Maximum likelihood estimation with missing outcomes: From simplicity to complexity.

Area of Science:

• Biostatistics
• Clinical Research Methodology
• Data Analysis

Background:

• Clinical and prevention studies frequently encounter missing or censored outcome data.
• Accurate estimation with incomplete data is crucial for reliable study findings.
• Maximum Likelihood (ML) methods present a robust framework for handling such data.

Purpose of the Study:

• To review and categorize various Maximum Likelihood (ML) methods for handling missing or censored outcomes.
• To describe the implementation complexity of different ML approaches based on data type and missing-data mechanisms.

Main Methods:

• Categorization of ML methods based on complexity and applicability to ignorable (Missing At Random) and nonignorable missing-data mechanisms.
• Description of simple ML methods: complete-case analysis, covariate adjustment, survival analysis, and propensity-to-be-missing scores.
• Overview of complex ML methods: longitudinal dropout analysis (marginal/conditional models) and perfect fit analysis for categorical data.
• Introduction to composite linear models for categorical data with ignorable or nonignorable mechanisms.

Main Results:

• ML methods provide a conceptually straightforward approach to estimation with partially missing outcomes.
• Implementation difficulty ranges from easy to moderate for most ML methods, with composite linear models being the most complex.
• Specific methods are suitable for different data types (continuous, categorical) and missing-data assumptions (ignorable vs. nonignorable).

Conclusions:

• A spectrum of ML methods exists to address missing and censored data in clinical research.
• The choice of ML method depends on the specific characteristics of the data and the nature of the missingness.
• Understanding the complexity and applicability of these methods aids researchers in selecting appropriate analytical strategies.