A true score imputation method to account for psychometric measurement error.

Abstract

Scores on self-report questionnaires are often used in statistical models without accounting for measurement error, leading to bias in estimates related to those variables. While measurement error corrections exist, their broad application is limited by their simplicity (e.g., Spearman's correction for attenuation), which complicates their inclusion in specialized analyses, or complexity (e.g., latent variable modeling), which necessitates large sample sizes and can limit the analytic options available. To address these limitations, a flexible multiple imputation-based approach, called true score imputation, is described, which can accommodate a broad class of statistical models. By augmenting copies of the original dataset with sets of plausible true scores, the resulting set of datasets can be analyzed using widely available multiple imputation methodology, yielding point estimates and confidence intervals calculated with respect to the estimated true score. A simulation study demonstrates that the method yields a large reduction in bias compared to treating scores as measured without error, and a real-world data example is further used to illustrate the benefit of the method. An R package implements the proposed method via a custom imputation function for an existing, commonly used multiple imputation library (mice), allowing true score imputation to be used alongside multiple imputation for missing data, yielding a unified framework for accounting for both missing data and measurement error. (PsycInfo Database Record (c) 2023 APA, all rights reserved).