Informed healthcare policy decisions must be driven by consideration of an intervention's effectiveness as well as its cost. Cost-effectiveness analyses provide a framework for decision making that balances these joint outcomes in some optimal way. However, because these studies often use data from observational sources, results may be biased due to unmeasured or time-varying confounding, informative cost censoring, and skewed or zero-inflated data. The goals of this dissertation are two-fold; we aim to (1) elucidate the conditions under which causal conclusions can be drawn from cost-effectiveness data, and (2) develop novel statistical methods for identifying cost-effective treatments while accounting for confounding and other data irregularities. We discuss three such developments: regression methodology for a novel probabilistic measure of cost-effectiveness, interpretable Q-learning based methods for identifying cost-effective treatment strategies, and a flexible and efficient influence function based estimator of average treatment cost that is robust to unmeasured confounding given a valid instrumental variable. We evaluate the operating characteristics of our proposed methods under several realistic data scenarios through simulation studies. We also illustrate usage by identifying cost-effective adjuvant treatments for early-stage endometrial cancer patients as well as assessing differences in costs between surgical and non-surgical interventions for gallstones and hemorrhaging using observational data.