Matching is a common approach to reduce bias in observed covariates to draw reliable causal inferences in observational studies. This thesis consists of three papers discussing new methods for conducting, evaluating, and improving matching designs in observational studies. The first paper presents new optimal matching techniques for large-scale observational data. This new method reduces the computational complexity and preserves appealing properties in terms of balancing covariates. After constructing a matched sample, it is essential to assess the covariate balance of the matched data since lack of balance in covariates can induce a bias of the estimated treatment effect. The second paper discusses a formal evaluation of covariate balance. This new assessment evaluates whether the match is adequate compared to randomized experiments and identifies the major problems, guiding how to improve the covariate balance. If diagnostics suggest that the current match is not satisfactory, how can we improve the quality of matched samples? The final paper utilizes the idea of directional penalties, which can improve covariate balance in a matched sample effectively, even for a large observational study.