Age Standardization

Definition

Age standardization is the procedure of applying a fixed external age distribution (the "standard population") to a set of observed age-specific death rates, producing an aggregate mortality summary that removes confounding by differences in population age structure. The age-adjusted death rate (AADR) is computed as $\text{AADR} = \sum_i w_{si} \cdot R_i$ , where $w_{si}$ is the standard population's weight for age group $i$ and $R_i$ is the observed age-specific death rate. AADRs from different populations or time periods are directly comparable because they hold the age structure constant; crude death rates are not, because they conflate mortality change with demographic aging.

Key Ideas

Direct vs. indirect standardization: Direct standardization applies observed rates to a standard population: $\text{AADR} = \sum_i w_{si} \cdot R_i$ . Indirect standardization applies standard rates to the observed population and computes the Standardized Mortality Ratio (SMR): $\text{SMR} = O \big/ \sum_i R_{si} \cdot P_i$ . Direct is preferred for large populations with stable age-specific rates; indirect is preferred for small areas or sparse cause-of-death cells where age-specific rates are unstable.
The standard is arbitrary: Any reference population can serve as the standard. The choice does not change which population has higher mortality, but it does change absolute AADR levels, the magnitude of cause-specific rates, and racial/group ratios. Rates computed under different standards are not directly comparable.
US standard switch (1940 → 2000): The US switched from the 1940 standard population to the 2000 standard population for all national vital statistics, effective for data year 1999. The 2000 standard is substantially older (65+ weight: $0.068 \to 0.124$ ), raising the all-cause AADR from 503.9 to 918.5 per 100,000 in 1995 — an 82% level increase. Pre-1999 and post-1999 AADR series are not directly comparable.
Life expectancy immunity: Period life expectancy at birth ( $e_0$ ) is derived from the life table's own internal stationary population, not from an external standard. It is therefore unaffected by the choice of standard population and remains consistent across the 1999 series break.
Cause-specific amplification: Causes of death concentrated at older ages show the largest changes. Stroke AADR: 26.7 (1940 standard) → 63.9 (2000 standard) per 100,000 in 1995 — a 2.4× increase. Causes primarily affecting younger ages (infant mortality, accidents) change less.
Race ratio compression: Because the black/white mortality ratio is highest at young ages ( $\approx 2.0$ for ages 0–24) and lowest at old ages ( $\approx 1.1$ for ages 65+), the older-weighted 2000 standard places less emphasis on the ages where racial disparities are largest — compressing the aggregate black/white AADR ratio from $1.60 \to 1.41$ . This is a statistical artifact of the standard, not a real convergence in mortality.

How It Works

Direct standardization: Obtain age-specific death rates $R_i$ for the population of interest. Multiply each $R_i$ by the standard population's age-group weight $w_{si}$ . Sum: $\text{AADR} = \sum_i w_{si} \cdot R_i$ .
Indirect standardization (SMR): Apply standard age-specific rates $R_{si}$ to the local population $P_i$ to compute expected deaths: $E = \sum_i R_{si} \cdot P_i$ . Then $\text{SMR} = O / E$ where $O$ is observed deaths. $\text{SMR} > 1$ indicates higher-than-expected mortality.
Statistical inference: Each age-group death count is approximately Poisson. The AADR — a weighted sum of approximately Poisson variables — is not itself Poisson. Approximate confidence intervals (CIs) use the gamma distribution: if the observed AADR $= x$ with variance $v$ , CIs are constructed using chi-squared quantiles at $x^2/v$ (Chiang 1961; Fay and Feuer 1997).

Why It Matters

All US Centers for Disease Control and Prevention (CDC) / National Center for Health Statistics (NCHS) mortality surveillance after 1999 uses the 2000 standard. Published figures before 1999 used the 1940 standard. Any time-series analysis spanning the break must account for the standard switch explicitly.
The $\approx 82\%$ level increase in all-cause AADR at the 1999 break can mislead non-specialists into thinking US mortality nearly doubled — it did not. Only the aggregate summary measure changed.
The racial ratio compression artifact ( $1.60 \to 1.41$ ) means studies comparing black/white AADR ratios across the 1999 break will observe a spurious narrowing ( $\approx 0.19$ , or $\approx 12\%$ of the ratio) that does not reflect any real change in age-specific mortality disparities.
Healthy People 2000 goals were set using the 1940 standard; Healthy People 2010 goals required recalibration to the 2000 standard.
For small geographic units or rare causes of death, direct standardization becomes unreliable due to sparse numerators. Indirect standardization (SMR) is preferred in these settings.

Denominator Revision at Decennial Census Transitions (Anderson and Arias 2003)

A second source of non-comparability across years — distinct from the standard-population switch — is denominator revision. Death rates require a population-at-risk denominator. Between censuses, NCHS uses postcensal estimates (projected forward from the prior census). These can diverge substantially from the true population enumerated in the next census. At the 1990→2000 transition, the 1990-based estimate for July 1, 2000 underestimated the true April 1, 2000 enumeration by >6 million persons. Revising the denominator changes published death rates (often by 1–3% in aggregate, but up to 18–20% for specific race-age groups). Key patterns from the 2000 revision:

Most rates lower (1990-based denominator was too small → rates were inflated)
Exception: rates for infants and the very old (85+) higher (postcensal estimates overestimated these groups)
American Indian crude rate: −18.4% (working-age pop. severely underestimated)
Hispanic age-adjusted rate: +13.7% (elderly Hispanics overestimated; the Hispanic mortality paradox was substantially smaller than previously published)
Black infant rate: −5.3% (black infants overestimated in postcensal) This creates a trend discontinuity at every census year. Data for 2000 published with 1990-based denominators cannot be directly linked to 2001+ data (published with 2000-based denominators) without revision. Intercensal bridged estimates are produced retroactively to reconstruct a continuous series.

Open Questions

The 2000 standard will itself eventually require replacement as the US population continues to age; the question of when to switch and how to manage historical series continuity will recur.
No fully retroactively recalculated historical AADR series applying the 2000 standard to pre-1999 data exists for all causes from CDC, limiting clean long-run time-series analysis.
Whether the trend attenuation (less apparent improvement 1979–1995 under the 2000 standard vs. the 1940 standard) reflects real age-pattern heterogeneity in improvement rates or is purely compositional remains underexplored.