Use the binomial distribution when you’re counting successes out of a fixed number of trials. Use the Poisson distribution when you’re counting how many times something happens over a stretch of time or space, with no fixed number of trials. That single distinction drives the choice in most practical situations, but the details matter when your data sits in a gray area.
What the Binomial Distribution Requires
The binomial distribution models a specific setup: you run a fixed number of independent trials, each with two possible outcomes (success or failure), and the probability of success stays the same every time. All four conditions must hold.
Think of flipping a coin 20 times and counting heads. You know in advance there will be exactly 20 flips. Each flip has the same probability. The result of one flip doesn’t change the next. The thing you care about is how many of those 20 flips came up heads. That’s a textbook binomial scenario. In clinical research, the same logic applies: enroll a fixed group of patients, give them a treatment, and count how many respond. Each patient either responds or doesn’t, and the number of “successes” out of the fixed group follows a binomial distribution.
The two parameters you need are n (the number of trials) and p (the probability of success on each trial). If either one is undefined for your problem, the binomial isn’t the right fit.
What the Poisson Distribution Requires
The Poisson distribution counts events that happen at some average rate across a continuous interval of time, area, or volume. Instead of “how many successes in n trials,” the question becomes “how many events in this window.” There’s no cap on the count, and there’s no defined number of trials.
Three conditions define the Poisson setup. Events in non-overlapping intervals must be independent of each other. The probability of an event depends only on the length of the interval, not on when you start watching. And two events can’t happen at exactly the same instant. The distribution has a single parameter: the average rate of occurrence multiplied by the length of your observation window.
Classic examples include the number of calls a help desk receives per hour, the number of typos per page, or the number of car accidents at an intersection per month. In epidemiology, the CDC uses Poisson models when tracking rare conditions like traumatic brain injuries across states, because the cases are sparse, independent, and spread across large populations and time periods.
A Quick Decision Framework
Ask yourself two questions about your data:
- Do you have a fixed number of trials? If yes, and each trial has a yes/no outcome with constant probability, use the binomial. Examples: 50 coin flips, 200 patients in a drug trial, 30 items on a quality inspection line.
- Are you counting events in a continuous interval? If yes, and events occur independently at a roughly constant rate, use the Poisson. Examples: emails per day, mutations per DNA segment, emergency room visits per week.
If you find yourself unable to define what a single “trial” looks like, that’s a strong signal you’re in Poisson territory. When you count the number of potholes on a mile of road, there’s no meaningful concept of a “trial” that could produce a pothole. The events just occur along a continuum.
When Poisson Approximates Binomial
The two distributions aren’t unrelated. When you have a very large number of trials and a very small probability of success, the binomial starts to look almost identical to a Poisson. A common rule of thumb: if n is at least 100 and the product n times p is 10 or less, a Poisson distribution with its rate parameter set to n times p gives an excellent approximation of the binomial.
This comes up more often than you’d think. Suppose you manufacture 10,000 components per day and the defect rate is 0.0003. Technically, each component is a trial with a success/failure outcome, so the situation is binomial. But with n = 10,000 and p = 0.0003, the expected number of defects per day is just 3. The Poisson model with an average rate of 3 gives nearly identical probabilities and is far simpler to work with. You only need one parameter instead of two, and you avoid calculating enormous factorials.
This approximation is why many “rare event” problems get modeled as Poisson even when, strictly speaking, they involve a finite number of trials. It’s not wrong. It’s a well-justified shortcut.
Where Each Distribution Breaks Down
The binomial falls apart when its assumptions crack. The most common violation is non-independence: if the outcome of one trial changes the probability for the next, the binomial won’t give you accurate results. Drawing cards from a deck without replacement is a familiar example. Each draw changes the composition of the deck, so the probability shifts. For small samples from a large population (say, sampling 10 items from a warehouse of 100,000), the shift is negligible and the binomial still works. For larger sampling fractions, the hypergeometric distribution is more appropriate.
The other assumption to watch is constant probability. If your “success rate” varies across trials because some subgroups respond differently, a single binomial model will underestimate the spread of your results.
The Poisson has its own weak point: it assumes the mean and variance of the count are equal. In real data, the variance often exceeds the mean, a pattern called overdispersion. This can happen when events cluster (disease outbreaks aren’t evenly spread across weeks) or when an important variable has been left out of the model. When overdispersion is present, the Poisson model will make you overconfident in your estimates, producing standard errors that are too small. The typical fix is a negative binomial model, which adds a parameter to handle the extra variance. Cornell’s statistical consulting unit notes that overdispersion should only be addressed after confirming the model is correctly specified, because a poorly structured model can mimic overdispersion even when the data actually follows a Poisson process.
Concrete Examples Side by Side
Seeing matched examples often makes the distinction click faster than definitions alone.
You survey 500 people and ask whether they’ve been diagnosed with a particular condition. Each person is a trial, the outcome is yes or no, and you want to know how many out of 500 say yes. That’s binomial, with n = 500 and p equal to the prevalence rate.
Now suppose you’re monitoring a hospital’s emergency department and counting how many patients arrive with that same condition per week. There’s no fixed number of “trials.” People show up (or don’t) continuously. The count per week follows a Poisson distribution, with the rate parameter reflecting average weekly arrivals.
A software team tests 1,000 lines of code and flags each line as containing a bug or not. Binomial. The same team counts the total number of bugs found in a codebase over a month, without defining individual trials. Poisson.
In each pair, the data might even produce similar numbers. The difference is structural: whether your counting process is bounded by a known number of trials or open-ended across a continuous window.