The purpose of regressions is to obtain conditional means (or medians), in part so that we don't draw incorrect inferences from spurious correlations. Thus regressions are especially useful when we see a significant bivariate correlation, and then the significance goes away after conditioning on other variables.
But what if the sign on an explanatory variable flips from significant and positive before conditioning to significant and negative after conditioning on other variables? Should we take that negative sign seriously? If the model is specified properly, the answer is yes, but my gut tells me such models are rarely specified properly.
No comments:
Post a Comment