The Poisson Distribution for Counting Rare Events

When events are rare, independent, and steady

The Poisson distribution models the number of times an event occurs within a fixed interval of time, space, or another continuous measure, when three conditions roughly hold: events happen independently of each other, they occur at a constant average rate over the interval, and two events cannot occur at exactly the same instant. Under these assumptions, the count of events X follows a Poisson distribution with a single parameter λ (lambda), the average number of events per interval.

Typical examples include the number of customers arriving at a shop counter in an hour, the number of typos on a printed page, the number of cosmic ray particles hitting a detector per minute, and the number of insurance claims filed in a region per month. What unites them is not that the events are rare in an absolute sense, but that they are individually unpredictable, roughly independent, and occur against a steady background rate.

The Poisson probability mass function

The probability of observing exactly k events is

P(X = k) = e^−λ · λ^k / k!

where e is Euler's number (≈ 2.71828) and k! is k factorial. Unlike the binomial distribution, which requires knowing both a number of trials n and a per-trial probability p, the Poisson distribution needs only λ, which conveniently equals both the mean and the variance of the distribution: E[X] = Var(X) = λ. This single-parameter simplicity is exactly why the Poisson model is so widely used — a single count from historical data (the average rate) is enough to fully specify the entire probability distribution.

Poisson as a limit of the binomial distribution

The Poisson distribution is not an arbitrary alternative to the binomial — it is its limiting case. Imagine subdividing a fixed time interval into an enormous number of tiny sub-intervals n, each so short that at most one event could plausibly occur in it, with a small probability p of an event in each sub-interval. The total count over the full interval is then binomial with parameters n and p. As n → ∞ and p → 0 while their product np is held fixed and equal to λ, the binomial probability mass function converges exactly to the Poisson formula above. This is why the Poisson distribution is the natural model whenever an event could in principle happen at any of a huge number of possible moments, each individually unlikely, but the total expected count over the interval is a modest, known number λ.

As a practical rule of thumb, when n is large (roughly n ≥ 100) and p is small (roughly p ≤ 0.01) with np = λ moderate, the Poisson distribution provides a very close and computationally simpler approximation to the binomial, avoiding the need to compute large binomial coefficients.

Mean equals variance: a useful diagnostic

Because a genuine Poisson process forces the mean and variance to be exactly equal, real datasets are often checked against this property as a first diagnostic test. If the sample variance is noticeably larger than the sample mean, a condition called overdispersion, that is a signal that the simple independence and constant-rate assumptions behind the Poisson model are being violated — perhaps events cluster together (a single hospital outbreak generating many related admissions, rather than independent ones), or the underlying rate itself varies over time or across the population being studied. In those cases, statisticians typically switch to a negative binomial distribution, which has an extra parameter specifically to accommodate that excess variance.

Applications: call centres, decay, and hospital admissions

The Poisson distribution's most classic application is queueing theory: if customer arrivals at a call centre or checkout line genuinely follow a Poisson process with rate λ calls per minute, staffing models can compute the probability of any given number of simultaneous callers, directly informing how many staff are needed to keep wait times acceptable. In physics, radioactive decay counts follow a Poisson distribution almost exactly, since each atom decays independently of every other atom and the population is enormous, making it one of the cleanest real-world Poisson processes available for teaching. In healthcare, the number of emergency department admissions in a given hour, or the number of rare adverse drug reactions reported per month across a population, are commonly modelled as Poisson, which then feeds directly into staffing decisions and pharmacovigilance signal detection — flagging when an observed count is high enough to be statistically surprising under the assumed background rate.

How the shape changes as lambda grows

For small λ (say λ = 0.5), the distribution is heavily right-skewed, with "zero events" as by far the most likely outcome and a rapidly thinning tail. As λ increases, the distribution's peak shifts rightward and the shape becomes progressively more symmetric. By roughly λ ≥ 10, the Poisson distribution is already visually close to a normal distribution with mean λ and standard deviation √λ — another instance of the same central-limit-style convergence toward the bell curve that appears throughout probability theory, since a Poisson(λ) random variable can itself be constructed as the sum of many independent Poisson variables with smaller rates.