Urgh. I made the mistake of thinking I'd finished a chapter, which of course is the cue for my supervisor to point out how I should have done it! So, for those whose brains are still functioning, may I seek your advice on a bit of basic analysis?

Part of my research has involved testing antibody levels in patients and controls, by ELISA. All kits came with multiple controls, from which I created the standard curve and determined sample concentrations as instructed. So, I have mean concentrations (in RU/ml or U/ml) for each sample and want to use a t-test to check for any significant differences between the two groups. Can I just do this with the U/ml data I have, or do I need to log transform the numbers first?

Sup is keen on my logging like a timber merchant, but I'm not sure and my brain is too addled to figure out the logical reasons for either option :-( any help massively appreciated.

Hi Teek :)

ok, Sneaks being the PGF Stats wizz will prob know best, but first off you need to check your data to see if they are normally distributed, which will tell you if you need to log transform. So, I use SPSS for this stuff, and depending on what the sample size is, I use the Kolmogorov-Smirnoff test or the Shapiro Wilk test if it's less than 50ish. if it's normally distributed i.e most data points round the centre and not skewed to either side of the curve, then you can use the t-test without log transforming. If it's not normally distributed, you'll need to use a non-parametric test.
It looks like your data is independent having only one value for each patient, so the independent t-test is ok if you are not transforming, and the Mann Whitney is the non-parametric equivalent of this if you are transforming. One thing though, if you have lots of zero values for your data you may have problems with log transforming. You can't log transform zero data so you just have to change it all to 1, which is acceptable, but I personally don't like doing it when there are lots of zero values.

I hope that helps!!

Thanks algae!

I'd already checked the distribution and used t-test for normals and mann-whitney for abnormals. Does using mann-whitney require you to log transform values? Ifeel as if I ought to do it for all tests or none but perhaps that doesn't matter?

teek - if you transform the data, then make sure you do it for all the data within the same test i.e. don't compare x (transformed) and y (untransformed). But you don't have to do it across separate tests, unless you then go on to talk about/plot values e.g. Mr Man 1 scored 3 and Mr Man 8 scored .0001 (because you transformed Mr Man 8's data).

Anyhoo, if you have SPSS 18 with the bootstrapping module you can click the 'bootstrap' option on the t-test, which can correct for non-normal data which means you don't have to use mann-whitney

Yea its fine to use both tests in your results, there's no point log transforming for a mann whitney as then you can just use the t-test. But it;s up to you if you want to keep all your tests constant, i.e. use only parametric or non-parametric tests. Although as there are different tests for different types of data I don't see any problem using different tests.

Thanks guys! This may just give me what I need to ignore my sup and NOT redo two dozen graphs! :-)