Multilevel regression and poststratification as a modelling approach for estimating descriptive population parameters from highly selected survey samples in large-scale health studies

Abstract

Recruiting a representative sample of participants is becoming increasingly difficult in large-scale population health and epidemiological surveys, even for studies using a well-documented sampling frame and a sound sampling process. In view of this and with the increasing appeal of online recruitment, due to significantly lower cost and rapid accrual, many researchers undertaking health surveys are faced with the challenge of analysing data obtained from a sample that is not representative of the target population of interest.

Statistical methods for appropriately addressing this selection or participation bias are therefore critical to ensuring reliable and accurate population inference from large-scale complex health surveys. This is particularly important as results of such surveys often influence health care decision making and policy development.

Historically, inverse-probability weighting using survey sampling weights has been the standard method for adjusting for known or expected discrepancies between sample and population when estimating descriptive population parameters in complex health surveys.

A recently developed model-based approach is multilevel regression and poststratification (MRP). MRP was first described in the context of political polling and social research in the US. MRP first uses multilevel regression to model individual survey responses for the outcome measure of interest as a function of individual-level demographic and area-level geographic covariates. The resulting estimates of the target parameter for each demographic-geographic respondent subtype are then combined using a weighted average across the subtypes (poststratification cells), weighting by the proportions of each subtype in the actual population, to produce an overall population-level estimate.

The research of this PhD investigated the use of MRP for producing valid and accurate inference for descriptive population parameters in large-scale population health and epidemiological surveys where samples may not be representative of the target population of interest. This approach was evaluated in comparison to inverse-probability weighting using a combination of simulation experiments and a case-study analysis of data from the baseline wave of Ten to Men: The Australian Longitudinal Study on Male Health.

MRP was consistently found to achieve greatly superior precision as well as increased uniformity of estimates across population subsets relative to inverse-probability weighting. In simulation studies, while sampling weights produced estimates with smaller bias on average, the reduced variance associated with MRP was shown to result in estimates that were more often closer to the true population parameter values.

As well as establishing MRP as a valuable analytic approach for large-scale health surveys where samples may not be representative of the target population of interest, this research explored the practical challenges associated with the application of MRP to real survey data and provided a number of recommendations to support future applications of MRP to health-related outcomes.