Contents
On this page
- Why statistics decides whether a trial answers anything
- How often drug development succeeds
- Where the attrition is
- ICH E9, and the estimand addendum
- Intention-to-treat, and what replaced the simple version
- Randomisation, blinding and sample size
- Endpoints and multiplicity
- Registration is a legal requirement
- Working practices
Insights · Study guide · Clinical research and biostatistics
Biostatistics in clinical research: what is decided before enrolment
The sample size, the randomisation, the primary endpoint and the analysis population are all fixed before the first patient. Analysis cannot repair a design.
Biostatistics is not a stage of a clinical trial; it is the structure the trial is built on. The sample size, the randomisation, the definition of the primary endpoint and the analysis population are all decided before the first patient is enrolled, and a trial that gets them wrong cannot be repaired by analysis afterwards.
Why statistics decides whether a trial answers anything
Every substantive constraint on what a trial can conclude is fixed at the design stage. ICH E9 puts the requirement plainly: the statistical considerations “should be clearly specified in a protocol written before the trial begins”1
That is the whole argument for involving a statistician early rather than at analysis. By the time data exist, the sample size, the allocation, the outcome definitions and the available comparisons are all settled. Our page on the biostatistics concepts that get misstated covers the inferential reasoning itself.
How often drug development succeeds
Correcting a figure this page used to assert
The live page asserted, with no source, that “only around 10% of novel compounds make it to market”. Two things are wrong with that sentence beyond the missing citation. First, the figure is not settled: verified estimates run from 9.6% to 13.8%, and the largest and most recent study gives the highest number, which its authors' own institution described as “much higher than previous studies indicate”. Quoting the low end as though it were consensus reproduces the older literature. Second, the construction is imprecise in a way that changes the meaning: every one of these estimates measures the probability of regulatory approval for a programme entering Phase I clinical trials — not of “novel compounds” generally, which would include preclinical attrition and give a far lower figure, and not of “reaching market”, since approval is not marketing.
Table 1. Published estimates of the likelihood of approval from Phase I
| Estimate | Source | Database | Date range | Scope |
|---|---|---|---|---|
| 10.4% | Hay et al. 2014 | BioMedTracker | 2003–2011 | 4,451 drugs; 5,820 phase transitions. All indications. |
| 15.3% | Hay et al. 2014 | BioMedTracker | 2003–2011 | The same dataset, counting lead indications only. |
| 9.6% | Thomas et al. 2016 (BIO) | Biomedtracker | 2006–2015 | 9,985 phase transitions; 7,455 programmes. |
| 11.8% | DiMasi et al. (Tufts CSDD) | Commercial pipeline databases | First-in-human 1995–2007 | 1,442 self-originated compounds. |
| 13.8% | Wong, Siah & Lo 2019 | Citeline | 2000–2015 | 185,994 trials; 21,143 compounds. The largest and most recent. |
The spread is not noise; it reflects genuinely different datasets and inclusion rules. The lowest figure comes from an industry analysis of 9,985 phase transitions5; the Tufts estimate of about 11.8% rests on 1,442 self-originated compounds first tested in humans between 1995 and 20076; and Hay and colleagues report two different headline numbers from the same data depending on whether all indications or only lead indications are counted4.
The largest and most recent of these reports that “13.8% of all drug development programs eventually lead to approval.”3 — with a standard error of 0.2 percentage points3. The authors' own institution described this as “nearly 14% of all drugs in clinical trials eventually win approval from the FDA,”7 and as “much higher than previous studies indicate”7
Two notes on using these numbers. No study we opened reported a confidence interval for an overall Phase I to approval figure, so the 13.8% carries a standard error rather than an interval — writing it with a confidence interval would be an invention. And every one of these four sources was read on a repository or mirror rather than the publisher's own site, which is noted against each reference below.
Where the attrition is
Table 2. Phase transition success rates, two independent datasets
| Transition | Hay et al. (all indications) | BIO analysis |
|---|---|---|
| Phase I to Phase II | 64.5% | 63.2% |
| Phase II to Phase III | 32.4% | 30.7% |
| Phase III to filing | 60.1% | 58.1% |
| Filing to approval | 83.2% | 85.3% |
The two independent datasets agree closely, and they agree on where the attrition is: Phase II. Roughly two-thirds of programmes that complete Phase II do not proceed to Phase III. That is the proof-of-concept stage, where a compound meets a real efficacy question in patients for the first time, and it is where statistical design earns its keep — an underpowered or poorly specified Phase II either kills a workable drug or advances an unworkable one into the most expensive phase.45
Why one source is absent from this table
No phase-by-phase figures are taken from Wong and colleagues here, deliberately. A published corrigendum to that paper corrected values in exactly those columns, attributing the error to a spreadsheet formatting problem, and we could not open the corrected tables. The uncorrected figures also fail an internal consistency check against the paper's own headline estimate. The phase transitions above are from Hay and from the BIO analysis, which we could verify.
ICH E9, and the estimand addendum
ICH E9, Statistical Principles for Clinical Trials, reached Step 4 in March 1998 and came into operation that September1. It remains the foundational document, and two of its requirements are the ones most often loosely described.
The protocol and the analysis plan have different deadlines. The statistical considerations “should be clearly specified in a protocol written before the trial begins”1, while the statistical analysis plan “should be finalised before breaking the blind”1
Two deadlines, often conflated
Note the two different deadlines, which are easy to conflate. The protocol, including the sample size calculation, comes before the trial begins. The statistical analysis plan must be finalised “before breaking the blind” — which is later, but still before anyone can see which arm did what.
The estimands addendum. ICH E9(R1) was endorsed at Step 4 in November 2019 and introduced the estimand — “A precise description of the treatment effect reflecting the clinical question posed by the trial objective.”2 The addendum states that “The framework outlined in this addendum gives a basis for describing different treatment effects.”2, together with strategies for intercurrent events such as treatment discontinuation, rescue medication and death.
Intention-to-treat, and what replaced the simple version
ICH E9's statement of the principle: “The intention-to-treat principle implies that the primary analysis should include all randomised subjects.”1
Where this subject has moved on
This is where a page on clinical biostatistics most easily goes out of date. ICH E9 states the intention-to-treat principle, and it remains the reference point. But the 2019 addendum introduced the estimand — “a precise description of the treatment effect reflecting the clinical question posed by the trial objective” — and a framework for intercurrent events such as treatment discontinuation, rescue medication and death. Under it, the treatment-policy effect that intention-to-treat estimates is one well-defined choice among several, and not always the most decision-relevant one. Presenting intention-to-treat as simply the right analysis is a 1998 position.
Randomisation, blinding and sample size
Randomisation does two separate things, and ICH E9 names both. It “helps to avoid possible bias in the selection and allocation of subjects arising from the predictability of treatment assignments.”1 And, “it provides a sound statistical basis for the quantitative evaluation of the evidence relating to treatment effects.”1 — the second is the one that is usually forgotten: randomisation is not only a fairness device, it is what makes the statistical inference valid in the first place.
Blinding exists to “limit the occurrence of conscious and unconscious bias in the conduct and interpretation of a clinical trial.”1 with the aim that “The essential aim is to prevent identification of the treatments until all such opportunities for bias have passed.”1
Sample size. ICH E9 requires that “The number of subjects in a clinical trial should always be large enough to provide a reliable answer”1 and that “The method by which the sample size is calculated should be given in the protocol”1
A correction to a common claim
One correction to a claim often attributed to ICH E9: it does not say the sample size can never change. It requires that the calculation be given in the protocol, and it notes that “methods for sample size reestimation are available without unblinding data or making treatment comparisons”, with any revision justified and documented in a protocol amendment. The requirement is pre-specification and blinding, not immutability. Nor does the guideline contain a warning against post hoc power calculation, so that argument should not be attributed to it.
Endpoints and multiplicity
The primary variable has to be named in advance: “The primary variable should be specified in the protocol, along with the rationale for its selection.”1 And the consequence of not doing so is stated as firmly as anything in the guideline: “Redefinition of the primary variable after unblinding will almost always be unacceptable.”1
On multiplicity, the guideline requires that “the method of controlling Type I error should be given in the protocol.”1 The reason is the one covered on our biostatistics page: every additional comparison is another opportunity for a false positive, so testing many endpoints without controlling the error rate produces significant findings whether or not anything is there.
A sourcing note: the text we were able to read was truncated before ICH E9's dedicated section on multiplicity adjustment, so the quotations above are from its sections on endpoints rather than from that section. If you are designing a multiple-comparison strategy, read that section in the guideline itself.
Registration is a legal requirement
ClinicalTrials.gov is maintained by the National Library of Medicine, part of the National Institutes of Health. It was created under the FDA Modernization Act of 1997, went public in February 2000, and added a results database in September 2008.8
In the United States registration is not merely an expectation of journals. Under 42 CFR Part 11, “Any applicable clinical trial that is initiated after September 27, 2007, must be registered.”9 — with registration due no later than “21 calendar days after the first human subject is enrolled, whichever date is later.”9 The rule “implements section 402(j) of the Public Health Service Act (42 U.S.C. 282(j)).”9
Citing the requirement correctly
A citation note worth getting right in a regulatory document. The rule implements section 402(j) of the Public Health Service Act, 42 U.S.C. 282(j), as amended by Title VIII of the Food and Drug Administration Amendments Act of 2007. Writing “FDAAA 801” on its own is loose shorthand; pair it with the statutory citation. The Final Rule took effect 18 January 2017, with compliance required from 18 April 2017.
The journal-side requirement is separate and earlier: the ICMJE expects registration at or before the first patient's consent as a condition of publication. Meeting the legal deadline of 21 days after first enrolment can still leave a trial inadmissible to those journals, so the two deadlines both have to be planned for.
Working practices
How to apply this to your own work
- Bring statistical input in at protocol stage. Everything that limits what the trial can conclude is decided before enrolment.
- Specify the estimand, not just the analysis population. ICH E9(R1) asks what treatment effect you are estimating and how intercurrent events are handled.
- Name the primary variable and its rationale in the protocol. Redefining it after unblinding is, in the guideline’s words, almost always unacceptable.
- State the method of controlling Type I error in the protocol when you have multiple comparisons.
- Finalise the statistical analysis plan before the blind is broken — a later deadline than the protocol, but still before anyone can see the results.
- Treat Phase II as the decision point it is. Roughly two-thirds of programmes stop there, so that is where design quality pays.
- Register the trial before first patient consent for ICMJE journals, and within 21 days of first enrolment for 42 CFR Part 11. Plan for the earlier of the two.
- Quote development success rates with their source. The published range is 9.6% to 13.8%, and a bare round number misrepresents it.
Take-home message
Statistical design decides what a trial can conclude, and all of it is fixed before enrolment: ICH E9 requires the plan in the protocol, the primary variable named with its rationale, and the method of Type I error control stated. ICH E9(R1) added the estimand, which makes intention-to-treat one well-defined choice rather than simply the right answer. Phase II carries the heaviest attrition. And the success rate of drug development is a range of 9.6% to 13.8% across four studies, not a round 10% — the largest and most recent gives the highest figure.
The published estimates do not agree, and the range matters. Verified figures run from 9.6% to 13.8%: 10.4% from Hay and colleagues for all indications and 15.3% for lead indications, 9.6% from the BIO analysis, about 11.8% from the Tufts briefing, and 13.8% from Wong, Siah and Lo — the largest and most recent study, covering 21,143 compounds. Quote a figure with its source, not as a round number.
That sentence has two problems. The figure is the low end of a range that runs to 13.8%, and the most recent and largest study is the highest. And the construction is imprecise: these estimates measure the probability of regulatory approval for a programme entering Phase I — not of novel compounds in general, which would include preclinical attrition and be far lower, and not of reaching market, since approval is not marketing.
Phase II, in both datasets we could verify. Only about 32% of programmes move from Phase II to Phase III in Hay and colleagues’ data and about 31% in the BIO analysis — so roughly two-thirds stop there. It is the stage where a compound first meets a real efficacy question in patients, which is why the statistical design of a Phase II study carries so much weight.
The ICH guideline Statistical Principles for Clinical Trials, which reached Step 4 in March 1998 and came into operation that September. It is the foundational statistical guideline for trials, covering trial design, randomisation and blinding, sample size, the choice of primary variable, and the analysis population. ICH E9(R1), the estimands addendum, was endorsed in November 2019.
In ICH E9(R1)’s words, “a precise description of the treatment effect reflecting the clinical question posed by the trial objective”. The addendum introduced it along with a framework for intercurrent events — treatment discontinuation, rescue medication, death. The effect of specifying an estimand is that you state exactly which treatment effect you are estimating before you estimate it.
It remains the reference point, but the simple version of the claim is out of date. ICH E9 states that the principle “implies that the primary analysis should include all randomised subjects”. Under the 2019 addendum, the treatment-policy effect that intention-to-treat estimates is one well-defined choice among several, and not always the most decision-relevant one.
Yes, under conditions, and ICH E9 says so. It requires the calculation to be given in the protocol, notes that “methods for sample size reestimation are available without unblinding data or making treatment comparisons”, and asks that any revision be justified and documented in a protocol amendment. The requirement is pre-specification and blinding, not immutability.
An applicable clinical trial initiated after 27 September 2007 must be registered, no later than 21 days after the first human subject is enrolled, under 42 CFR Part 11. The rule implements section 402(j) of the Public Health Service Act, 42 U.S.C. 282(j), as amended by Title VIII of FDAAA 2007. It took effect on 18 January 2017, with compliance required from 18 April 2017.
References and sources
- ICH Harmonised Tripartite Guideline E9: Statistical Principles for Clinical Trials. Step 4, March 1998; in operation September 1998. CPMP/ICH/363/96. read 15 October 2026. View on FDA guidance (HHS)The foundational statistical guideline for clinical trials. Read from the FDA's published guidance version. Section numbering differs between the ICH and FDA renderings, so both forms are given where cited.
- ICH E9(R1): Statistical Principles for Clinical Trials: Addendum: Estimands and Sensitivity Analysis in Clinical Trials. ICH Step 4, November 2019; FDA guidance May 2021. read 15 October 2026. View on FDA guidance (HHS)The estimands addendum. It is what makes a 1998-era presentation of intention-to-treat as simply the correct analysis out of date.
- Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20(2):273–286. read 15 October 2026. View on MIT DSpace recordThe largest and most recent of the four: 185,994 trials and 21,143 compounds from 2000 to 2015, giving 13.8%. Note a published corrigendum corrected several phase-transition values, so no phase-by-phase figure is taken from this paper.
- Hay M, Thomas DW, Craighead JL, Economides C, Rosenthal J. Clinical development success rates for investigational drugs. Nature Biotechnology. 2014;32(1):40–51. doi:10.1038/nbt.2786 read 15 October 2026. View on a third-party full-text mirror4,451 drugs and 5,820 phase transitions from 2003 to 2011. The source of the 10.4% figure, and also of a 15.3% figure for lead indications from the same dataset. Read from a mirror, not the publisher.
- Thomas DW, Burns J, Audette J, Carroll A, Dow-Hygelund C, Hay M. Clinical Development Success Rates 2006–2015. Biotechnology Innovation Organization, Biomedtracker and Amplion, June 2016. read 15 October 2026. View on a university-hosted copy9,985 phase transitions across 7,455 programmes, 2006 to 2015. The lowest of the verified estimates at 9.6%. Read from a university mirror rather than the organisation's own site.
- DiMasi JA, Grabowski HG, Hansen RW. Cost of Developing a New Drug. Tufts Center for the Study of Drug Development briefing, 18 November 2014. read 15 October 2026. View on a third-party copy of the Tufts briefing1,442 self-originated compounds first tested in humans between 1995 and 2007. This is the 2014 briefing deck, not the later peer-reviewed article, which we could not open. Cited as the briefing.
- MIT Sloan School of Management. Measuring the risks and rewards of drug development. 31 January 2018. read 15 October 2026. View on MIT SloanThe authors' institution describing the finding, and the source for the framing that the figure is higher than earlier studies suggested.
- U.S. National Library of Medicine. What kind of web site is ClinicalTrials.gov? NLM Support Center, article KA-03869. read 15 October 2026. View on NLM Support CenterRead from NLM's own support site rather than ClinicalTrials.gov itself, which did not surface as a fetchable page. The registry is maintained by NLM, part of NIH.
- Clinical Trials Registration and Results Information Submission, 42 CFR Part 11. 81 FR 65138, 21 September 2016; amended 89 FR 97559, 9 December 2024. read 15 October 2026. View on eCFRThe codified US requirement. It implements section 402(j) of the Public Health Service Act, 42 U.S.C. 282(j), as amended by Title VIII of FDAAA 2007 - so citing FDAAA 801 alone is loose shorthand.
Pubrica AcademyWritten and maintained by the subject matter experts of Pubrica who do this work. Articles are revised when the standard underneath them moves, and each one carries the date it was last checked.
Related: biostatistics concepts that get misstated · CONSORT 2025 for reporting trials · ethics in research design.
How Pubrica helps
Our biostatisticians work on clinical trial design and analysis: sample size and power at protocol stage, estimand and analysis-population definitions to ICH E9(R1), statistical analysis plans, the analysis itself, and the statistical sections of the clinical study report.