Incidence Rate Ratio: Worked Example

An incidence rate ratio (IRR) compares two incidence rates — the number of new events divided by the total person-time at risk that produced them — rather than two fixed-denominator proportions. That distinction sounds small until follow-up time differs between the groups being compared, which it almost always does in real cohort data: people enroll on a rolling basis, drop out, die of unrelated causes, or are simply observed for different lengths of time. A risk ratio (relative risk) divides events by the number of people at the start; an IRR divides events by the person-time those people actually contributed. This guide works through why that difference matters, walks through a fully worked calculation, and covers the Poisson-based confidence interval used to report an IRR correctly.

What an Incidence Rate Ratio Actually Compares

An incidence rate is events per unit of person-time, not events per person:

Incidence rate = Number of new events ÷ Total person-time at risk

Person-time is the sum, across every subject, of the time each one was actually observed and at risk of the event — a person followed for 2 years contributes 2 person-years, a person followed for 6 months contributes 0.5 person-years, and someone who leaves the study (or has the event) partway through contributes only the time up to that point, not the full study duration. The incidence rate ratio is simply one group’s incidence rate divided by another’s:

IRR = (Events_exposed ÷ Person-time_exposed) ÷ (Events_unexposed ÷ Person-time_unexposed)

This is the standard summary measure from a Poisson regression model or a directly calculated rate comparison, and it is the measure CASRAI’s cohort study guide and confounding variable entry both assume when they refer to a “rate” rather than a “risk.”

Why Person-Time Matters: A Worked Example

Consider two cohorts followed for an outcome, where average follow-up differs between arms — a routine situation whenever enrollment is staggered, attrition differs by group, or one arm includes a higher-mortality population that is naturally observed for less time.

	Group A (exposed)	Group B (unexposed)
Subjects at baseline	100	200
Average follow-up per person	5.0 years	4.0 years
Person-time at risk	500 person-years	800 person-years
Incident events observed	15	10
Incidence rate	0.03 / person-year (30 per 1,000 person-years)	0.0125 / person-year (12.5 per 1,000 person-years)

The incidence rate ratio is 0.03 ÷ 0.0125 = 2.4 — Group A’s event rate is 2.4 times Group B’s, per unit of time at risk.

Now compare that with what a naive, fixed-denominator calculation would report if it used events divided by starting headcount instead of person-time — effectively treating this as a risk ratio and ignoring that Group A was followed a full year longer, on average, than Group B:

	Group A	Group B
“Risk” (events ÷ starting N)	15 ÷ 100 = 0.15	10 ÷ 200 = 0.05

That fixed-denominator ratio comes out to 0.15 ÷ 0.05 = 3.0 — a full 25% higher than the IRR of 2.4, on identical event counts and identical starting group sizes. The gap exists entirely because the naive calculation credits both groups with the same exposure time when they didn’t have the same exposure time: Group A’s members were, on average, at risk for an extra year each, so a fixed-denominator comparison overstates how much more frequently events actually happened per unit of time in Group A relative to Group B. This is exactly the scenario person-time adjustment exists to correct, and it is why a risk ratio computed on cohorts with unequal follow-up is not simply a less-precise version of the IRR — it is answering a subtly different question and can give a materially different number.

Incidence Rate Ratio vs. Risk Ratio

	Incidence Rate Ratio	Risk Ratio (Relative Risk)
Denominator	Person-time at risk (person-years, person-months)	Number of people at the start of follow-up (fixed)
Handles unequal follow-up?	Yes — built into the person-time denominator	No — assumes everyone was observed for the same, complete period
Handles staggered entry / loss to follow-up?	Yes, naturally, since each subject contributes only the time actually observed	Requires a fixed cohort with complete follow-up, or a separate survival-analysis adjustment
Typical source	Poisson regression, or events ÷ person-time directly	Cumulative incidence in a fixed cohort, 2×2 table
When the two converge	When the event is rare and every subject in both groups is followed for the same, complete duration, an IRR and a risk ratio computed on the same data will be numerically close. As follow-up duration diverges between groups (as in the example above) or the event stops being rare, they diverge — and the IRR is the statistically appropriate measure whenever person-time itself varies across subjects or groups.

A related but distinct measure is the hazard ratio, which also uses a time-based framework but comes from a Cox proportional-hazards model rather than a rate-per-person-time calculation; see CASRAI’s guide to interpreting a hazard ratio for how that measure differs from both the IRR and the risk ratio.

Computing the 95% Confidence Interval

An IRR without a confidence interval is as incomplete as any other point estimate. For two independent event counts each following an approximate Poisson distribution, the standard large-sample approximation (Rothman & Greenland) works on the log scale:

SE(ln IRR) = √(1 ÷ Events_exposed + 1 ÷ Events_unexposed)

Applying this to the worked example above (15 and 10 events):

ln(IRR) = ln(2.4) = 0.8755
SE(ln IRR) = √(1/15 + 1/10) = √0.1667 = 0.4082
95% CI on the log scale: 0.8755 ± (1.96 × 0.4082) = 0.0753 to 1.6756
Exponentiating back: 95% CI for the IRR = 1.078 to 5.342

Because this interval excludes 1.0, the rate difference between the two groups would conventionally be reported as statistically significant at α = 0.05 — though note how wide the interval is (1.08 to 5.34) relative to the point estimate of 2.4: with only 15 and 10 events, precision is limited, which is a direct consequence of event counts driving Poisson precision rather than the size of the underlying cohorts. See CASRAI’s confidence interval guide for how to read that interval width, and its confidence interval entry for the underlying definition.

When to Report an Incidence Rate Ratio

An IRR is the appropriate measure whenever person-time genuinely varies across subjects or groups being compared — which covers most real cohort studies, not just an edge case:

Open or dynamic cohorts, where subjects enter and exit on a rolling basis rather than all starting together (occupational cohorts, disease registries, ongoing surveillance systems).
Differential loss to follow-up or competing mortality between comparison groups, which shortens observed person-time unevenly even in a study that started with a fixed enrollment.
Recurrent-event outcomes, where a subject can experience the event more than once and total events (not just “did it happen, yes/no”) is the quantity of interest — a risk ratio’s binary framing does not fit this case at all.
Rare, chronic-exposure comparisons such as incidence density in occupational or environmental epidemiology, where exposure duration is itself part of what’s being compared.

A risk ratio remains the right choice for a fixed cohort with complete, equal-length follow-up and a single-occurrence outcome — a randomized trial with full follow-up and no loss to follow-up, for instance, where every subject genuinely was at risk for the same period. See CASRAI’s cohort study and case-control study guides for how study design determines which measures are available in the first place; a case-control study, notably, cannot produce an IRR or a risk ratio directly at all, since it does not sample on person-time or a fixed cohort denominator — only an odds ratio.

Common Pitfalls

Reporting an IRR as if it were a risk ratio in plain language. “2.4 times the rate” and “2.4 times the risk” are not interchangeable statements once follow-up time differs across groups, exactly as shown in the worked example above.
Assuming a constant rate across the whole follow-up period. An overall IRR, like a hazard ratio, assumes the underlying rate ratio does not itself change meaningfully across the exposure/follow-up period; if it plausibly does (e.g., risk that changes with time since exposure), consider reporting rates for defined follow-up intervals rather than one overall IRR.
Confusing incidence rate with cumulative incidence. Cumulative incidence (a proportion, bounded 0-1) answers “what fraction of the cohort had the event by time T” and pairs with a risk ratio; incidence rate (which can technically exceed 1 depending on the time unit) answers “how fast are events occurring per unit of time at risk” and pairs with the IRR. They are not the same quantity measured two ways.
Ignoring overdispersion. The Poisson-based CI above assumes event counts follow a Poisson distribution (variance equals the mean); if the actual data are overdispersed (variance exceeds the mean, common with clustered or recurrent events), a negative binomial model produces a more honest, typically wider, confidence interval than the Poisson approximation.

Frequently Asked Questions

Is incidence rate ratio the same as relative risk?

No. Relative risk (risk ratio) divides events by a fixed number of people at the start of follow-up; incidence rate ratio divides events by person-time at risk. They converge numerically when the event is rare and follow-up is complete and equal across groups, and diverge as follow-up time varies between groups or the event becomes more common, as the worked example above shows directly.

Can an incidence rate ratio be greater than any fixed upper bound, the way a proportion cannot exceed 1?

Yes. Because an incidence rate is events divided by person-time rather than a proportion of people, it is not bounded at 1 the way cumulative incidence or a risk is; a rate (and therefore an IRR) can in principle be any non-negative number, including one where the numerator rate exceeds 1 event per person-year if events recur within that time unit.

What model produces an IRR directly?

Poisson regression (or negative binomial regression when event counts are overdispersed) with person-time entered as an offset term produces incidence rate ratios directly as exponentiated coefficients, alongside adjustment for covariates the way any regression model allows. See CASRAI’s regression analysis guide for the general mechanics of interpreting model coefficients and their reported intervals.

Why does the confidence interval in the worked example look so wide relative to the point estimate?

Poisson-based precision is driven by the number of events observed, not the size of the underlying cohorts (100 and 200 subjects here) — with only 15 and 10 events, the SE(ln IRR) term is large, which is why the interval (1.08 to 5.34) spans a wide range even though the point estimate itself (2.4) is a single, precise-looking number. More events, not more enrolled subjects on their own, is what narrows this interval.

Incidence Rate Ratio: Why It Isn’t a Risk Ratio (Worked Example)

Ask about Incidence Rate Ratio: Why It Isn’t a Risk Ratio (Worked Example)