On the use of the not-at-random fully conditional specification (NARFCS) procedure in practice

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Missing data imputation is crucial in statistical analysis.
• The not-at-random fully conditional specification (NARFCS) procedure handles missing data under missing-not-at-random (MNAR) conditions.
• Eliciting sensitivity parameters for NARFCS is challenging due to their conditional nature, potentially leading to inconsistent imputations.

Purpose of the Study:

• To clarify the importance of correctly conditioning NARFCS sensitivity parameters.
• To develop procedures for calibrating these parameters using simpler models.
• To provide guidance on including missingness indicators in NARFCS imputation models.

Main Methods:

• Calibration of NARFCS sensitivity parameters by relating them to parameters from pattern mixture models.
• Development of algorithms to implement the calibration procedure.
• Inclusion of missingness indicators as part of the imputation models, recommending their inclusion in each model by default.

Main Results:

• Demonstrated a method to calibrate NARFCS sensitivity parameters, making them more accessible.
• Showcased the practical application of the calibration procedure using real-world data (Avon Longitudinal Study of Parents and Children).
• Validated the approach through simulation studies, confirming its utility and robustness.

Conclusions:

• Correctly conditioned sensitivity parameters are vital for accurate NARFCS imputations under MNAR conditions.
• The proposed calibration method simplifies parameter elicitation and improves imputation consistency.
• Including all missingness indicators by default in NARFCS models is recommended practice.