Assessing Alternative Imputation Strategies for Infrequently Missing Items on Multi-item Scales

Area of Science:

• Health Sciences
• Psychometrics
• Biostatistics

Background:

• Health researchers frequently use multi-item scales to measure psychological constructs.
• Missing data on these scales is a common challenge.
• Multiple imputation (MI) is a theoretically motivated approach to handle missing data, unlike ad-hoc methods like mean substitution.

Purpose of the Study:

• To compare the statistical properties of various multiple imputation (MI) implementations against ad-hoc methods for handling missing items on multi-item scales.
• To investigate how item-level vs. scale-level imputation and the use of auxiliary variables impact results.
• To assess the performance of these methods in the context of an HIV study measuring depression and anxiety.

Main Methods:

• Empirical investigation contrasting ad-hoc methods with different MI strategies (item-level vs. scale-level imputation).
• Analysis of how auxiliary variables were incorporated into imputation models.
• Utilized data from an HIV study with multi-item scales for depression and anxiety.

Main Results:

• Findings align with previous research favoring item-level imputation when feasible.
• Observed only subtle differences in statistical properties across the compared methods.
• The weaknesses of ad-hoc procedures may be less pronounced with modest percentages of missing data.

Conclusions:

• Item-level imputation is generally preferred for handling missing data on multi-item scales when feasible.
• The choice of method may have less impact than anticipated when missing data is not extensive.
• Further research may be needed to explore optimal strategies under varying missing data conditions.