Reference-class base rate
46% – 51%
historical share of Phase 3 trials in this class that reached approval
This is a base rate, not a prediction about this trial.
It is what happened to comparable trials historically. It is not a model
fitted on outcomes — no outcome labels exist for trials that have not read
out — and it uses none of this trial's own results.
How this number was reached
- Published Phase 3 → approval rate for neurology: 51.1% (Wong/Siah/Lo 2019) and 46.0% (BIO/Informa 2011–2020).
- Biomarker selection is present but is NOT applied: outside oncology the same split is ~null (60.2% vs 59.0%), so adjusting would invent an effect the evidence does not support.
Design flags
- • Subjective endpoint with little or no blinding — the combination where treatment effects are most prone to inflation (Wood, BMJ 2008: ROR 0.69).
- • Not randomised.
- • Enrolment is complete, so the readout date carries less recruitment risk.
Flags are shown separately rather than folded into the
number. Wood et al. (BMJ 2008) measured effect-size inflation, not approval
probability, so multiplying it into an approval rate would misuse the
finding.
Deliberately not adjusted for
| Enrollment size | no evidenced association with Phase 3 success |
| Number of sites | no evidenced association |
| Endpoint type | no evidenced association with success once area is held |
| Phase 2 design | single-arm vs randomised Phase 2 does not predict Phase 3 |
Each of these is visible in the registry and each is
evidenced to be null for Phase 3 success. Using them would add
false precision.
Evidenced, but not in the registry
| Sponsor approval track record | 51%→81% (Danzon 2005) — not a registry field |
| Orphan designation | aOR 2.3 — zero CT.gov fields carry FDA designations |
| Phase 2 effect-size shrinkage | rHR 0.73 — requires prior-trial results |
The standing criticism, stated plainly. Hampson et al.
(Pharmaceutical Statistics 2022;21(2):439–459, eight of nine authors at
Novartis) argue that starting from industry rates and applying subjective
adjustments "lacks transparency and fails to make full use of data from
previous clinical trials". That criticism applies to this figure. The response
here is to show every input and to refuse unsupported adjustments, not to
claim the objection does not apply.
Rates: Wong, Siah & Lo, Biostatistics 2019 (CC BY 4.0; corrigendum
values) and BIO/Informa/QLS Clinical Development Success Rates 2011–2020
(cited, not reproduced). MIT Project ALPHA is not used — its terms
restrict use to governmental, charitable or social purposes and bar
redistribution.