Rescaling Poisson Counts Across Observation Windows

Why the Observation Window Matters

A count of events is meaningful only alongside the amount of opportunity to observe them. Three events in one hour and three events in a month may be the same count, but they suggest very different event rates. When the rate is stable, changing the observation window changes the expected count—and therefore the Poisson distribution used to describe the count.

The key operation is to change the Poisson parameter in proportion to exposure. If the amount of exposure doubles while the rate stays the same, the expected count doubles. But the probability of any particular count does not generally double: it must be recalculated from the new Poisson distribution.

Rate, Exposure, and the Poisson Parameter

Definition: Let \(w\) be the amount of exposure in a specified unit, such as days, square metres, or patient-hours. Let \(r\) be a constant event rate per unit of that exposure. Under a Poisson count model, the count \(X_w\) in an exposure amount \(w\) has distribution \(X_w\sim\operatorname{Poisson}(\lambda_w)\), where the parameter is \(\lambda_w=rw\). The parameter \(\lambda_w\) is the expected count in that window.

For example, if the rate is \(0.8\) events per day, then a five-day window has parameter \(\lambda=0.8(5)=4\). The rate and exposure units must match: multiplying events per day by days gives an expected number of events.

This model requires more than a multiplication rule. The event rate must be reasonably stable over the windows being compared, and the exposure measure must represent the relevant opportunity for events. The Poisson model must also be appropriate for the count process: substantial clustering, changing rates, or other systematic departures can make it unsuitable. Rescaling does not repair a poor model; it carries the same rate and model assumptions to a different window.

For any nonnegative integer \(k\), a Poisson count with parameter \(\lambda\) has probability

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

The mean of this count is \(\lambda\). Thus, the parameter changes directly with exposure, while the probabilities change according to the full probability formula.

The Rescaling Rule

Suppose a window with exposure \(w_0\) has Poisson parameter \(\lambda_0=rw_0\). If the new window has exposure \(w_1\), and the rate \(r\) remains the same, then its parameter is \(\lambda_1=rw_1\). Dividing the two expressions gives the rescaling rule:

\lambda_1=\lambda_0\frac{w_1}{w_0}.

The ratio \(w_1/w_0\) is the exposure multiplier. A window half as long has half the parameter; a window three times as large has three times the parameter. These statements concern the distributions and their means, not a promise that an individual observed count will scale by the same factor.

The rule also works when exposure is measured in different but convertible units. Convert the amounts to a common unit first, or use the rate in units that match the new exposure. For instance, a rate per 1,000 hours multiplied by 2,500 hours means multiplying by \(2.5\), not by \(2,500\).

Key result: If a Poisson model has rate \(r\) per unit of exposure, then the count over exposure \(w\) has parameter \(rw\). Changing exposure by a factor \(c\) changes the parameter and expected count by that factor: \(\lambda_{\mathrm{new}}=c\lambda_{\mathrm{old}}\).

Worked Applications

Worked Example: Rescaling an Event Rate Over Time

Suppose a hypothetical monitoring system records events at a stable rate of \(0.75\) events per day. Under a Poisson model, find the expected count over 10 days, the probability of exactly two events, and the probability of zero or one event.

Check the model and units. Assume the event rate stays at \(0.75\) per day throughout the 10-day window, and that a Poisson count model is appropriate. The exposure is 10 days, so the parameter is

\lambda=(0.75\ \text{events/day})(10\ \text{days})=7.5\ \text{events}.

The expected count is \(7.5\) events. The expectation need not be a whole number; actual counts are whole numbers, but their long-run average can be fractional. For exactly two events, substitute \(k=2\) and \(\lambda=7.5\):

P(X=2)=e^{-7.5}\frac{7.5^2}{2!} =e^{-7.5}(28.125) \approx 0.01556.

Thus, the probability of exactly two events in 10 days is about \(1.556\%\). For zero or one event, add the probabilities for those two counts:

P(X\leq 1)=P(X=0)+P(X=1) =e^{-7.5}\left(1+7.5\right) \approx 0.004701.

The probability of zero or one event is about \(0.4701\%\). Both probabilities use the 10-day parameter; using the one-day parameter would answer a different question.

Worked Example: Rescaling Across Areas

In a hypothetical production setting, assume a certain type of flaw occurs at a stable rate of \(0.04\) flaws per square metre, and that a Poisson model is appropriate. Compare the expected counts and probabilities of exactly one flaw for areas of \(12.5\) and \(25\) square metres.

Check the exposure scale. The rate is per square metre, so multiply it by area in square metres. For \(12.5\) square metres,

\lambda_{12.5}=(0.04)(12.5)=0.5.

The expected count is \(0.5\) flaws, and the probability of exactly one is

P(X_{12.5}=1)=e^{-0.5}\frac{0.5^1}{1!} \approx 0.3033.

For \(25\) square metres, the exposure is twice as large, so the parameter is twice as large:

\lambda_{25}=(0.04)(25)=1, \qquad P(X_{25}=1)=e^{-1}\frac{1^1}{1!}\approx 0.3679.

The expected count rises from \(0.5\) to \(1\) flaw. The probability of exactly one flaw rises from about \(30.33\%\) to about \(36.79\%\), not to twice its original value. Probabilities for individual counts depend on the entire Poisson formula, not just on the exposure multiplier.

Worked Example: Comparing Shorter and Longer Observation Windows

A hypothetical service records events at a stable rate of \(1.2\) events per 1,000 patient-hours. Assume the Poisson model is suitable. Compare the probability of at most one event in 2,500 patient-hours and in 5,000 patient-hours.

Convert exposure to the rate's units. The first window contains \(2.5\) units of 1,000 patient-hours, so its parameter is \(\lambda_1=(1.2)(2.5)=3\). The longer window contains \(5\) such units, giving \(\lambda_2=(1.2)(5)=6\). The longer window has twice the exposure and twice the expected count.

For a Poisson count, the probability of at most one event is the sum of the probabilities of zero and one. With \(\lambda=3\),

P(X\leq 1)=e^{-3}\left(1+3\right) \approx 0.1991.

With \(\lambda=6\),

P(X\leq 1)=e^{-6}\left(1+6\right) \approx 0.01735.

Under the stated model, the chance of at most one event is about \(19.91\%\) in 2,500 patient-hours and about \(1.735\%\) in 5,000 patient-hours. The rate is unchanged; the longer exposure makes a low count less likely.

Why Expected Counts and Probabilities Behave Differently

The expected count is especially simple to rescale because it equals the parameter: multiplying exposure by \(c\) multiplies the expected count by \(c\). Probabilities are not generally proportional to exposure. For instance, the probability of no events in a window is

P(X=0)=e^{-\lambda}=e^{-rw}.

Consequently, when exposure is multiplied by \(c\), the probability of no events becomes \(e^{-crw}\), which is not \(c\) times the original probability. Similarly, the probability of at least one event follows by taking the complement of zero events:

P(X\geq 1)=1-e^{-rw}.

This result is useful when the question is whether any event occurs, but it still depends on the new parameter. For exact counts or ranges of counts, use the Poisson probability formula or add the relevant individual probabilities.

A Short Justification of the Rescaling

The rate \(r\) is defined as the expected number of events per unit of exposure. If the expected count over \(w_0\) units is \(\lambda_0\), then \(r=\lambda_0/w_0\). Holding that rate fixed, the expected count over \(w_1\) units is \(rw_1\). Substitution gives

rw_1=\frac{\lambda_0}{w_0}w_1 =\lambda_0\frac{w_1}{w_0}.

Because a Poisson distribution is specified by its parameter, this new expected count is also the new Poisson parameter. That establishes the rescaling rule. The justification depends on keeping the rate constant; if the event rate changes between windows, the same multiplier need not describe the new count.

Common Mistakes and Model Checks

Keeping the old parameter. If the observation window changes, the Poisson parameter changes too. Recalculate it using the rate and the new exposure, or multiply the old parameter by the exposure ratio.
Scaling a probability like a mean. Doubling exposure doubles the expected count under the model, but it does not generally double the probability of exactly \(k\) events. Recompute the probability using the new parameter.
Confusing an expected count with a guaranteed count. A parameter of \(7.5\) means an expected count of \(7.5\), not that the observed count must be near \(7.5\) in every window.
Mixing exposure units. A rate per 1,000 hours must be multiplied by the number of 1,000-hour units, not by the raw number of hours.
Assuming the rate stays fixed without checking. Rescaling is not justified if the rate changes systematically across time, location, or operating conditions, or if the Poisson model otherwise fails to represent the counts.

A useful check is to write the rate units and exposure units beside each other. They should cancel, leaving an expected number of events. Then state the assumption that makes the same rate applicable to the new window.

Check Your Understanding

Use the rate-times-exposure rule and state any assumption needed for each comparison.

A Poisson model has rate \(0.6\) events per hour. What parameter and expected count correspond to a 4-hour window?
A window with exposure \(w_0\) has parameter \(5\). If a new window has three times the exposure and the rate is unchanged, what is its parameter?
Explain why doubling exposure does not mean the probability of exactly one event doubles.
A rate is \(2\) events per 1,000 machine-hours. Find the Poisson parameter for 3,500 machine-hours.
Give one reason why multiplying the original parameter by an exposure ratio might fail to describe a new window.