The txshift R package is designed to provide facilities for the construction of efficient estimators of a causal parameter defined as the counterfactual mean of an outcome under stochastic mechanisms for treatment assignment (Dı́az and van der Laan 2012). txshift implements and builds upon a simplified algorithm for the targeted maximum likelihood (TML) estimator of such a causal parameter, originally proposed by Dı́az and van der Laan (2018), and makes use of analogous machinery to compute an efficient one-step estimator (Pfanzagl and Wefelmeyer 1985). txshift integrates with the sl3 package (Coyle et al. 2020) to allow for ensemble machine learning to be leveraged in the estimation procedure. For many practical applications (e.g., vaccine efficacy trials), observed data is often subject to a two-phase sampling mechanism (i.e., through the use of a two-stage design). In such cases, efficient estimators (of both varieties) must be augmented to construct unbiased estimates of the population-level causal parameter. Rose and van der Laan (2011) first introduced an augmentation procedure that relies on introducing inverse probability of censoring (IPC) weights directly to an appropriate loss function or to the efficient influence function estimating equation. txshift extends this approach to compute IPC-weighted one-step and TML estimators of the counterfactual mean outcome under a shift stochastic treatment regime. The package is designed to implement the statistical methodology described in Hejazi et al. (2020).