看看woodridge的计量经济学的书

同问……

if you suppose a treatment has x1 and x2 variant effect, then you should use interaction term to control for such heterogeneity. As OLS are mean regression, higher powered terms are better to separate the mixed effects of certain treatment.

You don't see if there's any sort of linearity by drawing the data, especially when you have hundreds of explanatory variables. Always go with your hypothesis first.

第三版的第6、7章