The correct Statistical test - (Jul/01/2013 )
I've got some cytokine ELISA data and I'm just trying to work out the best wy to present it and with the correct stats test.
A little background, on the test, there are 4 conditions of Macrophages;
1) -ve Control- media alone
2) +ve Control- treated with LPS
3) Macs fed apoptotic cells
4)Macs fed apoptotic cells + LPS
Now the graph I've been adivsed by my supervisor best to use is making everything a % relative to the +ve control, with the main test well being Macs+apoptotic cells+LPS. Effectively making condition 2 always 100% for each experiment.
My first thoughts for stats is simply to do an One way ANOVA with Tukey post hoc, I use Graphpad for my stats as well.
However what I really want is in someway to foucs the stats comparing conditions 2 and 4, they're capable of being paired (is making Anova's repeated measures a similar thing), a t-test won't work as well because conditions 2 equals the same. The other option I have is to use absolute values, but the problem with that is a couple of results, although show the same effect as we expect, can have a high variability in the absolute values of cytokines secreted.
Any help or observations would be great.
Check out non-parametric test statistics - unless you have a lot of samples (30+) and can expect a normal distribution of the data, parametric tests like the ANOVA are not the ones to be doing. You will probably want the Kruskal-Wallis with a Mann-Whitney U or WIlcoxon post hoc test for the equivalent to ANOVA and Tukey.
If I’ve interpreted your design correctly you have a 2 way ANOVA: ±LPS, ± apoptotic cells. This will incorporate the paired effect you are looking for.
I’ll cautiously argue with Bob and reckon that there is a good chance the data, relative to 2, is acceptably normal assuming you have replicates for each run/plate. But you need to check and Graphpad won’t (I think) be able to tell you.
I'd say that you know more about stats than I do DRT, but I have found little applicablity for parametric statistics in typical lab based biological setups, mostly due to the lack of replicates.
That's a great novel argument I hadn't considered: that non-parametric analysis is more efficient (less replicates required). I suspect it is true in a lot of simple designs by removing the need to identify the distribution. This will give me another topic to chase on a boring TV night
And as a second thought to Ross’s experiment the negative control is likely to be close to 0% with very small variance compared to the other treatments. Another point to nonparametic analysis.
Yes the n of these epxeriments is generally around 5-8.
The negative control varies slghtly between the different cytokines, for the most part it is closer to zero for the proinflammatory cytokines we've analysed (such as TNF-a), but for anti inflammatory cytokines little induction occurs by LPS stimulation (as we expect).I think there is a normalish distribution (but I haven't actually checked that).