Difference in differences analyses
Department of Public Health Sciences, School of Medicine, University of California, Davis
Many approaches to causal inference assume exchangeability and exploit its consequence:
\[{\mathbb{E}\mathopen{}\left[Y_{x} \mid X = x'\right]\mathclose{}} = {\mathbb{E}\mathopen{}\left[Y_{x} \mid X = x\right]\mathclose{}}\]
Write \(Y_{0,t}\) for the potential outcome at time \(t\) without treatment. Difference-in-differences makes a weaker exchangeability assumption:
\[{\mathbb{E}\mathopen{}\left[Y_{0,t} - Y_{0,t'} \mid X = 1\right]\mathclose{}} = {\mathbb{E}\mathopen{}\left[Y_{0,t} - Y_{0,t'} \mid X = 0\right]\mathclose{}}\]