Addressing Item-Level Missing Data: A Comparison of Proration and Full Information Maximum Likelihood Estimation

Area of Science:

• Psychometrics
• Statistical Methods
• Health Services Research

Background:

• Researchers often use prorated scale scores to handle missing item data, averaging available responses.
• Methodological concerns exist regarding proration's assumptions about scale item means and covariances.
• Proration can introduce statistical bias, even when missing data is completely random (MCAR).

Purpose of the Study:

• To empirically investigate the impact of score proration on statistical analyses.
• To introduce and advocate for a full information maximum likelihood (FIML) approach for item-level missing data.
• To demonstrate how FIML mitigates power loss and utilizes available data without altering substantive analyses.

Main Methods:

• Empirical investigation of score proration bias under missing completely at random (MCAR) conditions.
• Description of a full information maximum likelihood (FIML) method for handling item-level missing data.
• Proposed FIML approach treats scale scores as missing if any item is missing, using items as auxiliary variables.

Main Results:

• Proration was found to introduce bias, even under missing completely at random (MCAR) assumptions.
• Simulations indicated that item-level missing data handling via FIML significantly increases statistical power compared to scale-level handling.
• The FIML approach effectively utilizes available item data and mitigates power loss.

Conclusions:

• Researchers should avoid score proration due to its potential for bias.
• Full information maximum likelihood (FIML) offers a statistically sound and powerful method for handling item-level missing data.
• This approach has significant practical implications, particularly when participant recruitment is challenging or costly.