Why Exposure Belongs in the Denominator
A rate relates how many events occurred to how much opportunity there was for those events to occur. For example, comparing event counts between groups can be misleading if one group was observed for much longer, or contains more people. A rate such as events per person-year makes the amount of observed exposure explicit.
In a survey, the observed records may represent different numbers of population units. The survey weight accounts for that representation, while the exposure denominator accounts for differing time or other opportunity at risk. These are distinct roles: the weight expands a record to the target population, and the exposure measures the amount of observation associated with that record.
Suppose a survey sample contains records indexed by \(i\), with survey weight \(w_i\). A weight is used to estimate population totals from sampled records. The rate estimator divides the weighted estimate of the event total by the weighted estimate of the exposure total. Both totals must refer to the same target population, event definition, and observation period.
This is a descriptive ratio of estimated totals. It does not require every sampled unit to have the same rate, nor does the formula itself assume that event counts follow a particular probability distribution. For the estimator to describe the intended population, the weights must appropriately represent that population, and the event and exposure measurements must be defined consistently. Records used for the numerator and denominator should cover the same analytic population; missing or ineligible records need to be handled deliberately rather than omitted differently from the two totals.
Calculating the Weighted Rate
For each record, multiply the event count by the survey weight to obtain its weighted contribution to the event total. Separately, multiply exposure by the same weight to obtain its contribution to the exposure total. Sum each set of contributions, then divide the event total by the exposure total. If a more readable scale is wanted, multiply the resulting rate by a stated constant, such as 1,000 to report events per 1,000 person-years.
Worked Example: A Survey Rate From Three Records
Consider a hypothetical survey with three sampled people. The event is a specified health event during the observation period, exposure is time observed in person-years, and the weights are assumed to represent the target population appropriately.
| Person | Weight \(w_i\) | Events \(e_i\) | Exposure \(t_i\) (person-years) |
|---|---|---|---|
| A | 2 | 1 | 2 |
| B | 1 | 2 | 4 |
| C | 3 | 0 | 1 |
The weighted event total is \(2(1)+1(2)+3(0)=4\) events. The weighted exposure total is \(2(2)+1(4)+3(1)=11\) person-years. Therefore, $$ \hat{R}=\frac{4}{11}=0.3636 \quad\text{events per person-year}. $$
Equivalently, the estimated rate is \(0.3636(100)=36.36\) events per 100 person-years. This is a rate for the represented population over the specified observation period, not a count of 0.3636 events in the sample.
The division is essential to the interpretation. The weighted event total alone is an estimated number of events, while the weighted exposure total is an estimated amount of exposure. Dividing them puts the event total on a common exposure scale. Reporting the numerator without its denominator, or calling a weighted event total a rate, changes the quantity being described.
A Pooled Rate Is Not Usually a Weighted Mean of Person-Specific Rates
When \(t_i>0\), a unit-specific rate is \(r_i=e_i/t_i\). One might average these rates using the survey weights alone: $$ \bar{r}_{w}=\frac{\sum_{i\in s}w_i r_i}{\sum_{i\in s}w_i}. $$ That calculation summarizes the weighted mean of the units’ individual rates. The pooled population rate instead divides total events by total exposure. Units with more exposure contribute more to the pooled rate because they contribute more exposure to its denominator.
The identity follows by substituting \(r_i=e_i/t_i\). For each record, \(w_i t_i r_i=w_i t_i(e_i/t_i)=w_i e_i\). Summing gives \(\sum_i w_i t_i r_i=\sum_i w_i e_i\), while the sum of the exposure weights is \(\sum_i w_i t_i\). Their ratio is exactly the weighted rate. This algebra also shows why the two summaries can differ: they use different weights on the individual rates.
Worked Example: Why an Average of Individual Rates Can Mislead
Use the hypothetical records in the first example. Person A's rate is \(1/2=0.5\) events per person-year, person B's rate is \(2/4=0.5\), and person C's rate is \(0/1=0\).
The survey-weighted average of the individual rates, using the survey weights alone, is $$ \bar{r}_{w} = \frac{2(0.5)+1(0.5)+3(0)}{2+1+3} = \frac{1.5}{6} = 0.25 \quad\text{events per person-year}. $$
The pooled rate is \(4/11=0.3636\) events per person-year, as calculated earlier. The two results answer different questions. The value \(0.25\) is a survey-weighted mean of the three unit-specific rates; the value \(0.3636\) is total weighted events per total weighted person-years.
To verify the identity, the exposure-adjusted weights \(w_i t_i\) are 4, 4, and 3. The pooled rate as an exposure-weighted mean is $$ \frac{4(0.5)+4(0.5)+3(0)}{4+4+3} = \frac{4}{11} = 0.3636 \quad\text{events per person-year}. $$
If all included units have the same exposure, weighting the individual rates by \(w_i\) gives the same result as the pooled rate: the common exposure factor cancels. With unequal exposures, that equality generally fails. The appropriate summary depends on the scientific question. Use the pooled rate when asking how many events occurred per total amount of exposure. Use a survey-weighted mean of individual rates when the target is the mean of unit-specific rates, with each unit represented according to its survey weight rather than its exposure.
Combining Groups and Checking the Denominator
A population rate for several groups combined should be calculated from their combined event and exposure totals. It is generally not the arithmetic average of the group rates. A group contributing much more exposure also contributes more to the combined denominator, and the group event totals must be combined on the same basis.
Worked Example: Combining Two Groups
Suppose, hypothetically, that weighted survey totals for Group A are 12 events and 100 person-years, while totals for Group B are 8 events and 400 person-years. The group rates are \(12/100=0.12\) and \(8/400=0.02\) events per person-year.
The combined rate is found by adding the totals first: $$ \hat{R}_{\text{combined}} = \frac{12+8}{100+400} = \frac{20}{500} = 0.0400 \quad\text{events per person-year}. $$
The unweighted average of the two group rates is \((0.12+0.02)/2=0.0700\), which is not the combined rate. It gives the two groups equal influence despite their very different exposure totals. The combined rate can also be checked as an exposure-weighted average: \(\{100(0.12)+400(0.02)\}/500=20/500=0.0400\).
The denominator deserves a substantive check, not just an arithmetic one. For example, if the numerator counts events during a calendar year, the exposure should represent the corresponding observation time for the same eligible population. A denominator of enrolled people, clinic visits, or person-years answers different questions. Before calculating, specify what counts as an event, what exposure is measured, the time window, and which population is included.
Specify the population, event, observation period, and exposure unit. State the rate units that will be reported.
Confirm that each included record has compatible event and exposure measurements and an appropriate survey weight. Resolve missingness and eligibility consistently.
Calculate \(\sum_i w_i e_i\) and \(\sum_i w_i t_i\) separately, retaining their event and exposure units.
Divide the weighted event total by the positive weighted exposure total. Report the rate with its units and, if rescaled, the scale factor.
Common Mistakes and Why the Distinction Matters
- Using a count as a rate. A weighted event total estimates events, not events per person-year. Include the appropriately weighted exposure total in the denominator.
- Averaging individual rates without checking the target. The survey-weighted mean of individual rates and the pooled rate are different summaries when exposures vary. Decide which one answers the question before choosing the formula.
- Using incompatible records or periods. A numerator and denominator from different populations or time windows do not form the intended rate. Ensure both totals refer to the same analytic domain and observation period.
- Ignoring rate units. The numerical result is not meaningful without the event and exposure units. State, for example, “events per 1,000 person-years,” rather than reporting only a decimal.
Check Your Understanding
For each question, identify the quantity being summarized and show the key calculation or reasoning.
- A survey estimates 30 weighted events over 600 weighted person-years. What is the rate per person-year and per 1,000 person-years?
- In what sense does the survey-weighted mean of individual rates differ from the pooled rate?
- Why is the combined rate for several groups calculated by combining their weighted event and exposure totals rather than taking an unweighted average of their rates?
- What should be checked if event counts cover a calendar year but the exposure measurements cover a different observation period?
- When can a survey-weighted average of individual rates equal the pooled weighted rate even if the two formulas appear different?