9.2 - Difference-in-Differences Overview

[@JP-zz4xx]

I am use to looking for a step function change in the trend after the treatment, therefore the change in slope pre and post treatment signify a measure of the ATE. Why doesn't that work? Also, is it advisable for the treatment data slope and the control data slope be roughly the same? I am thinking it implies the two groups have the same composition with no bias where if the slopes differ drastically one of the groups may be bias, e.g. drunk people. If you have a trend where you have considerable trend data prior to the treatment time couldn't you use those data points as your control group and work out the same type of math to DiD but based on pre-treatment slope?

[@BradyNealCausalInference]

"I am use to looking for a step function change in the trend after the treatment, therefore the change in slope pre and post treatment signify a measure of the ATE. Why doesn't that work?" I believe this also works, under the parallel trends assumption (assumption that treatment and control groups have the same slope). However, consider the case where the control group curve also jumps at the time the treatment group is getting treatment. There, having the control group allows you to detect that there might be something other than the treatment that is effecting things (the control group too). "Also, is it advisable for the treatment data slope and the control data slope be roughly the same?" Yes, that's what the parallel trends assumption assumes, if you're talking about lines. "If you have a trend where you have considerable trend data prior to the treatment time couldn't you use those data points as your control group and work out the same type of math to DiD but based on pre-treatment slope?" Yes, I believe so. Different assumptions.