The Demographic Granularity Gap in Algorithmic and Clinical Evaluation

30 concepts27 papers3 proposals

Imagine auditing a medical diagnostic model for fairness. You evaluate its performance across broad racial categories—White, Black, Asian, and Hispanic—and find that the error rates are roughly equal. You conclude the model is fair.

This conclusion, however, rests on a dangerous assumption: that these broad categories represent homogeneous populations. In reality, they are administrative conveniences masking massive internal diversity. When a model or clinical risk score is evaluated only on these "coarse" labels, severe performance disparities affecting specific sub-populations can remain completely hidden. This phenomenon is known as subgroup masking, and it creates a profound demographic granularity gap across machine learning, public health, and clinical medicine.

The Intuition of Subgroup Masking

Why does demographic aggregation fail? When we group individuals into massive continental or bureaucratic categories, we compute an average performance metric. If a specific subgroup within that category faces a severe disadvantage, its distinct signal is washed out by the larger group's average.

Consider how this plays out in public health. The category "Asian/Pacific Islander" aggregates dozens of distinct ethnic groups with entirely different genetic backgrounds, socioeconomic realities, and environmental exposures. As Robert Hsu et al. demonstrated in Characterization of mortality and high-risk characteristics of thyroid cancer in Filipinos using the California Cancer Registry, Filipino patients exhibit significantly higher incidence and mortality rates for thyroid cancer compared to other Asian groups, often presenting with advanced Stage IV disease. By lumping all Asian ethnicities together, traditional public health models mask this critical disparity, preventing targeted clinical interventions.

Similarly, Huafu Li discovered a striking "prognostic reversal" in gastric cancer survival in Differences in the prognosis of gastric cancer patients of different sexes and races and the molecular mechanisms involved: while White females generally outlive White males, the opposite is true in Asian populations, where males show superior survival. Aggregating these groups into coarse categories completely hides these intersectional, molecularly driven variations.

The Mathematical Reality: Intra-Group vs. Inter-Group Variance

In machine learning, fairness is frequently evaluated by minimizing performance discrepancies between broad groups. However, the authors of Coarse race data conceals disparities in clinical risk score performance (Rajiv Movva et al.) exposed a fundamental flaw in this approach: the variance in predictive performance within a single coarse demographic category can actually exceed the variance observed between different categories.

When intra-group variance is high, an aggregate metric becomes a deceptive proxy. A model can perform exceptionally well on one dominant subgroup while failing catastrophically on a smaller minority subgroup within the same category.

Coarse Fairness Audit (Deceptive):
[ Broad Category: Asian ] ---> Average Error: 5% (Looks Fair)

Granular Subgroup Disaggregation (Actual):
├── Subgroup A (e.g., East Asian)   ---> Error: 2%
└── Subgroup B (e.g., South Asian)  ---> Error: 18%  <-- Hidden Algorithmic Failure

This mathematical masking is vividly illustrated in computer vision. Ioannis Sarridis et al. analyzed face verification systems in Towards Fair Face Verification: An In-depth Analysis of Demographic Biases and found that while overall accuracy metrics appeared high, specific intersectional subgroups—such as African females over the age of 60—suffered from disproportionately high mistreatment rates. Treating diverse continental populations as a single group ignores these internal variations, rendering standard fairness evaluations blind to localized algorithmic failures.

The Pervasiveness of the Granularity Gap

The demographic granularity gap is a systemic issue spanning multiple domains:

  • Medical Imaging Foundation Models: Muhammad Shujaat Mirza et al. showed in How Fair are Medical Imaging Foundation Models? that while pre-training on massive medical datasets boosts diagnostic performance, it introduces hidden fairness trade-offs across demographic subgroups that are completely obscured by broad racial categories.
  • Natural Language Processing: In text-based prediction, Lyle Ungar et al. demonstrated in User-Level Race and Ethnicity Predictors from Twitter Text that broad demographic categories fail to capture the rich internal linguistic and cultural variations of specific populations, limiting a model's ability to accurately represent distinct communities.
  • Epidemiology and Geography: In The spatial epidemiology of sickle-cell anaemia in India, Carinna Hockham et al. built a national map of allele frequencies by accounting for complex social stratification (such as "scheduled" vs. "non-scheduled" populations). They demonstrated that reducing this diversity into binary categories completely obscures unique epidemiological profiles.

Methodologies for Disaggregation

If coarse categories are fundamentally flawed, how do we fix them? The primary challenge is that real-world administrative and electronic health record (EHR) datasets rarely collect granular subgroup data. To overcome this, researchers are developing sophisticated algorithmic and statistical techniques to infer and disaggregate these hidden populations.

1. Surname and Linguistic Filtering

When standard EHR systems fail to capture granular data, secondary identification methods become essential. Jeannette M. Beasley demonstrated in Diabetes and hypertension among South Asians in New York and Atlanta leveraging hospital electronic health records that combining surname analysis with language filtering can correct systematic misclassifications. This approach revealed that cardiovascular risk is highly non-uniform across different regional diasporas, showing a much higher rate of co-occurring diabetes and hypertension in Atlanta compared to New York City.

2. Open-Source Name Mapping

To enable large-scale disaggregation, researchers are building specialized open-source resources. For instance, Qiwei Lin et al. developed a dataset of Wikidata names specifically designed for disparity estimation in Enabling disaggregation of Asian American subgroups: a dataset of Wikidata names for disparity estimation. This allows researchers to decompose broad categories into specific subgroups and uncover hidden variations in chronic disease prevalence that are masked when diverse populations are treated as a single monolithic group.

3. Bayesian Instrumental Variables

When direct demographic labels are missing entirely, naive estimation methods fail because they assume demographic attributes are independent of outcomes once controlling for basic covariates. To solve this, Cory McCartan et al. introduced BIRDiE in Estimating Racial Disparities When Race is Not Observed. This scalable Bayesian framework treats surnames as an instrumental variable to produce unbiased estimates of racial disparities in massive administrative datasets, preventing the bias that occurs when broad categories obscure subgroup differences.

Conclusion

Whether building clinical risk scores, auditing foundation models, or mapping public health crises, relying on coarse demographic categories introduces a systematic blind spot. If the variance within a category is ignored, fairness metrics become a form of theater—protecting the average while neglecting the vulnerable. True equity requires a shift toward granular ancestral mapping and algorithmic disaggregation, ensuring that the unique vulnerabilities of distinct sub-populations are no longer averaged out of existence.

Key Foundational Work

Proceedings of the International Workshop on Spatio-Temporal Database Management (1999) — contributed 1 concept elements to this direction.
Inter-Ethnic/Racial Facial Variations: A Systematic Review and Bayesian Meta-Analysis of Photogrammetric Studies (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_subgroup_masking_e3f7e8bb,
  title = {subgroup masking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subgroup-masking/rdr_e3f7e8bbc9c14420b3d1fa88f8105cc4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_subgroup_masking_e3f7e8bb_v1,
  title = {subgroup masking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subgroup-masking/rdr_e3f7e8bbc9c14420b3d1fa88f8105cc4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake