Breathing the Divide — COPD Prevalence, Air Pollution & Smoking Across U.S. States

Chronic obstructive pulmonary disease affects roughly 6% of American adults, but that single national number hides a lot. This study asks two things: how much does COPD prevalence vary across U.S. states and regions, and does breathing dirtier air actually cause more of it? Using CDC survey data, census-tract PM2.5 predictions and state demographics for 2020, it answers the first question clearly and the second one honestly — which is to say, in the negative.

6.05% Modeled national COPD prevalence 94% HDI 4.28 – 7.29

1.6 pp Spread between the highest and lowest regional baseline ≈27% relative variation

p = 0.127 Causal effect of elevated PM2.5 — not significant 95% CI −0.16 to 1.22 pp

Δ4.71 WAIC gain of the hierarchical model over a pooled one Lower WAIC = better fit

01Why this question

COPD is a progressive lung disease, and smoking is its dominant risk factor. But a CDC surveillance report found that about 25% of cases occur in people who have never smoked — which leaves a substantial share of the burden unexplained by tobacco alone. Air quality is the obvious suspect, and a large meta-analysis found that every 10 µg/m³ increase in PM2.5 (fine airborne particulate matter) was associated with an 18% higher incidence of COPD.

That motivates two distinct research questions, which need two different statistical tools:

RQ1 — How does COPD prevalence vary across states and regions? Answered with a Bayesian hierarchical model, because the data has a natural nesting: states sit inside regions, and regions share air quality and policy environments.
RQ2 — Does elevated PM2.5 cause higher COPD prevalence? Answered with causal inference via outcome regression, because association alone would not justify the air-quality policy conclusions the question invites.

02The data

The outcome comes from the CDC's U.S. Chronic Disease Indicators, filtered by API to the COPD topic. The underlying source is the BRFSS — a health-related telephone survey, not a census, which matters for how far the results generalise.

The analysis uses age-adjusted rather than crude prevalence. Standardising the age distribution across states makes them comparable, and it also blunts a real selection-bias worry: a telephone survey is likely to over-sample particular age groups, and age adjustment strips out that weighting.

Exposure comes from the CDC's Daily Census Tract-Level PM2.5 Concentrations dataset (2016–2020), averaged up to state level. That aggregation is a genuine cost — a state average can hide the specific counties driving a high prediction. Smoking rates, urban household share, median household income and 2020 census populations round out the covariates.

03How it was modeled

RQ1 — Bayesian hierarchical model

For each state i in region j, expected prevalence is a regional baseline plus a regional PM2.5 slope plus a gender-gap term: mu = alpha_region[j] + beta_pm25_region[j] × pm25 + beta_gender × gender_gap.

The important choice here is partial pooling. Rather than estimating each region independently (which overfits when some regions hold only a few states) or forcing them all to share one value (which erases real differences), regional parameters are drawn from a shared population distribution. Regions with less data or more extreme values get pulled harder toward the national mean. Priors were kept weakly informative — the population intercept centred at 6% with a standard deviation of 2 — so the data, not the prior, drives the answer. Sampling used PyMC's NUTS sampler: 2 chains, 1,000 tuning steps and 2,000 draws each, for 6,000 total draws, seed fixed at 42.

RQ2 — Causal estimate

Treatment is a binary flag for states whose 2020 PM2.5 sits above the 75th percentile of the state distribution. The estimate is an OLS outcome regression adjusting for smoking prevalence, urban household share and median household income — all standardised, which resolved a multicollinearity warning flagged by a very high condition number in the unscaled fit.

04What the models found

The headline result is easiest to read as effect sizes against zero. Every estimated effect in this study is positive in expectation — and every one of them has an interval that crosses zero.

Fig. 01

Every estimated effect on COPD prevalence includes zero

RQ1 · Bayesian hierarchical, 94% HDI RQ2 · OLS, 95% CI

All three point estimates lean positive, and all three intervals contain zero. The PM2.5 effect (0.30 pp per standard deviation) and the causal estimate (0.53 pp) point the direction you would expect — dirtier air, more COPD — but neither is distinguishable from no effect at all. The gender-gap term is close to zero outright: gender disparities in COPD exist, but the size of a state's gap does not predict its overall prevalence. Interval widths are not comparable across the two models; the treatment definitions differ, as the row labels note.

View as table

Parameter	Model	Mean	SD	Lower	Upper
Global intercept (%)	Hierarchical	6.054	0.688	4.282	7.285
PM2.5 effect (pp / SD)	Hierarchical	0.303	0.330	−0.269	0.983
Gender-gap effect (pp / SD)	Hierarchical	0.060	0.225	−0.356	0.471
Regional variation (σ)	Hierarchical	0.976	0.465	—	—
Elevated PM2.5 (pp)	OLS	0.533	0.343	−0.16	1.22

Source: posterior summaries from the Bayesian hierarchical model (94% HDI) and the OLS outcome regression (95% CI, p = 0.127), 48 states, 2020.

Where the 6.05% actually sits

The model's national estimate lands almost exactly on the CDC's reported figure, which is a good sign that it is calibrated. But putting it next to the real observed spread shows what partial pooling costs:

Fig. 02

A calibrated national estimate sits inside a much wider real spread

Modeled in this study External reference

The modeled intercept lands on the CDC's national figure, so the model is calibrated where it should be. The two band rows are deliberately not like-for-like: the upper band is four shrunk regional posterior means, the lower is the observed range across individual states. That contrast is the point — partial pooling compresses regional estimates into a 1.6 pp window, while real state-to-state variation runs nearly 9 pp wide. The model is built to describe regions, not to reproduce the extremes of any single state.

View as table

Quantity	Kind	Point	Lower	Upper
Modeled global intercept	Posterior, 94% HDI	6.054	4.282	7.285
Regional intercepts (4 regions)	Posterior span	—	≈5.3	≈6.9
CDC national prevalence	Reference	≈6.0	—	—
Observed state prevalence	Reported range	—	3.0	11.8

Source: posterior summaries from this study; national and state-range reference values from Liu et al. (2023), CDC MMWR 72(46). All values are age-adjusted prevalence.

Robustness

The PM2.5 result is not an artefact of one prior. Re-running the model with the regional slope prior set to HalfNormal(0.25), HalfNormal(0.5) and HalfNormal(1.0) kept the population-level PM2.5 effect in the 0.25 – 0.35 range every time, with credible intervals that always included zero. Swapping regional random effects for state-level fixed effects made the coefficient less certain, not more — which is itself an argument for the hierarchical structure.

Where this disagrees with prior work

Park et al. (2021) found a statistically significant 18% increase in COPD incidence per 10 µg/m³ of PM2.5. This study finds nothing significant. The two are less contradictory than they look: that meta-analysis pooled individual-level data with follow-up over time, while this analysis has 48 state averages from a single year. COPD develops over decades, so one year of air quality is a poor proxy for the exposure that actually caused it.

Model comparison and diagnostics

The hierarchical model beat a fully pooled alternative on WAIC by 4.71, so the added complexity earns its place. Diagnostics were acceptable but not spotless: some divergences appeared during sampling, R-hat sat around 1.02 for the global intercept and regional variance, and tail effective sample sizes were low for those parameters. With only 4 regions, some of them small, that is a familiar pattern. Posterior predictive checks against 100 simulated datasets tracked the observed data's centre and spread reasonably well.

05What would change the answer

The honest summary is that this analysis is well-powered for description and under-powered for causation. Four things drive that:

One year of data. 2020 was chosen because it maximised overlap between datasets, but COPD accumulates over decades — and 2020 was not a typical year for respiratory health or behaviour. Historical 10–20 year PM2.5 averages would align exposure with the disease's actual timescale.
Missing confounders. Occupational exposure in mining, manufacturing and agriculture; healthcare access, insurance coverage and pulmonologist density; indoor air quality from cooking, heating and secondhand smoke. The unconfoundedness assumption behind the causal estimate almost certainly does not hold, and the write-up says so directly.
Ecological aggregation. State averages can mask intra-state disparities, and a state-level association need not hold for individuals. Individual BRFSS microdata or county-level modeling would address this.
Treatment definition. "Elevated" PM2.5 was set relative to other states, at the 75th percentile, to keep the comparison groups balanced. Measured against the EPA standard instead, only a few states would qualify — a more meaningful threshold, but a much more lopsided design.

None of that makes the descriptive finding go away. Regional clustering in COPD prevalence is real and substantial, the point estimates for pollution consistently lean positive, and the wider epidemiological literature supports acting on air quality regardless of what one cross-section can prove. Targeting screening, smoking-cessation resources and clean-air investment where the burden concentrates does not require this particular p-value to cooperate.

06The full report

Everything above is a summary. The complete write-up includes the model's plate diagram, all seven exploratory figures — prevalence by state, regional box plots, the PM2.5 scatter, prevalence by race and ethnicity, the gendered comparison across the top 15 states, and the smoking-equality plot — plus the full methods, sensitivity analyses and references.

Read in full Breathing the Divide — complete report → Full PDF, ~20 pages, viewable in the browser or downloadable.

07References

Liu, Y., Carlson, S. A., Watson, K. B., Xu, F., & Greenlund, K. J. (2023). Trends in the prevalence of chronic obstructive pulmonary disease among adults aged ≥18 years — United States, 2011–2021. Morbidity and Mortality Weekly Report, 72(46), 1251–1256. doi.org/10.15585/mmwr.mm7246a1
Park, J., Kim, H.-J., Lee, C.-H., Lee, C. H., & Lee, H. W. (2021). Impact of long-term exposure to ambient air pollution on the incidence of chronic obstructive pulmonary disease: A systematic review and meta-analysis. Environmental Research, 194, Article 110703. doi.org/10.1016/j.envres.2020.110703
Centers for Disease Control and Prevention (2024). U.S. Chronic Disease Indicators; Behavioral Risk Factor Surveillance System; Daily Census Tract-Level PM2.5 Concentrations, 2016–2020.