I have a set of samples and I performed qPCR to check the level of my transcript of interest. Now I have a series of Relative Expression data and I want to identify a cut off above which the expression might be considered "high" and below which it might be considered "low". I don't want to set the cut off arbitrarily so I was thinking about ROC analysis perhaps?
I have previously used a pool of 'normals' as cut-off. Samples greater than normal (statistically) are elevated (or depressed if going the other way), other wise they are baseline.
If you are doing a comparison just among the sample and If you have a normal distribution, could do a similar thing, greater than 1 sigma (or more if you want to be more stringent) from mean.
I don't know enough about ROC to give input on that front though.
A threshold is either relative to something given or it is arbitrary.
I possible I would avoid such dichotomous distinctions if quantitative data is available. The Rel.Expr. value is a measure of the expression. Take it as it is.
If you want to say that some value indicates a "low expression" you need some reference to benchmark what a "low expression" means. This is difficult, and the choice will finally be arbitrary anyway.
For instance, consider your gene is a transcription factor (TF). Then you could measure rel.expr values of many DIFFERENT TFs and look at their distribution (you really need to measure them, because the rel. expr. values are specific for your assay!). You can then for instance define that the 10% percentile indicates a "low expression". Then you have your (more or less arbitrary) cut-off.
Hi
in my experience if you don't have any previous cut off, the best choice is to decide a cut off value on the basis of a ROC curve (if you have all clinical data) or on the basis of median or quartile values. If you want to compare two different groups the best way is always to consider PCR value as continuos variable using for example mann whitney test.
I have all the clinical data and basically I need to do this dichotomous distinction because I have to plot a Kaplan Meier curve. I was just wondering, if I use the survival data to predict a cut off and then I use the same survival data to plot a survival curve isn't it a little bit biased? I don't know, maybe it is a silly question suggested by my lack of experience in the field.
You can do a ROC analysis to find a cut-off that will give you a somehow optimal (by specificity or by sensitivity or by some compromise) test to predict the clinical outcome. But to say that this cut-off distinguishes a "high expression" from a "low expression" (or something like this) is, to my understanding, inappropriate (and not required!). All you need for diagnosis is a cut-off. Just call it the "critical value", or the "cut-off value". Values below will simply be values lower than the critical value. Whether or not this is a "low expression" is irrelevant and can't be concluded from this fact.
I have already calculated the fold changes using the Livak method since in my experiment I have a reference gene and a calibrator (normal) sample.
I have calculated the fold change using Livak method too. When can i say that the gene of interest is over or underexpressed? Is there a preset cut off i can use? or i should compare each tumor sample with its normal tissue sample?
- Badges
- Science method