Objective: Properly handling missing data is a challenge, especially when working with older populations that have high levels of morbidity and mortality. We illustrate methods for understanding whether missing values are ignorable and describe implications of their use in regression modeling.
Study design and setting: The use of missingness screens such as Little's missing completely at random "MCAR test" (1988) and the "Index of Sensitivity to Nonignorability (ISNI)" by Troxel and colleagues (2004)introduces complications for regression modeling, and, particularly, for risk factor selection. In a case study of older patients with simulated missing values for a delirium outcome set in a 14-bed medical intensive care unit, we outline a model fitting process that incorporates the use of missingness screens, controls for collinearity, and selects variables based on model fit.
Results: The proposed model fitting process identifies more actual risk factors for ICU delirium than does a complete case analysis.
Conclusion: Use of imputation and other methods for handling missing data assist in the identification of risk factors. They do so accurately only when correct assumptions are made about the nature of missing data. Missingness screens enable researchers to investigate these assumptions.