Abstract
BACKGROUND: Liquid bead microarray antibody (LBMA) assays are used to assess pathogen-cancer associations. However, studies analyze LBMA data differently, limiting comparability. METHODS: We generated 10,000 Monte Carlo-type simulations of log-normal antibody distributions (exposure) with 200 cases and 200 controls (outcome). We estimated type I error rates, statistical power, and bias associated with t-tests, logistic regression with a linear exposure and with the exposure dichotomized at 200 units, 400 units, the mean among controls plus two standard deviations, and the value corresponding to the optimal sensitivity and specificity. We also applied these models, and data visualizations (kernel density plots, receiver operating characteristic (ROC) curves, predicted probability plots, and Q-Q plots), to two empirical datasets to assess the consistency of the exposure-outcome relationship. RESULTS: All strategies had acceptable type I error rates (0.03 ≤ P ≤ 0.048), except for the dichotomization according to optimal sensitivity and specificity, which had a type I error rate of 0.27. Among the remaining methods, logistic regression with a linear predictor (Power=1.00) and t-tests (Power=1.00) had the highest power to detect a mean difference of 1.0 MFI (median fluorescence intensity) on the log scale and were unbiased. Dichotomization methods upwardly biased the risk estimates. CONCLUSION: These results indicate that logistic regression with linear predictors and unpaired t-tests are superior to logistic regression with dichotomized predictors for assessing disease associations with LBMA data. Logistic regression with continuous linear predictors and t-tests are preferable to commonly used LBMA dichotomization methods.