The Poisson distribution describes the probability of a given number of independent events occurring in a fixed interval of time, space, or some other continuous unit — when those events happen at a known, constant average rate. It is the standard model for count data: the number of citations a paper receives in a year, the number of adverse events reported per clinical trial site, the number of mutations per DNA segment, the number of support tickets per hour, the number of typos per page. If your outcome variable is a non-negative integer (0, 1, 2, 3…) counting how many times something happened, the Poisson distribution is usually the first model to test.
The formula
The probability of observing exactly k events is:
P(X = k) = (λk · e−λ) / k!
Where:
- λ (lambda) is the single parameter of the distribution — the average (expected) number of events in the interval.
- k is the specific count you’re finding the probability for (0, 1, 2, 3…).
- e is Euler’s number (≈ 2.71828).
- k! is k factorial.
Unlike the normal distribution, which needs two parameters (mean and standard deviation) to be fully specified, the Poisson distribution needs only one: λ. That single parameter is simultaneously the distribution’s mean and its variance — a defining property covered below.
The defining property: mean equals variance
For a true Poisson-distributed variable:
E(X) = Var(X) = λ
This is called equidispersion, and it is the single most useful diagnostic check when deciding whether count data actually fits a Poisson model. In practice, sample variance in real count data is very often larger than the sample mean — a condition called overdispersion — which signals that a plain Poisson model will understate uncertainty and produce overconfident (too-narrow) standard errors and confidence intervals. See the section on overdispersion below for what to use instead when that happens.
The four assumptions
A count variable is well-modeled by a Poisson distribution only when all four of these hold:
- Events are counted over a fixed, well-defined interval. Time (per hour, per year), space (per square meter, per DNA segment), or another fixed unit (per manuscript, per site, per patient-year).
- Events occur independently. One event happening does not change the probability of another event happening in the same interval. (This is frequently violated in practice — see the counter-example below.)
- The average rate (λ) is constant over the interval. If the underlying rate changes systematically — e.g., citation rate rising sharply after a paper wins an award — a single Poisson model across the whole period is misspecified.
- Two events cannot occur at literally the same instant (formally called “orderliness”). In continuous time this is a technical requirement rather than a practical concern for most research data.
Worked examples
Example 1: Citations per paper
A department finds that papers in a given sub-field average λ = 3 citations in their first year. What is the probability that a randomly selected paper receives exactly 5 citations?
P(X = 5) = (35 · e−3) / 5! = (243 × 0.0498) / 120 ≈ 0.1008
About a 10.1% chance. The same formula gives the probability of zero citations (P(X=0) = e−3 ≈ 0.0498, about 5%), which is a common and useful sanity check when reporting Poisson-based results: it tells you what share of the population is expected to have no events at all.
Example 2: Adverse events per trial site
A multi-site clinical trial logs adverse events at a known average rate of λ = 1.5 per site per month. The probability of a site reporting zero adverse events in a given month is e−1.5 ≈ 0.223, or about 22.3%. A site logging 6 or more in one month is a low-probability event under this model (well under 2%) and is exactly the kind of tail observation that would prompt a data-quality or safety review in clinical trial monitoring.
A counter-example: when count data is NOT Poisson
Number of retractions per journal issue looks like classic Poisson count data at first glance — a non-negative integer counted over a fixed interval. But retractions frequently violate the independence assumption: a single misconduct investigation at one lab can produce several retractions from the same journal in a short window, all triggered by the same underlying cause rather than occurring independently. That clustering inflates the variance well above the mean (overdispersion) and means a plain Poisson model will misrepresent the true uncertainty. The same logic applies to disease outbreak counts, citation cascades after a paper goes viral, and equipment failures with a common root cause — anywhere one event mechanically raises the probability of another, treat Poisson as a starting hypothesis to test, not an assumption to take for granted.
Poisson vs. binomial distribution
Both distributions model counts, and they are mathematically related, but they answer different questions:
| Poisson | Binomial | |
|---|---|---|
| Question it answers | How many events occur in a fixed interval, when the number of opportunities is effectively unbounded? | How many successes occur in a fixed, known number of independent trials (n)? |
| Parameters | One: λ (mean rate) | Two: n (number of trials) and p (probability of success per trial) |
| Typical use | Counts per unit time/space with no natural upper bound (citations, defects, arrivals) | Counts with a known denominator (successes out of n survey respondents, n coin flips) |
| Relationship | The Poisson distribution is the limiting case of the binomial distribution as n becomes very large and p becomes very small, with n×p held constant and equal to λ. This is why Poisson is the natural model for “rare events with many opportunities” — e.g., a rare adverse reaction across a large patient population. | |
Poisson regression
When you want to model how a count outcome depends on one or more predictor variables — e.g., number of publications as a function of career stage, funding level, and institution type — ordinary linear regression is the wrong tool: it assumes a continuous, normally distributed, unbounded outcome, and will happily predict a negative number of publications, which is meaningless. Poisson regression is the generalized linear model built for this: it uses a log link function to keep predicted counts non-negative and models the outcome’s expected value as a function of the predictors.
Poisson regression carries the same equidispersion assumption as the base distribution — the model assumes conditional variance equals the conditional mean. Researchers report coefficients as incidence rate ratios (IRRs) (exponentiated coefficients), analogous to how logistic regression coefficients are reported as odds ratios.
Overdispersion: when Poisson isn’t enough
Real count data is overdispersed far more often than it is cleanly equidispersed. Standard diagnostics: compare the residual deviance to its degrees of freedom (a ratio well above 1 signals overdispersion), or directly compare sample variance to sample mean. When overdispersion is present, two standard alternatives are used instead of plain Poisson regression:
- Negative binomial regression — adds a dispersion parameter that lets variance exceed the mean, and is the most common fix for overdispersed count data in practice.
- Quasi-Poisson regression — keeps the Poisson mean structure but corrects the standard errors for the extra variance, without fully modeling a separate distribution.
A related but distinct problem is zero-inflation — when a dataset has far more zero counts than a Poisson (or negative binomial) model would predict, because a separate process is generating “structural” zeros (e.g., researchers who could never have an adverse event because they were never enrolled long enough to be at risk). Zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) models handle this by explicitly modeling the excess-zero process separately from the count process.
The Poisson process and inter-arrival times
The Poisson distribution counts how many events happen in an interval. Its close relative, the exponential distribution, describes how long you wait between consecutive events in the same underlying process (a “Poisson process”). Both share the same rate parameter and the same independence assumption — if event counts per interval are Poisson-distributed with rate λ, the time between successive events is exponentially distributed with the same λ. This pairing shows up often in reliability and survival-type analyses of equipment failures, patient arrivals, or system outages.
Frequently asked questions
What does the lambda (λ) parameter mean in a Poisson distribution?
Lambda is the average number of events expected in the fixed interval you’re measuring. It is both the mean and the variance of the distribution, and it’s the only parameter you need to fully specify a Poisson distribution — unlike a normal distribution, which needs both a mean and a standard deviation.
Can lambda be a decimal, like 2.5?
Yes. Lambda is an average rate, so it is very often not a whole number (e.g., 2.5 adverse events per site per month on average), even though any individual outcome (X) the distribution predicts is always a non-negative integer.
How do I know if my count data is Poisson-distributed?
Check whether the sample variance is approximately equal to the sample mean (the equidispersion property). If variance is meaningfully larger than the mean, the data is overdispersed and negative binomial or quasi-Poisson regression is usually more appropriate than plain Poisson. Also check the independence assumption conceptually — do you have a mechanism by which one event could trigger or cluster with another?
What’s the difference between Poisson regression and linear regression for count outcomes?
Linear regression assumes a continuous, unbounded, normally distributed outcome and can predict impossible negative counts. Poisson regression uses a log link function so predictions stay non-negative, and it models the outcome according to the Poisson (or, when overdispersed, negative binomial) distribution rather than the normal distribution.
Is the Poisson distribution the same as a Poisson process?
Related but not identical. The Poisson distribution is a probability distribution describing counts in a single fixed interval. A Poisson process is the underlying stochastic model — events occurring continuously and independently over time or space at a constant rate — from which the Poisson distribution (for counts) and the exponential distribution (for inter-arrival times) are both derived.
Related reading
- Normality of Distribution — how to check the normal-distribution assumption, the usual alternative for continuous rather than count data.
- Sampling Methods — probability and non-probability sampling approaches for the data you’re modeling.
- Regression Analysis — choosing between linear, logistic, and other regression models, including when a generalized linear model like Poisson regression is the right fit.
- Logistic Regression — the GLM for binary outcomes, useful comparison for how link functions and exponentiated coefficients work in Poisson regression.
- Chi-Square Test — another test frequently used with categorical/count data.
- What Is a P Value? — interpreting significance for the hypothesis tests built on these distributions.
