THE DEVELOPMENT, CALIBRATION, AND VALIDATION OF THE US POPULATION HEALTH MICROSIMULATION MODEL

Purpose:  Develop and validate a microsimulation model that forecasts the chronic disease burden for the US population and captures the risk of developing a specific chronic condition as a function of other disease as well as demographic factors, for example the risk of diabetes increases with age and in the presence of high blood pressure. This model can be used to forecast the long term health implications of a variety of public health interventions across multiple conditions.

Methods:  For individuals over 50, we used the Health and Retirement Survey to estimate cox-proportional hazard models for hypertension, diabetes, cancer, COPD, heart disease, stroke and death. We adjusted for time invariant characteristics (sex, race/ethnicity and education) and time-dependent characteristics (health insurance, underweight, obesity, and disease indicators). For individuals under 50, we used the 2008-2011 Medical Expenditure Panel Study (MEPS), to estimate survival models using a discrete time models for hypertension, diabetes, cancer, heart disease, stroke adjusting for the same covariates as used in the over 50 models. For the under 50 discrete time survival model, we use a logistic regression and modelled age as a cubic function. We then combined these estimates into one integrated model. Using this model, we externally validated and calibrated our model using the 2005-2014 National Health Interview Survey (NHIS). We used the NHIS to forecast population level prevalence rates for a cohort of patients in the data in 2005 until 2014. We used a random search algorithm to update our model parameters and determined the optimal parameter set using the weighted mean percentage deviation method.

Results:  Our uncalibrated model overestimated heart disease, stroke, and cancer prevalence for those over 50 and underestimated stroke, high blood pressure and cancer prevalence for those under 50. After calibration our model’s weighted absolute mean percent deviation was 3.95%, indicating that the forecast was quite accurate. Furthermore, our predictions were within the confidence intervals of the NHIS estimates.

Conclusions:  Many chronic conditions interact and modify future risk of developing additional conditions and mortality, alongside demographic characteristics. This model can accurately forecast a population’s disease burden for multiple chronic conditions. We plan further work to extend the model to forecast population level health care spending.