Replies to post #3444 on NorthWest Biotherapeutics Inc (NWBO)

[gpb]

I get your concept but would make a few suggestions.

Visualization: First, I was always bad with scissors and paste (that's why I repeated kindergarten several times), so instead for a visualization I suggest this. Go into Gnumeric (or spreadsheet application of your choice), and create a table with per-patient PFS, censorship (if you intend to fill out an entire timeline or play with the window/group-split), and group. In your case the starting values are to model off the earlier phase results for both treatment and control, so copy/estimate those in based on the dots (or if you want just generate a sequence from a normal distribution based on the median and a standard deviation that approximately fits the real plot - since we're going to toy with these numbers anyway, fidelity isn't crucial). Go to [your application's equivalent of] Statistics -> Dependent Observations -> Kaplan-Meier Estimates. If you don't have an equivalent built-in tool or extension for your application, this is when you download Gnumeric and paste your numbers over. Put in the parameters where directed to map your columns to time, censorship, and group. Let it output to another sheet or change that as well if you want (but leave some room on this sheet).

Now, create a new column where the cell formula is equal to your original time column multiplied by 1.43 (the stretch you gave earlier). Repeat the K-M output using the new time column. You'll note a few things in the chart and table. First, there's your new graph and median pfs advantage stretched exactly as you told it. Second, because you scaled everything uniformly (including the variance in each curve), despite the larger treatment advantage, the p-value of the advantage is unchanged.

Another thing you probably want to do is see how the p-value changes in different scenarios. Try changing everything below the median in each group to something absurdly low like 1 month and everything above the median to something absurd like 24 months for control or 84 months for treatment. Keep the medians themselves the same. This is your worst-case scenario for significance of advantage, similar to if one subgroup was perfectly cured every time and another subgroup was entirely non-responsive or even made worse. You should see now that your K-M sheet shows a much, much higher p-value. For that matter, had you made it so the range of all control events was 8-10 months and all treatment was 17-19 months, still with the same median, the p-value would be incredibly tiny.

Okay, so plugging in the event times was a bit tedious, but wasn't that a lot more fun than printing a chart on plastic wrap and stretching it over progressively larger salad bowls?

Magnitude of GTR effect: The second suggestion is that I think you're overestimating the effect of the trial's "intent of resection" on PFS. Let's take a look at three recent trials using different techniques which measured the prognostic value of EOR under various circumstances. The first trial, Sanai '11, focused on microsurgery technique and the difference between total or highly extensive resection and average or poor resection was about 135 days (NON-weighted average).
The second trial, Kuhnt '11, focused on intraoperative MRI and the difference between total resection and non-total resection (threshold 98%) was 150 days.
The third trial, Stummer '06, focused on use of the 5-ala contrast agent and the difference between 'complete' and 'incomplete' resection was 147 days.
All of these were retrospective analysis of what EOR was ACTUALLY achieved, not what was the goal (which was always 100%). Also, note how including the almost-but-not-quite group in the first trial drags down the advantage (confirming how important that difficult final two percent really is). Therefore, it's the theoretical upper limit of what's at-all possible with the right intent and I believe the actual increase in survival from requiring the 'intent' of complete resection to be maybe half that, on top of global SoC (from Stupp, 6.9 months, bringing us to 9.3 months).
For a more complete review of 10 such studies, check out Eyüpoglu '13

Your percentage, in the SoC group, would yield about 104 days (an extra month or so over my 2.4), and in your example you added that to the 8.1 months seen in the UCLA data rather than the 6.9 months seen in Stupp. My problem here is that the UCLA data includes matched concurrent controls from a study site already priding itself on achieving better than average results regardless of particular therapy, no doubt due in part to aggressive surgical resection by a well-informed leader in the field, and the timeframe of the study was in the modern era where it's generally well-accepted that greater resection means greater survival. This is in contrast to ten years earlier when 'resection of convenience' was preferred due to fear of collateral damage caused by inferior surgical and imaging techniques, and 'just biopsy' was common as well. In other words, the goal in most cases (depending on exact tumor location and infiltration) probably already was GTR, and this is factored into the matched PFS number (8.1) already. So, for these two reasons (intent != success and intent already somewhat factored into earlier data), I believe your visualization of year-long PFS in the control arm is going to turn out too high.

Also, I believe if we did have such great results in the control arm, it would imply either a later date of triggering the interim analysis, or a significantly shorter PFS advantage to treatment (thus negating the core premise of your example), due to the dearth of control patients from 2008.

So, that's why I'm sticking with ~9 months for control PFS.

Of course, for the patients' sake, I hope you're right that it's 12 months or longer, but then I also hope it's 36 months in the treatment arm too, and I don't think my hoping is enough to make it so.


On a totally unrelated note, if it seems like a disproportionate number of my posts are criticizing yours, please don't be offended. I'm just responding to things I think are worthy of response and discussing topics I think are worthy of discussion.

[Astavakra]

Flipper,
Again, thanks for this and all of your posts. I must say, this is THE most informative and true-to-purpose board I have visited in the biotech space. Because the sector is so volatile, I think, investors/posters are subject to wild imaginings, fears, hopes, hype, and all of this is written in the posts. It is such a pleasure to read intelligent, well thought out and written conversations among posters with a common goal,i.e., to learn and understand as much as possible about a treatment that may change lives immeasurably.

My remark about your example was meant more as an admission that it went over my head than that it was in some way flawed.