Written and maintained by CASRAI Editorial Board
Last updated
The standardized mortality ratio (SMR) compares the number of deaths actually observed in a study population to the number that would be expected if that population had died at the age-specific rates of a chosen reference population. It is the classic epidemiological instance of an observed-to-expected ratio, built using a technique called indirect standardization. Occupational and environmental epidemiologists, demographers, and health services researchers reach for it whenever a study population is too small, or too unevenly distributed across age groups, to compute reliable age-specific rates of its own — which is most of the time a cohort is defined by something other than the general population (a workforce, a clinical registry, a geographic area with a skewed age structure).
SMR = Observed deaths ÷ Expected deaths, conventionally multiplied by 100 so that 100 reads as “exactly as expected,” above 100 as excess mortality, and below 100 as a mortality deficit relative to the reference population.
What “Expected” Means: Indirect Standardization
The expected count in an SMR is not the reference population’s overall death rate applied as a single flat number. It is built band by band, using the reference population’s age-specific death rates (mortality almost always varies enough by age that a single overall rate would confound any comparison), applied to the study population’s own age structure:
- Split the study population into age bands and get each band’s person-time at risk (person-years, typically) or population count.
- Take the reference population’s age-specific death rate for each of those same bands — a national or regional rate is the usual source, precisely because it is large enough to produce stable age-specific rates the study population itself often can’t.
- Multiply each band’s person-time by the matching reference rate to get that band’s expected deaths, then sum across bands for the total expected count.
- Divide the study population’s actual observed deaths by that total.
This is why it’s called indirect standardization: the reference population supplies the rates, and the study population supplies the weights (its own age structure). Age is the confounding variable being controlled for — the SMR answers “how does this population’s mortality compare to the reference population’s, once we account for the fact that this population might simply skew older or younger?”
How to Calculate an SMR: Worked Example
The figures below are an illustrative, hypothetical worked example built to show the mechanics clearly — they are not data from any real cohort or published study.
Say a hypothetical occupational cohort study covers 5,000 person-years of follow-up in the 40–49 age band, 3,000 in the 50–59 band, and 1,000 in the 60–69 band. A national reference population’s age-specific annual death rates for those same three bands are 120, 350, and 900 per 100,000 person-years, respectively.
| Age band | Cohort person-years | Reference rate (per 100,000/yr) | Expected deaths |
|---|---|---|---|
| 40–49 | 5,000 | 120 | 6.0 |
| 50–59 | 3,000 | 350 | 10.5 |
| 60–69 | 1,000 | 900 | 9.0 |
| Total | 9,000 | — | 25.5 |
Total expected deaths (E) = 25.5. Suppose the cohort’s actual, observed death count (O) over the same follow-up period was 34.
SMR = 34 ÷ 25.5 = 1.33, or 133 (×100). Read as: this cohort experienced about 33% more deaths than a population with the reference population’s age-specific mortality, but with this cohort’s own age structure, would be expected to produce.
Reading the Ratio and Its Confidence Interval
A point estimate of 133 says nothing about precision on its own — it needs a confidence interval, because the observed count is a count of discrete, comparatively rare events and carries real sampling variability, especially at low event counts. The standard approach treats the observed count as Poisson-distributed and calculates an exact or approximate confidence interval on O, then divides both bounds by the fixed expected count E to get the interval for the SMR itself.
A widely used closed-form approximation is Byar’s method (Byar, 1980, as used in NHSN/PHE-style surveillance reporting and implemented in standard epidemiology software such as R’s epitools::pois.byar):
Lower bound on O = O × (1 − 1/(9O) − z/(3√O))³
Upper bound on O = (O+1) × (1 − 1/(9(O+1)) + z/(3√(O+1)))³
For the worked example above (O = 34, z = 1.96 for 95%), Byar’s approximation gives a 95% interval on the observed count of roughly 23.5 to 47.5 deaths. Dividing both bounds by the expected count (25.5) gives a 95% confidence interval for the SMR of approximately 0.92 to 1.86 (92 to 186 ×100). Because that interval spans 1.0 (100), this particular result is not statistically distinguishable from “exactly as expected” at the 95% level, even though the point estimate of 133 looks elevated on its face — a routine reminder that an SMR built on a modest number of observed deaths deserves the interval read alongside the ratio, not the ratio alone. As the observed count grows, the interval narrows around the point estimate; SMRs built on very small observed counts (roughly under 10 events) should be reported and interpreted especially cautiously, since a single additional or missing death can move the ratio substantially.
The Comparability Trap: Why Two SMRs Can’t Be Compared Directly
The single most common misuse of the SMR is treating it as if it were portable — lining up one study population’s SMR of 155 against a different study population’s SMR of 126 and concluding the first population has the worse relative mortality. That comparison is not statistically valid, even when both SMRs were calculated against the exact same reference population.
The reason is built into the mechanics above: indirect standardization uses each study population’s own age structure as the weights applied to the reference rates. Two study populations with different age distributions apply the reference rates with different weights, so their two SMRs are not adjusted onto a common, shared basis the way directly standardized rates are — each SMR only tells you how that one population compares to the reference population, not how it compares to the other study population.
The figures below are again an illustrative, hypothetical demonstration — not real data — constructed specifically to make the mechanism visible. Take two hypothetical cohorts of 10,000 person-years each, standardized against the same reference rates (1.0, 8.0, and 30.0 deaths per 1,000 person-years in a young, middle, and old age band). Cohort North is weighted toward the young band (6,000 / 3,000 / 1,000 person-years); Cohort South is weighted toward the old band (1,000 / 3,000 / 6,000 person-years). Now suppose each cohort’s true underlying age-band-specific mortality, relative to the reference rate in that same band, is actually higher in South than in North in every single age band (South’s true relative risk is 4.5× in the young band, 1.6× in the middle band, and 1.2× in the old band; North’s is 4.0×, 1.5×, and 1.1× — South at or above North everywhere):
| Cohort | Observed deaths | Expected deaths | SMR (×100) |
|---|---|---|---|
| North | 93 | 60.0 | 155 |
| South | 259 | 205.0 | 126 |
Even though South’s age-band-specific relative mortality is equal to or higher than North’s in every age band, North’s overall SMR (155) comes out higher than South’s (126) — the reversal of what the band-by-band comparison would suggest. That happens because North’s elevated risk sits disproportionately in the young band, which the reference rate treats as low-risk to begin with, and North’s own age structure gives that band far more weight (6,000 of its 10,000 person-years) than South’s does (1,000 of its 10,000). The overall ratio is dominated by where each cohort’s own person-time sits, not just by how elevated its risk is. A reader comparing “155 vs. 126” without knowing the underlying age structures would draw exactly the wrong conclusion about which cohort has the more elevated mortality pattern.
The Correct Alternative When You Need to Compare Populations
When the actual research question is “which of these populations has higher mortality, holding age constant” — a head-to-head comparison rather than each population’s comparison to a single reference — the standard fix is direct standardization instead: apply both study populations’ own age-specific rates to one shared, common standard population (a fixed set of weights, such as a published standard population), producing two age-standardized rates that were weighted identically and therefore genuinely are comparable to each other. Indirect standardization (the SMR) remains the right tool when the study population is too small for its own age-specific rates to be stable, or when the question is genuinely “how does this one population compare to the reference,” not “how do these two populations compare to each other.”
SMR vs. Related Measures
The same observed-to-expected mechanic under a different name shows up in adjacent fields. Hospital quality and patient-safety programs use the analogous standardized infection ratio (SIR) and other O/E-style risk-adjustment ratios to compare a facility’s observed adverse-event count to a model-predicted expected count — conceptually the same indirect-standardization logic, applied to infections or complications instead of deaths, and built for a hospital quality-improvement audience rather than a population-epidemiology one. Relative risk, by contrast, compares two directly measured rates to each other and doesn’t involve standardization at all. Cohort studies and case-control studies are the study designs that most commonly generate the person-time and death counts an SMR is calculated from; the Poisson distribution is the statistical model underlying both the observed-death count itself and the confidence-interval methods (including Byar’s approximation) used to bound it.
Frequently Asked Questions
What does an SMR of 100 mean?
It means the study population’s observed deaths exactly matched the number expected, given the reference population’s age-specific rates applied to this population’s own age structure — no excess and no deficit relative to the reference.
Can I compare an SMR to a national average directly?
Yes — that comparison (this population vs. the single reference population it was standardized against) is exactly what an SMR is built to support. The invalid comparison is SMR-to-SMR between two different study populations, covered above.
Why not just compute age-specific rates for the study population itself and skip standardization?
Because many study populations are too small to produce stable age-specific rates of their own — a handful of deaths split across several age bands produces wildly unstable per-band rates. Indirect standardization borrows the reference population’s much larger, more stable age-specific rates and applies them to the study population’s actual age structure instead.
What is the difference between SMR and the standardized incidence ratio (SIR)?
Same indirect-standardization mechanic, different outcome: SMR counts deaths, while a standardized incidence ratio (or, in hospital infection surveillance, a standardized infection ratio) counts new cases of disease or a defined adverse event over the same kind of observed-vs-expected comparison.
Does a wide confidence interval mean the SMR is wrong?
No — it means the observed event count the SMR is built on isn’t large enough yet to pin the true ratio down precisely. The point estimate can still be the best available estimate; the interval just quantifies how much sampling uncertainty surrounds it, which is exactly why the interval should be read alongside the ratio rather than the ratio alone.
