Abstract
In population studies on the etiology of disease, one goal is the estimation of the fraction of cases attributable to each of several causes. For example, pneumonia is a clinical diagnosis of lung infection that may be caused by viral, bacterial, fungal, or other pathogens. The study of pneumonia etiology is challenging because directly sampling from the lung to identify the etiologic pathogen is not standard clinical practice in most settings. Instead, measurements from multiple peripheral specimens are made. This paper introduces the statistical methodology designed for estimating the population etiology distribution and the individual etiology probabilities in the Pneumonia Etiology Research for Child Health (PERCH) study of 9,500 children for 7 sites around the world. We formulate the scientific problem in statistical terms as estimating the mixing weights and latent class indicators under a partially-latent class model (pLCM) that combines heterogeneous measurements with different error rates obtained from a case-control study. We introduce the pLCM as an extension of the latent class model. We also introduce graphical displays of the population data and inferred latent-class frequencies. The methods are tested with simulated data, and then applied to PERCH data. The paper closes with a brief description of extensions of the pLCM to the regression setting and to the case where conditional independence among the measures is relaxed.
Introduction
Determining the specific cause of childhood pneumonia—whether viral, bacterial, or fungal—is a critical public health challenge. However, the "gold standard" for diagnosis, which involves sampling fluid directly from the lung, is invasive and rarely performed in clinical practice. Instead, researchers and clinicians must rely on peripheral measurements, such as blood cultures or nasopharyngeal swabs.
These peripheral tests are inherently noisy. They suffer from imperfect sensitivity (failing to detect the actual pathogen in the lung) and imperfect specificity (detecting pathogens that are colonizing the upper respiratory tract but not causing the actual infection). Traditional statistical methods, such as logistic regression, often fail to account for these measurement errors or integrate disparate data types effectively. This paper introduces the Partially-Latent Class Model (pLCM), a Bayesian framework specifically designed to address these limitations by leveraging case-control data.
Methodology
The pLCM Framework
The pLCM is an extension of traditional Latent Class Models (LCM). The "partially-latent" designation refers to the model's treatment of the case-control study design:
- Known States: In the control group (healthy children), the latent state is known to be zero (no lung infection). This allows the model to estimate the "background noise" or false-positive rates (colonization) directly from the data.
- Latent States: In the case group (sick children), the specific cause of the pneumonia is unknown and treated as a latent variable ($j$).
Measurement Tiers
To handle varying levels of diagnostic accuracy, the model categorizes data into three distinct tiers:
- Gold Standard (GS): Tests with near-perfect sensitivity and specificity (rarely available).
- Silver Standard (SS): Tests with high specificity but imperfect sensitivity, such as blood cultures.
- Bronze Standard (BrS): Tests with imperfect sensitivity and specificity, such as PCR swabs, which are highly sensitive but prone to detecting background colonization.
Bayesian Inference and Identifiability
The framework employs a hierarchical Bayesian approach using Markov Chain Monte Carlo (MCMC) sampling to estimate the etiology fraction (the proportion of cases caused by each pathogen). A significant challenge in these models is identifiability; without gold standard data, the model may struggle to distinguish between a pathogen being rare but easy to detect versus common but difficult to detect. To resolve this, the authors incorporate "expert priors"—informed scientific estimates of test sensitivity—to anchor the model's inferences.
Application to the PERCH Study
The pLCM was applied to data from the Pneumonia Etiology Research for Child Health (PERCH) study, a massive global undertaking involving 9,500 children. For this analysis, a subset of 432 cases and 479 frequency-matched controls from a single site was utilized. The study focused on 11 primary pathogens, including seven viruses and four bacteria.
Key Findings
- Viral Dominance: The model estimated that viruses were the primary drivers of pneumonia in the studied cohort, with a total viral etiologic fraction of approximately 70% (95% CI: 57.0%–79.2%).
- Primary Pathogens: Respiratory Syncytial Virus (RSV) was identified as the single largest contributor (~29.4%), followed by Rhinovirus (~16.7%) and Streptococcus pneumoniae (~14.8%).
- Diagnostic Refinement: The model successfully differentiated between colonization and infection. For instance, while Rhinovirus was frequently detected in both cases and controls, the pLCM's joint distribution analysis correctly identified it as a significant cause in a specific subset of cases.
Discussion and Future Work
The pLCM provides a robust statistical tool for estimating disease etiology when direct observation is impossible. By using healthy controls to define background noise, the model provides a more accurate picture of population-level disease distribution than simple attributable fraction calculations.
However, the model currently assumes conditional independence, meaning the presence of one pathogen does not affect the detection of another. Future extensions of the pLCM aim to relax this assumption to account for biological competition or seasonality. Additionally, the authors propose extending the model into a regression setting to examine how etiology fractions change based on covariates such as age, geography, or season.
Summary updated