Use of Biostatistical Models to Manage Replicate Error in Concussion Biomarker Research

Advancing research on fluid biomarkers associated with sport-related concussion (SRC) highlights the importance of detecting low concentrations using ultrasensitive platforms. However, common statistical practices may overlook replicate errors and specimen exclusion, emphasizing the need to explore robust modeling approaches that consider all available replicate data for comprehensive understanding of sample variation and statistical inferences.

To evaluate the impact of replicate error and different biostatistical modeling approaches on SRC biomarker interpretation.

This cross-sectional study within the Surveillance in High Schools to Reduce the Risk of Concussions and Their Consequences study used data from healthy youth athletes (ages 11-18 years) collected from 3 sites across Canada between September 2019 and November 2021. Data were analyzed from November 2022 to February 2023.

Demographic variables included age, sex, and self-reported history of previous concussion.

Outcomes of interest were preinjury plasma glial fibrillary acidic protein (GFAP), ubiquitin C-terminal hydrolase-L1 (UCH-L1), neurofilament-light (NFL), total tau (t-tau) and phosphorylated-tau-181 (p-tau-181) assayed in duplicate. Bland-Altman analysis determined the 95% limits of agreement (LOAs) for each biomarker. The impact of replicate error was explored using 3 biostatistical modeling approaches assessing the associations of age, sex, and previous concussion on biomarker concentrations: multilevel regression using all available replicate data, single-level regression using the means of replicate data, and single-level regression with replicate means, excluding specimens demonstrating more than 20% coefficient variation (CV).

The sample included 149 healthy youth athletes (78 [52%] male; mean [SD] age, 15.74 [1.41] years; 51 participants [34%] reporting ≥1 previous concussions). Wide 95% LOAs were observed for GFAP (-17.74 to 18.20 pg/mL), UCH-L1 (-13.80 to 14.77 pg/mL), and t-tau (65.27% to 150.03%). GFAP and UCH-L1 were significantly associated with sex in multilevel regression (GFAP: effect size, 15.65%; β = -0.17; 95% CI, -0.30 to -0.04]; P = .02; UCH-L1: effect size, 17.24%; β = -0.19; 95% CI, -0.36 to -0.02]; P = .03) and single-level regression using the means of replicate data (GFAP: effect size, 15.56%; β = -0.17; 95% CI, -0.30 to -0.03]; P = .02; UCH-L1: effect size, 18.02%; β = -0.20; 95% CI, -0.37 to -0.03]; P = .02); however, there was no association for UCH-L1 after excluding specimens demonstrating more than 20% CV. Excluding specimens demonstrating more than 20% CV resulted in decreased differences associated with sex in GFAP (effect size, 12.29%; β = -0.14; 95% CI, -0.273 to -0.004]; P = .04) and increased sex differences in UCH-L1 (effect size, 23.59%; β = -0.27; 95% CI, -0.55 to 0.01]; P = .06), with the widest 95% CIs (ie, least precision) found in UCH-L1.

In this cross-sectional study of healthy youth athletes, varying levels of agreement between SRC biomarker technical replicates suggested that means of measurements may not optimize precision for population values. Multilevel regression modeling demonstrated how incorporating all available biomarker data could capture replicate variation, avoiding challenges associated with means and percentage of CV exclusion thresholds to produce more representative estimates of association.