Estimation of parameters and missing values under a regression model with non-normally distributed and non-randomly...

Area of Science:

• Statistics
• Data Analysis
• Computational Statistics

Background:

• Missing data is a common challenge in statistical analysis, potentially biasing results.
• Various algorithms exist for handling missing data, each with potential strengths and weaknesses.
• Understanding algorithm performance under different data conditions is crucial for reliable analysis.

Purpose of the Study:

• To compare the performance of three algorithms: complete cases, ALLVALUE, and expectation maximization (EM).
• To evaluate their effectiveness in estimating regression parameters and imputing missing values.
• To assess performance across varying missing data percentages, distributions, patterns, and correlational structures.

Main Methods:

• A simulation study was conducted.
• Three algorithms (complete cases, ALLVALUE, expectation maximization) were compared.
• Performance was evaluated under diverse conditions of missing data (5% and 25%), distributions (normal, mixture of normals, lognormal), patterns (random, related, censored), and correlations.

Main Results:

• At 5% missing data, EM and complete cases performed equally well, irrespective of correlational structure.
• At 25% missing data, EM generally provided the best parameter estimation, while complete cases offered a conservative approach.
• ALLVALUE was competitive only with weak correlations or minimal missing data; EM excelled in imputation, even with censored or log-normal data.

Conclusions:

• The expectation maximization (EM) algorithm is recommended for robust parameter estimation when missing data increases.
• The complete cases method serves as a reliable, conservative option.
• The EM algorithm demonstrates optimal performance for imputing missing values across various challenging data scenarios.