At a Glance
Assessment of Methods in Population Health
Kieron Barclay, Maarten Jacob Bijlsma, Christina Bohk-Ewald (MPIDR / University of Helsinki, Finland), Roland Rau, Pekka Martikainen (MPIDR / University of Helsinki, Finland); in Collaboration with Nikkil Sudharsanan (Heidelberg University, Germany)
One key objective of the population health sciences is to understand why one social group has different levels of health and well-being compared to another. A number of methods have been developed to answer these types of research questions, such as Oaxaca-Blinder, Arriaga, Kitagawa, and line-integral decomposition. While these decomposition approaches produce results that have clear mathematical interpretations, they often have ambiguous causal interpretations, i.e., they are not mapped to specific counterfactual scenarios.
We develop and demonstrate an implementation of the Jackson and Vander Weele counterfactual decomposition that uses parametric models and Monte Carlo integration of the g-formula. This approach produces a clearly interpretable decomposition result and has increased statistical flexibility compared to other approaches. Compared to established demographic decompositions, this approach allows for the adjustment of confounding variables without incurring dimensionality problems. Compared to individual-level data-based decompositions, this approach allows for the decomposition of any contrast of any aggregate population measure. This substantially extends the decomposition analyses to common population health measures such as life expectancies, prevalence ratios, and medians.
The Monte Carlo integration of the g-formula was also extremely useful for other approaches to understanding health, such as in the study of unemployment and depression. Depression is a leading cause of disability in both men and women, and a major public health concern globally. Depression and unemployment are risk factors for one another; such interdependencies (more technically known as intermediate and time-varying confounding) are difficult to unravel using established methodologies. A g-formula approach allowed us to do so, but we still faced the issue of unmeasured time-constant confounding; some individuals are more likely to both be unemployed and suffer from depression due to more time-stable factors that were unmeasured, such as personality traits. Through Monte Carlo integration, we were able to combine the g-formula with individual-level fixed-effects intercepts, which additionally allowed us to adjust for such unmeasured time-stable factors. Combining these two approaches to causal inference allows us to adjust for reverse causality, time-varying confounding, and unobserved selection in a more flexible, generalizable, and robust way than ever before.
Data and Surveys, Health Care, Public Health, Medicine, and Epidemiology, Projections and Forecasting, Statistics and Mathematics
Epidemiology 30:3, 388–395. (2019)
MPIDR Working Paper WP-2019-004. (2019)
International Journal of Environmental Research and Public Health 15:10, 1–11. (2018)
Demography 54:2, 721–743. (2017)
Demography 54:4, 1559–1577. (2017)
BMC Medical Research Methodology 17:68, 1–17. (2017)