Thanks for prompt response. I am not sure if I am following. As a concrete example, say \mu ~ N(0,1), observed Y | \mu ~ N(\mu, 1). Inference would be done based on p(\mu | Y) while Y has no missing.
If Y has missing (denote missing part as Y_m and non-missing as Y_nm. Would pymc3 do p(\mu | Y_nm) and p(Y_m | Y_nm)? And how to understand the likelihood (Y_m, Y_nm) | \mu ~ N(\mu, 1)? Don’t we need Y_nm | Y_m, \mu instead as likelihood?