Measuring the Impact of Nonignorable Missing Data (ADDED FEE) — Professional Development Continuing Education Course

ASA, Biometrics Section

The popular but typically unverifiable assumption of ignorability greatly simplifies analyses with incomplete data, both conceptually and computationally. We say that missingness is ignorable when the probability that an observation is missing depends only on fully observed information, and nonignorable when the probability that an observation is missing depends on the value of the observation, even after conditioning on available design variables and covariates. For example, in a clinical trial the data are plausibly nonignorably missing when the subjects who drop out are those for whom the drug is either ineffective or excessively toxic. The possibility that the missing observations in a study are the result of a nonignorable mechanism casts doubt on the validity of conclusions based on the assumption of ignorability. Unfortunately, it is generally impossible to robustly assess the validity of this assumption with just the data at hand.
One way to address this problem is to conduct a local sensitivity analysis: Essentially, re-compute estimated parameters of interest under models that slightly violate the assumption of ignorability. If the parameters change only modestly under violation of the assumption, then it is safe to proceed with an ignorable model. If they change
drastically, then a simple ignorable analysis is of questionable validity.
To conduct such a sensitivity analysis in a systematic and efficient way, we have developed a measure that we call the index of local sensitivity to nonignorability (ISNI), which evaluates the rate of change of parameter estimates in the neighborhood of an ignorable model. Computation of ISNI is straightforward and avoids the need to estimate a nonignorable model or to posit a specific magnitude of nonignorability. We have developed a suite of statistical methods for ISNI analysis, now implemented in an R package named isni.
In this half-day short course we will describe these methods and train users to apply them to inform evaluations of the reliability of empirical findings when data are incomplete.

Instructor(s): Daniel Heitjan, Southern Methodist University; Hui Xie, Simon Fraser University