I am in the middle of a large screen using an ELISA based assay and I find that even though I use the same standards and conditions for each plate there still exists some variation between my plates. This became more apparent to me when I ran the assay twice for the same samples (one assay showed an average of 1.6x higher results after I fitted the readings to the standard curve but the variation between the samples was similar in both the plates - the results seem to cluster for each plate). There may be many reasons for this and obviously some may be out of my control. My main question is that is there a statistical method to correct such variations between two plates (inter-assay)?
Note: all samples are evaluated in a triplicate manner and intra-assay CV values are below 5 percent.
I think you can edit your data as percentage response compared to the control/untreated group and then performed a statistical test such as ANOVA coupled to dunnett's post hoc or Sidak
Showing your data as a relative expression (compared to control) might be a solution but then it will be semi-quantitative way and your control would always be 1 or 100% (no S.E.M for control group when you would show the results of few independent experiments and this may challenge your statistical analysis).
The statistically "most correct" approach is to appreciate the variation between plates using an appropriate model. One possibility here is a mixed model where "plate" is used as a random factor.
Going back to "percent of control" is an ancient if not medival solution. It works, though, but it ignores all progress made during the last centuries and lacks to provide an absolute measure that can be valuable in itself.
I agrre with Wihelm but although I can understand some variation for a biological replicate, but if it a technical replicate gives much variation, that is undesireable - so if you are using the same standard concenration then there should be minimal reading variation both inter and intra plate taking into consideration pipeting errors etc. Otherwise, have a technician have a look at your machine or check with the manufacturer of plates
One of the big fallacies in science is that there is no "day to day" variation. In Industry, we will use "blocks" to control for these types of variation. If you ran only one assay per day, then I would call an assay a block. If you ran multiple assays in the same day, then ran more assays on a later day, I would use "day" as a block.
When you say you ran samples in replicate, were they all in the same assay? Did you ever run a pair of assays on the same day?
I follow Preeti Bansal question. The best way is to normalize with a standard curve made in each plate.
Hey Jochen,
Thank you for your reply: "The statistically "most correct" approach is to appreciate the variation between plates using an appropriate model. One possibility here is a mixed model where "plate" is used as a random factor.".
It sounds like something I would want to do with my data but how can I generate such a model?
Dear Andrew,
My replicates are on the same assay but I repeated one of my assay on another day to see the variation between the two assay on different days and there was an average increase of 1.6x in the second assay.
Dear Jean-Marc,
Yes you're right but when I re-ran the assay as I said before the average 1.6x increase got me worried.
Thank you so much Hamid, for your reply! I would ask you to do a small experiment, with few samples only. This time do an ELISA. prepare the standards in such quantity that you can spare for next day repeat. And run the assay the next day with the same samples with the last time made standards. Might be you will be able to reproduce the results. As if there is change in conc. or degradation, there will be same in both the cases. Besides, are the storing conditions for standards and the samples are same; etc. It might be carzy! But, it may answer you the reasons of variation! Besides, I would like to know exactly which ELISA you are doing. I mean is there chance of very rapid degradation of the target molecule in your sample? or the reaction between the molecule of your interest and sample'other components. Which samples are these exactly?
If the assay (including the calibration) gives you different answers each time, then I think that the assay is not suited to provide absolute measures. You may still use it to investigate relative changes within an assay. But to compeare results between assays you need then need to replicate the whole assay many times to get "mean result" (averaged over assays) and to assess the variability associated with your final estimates. It could turn out in such a case that a sophisticated calibration is a waste of time (and money).
Dear all,
After some research I've come up across "Randomized Block Design in regression analysis notation" (http://www.socialresearchmethods.net/kb/statblck.php) and I think this is the answer I've been looking for but I'm not sure how to implement it to my data and I'm yet to find a software solution that can perform such analysis. It would be great if someone could guide me towards a suitable tool.
Best regards and thank you, Hamid
- Badges
- Science method
- Similar topics
- Medicine
- Immunology
- Immunology Techniques
- Immunoassay
- ELISA