Hi All, I am wondering, which statistics is the most reliable, in your experience, in the following situation: We're working on a human intervention study dataset with microarrays from 16 patients before and after treatment, 32 samples altogether. Eight of the patients were treated with a drug and the other 8 were treated with placebo. The situation is, what we also saw in many other similar cases, that the patient-to-patient variance is much larger than the treatment effect. As a solution, we are experessing the RMA(after treatment)-RMA(before treatment) differences. The resulted 16 samples are analyzed with all the methods in mulltest plus the maxT/minP methods. None of these methods selected any genes at p<0.05. At the same time, t-test pointed out about 800 "significant" genes. Unfortunately, these 800 genes are not really the ones that are biologically plausible. Q1: Do you have any arguments against the RMA(after)-RMA(before) approach? RMA is a logarithm, so the difference should express a fold change. Q2: Do you have another method for similar situations when the patient-to-patient variance is bigger than the treatment effect? Q3: Seven tests say NO, is it really a NO? Should I conclude that the treatment was ineffective? Q4: Can you suggest another method that is more likely to find a "real" change in response to treatment. Thank you in advance for your input. I feel that the basic quesion is important in all human studies where multiple samples come from the same patient. We need a method to control for the patient-to-patient variance. Statistical and philosphical ideas are all welcome. Thank you, Kevin

Hi Kevin, did you do in vitro amplification of your samples to increase the cRNA available for hybridization? This is a common source of the problem you are describing. There are no mathematical methods for retrofixing this problem. > Q1: Do you have any arguments against the RMA(after)-RMA(before) approach? With RMA you loose some of your differentially expressed genes. If your intensities on the chip are not all on the very low end, VSN does a much better job. > RMA is a logarithm, so the difference should express a fold change. Almost, its the log ratio. Fold change is the log-ratio +1 or -1, respectively. > > Q2: Do you have another method for similar situations when the > patient-to-patient variance is bigger than the treatment effect? Can you do the measurements on the same patients: like before and after? In that case you should used paired t-tests. They account for that problem. > > Q3: Seven tests say NO, is it really a NO? Should I conclude that the > treatment was ineffective? First, you have to sort out your technological problems. If all stays the same afterwards, you have an indication for ineffectiveness. Not more. > Q4: Can you suggest another method that is more likely to > find a "real" change in response to treatment. Could you plot the standard error of your estimator for the difference versus the absolute value of this estimator? This could help in deciding if there is more in store for you. Cheers, Hinnerk