Population Denominators In Rate Models

Why Population Denominators Matter

A population count is hard to interpret without knowing how many people could have contributed events. Forty events in a population of 800 people is not the same population rate as forty events in a population of 8,000. If follow-up time also differs, the denominator must reflect both the number at risk and the time they contribute. In a Poisson rate model, this denominator is exposure: it scales the expected count while the rate describes the event intensity per unit of exposure.

This distinction is important because a rate is not, by itself, an individual probability. A fitted rate can be used to calculate an expected count for a population of a stated size and duration. It does not directly say what fraction of individuals will experience an event.

Population Exposure in a Count Model

Definition: For a group with \(N_i\) people at risk over a common duration \(t_i\), its population exposure is \(E_i=N_i t_i\), measured in person-time units. More generally, \(E_i\) is the sum of the actual time at risk contributed by the people in group \(i\). A population event rate is a count per unit of this exposure.

For example, 400 people each followed for 3 years contribute 1,200 person-years if everyone remains at risk for all 3 years. If people contribute different durations, the denominator is the sum of those durations, not the initial population size multiplied by a duration that they did not all complete.

Let \(Y_i\) be the number of events recorded for group \(i\). A Poisson rate model uses a rate \(\lambda_i\) per person-time unit and specifies

Y_i\sim\operatorname{Poisson}(E_i\lambda_i),\qquad E[Y_i]=E_i\lambda_i.

The population size and duration determine \(E_i\); covariates describe how the rate varies across groups. In the log-rate form, the model is

\log E[Y_i]=\log E_i+x_i^\mathsf{T}\beta, \qquad \lambda_i=\exp(x_i^\mathsf{T}\beta).

Here \(\log E_i\) is the known exposure contribution, and the linear predictor \(x_i^\mathsf{T}\beta\) determines the log rate. The fitted mean count for a specified group is \(\hat{\mu}_i=E_i\exp(x_i^\mathsf{T}\hat{\beta})\), so dividing that mean by \(E_i\) recovers the fitted rate. The denominator is not an extra event count and is not the quantity being predicted.

Worked Example: Converting a Fitted Rate to Population Counts

Suppose a fitted model has log rate \(\log(\hat{\lambda})=-5+0.4x\), where \(x=0\) denotes the reference group and \(x=1\) denotes the comparison group. Calculate each fitted rate per person-year and the fitted count for each group if each is followed for 1.5 years.

Reference group. With \(x=0\), the fitted log rate is \(-5\), so \[ \hat{\lambda}_0=\exp(-5)=0.006738 \] events per person-year. This is \(6.738\) fitted events per 1,000 person-years.

Comparison group. With \(x=1\), the fitted log rate is \(-5+0.4=-4.6\), so \[ \hat{\lambda}_1=\exp(-4.6)=0.010052 \] events per person-year, or \(10.052\) per 1,000 person-years.

For a population of 2,000 people followed for 1.5 years, the exposure is \(2{,}000(1.5)=3{,}000\) person-years. The fitted count is \(3{,}000(0.006738)=20.214\) events in the reference group and \(3{,}000(0.010052)=30.156\) in the comparison group. These are expected counts, so they need not be integers.

The two populations have equal size and follow-up, so their fitted counts reflect their different fitted rates. If their rates were equal but their sizes differed, the larger population would instead have the larger expected count in proportion to its exposure.

How an Aggregate Count Reflects Individual Contributions

The population formulation has a direct interpretation when a group consists of people whose counts are individually modeled. Suppose the \(N\) people each contribute the same duration \(t\), and person \(j\)'s event count \(X_j\) has a Poisson distribution with mean \(t\lambda\). The group count is \(Y=\sum_{j=1}^N X_j\). The following result explains the aggregate mean and, with independence, its Poisson distribution.

Result: If \(X_1,\ldots,X_N\) are independent Poisson random variables with means \(\mu_1,\ldots,\mu_N\), then their sum is Poisson with mean \(\sum_{j=1}^N\mu_j\). In particular, if each person has mean count \(t\lambda\), the population count is Poisson with mean \(Nt\lambda\).

Proof. For a Poisson random variable \(X_j\), its probability-generating function is \(G_j(s)=E[s^{X_j}]=\exp\{\mu_j(s-1)\}\). Independence gives the generating function of the sum:

E\left[s^{\sum_{j=1}^N X_j}\right] =\prod_{j=1}^N E[s^{X_j}] =\prod_{j=1}^N \exp\{\mu_j(s-1)\} =\exp\left\{\left(\sum_{j=1}^N\mu_j\right)(s-1)\right\}.

The final expression is the probability-generating function of a Poisson random variable with mean \(\sum_j\mu_j\). When \(\mu_j=t\lambda\) for every person, this sum is \(Nt\lambda\), which is \(E\lambda\) for exposure \(E=Nt\). Independence is needed for this distributional conclusion. The expected total count is \(\sum_j E[X_j]\) even without independence, but dependence can invalidate the Poisson distribution for the total. \(\square\)

A related rate calculation applies when people contribute unequal amounts of person-time. If person \(j\) contributes \(t_j\) and has the same rate \(\lambda\), then \(E[X_j]=t_j\lambda\), and linearity of expectation gives

E\left[\sum_{j=1}^N X_j\right] =\sum_{j=1}^N t_j\lambda =\left(\sum_{j=1}^N t_j\right)\lambda.

Thus the total expected count is total person-time multiplied by the common rate. If people instead have different rates, the total expected count is \(\sum_j t_j\lambda_j\). Dividing by total person-time then gives an exposure-weighted average rate, not necessarily a rate that describes every individual.

Worked Example: Estimating a Common Population Rate

Across several groups assumed to share a common rate, there are \(y=18\) observed events in total exposure \(E=6{,}000\) person-years. Find the maximum-likelihood estimate of the common rate and the fitted count for 1,200 people followed for 2 years.

For independent Poisson counts with a common rate \(\lambda\), the likelihood across groups, omitting factors that do not depend on \(\lambda\), is proportional to \[ L(\lambda)\propto \lambda^{\sum_i y_i}\exp\left(-\lambda\sum_i E_i\right). \] Its log-likelihood, up to an additive constant, is \[ \ell(\lambda)=\left(\sum_i y_i\right)\log\lambda-\lambda\sum_i E_i. \] Differentiating and setting the score equal to zero gives \[ \frac{d\ell}{d\lambda}=\frac{\sum_i y_i}{\lambda}-\sum_i E_i=0, \qquad \hat{\lambda}=\frac{\sum_i y_i}{\sum_i E_i}. \] Here \(\hat{\lambda}=18/6{,}000=0.003\) events per person-year, or 3 events per 1,000 person-years. The second derivative is \(-(\sum_i y_i)/\lambda^2\), which is negative for \(\lambda>0\), so this stationary point is a maximum.

The new population contributes \(1{,}200(2)=2{,}400\) person-years. Its fitted expected count at the estimated common rate is \(2{,}400(0.003)=7.2\) events. The value 7.2 is a model-based expected count, not a claim that a realized count must equal 7.2.

Rates Are Not Individual Risks

An individual risk is a probability, such as the probability that a person experiences at least one event during a stated interval. A population rate is an expected event count divided by person-time. The numerator of a rate can include repeated events for one person; a risk of experiencing at least one event cannot count the same person more than once. The two quantities also have different units: risk is a probability, while rate is measured per unit of time at risk.

For a simple illustration, suppose an individual event count over a specified interval follows a Poisson distribution with mean \(\mu\). The expected count is \(\mu\). The probability of at least one event is instead

P(X\geq 1)=1-P(X=0)=1-\exp(-\mu).

If \(\mu=1.2\), the expected count is 1.2 events, whereas the probability of at least one event is \(1-\exp(-1.2)\approx0.699\). The model permits multiple events, which is why an expected count can exceed 1 even though a probability cannot. Other event and follow-up definitions can lead to other relationships; a rate alone does not determine an individual risk without additional assumptions about the event process and interval.

Worked Example: A Rate and a Risk Can Differ

A model assigns a rate of \(0.5\) events per person-year to a group. For an illustration, assume each person's event count over 2 years follows a Poisson distribution with mean equal to rate times duration. Find the expected count per person and the probability of at least one event in 2 years.

The expected count per person is \(\mu=(0.5)(2)=1\) event. The probability of at least one event is \[ P(X\geq1)=1-\exp(-1)\approx0.6321. \] Thus the rate multiplied by the interval gives the expected number of events under this model, while the chance of any event is about 63.2%. They are not the same quantity. This calculation relies on the stated Poisson count assumption and should not be treated as a general conversion from rate to risk.

Comparing Populations and Avoiding Denominator Errors

When two populations have different sizes, compare their fitted rates rather than their fitted counts if the question concerns event intensity per unit of exposure. For example, a population with 10,000 person-years and a rate of 4 per 1,000 person-years has an expected count of 40. A population with 2,000 person-years and a higher rate of 8 per 1,000 person-years has an expected count of 16. The first population has more expected events because its exposure is much larger, even though its rate is lower.

A useful check is to express the rate and exposure in compatible units before multiplying. If the rate is reported per 1,000 person-years, divide the number of person-years by 1,000; equivalently, multiply the rate per person-year by total person-years. For a rate of 4 per 1,000 person-years and exposure of 10,000 person-years, the expected count is \((4/1{,}000)(10{,}000)=40\).

A common mistake is to read a fitted count as a fitted risk: “the model predicts 40 events, so each person has a 40% chance of an event.” That conclusion does not follow. Forty is a population expected count, and it may include repeated events. To obtain a risk, the outcome must be defined at the individual level and the relevant probability model and time interval must be specified. Another mistake is to compare raw counts while ignoring population denominators. Counts combine the rate with exposure, so a larger count does not by itself indicate a higher rate.

Key takeaway: In a Poisson rate model, expected population count equals exposure multiplied by rate. Use the count to describe expected events at a given population size and duration; use the rate to compare event intensity per unit of person-time. Neither quantity should be called an individual risk without an appropriate probability interpretation.

Check Your Understanding

Answer each question using the distinction between exposure, rate, count, and risk.

A group of 500 people is followed for 2 years, with everyone at risk throughout. What is its exposure in person-years?
A fitted rate is 6 events per 1,000 person-years. What is the fitted expected count for 4,000 person-years?
Why does independence matter for concluding that a sum of Poisson individual counts is itself Poisson? Does it matter for adding their expectations?
In a common-rate model, 12 events occur in 3,000 person-years. What is the estimated rate per person-year and per 1,000 person-years?
Explain why an expected count of 1.2 events per person over an interval is not an individual probability of 1.2.