An interpretable approach to modeling the impact of an environmental risk factor on correlated neurobehavioral subscales

Abstract

PDS 69: Methods and statistics, Johan Friso Foyer, Floor 1, August 26, 2019, 4:30 PM - 5:30 PM Background/Aim: Neurobehavioral tests are often correlated and tap into overlapping domains, but are commonly analyzed separately. Statistical approaches that aggregate test data could increase power. We introduce easily implementable approaches to assess how an environmental risk factor predicts correlated neurobehavioral subscales, and identify which subscales are most sensitive to the risk factor. We illustrate these approaches by identifying a neurobehavioral signature for arsenic, a known neurotoxicant. Methods: We modeled the prospective association between third trimester arsenic and 19 clinical, adaptive and content t-scored subscales of the Behavioral Assessment System for Children (BASC-2) Parent Rating Scales, assessed at 4 years of age in a birth cohort study (PROGRESS) in Mexico City (n = 425). We used three different approaches to quantify the global neurobehavioral burden using all subscale data: (1) clinical burden (sum-score based on clinical significance cutoffs, [t-score >= 70 or <= 30], adjusting for the subscale’s directionality), (2) sample-based burden (number of subscales in the highest risk quartile); and (3) the Mazziotta-Pareto Index (MPI), a nonlinear composite of continuous subscale data. We used generalized linear models adjusted for child sex, socio-economic status, maternal IQ and maternal age. To identify subscales most sensitive to arsenic, we explore a leave-one-subscale-out step and compare model performance. We also compared results from separately modeled subscales. Results: We detected a positive association between arsenic and clinical burden index (p = 0.0081), as well as with the MPI (p = 0.043) and the sample-based burden score (p = 0.049). However, arsenic was not associated with any BASC-2 individual subscale or composite scale after adjusting for multiple comparisons using the Benjamini-Hochberg method. Conclusions: This easy-to-implement framework could be used to increase power and uncover relationships not seen when subscales are separately modeled, providing a succinct way to account for correlated subscales.