Strategies for analyzing multiple inflammatory biomarkers: Impacts on the strength and replicability of findings in
Phoebe H Lam1, Gregory E Miller2
1Department of Psychology, Carnegie Mellon University, Pittsburgh, PA, USA.
Abstract:
Circulating biomarkers of low-grade inflammation are increasingly used in behavioral medicine and related disciplines as subclinical indicators of health problems. However, strategies for analyzing these biomarkers vary widely, and it is unclear whether such analytical variability affects the assumed inflammation-health association. Using the Midlife in The United States (MIDUS) cohorts as large case examples, we examined whether three analytical strategies - (1) using raw vs. log-transformed data; (2) analyzing biomarkers individually vs. forming composites; and (3) if the latter, how they were constructed- affected the magnitude and replicability of inflammation-health associations. The Core (N = 1,201) and Refresher (N = 709) cohorts included data on 8 inflammatory biomarkers and 8 health problems. A subset of Core participants (N = 643) had longitudinal data about 9 years later. Three composites were constructed: reflective (via factor analysis), formative (via PLS-SEM), and agnostic (standardized mean scores) scales using both raw and log-transformed data. These variations yielded 24 operationalizations of inflammation (2 transformation methods × 12 scoring approaches [8 individual biomarkers + 4 composites]). We then evaluated each operationalization based on (a) its cross-sectional and longitudinal associations with health outcomes, in terms of statistical significance and effect size; and (b) the replicability of the cross-sectional associations across the two samples, in terms of significance agreement and effect size consistency. In MIDUS, operationalizations based on log-transformed (vs. raw) biomarker data yielded stronger and more replicable associations with health outcomes, and so did composite-based operationalizations (vs. individual biomarkers). Patterns were evident in cross-sectional models, and to a lesser extent, in longitudinal models. These findings are not intended to prescribe a single best-practice for analyzing inflammatory biomarkers, but rather to illustrate how analytical decisions can shape estimated inflammation-health links, and thus, how well these biomarkers function as subclinical disease indicators within a given sample.