About method comparison

Method comparison, instrument comparison and reagent lot comparison measure the same specimens two ways and judge every specimen's difference. Bias against an assigned value compares results with an assigned value.

A method comparison fits four regressions to the specimen means. The Primary regression supplies the estimates for the limits, the decision-level table, the residual plot and the band. Each fit suits a different pattern of measurement error. Passing-Bablok, the default, suits two methods with similar error and is less affected by a few unusual specimens. Deming suits two methods that both have roughly constant error SD. Weighted Deming suits error that grows in proportion to concentration. Ordinary least squares suits a comparative method whose error is negligible. What each fit needs is on the reference page.

The residual plot shows whether the chosen fit matches the data. Curvature or a widening spread means the fit or its error model is a poor match for the data.

Both Deming fits need λ, the ratio of the two methods' error variances. Its value can come from measurement evidence or a stated assumption. Replicate pairs measured in the comparison itself give λ from each method's scatter within a specimen. Linked repeatability or multi-day precision studies of the same analyte and unit give it from each method's repeatability SD. A stated assumption uses the value entered, or 1, which treats both methods as equally imprecise.

A link to a precision study keeps the chosen result until you select a newer one. Linked precision results describes what happens when the linked study changes.

The comparison studies decide agreement specimen by specimen. Each specimen's difference is checked against the allowable difference at its own concentration, and at least 95% of specimens, rounded up, must be within their limits, whatever the mean difference. With fewer than 20 specimens that means every one. Method comparison, instrument comparison, reagent lot comparison and the PT/INR lot comparison share this count rule.

A mean difference inside the limit can hide specimens outside it. In the instrument comparison example, Analyzer B's mean difference is within ±4 mg/dL, yet with a ±4 mg/dL limit three specimens are outside it and B is Not met. In the lot comparison example, a ±1 mg/dL limit fails one specimen although the mean difference is well inside ±1. Specimen allowances alone decide acceptance; regression statistics and the mean difference support the comparison. Counting specimens explains how a specimen without both results is counted.

A bias estimate comes with a confidence interval that shows how well the average is known. The 95% limits of agreement answer a different question: where about 95% of individual differences fall. A narrow confidence interval with wide limits of agreement means the average is estimated precisely but individual differences vary widely.

The bias interval depends on whether the results are independent. Replicates of one material in one run give an interval from the replicate SD. It describes that run's repeatability only. Results from one material on different days aren't independent, and SD / √n would understate the uncertainty. A bias check therefore reports repeats of one material across days or runs without an interval. Separate specimens that share one assigned value are independent results. Intervals by how results relate lists each case.

Bias at a medical decision level is read from the primary fit. The intercept is judged on the fitted line itself and needs no specimens near zero. Without low specimens, the intercept is an extrapolation. Read it together with the bias at your lowest decision level.