Appropriate To Use Roc To Evaluate Validation Of Microarrays With Qpcr?

Question

[Adam Cornwell]

Hello,

Somewhere before I think I have seen ROC used as a method of evaluating the success of validating a microarray experiment via qPCR. It seems like this could work, if you view the PCR results as a source of true positives/true negatives. That's not necessarily the case though, and so I have some reservations about apply ROC in such a manner. Does it make sense to use it that way? I typically just look at correlations between the array and the qPCR results, but AUC from ROC would similarly be a nice summary statistic.

Thanks.

[brentp]

I don't think it's a good idea to lose information by forcing a continuous variable to a true/false. What is the problem with using correlation?

[pld]

In the end though, expression arrays are used for hypothesis testing. It would be interesting to generate a surface representing the ROC over the space of p-value and absolute fold change.

At first glance, I don't think there would be anything out of normal you would have to do. Pick a set of genes and assign them to the classes "changed" and "unchanged" given your qPCR data, this is your control set. Go, back to the microarray data and make calls on the same genes. Calculate the TP/FP/TN/FN rates. Repeat this a few times across a range of thresholds, and you have your ROC data.