Population Health


Assessment of Methods in Population Health

Maarten Jacob Bijlsma, Karen van Hedel, Mathias Lerch, Christina Bohk-Ewald, Timothy Riffe

Detailed description:

APC methods are commonly applied in demography and related fields to decompose an outcome into constituent effects associated with, or even caused by, age (e.g., effects associated with aging), period (effects associated with calendar time), and cohort (effects associated with the formative experiences of a generation). However, the application of APC methods is troubled by the APC linear identification problem (A = P – C). Through a simulation study mimicking APC data on cardiovascular mortality, we tested and demonstrated possible pitfalls that users of the mechanism-based approach to APC analysis may encounter under realistic conditions. We found that the approach structurally misestimates the effects when the set of intermediate variables is incomplete, which biases the results when an intermediate variable is affected by more than one APC variable (and this feature is not acknowledged in the analysis), and when unaccounted confounding is present between intermediate variables and the outcome. Based on these findings, we provided guidelines for minimizing bias for users of the mechanism-based approach. Further, we extended the method to allow it to be used with any fully parametric model for the outcome and intermediate variables by approximating the estimation of the APC parameters through Monte Carlo integration, and we provided R-code demonstrating this extension. This extended method thus helps users apply the method with Poisson regression (for count and rate outcomes) and logistic regression, both of which are commonly used in the social and health sciences.

Over the past three decades, the prevalence of obesity has risen tremendously across the globe, to the point that it is now considered a global pandemic. In quantifying the health burden of obesity at the population level, the population attributable fraction (PAF) is commonly used. Many methodologies for estimating obesity-attributable mortality fractions exist. We assessed the impact of using different techniques to estimate both the levels of and trends in obesity-attributable mortality for the Netherlands, 1981-2013. We used data from Statistics Netherlands, the Human Mortality Database, and the WHO mortality data and applied three oft-used approaches to these data: the partially adjusted method, the weighted sum method, and the comparative risk assessment approach (CRA). All of these approaches estimated an increase in the obesity-attributable mortality rate over the period 1981-2013, with minor exceptions, but with (strongly) varying levels and trends. Our findings were in line with previous research showing that PAF estimates vary widely according to the method used. Overall, the weighted-sum method seems to be the most appropriate method to use because it deals with confounding and effect modification better than the other methods we examined. Taken together, our findings indicate that the obesity epidemic is accelerating in the Netherlands, which contradicts findings from the influential Global Burden of Disease (GBD) study that obesity-related mortality has been declining. We suggest that the GBD’s conclusions are driven by its application of indirect estimation of BMI values, whereas the methods tested in our study use direct estimation.

Research keywords: Statistics and Mathematics; Projections and Forecasting; Health Care, Public Health, Medicine, and Epidemiology; Data and Surveys


Bijlsma, M. J.; Wilson, B.; Tarkiainen, L.; Myrskylä, M.; Martikainen, P.:
Epidemiology 30:3, 388-395 (2019).
Sudharsanan, N.; Bijlsma, M. J.:
MPIDR Working Paper WP-2019-004. (2019).
Bijlsma, M. J.; Daniel, R. M.; Janssen, F.; De Stavola, B. L.:
Demography 54:2, 721-743 (2017).
Bohk-Ewald, C.; Ebeling, M.; Rau, R.:
Demography 54:4, 1559-1577 (2017).
Hu, Y.; van Lenthe, F. J.; Hoffmann, R.; van Hedel, K.; Mackenbach, J. P.:
BMC Medical Research Methodology 17:68 (2017).