scenarios with non-sparse dependencies (where all units are correlated), particularly for the r/m → 0 asymptotic regime, which is relevant to their work.
Summary updated