TL;DR
This paper introduces a novel Bayesian factor model to improve the accuracy of determining causes of death from verbal autopsy (VA) data, particularly by accounting for complex symptom dependencies. This method offers a more nuanced understanding of disease epidemiology in settings where death registration is poor, outperforming existing methods in key metrics.
What Problem Does This Paper Address?
Globally, a significant portion of deaths (around 60%) lack proper registration or a determined cause, hindering public health efforts. Verbal autopsy (VA) surveys, which collect information on symptoms and medical history from relatives of the deceased, are a crucial tool in low-resource settings. However, existing VA analysis methods often assume symptoms are independent given a cause, which is unrealistic and can lead to inaccurate cause-of-death estimations. This paper tackles the challenge of modeling these symptom dependencies to improve the probabilistic assessment of causes of death from VA data.
Methods
The authors propose a Bayesian factor model for VA data. The core idea is to model the observed binary symptoms as manifestations of underlying continuous latent variables. These latent variables are then modeled using a multivariate probit model conditional on the cause of death. Crucially, to handle the high dimensionality of VA questionnaires and avoid overfitting, they introduce a K-dimensional factor model where K is much smaller than the number of symptoms. This allows for a sparse covariance structure, effectively capturing symptom dependencies through a few latent factors. The model also naturally handles missing data, a common issue in VA surveys, by integrating over missing symptoms under a missing-at-random assumption. They developed an efficient Markov Chain Monte Carlo (MCMC) algorithm for posterior computation. Additionally, they propose a method to quantify the importance of each symptom for specific causes of death using a model-based version of Cramér's V.
Datasets
The primary dataset used is the Population Health Metrics Research Consortium (PHMRC) labeled data. This dataset includes verbal autopsy (VA) data along with medically certified causes of death from six study sites across four countries: Andhra Pradesh, India; Bohol, Philippines; Dar es Salaam, Tanzania; Mexico City, Mexico; Pemba Island, Tanzania; and Uttar Pradesh, India. This dataset is valuable because it provides a "gold standard" for evaluating VA analysis methods.
Key Insights and Ideas
- Symptom Dependencies Matter: The paper strongly suggests that ignoring symptom dependencies leads to suboptimal cause-of-death assignment. Their factor model explicitly captures these relationships, which are often missed by simpler methods.
- Latent Factors as Summarizers: The use of a small number of latent factors is an elegant way to reduce the complexity of symptom interactions, making the model more parsimonious and interpretable while still capturing important dependencies.
- Handling Missing Data Gracefully: The Bayesian framework's ability to integrate over missing data without explicit imputation is a significant advantage, especially given the high rates of missingness in VA surveys.
- Quantifying Symptom Importance: The proposed measure of symptom-cause association (a model-based Cramér's V) is a valuable tool for understanding which questions are most informative, potentially leading to shorter and more efficient future VA questionnaires.
Key Results
- The proposed Bayesian factor model consistently outperformed several state-of-the-art VA classification methods (including conditionally independent models, InSilicoVA, InterVA, King-Lu, and Tariff) in terms of Cause-Specific Mortality Fraction (CSMF) accuracy.
- In scenarios with significant differences between training and testing data distributions (indicated by a large L1 distance), the proposed method still performed well, often better than competitors.
- The method demonstrated robustness across different scenarios and sensitivity analyses for prior distributions and Monte Carlo simulation settings.
- The analysis identified specific symptoms (e.g., related to medical history like cancer and TB) that show high dependence with certain causes of death, offering insights for questionnaire design.
Limitations and Future Work
- Spatial Information: The authors suggest incorporating spatial information into the model, as factors affecting causes of death can vary geographically, and nearby areas might share similar patterns.
- Survey Weights: Generalizing the framework to incorporate survey weights is identified as a future direction, which is crucial for handling data from stratified or biased sampling designs.
- Interpretation of Factors: While the factors are used for dimensionality reduction, their direct interpretation is not a primary goal, and constraints for identification were deliberately avoided to prevent computational inefficiencies. Future work could explore methods for more direct factor interpretation.
Summary updated