| Phase | Typical Sample Size (N) | Expected Events | Detection Power P(≥1) | Surveillance Status |
|---|
Each dot represents a vaccinated individual in a scaled clinical micro-cohort. Green indicates prevented severe disease cases, blue indicates healthy protected recipients, and red indicates observed serious adverse reactions.
| Clinical Metric | Value per 1,000,000 | Interpretation |
|---|---|---|
| Baseline Disease Cases | 12,000 | Without vaccination / intervention |
| Prevented Severe Cases | 11,040 | Cases averted via efficacy |
| Expected Rare Adverse Events | 25 | True rate: 2.5 / 100k |
| Net Prevented Burden | +11,015 | Lives protected minus severe reactions |
| Number Needed to Treat (NNT) | 91 | Patients treated to prevent 1 disease case |
| Number Needed to Harm (NNH) | 40,000 | Patients treated before 1 adverse event |
Why do randomized clinical trials require millions of real-world recipients in Phase IV pharmacovigilance to detect rare adverse events?
Clinical trials evaluate efficacy and common safety outcomes, but detecting rare adverse reactions requires substantial participant cohorts. The probability of observing at least one event follows the binomial model P(k ≥ 1) = 1 - (1 - p)^N. Under the biostatistical Rule of Three, ruling out an adverse event rate p with 95% confidence when zero events are seen requires a sample size of N ≈ 3 / p. A pivotal Phase III trial of 40,000 individuals has roughly 63.2% power to detect an event occurring at 1 in 40,000 (2.5 per 100,000), leaving a 36.8% chance of observing zero cases prior to regulatory approval.
The on-page 250-dot cohort simulation applies visual scaling multipliers (50x for adverse events and 4x for prevented cases) solely for demonstration rather than literal event frequencies. Additionally, the population ledger performs unadjusted linear subtractions that assume zero background incidence of the adverse event, uniform exposure, and independent binomial trials without confounding or competing risks.
Click the 'Acute Therapeutic' preset button. The True Adverse Event Rate shifts to 45.0 per 100k (p = 0.00045, or 1 in 2,222) and the Phase III slider updates to 7,500 due to its 2,500 step increments. The Rule of 3 Sample Needed recalibrates to 6,667 (3 / 0.00045), and Phase III Detection Power increases to 96.58% (1 - (1 - 0.00045)^7500).
When a rare event with incidence rate p occurs 0 times across N independent subjects, the probability of zero occurrences is (1 - p)^N. Setting this non-detection probability to an upper alpha risk of 5% yields (1 - p)^N = 0.05. Taking natural logarithms gives N · ln(1 - p) = ln(0.05) ≈ -2.9957. Using the first-order Taylor approximation ln(1 - p) ≈ -p for very small p yields -N · p ≈ -3, which simplifies directly to N ≈ 3 / p.
Pre-approval clinical trial databases face inherent sample size limitations and cannot rule out ultra-rare adverse drug reactions before licensing. Regulatory guidance on pharmacovigilance planning outlines the necessity of prospective safety specifications and post-approval observational surveillance to monitor real-world populations across millions of doses. ICH E2E Pharmacovigilance Planning Guidance for Industry
Clinical safety database limitations regarding sample size and rare adverse drug reactions, and the structure of post-approval pharmacovigilance.
The page converts entered two point five events per one hundred thousand to probability point zero zero zero zero two five. Assuming independent events at a constant probability, no event in forty thousand trials has probability one minus p raised to forty thousand. Subtracting that from one gives about sixty three point two percent for at least one event. The bars compare this probability with its thirty six point eight percent complement at four pixels per percentage point. This is a mathematical scenario, not proof that a clinical event occurred, was caused by a treatment, or reflects a verified real event rate. At ninety five percent the source required sample branch uses three divided by p, rounded, giving one hundred twenty thousand. That is a rule of three approximation, not the exact solution obtained by dividing logarithms of no event probability. Its other confidence choices use saved logarithmic factors. The bars compare forty thousand and one hundred twenty thousand samples at point zero zero three pixels per sample. Expected count at forty thousand is one, but observing at least one remains only around sixty three percent. Detection probability also does not establish causality, severity, statistical comparison power or suitability of a real study design. The stochastic display samples two hundred fifty dots, each with the entered event probability. At point zero zero zero zero two five its expected number of rare event dots is point zero zero six two five, so most illustrative displays have none. The million person table independently multiplies that same assumed rate by one million, producing twenty five expected events. The bars compare two hundred fifty display dots with an illustrative one thousand trials at point four pixels per trial. A dot display without red dots therefore cannot negate the separate expected count. These outputs are simulated arithmetic and do not support a clinical recommendation or validate actual safety claims.