Model Calibration \(\rightarrow\) Calibration Overview
Many processes in a woman’s reproductive life are unobserved, and those that are observed are surrounded with much uncertainty. Model calibration offers a way to systematically explore the uncertainty around each input parameter, and generate parameter sets that fit well to observed data. In this way, we can reflect the uncertainty that surrounds each model input by presenting a range of plausible (calibrated) estimates.
We used a directed search algorithm (simulated annealing) in which we reduced the variance of the proposal distributions following an exponential cooling schedule. This allowed us to more widely explore the parameter space near the beginning of each search chain and then narrow the proposed jumps around the best set found so far, allowing us to fine-tune the calibrated parameters. Parameter sets were scored based on the squared distance between the model predictions and the observed data. For targets for which confidence intervals were available, we re-weighted the targets by the inverse of the width of the 95% CIs to allow more precise estimates to have more influence in the calibration scoring.
We performed iterative rounds of search chains, using the best sets of hyperparameters found from the previous round of calibration for computational efficiency.
GMatH (Global Maternal Health) Model - Last updated: 28 November 2022
© Copyright 2020-2022 Zachary J. Ward