Measuring under-five mortality and fertility through mobile phone surveys: an assessment of selection bias in 34 low- and middle-income countries

In countries lacking comprehensive systems of civil registration and vital statistics, the collection of demographic data through mobile surveys is increasingly feasible, as mobile phone coverage has rapidly expanded. However, mobile phone users are often younger, male, with higher education, and wealthier, potentially biasing demographic estimates based on phone surveys. We assess selection bias in estimates of under-five mortality and fertility using data from 34 Demographic and Health Surveys. We use logistic regressions to characterize individuals who own or have access to a mobile phone in terms of sociodemographic background characteristics, and Poisson regressions to estimate the association between mobile phone ownership (or access)and total fertility and under-five mortality rates. We further evaluate the potential reasons why estimates by mobile phone ownership may differ using a set of behavioral characteristics. Fertility and mortality rates would be biased downwards if estimates are only based on women who own or have access to mobile phones. Under-five mortality estimates can be adjusted through post-stratification using age, educational level, area of residence, wealth, and marital status as weights. However, fertility estimates remain biased even after adjusting for these covariates. This difference is associated with behavioral factors that are not captured by the post-stratification variables, but for which there are also differences between mobile phone owners and non-owners. In many countries, modern contraceptive use is positively associated with mobile phone ownership, which may complicate fertility estimates based on mobile phone surveys. Mobile phone surveys need to include background questions on socioeconomic status to be able to weight and adjust mortality estimates ex post-facto. Fertility estimates from mobile phone surveys will be biased unless further research uncovers the mechanisms driving the bias.