OBJECTIVES: High-dimensional propensity score (PS) methods have been used in health care claims data to improve control of confounding by adjusting for a large number of covariates that may be proxies for unobserved factors.  We have previously shown that PS models are biased for non-linear link functions when confounders were included. We conducted a simulation study to understand whether inclusion of covariates that are not confounders may also bias the association by estimating Monte Carlo mean bias, relative efficiency (RE) and coverage probability (CP) of log odds ratios when covariates only related to the exposure or only related to the outcome were included. METHODS: We conducted 1000 Monte Carlo simulations, and estimated effect of exposure using logistic regression models.  The propensity score was included in the logistic model as a linear predictor or as a smoothed covariate using restricted cubic splines. Simulations were conducted for scenarios including 5, 15, and 25 covariates.  RESULTS: Using the PS with 25 covariates related only to the binary exposure, Monte Carlo bias, standard error (SE), RE and CP were -0.002, 0.015, 1.34, and 0.94 when the PS was included as a smoothed covariate, and –0.002, 0.015, 1.31, and 0.94 when the PS was included as a linear covariate. The bias, SE, RE and CP for 25 covariates related to the binary outcome were 0.307, 0.096, 21.6, and 0 when the PS was included as a linear covariate.  Bias tended to increase with more covariates.   CONCLUSIONS: We observed minimal bias when using PS models where covariates were related only to the exposure, but substantial bias when the covariates were related to the outcome.  PS models may not be appropriate for logistic models because these models do not adequately deal with errors in the outcome due to the covariate.