OBJECTIVES: To compare the propensity score (PS) matching approaches with PS covariate adjustment using zero-inflated Poisson (ZIP) models for estimating log(RR) and log(OR) from zero-inflated count data in observational studies. METHODS:  We compare three PS approaches: matching with/without replacement (WiR,WoR) and the PS covariate adjustment (CoA) with ZIP models in terms of bias, mean squared error (MSE), empirical standard error (ESE) and coverage probability of the 95% Wald confidence interval (CP) for estimating log(RR) and log(OR) under three simulated scenarios: weak, moderate, and strong associations between baseline covariates and the outcome variable. RESULTS: Simulation results show that CoA has the smallest bias of log(RR) regardless of the association between the baseline covariates and the outcome variable. However, WiR has the smallest bias of log(OR) when the associations between the baseline covariates and the outcome variable are moderate and strong. Similar pattern is also observed for the MSE. CoA has the smallest MSE of log(RR) regardless the associations, and in addition, it also has the smallest MSE of log(OR) when associations are weak and moderate. For ESE of the estimates, CoA and WoR have the smallest ESE of log(RR) and log(OR), regardless of the associations between the baseline covariates and the outcome variable. CONCLUSIONS: Although zero-inflated count data are commonly seen in practice, traditional ZIP models may not be able to be applied directly if the zero-inflated count data are collected from observational studies. Thus, in this paper, we compare the performance of three various approaches, WiR, WoR, and CoA, for estimating log(RR) and log(OR) when the associations between the baseline covariates and the outcome variable are weak, moderate, and strong in terms of bias, MSE, ESE, and CP through a Monte Carlo simulation study. Simulation results show that overall CoA has a better performance compared to WiR and WoR in estimating log(RR) and log(OR).