BAYESIAN NESTED LATENT CLASS MODELS FOR CAUSE-OF-DEATH ASSIGNMENT USING VERBAL AUTOPSIES ACROSS MULTIPLE DOMAINS.

Area of Science:

• Public Health
• Biostatistics
• Epidemiology

Background:

• Accurate cause-specific mortality rates are vital for global health monitoring and interventions.
• Two-thirds of global deaths lack an assigned cause, hindering public health efforts.
• Verbal autopsy (VA) is used in low- and middle-income countries to determine causes of death but faces challenges with data distribution shifts.

Purpose of the Study:

• To propose a novel latent class model framework for VA data (LCVA).
• To address the challenge of assigning causes of death when training and target populations differ.
• To estimate cause-specific mortality fractions for new populations using existing VA data.

Main Methods:

• Developed a latent class model framework (LCVA) to jointly model VA data from multiple domains.
• Introduced a parsimonious representation of symptom distribution using nested latent class models.
• Created a computationally efficient algorithm for posterior inference.

Main Results:

• LCVA demonstrates superior predictive performance compared to existing methods.
• The proposed framework shows improved scalability for analyzing large VA datasets.
• LCVA effectively assigns causes of death for out-of-domain observations.

Conclusions:

• LCVA offers a robust solution for cause-specific mortality estimation from VA data.
• The method overcomes limitations of traditional algorithms vulnerable to distribution shifts.
• LCVA enhances the accuracy and scalability of global health surveillance through improved cause-of-death assignment.