This paper proposes an empirical likelihood inference method for a general framework that covers various types of treatment effect parameters in regression discontinuity designs (RDD) . Our method can be applied for standard sharp and fuzzy RDDs, RDDs with categorical outcomes, augmented sharp and fuzzy RDDs with covariates and testing problems that involve multiple RDD treatment effect parameters. Our method is based on the first-order conditions from local polynomial fitting and avoids explicit asymptotic variance estimation. We investigate both firstorder and second-order asymptotic properties and derive the coverage optimal bandwidth which minimizes the leading term in the coverage error expansion. In some cases, the coverage optimal bandwidth has a simple explicit form, which the Wald-type inference method usually lacks. We also find that Bartlett corrected empirical likelihood inference further improves the coverage accuracy. Easily implementable coverage optimal bandwidth selector and Bartlett correction are proposed for practical use. We conduct Monte Carlo simulations to assess finite-sample performance of our method and also apply it to two real datasets to illustrate its usefulness.