Batch differences in biochemical measurements - statistical solutions? I'm working with inflammatory markers from human participants in a current study and have been stumped by between-batch differences. One third of the participants had their inflammatory markers measured in one batch, the next third in a second batch etc etc. 
Non-parametric tests tell me there are substantial differences in mean ranks between the batches. This could be due to measurement factors (length of time in storage/time of day assessed/different technician) or it could be due to real differences in the participants that make up the sample. These differences remain after controlling obvious confounders (e.g. BMI). Rsquared values tell me that 10% of variance in inflammatory markers is due to batch differences.
I imagine this is quite a common problem in this field - however, being new to it, I can't seem to figure out how to deal with it. I don't want these batch differences to confound results when later running a regression analysis. Is there a statistical method or some kind of technique that could help me address this issue (or normalise these batch differences)?? Could anyone suggest a paper to read? The worst case scenario I see is that I have to analyse each batch independently (thus losing a heap of power). Thanks!
 A: There are many ways in which assays for inflammatory markers, often measured with immunochemical techniques, can go wrong. This manual describes the steps involved in developing and validating such methods; you would probably be most interested in the sections on Method Validation. The problems posed by immunoassays on biological specimens are nicely described here. This guide illustrates the steps for developing quality control systems for clinical laboratories. For one of the most widely used inflammation markers, C-reactive protein (CRP), you might consult the US FDA review criteria for approving new assays or the US CDC manual for one standard test to see the level of care needed to develop and perform such assays properly. 
One very good way to deal with batch-difference problems is to incorporate biological control samples having known concentrations of the analyte, possibly including samples "spiked" with extra known amounts, into each assay batch. This provides information on recovery of the analyte through the multiple steps of processing, and it provides checks on whether the "standard curve" might have changed from assay to assay due to technical problems. If such controls were included in these assays and you can get the results for them, then they provide you a defensible way to correct directly for batch differences.
Otherwise I would be reluctant to try to correct for batch differences without independent evidence of them. This is particularly true if you think that only 10% of the variance might be due to batch differences, which is not a large amount of variance in most biomedical studies and is probably much less than the variance that you would find simply among independent laboratories analyzing the same samples. I suspect that a reviewer would be highly skeptical of any results that seem to depend on correcting for such a small proportion of the variance.
There is also a danger in trying to adjust for batch differences. As this paper  puts it, "Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses." Quoting from the abstract:

Many methods and tools exist for removing batch effects from data. However, when study groups are not evenly distributed across batches, actual group differences may induce apparent batch differences, in which case batch adjustments may bias, usually deflate, group differences. Some tools therefore have the option of preserving the difference between study groups, e.g. using a two-way ANOVA model to simultaneously estimate both group and batch effects. Unfortunately, this approach may systematically induce incorrect group differences in downstream analyses when groups are distributed between the batches in an unbalanced manner. 

This paper does, however, provide references to several methods used to correct for batch differences if you still wish to try them.
