To assess the impact of home-delivered meals offered by Meals on Wheels (MOW) programs, a dataset that includes both enrollment into MOW and healthcare utilization is required. Due to confidentiality restrictions, this data is dispersed over MOW internal systems and Medicare claims. Probabilistic linking algorithms may erroneously link MOW recipients with Medicare beneficiaries, which can result in biased treatment effect estimates and suboptimal interval estimates. We propose a two-stage multiple imputation framework to estimate causal effects when the covariate and outcome information are stored in a separate data source from the treatment assignment. In the first stage, we create multiple datasets in which MOW recipients are linked to Medicare beneficiaries using a Bayesian record linkage technique. In the second stage, Medicare beneficiaries who were not enrolled in MOW are matched to those who were, and the unobserved healthcare utilization for each MOW recipient had they not received MOW is multiply imputed. This procedure propagates the error in the linking and matching processes, and can be used to estimate effects of interventions in other linked datasets.