Comparison Of Propensity Score Analysis Techniques And Application In Health Research Of The Elderly

objectives:To compare the ability of propensity score analysis techniques based on four machine learning algorithms to balance covariates and estimate real intervention effects so as to control the confounding factors and reduce the sample selection bias in the China Health and Retirement Longitudinal Study?CHARLS?data.And in order to understand the changes in quality of life of the elderly as well as provide a reasonable and scientific reference for the formulation of relevant policies about the gradual delay of retirement in China by exploring the potential effect of retirement on health of the elderly with this method.Methods:1 Simulation Study:Based on the selection on observables assumption,Hypothetical studies of varying sample sizes?n=500,2000?with a binary intervention,continuous outcome,and six covariates were simulated under scenarios differing by degree of non-linear and non-additive associations between covariates and the intervention as well as different correlation strengths betweenand.This paper evaluated the performance of propensity score analysis techniques?nearest neighbor matching,optimal matching,genetic matching?based on four machine learning algorithms?classification and regression trees,random forests,generalized boosted model,neural networks?using simulated data.Performance metrics included covariate balance,relative bias,standard error,mean square error and coverage rate of 95%confidence interval.2 Case Study:This method was applied to observational data after simulation.It could control the confounding factors in the CHARLS data and reduce the sample selection bias so as to explore the potential effect of retirement on health of the elderly.Results:1 Results of Simulation Study?1?Compared with the other three machine learning algorithms,GBM had the best ability to balance covariates regardless of the complexity of the intervention-selection model.Generally speaking,better covariate balance led to the lower bias,and genetic matching could achieve better covariate balance,a smallest bias and mean square error.In other words,genetic matching technology with GBM performed best across the scenarios.?2?Although the average absolute distance between matched pairs is lower,the relative performance between optimal matching and nearest neighbor matching changed little.2 Results of Case StudyBefore matching,75%of the covariates were unbalanced between retired and employed groups with the largest ASMD of 1.150 and average ASAM of 0.318.The balance of baseline characteristics was improved after propensity score matching.The largest ASMD and average ASAM for nearest neighbor matching with GBM were 0.226 and 0.071 respectively,which decreased by 80.34%and 77.67%.And there was no significant difference in other covariates between two groups except education level?P>0.05?.In contrast,genetic matching with GBM had a maximum ASMD of0.104 and average ASAM of 0.057,which decreased by 90.52%and 82.08%respectively,in addition there was no statistical difference in all covariates between two groups?P>0.05?.Genetic matching performed better than nearest neighbor matching in balancing covariates.The results showed that the retired group had a rate of self-reported body pains of 0.57?95%CI:0.46⁰.71?times that of the employed group,and the retired group had a 1.27-point?95%CI:1.04¹.52?improvement in mental state scores and a 1.75-point?95%CI:-2.26^-1.24?decrease in depression scores compared to the employed group.and these differences were statistically significant?P<0.001?,while the other health indicators had no significant difference?P>0.05?.The results of further comparison between two groups in terms of gender were basically consistent with that of the whole sample.Conclusions:?1?The real shape between the intervention-selection and observed covariates is usually unknown,so nonparametric methods had superior performance in modeling propensity scores.The combination of GBM and genetic matching has a good result in reducing bias and covariate balance,which is not only a technical attempt to be popularized,but also helps to systematically select methods with better performance rather than the most familiar or easy to implement methods.As any new combined method is based on the selection on observables,it is impossible to be perfect.Therefore,more verification researches based on real data are needed to determine the performance of the new method.?2?Overall,retirement can exert a beneficial effect on health of the elderly.Retirement may be an appropriate time for the elderly to carry out health promotion activities and support healthy aging,which can help to curb the increasing cost of health care.Raising retirement ages may be harmful to the overall health status of the elderly.Therefore,the government and relevant departments should take full account of the potential effect when instituting these retirement delaying policies,and intervention programs focused on health of retired people need to be actively developed and widely implemented.