Nested Partially-Latent Class Models for Dependent Binary Data; Estimating Disease Etiology

Zhenke Wu, Maria Deloria Knoll, Scott L. Zeger

TL;DR

This paper introduces a novel statistical model, the "nested partially-latent class model" (npLCM), to better understand the causes of diseases like pneumonia by accounting for complex dependencies in diagnostic test results. This new approach leads to more accurate estimations of disease origins and improved diagnostic inference compared to existing methods.

What Problem Does This Paper Address?

Diagnosing diseases, especially those with multiple potential causes like childhood pneumonia, is challenging. Doctors and researchers often rely on a battery of tests, each producing binary (yes/no) results. The problem is that these test results aren't always independent; the presence of one pathogen might influence the likelihood of detecting another, or the tests themselves might have correlated errors. Traditional statistical models, like latent class models (LCMs), often assume these test results are independent given the underlying disease cause (this is called the "local independence" assumption). This assumption can lead to biased estimates of disease causes and less reliable diagnoses, especially when these dependencies are strong. The PERCH study, which aims to identify the specific pathogens causing pneumonia in children, highlights this challenge, as they have numerous binary measurements (PCR results for different pathogens) for each child.

Methods

The authors propose a "nested partially-latent class model" (npLCM). The core idea is to extend existing latent class models by introducing "subclasses" within each disease class. Imagine each disease (e.g., pneumonia caused by pathogen A) has several "sub-states" or "subclasses." The measurements (test results) are assumed to be independent within these subclasses. By nesting these subclasses within the main disease classes, the model can capture the residual dependence (the "local dependence") between measurements that the simpler LCMs miss.

Key methodological aspects include:

  • Partially-Latent Structure: They leverage case-control data. Control subjects (known to be disease-free) help estimate the "false positive rates" of the tests. This information is then "borrowed" to inform the model for cases.
  • Nested Subclasses: This is the novel part. Instead of assuming independence within a disease class, they introduce multiple subclasses within each class. This allows for complex correlations between test results.
  • Bayesian Framework: The model is estimated using Bayesian methods, specifically Markov Chain Monte Carlo (MCMC) via Gibbs sampling. This allows for incorporating prior knowledge (e.g., from experts on test sensitivities) and quantifying uncertainty in the estimates.
  • Stick-Breaking Priors: To avoid pre-specifying the number of subclasses (which can be difficult), they use "stick-breaking priors." This Bayesian technique allows the model to infer the most appropriate number of subclasses from the data, promoting parsimony and preventing overfitting.

Datasets

The primary dataset used for evaluation is from the Pneumonia Etiology Research for Child Health (PERCH) study. This study collected data on children hospitalized with severe pneumonia, including PCR assays on nasopharyngeal specimens for 6 different pathogens. The authors also used simulated data to rigorously test their model's performance under various conditions, including different degrees of local dependence.

Key Insights and Ideas

  • Local Dependence Matters: The paper strongly demonstrates that ignoring the dependencies between diagnostic test results (local dependence) can lead to significant biases in estimating disease causes, especially when these dependencies are substantial.
  • Subclasses as a Solution: The "nested subclass" approach is a powerful way to model these complex dependencies without making overly simplistic independence assumptions. It's like adding more nuanced states to the model to capture the real-world correlations.
  • Leveraging Controls: Using control data to inform the measurement characteristics (like false positive rates) is an efficient way to improve the model's ability to estimate disease etiology in cases, even when the true disease state is unknown.
  • Parsimony Through Priors: The use of stick-breaking priors is a clever way to let the data guide the complexity of the model (i.e., the number of subclasses), preventing overfitting and leading to more robust inferences.

Key Results

  • Reduced Bias: The npLCM significantly reduces estimation bias compared to simpler models that assume local independence, particularly in scenarios with strong local dependence. For example, in Scenario II with strong LD, the bias for one etiologic fraction was reduced from 26.2% to 13.6%.
  • Improved Efficiency: While accounting for LD, the npLCM maintains good estimation efficiency, meaning it doesn't lose much precision compared to simpler models when the underlying data is close to independent.
  • Better Coverage: The 95% credible intervals produced by the npLCM show better empirical coverage rates, meaning they more accurately reflect the true uncertainty in the estimates.
  • Model Fit: The npLCM demonstrated a better empirical fit to the PERCH data, as evidenced by posterior predictive checks for pairwise odds ratios.

Limitations and Future Work

  • Computational Cost: While the authors developed an efficient Gibbs sampler, complex Bayesian models can still be computationally intensive, especially with large datasets or many subclasses.
  • Single Primary Cause Assumption: The current framework assumes a single primary cause for pneumonia. The authors suggest extending the model to handle multiple co-occurring pathogens as causes, which would involve a more complex latent structure.
  • Covariate Effects: The authors mention that regression versions of the npLCM, allowing etiology distributions and false positive rates to depend on covariates (like season, age, HIV status), are a subject of current study. This would allow for more personalized and context-aware disease etiology estimation.
  • "Other" Causes: The model includes an "other" category for causes not explicitly measured, which is important for completeness but represents a simplification.

Summary updated

Abstract

The Pneumonia Etiology Research for Child Health (PERCH) study seeks to use modern measurement technology to infer the causes of pneumonia for which gold-standard evidence is unavailable. The paper describes a latent variable model designed to infer from case-control data the etiology distribution for the population of cases, and for an individual case given his or her measurements. We assume each observation is drawn from a mixture model for which each component represents one cause or disease class. The model addresses a major limitation of the traditional latent class approach by taking account of residual dependence among multivariate binary outcome given disease class, hence reduces estimation bias, retains efficiency and offers more valid inference. Such "local dependence" on a single subject is induced in the model by nesting latent subclasses within each disease class. Measurement precision and covariation can be estimated using the control sample for whom the class is known. In a Bayesian framework, we use stick-breaking priors on the subclass indicators for model-averaged inference across different numbers of subclasses. Assessment of model fit and individual diagnosis are done using posterior samples drawn by Gibbs sampling. We demonstrate the utility of the method on simulated and on the motivating PERCH data.
BibTeX
@inproceedings{wu2016nested,
  author = {Zhenke Wu and Maria Deloria Knoll and Scott L. Zeger},
  title = {{Nested Partially-Latent Class Models for Dependent Binary Data; Estimating Disease Etiology}},
  booktitle = {Biostatistics},
  year = {2016},
  doi = {10.1093/biostatistics/kxw037},
}

Loading related content...

Figure