Validation of bioassays for quality control.

For a biological assay to be useful for quality control it should fail bad lots, pass good lots, and estimate relative potency with high accuracy and precision. To fail a lot we rely most heavily on the test for parallelism. For the parallelism test and other preliminary tests as well as for inference, appropriate estimates of assay variation are crucial. Location effects on 96 well plates and serial dilution of samples using multichannel pipettes make it difficult to obtain good estimates of assay variation. This paper develops the use of a split-block design and analysis where blocks are reasonably consistent regions of a plate; this approach removes some location effects, allows other location effects to be treated as assay variation and provides appropriate measures of assay variation. Randomization, even within the split-block design, is difficult without robots to reduce the likelihood of procedural errors. There are hardware, software, and validation obstacles to implementation of robots in the bioassay laboratory. More generally, validation of a bioassay should be reported on log relative potency and must address between- and within-assay variation. When between assay variation is not small, the usual weighted approach to combining relative potency estimates (which ignores between-assay variation) is inappropriate; a simple sampling average and standard deviation is a better solution.