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 analyzing complex, correlated binary measurements. This new approach accounts for "local dependence" between these measurements, leading to more accurate and less biased estimates of disease origins compared to traditional methods.
What Problem Does This Paper Address?
Diagnosing the cause of a disease, especially when direct sampling is difficult (like pneumonia in children), relies on interpreting multiple, often imperfect, diagnostic tests or measurements. Traditional statistical models, like latent class models (LCMs), often assume these measurements are independent once the underlying disease "class" is known. However, in reality, these measurements can be dependent even within a disease class, a phenomenon called "local dependence" (LD). Ignoring LD can lead to biased estimates of disease prevalence (etiologic fractions) and less reliable diagnoses. This paper aims to address this limitation by developing a model that explicitly accounts for LD in multivariate binary data, using case-control studies as a primary application.
Methods
The core of this paper is the development of the nested partially-latent class model (npLCM). This model builds upon existing "partially-latent class models" (pLCMs) by introducing a hierarchical structure. Here's a breakdown:
- Latent Classes: The model assumes that each case of pneumonia belongs to one of several latent "disease classes," each representing a specific pathogen or cause.
- Subclasses (Nesting): Within each disease class (and for the control group), the model introduces "subclasses." Measurements within a subclass are assumed to be independent. This nesting allows for the modeling of local dependence.
- Borrowing Strength: Crucially, the model uses data from the control group (where the "no disease" class is known) to inform the measurement characteristics (like false positive rates) for the case groups. This "borrowing strength" helps to estimate parameters more efficiently and robustly.
- Bayesian Framework: The model is implemented within a Bayesian framework, using Markov Chain Monte Carlo (MCMC) methods (specifically, a blocked Gibbs sampler) for posterior inference.
- Stick-Breaking Priors: To handle the uncertainty in the number of subclasses and encourage parsimony, stick-breaking priors are used. This allows the model to infer the optimal number of subclasses from the data.
The npLCM extends the pLCM by allowing for this nested structure, thus explicitly modeling local dependence.
Datasets
The primary dataset used for evaluation and demonstration is from the Pneumonia Etiology Research for Child Health (PERCH) study. This study collected nasopharyngeal PCR assay data on cases of severe and very severe pneumonia and controls from multiple countries. The paper uses preliminary data from one site for its illustrative analysis, including data for 592 cases and 613 controls across 6 pathogens.
Key Insights and Ideas
- Local Dependence Matters: The paper strongly emphasizes that ignoring local dependence between diagnostic measurements can lead to significant bias in estimating disease prevalence and potentially flawed inferences.
- Nesting as a Solution: The novel "nesting" of subclasses within disease classes is a key innovation that effectively captures local dependence without requiring an overly complex model.
- Controls Inform Cases: Using control data to estimate measurement properties (like false positive rates) is a powerful strategy that improves the efficiency and validity of inferences about disease causes in cases.
- Parsimony via Stick-Breaking: The use of stick-breaking priors is an elegant way to allow the model to learn the appropriate number of subclasses from the data, preventing overfitting and promoting parsimonious models.
Key Results
- Reduced Bias: In simulations with strong local dependence, the npLCM significantly reduced estimation bias for etiologic fractions compared to a simpler model that ignored LD (pLCM). For example, biases were reduced from over 100% to around 40% in some scenarios.
- Improved Efficiency: The npLCM generally showed comparable or smaller Mean Squared Errors (MSEs) than the pLCM, indicating good efficiency even when the truth was close to local independence.
- Better Coverage: The npLCM produced 95% credible intervals with near-nominal empirical coverage rates, suggesting more reliable uncertainty quantification.
- PERCH Data Application: Analysis of the PERCH data revealed that RSV was estimated as the most prevalent infectious cause of pneumonia (excluding "other" causes). The npLCM provided a better empirical fit to the data, accurately predicting observed measurement patterns and pairwise associations.
Limitations and Future Work
- Computational Cost: While the Gibbs sampler is designed to be efficient, fitting complex models with many subclasses can still be computationally intensive.
- Expert Knowledge: The model benefits from expert prior knowledge on measurement sensitivities, which might not always be available.
- Single Primary Cause Assumption: The current framework assumes a single primary cause for pneumonia. The authors suggest extending the model to accommodate multiple pathogens causing disease in a single case.
- Covariate Dependence: Future work could involve extending the npLCM to allow etiology distributions and false positive rates to depend on covariates like season, age, and HIV status.
- Supervised Learning: The npLCM can be adapted for supervised learning scenarios where gold-standard measurements are available for some cases, which could further improve inference.
Summary updated