An Investigation of Factored Regression Missing Data Methods for Multilevel Models with Cross-Level Interactions

Area of Science:

• Statistics
• Statistical Modeling
• Data Science

Background:

• Missing data methods often factorize joint distributions.
• Limited research exists on factored regressions for multilevel models with interactive effects.

Purpose of the Study:

• Investigate Bayesian and multiple imputation strategies using factored regressions for multilevel models.
• Assess the performance of these methods under various distributional assumptions.

Main Methods:

• Monte Carlo computer simulations were employed.
• Bayesian and multiple imputation techniques based on factored regressions were simulated.
• The study examined performance under satisfied and misspecified distributional assumptions.

Main Results:

• Factored regression imputation methods generally yield unbiased estimates and good coverage when distributional assumptions are met.
• Severe misspecifications, particularly with non-normal distributions, can lead to biased estimates and poor coverage.
• A Yeo-Johnson transformation demonstrated potential in mitigating these biases.

Conclusions:

• Factored regression imputation is a viable strategy for multilevel models, provided distributional assumptions are reasonably met.
• Data transformations, such as Yeo-Johnson, can improve the robustness of these methods when dealing with non-normal data.
• Further research is recommended to explore the nuances of these methods in complex statistical modeling scenarios.