Skip to content
Back to search results

Indirect Age Standardization

Introduced by
Classical demography / vital statistics (formalized in Preston, Heuveline & Guillot)
Year
2001
Type
Rate-standardization pipeline for comparing populations under unstable stratum rates
Family
Demographic standardization / mortality analysis
Font size

What does it do?

Indirect age standardization is a demographic technique for comparing the overall event rate (most often mortality) of a study population to a reference, when the population's own age-specific rates are too sparse or unstable to standardize directly.

Instead of applying the study population's rates to a standard age structure, it does the reverse: it applies a stable set of standard age-specific rates to the study population's age distribution to compute the number of events that would be expected under the standard schedule. The ratio of observed to expected events is the standardized mortality (or morbidity) ratio, the SMR, and multiplying it by the standard's crude rate yields an indirectly standardized rate. The method is a staple of vital statistics and occupational and small-area epidemiology, and is developed from first principles in Preston, Heuveline and Guillot's demography text.

The intuition behind the method

Suppose you want to know whether workers in a particular factory die at unusually high rates, but the factory has only a few hundred people, so in some age bands there are only one or two deaths. Computing the factory's own age-specific death rates and applying them to a standard population — the direct method — would give wildly unstable numbers, because a single extra death in a small age band swings the rate enormously. Indirect standardization sidesteps this. You borrow trustworthy age-specific death rates from a large reference population (say the national rates), and ask: given the factory's age structure, how many deaths should we have expected if these workers died at exactly the national rates? That expected count is stable because it leans on the big reference, not the tiny sample. You then compare the deaths you actually observed to that expected count. If the ratio is 1.0 the factory is dying at the national pace; 1.3 means thirty percent more deaths than expected; 0.8 means twenty percent fewer. The whole appeal is that it needs only the total number of deaths from the small group, not a reliable rate in every age band.

History

Standardization of rates dates to nineteenth-century vital statistics, where demographers and public-health pioneers sought to compare mortality across populations with different age compositions without being misled by age structure alone. The indirect method, built around the standardized mortality ratio, became the workhorse of occupational and small-area epidemiology precisely because such studies routinely involve small groups and rare events. Preston, Heuveline and Guillot's 2001 text Demography: Measuring and Modeling Population Processes presents the direct and indirect methods from first principles and situates them within the broader machinery of rate analysis and life tables. The later recognition, developed in the rate-modeling tradition exemplified by Frome's 1983 Biometrics paper, that the SMR is the natural output of a Poisson regression with a log-expected-count offset, unified classical standardization with modern generalized linear models and is how the calculation is performed today.

When to use it

Use indirect age standardization when you must compare a population's overall event rate to a reference but the population is too small or the events too rare to yield reliable age-specific rates of its own — the classic settings being occupational cohorts, small geographic areas, and rare causes of death. It is the method of choice when you can obtain the total event count and the age structure of the study group plus a trustworthy external standard, but cannot trust the study group's own stratum rates. Prefer direct standardization instead when stratum-specific rates are stable in every population being compared and you want to compare several populations to each other, because indirectly standardized rates from different study populations are not strictly comparable to one another (each is anchored to its own age structure). Avoid relying on a single summary SMR when the underlying stratum-specific rate ratios vary strongly with age, since the SMR can then mask important age-patterned effects.

Limitations and considerations

  • Indirectly standardized rates from different study populations are anchored to different age structures and are not, strictly speaking, comparable to one another.
  • A single SMR can be misleading when the rate ratio between study and standard varies across age groups, hiding age-specific heterogeneity.
  • Results depend on the choice of standard, and an ill-matched standard can distort the expected count and the resulting ratio.
  • The method assumes the standard schedule is essentially error-free; when the standard itself is uncertain, the simple Poisson interval understates total uncertainty.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Remains stable when the study population is small or events are rare, because it needs only the total observed count, not reliable stratum-specific rates.
  • Requires minimal data from the study group — age structure and total events — which is often all that is available for occupational or small-area studies.
  • Produces an intuitive, widely understood summary (the SMR) with a simple Poisson-based confidence interval.
  • Connects directly to Poisson regression via an offset, allowing covariate adjustment and formal modeling once the basic SMR is computed.

Applications

  • Comparing mortality in an occupational cohort to national rates to detect excess risk from a workplace exposure.
  • Producing standardized mortality or morbidity ratios for small geographic areas where direct standardization is too unstable.
  • Assessing whether a specific cause of death is elevated in a defined subpopulation relative to a reference schedule.
  • Summarizing relative mortality experience in registry and surveillance reports where only totals and age structure are reliable.
  • Serving as the descriptive first step before fitting Poisson rate regression with covariate adjustment.

How it works

How it works

1

The pipeline begins by partitioning both the study population and a reference population into the same age strata (often five-year age groups, sometimes crossed with sex). From the study population you need only the person-time or population count in each stratum and the total observed events; you do not need reliable stratum-specific rates, which is exactly the situation indirect standardization is designed for. From the reference you take a stable set of standard age-specific rates, typically derived from a large population such as a national vital-statistics series. Preston, Heuveline and Guillot stress that the choice of standard is a substantive decision: it defines the schedule against which the study population is judged, and comparisons across studies are only coherent if they share a standard.

Formula
\text{study: } \{N_i\}_{i=1}^{K},\ D=\sum_i D_i;\quad \text{standard: } \{m_i^{s}\}_{i=1}^{K},\ m_i^{s}=\tfrac{D_i^{s}}{N_i^{s}}
2

With the standard rates and the study population's age structure in hand, the analyst computes the expected number of events by multiplying each stratum's study population by the corresponding standard rate and summing across strata. This expected count is the number of deaths or cases the study population would have generated had it experienced the standard's age-specific risks. Because the standard rates come from a large, stable source, the expected total is well determined even when the study group is small, which is the key advantage that makes the method usable where direct standardization fails.

Formula
E=\sum_{i=1}^{K} N_i\, m_i^{s}
3

The core comparison is the standardized mortality (or morbidity) ratio, formed by dividing the total observed events by the total expected events. An SMR of one means the study population experiences events at exactly the rate implied by the standard schedule; values above one indicate excess risk and values below one indicate a deficit. Crucially, the SMR uses only the study population's total event count, so it remains stable even when individual strata are sparse. Preston, Heuveline and Guillot present the SMR as the natural summary that indirect standardization produces.

Formula
\mathrm{SMR}=\dfrac{O}{E}=\dfrac{\sum_i D_i}{\sum_i N_i\, m_i^{s}}
4

To express the result on the familiar scale of a rate rather than a ratio, the SMR is multiplied by the crude rate of the standard population. The product is the indirectly standardized rate: an estimate of what the study population's overall rate would be if it shared the standard's age structure and its event level were scaled by the observed SMR. This converts the dimensionless ratio into a comparable per-person-time figure, which is useful when reporting alongside crude and directly standardized rates.

Formula
\hat{m}^{\,ISR}=\mathrm{SMR}\times \bar{m}^{s},\qquad \bar{m}^{s}=\dfrac{\sum_i N_i^{s} m_i^{s}}{\sum_i N_i^{s}}
5

Finally, uncertainty is quantified by treating the observed event count as a Poisson variable, since the expected count is regarded as fixed. The variance of the log-SMR is approximately the reciprocal of the observed count, giving a simple confidence interval that widens sharply when events are few. As Frome shows, this Poisson footing connects indirect standardization directly to Poisson rate regression: the SMR is essentially the exponentiated intercept of a Poisson model with the log of expected counts as an offset, which is how analysts now compute it and add covariates.

Formula
O\sim\mathrm{Poisson}(\mu),\quad \widehat{\mathrm{Var}}(\ln \mathrm{SMR})=\tfrac{1}{O},\ \ \mathrm{CI}_{95\%}=\mathrm{SMR}\,e^{\pm 1.96/\sqrt{O}}

Common pitfalls

  • Comparing SMRs from two study populations directly as if they were comparable rates, ignoring that each is conditioned on its own age distribution.
  • Choosing a standard population whose age-specific rate pattern is very different from the study group, which makes the implicit proportionality assumption untenable.
  • Reporting an SMR without its confidence interval when events are few, where a count of three versus six can swing the estimate dramatically.
  • Confusing indirect with direct standardization, or mixing the two within a single comparison and misinterpreting the resulting figures.

Frequently asked questions

How does indirect standardization differ from direct standardization?

Direct standardization applies the study population's own age-specific rates to a standard age structure, so it needs reliable stratum-specific rates in every population. Indirect standardization does the opposite: it applies standard age-specific rates to the study population's age structure to obtain an expected count, then compares observed to expected as the SMR. Indirect is used when the study group is small or events are rare, so its own stratum rates are unstable. Preston, Heuveline and Guillot present both as complementary tools, with the choice driven by data stability.

Why can't I just compare two SMRs directly?

Each indirectly standardized result is conditioned on its own population's age structure, because the expected count is computed from that structure. Two SMRs therefore answer slightly different questions — each compares its own group to the common standard, not to each other. When age-specific rate ratios are roughly constant across age the comparison is approximately valid, but when they vary by age, comparing SMRs can be misleading. If direct comparison of several populations is the goal and stratum rates are stable, direct standardization is the cleaner choice.

How do I get a confidence interval for the SMR?

Treat the observed event count as Poisson with the expected count held fixed. The variance of the log-SMR is approximately one over the observed count, giving an interval of SMR times exp(plus or minus 1.96 over the square root of observed events). This interval widens sharply when events are few, which is exactly when caution is needed. Equivalently, as Frome's rate-regression framework shows, you can fit a Poisson model with the log of expected counts as an offset and read the SMR and its interval directly from the exponentiated intercept.

Related methods

Sources

  1. Preston, S. H., Heuveline, P., & Guillot, M. (2001). Demography: Measuring and Modeling Population Processes. Blackwell Publishers.

    ISBN: 9781557864512
    Copied.
  2. Frome, E. L. (1983). The Analysis of Rates Using Poisson Regression Models. Biometrics, 39(3), 665-674.

    Open DOI reference
    Copied.