In our laboratory we have tissue pieces that secrete a certain protein. After we subject them to treatment we measure the secreted protein again. The number of pieces is usually 4 or 5 in a group.
Up until now the data was analyzed by calculating the percent of difference; if following the treatment there's more than 15% reduction (average for the group) in the level of protein - the treatment is rejected.

I believe that using a paired t-test is a better a way to look at the data, but a I do not know how to translate the 15% into p-value or any other statistical threshold.

hi.
I think you are mixing the significance ( where p would be relevant) with effect size, where the 15% would be relevant.

the paired t test does seem to be a good idea, provided you can build pairs of data, which was not quite clear to me from your description BTW.

If you can, you will get two outcomes, the p value will tell you whether the test is significant ( i.e. there is a low chance that the difference you see is due to chance only) and the confidence interval for the difference (the range of possible differences ). This you will probably need to express in percentages.

Maybe a warning, your sample sizes do not seem to be too big. I think you should check the power of the present test and of the t test as well and see if you need larger samples?

I think the issue is that maxina wants to do a test on the ratios/percents instead of the raw protein level differences.

Maxina, is there a reason you can't compare the raw protein levels, perhaps because the tissue pieces are different sizes?

It's preferable to do the t-test with the protein levels themselves if the pieces are similar sizes, but if you want to do the ratios, here is something I found:

Which basically tells you to do the paired T on the log of the values. It's a little more complicated than a paired t-test, because as you seem to know, ratios and percents by themselves are a little sketchy for comparisons.

hi Panurge,
I think the ratio test (thanks for the link) is not addressing the problem of different sizes but the problem that the change might be proportional to the original value e.g. the effect of the treatment is as an example a consistent increase by 10% . If the starting value is small, then the change in absolute values will be small if the starting value is large then the absolute change will be large although we have the same effect.

if the sizes are different rhe scaling according to the size would work, I think.

regrds and thanks for the pointer!
rogojel

Last edited by rogojel; 08-11-2013 at 04:44 PM.
Reason: correction