The accurate quantification of trends in the distribution of risk factors is critical in public health. Reliable estimates are key for planning prevention activities and treatment services, especially in low-income countries where the optimal allocation of limited resources is a priority. However empirical data – usually self-report from population surveys – are often sparse (available for selected subpopulations and time points), heterogeneous (collected with inconsistent methods across data sources), and of varying characteristics in terms of precision and risk of bias. Bayesian meta-regression is an alternative to frequentist approaches to make sense of sub-optimal data by integrating in a principled way additional sources of information and broad epidemiological and biological evidence. We present an application of Bayesian meta-regression to estimate age- and sex-specific trends in alcohol consumption – a major risk factor for cardiovascular and other diseases – in the South African adult population. The model accounts for the censored nature of the consumption data and the ubiquitous under-reporting of alcohol use in surveys. It allows for time and age non-linearity and for complex constraints in the parameter space, derived from biological knowledge and administrative records on alcohol sales. Mild assumptions of smoothness in age and time trends and relationship with auxiliary variables allow the model to make estimates where data are sparse or unreliable. The Bayesian estimator – implemented using Stan Modelling Language and its default NUTS sampling algorithm – accounts for uncertainty beyond sampling error, and the availability of draws from the posterior distribution makes it straightforward to recover estimates of various linear and non-linear functions of the model parameters. We show how this approach compares favourably to classical rescaling methods used to recover estimates of population alcohol consumption from downward-biased survey data.
sssss
