Description
Model the chance that each value is missing as depending on the value itself, fit it jointly with the substantive model, and compare it with a missing-at-random assumption.
Data are nonignorable missing when the probability of a value being missing depends on that value, as when dissatisfied users skip satisfaction items or struggling users drop out of a diary study. Multiple imputation and full-information ML assume missing at random and can be biased then. Lee's approach defines a missingness indicator for every item and models each by logistic regression on the item values and optionally the latent variables, assuming indicators are conditionally independent given the latent variables. Missing values, latent variables and parameters are sampled jointly by Gibbs and Metropolis-Hastings steps. Bayes factors compare a MAR model with several nonignorable mechanisms, then substantive models are compared under the chosen one. Because the mechanism's form can never be verified from the data, and complex mechanisms become unidentifiable, the method is best used as a sensitivity analysis showing how conclusions change under plausible departures from MAR.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define a missingness indicator for every item or wave.
- Specify a simple logistic model for each indicator on the item values and, optionally, the latent variables, keeping it identified.
- Augment the data with missing values and latent variables and sample all unknowns jointly.
- Fit a MAR version and several nonignorable mechanisms and compare them by Bayes factor.
- Compare substantive models under the selected mechanism.
- Report how key estimates change between MAR and nonignorable assumptions, framing the result as sensitivity analysis rather than proof.
Sources
- Structural Equation Modeling: A Bayesian Approach (Lee, 2007)
- The EM Algorithm and Extensions (McLachlan & Krishnan, 1997)
- Applied Bayesian Modelling (Congdon, 2003)
- Bayesian Statistical Modelling (Congdon, 2006)
Also known as
- MNAR modelling
- Missing not at random
- Selection model for missing data
- Nonignorable missingness mechanism
- MNAR modelling in SEM
Categories
- Analytics & Behaviour
- Statistics & Modelling