Upon reflection, and with the aid of other discussions of this very point, I have come to the conclusion that I’ll need to amplify Senator Grassley’s Febrary 11, 2008 concerns, in order to make them plain — to a wider-audience.
So, to amplify the Senator’s point:
If one is looking for NON-inferiority, one should expect to see statistically-significant higher values in some of the blinded data, and statistically-significant lower values in some of the blinded data. Said another way, there should be some set of STATISTICALLY-SIGNIFICANT differences — in at least some groupings — of the “blinded” data.
Now, if the “blinded” data ALL comes back (for the sake of a simple, clear example) with wall-thickness values between, say, .682 and .690, and there were about 640 patients in the study, all of them between .682 and .690 (on whatever thickness scale we are measuring), one may readily, and confidently, deduce that there are NO STATISTICALLY-SIGNIFICANT differences in ANY possible COMBINATION of groupings — it’s a sort of regression analysis, built into all Excel spreadsheet software (and/or science-based, off-the-shelf, data analysis programs) — which would allow anyone, anyone with access to NOTHING BUT THE “BLINDED” DATA-SET, and a computer, to know that ENHANCE had failed.
As irony would have it, the above-made-up example is, in actuality, fairly close to what happened with Schering-Plough’s ENHANCE study — now, straight from the New England Journal of Medicine‘s just-published-chart on the outcome [Click image to view full-size.]:
One could have deduced that NO set of roughly-equally populated groupings (of about 320 patients each) WOULD EVER YIELD a statistically-significant difference. That is Senator Grassley’s — and, now, my — point.
All one with the “blinded” data-set need do, is enter the values into an Excel spreadsheet, then choose the “auto-sort column data” sub-menu, select “lowest to highest“, and then, start splitting the values — take the top of the column (lowest — assume that is a Vytorin/Zetia patient), and match it to the bottom of the column (highest — assume that is a statin-only patient). Now take the second-lowest data point, and match it against the second highest. . . . and so on, and so on, until one has reached the middle of the 642 patients. This simulation/permutation would assume that all the thinnest-walled patients were given Vytorin/Zetia, and all the thickest-walled patients were given statins only. In other words, this artificial sorting would ASSUME the best possible universe of outcomes for Vytorin/Zetia from the ENHANCE study. Looking at the above-chart, I think it almost a mathematical certainty that any variances between the two “best-case” groups would be well-within the margin of error for the sample size.
Anyone with the “blinded” data could have done this, at Schering-Plough, in about one hour’s time — that’s all they’d have needed. And that would have been in January of 2007 — a full year earlier than SGP presently admits it “knew anything“.
Might this explain the hurried effort in January 2007 to query the ENHANCE “outlier data” — per the time-line SGP submitted to the SEC on January 25, 2008?
[Much Later — it seems Melissa Davis (of “The Street“) reads this blog.]