MC-V2-09

## The Inferential Error: Claiming Causation from Observational Data The researcher commits the **causal inference fallacy** — concluding that breakfast *causally reduces* cardiovascular mortality from what is, by design, an **observational cohort study**, not a randomized controlled trial. ### Why the conclusion doesn't follow from the evidence **1. Study design cannot establish causation.** This is a prospective cohort study. Participants self-selected into breakfast-eating vs. non-eating groups. No intervention was performed — no one was *randomly assigned* to eat or skip breakfast. Without randomization, the observed association cannot be elevated to a causal claim, no matter how statistically significant it is. **2. Residual and unmeasured confounding.** The researchers adjusted for age, sex, BMI, smoking, exercise, and socioeconomic status — but these are only *some* potential confounders. Plausible unmeasured common causes of both breakfast eating and cardiovascular health include: - **Overall health consciousness** — people who eat breakfast may also sleep better, manage stress more effectively, attend medical checkups, and adhere to medications - **Mental health status** — depression and anxiety are linked to both meal-skipping and cardiovascular risk - **Meal quality and overall dietary pattern** — breakfast eaters may consume healthier diets overall - **Shift work / sleep disruption** — associated with both skipping breakfast and cardiovascular disease - **Alcohol consumption patterns** These backdoor paths between breakfast and mortality remain unblocked. Statistical adjustment for a handful of covariates does not simulate an interventional distribution. **3. Healthy user bias.** Breakfast consumption is a marker of a generally regulated, health-conscious lifestyle. The association likely reflects this broader behavioral cluster rather than an isolated causal effect of one meal. **4. Impressive-looking features don't rescue the design.** The large sample size (50,000), long follow-up (20 years), tight confidence intervals, dose-response gradient, and consistency across subgroups make the *association* more robust — but none of these features can convert an observational association into a causal effect. Dose-response and consistency are part of the Bradford Hill criteria, which are *guidelines for plausibility*, not a causal identification strategy. ### What the conclusion should say > "Habitual breakfast consumption is **associated with** lower cardiovascular mortality. However, as this is an observational study, **causation cannot be inferred**. The observed association may be partly or wholly explained by unmeasured confounding, particularly healthy-user behaviors correlated with breakfast eating. Randomized trials or natural experiments would be needed to establish a causal effect before issuing public health recommendations." ### Bottom line The specific error is **treating a well-adjusted observational association as if it were an interventional causal estimate**. No amount of covariate adjustment in a cohort study can substitute for the confounding elimination that randomization provides, because you can never rule out what you didn't measure.

## The Inferential Error: Causal Claim from Observational Data **Step 1 — State the claim:** Breakfast eating (X) *causally* reduces cardiovascular mortality (Y). **Step 2 — Classify the design:** Prospective cohort. Participants *self-selected* into breakfast or no-breakfast groups. No randomization. No intervention. This estimates P(Y|X), not P(Y|do(X)). ### The specific error The researcher commits **causal inference from observational association without a valid identification strategy** — colloquially the correlation-causation fallacy, but the precise failure is this: **the adjusted hazard ratio cannot be interpreted as a causal effect unless all backdoor paths are blocked**, and they are not. **Why the adjustments are insufficient:** The HR dropped from 0.73 → 0.81 after adjustment. That *attenuation* confirms confounding is present. But the residual association cannot be attributed to breakfast itself because unmeasured common causes remain: | Latent confounder | Drives breakfast eating | Drives lower CV mortality | |---|---|---| | Health-conscious lifestyle | Yes | Yes | | Circadian regularity / stable routine | Yes | Yes | | Food security / socioeconomic stability (beyond crude SES) | Yes | Yes | | Absence of subclinical illness (reverse causation) | Yes — sick people lose appetite | Yes — they die sooner | The last row is especially important: **reverse causation**. People with early-stage cardiovascular disease may skip breakfast due to poor appetite, nausea, or fatigue. The causal arrow could run *Y → X*, not *X → Y*. ### Why the public health recommendation compounds the error Recommending breakfast as an *intervention* requires answering: "If we force non-breakfast-eaters to eat breakfast, will their mortality fall?" That requires estimating a *do*-distribution — which this study cannot provide. It would need: - A randomized controlled trial, or - A natural experiment / instrumental variable that shifts breakfast behavior exogenously The dose-response finding (4% per additional day) and subgroup consistency are often cited as causal evidence, but **neither satisfies the backdoor criterion**. Both are consistent with a confounded association that scales smoothly. ### Summary The researcher conflates **statistical association (even a robust, adjusted one) with causal effect**, then builds a public health intervention recommendation on that conflation. The conclusion should read: *"Breakfast consumption is associated with lower cardiovascular mortality; randomized evidence is needed before causal conclusions or interventions are warranted."* The formal label: **confounding + reverse causation in observational epidemiology, producing an unwarranted causal leap.**