Health Equity-Stratified Analysis in RWE

Health equity-stratified analysis

designs RWE so that differences across socially defined groups are measured and interpreted, not merely adjusted away. Where conventional practice treats race/ethnicity and socioeconomic position as nuisance covariates, equity analysis asks: does the treatment work equally well, is it equally accessible, and is the evidence itself equally representative — across these groups?

Why it matters for RWE

Three forces make this a first-class design concern rather than an optional sensitivity analysis. First, regulation: FDORA (2022) directed FDA to publish diversity action plans for trials, and external-control RWE inherits the expectation that its populations reflect intended treatment populations. Second, validity: differential outcome ascertainment (fewer labs in marginalized groups), differential censoring (insurance churn concentrated in low-income enrollees), and differential treatment allocation all create effect-measure modification and bias that marginal estimates conceal. Third, policy use: HTA bodies and payers making coverage decisions need to know whether evidence generalizes to the populations they cover.

Design elements

• Pre-specified strata: define equity-relevant subgroups (race/ethnicity, SES measures like area deprivation index or dual-eligibility, geography/rurality, language, insurance type) in the protocol with power justification — not post hoc fishing.
• Interaction testing with humility: test effect-measure modification on additive and multiplicative scales; additive-scale differences matter most for policy because they express absolute benefit gaps.
• Differential-bias diagnostics: evaluate missingness, censoring, and misclassification by subgroup; a validated algorithm overall can be badly calibrated within groups (e.g., pulse oximetry, eGFR race corrections).
• Representativeness accounting: compare study population composition against the target population and report who was excluded and why.

Measurement cautions

• Race is social, not biological: model it as exposure to structural forces (racism, segregation, access), never as intrinsic risk absent that framing; avoid race-adjusted clinical algorithms without scrutiny.
• SES measurement choice matters: individual income is rarely available; area-level indices (ADI, SVI) introduce ecological misclassification — state which and why.
• Small-stratum instability: rare groups yield unstable estimates; report precision limits honestly instead of collapsing categories silently.

Common pitfalls

• Adjusting for mediators (e.g., adjusting away insurance status when studying access interventions) erases the mechanism of interest.
• Post-hoc subgroup trawling presented as pre-specified equity analysis.
• Treating 'no significant interaction' as absence of inequity when power was inadequate.
• Ignoring differential follow-up: shorter observation in marginalized groups masquerades as lower event rates.

Pros, cons, and trade-offs

• vs pooled marginal estimation: reveals distributional effects and policy-relevant gaps vs simpler models and tighter precision.
• vs dedicated equity studies: embeds equity in existing evidence workflows vs purpose-built cohorts with richer social variables.
• Trade-off: multiple-strata inference multiplies Type I error surface — pre-specification and hierarchical testing discipline are the price of credibility.

When NOT to use

Do not perform underpowered subgroup theater for compliance optics; if the data cannot support stratified inference, say so and scope a fit-for-purpose study. And never present race-stratified associations without a structural interpretation framework.

Decision diagram

flowchart LR
  P[Protocol] --> SP[Pre-specified equity strata\n+ power + multiplicity]
  SP --> A[IPTW analysis overall]
  SP --> B[Differential-bias audit\nascertainment - censoring - missingness by stratum]
  A --> C[Stratified effects:\nmultiplicative + additive scale]
  B --> C
  C --> I[Interpretation through\nstructural mechanisms]

Worked example

Scenario

Evaluate whether a heart-failure SGLT2i effectiveness signal differs by area-deprivation quintile in a claims cohort.

Dataset

Effectiveness by ADI quintile with interaction tests.

Steps

Result

Treatment reduced hospitalization in all strata but absolute benefit was smaller in Q5 (-1.8 vs -3.1 pp); additive-scale interaction p=0.04, driven partly by differential follow-up length corrected by censoring weights.

Trade-offs

Runnable example

\
import pandas as pd
import statsmodels.formula.api as smf

def equity_stratified(df, treatment, outcome, strata="adi_q", ps_covs=None):
    # IPTW via logistic PS
    import statsmodels.api as sm
    X = sm.add_constant(df[ps_covs])
    df = df.assign(ps=sm.Logit(treatment, X).fit().predict(X))
    df = df.assign(w=df[treatment]/df.ps + (1-df[treatment])/(1-df.ps))
    # Outcome model with treatment x strata interaction (additive via GLM identity)
    m = smf.glm(f"{outcome} ~ {treatment} * C({strata})", data=df,
                family=sm.families.Gaussian(), freq_weights=df.w).fit(cov_type="HC1")
    return m  # interaction terms give stratum-specific absolute RDs

\
library(survey)

equity_stratified <- function(df, treatment, outcome, strata = "adi_q") {
  des <- svydesign(ids = ~1, weights = ~w, data = df)
  m <- svyglm(as.formula(paste(outcome, "~", treatment, "*", strata)),
              design = des)
  # Additive-scale interaction: linear model coefficients directly give RD contrasts
  regTermTest(m, strata)   # joint interaction p-value
  summary(m)$coefficients
}

\
proc genmod data=coh;
  class adi_q(ref='Q1') treatment(ref='0') / param=reference;
  weight w;
  model y = treatment adi_q treatment*adi_q / dist=normal link=identity;
  lsmeans treatment*adi_q / diff oddsratio cl;
  /* Interaction LRT */
run;

/* Differential censoring check */
proc lifetest data=coh notable;
  time followup_days*dropout(0);
  strata adi_q / testcmd=logrank;
run;

Citations

1. [1]U.S. Food and Drug Administration. FDORA Diversity Action Plans guidance activities.
2. [2]ISPOR/NPC Health Equity initiatives: good practices for incorporating health equity considerations in RWE and HEOR.
3. [3]NCI SEER-Medicare: patterns-of-care and disparities research applications.

1. [4]CDC/ATSDR Social Vulnerability Index documentation.