MerQuR+Statistics Glossary

Statistics Glossary → Age Standardization and SMR

Age Standardization and SMRanalysis

A method that makes disease or death rates of populations with different age (and optionally sex) structures comparable by removing the effect of the age difference.

Concept

Death and many diseases rise steeply with age. A region with an older population can therefore have a higher crude death rate than a young region even if its health conditions are better. Age standardization answers the question 'how would the rates compare if the two populations had the same age structure?'

In the direct method, each population's age-specific rates are combined using the age distribution of a common standard population as weights: standardized rate = Σ (share of the standard population in the age group × rate in that age group). All populations are thus summarised over the same age structure and compared with a rate ratio or rate difference. In the indirect method, the standard population's age-specific rates are applied to the study population to obtain the expected number of events, which is compared with the observed number: SMR = observed / expected. An SMR of 1.25 means the study population had 25% more events than expected given its age structure.

The direct method needs enough events in every age group; in small populations age-specific rates are unstable, so the indirect method (SMR) is preferred. A standardized rate has no meaning on its own: its value depends on the standard population chosen, and only rates computed with the same standard should be compared. SMRs of different study populations are each computed with their own age structure and are generally not compared directly.

When to use it

  • To compare death or disease rates fairly across regions, periods or groups with different age structures.
  • If the study population is small or its age-specific rates are unreliable, use indirect standardization (SMR) instead of the direct method, for example when comparing an occupational group's mortality with the general population.
  • If other factors besides age (income, smoking) are to be controlled as well, multivariable models such as Poisson regression are a more flexible alternative.

An everyday example

Suppose one seaside town is popular with retirees and another town is home to a university. The seaside town has 1200 deaths per 100,000 people a year, the university town 600; at first glance the seaside town looks twice as unhealthy. But 35% of the seaside town's population is over 65, against only 8% in the university town. When each town's age-specific death rates are weighted by the age distribution of the same standard population, the standardized rates might come out at about 750 and 790: the whole crude difference came from age structure, and once age is accounted for the university town is in fact slightly worse off.

Parameters in MerQur

Stratum (age group) and Second stratum (optional)

Concept: The subgroups (usually age groups) within which rates are computed separately and then combined with standard weights. When a second stratum (e.g. sex) is added, standardization is done over age-by-sex combinations and the structural differences in both variables are removed.

Everyday example: In the towns example the strata are age groups such as 0-14, 15-44, 45-64 and 65 and over. If the seaside town has many retired women, sex is added as a second stratum so that the difference in sex distribution does not distort the comparison either.

What it does: Defines the strata in which standardization is done (e.g. age group); the optional second stratum (e.g. sex) is combined with the first. Strata are reported in order of first appearance in the data.

Effect on the result: Determines the cells in which weights and expected events are computed.

On the validation data: Sex x age strata in the Alaska/Florida data (10 strata); age strata in the Florida cancer data; 10-year age groups in Doll-Hill.

→ Direct and indirect standardization (SMR)

Group (study population)

Concept: The variable defining the study populations to be compared (regions, periods, exposed and unexposed groups). A separate standardized rate or SMR is computed for each group.

Everyday example: In the towns example the group variable is the town's name: the seaside town and the university town. If a third town is added, it is summarised with the same standard and can be compared with the others.

What it does: Determines the study populations to be compared (region, exposure); if empty, all the data are a single population.

Effect on the result: Separate standardized rates/SMR and between-group comparisons are produced for each group.

On the validation data: Alaska and Florida; smoking and non-smoking doctors.

Number of events and Person-time / population

Concept: Sets the numerator of the rate (number of events: cases, deaths) and its denominator (person-time or population at risk). Person-years are the total time each person spent under observation; when follow-up times differ, this is a more accurate denominator than a head count.

Everyday example: If the over-65 group in the seaside town has 420 deaths and 14,000 person-years in a year, its death rate is 3000 per 100,000 person-years. Retirees who spend only the summer in the town count in the person-years only for the time they stay.

What it does: Selects the number of events (cases, deaths) in the stratum x group row and the denominator, person-years or population.

Effect on the result: Determines the stratum rates (events / person-time).

On the validation data: Alaska events 2924, person-time 626 932; Florida events 162 795, person-time 15 658 229. With a multiplier of 1 000 the crude rates are 4.663982697 and 10.39676965; with the default 100 000 they are shown 100 times larger.

→ Person-time (person-years) denominator

Method (Direct standardization, Indirect standardization (SMR))

Concept: The direct method applies the study population's age-specific rates to the age structure of the standard population; the indirect method applies the standard population's age-specific rates to the study population's age structure and compares observed with expected events (SMR). The indirect method is preferred in small populations whose age-specific rates are unreliable.

Everyday example: The towns have thousands of residents, so the direct method suits them. A village of only 300 people, though, has just a few deaths per age group; here national rates are applied to the village's age structure, and if 15 deaths are observed against 12 expected, SMR = 1.25.

What it does: In the direct method the study stratum rates are combined with standard population weights. In the indirect method the standard rates are applied to the study population and SMR = observed / expected is computed. Only the rate statistic is supported.

Effect on the result: Determines the output set (standardized rate and rate ratio, or expected events, SMR and the SMR = 1 test).

On the validation data: Direct (per 1 000 person-years): Alaska 8.069284378 (SE 0.1643185522), Florida 7.734186126; rate ratio (Alaska/Florida) 1.043326893, rate difference 0.3350982524. Indirect: Florida C43 SMR 1.01753964, C16 SMR 0.8612653925.

→ Direct and indirect standardization (SMR)

Standard (Standard population column, Standard population entered manually, Total population of the groups), Standard population, Standard rate, Standard events, Standard values

Concept: The reference serving as the common yardstick: in the direct method the standard population that supplies the age weights, in the indirect method the standard rates used to compute expected events. The values of standardized rates depend on this choice; only rates computed with the same standard can be compared.

Everyday example: The towns could be standardized to the national population or to an international standard population. Choosing an older standard raises both towns' standardized rates; but as long as the same standard is used, which town is worse off usually stays the same.

What it does: Determines the source of the standard population (direct) or the standard rates (indirect): a column, manually entered values (in order or "label=value") or the total of the groups. The standard population is a standard such as US 2000, Europe 2013 or WHO.

Effect on the result: Changes the weights, expected events and hence all standardized measures.

On the validation data: US 2000 person-years (column) in Alaska/Florida; SEER 2000 US standard million in Doll-Hill: 456.610386 in non-smokers, 581.5283978 in smokers (per 100 000 person-years). Manually entered values and the column source give the same result.

→ Direct and indirect standardization (SMR)

Confidence interval method (Gamma (Fay-Feuer), Gamma (Tiwari), Normal, Lognormal)

Concept: Sets how the confidence interval for the standardized rate or SMR is computed. Because event counts follow a Poisson distribution, intervals based on the gamma or exact Poisson distribution are more reliable than the normal approximation when events are few; as the number of events grows the methods converge.

Everyday example: With thousands of deaths in the seaside town, the intervals are nearly identical whichever method is chosen. The village of 300 has only 15 deaths; there the normal approximation forces an artificially symmetric interval and can shift its limits, and the exact Poisson interval is more accurate.

What it does: In the direct method a gamma (Fay-Feuer, default), gamma Tiwari, normal or lognormal interval is selected. In the indirect method the same field controls the exact Poisson, Byar, normal and lognormal options.

Effect on the result: Changes only the confidence interval limits and the type of the SMR = 1 test; the point estimate does not change.

On the validation data: Alaska (per 1 000 person-years, 8.069284378): normal 7.747225934 to 8.391342822; gamma (default) 7.750406818 to 8.399483247; Tiwari upper limit 8.396736361; lognormal 7.753568218 to 8.397856127. Per 100 000 person-years: normal [774.72, 839.13], gamma (default) [775.04, 839.95], Tiwari [775.04, 839.67], lognormal [775.36, 839.79].

→ Direct and indirect standardization (SMR)

Rate multiplier (1, 100, 1,000, 10,000, 100,000, 1,000,000)

Concept: The scale multiplier setting per how many people or person-years rates are shown. It only affects readability; for rare events a large base such as 100,000 makes the numbers easy to grasp.

Everyday example: A death rate written as 0.012 is hard to read; written as 1200 per 100,000 people, the same information is much easier to take in.

What it does: Determines per how many person-time the rates are shown (default 100 000).

Effect on the result: Changes only the scale.

On the validation data: With the default 100 000 the Alaska standardized rate is 806.9284378; with 1 000 it is 8.069284378.

Reference group ((first group), Alaska, Florida)

Concept: The group taken as the denominator when groups are compared: the rate ratio divides the other groups' standardized rates by this group's, and the rate difference subtracts this group's rate. The choice changes the direction in which the result is expressed, not its meaning.

Everyday example: With the seaside town as reference, the university town's rate ratio is 790/750 ≈ 1.05, '5% higher'. With the university town as reference, the seaside town's rate ratio would be about 0.95, read as 'about 5% lower'.

What it does: Selects the denominator group in comparisons: other groups are divided by this group (rate ratio) and subtracted from this group (rate difference).

Effect on the result: Determines the sign/direction of the rate ratio and rate difference.

On the validation data: The default reference is the first group, Alaska: Florida / Alaska rate ratio 0.9585 (z = -2.067, p = .0387). With Florida selected as the reference, the Alaska / Florida rate ratio is 1.043326893, log rate ratio 0.04241454273 (SE 0.02051950695).

→ Rate ratio and rate difference

Confidence Level (%90, %95, %99, Other (custom))

Concept: Sets how 'sure' all the confidence intervals for standardized rates, rate ratios, rate differences and SMRs are. The point estimates do not change.

Everyday example: If the university town's rate ratio is 1.05 with a 95% confidence interval of 0.98 to 1.12, the difference is not reliable; at 99% confidence the interval widens further.

What it does: Determines the level of all confidence intervals.

Effect on the result: Changes only the interval limits.

On the validation data: Compared at 95%.

→ Confidence level and confidence interval

Save row values

Concept: Adds the rate and expected number of events computed for each stratum row to the dataset as new columns. It does not change the calculation; it makes the results easier to use in other analyses or charts.

Everyday example: If the town council wants to see the death rate and expected number of deaths for each age group in its own table, this option writes them into the dataset without copying by hand; which age group drives the difference can then be charted separately.

What it does: Adds the stratum rate (ASR_STRATUM_RATE) and the expected number of events (ASR_EXPECTED) to the dataset as new columns.

Effect on the result: Does not change the calculation.

Example on real data

Data: Alaska and Florida 2000 cancer deaths, SAS STDRATE Example 113.1

Variables: strata Age (age group) and Sex, group State, events Death, person-years PYear, standard population US_PYear (US 2000)

  • Question: How do the death rates of Alaska and Florida, which differ in age and sex structure, compare when standardized to the US population?
  • Crude rates: Alaska 466.40, Florida 1039.68 (per 100 000 person-years); Florida's older population inflates its crude rate.
  • Standardized rates: Alaska 806.93 (95% CI [775.04, 839.95]), Florida 773.42 ([769.59, 777.26]); Gamma (Fay-Feuer) interval.
  • Comparison: Rate ratio (Florida / Alaska) = 0.959, 95% CI [0.921, 0.998], z = -2.07, p = .039; rate difference = -33.51 (SE 16.55), p = .043.
  • Interpretation: The more than twofold difference in crude rates reverses once the age-sex structure is equalised: standardized, Florida's death rate is about 4% lower than Alaska's.
  • APA: Direct age-standardized rates: Alaska 806.93, 95% CI [775.04, 839.95]; Florida 773.42, [769.59, 777.26]; rate ratio = 0.96, [0.92, 1.00], z = -2.07, p = .039.

Related concepts

Confidence level and confidence intervalDirect and indirect standardization (SMR)Rate ratio and rate differencePerson-time (person-years) denominatorStratified analysis (stratum variable)Exact test and exact p-valueReference category (comparison level)

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.