My view on this is that the statistical significance aspect of the WOMAC scores can not be understated. Its a monumental outcome (with one caveat - that it was for the 2mg 2x6 group which is not part of our P3 - but lets put that aside for a moment and come back later).
Layman statistician here but the basics hold I think.
You have 2 populations, placebo and treatment, and each has a bell-curve of readings. The aim is to prove that in the
entire population of each group, the mean of the treatment bell curve is meaningfully better than the mean of the placebo (even if the bell curves overlap which they will always do to some extent).
Establishing the means of each population is not trivial though, unless you want to test the entire population! So you take a handful of samples inside each large population, and you can measure the mean of those samples. However, this sample mean will be different to the mean of the population, just due to random chance. How different? The population mean could be either side of the sample mean and its related to the spread in the data (standard deviation) and inversely proportional to the SQRT(n). As n tends to a very large number, the difference becomes zero (because your sample IS the population). Its called the Standard Error.
The error bars are symmetrical because its a statistic, not a measure of the spread of the actual measurements (which would be assymmetric). Its a measure of where the mean of the data COULD actualy be so what you need to see is a gap between them so that the mean of the treatment group is always better than the mean of the placebo, by a meaningful margin.
So having statistical significance with an n of only 20 is quite incredible as
@Mozzarc writes on another thread and it can be brought about if the treatement effect is large (big difference between population means), and consistent (small spread within each population) = less overlappling of bell curves with just a small number of samples.
In terms of the P3 and 1.5mg vs 2mg (/kg x 2 x 6)....lets say the treatment effect is propotional to the dose, so we have 25% less effect (ie the difference between the placebo and treatment mean is 75% of the one we have just measured. We are sampling 234 patients, as opposed to 20, and what this means is that the Standard Error should decrease by a factor of over 3 (SQRT(234/20)), so those error bars tighten up a lot remembering its not the spread of the data, its the uncertainty in where the mean is. It will narrow down a lot. If you work it out, the large increase in candidate numbers and reduced Standard Error could
easily compensate for a proportional loss in dose effect! In fact you could theoretically compensate for a 25% reduction in treament by simply doubling the number candidates. Beware, this is a very basic illustrative example, but I think the general principle holds.
So when Marco wryly states he is "hopeful" that the statistical signifcance is repeated in the P3, the wryness is because well, if you believe in statistics we have basically proven the P3 with 234 patients WILL pass. It reminds me of this....
(20min delay)