9 Phase III: Confirmatory
The transition from Phase II to Phase III often commits a sponsor to hundreds of millions of dollars on top of 5-7 years of prior work, all of it staked on a therapeutic benefit that Phase II could only suggest. Phase II asks whether a drug might work. Phase III has to demonstrate that it does, with the statistical rigor and methodological quality that regulatory agencies require for approval.
These large, well-controlled pivotal trials, classified as confirmatory under ICH E8(R1), generate the primary evidence regulators require for marketing approval and establish the safety profile with sufficient precision to support labeling (International Council for Harmonisation 2021).
9.1 Characteristics of Phase III Trials
Table 9.1 sets the two phases side by side (International Council for Harmonisation 2021).
| Feature | Phase II | Phase III |
|---|---|---|
| Sample Size | Often 20-300 patients | Hundreds to thousands |
| Primary Objective | Explore efficacy; find optimal dose | Confirm efficacy with statistical rigor |
| Design | May use adaptive or exploratory designs | Rigorous randomized controlled trials |
| Sites | Often single-center or limited multi-center | Multi-center, often multinational |
| Setting | Specialized research centers | Typical clinical practice settings |
| Duration | Months | Often years (long-term follow-up) |
| Control Group | Placebo or active comparator | Placebo (if ethical) or active comparator |
| Endpoints | May use surrogates | Clinically meaningful outcomes preferred |
| Regulatory Role | Generate “go/no-go” decision | Provide pivotal evidence for approval |
| Cost | About $7-20 million per study on average (Sertkaya et al. 2016) | About $12-53 million per study on average, and up to about $350 million for the largest pivotal trials (Sertkaya et al. 2016; Moore et al. 2018) |
Size is the first difference. Enrolling hundreds to thousands of patients, depending on the indication and endpoints, provides the statistical power to detect clinically meaningful treatment effects with high confidence. Rigor is the second: randomized, controlled designs minimize bias and confounding, and placebo controls (when ethically appropriate) or active comparators allow observed effects to be attributed to the investigational drug and not to other factors.
Phase III also moves closer to routine care. Early studies occur in specialized Phase I units; Phase III takes place in hospitals, clinics, and medical offices where patients receive routine care, under conditions intended to approximate the drug’s eventual clinical use. Trials typically run at multiple sites, often across multiple countries, and this multicenter approach helps ensure that the study population is representative of the intended patient population and that results are generalizable across different healthcare settings and geographical regions.
9.2 Trial Design Considerations
The choice of control group is among the most consequential design decisions. Placebo controls provide the clearest evidence of efficacy but may be unethical when effective treatments exist. Active controls demonstrate that the new drug is not worse than existing therapy by more than a prespecified margin (non-inferiority) or that it is better (superiority), though they require larger sample sizes.
The primary endpoint should be clinically meaningful and clearly defined in the protocol, and regulatory agencies increasingly expect trials to demonstrate effects on outcomes that matter to patients: survival, symptom improvement, functional status. When those outcomes require long follow-up, surrogate endpoints that are reasonably likely to predict clinical benefit may be acceptable, particularly for serious conditions with unmet medical need.
The ICH E9(R1) estimands framework defines the primary endpoint precisely enough to align the clinical question of interest with the statistical analysis that answers it (International Council for Harmonisation 2019). An estimand specifies five attributes: (1) the treatment being evaluated, (2) the target population, (3) the variable (endpoint) to be measured, (4) how intercurrent events (such as treatment discontinuation, rescue medication use, or death) will be handled, and (5) the population-level summary measure. For example, a Phase III cardiovascular trial might define its estimand as the effect of Drug X versus placebo on time to cardiovascular death in patients with heart failure, summarized as a hazard ratio, using a treatment policy strategy that counts events regardless of treatment adherence and stating how deaths from other causes are handled.
The statistical analysis plan must be fully specified before unblinding occurs: the primary analysis population, handling of missing data, multiplicity adjustments for multiple endpoints or comparisons, and the specific statistical tests.
Increasingly, sponsors do not stop at the Phase II and Phase III boundary at all. A seamless design runs both phases as a single trial: it may begin in Phase II exploratory mode (selecting doses, confirming the target population, generating hypotheses) and then transition without a stop into a confirmatory Phase III stage, with the Phase II data contributing to the final analysis under pre-specified conditions. Designed carefully, it reduces calendar time and can preserve patient data across both stages. The design and operational requirements are covered at Section 12.4.
The master protocol goes further still: one overarching design evaluating multiple therapies, disease subtypes, or both under shared infrastructure. Platform trials such as RECOVERY, REMAP-CAP, and I-SPY 2 add and drop treatment arms against a common control as evidence accumulates, and several have produced confirmatory, practice-changing results. Master protocols and platform designs are developed in detail at Section 13.6.
9.3 Interim Analyses and Group-Sequential Designs
Many Phase III trials include one or more pre-planned interim analyses, reviewed by the DSMB. The point is to catch harm early enough to stop the trial or modify it. Two other findings justify an interim look: efficacy convincing enough that continued randomization to the control group becomes unethical, and futility, where accumulating data make a positive outcome at final analysis implausible.
Each interim look gives the trial an additional chance to cross the significance threshold by chance. A trial with one interim analysis and a final analysis, both tested at the conventional \(\alpha = 0.05\), has an overall Type I error rate of roughly 8%. Group-sequential designs solve this by pre-allocating the significance budget across analysis time points, so the cumulative probability of a false positive never exceeds the target \(\alpha\).
The O’Brien-Fleming boundaries allocate very little alpha to early analyses, requiring near-overwhelming evidence to stop, and more to later ones (O’Brien and Fleming 1979); the conservatism at early time points reflects the greater uncertainty when fewer patients have been followed. Pocock boundaries apply the same nominal threshold at every look, allowing earlier stopping but requiring a more stringent final threshold (Pocock 1977). The Lan-DeMets generalization allows these boundaries to be expressed as a continuous function of the information fraction, so that unplanned or irregularly spaced interims can still be accommodated without inflating error rates, provided their timing is not chosen in response to the accumulating treatment comparison (Lan and DeMets 1983; U.S. Food and Drug Administration 2019a). The full technical treatment is at Section 11.5.
Group-sequential boundaries control Type I error; they do not guarantee that an early stop reflects the true treatment effect. When a trial stops early for efficacy, the observed effect size is typically inflated relative to the true effect because extreme results are more likely to cross a stopping threshold when sample size is small. This overestimation bias is relevant to how results from stopped trials should be interpreted and communicated.
Futility stopping asks the opposite question. Efficacy boundaries ask whether the evidence is strong enough to conclude benefit; futility rules ask whether it is weak enough that continuing is not worth the cost. Conditional power and predictive probability of success are the two common metrics, described at Section 12.7.1. Futility rules are often specified as nonbinding guidance rather than strict triggers, since the costs of a false futility stop (abandoning an effective drug) are high and the statistical uncertainty at interim points can be substantial.
9.4 The Two-Pivotal-Study Paradigm and the Single-Trial Standard
Regulatory agencies have traditionally interpreted substantial evidence of effectiveness, as required under the Federal Food, Drug, and Cosmetic Act, to mean at least two adequate and well-controlled clinical trials (U.S. Food and Drug Administration 2019b). A single trial reaching statistical significance could be a false positive. Replication across independent trials, patient populations, and sets of investigators provides meaningful protection against this. Regulators evaluate each pivotal trial at a one-sided \(\alpha = 0.025\) (equivalent to a two-sided 0.05), so two independent trials that both succeed have a combined false-positive probability of roughly \(0.025 \times 0.025 = 0.000625\), a much higher evidentiary standard.
The FDA Modernization Act of 1997 codified a formal exception (United States Congress 1997): FDA may accept a single pivotal trial when accompanied by confirmatory evidence. The strongest single-trial cases rest on a trial that is large, multicenter, and internally consistent across pre-specified subgroups, with a hard clinical outcome as the primary endpoint (death, major cardiovascular event) rather than a surrogate, a p-value substantially below the conventional 0.05 threshold (typically \(p < 0.001\) or even \(p < 0.0001\) in recent large outcomes trials), and supportive evidence from mechanistic studies, surrogate endpoint data, or other development programs.
Large cardiovascular outcome trials meeting these criteria have supported single-trial submissions, and some oncology approvals in rare diseases and serious unmet-need settings have relied on single trials with strong effect sizes. For a sponsor, one very large, high-quality trial with a hard primary endpoint can therefore be more regulatory-efficient than two moderately sized studies, provided the effect size and evidence quality are sufficient.
In June 2026, FDA issued a revised draft guidance on demonstrating substantial evidence of effectiveness that moves further in this direction (U.S. Food and Drug Administration 2026). The revision repositions a single adequate and well-controlled trial supported by confirmatory evidence from an exception toward the default expectation, as FDA’s leadership has framed it (Prasad and Makary 2026), and it drops the earlier framing that treated one such trial as equivalent to two independent trials. Because it remains a draft (the public comment period closed on September 22, 2026) rather than final policy, sponsors should read it as a signal of regulatory direction rather than a settled standard when planning pivotal programs.
9.5 Safety Evaluation in Phase III
Phase III substantially expands the safety database. For drugs intended for long-term treatment of non-life-threatening conditions, the ICH E1 guideline recommends a minimum extent of population exposure to characterize the safety profile (Table 9.2) (International Council for Harmonisation 1994):
| Exposure Metric | ICH E1 Recommendation |
|---|---|
| Total patients exposed | About 1500, including short-term exposure |
| Exposed for \(\geq 6\) months | 300-600 patients |
| Exposed for \(\geq 12\) months | At least 100 patients |
Even with 1500 patients exposed, the safety database is not expected to characterize events occurring in fewer than 1 in 1000 patients (0.1%) (International Council for Harmonisation 1994): at exactly that rate only 1.5 cases are expected, too few to estimate the risk with any precision, and much rarer events may not appear at all. The “rule of 3” states that if zero events are observed in N patients, the upper 95% confidence limit for the true rate is approximately 3/N. Thus, 1500 patients provides only 95% confidence that the true rate is below 0.2%.
That exposure is necessary to detect adverse events occurring infrequently or developing only after extended use. The events it still misses before approval are why post-marketing surveillance remains important.
9.6 Data Safety Monitoring Boards
Given their size and duration, many Phase III trials are overseen by a Data Safety Monitoring Board (DSMB), also known as a Data Monitoring Committee (DMC) or Independent Data Monitoring Committee (IDMC): a group of independent experts who periodically review unblinded data to assess safety and, in some cases, efficacy.
The DSMB may recommend that a trial continue as planned, that it be modified (by changing eligibility criteria, say, or stopping a specific treatment arm), or that it be stopped early for any of the reasons interim analyses exist to detect: harm, clear benefit, or futility.
The operational difficulty is keeping the sponsor and investigators blinded while the DSMB works from unblinded data. A formal charter describes the board’s responsibilities, composition, and operating procedures.
9.7 Global Considerations
Phase III trials are increasingly global, with sites across North America, Europe, Asia, Latin America, and elsewhere. Globalization speeds enrollment, gives access to treatment-naive populations, and allows efficacy to be demonstrated across diverse populations. The complications are regulatory and clinical at once: requirements differ across regions and investigators must ensure compliance with all applicable regulations, while differences in medical practice, standard of care, and patient populations may affect results. The ICH E5 guideline addresses the acceptability of foreign clinical data and the concept of bridging studies to assess whether results from one region apply to another (International Council for Harmonisation 1998).
For pivotal trials specifically, representative enrollment is now a U.S. regulatory expectation: under the Food and Drug Omnibus Reform Act of 2022 (FDORA), sponsors are directed to submit Diversity Action Plans for Phase III and other pivotal studies that begin enrolling more than 180 days after FDA publishes final guidance on the plans (U.S. Congress 2022). The Diversity Action Plans implementing guidance has been unsettled since early 2025: FDA briefly withdrew and, under court order, restored its draft, then let the statutory deadline for a final version pass, so the plan requirement has not yet taken effect (Grossi 2025). Separately, FDA’s December 2025 final guidance on trial participation (a revision of its earlier eligibility-criteria guidance) keeps the representative-enrollment recommendations while dropping the diversity framing (U.S. Food and Drug Administration 2025). The scientific rationale is unchanged; the operational mechanics of meeting enrollment goals are covered at Chapter 18.
9.8 Example Phase III Protocol Structure
A Phase III protocol follows the structure of earlier phases with greater detail and rigor. Background summarizes earlier clinical development; objectives state the primary and secondary endpoints; the design section describes the randomization scheme, blinding procedures, treatment duration, and visit schedule; subject selection provides the detailed inclusion and exclusion criteria designed to enroll a population representative of the intended use while ensuring patient safety; treatment procedures specify exactly how the investigational product and comparator are to be administered, stored, and accounted for; and the efficacy and safety assessment sections detail how those data are collected and analyzed.
The statistical section carries particular weight: analysis populations, primary analysis method, sample size justification, interim analyses (if planned), and methods for handling multiplicity.
A Phase III that succeeds produces the definitive evidence behind the label, and with it the information prescribers need and the therapy patients eventually receive.