Chapter 2 — Key Takeaways

The reference card for the most important chapter in the book. If you keep one page from these 38 chapters, keep this one.

The Ladder

Weakest → strongest as a basis for changing what you eat:

Rung Design Bottom line
1 Mechanism / in vitro What could happen. Cells in a dish aren't a person.
2 Animal study Hypotheses and doses. Mice aren't tiny people.
3 Case report / anecdote Includes every testimonial ever filmed.
4 Cross-sectional A snapshot. Can't tell you what came first.
5 Prospective cohort Real outcomes, real timescales — and confounding.
6 RCT Causation. Usually short, small, surrogate endpoints.
7 Meta-analysis of RCTs Strongest — if the underlying trials are good.
8 Consensus of independent bodies Slow, conservative, usually right.

Place any study in ten seconds: Species? → Randomized? → Disease or marker?

🚪 The Threshold Concept

In free-living populations, the people who do one healthy thing tend to do all the others too — and this is very hard to separate from the effect of the thing being studied.

That's healthy-user bias, and it is the single biggest reason nutrition headlines are wrong.

Beta-carotene didn't prevent lung cancer. Blood beta-carotene was a marker of eating vegetables, which was a marker of not smoking, exercising, having money, and having a doctor. Two randomized trials found more lung cancer in the supplemented group. One was stopped early.

Same shape: vitamin E. Hormone therapy. Multivitamins. Most of what you've read about food.

The Five Failure Modes

Failure mode The question that catches it
Healthy-user bias What else is true about people who do this?
Residual confounding They adjusted for what they measured — how well did they measure it?
Reverse causation Could the arrow point the other way?
Unspecified substitution Compared to what?
Relative-risk inflation How big in absolute terms? Per hundred people?

Plus: surrogate endpoints (a number is not a disease) · publication bias · the garden of forking paths · funding effects in every direction · significance ≠ size.

The Claim Filter

  1. What kind of study? What rung? Species → randomized? → disease or marker?
  2. Compared to what? No comparator, no claim.
  3. Could the arrow point the other way?
  4. What else is true about people who do this? Name four in fifteen seconds.
  5. How big, in absolute terms? Convert to per-hundred.
  6. Who benefits if I believe this? Every direction. Including me.

Most claims fail at 2 or 5 — and they fail because you cannot answer. Recognizing "I don't have enough information to have an opinion" is the skill.

Reading an Abstract: the red flags

In the abstract What it means
Sample under ~50 One unusual participant drives the result
Many outcomes listed ~1 in 20 hits significance by chance
P just under .05 on one outcome Weak. Ask for the effect size.
"Post-hoc," "exploratory," "subgroup" Hypothesis, not a finding
Healthy young volunteers May not transfer to the people being sold to
Extract / isolate / standardized dose It's about the compound, not the food
Markers rather than disease Surrogate endpoint
Conclusion broader than results The gap is where the press release lives

Absolute vs. Relative — worked

Processed meat: ~18% relative increase in colorectal cancer per 50 g/day.

Baseline lifetime risk   ≈  5.0%
× 1.18                   =  5.9%
Difference               =  0.9 percentage points
                         ≈  1 extra case per 100 people, lifetime

IARC Group 1 = confidence that it causes cancer. Not magnitude. Tobacco and processed meat are both Group 1 and differ by orders of magnitude.

Relative risk without a baseline is uninterpretable.

What Good Evidence Looks Like

  1. It converges across designs with non-overlapping weaknesses — cohorts are confounded, trials aren't; trials are short, cohorts aren't. Agreement is hard to explain away.
  2. Dose-response.
  3. It survives new methods — Mendelian randomization uses genetic variants as natural randomization, sidestepping healthy-user bias.
  4. It's boring and old. Certainty and excitement are inversely related.
  5. The field corrects itself in public — PREDIMED was retracted and republished; conclusions largely held; the mechanism worked at real reputational cost.

Common Mistakes (and the fix)

Mistake Fix
Trusting a big cohort over a small trial for causal claims Size doesn't fix bias — it gives you a precise estimate of a biased quantity
"They adjusted for it, so it's fine" Adjustment removes what you measured, as well as you measured it
Reading relative risk as absolute risk Always ask: increase from what?
Treating "reduces inflammation" as a health outcome It's a lab value with an unvalidated chain to disease
Accepting a subgroup finding as a result Post-hoc subgroups are hypotheses
Concluding "only RCTs count" You'd have to discard the evidence on smoking
Concluding nothing is knowable Over-skepticism is credulity with extra steps

Verdict Summary

Claim Verdict Why
Breakfast is the most important meal; skipping causes weight gain 🟠 Probably false Strong observational signal, healthy-user bias + reverse causation; trials matching calories don't reproduce it
Processed meat is Group 1 — same as smoking, no safe amount 🟡 Unclear / it depends Classification correct, inference wrong; Group 1 is confidence, not magnitude
Observational studies are worthless; only RCTs count 🟠 Probably false Discards the only design that measures hard outcomes over decades; RCTs have their own failure modes

What We Still Don't Know

We cannot quantify residual confounding in nutritional epidemiology. We know it's there, we know its usual direction, and we cannot say how much of a reported 20% risk reduction is real.

That is the unresolved problem in the field. It's why serious people disagree about red meat, eggs, dairy, and moderate alcohol — not because they read different data, but because they judge differently how much signal survives the confounding.

Anyone who tells you it's settled, in either direction, is telling you about their confidence rather than about the evidence.

One Thing to Remember

The people who do one healthy thing tend to do all of them.

Once you can see that, you can't unsee it — and about half of everything you've ever read about food rearranges itself.