Abstract

The total number of deaths of children aged below 5 years per every 1,000 live births provide a picture about the health of the population, and policy directions for the health sector of a country. Although this needs regular estimation, it is affected by inaccurate death registers in some parts of the world, such as the sub-Saharan Africa. This makes the use of national surveys as a convenient approach for determining the under-five mortality rate for countries in the region. But, the survey design effect is often disregarded by researchers in such applications. If regression methods are used for prediction of under-five mortality, not much is known about the effect of ignoring the sample design in the estimates. This chapter concerns with estimating and comparing the bias a researcher commits when using unweighted and weighted logistic regression methods to predict under-five mortality rate in Malawi through national survey. We used data from 2004, 2010, and 2015–16 Malawi demographic and health surveys, as well as UNICEF annual mortality monitoring records. The survey weights were considered at two stages: during model fitting and when computing overall predicted mortality rate. The computations were done using R version 3.6.3 and Stata version 12.0. The results showed that there was higher accuracy in estimation when the weights were applied during calculation of overall predicted probability of child death given a fitted logit model, than during model fitting. We recommend incorporating survey cluster weights when computing the overall predicted probability of event, without regard of the weights during model fitting, for binary data models whose goal is prediction of event probability.

Original language	en
Pages (from-to)	257-279
Publication status	Published - 2024