Date Published: November 15, 2018
Publisher: Public Library of Science
Author(s): René Kooiker, Hendriek C. Boshuizen, Brecht Devleesschauwer.
Micro-simulation models of risk-factors and chronic diseases are built increasingly often, and each model starts with an initial population. Constructing such populations when no survey data covering all variables are available is no trivial task, often requiring complex methods based on several (untested) assumptions. In this paper, we propose a method for evaluating the merits of construction methods, and apply this to one specific method: the construction method used in the DYNAMO-HIA model.
The initial population constructed using the DYNAMO-HIA method is compared to another population constructed by starting a simulation with only newborns and simulating the course taken by one risk-factor and several diseases. In this simulation, the age- and sex-specific prevalence of the risk-factor is kept constant over time.
Our simulations show that, in general, the DYNAMO-HIA method clearly outperforms a method that assumes independence of the risk-factor and the prevalence of diseases and independence between all diseases. In many situations the DYNAMO-HIA method performs reasonably well, but in some the proportion with the risk-factor for those with a disease is under- or overestimated by as much as 10 percentage points. For determining comorbidity between diseases linked by a common causal disease or a common risk-factor it also performs reasonably well. However, the current method performs poorly for determining the comorbidity between one disease caused by the other.
The DYNAMO-HIA methods perform reasonably well; they outperform a baseline assumption of independence between the risk-factor and diseases in the initial population. The method for determining the comorbidity between diseases that are causally linked needs improvement. Given the existing discrepancies for situations with high relative risks, however, developing more elaborate methods based on running simulation models to generate an initial population would be worthwhile.
Demographic population health models can be used to model the effects of risk-factors on the transitions between health states (e.g. being healthy, diseased, or disabled), on their respective prevalence rates and durations, and on the population measures these health states affect, such as the life expectancy, structure, and size of the population. These models can be applied in Health Impact Assessment (HIA) in order to support decision-making in health policy, since they quantify population outcomes of policy interventions that influence risk-factor prevalence (e.g. the proportion of smokers) or transition rates between risk-factor states (e.g. start/stop rates in smoking).
Fig 1 gives a schematic overview of the modeling process and the proposed validation method for the initial population.
Fig 2 shows the proportion in each risk-factor category (never, current, and former smokers) in the entire population (bottom right) and in populations with a particular disease (other plots).
Our simulations show that, in general, the approximation method for constructing an initial population from marginal data implemented in DYNAMO-HIA clearly outperforms assuming independence. In many situations it performs reasonably well, but in some the proportion of the risk-factor is under- or overestimated by as much as 10 percentage points. For determining comorbidity between diseases that are linked by a common causal disease or common risk-factor the method also performs reasonably well. The method performs poorly in determining the comorbidity between a causal disease and the disease it causes.
DYNAMO-HIA uses the following algorithm to simulate the initial population (adapted from Boshuizen et al. 2012 ):
DYNAMO-HIA applies an epidemiological model that defines transition rates from being without a disease to having the disease (incidence rate) and from being alive to death (mortality rate). The epidemiological model of DYNAMO-HIA parameterizes two update rules: one using incidence rates and one using mortality rates. These are described below (adapted from Boshuizen et al. 2012 ):