On Item Response Theory models in allostatic load scoring

To the Editors

We compliment [1] who used Item Response Theory (IRT), a statistical technique widely employed in educational and psychological assessments, to evaluate allostatic load (AL). This allows an advancement in the field compared with the traditional AL index scoring method: dichotomizing scores for various AL biomarkers based on clinically- or empirically-determined cutoffs and then adding one point for each marker that exceed these thresholds. The use of this approach in coming studies, however, should consider the following three issues.

First, although Liu et al.’s supplementary material includes a script for the Confirmatory Factor Analysis (CFA) used to investigate the dimensionality of AL, model fit indices were not described (e.g., root mean square error of approximation, comparative fit index, etc.). This is needed to indicate the model is not misspecified [2], an essential step to verify whether IRT assumptions are met [3]: local independence of the items and the unidimensionality of the latent trait.

Second, despite the authors’ claim that item parameter scores are robust even when the unidimensionality assumption is violated, model fits must still be checked if IRT is to be used to build a common scale across different datasets. This allows any arising misfits/misspecifications to be tackled to propose psychometrically sound models [4]. Understanding why Liu et al. considered AL as a unidimensional construct when it is usually hypothesized as being multidimensional would also aid in this respect.

Third, when performing CFA it is fundamental to choose the proper estimator according to variables’ levels of measurement. In the example provided in Liu et al.’s supplementary file, it seems Maximum Likelihood (ML) estimator was used, which is adequate for scalar and not dichotomous items. To avoid erroneous conclusions, Weighted Least Squares Mean and Variance (WLSMV) should be used for dichotomous data [5]. R package lavaan can automatically select the best estimator if the type of data (dichotomous, scalar) is reported correctly.

We hope our suggestions assist the development of studies designed to explore AL indicators with latent model frameworks. Researchers might also consider using biomarkers’ original scalar units of measurement rather than the traditional point-based composite score, which will likely reflect more AL variance.

Funding

This study was supported by Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP – Process 2016/14750-0, 2017/02816-9), Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq – #301899/2019-3 and 302511/2018-0, and Associação Fundo de Incentivo à Pesquisa (AFIP).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.