A Nonlinear Mixed-Effects Model for Under-Five Mortality Estimation from Biased Data

Chelsea Liddell, University of Washington
Haidong Wang, University of Washington
Christopher J. L. Murray, University of Washington
Laura A. Dwyer-Lindgren, University of Washington
Katherine Lofgren, University of Washington
Alan D. Lopez, University of Queensland
Marie Ng, University of Washington

Accurate estimates of under-five mortality are crucial, and many countries rely on various, potentially biased, data for these estimates. Here we attempt to assess the level of bias and adjust data to an unbiased level. To do this, we fit a mixed-effects nonlinear regression for under five mortality rate, using lag distributed income, maternal education, and HIV crude death rate as covariates, as well as a country- source nested random effect and a fixed effect on source type category. For each country, mixed effects values for an expert-selected reference (unbiased) source are used to adjust the other points to the unbiased level. These adjusted points are used in space-time smoothing and Gaussian process regression to produce a final, unbiased, time series of under-five mortality estimates for each country. The result is an automated algorithm which will produce mortality estimates for future iterations of the Global Burden of Disease study.