Recoverability of target estimands under MNAR missingness

Determine whether and how a specified target estimand can be recovered (consistently estimated) from the observed-data distribution under missing-not-at-random (MNAR) mechanisms, namely when the conditional distribution of missingness indicators depends on partially observed or latent variables, and specify the conditions under which consistent estimation is possible.

Background

The paper discusses that while Missing At Random (MAR) is a common assumption enabling standard methods such as multiple imputation or inverse probability weighting, MAR is only sufficient and not necessary for unbiased estimation. When data are Missing Not At Random (MNAR), recoverability of target estimands is not guaranteed and must be assessed case-by-case using graphical models for missingness (m-DAGs).

This uncertainty motivates the development and application of m-DAGs to evaluate whether estimands can be recovered given a correct causal missingness model. The authors note the lack of general algorithms for recoverability, underscoring the need to establish concrete conditions and procedures for consistent estimation under MNAR.

References

Moreover, even though MAR is the weakest known condition under which the missingness process can be ignored (i.e., dealt with using the observed data), it is only a sufficient, but not a necessary condition for unbiased estimation; this means that under a missing not at random (MNAR) scenario, it is unclear if and how a target estimand can be recovered (estimated consistently).

— Recoverability of Causal Effects under Presence of Missing Data: a Longitudinal Case Study  (2402.14562 - Holovchak et al., 2024) in Section 1 (Introduction)