Causal Inference of Blood Pressure Reduction and Coronary Heart Disease Risk in the Framingham Study

By applying causal inference methods to the Framingham Heart Study, this paper demonstrates that standard observational risk calculators overestimate the absolute benefit of blood pressure reduction on coronary heart disease risk by approximately 21.8% due to confounding, highlighting the critical need to distinguish between conditional probabilities and interventional effects in clinical decision-making.

The Gist

Imagine you are trying to figure out how much a specific action—like lowering your blood pressure—actually helps prevent a heart attack. For decades, doctors and risk calculators have used a method that looks at people who naturally have low blood pressure and compares them to people who naturally have high blood pressure.

The problem, according to this study, is that this comparison is like judging a race by looking at the runners' shoes instead of their running ability.

Here is the breakdown of what the researchers found, using simple analogies:

The "Natural" vs. "Forced" Mistake

Think of blood pressure like the temperature in a house.

• The Old Way (Observational): Researchers looked at houses that are naturally cool and compared them to houses that are naturally hot. They noticed the hot houses had more broken windows (heart disease). They concluded: "If we cool the house down, we will save a lot of windows."
• The Flaw: The houses that are naturally hot often have other problems too: they are older, the insulation is worse, and the owners are less careful with maintenance. The heat isn't the only reason the windows are breaking; the age and poor maintenance are, too.
• The New Way (Causal): This study asked a different question: "If we take a house that is naturally hot and force the thermostat down (using medicine), how many windows do we actually save?"

The study found that the "Old Way" was overestimating the benefit by about 22%. It was blaming the heat for all the broken windows, when in reality, the age of the house and poor maintenance were doing a lot of the damage. When you fix the heat, you fix the heat, but you don't magically fix the age or the maintenance issues.

The "Map" Correction

To get this right, the researchers drew a new "map" (called a Directed Acyclic Graph or DAG) of how heart health works. They found four errors in the maps everyone else had been using:

1. Smoking: They removed a line suggesting smoking directly causes high blood pressure in this specific context, realizing it affects heart health through other paths (like cholesterol) first.
2. Age: They added a line showing that getting older causes diabetes, which then affects heart health.
3. Medication: They realized that taking blood pressure pills is a result of having high blood pressure, not a cause of it. Treating the pills as a cause was like blaming the fire extinguisher for starting the fire.
4. History: They stopped counting "past high blood pressure" as a separate cause, because it's just a symptom of the current high blood pressure.

By fixing these map errors, they could calculate the true effect of the "thermostat adjustment."

The Results: A More Honest Number

When they did the math correctly:

• The Old Guess: Lowering blood pressure by 20 points would prevent heart disease in about 4.14% of people.
• The New Truth: Lowering blood pressure by 20 points actually prevents heart disease in about 3.40% of people.

While 3.40% still sounds good, the study shows that the "Old Guess" was too optimistic. If a doctor uses the old number to decide whether to prescribe medicine, they might think the benefit is big enough to treat a patient, when the real benefit might be just below the threshold for treatment.

Who Benefits Most?

The study also looked at whether this "fix" works differently for different people:

• Age: Older people seem to get a bigger absolute benefit from lowering blood pressure than younger people. This makes sense because older people have more "broken windows" to begin with.
• Diabetes: The researchers tried to look at people with diabetes, but there were too few of them in the study to say anything reliable. They explicitly warn that we cannot draw conclusions about diabetics from this data yet.

The Bottom Line

This study doesn't say "don't lower blood pressure." It says: "Be careful about how much you think lowering blood pressure will help."

Standard risk calculators are like a weather forecast that predicts a storm based on how dark the sky looks right now, ignoring that the sky is dark because of a storm cloud that is already moving away. By using a more advanced method (called causal inference), the researchers showed that the "storm" (heart disease risk) is slightly less severe than the old forecasts predicted when you isolate the effect of blood pressure alone.

This helps doctors make more precise decisions, ensuring they aren't overestimating the "magic" of a single treatment when a patient's overall health picture is much more complex.

Technical Summary

Technical Summary: Causal Inference of Blood Pressure Reduction and Coronary Heart Disease Risk in the Framingham Study

Problem Statement
Standard cardiovascular risk calculators, such as the Framingham Risk Score and ACC/AHA Pooled Cohort Equations, estimate the conditional probability of Coronary Heart Disease (CHD) given a specific blood pressure profile, $P(\text{CHD} \mid \text{SysBP}=s, Z=z)$. This quantity reflects the observed frequency of disease among individuals who naturally present with certain characteristics. However, clinical decision-making regarding antihypertensive therapy requires the interventional quantity, $P(\text{CHD} \mid \text{do}(\text{SysBP}=s))$, which represents the risk if blood pressure were pharmacologically set to a target value, independent of the patient's underlying biological state (e.g., age, obesity, metabolic dysregulation).

The paper argues that conflating these two quantities leads to a systematic overestimation of the absolute benefit of blood pressure reduction. Individuals with naturally elevated blood pressure often possess a higher burden of comorbidities (older age, higher BMI, diabetes) compared to those with lower blood pressure. Observational estimates capture the risk associated with this entire profile, whereas antihypertensive therapy only severs the causal path from pressure to vascular events. Consequently, observational tools may inflate the perceived efficacy of treatment, potentially influencing prescribing thresholds and risk stratification.

Methodology
The study applies Pearl's do-calculus and structural causal modeling to the Framingham Heart Study Offspring Cohort ($n=4,240$; primary analysis on $n=3,776$ complete cases). The methodological pipeline includes:

1. Structural Causal Model (DAG) Specification: A Directed Acyclic Graph (DAG) was constructed with 11 nodes and 24 edges. The authors introduced four biologically motivated corrections to prior formulations:

   • Removal of Smoking $\to$ SysBP: Acknowledging that while nicotine causes acute vasoconstriction, it does not constitute a structural population-level causal path from smoking status to sustained systolic blood pressure in this cross-sectional context.
   • Addition of Age $\to$ Diabetes: Incorporating the sharp increase in Type 2 diabetes incidence with age, recognizing age as a shared cause of both blood pressure elevation and CHD.
   • Reclassification of Medication (BPMEDS): Treating antihypertensive medication as a descendant of Systolic Blood Pressure (SysBP) rather than a confounder. Including it in the adjustment set would induce collider bias.
   • Exclusion of Prevalent Hypertension: Removing prevalent hypertension from the primary adjustment set to avoid over-adjustment bias, as it is causally downstream of long-term blood pressure.

   The validity of the DAG was tested via conditional independence tests (partial correlations), all of which passed at $\alpha=0.05$.

2. Causal Effect Estimation:

   • Primary Estimator: G-computation (standardization) was used to estimate the Average Causal Effect (ACE) of a 20 mmHg reduction in SysBP. The back-door adjustment set $Z = \{\text{Age, Sex, BMI, Current Smoking}\}$ was used. Total cholesterol and glucose were included as precision covariates in the outcome regression.
   • Triangulation: Results were corroborated using sex-stratified Propensity Score Matching (PSM) and Inverse Probability Weighting (IPW).
   • Heterogeneity: Conditional Average Treatment Effects (CATE) were estimated using R-Learner and T-Learner metalearners with gradient-boosted nuisance models to assess effect modification by age, sex, and diabetes status.
   • Robustness: The study employed permutation refutation tests and placebo instruments (e.g., heart rate $\to$ age) to validate the causal estimates. E-values were calculated to quantify the robustness of the findings to unmeasured confounding.

Key Results

• Prediction-Causation Gap: The g-computation approach yielded an ACE of 3.40% absolute risk reduction (95% bootstrap CI: 2.64%–4.14%) for a 20 mmHg reduction in SysBP. In contrast, the naive observational estimate was 4.14%. This represents a relative overestimation of approximately 21.8% by the observational approach.
• Bias Source: The divergence is driven by the clustering of older age, higher BMI, and metabolic dysregulation with elevated blood pressure. The observational estimate captures the risk of the comorbid profile, not just the blood pressure reduction.
• Heterogeneity: Statistically significant heterogeneity was detected across age strata ($p < 0.001$) and diabetes status ($p < 0.001$). Older patients (45–64 years) showed higher estimated absolute risk reductions (~4.5–4.8%) compared to younger patients (<45 years, ~2.2%).
• Subgroup Limitations: Estimates for the diabetic subgroup ($n=109$) were unstable and underpowered, with a wide confidence interval crossing zero. The authors explicitly state that no reliable inference regarding differential benefit in diabetics can be drawn from this data without replication.
• Robustness: The E-value lower bound was 2.18, indicating that an unmeasured confounder would need to have a risk ratio of at least 2.18 with both SysBP and CHD to fully explain away the observed effect. This exceeds the strength of association of all measured covariates.

Significance and Claims
The paper claims to be the first to quantify the magnitude of bias in standard observational cardiovascular risk tools within the Framingham cohort, demonstrating that these tools may overestimate the absolute benefit of antihypertensive treatment by roughly one-fifth.

The authors argue that this overestimation has direct clinical consequences. For patients near prescribing thresholds, a 21.8% inflation in estimated benefit could lead to treatment decisions where the true causal benefit falls below a clinically meaningful threshold. The study suggests that clinical risk stratification should ideally rely on interventional estimates rather than conditional associations.

Methodologically, the paper contributes an end-to-end reproducible pipeline integrating DAG evaluation, g-computation, PSM, IPW, and semiparametric metalearners. It highlights specific structural errors in previous causal analyses of cardiovascular data, particularly the incorrect treatment of antihypertensive medication as a confounder rather than a descendant of blood pressure, which introduces collider bias.

Limitations Acknowledged by the Authors
The authors maintain a modest tone regarding their conclusions, noting several limitations:

• Unmeasured Confounding: The assumption of no unmeasured confounding is unverifiable; factors like diet, physical activity, and genetics could bias results.
• Cross-Sectional Nature: The analysis relies on baseline data, failing to account for time-varying confounding over the 10-year follow-up, which may lead to conservative estimates.
• DAG Simplification: The DAG is a simplified representation; failures in some placebo tests suggest latent shared causes not encoded in the model.
• Generalizability: The Framingham Offspring Cohort is a largely white, community-based sample from mid-20th century New England, limiting generalizability to contemporary, diverse, or higher-risk populations.
• Subgroup Power: The diabetic subgroup analysis is explicitly flagged as underpowered and unreliable for inference.

The study positions its findings as hypothesis-generating, calling for future validation against randomized trial data (e.g., SPRINT, ACCORD-BP) and replication in larger, more diverse cohorts.