Stratified Measurement in Practice

An earlier lesson in this module established the consensus that quality reporting should be stratified rather than adjusted, so gaps stay visible. This lesson covers what stratification requires in practice, because the obstacles are where most efforts stop.

The data problem comes first

You cannot stratify by a characteristic you have not recorded. Race, ethnicity, and language data is frequently missing, inconsistently categorized, or entered by staff guessing rather than by patients self-reporting.

Three failures recur:

• Missingness that is not random. The patients whose demographic data is absent are disproportionately those who move between organizations, which is correlated with the disadvantage being measured.
• Categories that do not fit. Standard race and ethnicity categories collapse groups with very different outcomes, and a single “Asian” or “Hispanic” category can hide a large disparity inside it.
• Collected but not usable. Demographic data entered in a free-text field or an unmapped local code is unavailable to any analytic system. Being recorded somewhere and being usable as structured data are different things.

Improving this is unglamorous work: asking patients to self-report, explaining why, training registration staff, and mapping local codes to standard values. It precedes everything else.

Small cells

Stratifying divides an already limited denominator. A practice with 400 attributed patients stratifying a measure across five groups may have 30 patients in some cells, and at that size a difference between groups reflects chance at least as much as performance.

Two responses are standard and both involve tradeoffs:

Approach	Gains	Costs
Suppress cells below a minimum size	Avoids reporting noise, protects privacy	The smallest groups, often the most marginalized, disappear from the report
Aggregate across time or across sites	Usable denominators	Slower feedback, less local specificity

The suppression tradeoff deserves explicit attention. A rule that hides any group under 30 patients systematically removes the smallest populations from view, which means the measurement designed to surface disparities can silence exactly the groups facing the largest ones.

What to stratify by

Race, ethnicity, and language are the standard starting point. Dual eligibility is often the most available and most predictive marker of social risk, since it is recorded reliably in claims. Others worth considering are primary language, disability status, and geographic markers of area-level deprivation.

The practical guidance is to start with what is reliably recorded rather than waiting for the complete picture. Dual eligibility alone will reveal real gaps in most populations, and it requires no new collection.

Equity provisions can be reversed

Medicare Advantage offers a cautionary example. A health equity index reward was finalized in 2023 to begin with the 2027 Star Ratings. In the Contract Year 2027 final rule issued April 2, 2026, “for the 2027 Star Ratings, CMS is not implementing the Excellent Health Outcomes for All reward (previously called the Health Equity Index reward) that was developed to incentivize improved performance for a subset of enrollees and will continue the historical reward factor that encourages consistently high performance for all enrollees across all quality measures.”

Worth remembering: an organization that stratified its measures only because a payment program required it now has no external requirement in that program. An organization that stratified because it wanted to know whether it delivers equivalent care to everyone it serves still has the same reason it had before. That difference determines which measurement survives a change in federal policy, and it is worth deciding which kind of organization you are before the requirement changes rather than after.

Key takeaways

• Demographic data quality is the binding constraint, and non-random missingness correlates with the disadvantage being measured.
• Stratification shrinks denominators, and cell suppression can hide the smallest and most marginalized groups.
• Dual eligibility is often the most reliably recorded marker of social risk and a practical place to start.
• The Medicare Advantage health equity reward was finalized in 2023 and not implemented for 2027, so external requirements can disappear.

Sources

• CMS, Contract Year 2027 Medicare Advantage and Part D Final Rule (opens in a new tab)
• CMS, Hospital Readmissions Reduction Program (opens in a new tab)