We study online inference for synthetic control, which is valid even if we test after every new post-treatment observation. Specifically, we introduce three tests that are ‘anytime-valid’ under different assumptions on the DGP and the treatment allocation mechanism.
While our theoretical results rely on strong exchangeability assumptions for exact validity, Monte Carlo simulations suggest the methods work well in more realistic settings.
We provide an empirical example through an analysis of the California Tobacco Control programme.
