Generalized additive selection models for the analysis of studies with potentially nonignorable missing outcome data

Area of Science:

• Statistics
• Biostatistics
• Econometrics

Background:

• Missing outcome data can bias study results, especially when missingness depends on unobserved factors.
• Previous methods (Rotnitzky et al., 1998) addressed this for longitudinal data using parametric selection models.
• Sensitivity analysis is crucial for assessing the impact of potential biases due to missing data.

Purpose of the Study:

• To extend existing sensitivity analysis methodologies for nonignorable missing data.
• To develop semiparametric selection models for cross-sectional, univariate outcome data.
• To propose robust estimation methods for the mean outcome in the presence of missing data.

Main Methods:

• Utilized generalized additive restrictions to define semiparametric selection models.
• Developed a backfitting algorithm for estimating parameters in the generalized additive selection model.
• Proposed three types of estimating functions for mean outcome: inverse weighted, doubly robust, and orthogonal.

Main Results:

• The proposed backfitting algorithm effectively estimates parameters for the generalized additive selection model.
• The data analysis and simulation study demonstrated the performance of the proposed methods.
• Doubly robust and orthogonal estimating functions showed promise for robust mean outcome estimation.

Conclusions:

• The developed semiparametric approach offers a flexible extension for sensitivity analysis of missing data in cross-sectional studies.
• The proposed estimation methods provide robust alternatives for handling nonignorable missingness.
• This work contributes to more reliable statistical inference when dealing with incomplete outcome data.