Marginal Event Rates And Rate Ratios

Why a Rate Needs a Denominator

For recurrent events, a count alone can be misleading: one person observed for a year has less opportunity to experience events than someone observed for five years. A person-time event rate accounts for this difference by relating the number of events to the time during which people could experience them. A rate ratio then compares two such rates. Its meaning depends on defining both the event count and the time denominator, as well as the population and follow-up period to which the comparison refers.

Definition: Let $N^a(\tau)$ be the number of recurrent events experienced by a member of a target population under treatment or exposure condition $a$, from time 0 through a specified horizon $\tau$. Let $T^a(\tau)$ be that person’s eligible time at risk over the same period. The marginal event rate under $a$ is $$ \lambda_a(\tau)=\frac{E\{N^a(\tau)\}}{E\{T^a(\tau)\}}. $$ The numerator is the population mean number of events per person, and the denominator is the population mean eligible person-time per person. The rate has units of events per unit of person-time. For two conditions, the marginal rate ratio is $RR(\tau)=\lambda_1(\tau)/\lambda_0(\tau)$, provided both rates are defined and $\lambda_0(\tau)>0$.

The expectation is taken over the specified target population. The rate is therefore a population-level ratio of mean totals, not necessarily the average of individual-level rates. The horizon $\tau$, event definition, and rules for counting eligible time must be stated. If a terminal event such as death ends eligibility for recurrent events, time after that event is not included in $T^a(\tau)$. The terminal event itself is not added to the recurrent-event numerator unless it is explicitly part of the event definition.

In a fully observed sample of $n_a$ people under condition $a$, a natural crude estimate is the total observed number of events divided by the total observed eligible person-time:

\widehat{\lambda}_a = \frac{\sum_{i=1}^{n_a}N_i(\tau)} {\sum_{i=1}^{n_a}T_i(\tau)}, \qquad \widehat{RR} = \frac{\widehat{\lambda}_1}{\widehat{\lambda}_0}.

If time is measured in person-years, for example, the estimated rate is reported as events per person-year. Multiplying a rate by 100 or 1,000 changes only the reporting scale. It does not turn the rate into a probability.

Assumptions and Interpretation

A rate is meaningful only when its numerator and denominator refer to the same people, event process, and observation period. The analysis must use a consistent event definition and a consistent rule for when a person contributes eligible time. Event ascertainment should be comparable across the conditions being compared. If a treatment changes the duration for which people remain alive and eligible, that change affects the denominator and can also affect the number of events observed.

With complete follow-up through $\tau$, the crude ratio of totals directly estimates the corresponding arm-specific rate. With censoring before $\tau$, observed person-time alone does not generally recover the full-horizon rate if censoring is related to future event experience. A rate estimate for the intended target then requires appropriate assumptions and handling of censoring; it should not silently treat censored people as if they had remained event-free and under observation through $\tau$. Independent censoring, possibly conditional on measured covariates, is a common identifying condition for censoring-adjusted approaches. The earlier material on censoring addresses such adjustments.

For a causal interpretation of a rate ratio, the two conditions must also represent well-defined interventions or exposure conditions, and the comparison must identify what would happen to the same target population under each. Randomization supports exchangeability; in an observational comparison, causal interpretation requires adequate control of confounding, along with consistency and positivity. Without those conditions, the rate ratio is an association between observed groups, not necessarily an effect of treatment.

Worked Example: Crude Rates and a Rate Ratio

Suppose a hypothetical trial follows two groups for one year, with complete event ascertainment and no loss to follow-up. In the intervention group, 48 events occur during 120 person-years. In the comparison group, 36 events occur during 150 person-years. Calculate the rates and rate ratio.

The intervention rate is total events divided by total eligible time:

\widehat{\lambda}_1=\frac{48\text{ events}}{120\text{ person-years}} =0.400\text{ events per person-year}.

The comparison-group rate is:

\widehat{\lambda}_0=\frac{36\text{ events}}{150\text{ person-years}} =0.240\text{ events per person-year}.

Thus the estimated rate ratio is:

\widehat{RR} = \frac{0.400}{0.240} = 1.67.

In this hypothetical sample, the intervention group experienced events at an estimated rate 1.67 times the comparison-group rate, or about 67% higher per person-year. This description is about rates, not the probability that an individual experiences an event. Because the example stipulates complete and comparable follow-up, the crude denominators are appropriate; censoring or unequal ascertainment could require a different analysis.

Why the Ratio of Totals Matters

The ratio of total events to total time weights each person’s contribution according to their eligible time. It is generally not the same as calculating each person’s event rate and then taking the unweighted average. The rate estimand in the definition is a ratio of population means, so its sample analogue is a ratio of sums.

The distinction follows from the relationship between totals and means. For a group of $n$ people, the sample mean event count is $\sum_i N_i/n$, and the sample mean time is $\sum_i T_i/n$. Their ratio is:

\frac{\sum_i N_i/n}{\sum_i T_i/n} = \frac{\sum_i N_i}{\sum_i T_i}.

By contrast, the average of individual rates is $\sum_i(N_i/T_i)/n$, which gives equal weight to each person’s rate regardless of how much time that person contributed. Unless the times are equal or a special relationship holds between counts and times, these quantities differ. The ratio-of-totals estimand answers how many events occurred per unit of time contributed by the target population.

Worked Example: Unequal Follow-Up and Individual Rates

Consider three hypothetical participants with complete observation until their respective study endpoints. Participant A has 2 events over 10 person-years, participant B has 4 over 20 person-years, and participant C has 6 over 10 person-years. Compare the ratio of totals with the unweighted average of individual rates.

The ratio of total events to total time is:

\frac{2+4+6}{10+20+10} = \frac{12}{40} = 0.300\text{ events per person-year}.

The individual rates are $2/10=0.200$, $4/20=0.200$, and $6/10=0.600$ events per person-year. Their unweighted mean is:

\frac{0.200+0.200+0.600}{3} = 0.333\text{ events per person-year}.

The results differ: 0.300 for the ratio of totals and 0.333 for the average of individual rates. The first uses all 40 person-years as the denominator; the second assigns equal weight to each of the three individual rates. For the marginal rate defined as population mean count divided by population mean time, the ratio of totals is the matching sample calculation.

Marginal Rates and the Target Population

A marginal rate averages over the covariate distribution of a named population. Consequently, two crude group rates can reflect both within-group event patterns and differences in who is represented in each group. If the goal is to compare conditions in a common target population, the rates should be defined for that same population under each condition. This is a standardization question, not a change in the definition of a rate.

Worked Example: Standardizing to a Common Population

Suppose a hypothetical target population is evenly divided between a lower-risk and a higher-risk stratum. Under condition A, the rates are 0.100 events per person-year in the lower-risk stratum and 0.300 in the higher-risk stratum. Under condition B, they are 0.080 and 0.240, respectively. Suppose each stratum contributes the same amount of eligible time per target-population member under both conditions. Calculate standardized marginal rates and their ratio.

With equal stratum weights and equal eligible-time contributions, the standardized rate under A is:

\lambda_A=0.5(0.100)+0.5(0.300) =0.200\text{ events per person-year}.

Under B it is:

\lambda_B=0.5(0.080)+0.5(0.240) =0.160\text{ events per person-year}.

The standardized rate ratio is $0.160/0.200=0.800$. For this target population, condition B has a rate 0.80 times the rate under A, or 20% lower. The common stratum weights make the target population explicit. If the conditions were observed in groups with different stratum compositions, their unadjusted rates need not produce this comparison. The calculation also relies on the stated equal person-time contribution; more generally, the marginal rate is a ratio of expected events to expected eligible time, so weighting event rates by population proportions alone is not sufficient when eligible times differ across strata or conditions.

What a Rate Ratio Does Not Say

A rate ratio compares event frequency per unit of eligible time. It does not directly compare the probability of at least one event by $\tau$, the mean number of events by $\tau$, or the time to a first event. Those are different summaries. The mean cumulative function, for example, describes the expected cumulative number of events by time; a rate summarizes events relative to time contributed over the chosen period. A constant rate over time is not required to define the period-specific marginal rate, but a single rate can hide important changes in event frequency over time.

Rates can also be affected by changes in exposure time. If one condition leads to earlier death, there may be less time in which recurrent events can occur. The resulting rate ratio describes events per eligible time under each condition; it does not by itself isolate a biological effect on the event process that would have occurred if everyone had equal survival. Similarly, a rate ratio above 1 does not mean that the probability of an event is multiplied by that value.

When repeated events from the same person are dependent, that dependence does not alter the arithmetic definition of the aggregate rate. It does matter for uncertainty estimates: treating every event as an independent observation can give misleading standard errors. Model-based or other appropriate inference must account for the data structure and target estimand. The definition of the marginal rate alone does not specify a particular inferential model.

Diagnostics and Common Pitfalls

Check the units. The numerator is a number of events and the denominator is person-time. Reporting “0.4 events” without a time unit leaves the rate incomplete.
Check who contributes time. Include only time during which the person is eligible to experience the event under the stated definition. Explain how terminal events and censoring end or alter that time.
Check comparable ascertainment. More frequent contact in one group can reveal more recurrent events, producing an apparent rate difference even when event experience is otherwise similar.
Check the target population. Differences in group composition may affect crude marginal rates. State whether the comparison is crude or standardized to a common population.
Do not call a rate a risk. A rate has units of inverse time; a risk is a probability over a specified period. They are not interchangeable.
Do not average individual rates by default. The stated marginal rate is the ratio of mean events to mean time. The unweighted average of person-specific rates targets a different quantity.

A useful audit is to report the event count, total person-time, rate, and rate ratio together. These quantities let readers verify the arithmetic and see whether the groups contributed substantially different exposure times.

Key takeaway: A marginal person-time rate is the population mean event count divided by the population mean eligible time. A rate ratio compares two such rates in a defined target population; its interpretation depends on the event definition, time denominator, follow-up, and assumptions supporting the comparison.

Check Your Understanding

Try each question, then compare your reasoning with the solutions.

A hypothetical group has 18 recurrent events over 60 person-years. What is its event rate, including units?
Two groups have rates of 0.12 and 0.08 events per person-year. What is the rate ratio for the first group relative to the second, and how should it be interpreted?
Why can an unweighted average of individual event rates differ from the ratio of total events to total person-time?
If death ends eligibility for a recurrent event, should post-death time be included in the person-time denominator? Explain.

Solutions. (1) $18/60=0.300$ events per person-year. (2) The rate ratio is $0.12/0.08=1.50$; the first group’s estimated event rate is 1.50 times the second group’s rate, or 50% higher per person-year. This does not mean its event probability is 50% higher. (3) The ratio of totals weights contributions through their person-time, whereas the unweighted average gives each person’s rate equal weight. The two calculations generally target different quantities. (4) No. If death ends eligibility, post-death time is not time at risk for the defined recurrent event and does not belong in its denominator.