Poisson Counts and the Rate Parameter

Counts, Windows, and Parameters

A count is the number of times an event occurs in a specified observation window: for example, the number of calls received in an hour or the number of defects found in a batch. A probability model for such data must assign probabilities to whole-number counts, including zero. The Poisson distribution is a standard model of this kind. Its parameter describes the expected count for the particular window being modeled.

The distinction between a count and a rate is essential. A count is dimensionless: it is a number of events. A rate is measured per unit of exposure, such as calls per hour or defects per metre. To use a rate in a Poisson model, multiply it by the amount of exposure. The resulting expected count is the distribution’s parameter.

Definition: A random variable $X$ has a Poisson distribution with parameter $\lambda$, written $X\sim\operatorname{Poisson}(\lambda)$, if its support is $\{0,1,2,\ldots\}$ and, for $\lambda>0$, $$ P(X=k)=\frac{e^{-\lambda}\lambda^k}{k!},\qquad k=0,1,2,\ldots. $$ For $\lambda=0$, the distribution is degenerate at zero: $P(X=0)=1$. In this model, $\lambda$ is the expected count in the observation window to which $X$ refers.

The support consists of nonnegative integers because $X$ counts occurrences. It cannot be negative or fractional. The parameter satisfies $\lambda\geq 0$. In applications with $\lambda>0$, the formula assigns positive probability to every nonnegative integer, although probabilities for very large counts can be extremely small.

The parameter’s interpretation depends on what window $X$ describes. If $X$ is the number of events in one hour, $\lambda$ is the expected number in one hour. If $X$ is the number in a 10-minute window, the parameter must refer to that 10-minute window instead. The symbol alone does not supply a time interval or exposure; those must be stated as part of the model.

Verifying the Probability Model

A probability mass function must assign nonnegative probabilities whose total is one. For $\lambda>0$, each Poisson probability is nonnegative. To check the total, use the power series for the exponential function, $e^\lambda=\sum_{k=0}^{\infty}\lambda^k/k!$:

\sum_{k=0}^{\infty}P(X=k) = \sum_{k=0}^{\infty}\frac{e^{-\lambda}\lambda^k}{k!} = e^{-\lambda}\sum_{k=0}^{\infty}\frac{\lambda^k}{k!} = e^{-\lambda}e^\lambda = 1.

When $\lambda=0$, the separate degenerate definition also sums to one. Thus the formula, together with that boundary case, specifies a probability distribution. This check matters: a plausible-looking expression is not a probability model unless its probabilities are valid.

Theorem: If $X\sim\operatorname{Poisson}(\lambda)$, then $E[X]=\lambda$. For $\lambda=0$, this follows because $X=0$ with probability one. For $\lambda>0$, the expectation is $$ E[X] = \sum_{k=0}^{\infty}k\frac{e^{-\lambda}\lambda^k}{k!} = e^{-\lambda}\sum_{k=1}^{\infty}\frac{\lambda^k}{(k-1)!} = \lambda e^{-\lambda}\sum_{j=0}^{\infty}\frac{\lambda^j}{j!} = \lambda. $$ Here $j=k-1$ in the last sum, and the exponential series gives $\sum_{j=0}^{\infty}\lambda^j/j!=e^\lambda$. Therefore the parameter is the expected count in the specified window. $\square$

This identity gives a precise meaning to the parameter, but it does not mean that every observed count equals $\lambda$, or that $\lambda$ must be an integer. An expected count can be fractional: it is an average across repetitions of the same observation setup, not a possible value that a single count must take.

From a Rate to an Expected Count

Suppose events occur at a constant modeled rate $\rho$ per unit of exposure, and the observation window contains $t$ units of exposure. For time exposure, $\rho$ might be measured in events per hour and $t$ in hours. The expected count over the window is rate multiplied by exposure:

\lambda=\rho t.

Thus a rate and a Poisson parameter are not interchangeable. The rate has units; the parameter is an expected count and has no physical units. The model for the count over exposure $t$ is $X\sim\operatorname{Poisson}(\rho t)$, not $X\sim\operatorname{Poisson}(\rho)$ unless the exposure is exactly one unit or $\rho$ is already the expected count for the stated window.

This conversion also applies when exposure is not time. If defects are modeled at a constant rate per metre, then the amount of exposure could be the length inspected in metres. If events are modeled per patient-year, exposure could be the total number of patient-years observed. In each case, the units must cancel correctly: (events per exposure unit) times (exposure units) gives an expected number of events.

Worked Example: A Count Over a One-Hour Window

A service desk models the number of calls received in one hour as Poisson with expected count $6$. Let $X$ be the number of calls in that hour. Identify the support, parameter, and probability of receiving no calls.

The count can be any nonnegative integer, so the support is $\{0,1,2,\ldots\}$. The parameter is $\lambda=6$, the expected number of calls in the specified one-hour window. For zero calls, substitute $k=0$ into the probability mass function:

P(X=0)=\frac{e^{-6}6^0}{0!}=e^{-6}\approx 0.00248.

The event count is an integer, but its expected value is $6$; that does not assert that exactly six calls will arrive in every hour. The probability calculation also refers specifically to a one-hour window, not to an unspecified period.

Worked Example: Converting a Rate Per Hour

A monitoring system has a modeled event rate of $3$ events per hour. The observation window is $20$ minutes. Find the Poisson parameter for the count in that window and state its interpretation.

First express the exposure in hours: $20$ minutes is $20/60=1/3$ hour. The rate is $\rho=3$ events per hour, so

\lambda=\rho t = \left(3\ \frac{\text{events}}{\text{hour}}\right) \left(\frac{1}{3}\ \text{hour}\right) = 1\ \text{event}.

The Poisson parameter is therefore $\lambda=1$, and the count in the 20-minute window is modeled as $X\sim\operatorname{Poisson}(1)$. The parameter means one expected event per 20-minute window. It is not the rate of three events per hour, even though both quantities describe the same event process on different scales.

Worked Example: Exposure Measured by Length

A manufacturer models flaws along a material at a constant rate of $0.4$ flaws per metre. A section of $7.5$ metres is inspected. Find the expected count and write a model for the number of flaws in the section.

The exposure is $7.5$ metres, so multiplying by the rate gives

\lambda = \left(0.4\ \frac{\text{flaws}}{\text{metre}}\right) (7.5\ \text{metres}) = 3\ \text{flaws}.

If $X$ is the number of flaws in the inspected section, the Poisson model is $X\sim\operatorname{Poisson}(3)$, with support $\{0,1,2,\ldots\}$ and expected count $3$. The unit “per metre” belongs to the rate; it cancels with the exposure measured in metres.

Modeling Choices and Common Pitfalls

Specifying a Poisson model requires more than naming a parameter. State what is counted and the observation window or exposure. If a rate is supplied, identify its units and multiply by the matching exposure. These details determine which random variable is being modeled and what its parameter means.

A constant-rate description is a modeling assumption: it says the expected count scales proportionally with exposure over the window under consideration. It does not by itself prove that observed counts follow a Poisson distribution. A Poisson model may be unsuitable when event rates change substantially across the window, events occur in clusters, or other important features of the sampling situation are not represented. The rate-to-parameter conversion is meaningful only with a clearly defined exposure and a model for the count.

A frequent error is to put a rate directly into the Poisson formula. For instance, if the rate is $3$ calls per hour but the window is 20 minutes, using $\lambda=3$ treats the 20-minute interval as if it had the expected count for a full hour. The correction is to convert the window to $1/3$ hour and calculate $\lambda=3(1/3)=1$. Another error is to interpret a value such as $\lambda=2.4$ as an impossible count. The parameter is an expected value, not an observed count, so it need not be an integer.

Key takeaway: A Poisson random variable counts events and takes values $0,1,2,\ldots$. Its parameter $\lambda$ is the expected count for a specified window. When the model is expressed using a constant rate $\rho$, convert it to the window-specific parameter by multiplying by the exposure: $\lambda=\rho t$.

Check Your Understanding

For each question, identify the count window or exposure before interpreting the parameter.

A count $X$ is modeled as Poisson with parameter $4.5$. What values can $X$ take, and what does $4.5$ mean?
An event rate is $8$ per day. What is the Poisson parameter for a six-hour observation window, assuming a constant rate?
A laboratory reports a rate of $0.2$ defects per centimetre and inspects $15$ centimetres. Find the expected count and state the model parameter.
Explain why a Poisson parameter need not be an integer, even though the modeled count is integer-valued.
A rate is stated as $5$ events per hour. Explain why this number is not automatically the Poisson parameter for a 10-minute count.

Poisson Counts and the Rate Parameter

What You'll Learn