Date of this Version
Psychoneuroendocrinology 129 (2021) 105274 https://doi.org/10.1016/j.psyneuen.2021.105274
Left censoring in salivary bioscience data occurs when salivary analyte determinations fall below the lower limit of an assay’s measurement range. Conventional statistical approaches for addressing censored values (i.e., recoding as missing, substituting or extrapolating values) may introduce systematic bias. While specialized censored data statistical approaches (i.e., Maximum Likelihood Estimation, Regression on Ordered Statistics, Kaplan-Meier, and general Tobit regression) are available, these methods are rarely implemented in biobehavioral studies that examine salivary biomeasures, and their application to salivary data analysis may be hindered by their sensitivity to skewed data distributions, outliers, and sample size. This study compares descriptive statistics, correlation coefficients, and regression parameter estimates generated via conventional and specialized censored data approaches using salivary C-reactive protein data. We assess differences in statistical estimates across approach and across two levels of censoring (9% and 15%) and examine the sensitivity of our results to sample size. Overall, findings were similar across conventional and censored data approaches, but the implementation of specialized censored data approaches was more efficient (i.e., required little manipulations to the raw analyte data) and appropriate. Based on our review of the findings, we outline preliminary recommendations to enable investigators to more efficiently and effectively reduce statistical bias when working with left-censored salivary biomeasure data.