Difference- and regression-based approaches for detection of bias

Objective: This simulation study was undertaken to assess the statistical performance of six commonly used rejection criteria for bias detection. 

Methods: The false rejection rate (i.e. rejection in the absence of simulated bias) and the probability of bias detection were assessed for the following: difference in measurements for individual sample pair, the mean of the paired differences, t-statistics (paired t-test), slope < 0.9 or > 1.1, intercept > 50% of the lower limit of measurement range, and coefficient of determination (R²) > 0.95. The linear regressions evaluated were ordinary least squares, weighted least squares and Passing-Bablok regressions. A bias detection rate of < 50% and false rejection rates of >10% are considered unacceptable for the purpose of this study. 

Results: Rejection criteria based on regression slope, intercept and paired difference (10%) for individual samples have high false rejection rates and/ or low probability of bias detection. T-statistics (α = 0.05) performed best in low range ratio (lowest-to-highest concentration in measurement range) and low imprecision scenarios. Mean difference (10%) performed better in all other range ratio and imprecision scenarios. Combining mean difference and paired-t test improves the power of bias detection but carries higher false rejection rates. 

Conclusions: This study provided objective evidence on commonly used rejection criteria to guide laboratory on the experimental design and statistical assessment for bias detection during method evaluation or reagent lot verification.