@DocMcstuffins
Re post 67434061
Statistical trawling does not make up for missing primary endpoints. Those who don't understand how trials or statistics work won't really understand that point. You and I have sort of agreed in passing - I said data patching you said trawling for correlation - I said you may be using more conventional terminology for the same thing.
But you know - I think there is actually two separate things.
I did a quick search for a paper or youtube video on trawling for correlation and didn't find a satisfactory one. Maybe you might know of such a paper or short primer that might help explain what "trawling for correlation" or "statistical trawling is" ?
What I had in mind with data patching was this slide cc-32 - from the ODAC panel briefing.
The slide shows 11 data points - each data point representing the mapping of an in vitro measurement and an in vivo measurement pair. The percentage reduction in Il2Ralpha measured in vitro on the horizontal (compared to the same set of activated anti-cd3/cd38 pbmcs but without remestemcel-L cells being added) versus percentage reduction from baseline blood levels of CD25 (ie IL2Ralpha) to day 28. So its a measure of the second potency assay matrix component in 11 cases against the measured reduction in IL2Ralpha in 11 patients. In vitro against in vivo.
MSB only had 11 data points because they couldn't do such a match up if multiple lots went into the same patient (how could you know which of the lots with the respective levels of IL2Ralpa inhibition was responsible for the observed in vivo change if you combined both lots? ) and they only had blood readings for 25 of the patients of the 54 or 55 in the GVHD001 trial.
So that's why only 11 data points.
We can actually put numbers on the data points that are reasonable percentage approximations.
So we can get a median IL2Ralpha inhibition level.
Sort the numbers from lowest to highest and we get a median value - the sixth value in a set of 11 - its about 82.
I think MSB is effectively taking 82 percent inhibition as the median value of IL2Ralpha from its set of 11 samples and then making an argument that just using the higher values. ie median of 82 percent inhibition and above they get a better relationship (correlation, association) with the in vivo data than if they use all the values (including the below the median percentage reduction in IL2Ralpha levels.
So in future they'd just use the higher values of IL2Ralpha inhibition. And voila - a better potency assay.
But is such an approach statistically valid?
I don't think so.
(20min delay)