USING A SHORTEST DISTANCE METHOD TO MATCH MICROSIMULATIONS TO RISK FACTOR CORRELATIONS IN CROSS-SECTIONAL DATA

Purpose: When longitudinal data are not available, alternate techniques must be used to inform microsimulation models that capture the relationship between a disease and multiple risk factors.  The distribution of a simulated cohort's risk factors can be approximated using repeated waves of cross-sectional data, but it is unclear how each simulated individual's risk factors should change over time. Previously published methods allow the simulated cohort's risk factor distributions to match the data but do not preserve the cross-sectional correlation observed between multiple risk factors.  Accurate modeling of cross-sectional correlation between risk factors may be important for prediction of diseases that depend on multiple risk factors.

Methods: We develop a shortest distance algorithm for modeling individual changes in risk-factors that preserves both the cohort distribution of each risk factor as well as the cross-sectional correlation between risk factors observed in repeated cross-sectional data. We demonstrate this method in a microsimulation of 1000 thirty-year-old white males that models multiple risk factors (i.e., total cholesterol, high-density lipoprotein, blood pressure, etc.) matched to NHANES data.

We contrast the performance of our method with an alternative published method where individuals' risk factors are only allowed to change within their quantile rank. We use each method to infer correlation and rates of change in risk factors for our model. To demonstrate clinical relevance, we use each method to predict the cohort's atherosclerotic cardiovascular disease (ASVCD) 10-year risk at age seventy.  Because ASCVD risk depends on multiple risk factors, each method's fidelity to the observed risk factor correlation may have important implications for prediction of cardiovascular risk and consequent management.

Results: We show that the correlation between risk factors is better preserved using our shortest distance method compared to the quantile method (Figure 1 panel a).  The sum of squared correlation differences from the NHANES data over all risk factors was smaller using the shortest distance method (0.11 vs 8.87).  This can have clinical implications, as shown in Figure 1 panel b, where the shortest distance method is better able to match the shape of the ASCVD distribution in the NHANES data.

Conclusions: When using microsimulation to predict health conditions which depend on multiple risk factors, modelers should consider using shortest distance methods that preserve the correlation between risk factors.

Figure 1