Sensitivity Analysis

In the body of the report, we focus our discussion of sensitivity analyses on sensitivity to the assumption about the underlying mechanism producing the missing values. There are other aspects of a statistical model for which sensitivity should be assessed. Here is an outline of the steps leading to a comprehensive sensitivity analysis for such models:

  1. Presumed mechanisms of missing data: steps would include identification of data likely to be missing, speculation on mechanisms leading to that missing data, and specification of analyses of missing data patterns.

  2. Planned analyses to deal with missing data: presumed model assumes either missing completely at random, missing at random, or missing not at random (as defined in Chapter 3); the population with available data that will be used (e.g., complete cases, all available data, etc.); the variables that will be used; how variables will be modeled; distributional assumptions; the statistical model; and the statistical paradigm (Bayesian, frequentist, likelihood).

  3. Sensitivity analyses: one will need (a) a framework for exploring effect of distributional assumptions, (b) a framework for exploring effect of variable modeling (e.g., linear, dichotomized, interactions), (c) a framework for exploring effect of considering other variables, (d) a framework for exploring effect of changing population used for modeling, (e) a framework for exploring effect of assumptions of missing at random or missing not at random, and, finally, (f) possible augmented data collection that can shed light on assumptions.

The National Academies of Sciences, Engineering, and Medicine
500 Fifth St. N.W. | Washington, D.C. 20001

Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement