I am not sure if this is the right venue for the question, but I'll give it a try. I am estimating the effect of a policy change. The policy affected industries within states different across time. So my variation is coming from industry, state, and calendar year. When I run the regressions I estimate five models. The first model has controls and all one-way fixed effects (state, industry, calendar year). In the second through fourth model in addition to controls and one-way fixed effects, I include two-way fixed effects (state-year, state-industry, industry-state) one at a time. In the fifth model I include controls, all one-way fixed effects and all two-way fixed effects. In all models the coefficients are significant and have the same sign. However, whenever the two-way state-year fixed effects are included (second and fifth model), the estimated coefficients blow up and are much higher than in the remaining models. This happens no matter how I define my independent variable (normalized or non normalized) or what outcome variable (dummy or continuous) I use. Has anyone and idea why this might be happening?