MODELING THE DETERMINANTS OF INDIVIDUAL MEDICAL EXPENDITURES AN ECONOMETRIC AND MACHINE LEARNING ANALYSIS OF THE ROLE OF AGE OBESITY AND SMOKING
Author(s)
Aazelarab BOUGHALEB, MASTER AND PHD CANDIDATE1, Amal Yassine, Jr.2, Omar Maoujoud, PhD, MD3.
1Laboratory of Economic Sciences and Public Policy Ibn Tofail University Kenitra, Morocco, Kenitra, Morocco, 2Médecin and Phd, Moroccan Society for Health Products Economy & ISPOR MAGHREB CHAPTER, Mohammedia, Morocco, 3Societe Marocaine de l'economie de produits de santé, Casablanca, Morocco.
1Laboratory of Economic Sciences and Public Policy Ibn Tofail University Kenitra, Morocco, Kenitra, Morocco, 2Médecin and Phd, Moroccan Society for Health Products Economy & ISPOR MAGHREB CHAPTER, Mohammedia, Morocco, 3Societe Marocaine de l'economie de produits de santé, Casablanca, Morocco.
OBJECTIVES: Rising healthcare expenditures represent a major challenge for health systems,insurers, and policymakers worldwide. While age, obesity, and smoking are widely recognized as important determinants of medical spending, existing studies generally treat obesity and smoking as independent and additive risk factors. This study aims to identify the determinants of individual medical expenditures, examine the potential interaction between obesity and smoking status, and compare the inferential performance of econometric models with the predictive performance of machine learning algorithms.
METHODS: The analysis uses the Medical Cost Personal Dataset, comprising 1,337 insured individuals after data cleaning. A log-linear ordinary least squares (OLS) model was estimated to identify expenditure determinants, followed by an extended specification including a BMI × smoking interaction term. Model diagnostics included multicollinearity, heteroscedasticity, and residual normality tests, with heteroscedasticity-robust standard errors used for inference. Predictive performance was assessed using Random Forest and Extreme Gradient Boosting (XGBoost) models on an 80/20 train-test split. Models were compared using RMSE, MAE, and R2.
RESULTS: Age, body mass index (BMI), smoking status, and number of dependent childrenwere significant positive predictors of medical expenditures. The interaction between BMI and smoking was positive and highly significant (β = 0.0456; p < 0.001), indicating that the marginal effect of BMI on expenditures was substantially greater among smokers than among non-smokers. Random Forest and XGBoost achieved superior predictive performance (R2 = 0.83) compared with the retransformed econometric model (R2 = 0.17-0.25).
CONCLUSIONS: The findings suggest that obesity and smoking act synergistically rather thanadditively in driving healthcare expenditures. Combining econometric and machine learning approaches provides complementary insights by balancing causal interpretation and predictive accuracy. Targeted interventions addressing both smoking and obesity may generate disproportionately large reductions in healthcare spending and should be considered in health policy and prevention strategies
METHODS: The analysis uses the Medical Cost Personal Dataset, comprising 1,337 insured individuals after data cleaning. A log-linear ordinary least squares (OLS) model was estimated to identify expenditure determinants, followed by an extended specification including a BMI × smoking interaction term. Model diagnostics included multicollinearity, heteroscedasticity, and residual normality tests, with heteroscedasticity-robust standard errors used for inference. Predictive performance was assessed using Random Forest and Extreme Gradient Boosting (XGBoost) models on an 80/20 train-test split. Models were compared using RMSE, MAE, and R2.
RESULTS: Age, body mass index (BMI), smoking status, and number of dependent childrenwere significant positive predictors of medical expenditures. The interaction between BMI and smoking was positive and highly significant (β = 0.0456; p < 0.001), indicating that the marginal effect of BMI on expenditures was substantially greater among smokers than among non-smokers. Random Forest and XGBoost achieved superior predictive performance (R2 = 0.83) compared with the retransformed econometric model (R2 = 0.17-0.25).
CONCLUSIONS: The findings suggest that obesity and smoking act synergistically rather thanadditively in driving healthcare expenditures. Combining econometric and machine learning approaches provides complementary insights by balancing causal interpretation and predictive accuracy. Targeted interventions addressing both smoking and obesity may generate disproportionately large reductions in healthcare spending and should be considered in health policy and prevention strategies
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
EE15
Topic
Economic Evaluation, Epidemiology & Public Health, Real World Data & Information Systems
Topic Subcategory
Cost/Cost of Illness/Resource Use Studies
Disease
Diabetes/Endocrine/Metabolic Disorders (including obesity)