Propensity scores in the presence of effect modification: A case study using the comparison of mortality on hemodialysis versus peritoneal dialysis
1 Program for the Assessment of Radiological Technology (ART Program), Department of Epidemiology & Biostatistics and the Department of Radiology, Erasmus University Medical Center Rotterdam, Dr. Molewaterplein 50, 3015 GE Rotterdam, the Netherlands
2 Department of Internal Medicine, Erasmus University Medical Center Rotterdam, Dr. Molewaterplein 50, 3015 GE Rotterdam, the Netherlands
3 Division of Clinical Decision Making, Informatics and Telemedicine, Department of Medicine, Tufts Medical Center, Tufts University School of Medicine, 800 Washington Street, #302, Boston, MA 02111, USA
4 Department of Health Policy and Management, Harvard School of Public Health, Kresge 4th Floor, 677 Huntington Avenue, Boston, Massachusetts 02115, USA
5 Dutch End Stage Renal Disease Registry RENINE, Postbus 2304, 2301 CH Leiden, The Netherlands
6 Division of Nephrology, Department of Medicine, Stanford University School of Medicine, 780 Welch Road, Suite 106, Palo Alto, CA 94304, USA
Emerging Themes in Epidemiology 2010, 7:1 doi:10.1186/1742-7622-7-1Published: 11 May 2010
To control for confounding bias from non-random treatment assignment in observational data, both traditional multivariable models and more recently propensity score approaches have been applied. Our aim was to compare a propensity score-stratified model with a traditional multivariable-adjusted model, specifically in estimating survival of hemodialysis (HD) versus peritoneal dialysis (PD) patients.
Using the Dutch End-Stage Renal Disease Registry, we constructed a propensity score, predicting PD assignment from age, gender, primary renal disease, center of dialysis, and year of first renal replacement therapy. We developed two Cox proportional hazards regression models to estimate survival on PD relative to HD, a propensity score-stratified model stratifying on the propensity score and a multivariable-adjusted model, and tested several interaction terms in both models.
The propensity score performed well: it showed a reasonable fit, had a good c-statistic, calibrated well and balanced the covariates. The main-effects multivariable-adjusted model and the propensity score-stratified univariable Cox model resulted in similar relative mortality risk estimates of PD compared with HD (0.99 and 0.97, respectively) with fewer significant covariates in the propensity model. After introducing the missing interaction variables for effect modification in both models, the mortality risk estimates for both main effects and interactions remained comparable, but the propensity score model had nearly as many covariates because of the additional interaction variables.
Although the propensity score performed well, it did not alter the treatment effect in the outcome model and lost its advantage of parsimony in the presence of effect modification.