Maine's crude death rate in 2021 was 49% above Alaska's. But the National Center for Health Statistics (NCHS) publishes a second death rate for every state each year, the age-adjusted rate, and by that rate Alaska's mortality is the higher. Here are both rates for four states (Murphy et al., 2024, Table 17). Which comparison should we believe, and how is the second rate made?
| State | Deaths | Crude rate | Age-adjusted rate | Share aged 65 and over |
|---|---|---|---|---|
| Maine | 17,269 | 1,258.4 | 889.4 | 21.7% |
| Alaska | 6,208 | 847.3 | 931.0 | 13.3% |
| Florida | 261,369 | 1,200.0 | 832.9 | 21.1% |
| Utah | 22,569 | 676.1 | 815.8 | 11.7% |
Rates are per 100,000; the shares aged 65 and over are from the Census Bureau's estimates for July 1, 2021. By the crude rates, Maine's mortality is 49% above Alaska's and Florida's 77% above Utah's. By the age-adjusted rates, both orders reverse. Alaska's rate is the higher of the first pair, and Florida's is only 2% above Utah's. The two older states, Maine and Florida, had the highest shares aged 65 and over of any state that year. Alaska and Utah were among the lowest.
Rates in Public Health showed why the crude rates mislead. A crude rate is a weighted average of age-specific rates, with the population's own age distribution as the weights. Death rates rise steeply with age. So an older population has a higher crude rate even if its residents are no more likely to die at any given age. This article develops the two classical remedies, and it is where the course defines them.
- Direct standardization, which produced the adjusted rates in the table, applies each population's age-specific rates to a single standard population.
- Indirect standardization applies one standard set of age-specific rates to each population. It then compares the deaths observed with the deaths that set of rates would produce.
First we define the directly standardized rate and work it on U.S. deaths in 2007 and 2013, when the crude and adjusted rates moved in opposite directions. Then we prove that two adjusted rates cannot reverse that way, and see how much the choice of standard still matters. Next we put a confidence interval on an adjusted rate. Finally we turn to indirect standardization, compare the four states, and ask when each method is the right one.
The worked examples use deaths of U.S. residents by age from CDC WONDER, the federal query system for vital statistics, and the 2021 state figures of NCHS. Every computed rate below reproduces the rate NCHS published, with one rounding exception noted where it arises.
Direct standardization
Here is the idea in one sentence. To compare two populations at the same ages, ask what each one's crude rate would be if it had the same age structure as some agreed reference population.
The formal version. Let a population be divided into $k$ age groups, with $d_i$ deaths in a population (or person-years) $n_i$ in group $i$. The age-specific rates are $r_i = d_i/n_i$. Choose a standard population with $P_{si}$ people in group $i$ and $P_s = \sum_i P_{si}$ in all. The standard weights $w_i = P_{si}/P_s$ are positive and sum to 1. The directly standardized rate (DSR), or age-adjusted rate, is
$$\text{DSR} = \sum_{i=1}^{k} w_i\, r_i = \frac{\sum_i P_{si}\, r_i}{P_s}.$$
The second form explains the name and the interpretation. $P_{si} r_i$ is the number of deaths the standard population would have in group $i$ if it experienced the study population's rate there. So the DSR is the crude death rate the standard population would have with the study population's age-specific rates. Every population being compared is given the same weights. Differences in the DSR can therefore come only from differences in the age-specific rates.
The crude rate has the same form, with the population's own shares $n_i/N$ as weights. A DSR is therefore a crude rate computed for a fictitious population, and its value depends on the standard chosen. NCHS (Murphy et al., 2024, Technical Notes) says that age-adjusted rates "should be viewed as relative indexes rather than as actual measures of mortality risk". Anderson and Rosenberg (1998) note that they are meaningful only in comparisons of groups or times adjusted to the same standard.
Which population should serve as the standard? For U.S. vital statistics, the question has an official answer.
The U.S. standard population
Since data year 1999, NCHS and the other agencies of the Department of Health and Human Services adjust death rates to the 2000 U.S. standard population. It is the Census Bureau's projection of the U.S. population for the year 2000, in eleven age groups (Anderson and Rosenberg, 1998; Klein and Schoenborn, 2001). The figures are those of Murphy et al. (2024, Table X):
| Age | Standard population | Weight $w_i$ |
|---|---|---|
| Under 1 | 3,794,901 | 0.013818 |
| 1–4 | 15,191,619 | 0.055316 |
| 5–14 | 39,976,619 | 0.145563 |
| 15–24 | 38,076,743 | 0.138646 |
| 25–34 | 37,233,437 | 0.135575 |
| 35–44 | 44,659,185 | 0.162614 |
| 45–54 | 37,030,152 | 0.134835 |
| 55–64 | 23,961,506 | 0.087249 |
| 65–74 | 18,135,514 | 0.066035 |
| 75–84 | 12,314,793 | 0.044841 |
| 85 and over | 4,259,173 | 0.015509 |
| All ages | 274,633,642 | 1 |
Klein and Schoenborn (2001, Table 2) publish six-decimal weights computed from the same projection rounded to thousands. They differ from these in the sixth decimal place (0.055317 for ages 1 to 4, for example). NCHS replaced the rounded weights with the full population counts beginning with 2003 data and reports that the effect is negligible (Murphy et al., 2024, Technical Notes). In the example below, both sets give the same adjusted rates to two decimals.
A worked example: 2007 and 2013
Now the method can meet national data. Between 2007 and 2013, the number of deaths of U.S. residents rose from 2,423,712 to 2,596,993. The crude death rate rose from 804.6 to 821.5 per 100,000, an increase of 2.1%.
Check your understanding. Over the same years the U.S. population grew older. Does the rise in the crude rate mean that the risk of dying at a given age rose?
Not necessarily: an older population has a higher crude rate even at unchanged rates by age. And here the risk fell. The age-specific rates tell a different story from the crude rate, and the age-adjusted rate fell by 5.6%:
| Age | Weight | Rate 2007 | $w_i r_i$ | Rate 2013 | $w_i r_i$ | Ratio 2013/2007 |
|---|---|---|---|---|---|---|
| Under 1 | 0.013818 | 702.5 | 9.71 | 594.7 | 8.22 | 0.847 |
| 1–4 | 0.055316 | 29.4 | 1.63 | 25.5 | 1.41 | 0.868 |
| 5–14 | 0.145563 | 15.2 | 2.21 | 13.0 | 1.89 | 0.855 |
| 15–24 | 0.138646 | 78.8 | 10.92 | 64.8 | 8.99 | 0.823 |
| 25–34 | 0.135575 | 107.2 | 14.53 | 106.1 | 14.39 | 0.990 |
| 35–44 | 0.162614 | 186.0 | 30.25 | 172.0 | 27.97 | 0.925 |
| 45–54 | 0.134835 | 420.3 | 56.67 | 406.1 | 54.75 | 0.966 |
| 55–64 | 0.087249 | 866.7 | 75.61 | 860.0 | 75.04 | 0.992 |
| 65–74 | 0.066035 | 1,976.0 | 130.48 | 1,802.1 | 119.00 | 0.912 |
| 75–84 | 0.044841 | 4,987.1 | 223.62 | 4,648.1 | 208.43 | 0.932 |
| 85 and over | 0.015509 | 14,160.9 | 219.62 | 13,660.4 | 211.85 | 0.965 |
| Adjusted rate | 775.3 | 731.9 | 0.944 |
The rates are per 100,000, computed from WONDER's deaths and populations. The products are shown to two decimals. Their sums, computed before rounding, are 775.25 and 731.92, and NCHS publishes 775.3 and 731.9 for these years (Murphy et al., 2024, Table 6). So the age-adjusted rate fell by 5.6% while the crude rate rose by 2.1%. Every one of the eleven age-specific rates fell, by between 0.8% (ages 55 to 64) and 17.7% (ages 15 to 24).
How can the crude rate rise when every rate inside it falls? Because the population aged. The share of residents aged 65 and over grew from 12.6% to 14.1%, and the share aged 85 and over from 1.67% to 1.91%. In the crude rate, the weights moved toward the age groups with the highest rates. They moved by enough to outweigh the decline in every one of those rates.
Lower at every age, lower when adjusted
The example has the structure of what is often called Simpson's paradox: a comparison that holds within every subgroup reverses when the subgroups are pooled. (The general form, and its link to confounding, is in Conditional Probability and the Law of Total Probability.) Could a different standard population produce the same kind of reversal between two adjusted rates? It cannot. The reason is that every adjusted rate uses the same weights, so a population that is lower in every term of the sum is lower in the sum.
Proposition. Let populations A and B have age-specific rates $r_{Ai}$ and $r_{Bi}$ with $r_{Ai} \le r_{Bi}$ for every age group $i$. Then for every set of standard weights $w_i \ge 0$ with $\sum w_i = 1$, $\text{DSR}_A \le \text{DSR}_B$. The inequality is strict if some $w_i > 0$ has $r_{Ai} < r_{Bi}$.
Proof. $\text{DSR}_B - \text{DSR}_A = \sum_i w_i (r_{Bi} - r_{Ai})$. Each term is a nonnegative weight times a nonnegative difference, and a term is positive if $w_i > 0$ and $r_{Bi} > r_{Ai}$. $\square$
The crude rates are not covered, because they give the two populations different weights. Write $a_i$ and $b_i$ for the age shares of A and B. Then $R_A - R_B = \sum_i (a_i r_{Ai} - b_i r_{Bi})$. If $a_i$ is larger where the rates are large, the sum can be positive even though $r_{Ai} < r_{Bi}$ in every term.
Check your understanding. Seven standard populations are tried in the table below, from the young 1940 U.S. population to the old 2013 one. Under which of them is the 2013 adjusted rate higher than the 2007 one?
None. Every age-specific rate fell, so the proposition says the adjusted rate must fall under every standard, and the script confirms it:
| Standard | 2007 | 2013 | Change |
|---|---|---|---|
| U.S. 2000 | 775.3 | 731.9 | −5.59% |
| U.S. 1940 | 413.2 | 388.3 | −6.01% |
| WHO World | 505.6 | 475.6 | −5.93% |
| Segi world | 433.8 | 407.2 | −6.13% |
| European | 651.4 | 614.5 | −5.65% |
| U.S. population, 2007 | 804.5 | 761.1 | −5.39% |
| U.S. population, 2013 | 867.8 | 821.5 | −5.34% |
The last two rows standardize both years to one year's actual population. Standardized to its own population, a year's adjusted rate is its crude rate (821.46 for 2013, computed from the deaths of known age). That is the weighted-average identity of the previous article, read in reverse. The three international standards, described below, group ages 0 to 4 together, so for them the script merges the first two age groups of the data.
The same arithmetic explains Maine and Alaska. NCHS's adjusted rates say that Alaska's age-specific rates are, on the whole, the higher. Maine's crude rate is the higher because 21.7% of its residents were 65 or over, against 13.3% of Alaska's. Anderson and Rosenberg (1998, Table A) illustrate the same reversal with two hypothetical groups. One has 60% of its members aged 65 and over and lower death rates at every age; it has the higher crude rate and the lower adjusted rate.
The proposition guarantees the direction of a comparison only when one population is lower at every age. When it is not, how much does the choice of standard matter?
Choosing the standard
Which standard to use is a convention, not a fact about the data, and the convention has a history.
- Ahmad et al. (2001) trace the method to the actuary Neison. In 1844 he answered a proposal to compare London's sanitary districts by their mean age at death. He set that crude mean beside a directly standardized one, and he introduced the ideas of direct and indirect standardization and the term "standard population".
- The Registrar General of England and Wales first used direct standardization in its report of 1883. It later fixed the 1901 census population as a permanent standard, and the United States used that 1901 standard until the early 1940s (Ahmad et al., 2001).
- NCHS and the states then used the U.S. population of 1940, from 1943 until the 2000 standard replaced it (Anderson and Rosenberg, 1998).
- Internationally, the World Health Organization adopted a WHO World Standard in 2001. It is the average age structure projected for the world's population over 2000 to 2025, in 5-year groups up to 100 and over. Two older international standards are the Segi "world" population, younger, and the Scandinavian or "European" population, older (Ahmad et al., 2001, Tables 1 and 4).
Ahmad et al. state the difficulty plainly: "There is clearly no conceptual justification for choosing one standard over another, hence the choice is arbitrary."
The standards differ most in the weight they give to old age:
| Standard | Weight on 65 and over | Weight on 85 and over |
|---|---|---|
| U.S. 2000 | 12.6% | 1.55% |
| U.S. 1940 | 6.85% | 0.28% |
| WHO World | 8.23% | 0.63% |
| Segi world | 7.0% | 0.50% |
| European | 11.0% | 1.00% |
The size of the adjusted rate
Because old age carries the highest rates, the standard sets the level of the adjusted rate. The 2019 U.S. adjusted death rate is 715.25 per 100,000 against the 2000 standard and 387.4 against the 1940 standard, a ratio of 1.85. (NCHS publishes 715.2 for 2019. WONDER's counts give 715.2517, just on the other side of the rounding boundary. It is the one computed rate in the article that does not round to the published figure.)
Anderson and Rosenberg (1998) report the same effect when the standard changed. The 1995 rate was 503.9 on the 1940 standard and 918.5 on the 2000 standard, much closer to the crude rate of 880.0. The level of an adjusted rate has no meaning on its own. That is why adjusted rates are compared only with others adjusted to the same standard.
The direction of a trend
When the age-specific rates move in parallel, the choice of standard changes the level of the adjusted rates but not the comparison between them. When they do not, the comparison depends on the standard. Anderson and Rosenberg (1998) make the point with cancer mortality from 1979 to 1995. It fell at young ages and rose at old ages, while the adjusted rate on the 1940 standard barely moved. The years 2010 and 2019 give a recent example.
From 2010 to 2019, death rates fell for infants (by 11%), young children, ages 45 to 54 and every group from 65 up. They rose for ages 5 to 44 and 55 to 64, by 25% at ages 25 to 34 and 17% at 35 to 44. Much of the rise at those ages came from drug-induced deaths (Murphy et al., 2024, Table 6):
- at ages 25 to 34, the drug-induced death rate went from 19.2 to 36.9 per 100,000, about 68% of the increase in the all-cause rate;
- at ages 35 to 44, it went from 21.7 to 42.3, about 72% of the increase.
Meanwhile the crude rate rose from 799.5 to 869.7, by 8.8%, as the population aged.
Check your understanding. Rates fell at the oldest ages and rose at many younger ones. Which standard will show the larger decline from 2010 to 2019: the old 2000 standard, or the young 1940 one?
The old one. The more weight a standard gives to the ages whose rates fell, the larger the decline it reports. The adjusted rates:
| Standard | 2010 | 2019 | Change |
|---|---|---|---|
| U.S. 2000 | 747.0 | 715.25 | −4.25% |
| U.S. 1940 | 394.8 | 387.4 | −1.90% |
| WHO World | 484.2 | 470.1 | −2.92% |
| Segi world | 414.2 | 404.2 | −2.41% |
| European | 626.0 | 604.8 | −3.39% |
| U.S. population, 2010 | 799.5 | 766.9 | −4.08% |
| U.S. population, 2019 | 909.1 | 869.7 | −4.33% |
| U.S. 2000, ages under 55 only | 149.9 | 156.8 | +4.58% |
Every standard that covers all ages shows a decline, but its size varies by a factor of more than two. It runs from 1.9% on the young 1940 standard to 4.2% on the old 2000 standard and 4.3% on the 2019 population. The last row restricts the 2000 standard to ages under 55, renormalizing its weights to sum to 1, as a summary of mortality below middle age would. It shows a rise of 4.6%.
Neither summary is wrong. They answer different questions, and the single adjusted rate hides the divergence that the age-specific rates show. Anderson and Rosenberg (1998) recommend that adjusted rates be supplemented with age-specific rates where the two can tell different stories. Breslow and Day (1987, §2.2) give an example, thyroid cancer in Connecticut, where an adjusted incidence rate hid a new peak among young women.
Two smaller choices also matter.
- A standard must use the same age groups as the data.
- More groups control the age difference more tightly, but small groups give unstable specific rates. Klein and Schoenborn (2001) note that two adult standards differing only in how ages 25 to 44 are grouped give slightly different adjusted rates.
The precision of an adjusted rate
An adjusted rate is built from counts of deaths, so it carries the random variation of Rates in Public Health. How much? Here is the idea. The DSR is a fixed weighted sum of independent rates, so its variance is the sum of the squared weights times the variances of the rates.
The formal version. Treat the deaths $d_i$ in the age groups as independent Poisson counts, with the populations $n_i$ and weights $w_i$ fixed. Then $\operatorname{Var}(r_i) = \operatorname{Var}(d_i)/n_i^2$, estimated by $d_i/n_i^2$. Since the DSR is a fixed linear combination of independent rates,
$$\operatorname{Var}(\text{DSR}) = \sum_i w_i^2 \operatorname{Var}(r_i) = \sum_i w_i^2\, \frac{d_i}{n_i^2}, \qquad \operatorname{SE}(\text{DSR}) = \sqrt{\sum_i w_i^2\, \frac{d_i}{n_i^2}}.$$
This is the formula NCHS uses (Murphy et al., 2024, Technical Notes; Anderson and Rosenberg, 1998, equations 6 to 8). Breslow and Day (1987, §2.2) give the same standard error, citing Chiang (1961) for treating the counts in different age groups as uncorrelated.
Back to 2007 and 2013. The SE is 0.500 per 100,000 for 2007 and 0.460 for 2013. The normal 95% intervals are (774.27, 776.23) and (731.02, 732.82). The two adjusted rates differ by
$$z = \frac{775.25 - 731.92}{\sqrt{0.500^2 + 0.460^2}} = 63.8$$
standard errors. With millions of deaths, random variation in national adjusted rates is negligible. Their differences are real differences in the age-specific rates (or in how deaths and populations were counted).
Small populations and the gamma interval
The formula has a weakness that Breslow and Day (1987, §2.2) point out: the weights take no account of how precisely each rate is estimated. A small age group with a large weight can dominate the variance. In a small population the adjusted rate is far from normally distributed, and a normal interval then has the defects of the normal interval for a single small count.
Fay and Feuer (1997) proposed an interval based on the gamma distribution. Here is the idea. A DSR is a weighted sum of independent Poisson variables. It is not itself Poisson, but it can be approximated by a gamma distribution with the same mean and variance, and the interval is read from that gamma distribution.
Write $x = \text{DSR}$ for the mean and $v = \operatorname{Var}(\text{DSR})$ for the variance. The matching gamma distribution has shape $x^2/v$ and scale $v/x$. The lower 95% limit is the corresponding gamma percentile,
$$L = \frac{v}{x}\, G^{-1}_{x^2/v}(0.025),$$
where $G^{-1}_a$ is the quantile function of the gamma distribution with shape $a$ and scale 1. For the upper limit, the mean and variance are first increased to allow for the discreteness of the counts. The increase is $k = \max_i (w_i/n_i)$, the increase in the DSR from one more death in the group where a death counts most:
$$U = \frac{v + k^2}{x + k}\, G^{-1}_{(x + k)^2/(v + k^2)}(0.975).$$
This is the form Anderson and Rosenberg (1998, Technical notes, equations 9 to 14) give for NCHS.
Two checks connect the interval to what we already know. When the study population has the same age structure as the standard, the DSR is a fixed multiple of a single Poisson count. The interval then reduces to the exact Poisson interval of the previous article. Fay and Feuer report that it is exact in that case and conservative, in their simulations, when the two populations differ. The quantity $x^2/v$ acts as an effective number of deaths, and Anderson and Rosenberg (1998) suggest the normal interval when it is 100 or more. For 2007 it is 2,405,972, and the gamma and normal intervals agree to the second decimal.
To see the difference in a small population, the script runs a simulation:
- it builds populations of 2,000, 5,000 and 20,000 with Alaska's 2021 age structure;
- it assigns them the U.S. 2019 age-specific death rates as their true rates;
- it draws Poisson deaths in each age group and adjusts to the 2000 standard.
The true adjusted rate is 715.25 per 100,000. A population of 5,000 expects 32.3 deaths a year. In one replicate it had 28, distributed over the eleven age groups as 0, 0, 0, 0, 1, 1, 3, 6, 6, 7, 4. Its crude rate is 560.0 and its adjusted rate 607.1, with SE 120.9. The normal interval is (370.3, 844.0), and the gamma interval (393.8, 909.2) is shifted upward. Here $k$ is 33.7 per 100,000, from the 85-and-over group, where each death in a group of about 46 people carries a weight of 0.0155.
Check your understanding. A population of 2,000 expects about 13 deaths. Will the normal interval cover the true adjusted rate more or less often than 95%?
Less often, as the normal interval for a single small count did in the previous article. The figure and the numbers below show by how much.
Over 30 seeds of 10,000 replicates each, the normal interval covered the true rate in a median of:
- 92.8% of replicates for populations of 2,000 (about 13 expected deaths);
- 94.0% for 5,000;
- 94.7% for 20,000 (about 129 deaths).
The gamma interval covered in 97.2%, 96.4% and 95.7%. Its conservatism has a price in width. The mean widths were 926 against 812 per 100,000 for 2,000 people, 559 against 518 for 5,000, and 269 against 260 for 20,000. For populations of 2,000, 0.35% of normal intervals had a negative lower limit.
What if a population's deaths by age are too few, or not known at all?
Indirect standardization
Direct standardization needs the age-specific death counts of every population compared. For many populations these are unavailable or too small to trust: a county, an occupational group, a cohort of workers exposed to a chemical. Indirect standardization needs only two things from the study population, its total number of deaths and its age structure. It also needs a set of standard age-specific rates $R_i$ from a large reference population.
Here is the idea in one sentence. Ask how many deaths the study population would have had at the standard rates, given its own ages, and compare that with how many it actually had.
The expected deaths are the deaths the study population would have if it experienced the standard rates at every age:
$$E = \sum_i n_i R_i.$$
The standardized mortality ratio compares the deaths observed, $O = \sum_i d_i$, with this expectation:
$$\text{SMR} = \frac{O}{E}.$$
An SMR of 1.10 means 10% more deaths than the standard rates predict for a population of that age structure. Multiplying by the standard population's crude rate $R_s$ gives the indirectly standardized rate, $\text{ISR} = \text{SMR} \times R_s$ (Anderson and Rosenberg, 1998, Technical notes, equations 4 and 5).
Breslow and Day (1987, §2.3) trace the SMR to the eighteenth century, and its use for comparing occupations to Farr, in the Registrar General's report of 1855. They also show that it is a weighted average of the age-specific ratios $r_i/R_i$, with weights equal to the expected deaths $e_i = n_i R_i$:
$$\frac{\sum_i e_i (r_i/R_i)}{\sum_i e_i} = \frac{\sum_i n_i r_i}{E} = \frac{O}{E}.$$
Suppose the study population's rates are a constant multiple $\theta$ of the standard rates. Then every ratio is $\theta$ and the SMR estimates it. These weights are the ones that minimize the variance of the estimate (Breslow and Day, 1987, §2.3).
Maine and Alaska against the United States
NCHS publishes each state's total deaths, but this article's data do not include deaths by age for the states. So the four states are compared by indirect standardization. The standard is the 2021 U.S. age-specific rates (Murphy et al., 2024, Table 2), applied to the Census estimates of each state's population by age.
For Maine, the expected deaths by age group, from under 1 to 85 and over, are
$$62.9,\ 12.8,\ 20.4,\ 137.3,\ 301.3,\ 475.8,\ 907.0,\ 2{,}383.2,\ 3{,}939.9,\ 4{,}303.2,\ 4{,}718.1,$$
summing to $E = 17{,}261.9$. Maine recorded $O = 17{,}269$ deaths, so $\text{SMR} = 17{,}269/17{,}261.9 = 1.0004$. For Alaska, the expected deaths are
$$51.6,\ 9.7,\ 14.7,\ 85.2,\ 205.4,\ 292.6,\ 441.6,\ 1{,}009.6,\ 1{,}429.0,\ 1{,}254.2,\ 1{,}060.8,$$
summing to $E = 5{,}854.3$ against $O = 6{,}208$ observed, so $\text{SMR} = 1.0604$. Maine's mortality was what the national rates predict for a population of its ages; Alaska's was 6% higher. These are the predictions Rates in Public Health made from the same national rates.
| State | Observed | Expected | SMR | Exact 95% interval | ISR | DSR / U.S. DSR |
|---|---|---|---|---|---|---|
| Maine | 17,269 | 17,261.9 | 1.0004 | (0.9855, 1.0154) | 1,044.2 | 1.0110 |
| Alaska | 6,208 | 5,854.3 | 1.0604 | (1.0342, 1.0871) | 1,106.8 | 1.0583 |
| Florida | 261,369 | 279,888.5 | 0.9338 | (0.9303, 0.9374) | 974.7 | 0.9468 |
| Utah | 22,569 | 24,973.7 | 0.9037 | (0.8920, 0.9156) | 943.3 | 0.9274 |
The indirectly standardized rates multiply each SMR by the U.S. crude rate in 2021, 1,043.8 per 100,000. They are on the scale of that crude rate rather than of the 2000 standard.
The last column divides each state's published adjusted rate by the U.S. adjusted rate, 879.7. Breslow and Day (1987, §2.3) call this ratio of directly standardized rates the comparative mortality figure (CMF). The SMR and the CMF tell the same story for every state, as Breslow and Day observe they usually do in practice. They never differ here by more than 0.025.
Check your understanding. Maine's SMR is 1.0004, and its exact interval contains 1. Does that show that Maine's death rates matched the nation's at every age?
It does not. Here is what the SMR does say. Maine's crude death rate was 21% above the nation's (1,258.4 against 1,043.8). The SMR says that the whole of that excess is accounted for by its age structure: for a population of its ages, Maine recorded almost exactly the deaths the national rates predict.
But the SMR is an average of Maine's age-specific ratios $r_i/R_i$, weighted by its expected deaths, and three quarters of those fall at ages 65 and over. A ratio above 1 at some ages can offset one below 1 at others. Nor does an interval that contains 1 show that the true ratio is 1. It shows only that random variation cannot tell it from 1, at a precision of about 1.5% either way.
The CMF averages the same ratios with different weights. Its weights are the deaths the 2000 standard population would have at the national rates ($w_i R_i$ in place of $n_i R_i$), and it puts Maine 1.1% above the nation.
A confidence interval for the SMR
Where did the exact intervals in the table come from? The idea is to treat the standard rates as fixed, which is reasonable when the standard population is much larger than the study population. Then $E$ is a constant and all the random variation is in $O$, a Poisson count. An interval for its mean, divided by $E$, is an interval for the SMR. Breslow and Day (1987, §2.3) use the exact Poisson limits. With the gamma quantiles of the previous article,
$$\text{SMR}_L = \frac{G^{-1}_{O}(0.025)}{E}, \qquad \text{SMR}_U = \frac{G^{-1}_{O+1}(0.975)}{E}.$$
A simpler approximation uses $\operatorname{SE}(\log \text{SMR}) = 1/\sqrt{O}$ (Breslow and Day, 1987, equation 2.9), giving $\text{SMR} \times e^{\pm 1.96/\sqrt{O}}$. For the states, with thousands of deaths, the two agree to three decimals; for Alaska the log-scale interval is (1.0344, 1.0871). Neither interval for Maine excludes 1. Alaska's, Florida's and Utah's do.
For small counts they diverge. Suppose an occupational cohort records 12 deaths from some cause where the standard rates predict 7.5. The SMR is 1.60. The exact interval is (0.827, 2.795) and the log-scale interval (0.909, 2.817), with the lower limit too high. Breslow and Day (1987, Example 2.8) find the same weakness in the log-scale lower limit. The exact one-sided probability of 12 or more deaths when 7.5 are expected is 0.0792. Doubling it, the two-sided exact test of $\text{SMR} = 1$ gives $p = 0.158$, consistent with the exact interval's inclusion of 1.
Comparing SMRs with each other
An SMR compares one population with the standard. Is it safe to compare two populations' SMRs with each other? Less so. Each SMR weights the age-specific ratios by its own population's expected deaths. An older population's SMR is dominated by the ratios at old ages, and a younger one's by those at younger ages.
Breslow and Day (1987, §2.3, Table 2.13) give an example, after Kilpatrick (1963), in which one cohort has the higher death rate at every age and yet the lower SMR. They quote Kilpatrick: "the ratio of two CMFs is a CMF but the ratio of two SMRs is not an SMR". Ratios of directly standardized rates are always weighted averages of the age-specific rate ratios, and so never show this reversal.
Back to the four states. For Maine and Alaska, the ratio of the SMRs is $1.0004/1.0604 = 0.943$, and the ratio of the published adjusted rates is $889.4/931.0 = 0.955$. For Florida and Utah, the two ratios are 1.033 and 1.021. Both measures agree that Alaska's mortality exceeded Maine's once age is taken into account, the reverse of the crude rates' ratio of 1.485. Both put Florida slightly above Utah.
When age-specific rate ratios are roughly constant across ages, SMRs can be compared. The close agreement of the SMRs and CMFs suggests the ratios are roughly constant here, though the states' age-specific deaths are not in this article's data. When the ratios are not constant, a directly standardized comparison, or the age-specific rates themselves, is the safer basis.
Which method, when
The two methods answer slightly different questions and have complementary weaknesses. Breslow and Day (1987, §2.3(f)) describe the choice as a trade-off between bias and variance.
- Direct standardization puts every population on the same weights. So its adjusted rates can be compared with each other and across time, and a population with lower rates at every age always has the lower adjusted rate. It needs age-specific deaths for each population, and it becomes unstable when some age groups have few deaths. It is the method of official vital statistics: NCHS's state, national and trend comparisons, and WHO's international ones.
- Indirect standardization needs only total deaths and the population by age. Its SMR is the minimum-variance estimate of a common rate ratio, so it is stable for small populations. It compares each population with the standard, not with each other, and comparisons of SMRs can mislead when rate ratios vary with age. It is the method of occupational cohorts, small areas and any study population compared with general-population rates.
- Neither replaces the age-specific rates when those rates move in different directions, as from 2010 to 2019. A single summary of a divergent pattern depends on its weights. Reporting the specific rates alongside the summary shows the reader what it averages.
Age is the usual adjustment variable because it is the strongest influence on most rates. But the arithmetic applies to any variable that divides the populations into strata: sex, race and Hispanic origin, or age and sex together. Adjusting for several variables at once leads to regression models for rates, which Breslow and Day (1987) prefer for analytical work. The library's Logistic Regression treats the corresponding model for proportions.
Summary
| Quantity | Formula |
|---|---|
| Directly standardized rate | $\text{DSR} = \sum_i w_i r_i$, $w_i = P_{si}/P_s$ |
| Variance of the DSR | $\sum_i w_i^2 d_i/n_i^2$ |
| Gamma interval (Fay and Feuer) | $L = \frac{v}{x} G^{-1}_{x^2/v}(0.025)$; $U = \frac{v + k^2}{x + k} G^{-1}_{(x+k)^2/(v+k^2)}(0.975)$, $k = \max_i w_i/n_i$ |
| Comparative mortality figure | $\text{DSR}/\text{DSR}_{\text{standard}}$ |
| Expected deaths | $E = \sum_i n_i R_i$ |
| Standardized mortality ratio | $\text{SMR} = O/E$; exact limits $G^{-1}_{O}(0.025)/E$ and $G^{-1}_{O+1}(0.975)/E$ |
| Indirectly standardized rate | $\text{ISR} = \text{SMR} \times R_s$ |
Corrections to the course notes
The unit's notes warn that an inappropriate standard population can bias the adjusted rate. Stated more precisely, no standard is correct or incorrect. An adjusted rate is a relative index whose level always depends on the standard. Comparisons between adjusted rates depend on it only when the age-specific rates do not move in proportion, as the 2010 to 2019 example shows.
The notes' observation that indirect standardization suits small populations, and that its adjusted rates are not directly comparable between populations, agrees with Breslow and Day (1987). The notes' formulas agree with this article's. These include the variance of the directly standardized rate and the variance $O/E^2$ of the SMR, which is $\text{SMR}^2/O$, the square of the standard error above. The course's formula sheets do not cover this unit.
What we have learned
A directly standardized rate is the crude rate a standard population would have with the study population's age-specific rates. Because every population gets the same weights, one that is lower at every age is lower when adjusted, under any standard. The crude rate gives no such guarantee, as 2007 and 2013 showed. The standard still sets the level of the adjusted rate, and when age-specific rates move in different directions it can set the direction of a trend. Its variance is a weighted sum of Poisson variances, and for small populations the gamma interval of Fay and Feuer keeps its coverage where the normal interval does not. Indirect standardization compares observed with expected deaths. Its SMR is stable for small populations but compares each one only with the standard. By both methods, Alaska's mortality in 2021 was higher than Maine's once age is taken into account.
Both methods rest on the same move: averaging rates over groups with a chosen set of weights. In the language of probability, that move is the law of total probability. The next unit, Probability, begins with The Rules of Probability. Its second article, Conditional Probability and the Law of Total Probability, returns to crude and standardized rates in those terms.
In Python
import numpy as np
from scipy import stats
w = np.array([3794901, 15191619, 39976619, 38076743, 37233437, 44659185,
37030152, 23961506, 18135514, 12314793, 4259173]) / 274633642
def dsr(d, n, w, level=0.95):
"""Directly standardized rate with normal and Fay-Feuer gamma intervals."""
d, n = np.asarray(d, float), np.asarray(n, float)
x, v = np.sum(w * d / n), np.sum(w**2 * d / n**2)
k, a = np.max(w / n), (1 - level) / 2
z = stats.norm.ppf(1 - a)
lo = v / x * stats.gamma.ppf(a, x**2 / v)
hi = (v + k**2) / (x + k) * stats.gamma.ppf(1 - a, (x + k)**2 / (v + k**2))
return x, (x - z * v**0.5, x + z * v**0.5), (lo, hi)
def smr(observed, expected, level=0.95):
a = (1 - level) / 2
lo = stats.gamma.ppf(a, observed) / expected
hi = stats.gamma.ppf(1 - a, observed + 1) / expected
return observed / expected, (lo, hi)
print(smr(6208, 5854.3)) # Alaska: 1.0604, (1.0342, 1.0871)
print(smr(12, 7.5)) # 1.6, (0.827, 2.795)
Every number in this article is computed by tools/experiments/applied-statistics/direct-and-indirect-standardization.py in the site's repository. The script queries CDC WONDER and the Census estimates, checks each adjusted and crude rate against the NCHS publications, and runs the simulation with fixed seeds.
References
- Ahmad, O. B., Boschi-Pinto, C., Lopez, A. D., Murray, C. J. L., Lozano, R., and Inoue, M. (2001). Age Standardization of Rates: A New WHO Standard. GPE Discussion Paper Series No. 31. World Health Organization. record · 2 pages
- Anderson, R. N., and Rosenberg, H. M. (1998). Age standardization of death rates: Implementation of the year 2000 standard. National Vital Statistics Reports, 47(3). National Center for Health Statistics. record · 3 pages
- Breslow, N. E., and Day, N. E. (1987). Statistical Methods in Cancer Research, Volume II: The Design and Analysis of Cohort Studies, Chapter 2 (Rates and rate standardization). IARC Scientific Publications No. 82. International Agency for Research on Cancer. record · 3 pages
- Centers for Disease Control and Prevention, National Center for Health Statistics. Underlying Cause of Death 1999–2020, CDC WONDER online database, from the Multiple Cause of Death Files compiled through the Vital Statistics Cooperative Program. https://wonder.cdc.gov/ucd-icd10.html (queried November 2026). record · 2 pages
- Chiang, C. L. (1961). Standard error of the age-adjusted death rate. Vital Statistics: Special Reports, 47(9), 273–285. National Office of Vital Statistics. Cited as Breslow and Day (1987) cite it. record
- Fay, M. P., and Feuer, E. J. (1997). Confidence intervals for directly standardized rates: A method based on the gamma distribution. Statistics in Medicine, 16(7), 791–801. record · 2 pages
- Kilpatrick, S. J. (1963). Mortality comparisons in socio-economic groups. Applied Statistics, 12(2), 65–86. Cited as Breslow and Day (1987) cite it. record
- Klein, R. J., and Schoenborn, C. A. (2001). Age Adjustment Using the 2000 Projected U.S. Population. Healthy People Statistical Notes No. 20. National Center for Health Statistics. record · 2 pages
- Murphy, S. L., Kochanek, K. D., Xu, J., and Arias, E. (2024). Deaths: Final data for 2021. National Vital Statistics Reports, 73(8). National Center for Health Statistics. Tables 2, 6, 17 and X and the Technical Notes. record · 4 pages
- Pagano, M., Gauvreau, K., and Mattie, H. (2022). Principles of Biostatistics (3rd ed.). CRC Press. The course text. record · 35 pages
- U.S. Census Bureau, Population Division (2022). Annual State Resident Population Estimates for 6 Race Groups (5 Race Alone Groups and Two or More Races) by Age, Sex, and Hispanic Origin: April 1, 2020 to July 1, 2021 (Vintage 2021; released June 2022), file sc-est2021-alldata6. https://www2.census.gov/programs-surveys/popest/datasets/2020-2021/state/asrh/ record · 3 pages