This paper studies settings where there is interest in identifying and estimating an average causal effect of a binary treatment on an outcome of interest, under complete randomization or selection on observables assumptions. The outcome does not get immediately realized after treatment assignment, a feature that is ubiquitous in empirical settings, creating a time window in between the treatment and the realization of the outcome. The existence of such a time window, in turn, opens up the possibility of other observed endogenous actions to take place and affect the interpretation of popular parameters, including the average treatment effect. In this context, we study several regression-based estimands that are routinely used in empirical work, and present five results that shed light on how to interpret them in terms of ceteris paribus effects, indirect causal effects, and selection terms. Our three main takeaways are the following. First, the three most popular estimands do not satisfy what we call strong sign preservation, in the sense these estimands may be negative even when the treatment positively affects the outcome for any possible combination of other actions. Second, the by-far most popular estimand that ``controlls'' for the other actions in the regression does not improve upon a simple comparisons of means in the sense that negative weights multiplying relevant ceteris paribus effects become more prevalent. Finally, while non-parametric identification of the effects we study is straightforward under our assumptions and follows from saturated regressions, we also show that linear regressions that correctly control for the other actions by stratifying lead to estimands that always satisfy strong sign preservation.