Standardized rates: compare groups under one declared population mix

Last updated: 7 Oct 20265 min read

tutorial

Direct standardization applies stratum-specific rates to one shared target mix so composition does not decide the comparison.

Define the common population

Two service centers handle different shares of routine and complex claims. Their crude escalation rates mix process performance with case composition, so a fair descriptive comparison needs a declared target distribution. Choose strata based on recorded severity before outcome review, define eligibility and outcome timing, and use the same target weights for both centers. The standardized rate is a weighted mean of each center’s stratum-specific rates. It is a comparison under the chosen mix, not the observed rate that either center actually experienced. The population lesson identifies who the weights represent.

Check support before multiplying

Every target stratum with positive weight needs an estimable rate in each center. A missing complex-claim group cannot be filled with zero or quietly dropped while retaining the original weights. Sparse cells have large uncertainty even when the final weighted average looks smooth. Show counts and intervals for each rate, and choose a coarser grouping or an explicit model if support is inadequate. The program below only calculates a weighted point estimate from supplied rates. Rare-proportion intervals help expose fragile cells.

State what standardization does not solve

Holding observed severity mix constant removes one compositional difference in the descriptive calculation. It does not remove unmeasured complexity, different coding rules, or selection into each center. If severity is affected by the intervention, standardizing on it may change the causal question or introduce bias. Show crude, stratum-specific, and standardized rates side by side rather than presenting only the adjusted figure. The reversal lesson demonstrates how crude ordering can contradict every within-stratum ordering.

Carry uncertainty and choice of weights

Specify whether the target mix comes from an agreed reference period, a pooled current population, or a planned future workload. A different target mix can change the summary, so freeze it before ranking centers and disclose sensitivity to plausible alternatives. Use a variance method respecting how stratum rates and target weights were estimated. The project catches an absent complex stratum and a changed severity coding rule before releasing a benchmark.

Implementation

def direct_standardized_rate(stratum_rates, target_weights):
    if set(stratum_rates) != set(target_weights):
        raise ValueError("rates required for every target stratum")
    if any(not 0 <= rate <= 1 for rate in stratum_rates.values()):
        raise ValueError("rates must lie between zero and one")
    if any(weight < 0 for weight in target_weights.values()) or        abs(sum(target_weights.values()) - 1) > 1e-9:
        raise ValueError("target weights must be nonnegative and sum to one")
    return sum(stratum_rates[group] * target_weights[group]
               for group in target_weights)

rate = direct_standardized_rate({"routine": 0.04, "complex": 0.12},
                                {"routine": 0.70, "complex": 0.30})
assert round(rate, 6) == 0.064

Performance and operating cost

The weighted sum is O(s) time and O(1) extra space for s strata. Estimating each stratum rate and its uncertainty costs more, especially with sparse cells. A crude comparison is cheaper but can answer a different question when case mix differs.

Common Mistakes

Changing target weights after seeing which center wins.
Setting an absent stratum rate to zero.
Calling a standardized rate the observed workload rate.
Treating severity adjustment as proof of a causal service-center effect.