Chapter 2 — Quiz

Twenty-two questions. This is the chapter the rest of the book assumes. Attempt each before opening the answer.

Multiple Choice

1. Which sequence correctly orders these from weakest to strongest as a basis for changing your diet?

  • a) Cohort study → RCT → animal study → meta-analysis of RCTs
  • b) Animal study → cohort study → RCT → meta-analysis of RCTs
  • c) RCT → cohort study → meta-analysis of RCTs → animal study
  • d) Anecdote → RCT → animal study → cohort study
Answer **b.** Rungs 2 → 5 → 6 → 7. The one people most often invert is cohort versus RCT: a 90,000-person cohort *feels* stronger than a 60-person trial, but size doesn't fix confounding.

2. The beta-carotene story is important primarily because:

  • a) The observational data was fraudulent
  • b) The observational data was accurate, but the association was not causal
  • c) The trials were poorly designed
  • d) Beta-carotene is toxic at any dose
Answer **b.** People with more blood beta-carotene really did get less lung cancer. Beta-carotene was a marker of vegetable intake, which was a marker of a whole cluster of protective behaviors. The association was real; the causal interpretation was backwards.

3. Healthy-user bias is best described as:

  • a) Researchers unconsciously favouring healthy foods
  • b) Participants lying about their diets
  • c) The tendency of people who do one healthy behavior to also do many others, confounding the effect of any single one
  • d) Healthier people being more likely to volunteer for studies
Answer **c.** (d) is *volunteer bias* — related, but distinct. (b) is measurement error. Healthy-user bias is specifically the clustering of health behaviors within individuals.

4. "They adjusted for smoking, income, and exercise" is insufficient reassurance mainly because:

  • a) Adjustment is a statistically invalid technique
  • b) You can only adjust for what you measured, and only as well as you measured it
  • c) Adjustment always removes the true effect
  • d) Confounders don't exist in large studies
Answer **b.** Both halves matter. Unmeasured confounders can't be adjusted for at all; measured ones are only partially removed because the measurement is coarse. What's left is residual confounding, which in nutrition is often larger than the effect under study.

5. IARC Group 1 classification indicates:

  • a) The magnitude of risk the agent poses
  • b) The strength of the evidence that the agent causes cancer in humans
  • c) The dose above which the agent becomes carcinogenic
  • d) That the agent should be banned
Answer **b.** It is a confidence statement, not a magnitude statement. Tobacco and processed meat are both Group 1; their magnitudes differ by orders of magnitude.

6. A relative risk increase of 18% applied to a 5% baseline gives an absolute risk of:

  • a) 23%
  • b) 18%
  • c) 5.9%
  • d) 0.9%
Answer **c.** 5% × 1.18 = 5.9%. The *difference* is 0.9 percentage points (d) — roughly one extra case per 100 people. (a) is the classic error of adding the relative percentage to the baseline.

7. Which of these is the weakest surrogate endpoint?

  • a) Blood pressure, for stroke risk
  • b) LDL cholesterol, for cardiovascular disease
  • c) "Antioxidant capacity," for anything
  • d) HbA1c, for microvascular diabetic complications
Answer **c.** The causal chain from a measured antioxidant capacity to a clinical outcome has never been established — and the beta-carotene trials are the demonstration of what happens when you assume it.

8. A trial measures fifteen biomarkers across four subgroups and reports two significant findings. The most important thing to note is:

  • a) Two findings is a strong result
  • b) Sixty comparisons were made, so a couple of hits is what chance alone predicts
  • c) The trial should be repeated with more biomarkers
  • d) Subgroup analysis strengthens a result
Answer **b.** The garden of forking paths. At the conventional threshold, roughly 1 in 20 comparisons hits by chance; sixty comparisons should produce about three false positives even if nothing is happening.

9. Mendelian randomization is useful in nutrition because:

  • a) It uses much larger samples than cohorts
  • b) Genetic variants are allocated essentially at random at conception and don't change with lifestyle, so they sidestep healthy-user bias
  • c) It measures diet more accurately than an FFQ
  • d) It eliminates the need for randomized trials
Answer **b.** It's a natural experiment. It has its own assumptions that can be violated, but confounding by lifestyle is not one of its main problems.

10. Which is the strongest signal that a body of evidence is trustworthy?

  • a) A very large number of studies
  • b) A very recent, very large study
  • c) Convergence across designs with non-overlapping weaknesses
  • d) A statistically significant result with a small p-value
Answer **c.** Cohorts suffer confounding; trials don't. Trials suffer short duration and unrepresentative samples; cohorts don't. When designs whose biases don't overlap agree, the agreement is hard to explain away.

11. In the constructed abstract in §2.9, the single most disqualifying feature for the headline "Improves Insulin Resistance in Overweight Adults" was:

  • a) The 8-week duration
  • b) That the insulin result was a post-hoc subgroup analysis
  • c) That participants were healthy
  • d) The 900 mg dose
Answer **b.** The trial was designed and powered for the whole group. Splitting it afterward by BMI, having already measured ten outcomes, is close to guaranteed to produce something. It is a hypothesis, not a finding. The other three compound the weakness but don't independently break the headline.

12. "Statistically significant" means:

  • a) The effect is large
  • b) The effect is clinically important
  • c) The result is probably not zero
  • d) The study was well designed
Answer **c.** In a study of 200,000 people, an effect too small to matter will still be highly significant. Always ask for the size, in units you can picture.

True / False

One-line justification each.

13. A prospective cohort of 90,000 people followed for 25 years provides stronger causal evidence than a randomized trial of 60 people over 12 weeks.

Answer **False** — for *causal* evidence. The cohort is better for hard outcomes over real timescales, but randomization is what breaks confounding, and no amount of size or duration substitutes for it. The two answer different questions, which is exactly why convergence between them matters so much.

14. Statistical adjustment for confounders eliminates confounding.

Answer **False.** It reduces the confounding captured by your measurements. Unmeasured confounders remain entirely; measured ones remain partially. Adjusting for a mediator or a collider can also make things worse.

15. The association between eating breakfast and lower body weight is a fabrication of the cereal industry.

Answer **False.** The observational association is real and replicated. What's unsupported is the *causal* interpretation — trials that hold calories comparable largely don't reproduce the effect. Industry promotion is real too, but the finding stands on its own; it's the inference that fails.

16. If a study's conclusion sentence is broader than its results section, that gap is where the press release lives.

Answer **True.** It's the most reliable single red flag in an abstract, and it's visible without any statistical training.

17. Because nutrition observational research is heavily confounded, dietary guidelines have no defensible evidence base.

Answer **False.** Guidelines rest on convergence — cohorts plus trials on intermediate outcomes plus mechanism plus consistency across populations — and are strongest for *dietary patterns* and *large effects*, which is exactly where confounding is least able to explain the signal.

18. Retraction of a major trial like PREDIMED shows that nutrition trials cannot be trusted.

Answer **False**, or at least not straightforwardly. The episode shows two things at once: flagship trials have flaws, *and* the correction mechanism functioned publicly at real reputational cost. A field that publishes its errors is more trustworthy than one that never reports any.

Short Answer

19. Explain, in three or four sentences, why "Is butter bad for you?" is a malformed question, and give two well-formed versions with different answers.

Answer Every dietary change is a substitution — you cannot remove butter and replace it with nothing — so the question has no answer until the comparator is specified. **Compared to olive oil, at 15% of calories, for LDL cholesterol, in adults, over six months:** worse. **Compared to refined-carbohydrate snacks, for cardiovascular events:** not clearly worse. Both answers are correct, which is why most public nutrition argument consists of two people answering different comparisons at each other.

20. A friend says: "Studies show people who take vitamin D have lower mortality, so I take vitamin D." Walk through the four objections you'd raise, in order.

Answer 1. **What design?** If observational, we're at rung 5 and confounding applies immediately. 2. **Healthy-user bias.** Supplement-takers are close to a definitional marker of health-consciousness — they exercise more, smoke less, see doctors more, adhere to medication. 3. **Reverse causation.** Low vitamin D status is *caused* by illness, obesity, indoor confinement, and low sun exposure. Sick people have low vitamin D partly because they're sick. 4. **What do the trials say?** Randomized vitamin D trials in generally well-nourished populations have not reproduced the dramatic mortality benefits the observational data suggested — which is the beta-carotene shape again. *(Chapter 13 and 16 handle the important exception: people who are genuinely deficient, where supplementation does help.)*

21. Name the five features of convincing evidence from §2.10 and explain why convergence is the most powerful of them.

Answer Convergence across designs; dose-response; survival under new methods (e.g. Mendelian randomization); being boring and old; and a field that corrects itself in public. Convergence is most powerful because the different designs have **non-overlapping weaknesses**. Cohorts are confounded but trials aren't; trials are short and unrepresentative but cohorts aren't; mechanism is untested in humans but the other two aren't. A single alternative explanation would have to defeat all of them simultaneously, and confounding — the usual culprit — cannot explain a trial result.

Applied Scenario

22. A headline reads:

"Daily Serving of Fermented Food Cuts Dementia Risk by 27%, Major Study Finds"

The article says researchers followed 24,000 adults over 60 for 14 years, recorded diet with an annual questionnaire, and adjusted for age, sex, education, smoking, alcohol, physical activity, and baseline cognitive score.

Run the full Claim Filter. Then state what you would actually do.

Answer **1. What kind of study, what rung?** Humans, not randomized, 14-year follow-up, real disease endpoint. **Prospective cohort — rung 5.** Reasonably good for what it is. **2. Compared to what?** Not stated. What did the low-fermented-food group eat instead? If they ate more ultra-processed food, that's one comparison; if they ate more fresh vegetables, quite another. Unanswerable from the article. **3. Could the arrow point the other way?** Yes, and seriously. Dementia has a prodromal phase lasting years to decades, during which appetite, food preferences, shopping, and cooking all change. People in early undiagnosed decline may eat fewer varied or fermented foods *because* of the disease process. The adjustment for baseline cognitive score helps but doesn't eliminate this; a lag analysis excluding the first 5–10 years would be more reassuring, and the article doesn't say whether one was done. **4. What else is true about people who eat fermented food daily?** In most Western populations: higher education, higher income, more varied diets, more cooking from scratch, more interest in health, more likely to be in the sort of social and cultural context that also protects cognition. They adjusted for education and smoking — coarsely — but not for diet quality overall, social connection, sleep, hearing loss, or occupational complexity, all of which are established dementia risk factors. **5. How big in absolute terms?** Unanswerable from the article — no baseline given. If lifetime dementia risk in this population is roughly 15%, a 27% relative reduction is about 4 percentage points, which would be substantial. If it's the annual incidence being reported, quite different. **The article has not given you enough to have an opinion**, which is itself the answer. **6. Who benefits?** Check for funding by a dairy, probiotic, or fermented-food body. Also note the press-release incentive and the journalist's. **What I'd actually do:** eat fermented foods if I like them, which many people do, because they're pleasant, they're generally part of a varied diet, and the downside risk is essentially zero. But that's a decision made on *cost and enjoyment*, not on this study — and I would not buy a probiotic supplement, take up a food I dislike, or tell anyone this prevents dementia. Chapter 27 has the fuller picture on fermented foods, and the honest verdict there is more modest than this headline.

Scoring

Score Reading
20–22 Excellent. You can read nutrition research better than most people who write about it.
16–19 Strong. Reread §2.3 (adjustment) and §2.6 (absolute risk) before continuing.
12–15 The core is there but shaky. Redo Exercises C1–C3 — the doing is what makes it stick.
Under 12 Reread this chapter before continuing. It is not optional; every remaining chapter assumes it.

Whatever you scored: make the Claim Filter card. Six questions, one index card. It is the single most durable thing this book will give you.