Guest Column | November 18, 2026

When Enrollment Data Says Your Forecast Is Wrong

By Uchemadu Nwachukwu

Woman analyzing data-GettyImages-2292822696

The forecast that matters is not the one signed off at protocol approval. It is the one that survives contact with actual sites, actual patients, and the pace of recruitment.

A trial’s enrollment forecast can look precise and still be wrong in the ways that matter: it may assume sites open together, patients arrive at a steady rate, and the eligible population is as accessible as it appears on paper. Then recruitment begins. Site openings stagger, first-patient timelines stretch, and the curve falls below plan. At that point, the right question is not whether the original forecast was reasonable. It is what the new data say about the remaining timeline — and what decisions should change as a result.

The scale of the planning problem is not trivial. In a study of 4,004 cancer trials registered on ClinicalTrials.gov, just 49.1% achieved actual enrollment of at least 85% of their target. The proportion meeting that threshold fell to 36.4% for trials activated in 2017 to 2019, down from 51.5% in 2008 to 2010.1 That is not a universal failure rate for every therapeutic area or sponsor. It is, however, a warning against treating feasibility as a settled fact rather than a testable hypothesis.

A Plan Is A Prior, Not A Promise

At planning, forecasts necessarily rely on incomplete information: the target population, proposed sites and countries, historical recruitment experience, and assumptions about start-up and enrollment. A simple calculation — target sample size divided by the number of sites and an average enrollment rate — can be useful as a first pass. But it assumes a stable rate across sites and time, and offers no uncertainty interval. Real recruitment is a sequence of distinct processes: country start-up, site activation, and patient enrollment. Each can vary, and each affects the final date.2

Historical data can improve the starting estimate, particularly when they are local and relevant to the proposed trial. In a retrospective analysis across 16 U.S. cancer centers, center-specific models predicted accrual at 99% of actual accrual and correctly classified whether a study would recruit at least four participants 73% of the time; disease-team estimates achieved 58% on that classification task.3 These results make a case for using institutional experience, not for assuming that one model will transport perfectly to another indication, protocol, site mix, or period.

A forecast should therefore be expressed as a distribution of plausible outcomes, not a single date with false precision. Zhong et al. describe a model that represents start-up, site activation, and participant enrollment sequentially and generates prediction intervals; in a selected set of 25 studies across four therapeutic areas, its median absolute error was 2.8 months, compared with 5.6 months for the prior internal approach. Its 95% prediction interval covered the observed duration in 92% of cases, while the comparison approach’s interval covered only 20%.2 The study is a useful demonstration, not a universal benchmark: it used a small, selected validation set and company-specific operational data.

The First Enrollment Curve Is New Evidence

Once the trial is recruiting, actual data should update the forecast. The initial estimate is not invalid simply because reality diverges; it is conditional on assumptions that can now be replaced by evidence. Cumulative enrollment, activated-site count, activation dates, and the lag from site opening to first participant can show whether the shortfall stems from start-up, site productivity, or both. The monitoring question changes from “How long did we plan?” to “Given what has happened so far, how long is left, and what is the probability we miss the target date?”4

That update must be more than extending a straight line from the first few weeks. Recruitment can accelerate as more sites open — or slow as the easiest-to-reach patients are enrolled. A useful forecast distinguishes sites that are not yet recruiting from sites that are open but not enrolling, and uses uncertainty intervals wide enough to reflect between-site and between-country variation. Real-time prediction methods have been developed precisely to use accumulating trial data to improve predictions, while the broader evidence base has historically had limited large-scale validation.4,5

The signal is not only the slope. A registry analysis found recruitment was associated with design features, sponsor experience, eligibility criteria, and competition from other trials. Trials with fewer than three inclusion criteria averaged 5.52 participants per week — about three times the rate of trials with more than eight — although this observational association does not establish that reducing criteria alone causes faster recruitment. More overlapping trials were also associated with lower recruitment rates and higher termination rates.6 When actual enrollment misses plan, these factors should prompt a diagnosis, not a reflexive instruction to “add sites.” More sites may help only if the bottleneck is reach or capacity; if site activation is delayed, eligibility is too restrictive, or competing studies are drawing from the same pool, the intervention should address that mechanism.

Make The Forecast Decision-Ready

A practical reforecast should answer four questions at each review point:

  1. What has changed? Compare observed site activation and recruitment trajectories with the planning assumptions, not only the aggregate enrollment total.
  2. What is the likely finish date now? Report a central estimate alongside a prediction interval and the probability of meeting the committed date. Do not convert a wide range into a falsely confident point estimate.
  3. Why is the forecast moving? Separate country start-up, site activation, time to first participant, and enrollment rate at active sites. The levers differ.
  4. What decision follows? Predefine thresholds for investigating sites, revising eligibility or outreach, opening additional sites, adjusting the timeline, or stopping recruitment. Tie each action to evidence and a review date.

The most useful forecast is not necessarily the most complex model. It is the one calibrated on data that resemble the current trial, updated on a defined cadence, transparent about uncertainty, and connected to decisions. A detailed model built on noncomparable historical trials can create the appearance of rigor without reliable prediction. Conversely, a modest model with well-maintained site-level data may reveal early that the planned trajectory is no longer plausible.

There is still a meaningful evidence gap: recruitment forecasting methods have not been consistently tested across large, diverse trial data sets, and published performance in one setting does not guarantee performance elsewhere. The RECRUIT-IT investigators specifically noted the lack of systematic empirical validation of existing models in their 2020 protocol.4 That makes calibration and local validation part of the work — not optional technical polish.

When enrollment data contradict the forecast, the forecast has done something valuable: it has exposed a decision that needs to be revisited. The failure is not that a plan met uncertainty. The failure is continuing to manage against an outdated estimate after the trial has supplied better evidence.

References

  1. Zhang S, Zhang JR, Liu S, Pang H, Stinchcombe T, Wang XF. Enrollment success, factors, and prediction models in cancer trials (2008–2019). JCO Oncology Practice. 2023;19(12):e1850–e1858.
  2. Zhong S, Xing Y, Yu M, Wang L. Enrollment forecast for clinical trials at the portfolio planning phase based on site-level historical data. Statistics in Biopharmaceutical Research. 2023. doi:10.1080/19466315.2022.2159324.
  3. Tate WR, Abraham I, Cranmer L. Center-specific modeling predicts cancer trial accrual more accurately than investigators and random effects modeling at 16 cancer centers. J Oncol Pract. 2019;15(8):e688–e696.
  4. Kasenda B, Liu J, Jiang Y, Gajewski B, Wu C, von Elm E, et al. Prediction of RECRUITment In randomized clinical Trials (RECRUIT-IT)—rationale and design for an international collaborative study. Trials. 2020;21:530. doi:10.1186/s13063-020-04666-8.
  5. Heitjan D, Ge Z, Ying G. Real-time prediction of clinical trial enrollment and event counts: a review. Contemp Clin Trials. 2015;40:185–195.
  6. Butler AM, Li Z, Liu C, Ta CN, Yuan C, Chen H, et al. Finding success in clinical trial recruitment: a trial registry analysis. Research Square [preprint]. 2019. doi:10.21203/rs.2.10112/v1.

About The Author:

Uchemadu Nwachukwu is a public health data analyst, health informatics professional, and clinical data management professional with over eight years of experience across healthcare analytics, clinical research, health informatics, and real-world evidence. He holds a First-Class MSc in applied clinical data analytics from the University of Galway and has expertise in statistical modeling, machine learning, epidemiological analysis, healthcare data management, interoperability, and clinical information systems. His work focuses on applying data, technology, and analytical methods to support clinical research, healthcare improvement, and evidence-based decision-making. His professional interests include digital health, health informatics, clinical research, healthcare analytics, epidemiology, and data-driven approaches to improving healthcare systems and outcomes.