Ignoring Non-ignorable Missingness

Area of Science:

• Statistics
• Biostatistics
• Data Analysis

Background:

• The missing at random (MAR) assumption is a cornerstone of statistical inference for incomplete datasets.
• Misinterpretations of MAR can lead to unnecessarily complex analytical requirements.
• Standard estimators are often avoided when data missingness is presumed non-ignorable.

Purpose of the Study:

• To demonstrate that the classical missing at random (MAR) assumption is not always necessary for valid statistical inference.
• To identify and explain strategies for handling non-ignorable missing data using standard estimation methods.
• To clarify common misunderstandings surrounding the MAR assumption.

Main Methods:

• Reviewing the theoretical underpinnings of missing data assumptions, particularly MAR.
• Proposing and detailing three distinct strategies to address non-ignorable missingness: conditioning on variables, data discarding, and parameter protection.
• Illustrating how these strategies permit the use of standard statistical estimators.

Main Results:

• The classical missing at random (MAR) assumption is frequently not required for valid inference when ignoring the missingness mechanism.
• Several assumptions often mistakenly linked to MAR are also unnecessary.
• Three practical strategies enable the use of standard estimators even with data typically deemed non-ignorably missing.

Conclusions:

• Valid statistical inference can be achieved without adhering to the strict missing at random (MAR) assumption in many scenarios.
• Understanding and applying specific strategies can simplify the analysis of incomplete data.
• Researchers can confidently use standard estimators by employing methods like conditioning on variables, data discarding, or parameter protection.