Direct Standardization Of Rates

Why Standardize Rates?

Rates for two groups can be difficult to compare when the groups have different distributions across strata. For example, age-specific rates may be higher in older people, and one group may contain a larger proportion of older people than another. Its crude rate can therefore be higher even if its rate is lower within every age stratum. Direct standardization addresses this compositional difference by applying each group’s stratum-specific rates to the same chosen standard population.

The standard population supplies the weights for the calculation. The group-specific rates come from the data or a fitted rate model. The resulting standardized rate is a summary for comparison, not the observed rate in either group and not, by itself, a causal effect.

Definition and Construction

Let \(g\) index the groups being compared and let \(s=1,\ldots,K\) index mutually exclusive strata. Strata might be age categories or other categories defined consistently across groups. Let \(\lambda_{gs}\) denote the rate for group \(g\) in stratum \(s\), and let \(\hat{\lambda}_{gs}\) be its estimated or model-fitted value. Let \(T_s^*\) be the standard population’s exposure in stratum \(s\), measured in units compatible with the rate denominator. Define the standard-population weights by \(w_s=T_s^*/\sum_{j=1}^K T_j^*\).

Definition: The directly standardized rate for group \(g\), using standard weights \(w_1,\ldots,w_K\), is

\hat{\lambda}^{\,\mathrm{std}}_g = \sum_{s=1}^K w_s\hat{\lambda}_{gs}, \qquad w_s\geq 0,\quad \sum_{s=1}^K w_s=1.

The weights are the proportions of standard-population exposure in the strata. If rates are reported per 1,000 units of exposure, the standardized rate is reported on that same scale.

The standard weights must be fixed and common to all groups in a comparison. The stratum definitions must also be common, and each group must have an estimable rate for every stratum with positive weight. A fitted rate may come from a Poisson model with an exposure offset: if \(\log E_{gs}\) is the offset for exposure \(E_{gs}\), then the fitted mean count divided by \(E_{gs}\) gives the fitted rate for that covariate pattern. The offset accounts for exposure; standardization is the separate operation of averaging the fitted rates with the chosen weights.

Worked Example: Standardizing Rates From Counts and Exposure

Suppose two groups have counts and exposure in two age strata. The standard population has 80% of its exposure in the younger stratum and 20% in the older stratum. Calculate the directly standardized rates per 1,000 exposure units.

Group A. In the younger stratum, the rate is \(12/6{,}000=0.002\) per exposure unit, or 2 per 1,000. In the older stratum, it is \(30/3{,}000=0.010\), or 10 per 1,000. Therefore,

\hat{\lambda}^{\,\mathrm{std}}_A =0.80(2)+0.20(10) =1.6+2.0 =3.6 \quad\text{per 1,000 exposure units.}

Group B. The younger-stratum rate is \(18/6{,}000=3\) per 1,000, and the older-stratum rate is \(24/3{,}000=8\) per 1,000. Applying the same weights gives

\hat{\lambda}^{\,\mathrm{std}}_B =0.80(3)+0.20(8) =2.4+1.6 =4.0 \quad\text{per 1,000 exposure units.}

Under this standard population, Group B’s standardized rate exceeds Group A’s by \(4.0-3.6=0.4\) per 1,000. The calculation uses the rates within each stratum, not the groups’ own proportions of exposure in those strata.

Why the Weighted-Sum Formula Works

The formula can be derived by imagining that each group’s stratum-specific rates applied to the same standard-population exposures. This is a mathematical construction: it shows what the formula summarizes, without asserting that the standard population actually experienced those rates.

Result: If the rate \(\hat{\lambda}_{gs}\) is applied to standard exposure \(T_s^*\), the corresponding fitted count is \(T_s^*\hat{\lambda}_{gs}\). The total fitted count divided by total standard exposure is exactly the directly standardized rate.

To verify the result, sum the fitted counts over strata and divide by the total standard exposure. Since \(T_s^*=w_sT^*\), where \(T^*=\sum_s T_s^*\), the calculation is

\frac{\sum_{s=1}^K T_s^*\hat{\lambda}_{gs}}{\sum_{s=1}^K T_s^*} = \frac{\sum_{s=1}^K (w_sT^*)\hat{\lambda}_{gs}}{T^*} = \sum_{s=1}^K w_s\hat{\lambda}_{gs} = \hat{\lambda}^{\,\mathrm{std}}_g.

This derivation also explains why the weights must correspond to the rate denominator. If the rate is per person-year, the standard exposure should be in person-years; if it is per person, use a compatible population denominator. The weighted average preserves the units of the stratum-specific rates.

Worked Example: From a Fitted Poisson Model to Standardized Rates

Consider younger and older strata and two groups, A and B. Suppose a Poisson model with an exposure offset gives these fitted rates per 1,000 exposure units: A has rates 2 and 10; B has rates 3 and 20. The standard population weights are \(w_{\mathrm{young}}=0.75\) and \(w_{\mathrm{old}}=0.25\). One way to represent the fitted values is a model with a younger-stratum baseline, an older-stratum effect, a group effect, and a group-by-old-stratum interaction:

\log(\hat{\lambda}_{gs}) = \beta_0 +\beta_{\mathrm{old}}I(s=\mathrm{old}) +\beta_B I(g=B) +\beta_{\mathrm{int}}I(g=B)I(s=\mathrm{old}).

For example, take \(\beta_0=\log(0.002)\), \(\beta_{\mathrm{old}}=\log(5)\), \(\beta_B=\log(1.5)\), and \(\beta_{\mathrm{int}}=\log(4/3)\), with rates expressed per exposure unit. For Group A, the younger rate is \(0.002\), or 2 per 1,000. Its older rate is \(0.002(5)=0.010\), or 10 per 1,000. For Group B, the younger rate is \(0.002(1.5)=0.003\), or 3 per 1,000. Its older rate is \(0.002(5)(1.5)(4/3)=0.020\), or 20 per 1,000.

The standardized rates are therefore

\hat{\lambda}^{\,\mathrm{std}}_A =0.75(2)+0.25(10)=1.5+2.5=4, \qquad \hat{\lambda}^{\,\mathrm{std}}_B =0.75(3)+0.25(20)=2.25+5=7.25,

both per 1,000 exposure units. The model supplies a fitted rate in each group-stratum combination; standardization then applies the same weights to those fitted rates. The interaction matters here because the group rate ratio differs by stratum.

Comparing Groups With Shared Weights

Using identical standard weights gives a direct comparison of the groups’ rates under one common stratum composition. The standardized difference is the weighted average of the stratum-specific differences. The standardized rate ratio is the ratio of the two weighted averages.

Result: For two groups A and B standardized to the same weights,

\hat{\lambda}^{\,\mathrm{std}}_B-\hat{\lambda}^{\,\mathrm{std}}_A = \sum_{s=1}^K w_s \bigl(\hat{\lambda}_{Bs}-\hat{\lambda}_{As}\bigr).

This follows by substituting the definition for each group and collecting terms:

\hat{\lambda}^{\,\mathrm{std}}_B-\hat{\lambda}^{\,\mathrm{std}}_A = \sum_{s=1}^K w_s\hat{\lambda}_{Bs} - \sum_{s=1}^K w_s\hat{\lambda}_{As} = \sum_{s=1}^K w_s \bigl(\hat{\lambda}_{Bs}-\hat{\lambda}_{As}\bigr).

Thus, the difference is not distorted by using one group’s stratum distribution for A and another’s for B. The standardized rate ratio, in contrast, is \(\hat{\lambda}^{\,\mathrm{std}}_B/\hat{\lambda}^{\,\mathrm{std}}_A\), when the denominator is positive; it is not generally the weighted average of the stratum-specific rate ratios.

Worked Example: A Crude Comparison Can Reverse After Standardization

Suppose Group A has rates 2 in the younger stratum and 10 in the older stratum. Group B has rates 3 and 8, respectively, all per 1,000 exposure units. In the standard population, the two strata each contribute half the exposure.

The standardized rates are

\hat{\lambda}^{\,\mathrm{std}}_A=0.5(2)+0.5(10)=6, \qquad \hat{\lambda}^{\,\mathrm{std}}_B=0.5(3)+0.5(8)=5.5.

Now suppose Group A’s own exposure is 90% younger and 10% older, while Group B’s is 10% younger and 90% older. Their crude rates, computed with those group-specific proportions, are

\hat{\lambda}_A^{\,\mathrm{crude}}=0.90(2)+0.10(10)=2.8, \qquad \hat{\lambda}_B^{\,\mathrm{crude}}=0.10(3)+0.90(8)=0.3+7.2=7.5.

The crude comparison is \(7.5>2.8\), whereas the standardized comparison is \(5.5<6\). These results are not contradictory: the crude rates use different stratum weights, while the standardized rates use the same weights. In this example, B’s older-stratum rate is lower than A’s, and the much larger older share in B’s own population strongly affects its crude rate.

Common Mistakes and Practical Checks

Using different weights for the two groups. That compares summaries based on different stratum compositions. Choose one standard population and apply its weights to every group in the comparison.
Averaging rate ratios instead of rates. Direct standardization first averages each group’s stratum-specific rates. The ratio is then formed from the two standardized rates; it is not generally the weighted average of stratum-specific ratios.
Confusing a standardized rate with an observed rate. The standardized rate is constructed using the standard population’s weights. It need not equal either group’s crude rate or describe its actual composition.
Ignoring units or unsupported strata. Check that rates and standard exposures use compatible denominators and that rates can be estimated for every stratum receiving positive weight.

The central calculation is simple, but its interpretation depends on stating the standard population and retaining the same weights across the comparison. Direct standardization is useful precisely because it holds the stratum composition fixed; it does not remove every possible difference between groups or establish why their rates differ.

Check Your Understanding

Use the definitions and calculations above to answer these questions.

A standard population has weights 0.6 and 0.4. A group’s fitted rates are 5 and 9 per 1,000. What is its directly standardized rate?
Why must a comparison of directly standardized rates use the same weights for both groups?
Explain how the weighted-sum formula follows from applying each stratum-specific rate to the standard exposure in that stratum.
Is the standardized rate ratio generally equal to the weighted average of the stratum-specific rate ratios? Explain.
Give one reason a standardized rate can differ substantially from the group’s crude rate.