Simulation Study Of Generalized Propensity Score Estimation Methods Under Different Confounding Structures And Its Application

Objective:By constructing exposure model and outcome model with different confounding structures,this study will compares three generalized propensity score(GPS)estimation methods for continuous exposure factor: GPS-OLS,GPS-Boosting,and The Covariate Balancing Generalized Propensity Score(CBGPS)methods to balance the effect of confounding factors and the estimation accuracy of exposure.This study will apply GPS to the field of nutritional epidemiology.Three GPS estimation methods are used to estimate GPS,with the aim of balancing confounding effects and finding a true association between the red meats consumption and dyslipidemia.Methods:In this study,Monte Carlo simulation was used to generate two cases of large samples(N = 1000)and small samples(N = 400).GPS models or outcome models with 16 different confounding structures were constructed,and weights were constructed by inverse variance weighting.This study compared the changes in the correlation coefficients between the covariates and the exposure after weighting to determine the ability of the three GPS estimation methods to balance the confounding effect,and compare the degree of bias and the mean square error to determine the effect on the estimation accuracy of the exposure effect.Results:The simulation results show that in the 16 scenes where different confounding structures exist,the GPS estimated by CBGPS method balanced the confounding effect is the best after constructing weights based on the inverse variance weighting method,which is better than GPS-OLS and GPS-Boosting methods.The estimation accuracy of exposure effect,the CBGPS method can also significantly reduce the mean square error and bias of the exposure effect estimation.The estimation accuracy is better than the GPS-OLS and GPS-Boosting methods,while the simulation results in two samples are same basically.A total of 3 919 study subjects were included in this study,of which 1 815(46.31%)were subjects with dyslipidemia.The results of multivariate analysis showed that there was no significant association between the red meats consumption and dyslipidemia after adjusting the confounding factors through the stepwise regression method(OR = 1.0004,95% CI:-0.0002-0.0011).The CBGPS method was used to balance the confounding effect,and the correlation coefficient between each covariate and the red meats consumption was close to zero.The consumption of red meats in the diet was significantly associated with higher risk of dyslipidemia(OR = 1.0007,95% CI: 0.0001-0.0013).After the GPS-OLS(OR = 1.0013,95% CI: 0.0007-0.0019)and GPS-Boosting methods(OR = 1.0011,95% CI:0.0005-0.0018)were used to balance the confounding effect,increasing the consumption of the red meats will increase the risk of dyslipidemia.Conclusion:In the scenes where varies confounding structures exist,the GPS estimated by the CBGPS method,weighted by inverse variance,has the best effect of balance the confounding factors,which is significantly better than the GPS-OLS and GPS-Boosting methods.So this study suggested that when applying GPS to balance confounding effect,it is best to choose the CBGPS estimation method.This study found that an increase consumption of red meats in daily diet associated with the higher risk of dyslipidemia.