Causal Inference of Blood Pressure Reduction and Coronary Heart Disease Risk in the Framingham Study
Suchibrata Patra · 2026
Read Causal Inference of Blood Pressure Reduction and Coronary Heart Disease Risk in the Framingham Study by Suchibrata Patra (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05666.txt.
Introduction
Risk calculators tell doctors the chance of heart disease among people observed at a given blood pressure. Suchibrata Patra argues that is the wrong number for deciding treatment. What treatment needs is the interventional probability: the risk if pressure were actually set to that value. Confounding can make the observational figure overstate the benefit of lowering pressure.
Patra applies Pearl's do-calculus to the Framingham Heart Study offspring cohort, with over four thousand participants and 574 coronary events in ten years. A corrected causal graph of the variables supports the analysis. The study separates genuine treatment effect from statistical association.
Catalog record
| CREATOR | Suchibrata Patra |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05666
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Causal Inference of Blood Pressure Reduction and Coronary Heart Disease Risk in the Framingham Study Suchibrata Patra Abstract Standard cardiovascular risk calculators, including the Framingham Risk Score and the ACC/AHA Pooled Cohort Equations, estimate the conditional probability P ( CHD ∣ SysBP = s ) P(\mathrm{CHD}\mid\mathrm{SysBP}=s) rather than the interventional quantity P ( CHD ∣ do ( SysBP = s ) ) P(\mathrm{CHD}\mid\mathrm{do}(\mathrm{SysBP}=s)) . When confounding is present, this distinction has direct clinical consequences: observational estimates may systematically overstate the absolute benefit of antihypertensive treatment. We applied Pearl’s do-calculus to the Framingham Heart Study Offspring Cohort ( n = 4,240 n=4{,}240 ; primary analysis on 3,776 3{,}776 complete cases; 574 ten-year coronary heart disease (CHD) events). A structurally corrected directed acyclic graph (DAG) was specified, incorporating four biologically motivated corrections, and subjected to conditional independence testing. The average causal effect (ACE) of a 20 mmHg systolic blood pressure (SysBP) reduction was estimated by g-computation with 1,500-iteration bootstrap confidence intervals, corroborated by sex-stratified propensity score matching (PSM) and inverse probability weighting (IPW). Conditional average treatment effects (CATE) were estimated using R-Learner and T-Learner metalearners with gradient-boosted nuisance models. G-computation yielded an ACE of 3.40% absolute risk reduction (95% bootstrap CI: 2.64%–4.14%), compared with a naive observational estimate of 4.14%, a relative overestimation of approximately 21.8% ( + 35.7 % +35.7\% on the log-odds scale). The E-value lower bound was 2.18. Statistically significant heterogeneity in treatment effect was detected across age strata (Kruskal–Wallis p < 0.001 p<0.001 ) and diabetes status (Mann–Whitney p < 0.001 p<0.001 ); however, diabetic subgroup estimates were unstable and underpowered ( n = 109 n=109 ), and no reliable subgroup inference can be drawn without replication. These findings suggest that observational cardiovascular risk tools may overestimate the absolute benefit of blood pressure reduction, with implications for clinical risk stratification and prescribing thresholds. keywords causal inference; directed acyclic graph; g-computation; heterogeneous treatment effects; coronary heart disease; systolic blood pressure Introduction Every major cardiovascular risk tool in routine clinical use, the Framingham Risk Score [ 1 ] , SCORE2 [ 2 ] , and the ACC/AHA Pooled Cohort Equations [ 3 ] , estimates a conditional probability of the form P(\mathrm{CHD}\mid\mathrm{SysBP}=s,\,Z=z) . This quantity describes the observed frequency of coronary heart disease (CHD) among individuals who happen to present with a given blood pressure profile. The quantity a prescribing clinician requires when deciding whether to initiate antihypertensive therapy is categorically different: P\!\left(\mathrm{CHD}\mid\mathrm{do}(\mathrm{SysBP}\leftarrow s)\right), (1) where Pearl’s do-operator [ 4 ] represents the act of pharmacologically setting blood pressure to a target value, regardless of the patient’s underlying biological state. Individuals who naturally present with \mathrm{SysBP}=140 mmHg are, on average, older, more obese, and more metabolically dysregulated than those presenting at 120 mmHg. Antihypertensive therapy severs only the causal path from elevated pressure to vascular events; it does not reverse these pre-existing comorbidities. Conflating the observational and interventional quantities therefore produces a systematically inflated estimate of treatment benefit, a gap with direct consequences for clinical decision-making and health economic evaluation [ 5 ] . This prediction–intervention gap is formally resolvable using structural causal models and the back-door adjustment theorem [ 4 ] . Under this framework, the interventional distribution ( 1 ) is identifiable from observational data provided that a valid directed acyclic graph (DAG) can be specified and that the positivity and conditional ignorability assumptions hold. A growing methodological literature has addressed components of this problem: marginal structural models have been applied to blood pressure trajectories [ 7 ] , inverse probability weighting (IPW) has been reviewed in cardiovascular epidemiology [ 8 ] , and the distinction between associational and interventional estimands has been extensively theorised [ 5 , 9 ] . Nevertheless, no prior study has delivered, within a single reproducible pipeline applied to the Framingham Heart Study, a biologically corrected and empirically evaluated causal DAG, multiple corroborating causal-effect estimators, and an explicit quantification of the bias incurred by standard observational approaches. The present study makes three contributions. First, we quantify, for the first time in this cohort, the magnitude by which the naive observational estimate of blood pressure treatment benefit exceeds the causal estimate (approximately 21.8% in relative terms). Second, we specify, correct, and empirically evaluate a structural causal DAG, identifying four errors in prior formulations that propagate into biased causal estimates. Third, we implement an end-to-end causal inference pipeline, covering DAG evaluation, g-computation, sex-stratified propensity score matching (PSM), IPW, refutation testing, and semiparametric treatment-effect heterogeneity estimation, within a single reproducible framework, and discuss the clinical implications of the resulting estimates. Results Baseline characteristics The analytical cohort comprised 3,776 complete-case participants (55.5% female; mean age 49.6\pm 8.6 years; mean systolic blood pressure (SysBP) 132.4\pm 22.0 mmHg; 2.7% with diabetes mellitus). As shown in Table 1 , participants who developed CHD within ten years were on average five years older ( 54.2 vs. 48.7 years; Mann–Whitney p<0.001 ), had 13.5 mmHg higher SysBP ( 143.8 vs. 130.3 mmHg; p<0.001 ), and had substantially higher prevalences of diabetes (6.3% vs. 2.1%; \chi^{2} , p<0.001 ) and prevalent hypertension (50.7% vs. 27.6%; p<0.001 ). These systematic differences illustrate the confounding structure that motivates causal rather than observational estimation: patients presenting with elevated SysBP are metabolically and cardiovascularly more compromised across multiple dimensions, independently of their blood pressure level. Table 1 : Baseline characteristics by 10-year CHD status. Data are mean \pm SD (continuous variables) or n (%) (binary variables). p -values from Mann–Whitney U test (continuous) or \chi^{2} test (binary). The systematic confounding structure, with older age, higher SysBP, higher body mass index (BMI), and greater comorbidity burden among CHD cases, motivates causal rather than observational estimation. Variable Overall No CHD CHD p (n=3{,}776) (n=3{,}202) (n=574) Age (years) 49.6\pm 8.6 48.7\pm 8.4 54.2\pm 8.0 <0.001 SysBP (mmHg) 132.4\pm 22.0 130.3\pm 20.4 143.8\pm 26.8 <0.001 DiaBP (mmHg) 82.9\pm 11.9 82.2\pm 11.3 87.2\pm 14.3 <0.001 Total cholesterol (mg/dL) 237.0\pm 44.7 235.3\pm 43.8 246.3\pm 48.1 <0.001 BMI (kg/m 2 ) 25.8\pm 4.1 25.7\pm 3.9 26.6\pm 4.6 <0.001 Glucose (mg/dL) 81.9\pm 23.8 80.7\pm 19.0 88.8\pm 40.8 <0.001 Heart rate (bpm) 75.7\pm 12.0 75.6\pm 11.9 76.4\pm 12.1 0.299 Male sex, n (%) 1,682 (44.5%) 1,362 (42.5%) 320 (55.7%) <0.001 Current smoker, n (%) 1,857 (49.2%) 1,561 (48.8%) 296 (51.6%) 0.231 Diabetes mellitus, n (%) 102 (2.7%) 66 (2.1%) 36 (6.3%) <0.001 Antihypertensive med., n (%) 114 (3.0%) 77 (2.4%) 37 (6.4%) <0.001 Prevalent hypertension, n (%) 1,176 (31.1%) 885 (27.6%) 291 (50.7%) <0.001 Observational prediction model A multivariable logistic regression achieved an AUROC of 0.721 (5-fold cross-validation: 0.721\pm 0.028 ; average precision: 0.339; Brier score: 0.116). The observational odds ratio per mmHg SysBP was 1.013 (95% CI: 1.008–1.018; p<0.001 ). Age (OR: 1.063), male sex (OR: 1.805), and current smoking (OR: 1.477) were the strongest associated predictors. These coefficients reflect the conditional distribution of CHD given observed covariates and carry no causal interpretation. Directed acyclic graph specification and evaluation The corrected DAG (Figure 1 ) encoded 11 nodes and 24 directed edges incorporating four biologically motivated structural corrections. Correction 1: Removal of CURSMOKE to SYSBP. Nicotine causes acute transient vasoconstriction [ 12 ] , but in this cross-sectional cohort this does not constitute a structural population-level causal path from smoking status to