2026-11-25 –, Cullen Room
Estimating the prevalence of smoking and vaping in culturally and linguistically diverse communities using Multilevel Regression and Poststratification
Dinah Lope, Andrew Ward, Emily Baker, Eve Mitsopoulos
Track: Computational social science
Keywords: Small Domain Estimation, Multilevel Regression and Poststratification
Traditional survey estimation techniques perform well when sample sizes are sufficiently large across broad geographic regions or demographic groups. In these settings, response proportions can be estimated reliably using established methods (Valliant et al., 2018). However, these approaches often perform poorly for small geographic areas or population subgroups, where limited sample sizes can lead to unstable, highly variable estimates that may not accurately reflect the target population.
To address this challenge, researchers increasingly use model-based approaches that combine survey data with auxiliary information from sources such as the Australian Bureau of Statistics Census of Population and Housing. This paper presents Multilevel Regression and Poststratification (MRP) as a method for producing small-domain estimates. MRP consists of two key stages (Gelman & Little, 1997; Park et al., 2006; Downes et al., 2018). First, a multilevel regression model is fitted to survey data to estimate the relationship between respondent characteristics and survey outcomes. Second, poststratification adjusts these model-based estimates using the population distribution of those characteristics within each domain. By combining survey responses with population-level demographic data, MRP borrows strength across related groups and domains, producing robust estimates for small populations and geographic areas where direct survey estimates may be unreliable.
We demonstrate this approach using the 2022 Victorian Smoking and Health Survey, a cross-sectional telephone survey of a population-representative sample of Victorians aged 18 years and older. Prevalence was estimated by main language spoken at home, country of birth and year of arrival in Australia.
Our contribution supports the ongoing development of methodologies aimed at improving the accuracy and relevance of population estimates in domains where traditional survey methods are constrained by declining response rates, limited coverage, and the challenges of reaching specific population groups. In particular, the findings can help identify priority CALD communities for targeted smoking cessation initiatives in Victoria – guiding community engagement and service delivery planning to maximise the reach and effectiveness of cessation efforts in populations with the greatest need.
Dinah Lope is a Senior Data Scientist and Statistician at The Social Research Centre. She holds a PhD in Statistics from RMIT University. Her research focuses on the development and application of Bayesian and spatiotemporal models for small-area and domain estimation. By leveraging auxiliary information from multiple data sources, her work seeks to improve the precision and relevance of population estimates in settings where traditional survey methods are constrained by declining participation, limited sample sizes, coverage issues, and growing difficulties in reaching underrepresented or hard-to-survey populations.