A Simulation Study For Propensity Score In Dealing With Collinearity Data

Objective The aim of our study was to find whether propensity score (PS) method is better in dealing with collinearity data, compared with common logistic regression and try to find some factors (including sample size, positive proportion of outcome variable and degree of collinearity) which can influence parameter estimations in PS regression.Methods A logistic model was used as a standard model in our study since it was designed to deal with non-collinearity data. Monte Carlo simulation was employed to compare parameter estimates between PS regression and common logistic regression in dealing with collinearity data under different conditions of different sample size, positive proportion of outcome variable and degree of collinearity. Also the mutual effects among three conditions mentioned above were detected. A real data from the hypertension study in Nantong was analyzed by use of PS regression to confirm our conclusions of simulation study.Results (1) Given4%as a positive proportion of outcome variable and0.92as a correlation coefficient between covariates and exposure factor the parameter estimates, either regression coefficient or its standard error from PS regression, are close to parameters estimated from standard regression model, compared to the common logistic regression. These different estimates are gradually disappeared along with increase of sample size.(2) Given sample size of1000and500and positive proportion of outcome variable of4%, we estimated regression coefficient and is standard error from three models along with degree of collinearity. The trend of parameters estimated from PS regression is parallel with the trend of standard model. It means the difference between these two models is consistent. However, the changes of regression coefficient and standard error from the common logistic regression are parallel with changes of two models mentioned above where r is in a low level and turn its direction at r=0.5(n=1000) and r=0.3(n=500).(3) Given a sample size of1000and correlation between covariates and exposure factors as0.92, the parameter estimates of PS regression are also close to those from the standard model, compared with the common logistic regression. But difference of parameter estimate between PS regression and common logistic regression is smaller when positive proportion of outcome variable increases.Conclusions Our data suggests that the parameters estimated from PS regression are more reliable than the common logistic regression, especially under the conditions of small sample size, low positive proportion of outcome variable and data with collinearity. Therefore, PS regression could be one of excellent methods in dealing with collinearity data.