OK, as requested I have done the analyses for 25% and 30% mortality rates in the treated group. The tables are for the sceptical and uninformed priors (one table each) at the 30%, 45% and 60% readout points.
Important: This model is done using assumed prior distributions that could be really different to the trial design, which means results could vary from reality. If someone wants to ask Mesoblast on twitter what prior they are using for their interim analysis, that would help! (though it may be confidential)
Sceptical prior, 25% treated mortality |
| 35% control
mortality | 40% | 50% | 60% | 65% | 70% | 75% | 80% | 85% |
|---|
| 1 | 30% readout | 0.069376 | 0.103193 | 0.180672 | 0.274534 | 0.330654 | 0.395361 | 0.470939 | 0.559612 | 0.662166 |
|---|
| 2 | 45% readout | 0.076687 | 0.131928 | 0.270299 | 0.435601 | 0.526312 | 0.620794 | 0.716373 | 0.808189 | 0.888601 |
|---|
| 3 | 60% readout | 0.082773 | 0.160975 | 0.366339 | 0.592439 | 0.699426 | 0.79544 | 0.875375 | 0.93504 | 0.97281 |
|---|
Only one path to early efficacy at the 60% readout.
Uninformed prior, 25% treated mortality |
| 35% | 40% | 50% | 60% | 65% | 70% | 75% | 80% | 85% |
|---|
| 1 | 30% readout | 0.430623 | 0.617767 | 0.870096 | 0.967423 | 0.984986 | 0.993308 | 0.997042 | 0.99865 | 0.999316 |
|---|
| 2 | 45% | 0.415246 | 0.643176 | 0.916221 | 0.988048 | 0.996062 | 0.998774 | 0.999625 | 0.999881 | 0.999956 |
|---|
| 3 | 60% | 0.402381 | 0.664118 | 0.944487 | 0.995447 | 0.998924 | 0.999765 | 0.99995 | 0.999989 | 0.999997 |
|---|
Interestingly, not much change in the results with extra power at 25% mortality. you're just getting a bit too close to the control group numbers to shift the needle at 35%, 40% etc. With a 75% readout the 50% box would go green... that's about it. (also this green block is nudging up against a 10% change and not a 5% change)
Sceptical prior, 30% treated mortality |
| 35% | 40% | 50% | 60% | 65% | 70% | 75% | 80% | 85% |
|---|
| 1 | 30% readout | 0.040635 | 0.065911 | 0.127826 | 0.205835 | 0.253275 | 0.308679 | 0.374557 | 0.45396 | 0.549723 |
|---|
| 2 | 45% | 0.036132 | 0.071771 | 0.175401 | 0.315085 | 0.397584 | 0.488294 | 0.586246 | 0.688623 | 0.789165 |
|---|
| 3 | 60% | 0.031996 | 0.076569 | 0.225918 | 0.430124 | 0.541739 | 0.653 | 0.757734 | 0.848757 | 0.918903 |
|---|
No paths to early efficacy.
Uninformed prior, 30% treated mortality |
| 35% | 40% | 50% | 60% | 65% | 70% | 75% | 80% | 85% |
|---|
| 1 | 30% readout | 0.249853 | 0.420464 | 0.731328 | 0.906167 | 0.948781 | 0.972954 | 0.985872 | 0.992436 | 0.995573 |
|---|
| 2 | 45% | 0.204219 | 0.402913 | 0.775015 | 0.946695 | 0.97726 | 0.990837 | 0.996392 | 0.998536 | 0.999327 |
|---|
| 3 | 60% | 0.169908 | 0.388266 | 0.808487 | 0.968787 | 0.989545 | 0.996774 | 0.99904 | 0.999704 | 0.999893 |
|---|
Looking at this case where we have 30% treated mortality, and still seeing early efficacy at 70% control group mortality with an uninformed prior, then I would be reasonably confident that we're not using an uninformed prior if there is no result at the 45% readout. Either that or I've got some bad mistakes in my code.
Some questions for you all on improving this analysis:
Does anyone know if we can access any of the details about the interim analysis? If they told us which prior they were using to determine overwhelming efficacy then I could do a pretty definitive table where you could pick the parameters you think will be observed in the trial to work out which readout may provide early efficacy.
Would it be useful to try and reproduce these numbers with a different algorithm to verify them? No promises on if I am able to do it of course, but I did find a promising lead on a version that would simulate outputs and run until convergence to get values (this is a bit more of a professional approach but in theory you get similar results). Would people be interested in that?
Is there any interest on more analysis on what the numbers mean? Or do the outputs make sense?
If you think about it, you can infer some things about the trial results from a continuation at the 30% readout!
Closing thoughts:
Anyway, I hope this helps a few people have realistic expectations on the possibility of a result at each readout, which is an advantage over institutional investors in decision making (unless they are reading this, in which case please let me know where to send my invoice for the analysis consulting!)
I quite enjoy owning this share, never a boring day in the market, the share price is always doing something crazy!
(20min delay)