Rates, Counts, And Exposure Time

Rates, Counts, and Exposure

A count tells how many events were observed; an event rate describes how that expected count relates to the amount of opportunity for events to occur. In clinical studies, exposure might be measured in person-years. In reliability studies, it might be operating hours. Comparing counts without accounting for exposure can be misleading: 12 events during 100 person-years and 12 events during 1,000 person-years are not the same event frequency.

The key distinction is that a rate is not a count and is not generally a probability. A rate has units such as events per person-year, while a count is a nonnegative integer and a probability is a number between zero and one. The connection between them requires an exposure amount and, when calculating probabilities, a probability model.

Definition: Let $N(e)$ be the number of events observed during exposure $e>0$. If the expected count is proportional to exposure, the event rate $\lambda$ is defined by $$ E[N(e)] = \lambda e. $$ Thus $\lambda=E[N(e)]/e$ is an expected count per unit of exposure. Its units are events per exposure unit. The count $N(e)$ is random and integer-valued; the expected count $\lambda e$ need not be an integer.

The proportionality assumption means that the same rate applies across the exposure amounts being considered. It is a modeling assumption, not a consequence of observing a count. For example, an event rate of 0.06 events per person-year implies an expected count of 0.06 during one person-year and 0.60 during ten person-years, under this assumption. It does not say that exactly 0.06 or 0.60 events will occur.

For observed data, if $y$ events occur over exposure $e$, the ratio $y/e$ is the observed rate or empirical count per exposure. It is a data summary, and often an estimate of an underlying event rate. It is not the same thing as the rate parameter itself: random variation means that $y/e$ can differ from $\lambda$.

Translating Between Rate, Exposure, and Expected Count

Once two of the three quantities are known, the third follows from $E[N(e)]=\lambda e$. Multiplication gives an expected count from a rate and exposure; division gives a rate from an expected count and exposure, or exposure from a rate and expected count. Keep the units visible while calculating: they provide a useful check that the answer has the right interpretation.

Identify the rate and its unit.
For example, $\lambda=0.04$ events per machine-hour.

Identify the exposure.
For example, $e=50$ machine-hours.

Multiply to obtain the expected count.
$E[N(e)]=0.04 \times 50=2$ events.

Worked Example: From a Clinical Event Rate to an Expected Count

Suppose a study uses a constant rate of 0.06 recurrent events per person-year. What is the expected number of events during 25 person-years of exposure?

The rate is $\lambda=0.06$ events per person-year, and the exposure is $e=25$ person-years. Therefore,

E[N(25)] = \lambda e = (0.06\ \text{events/person-year})(25\ \text{person-years}) = 1.5\ \text{events}.

The person-year units cancel, leaving an expected count of 1.5 events. This does not mean that the study will observe a fractional event. The realized count must be an integer; 1.5 is its mean under the model. Nor does this calculation alone specify the probability of any particular count. That requires a distributional assumption.

Worked Example: Calculating an Observed Rate

A monitoring system records 18 failures during 600 operating hours. Calculate the observed rate and use it to find the expected count during a further 250 operating hours if the same rate is assumed to apply.

The observed rate is the count divided by the observed exposure:

\frac{18\ \text{failures}}{600\ \text{hours}} =0.03\ \text{failures per hour}.

Treating this rate as the rate for the calculation, the expected count over 250 hours is

(0.03\ \text{failures/hour})(250\ \text{hours}) =7.5\ \text{failures}.

The result is an expected count, not a promise to observe 7.5 failures. Also, the rate 0.03 was calculated from data; it is not automatically known to equal the underlying rate. The calculation assumes that the rate remains applicable during the additional exposure.

When the Count Is Poisson

A rate and exposure determine an expected count under the proportionality assumption, but they do not by themselves determine a count distribution. For probability calculations, consider the Poisson model: for a fixed exposure $e>0$, the count $N(e)$ has a Poisson distribution with mean $\lambda e$. This model is a useful way to connect an event rate to probabilities, but its probability statements depend on the Poisson assumption.

Definition: Under the Poisson count model with rate $\lambda$ per unit of exposure, for $\lambda\geq 0$ and $e>0$, $$ N(e)\sim\operatorname{Poisson}(\lambda e), \qquad P(N(e)=k)=\exp(-\lambda e)\frac{(\lambda e)^k}{k!}, \quad k=0,1,2,\ldots. $$ The mean count is $\lambda e$. The probability of at least one event is obtained from the probability of zero events; it is not the rate itself.

The mean-count statement follows directly from the Poisson probability mass function.

Result: If $N(e)\sim\operatorname{Poisson}(\lambda e)$, then $E[N(e)]=\lambda e$.

Proof. Write $\mu=\lambda e$. If $\mu=0$, then $N(e)=0$ with probability one, so its expectation is zero. If $\mu>0$, using the Poisson mass function and the nonnegative terms in the expectation sum,

\begin{aligned} E[N(e)] &=\sum_{k=0}^{\infty} k\exp(-\mu)\frac{\mu^k}{k!}\\ &=\exp(-\mu)\sum_{k=1}^{\infty}\frac{\mu^k}{(k-1)!}\\ &=\mu\exp(-\mu)\sum_{j=0}^{\infty}\frac{\mu^j}{j!}\\ &=\mu\exp(-\mu)\exp(\mu)\\ &=\mu=\lambda e. \end{aligned}

In the third line, $j=k-1$; the final series is the exponential series, which converges to $\exp(\mu)$. This verifies that the Poisson model’s mean count agrees with the rate-times-exposure definition. $\square$

Worked Example: From an Expected Count to an Event Probability

An event rate is 0.06 events per person-year. Under the Poisson count model, calculate the probability of at least one event during 20 person-years.

First calculate the Poisson mean:

\mu=\lambda e=(0.06)(20)=1.2.

For a count to be at least one, it must not be zero. The probability of zero events is

P(N(20)=0)=\exp(-1.2)\frac{1.2^0}{0!}=\exp(-1.2).

Therefore,

\begin{aligned} P(N(20)\geq 1) &=1-\exp(-1.2)\\ &\approx 1-0.3012\\ &=0.6988. \end{aligned}

The probability is approximately 0.6988, or 69.88%. The rate 0.06 is not this probability: it is an expected count per person-year, while 0.6988 is the probability of at least one event over a specified 20 person-years under the Poisson model.

Rate Is Not Probability

Rates and probabilities answer different questions. A rate describes expected event frequency relative to exposure. A probability describes the chance of an event, or of a specified set of outcomes, in a defined situation. A rate can be greater than one when it is expressed in a chosen unit; it is therefore not constrained to the interval from zero to one. By contrast, every probability lies between zero and one.

Even when a rate is less than one, treating it as a probability is generally incorrect. For instance, under the Poisson model, a rate of 0.06 events per person-year gives the probability of at least one event in one person-year as $1-\exp(-0.06)$, approximately 0.0582. This is close to 0.06 but is not equal to it. For 20 person-years, the corresponding probability is approximately 0.6988, as calculated above. Exposure and the probability model matter.

This distinction is especially important for events that can occur repeatedly. A probability of an event over a period usually concerns whether it happens at least once. A count can instead record one, two, or more occurrences. The rate concerns the expected count per exposure, so it naturally accommodates repeated events.

Common Mistake

Wrong reasoning: “The rate is 0.06 events per person-year, so 6% of people will have an event.” This treats a count per exposure as a probability and leaves the time period and event model unspecified.
Correction: First state the exposure and the quantity of interest. For example, under a Poisson model, the chance of at least one event in one person-year is $1-\exp(-0.06)$, approximately 0.0582. For 20 person-years it is $1-\exp(-1.2)$, approximately 0.6988. These are probabilities for specified exposures, not alternative ways to write the rate.
Another warning: An observed count divided by exposure is an observed rate, not necessarily the true rate. Random variation can make it higher or lower than the underlying rate.

Key takeaway: Under a constant-rate model, expected count equals rate times exposure. A count is an observed integer, an observed rate is count divided by exposure, and an event probability requires a specified event and exposure, together with a probability model when one is needed.

Check Your Understanding

Use the distinction among rate, count, exposure, and probability to answer these questions.

A rate is 0.08 events per operating hour. What is the expected count over 30 operating hours under the constant-rate assumption?
A study observes 9 events over 300 person-years. What is the observed rate, with units?
Why can an expected count be non-integer even though the realized count must be an integer?
Under a Poisson model with rate $\lambda$ and exposure $e$, write the probability of zero events and the probability of at least one event.
Explain why a rate of 0.06 events per person-year is not, by itself, a 6% probability of an event.

Rates, Counts, And Exposure Time

What You'll Learn