phaserate
Reference-class base rates · clinical development method-v0 · snapshot 2026-07-23

The same trials. Two defensible answers. Nobody tells you which one you're reading.

Phase-transition base rates computed from ClinicalTrials.gov, served with the conventions that produced them — so you can change one and watch the number move.

What travels with every figure Sample size, 95% confidence interval, linkage coverage, methodology version, source lineage, and the three choices the rate is conditional on.
Figure 1 — Phase 2 → Phase 3, ophthalmology, 2015–2025

Identical data, identical numerator. Only the censoring rule differs.

0% 20% 40% 60% 70% A quiet programme counts as failed — 17.7% (95% CI 13.6–22.8) A quiet programme counts as failed our default · n=265 17.7% BIO/Informa, published — 35.5% BIO/Informa, published same transition, same area · n=200 35.5% A quiet programme leaves the denominator — 57.3% (95% CI 46.5–67.5) A quiet programme leaves the denominator BIO's stated convention · n=82 57.3% Bars are 95% confidence intervals (Wilson score). All three figures are correct. They answer different questions.
PhaseRate, convention varied Published reference
Why they differ Ophthalmology carries an unusually high share of programmes that stop reporting without formally terminating — 70%, against 58% across all indications. Under one rule they are failures; under the other they leave the denominator entirely.
Reproduce it Both figures are one API call apart. Neither is an illustration.

§ 1Published Phase 2 → 3 rates span 22% to 58%

Almost none of that is disagreement about the data. It is three undeclared choices, each measured on unchanged data, and each larger than any therapeutic-area or data-vendor effect.

Table 1 — effect of each choice, measured on unchanged data
ChoiceEffectMeasured by
Censoring rule16.9 pp Aryal et al. 2026
Estimator10.4 pp Wong, Siah & Lo 2019
Counting unit8.6 pp DiMasi et al. 2013
Nobody publishes them The 2016 and 2021 editions of BIO/Informa/QLS's Clinical Development Success Rates — the industry's reference — do not state what their censoring rule is.
References Aryal, Ciliberto, Farmer & Khmelnitskaya (2026). Valuing Pharmaceutical Drug Innovations. arXiv:2212.07384. Wong, Siah & Lo (2019). Estimation of clinical trial success rates and related parameters. Biostatistics 20(2):273–286. DiMasi et al. (2013). Clinical approval success rates for investigational cancer drugs. Hay et al. (2014). Clinical development success rates for investigational drugs. Nat Biotechnol 32:40–51.

§ 2One call, conventions attached

REST and MCP. Change censoring_rule and the same query returns the other answer, with the change recorded in the response. 

{
  "n": 978,
  "rate": 0.238,
  "confidence_interval_95": [0.213, 0.266],
  "coverage": 0.9649,
  "conditional_on": {
    "counting_unit": "asset_sponsor",
    "censoring_rule": "quiet_is_failure",
    "imputation_policy": "none"
  },
  "methodology": {
    "version": "method-v0",
    "snapshot_date": "2026-07-23"
  }
}
For assistants The MCP server instructs the model to quote the conventions alongside the rate, so it cannot hand a user a number stripped of its meaning.
Access Keys are argon2-hashed and shown once. Monthly quota is returned on every response; exceeding it returns 429 with the used and permitted counts.

§ 3What we refuse to do

  1. Thin cells refuse to quote. Below 20 evaluable programmes a cell returns its reason instead of a confident-looking number over a hollow base.
  2. Unverified is not failed. A trial whose registry status has not been confirmed in two years is reported separately, never counted as a failed transition.
  3. Non-efficacy stops leave the denominator. A trial halted for funding or enrolment says nothing about the asset. The count is reported either way.
  4. Unimplemented options error. Ask for imputation we do not perform and you get a 501 — never an unimputed figure wearing the wrong label.
  5. Limitations are published, not discovered. Registry-only visibility biases us downward; area rates are weak predictors of individual diseases. Both are in the documentation.
Standing rule Every figure on this page is a live response from the service. If a number cannot be obtained with one call, it does not appear here.
Coverage, this snapshot 97% of programmes that ran a Phase 2 could be dated and judged 80,476 candidate trials, industry-led, drug and biologic 20 evaluable programmes — the floor below which a cell refuses to quote

Know which number you're quoting.

Access is by request while the dataset is in its first release. Tell us the therapeutic area you care about and we will send a key — hello@phaserate.com.

Computed from the Aggregate Analysis of ClinicalTrials.gov (AACT) Database, Clinical Trials Transformation Initiative. Documentation · methodology method-v0.

Cite as Aggregate Analysis of ClinicalTrials.gov (AACT) Database. Clinical Trials Transformation Initiative (CTTI). Available at: https://aact.ctti-clinicaltrials.org/ (Accessed: 23 July 2026, single static snapshot).