This paper examines the properties of two-way fixed effects (FE) regressions when the assumption of constant treatment effects is violated. It shows that the FE regression coefficient estimates a weighted sum of average treatment effects (ATEs), with weights that may be negative. This can lead to biased estimates of the average treatment on the treated (ATT), as negative weights may cause the coefficient to be negative even when all ATEs are positive. The paper proposes a new estimator, DID_M, which is valid even when treatment effects are heterogeneous across groups or over time. DID_M estimates the ATE across all (g,t) cells where treatment changes from t-1 to t. It relies on common trends assumptions on both potential outcomes and is shown to be asymptotically normal. The paper also discusses the robustness of the FE coefficient to treatment effect heterogeneity, showing that the ratio of the absolute value of the FE coefficient to the standard deviation of the weights can indicate whether the coefficient and ATT may have opposite signs. The paper concludes that the FE coefficient may not be reliable in the presence of heterogeneous treatment effects and that DID_M is a more robust alternative. The paper also extends its results to first-difference regressions and fuzzy designs.This paper examines the properties of two-way fixed effects (FE) regressions when the assumption of constant treatment effects is violated. It shows that the FE regression coefficient estimates a weighted sum of average treatment effects (ATEs), with weights that may be negative. This can lead to biased estimates of the average treatment on the treated (ATT), as negative weights may cause the coefficient to be negative even when all ATEs are positive. The paper proposes a new estimator, DID_M, which is valid even when treatment effects are heterogeneous across groups or over time. DID_M estimates the ATE across all (g,t) cells where treatment changes from t-1 to t. It relies on common trends assumptions on both potential outcomes and is shown to be asymptotically normal. The paper also discusses the robustness of the FE coefficient to treatment effect heterogeneity, showing that the ratio of the absolute value of the FE coefficient to the standard deviation of the weights can indicate whether the coefficient and ATT may have opposite signs. The paper concludes that the FE coefficient may not be reliable in the presence of heterogeneous treatment effects and that DID_M is a more robust alternative. The paper also extends its results to first-difference regressions and fuzzy designs.