|
Abstract:
|
Computer-coded-verbal-autopsy (CCVA) algorithms used to generate burden-of-disease estimates rely on non-local training data and yield inaccurate estimates in local context. We present a general calibration framework to improve estimates of cause-specific-mortality-fractions from CCVA when limited local training data is available. We formulate a Bayesian hierarchical local calibration of discrete classifiers that updates a non-locally trained CCVA estimate using estimates of the misclassification rates of the CCVA algorithm on the local training data. This involves a novel transition matrix shrinkage for the misclassification matrix which theoretically guarantees that, in absence of any local data or when the CCVA algorithm is perfect, the calibrated estimate coincides with its uncalibrated analog, thereby subsuming the default practice as a special case. A novel Gibbs sampler using data augmentation enables fast implementation. We also present an ensemble calibration using predictions from multiple CCVA algorithms as inputs to produce a unified estimate. A theoretical result demonstrates how the ensemble calibration favors the most accurate algorithm. Simulation and real data anal
|
