Common Errors In Exposure And Rate Analysis

Why Exposure Errors Matter

A count does not describe risk or frequency on its own. Ten events observed over 20 person-years and ten events observed over 2,000 person-years are the same raw count, but they represent very different observed rates. In count models, exposure records how much opportunity there was for events to occur: person-time, distance traveled, area observed, or another relevant amount. Correct analysis must keep the count and its exposure connected.

This tutorial focuses on four recurring mistakes: treating exposure as an ordinary predictor despite assuming proportionality, comparing raw counts across unequal exposures, averaging rates without accounting for exposure, and confusing uncertainty about a rate with the variation of a future count. The examples use positive, fixed exposures and Poisson counts. Special handling is needed at zero exposure; the log-offset expression below is not defined there.

Definition: For record \(i\), let \(Y_i\) be a count and \(E_i>0\) its exposure. The observed rate is \(Y_i/E_i\). A proportional-exposure Poisson model specifies \(Y_i\) as Poisson with mean \(E_i\lambda_i\), where \(\lambda_i\) is the event rate per exposure unit. If \(\log(\lambda_i)=x_i^\mathsf{T}\beta\), the model can be written \(\log(E[Y_i\mid x_i,E_i])=\log(E_i)+x_i^\mathsf{T}\beta\); \(\log(E_i)\) is a fixed offset, not a coefficient estimated from the data.

Offset or Estimated Exposure Effect?

The offset expresses a specific scientific assumption: holding the rate predictors fixed, doubling exposure doubles the expected count. In the log-mean model, that assumption fixes the coefficient of \(\log(E_i)\) to one. In contrast, adding \(\log(E_i)\) as an ordinary predictor estimates its coefficient from the data. These choices answer different questions.

For example, the more general model \(\log(E[Y_i\mid x_i,E_i])=x_i^\mathsf{T}\beta+\gamma\log(E_i)\) implies that multiplying exposure by \(c\) multiplies the expected count by \(c^\gamma\). Proportionality corresponds to \(\gamma=1\). If the scientific assumption is proportionality, freeing \(\gamma\) changes the model rather than merely giving the offset a different name. Conversely, fixing the coefficient at one without justification can impose a poor assumption. Earlier material on checking the proportional exposure assumption addresses how to assess this issue.

Worked Example: An Offset Is Not an Estimated Exposure Effect

Suppose two records have the same covariates and exposures \(E_1=100\) and \(E_2=300\). Their common rate is \(\lambda=0.04\) events per exposure unit. Compare the expected counts under proportional exposure with a model that instead has \(\gamma=0.8\).

With a proportional-exposure offset, the expected counts are \(100(0.04)=4\) and \(300(0.04)=12\). Tripling exposure triples the expected count.

If the model uses \(\gamma=0.8\), the expected-count ratio is \((300/100)^{0.8}=3^{0.8}\), approximately \(2.408\), not 3. If the first expected count is 4, the second is \(4(2.408)\), approximately \(9.63\). The estimated exposure effect therefore relaxes proportionality; it is not equivalent to fixing an offset.

Raw Counts Do Not Adjust for Opportunity

Comparing counts alone ignores how long, how far, or how much was observed. When exposures differ, the larger count can correspond to the smaller rate. A rate comparison should use a consistent exposure unit and should be made in light of the intended comparison: crude observed rates and covariate-adjusted rates are not automatically interchangeable.

Worked Example: The Larger Count Has the Lower Rate

Clinic A records 12 events over 100 patient-years; Clinic B records 18 over 300 patient-years. Which clinic has the higher observed rate?

The raw counts are 12 and 18, so comparing counts alone would make Clinic B appear to have more events. The rates are \[ \frac{12}{100}=0.12 \quad\text{and}\quad \frac{18}{300}=0.06 \] events per patient-year. Clinic A has twice the observed rate of Clinic B: \(0.12/0.06=2\).

This calculation describes observed rates, not proof that the underlying clinic rates differ. Sampling variation and differences in patient mix may also matter. The correction is to state what is being compared and use an appropriate rate or model-based comparison, rather than infer relative risk from raw totals.

Why Rates Must Be Weighted by Exposure

When records are combined, the overall observed rate is total events divided by total exposure. This is an exposure-weighted average of the record-specific rates, not generally their simple arithmetic mean. Records with more exposure contribute more to the total rate because they represent more opportunity for events.

Result: For records with positive exposures \(E_1,\ldots,E_m\), counts \(Y_1,\ldots,Y_m\), and observed rates \(r_i=Y_i/E_i\), the pooled observed rate satisfies \[ \frac{\sum_{i=1}^m Y_i}{\sum_{i=1}^m E_i} = \sum_{i=1}^m \left(\frac{E_i}{\sum_{j=1}^m E_j}\right)r_i. \] The weights are nonnegative and sum to one.

To verify the identity, substitute \(Y_i=E_i r_i\) into its left-hand side:

\frac{\sum_i Y_i}{\sum_i E_i} =\frac{\sum_i E_i r_i}{\sum_i E_i} =\sum_i \left(\frac{E_i}{\sum_j E_j}\right)r_i.

Each weight is \(E_i/\sum_j E_j\), so the weights sum to \(\sum_i E_i/\sum_j E_j=1\). This proves the result. An unweighted mean gives each record the same influence, regardless of how much exposure it represents. It may be a meaningful summary of the average record, but it is not the pooled rate.

Worked Example: A Simple Average Can Misstate the Pooled Rate

One record has 8 events over 40 exposure units, and another has 30 events over 300 units. Calculate both the simple average of the two observed rates and the pooled rate.

The observed rates are \(8/40=0.20\) and \(30/300=0.10\). Their simple average is \((0.20+0.10)/2=0.15\).

The pooled rate is total events divided by total exposure: \[ \frac{8+30}{40+300}=\frac{38}{340}\approx 0.1118. \] Equivalently, the exposure weights are \(40/340\) and \(300/340\), so the weighted average is \((40/340)(0.20)+(300/340)(0.10)=0.1118\), to four decimal places.

The simple average is much higher because it gives the 40-unit record the same weight as the 300-unit record. Use the pooled rate when the target is events per unit of combined exposure. If the target instead concerns an equally weighted average of records, say so explicitly; that is a different estimand.

Rate Uncertainty Is Not Count Variation

A rate estimate and a future count are related, but their variability is on different scales. For a fixed exposure \(E\) and rate \(\lambda\), a Poisson count has variance \(E\lambda\). The observed rate \(Y/E\) has variance \(\lambda/E\). Thus, more exposure increases expected counts and their absolute count-scale variation, while decreasing the sampling variance of the rate.

Result: If \(Y\sim\operatorname{Poisson}(E\lambda)\) for fixed \(E>0\) and \(\lambda>0\), then \(E[Y/E]=\lambda\), \(\operatorname{Var}(Y)=E\lambda\), and \(\operatorname{Var}(Y/E)=\lambda/E\). These follow from the Poisson mean and variance, both equal to \(E\lambda\), and the rules \(E[aY]=aE[Y]\) and \(\operatorname{Var}(aY)=a^2\operatorname{Var}(Y)\): \[ E[Y/E]=\frac{E[Y]}{E}=\lambda,\qquad \operatorname{Var}(Y/E)=\frac{\operatorname{Var}(Y)}{E^2}=\frac{E\lambda}{E^2}=\frac{\lambda}{E}. \]

The standard deviation of a future count is \(\sqrt{E\lambda}\); the standard deviation of the observed rate is \(\sqrt{\lambda/E}\). Confusing them can lead to uncertainty intervals on the wrong scale or to treating ordinary count fluctuation as uncertainty about a fitted rate. When \(\lambda\) is unknown, the rate standard error is estimated by substituting an estimate, often giving \(\sqrt{Y}/E\) for a single Poisson observation. This is not the standard deviation of the future count, which is estimated on the count scale by \(\sqrt{E\hat{\lambda}}\).

Worked Example: Count Variation and Rate Precision

A process has a true rate of \(0.2\) events per unit exposure. At fixed exposure \(E=100\), compute the expected count, its standard deviation, and the standard deviation of the observed rate.

The count has mean and variance \(E\lambda=100(0.2)=20\). Its standard deviation is \(\sqrt{20}\approx4.47\) events. The observed rate is \(Y/100\), so its variance is \(\lambda/E=0.2/100=0.002\) and its standard deviation is \(\sqrt{0.002}\approx0.0447\) events per exposure unit.

If the observed count is \(Y=20\), the observed rate is \(20/100=0.2\). The plug-in standard error for that rate is \(\sqrt{20}/100\approx0.0447\), while the plug-in standard deviation of a future count at the same exposure is \(\sqrt{100(0.2)}=\sqrt{20}\approx4.47\). The numerical values differ by a factor of 100 because the quantities are measured on different scales.

A further distinction matters in regression: uncertainty in a fitted rate can include uncertainty in estimated model coefficients, while a future count also has Poisson variation conditional on those coefficients. A count prediction must account for count variation; a confidence interval for a rate parameter addresses uncertainty in the rate estimate. They answer different questions.

A Practical Error-Checking Sequence

Before reporting an exposure-based result, check the model assumption, the comparison, the weighting, and the uncertainty scale. These checks do not replace model diagnostics, but they catch many interpretation errors.

State the exposure and its units.
Confirm that each rate is events per the same exposure unit, and determine whether exposures are positive and fixed for the analysis at hand.

Check the proportionality assumption.
If proportionality is intended, use \(\log(E_i)\) as an offset. If its coefficient is estimated, describe that as estimating an exposure effect and justify the choice.

Match the summary to the target.
Do not compare raw counts across unequal exposures. For a pooled crude rate, divide total events by total exposure; do not silently substitute an unweighted mean.

Name the uncertainty quantity.
Specify whether an interval concerns a rate estimate or a future count, and keep its units and scale consistent with that target.

Common Mistakes

Wrong: “Exposure is included, so its coefficient should be estimated.” Correction: including \(\log(E_i)\) as a free predictor estimates a coefficient; an offset fixes it at one. Choose according to the intended proportionality assumption.
Wrong: “The group with more events has the higher rate.” Correction: divide by exposure before comparing rates, or use an appropriate adjusted model when groups differ in relevant covariates.
Wrong: “The average of the record-specific rates is the overall rate.” Correction: total events divided by total exposure is the exposure-weighted average. An unweighted average answers a different question.
Wrong: “The standard deviation of a Poisson count is the uncertainty of the estimated rate.” Correction: count variation is on the count scale; divide by exposure to express the corresponding rate variation, and distinguish it from uncertainty in model parameters.

Key takeaway: Exposure is not merely another number beside a count. It defines the opportunity for events, determines how counts translate to rates, supplies the weights for a pooled rate, and changes the scale of uncertainty. Always state the assumed exposure effect and the precise target of a comparison or interval.

Check Your Understanding

Check whether each conclusion respects the exposure, the target quantity, and its uncertainty scale.

A proportional-exposure model is intended. What is the role of \(\log(E_i)\), and what changes if its coefficient is estimated?
Group A has 24 events in 120 person-years; Group B has 30 in 300 person-years. Calculate both rates and identify which is larger.
Two records have rates 0.3 and 0.1, with exposures 10 and 90. Calculate their simple average and their pooled rate. Why do they differ?
If \(Y\sim\operatorname{Poisson}(E\lambda)\), write the variances of \(Y\) and \(Y/E\). Explain the difference in their units.
In one sentence, distinguish an interval for a rate estimate from uncertainty about a future count at fixed exposure.