I have been thinking about the statistical analysis plan they might be developing, which led to some scenarios I thought I would share. One caveat - I am a researcher and do my own statistics, but do not do KM survival analysis (but I know what it is). Although clinician biases in trying to figure out whether a drug works on a patient by patient basis are well know, I think the long history of DCVAX use by a highly experienced clinician (LL) who believes it works based on her clinical practice means that we do not have to consider the scenario where DCVAX proves entirely ineffective.
Scenario 1: PFS (primary outcome) is significant. No idea how likely this is, but if this were the case, that plus the long tail results would IMO be plenty to get DCVAX approved, regardless of whether traditional OS measures are significant.
These other scenarios all assume PFS is NOT significant.
Scenario 2: Significant test of the main effects of DCVAX based on the condition to which patients were assigned at baseline (ITT analysis). This pattern is what we would see if DCVAX works in newly diagnosed GBM, but does not work at all for recurrent GBM. Visualize a graph with a curve steeply sloping downward (placebo regardless of crossover status) and a clearly separate curve sloping downward above it less steeply (DCVAX). This is the most straightforward homerun result we could get for the key prespecified secondary outcome (OS). I don't think the FDA would provide much pushback against this even though it is technically a secondary endpoint, especially in context of the long tail results. Approval likely.
Scenario 3: This could be tested as the interaction between ITT group assignment and timing of DCVAX administration (e.g., placebo coded as 0, crossover coded as 1, and those assigned to DCVAX originally coded as a 2), or ignoring initial group assignment, simply as the main effect of DCVAX timing. A significant analysis like this would indicate that the efficacy of DCVAX depend on when it is given. This would be the most relevant analysis if DCVAX worked well for new GMB and worked somewhat but less well for recurrent GMB. Visualize the steepest downward sloping survival curve for placebos who did not cross over, a separate intermediate downward sloping curve for crossovers, and the shallowest curve for those originally assigned to DCVAX. If significant, the pattern of this effect would be biologically plausible, and because of this, I believe it would support the efficacy of DCVAX, and in context of strong long tail results, would also likely support approval.
Scenario 4: This is the best case scenario for patients but the worst case scenario in terms of interpretability. Visualize the curves for the patients originally assigned to DCVAX and crossover patients overlapping, with both showing a shallow downward sloping curve, with a much steeper downward sloping curve for the non-crossover placebo patients. Clearly DCVAX seems to be doing something (visually) but you would get a nonsignificant main effect of DCVAX in ITT analyses (because there are so many crossovers with long survival who would dilute the treatment effect) and the test in Scenario 3 would also not be significant (because the two curves with the most patients overlap). This is the only scenario in which comparison to SOC controls becomes crucial. There would be no direct statistics to show that DCVAX works, so approval could only be based on showing greater long-tail OS percentages at key milestones (e.g., 2 year, 3 year) than you would expect in SOC. These comparisons can be tested statistically, but how convincing the argument is in terms of approvability depends on whether an SOC number can be quantified for a group similar enough to the DCVAX trial population. While this may be difficult, it would be reasonable to handle this by doing statistical comparisons of milestone survival between DCVAX (initially assigned to DCAVX + crossovers) and prior SOC samples, including all relevant individual difference variables in the model as covariates (in this case, percentage of patients in each study with partial surgical resection, +MGMT status, etc). I think a meta-regression approach (like meta-analysis but looking at the influence of other variables) might be best for comparing the DCVAX trial results to a composite of other similar trials. In theory, if the DCVAX effect in these meta-regression models is significant even after controlling for all relevant sample characteristic differences between studies, then we would be justified in concluding that DCVAX works to increase long-tail survival. Given the orphan stats of GBM and the current regulatory climate at the FDA for orphan disease interventions, I think this has greater than 50/50 odds for approval even in this worst case scenario.
When I worked through these scenarios, I realized that the crossover issue is not necessarily an issue that is impossible to address. No idea what approach they actually plan on using, but I think a good statistician could figure out a way to show DCVAX works if indeed it does work.
