Measurement Error in Dietary Assessment

Why it matters

Diet is measured indirectly, by asking people. The resulting error is large enough to be the dominant consideration in reading almost any observational nutrition finding — and it does not push in one direction, which is what makes it hard to reason about.

The taxonomy that governs everything else

Error is random or systematic, within-person or between-person. The consequential split:

• Non-differential (unrelated to disease status) — biases relative risks toward 1, but the source scopes this: “toward the null value of one in univariate models without confounding”. With imperfectly measured covariates in the model the bias can run in any direction, so the true effect is at least as large as observed holds only in the unconfounded case.
• Differential (differs between cases and non-cases) — “Systematic differences in measurement error between these two groups… have serious consequences that are usually not amenable to correction. This… is the reason that typical case-control studies are treacherous.” [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 12, Types of Errors]

Also: systematic errors affecting everyone equally do not bias measures of association at all.

How much attenuation — the quantitative anchor

With validity γ (the regression of true on measured intake), observed RR ≈ true RR^γ. Willett’s Table 12-3:

true RR	γ=0.3	γ=0.5	γ=0.7
1.5	1.13	1.22	1.33
2.0	1.23	1.41	1.62
3.0	1.39	1.73	2.16

Verified against the rendered page (2026-07-26): all nine cells correct. The full table runs gamma 0.2-1.0 against true RR 1.5/2.0/3.0/5.0; the subset above is faithful. Two things the subset loses: at gamma=0.2 a true RR of 5.0 still observes as only 1.38, and Willett notes gamma can exceed 1.0 — “this can also result simply from a change in units” — so it is not a bounded validity score. [EXTRACTED (Willett - Nutritional Epidemiology 3e) Table 12-3 + following paragraph]

His read: “There is no obvious threshold of γ below which a measure of exposure is useless; however, for true relative risks of 1.5 or 2.0, epidemiologic effects are difficult to detect if γ is substantially less than 0.5.” And the sentence immediately after bounds any use of γ as a credibility threshold: “It will be apparent that the value of γ does not readily provide a generally interpretable measure of validity.” γ is scale-dependent, so is γ reported? is a disclosure test, not a pass/fail bar. [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 12, Correction of Relative Risks]

This is the bridge to any credibility threshold. A true RR of 2.0 measured at γ=0.3 shows up as 1.23. So a rule that dismisses observed relative risks below ~1.2 as within confounding range will also dismiss genuinely large effects that dietary measurement has flattened. Both readings are available for the same number, and only a validation study distinguishes them. [INFERRED (Willett - Nutritional Epidemiology 3e) — entailment from the attenuation table.]

Other magnitudes: a true OR of 2.67 observes as 1.71 at sensitivity 0.6 / specificity 0.9; energy under-reporting against doubly-labelled water runs 10-30%, worse in women and people with obesity; energy adjustment raised FFQ-vs-recall correlations from 0.30-0.57 up to 0.61-0.66 across four ethnic subgroups of men in the Multi-Ethnic Cohort.

Tests / indicators

• Was there a validation study, and is γ reported? Without it, attenuation cannot be bounded and a null is uninterpretable.
• Which direction could the error run? Non-differential → the true effect is at least as large as observed. Differential → the finding may be an artifact and no correction will fix it.
• Was energy handled? Failing to measure total energy can hide a real effect entirely: a saturated-fat/CHD study could find “an inverse or no association… even if high saturated fat composition of the diet truly caused coronary disease.” [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 11, Summary]
• Was correction applied to confounders too? Error in covariates “can distort relative risk estimates in any direction” — and in Willett’s Framingham example, correcting covariate error made an apparently significant BMI-CHD association non-significant.

The asymmetry worth internalising

Correcting for measurement error cannot manufacture an effect from a null, but can substantially strengthen a real one. In Willett’s worked table, an observed RR of 1.00 corrects to 1.00 with only a wider interval; an observed 1.50 corrects to 2.38. [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 12, Table 12-2]

So measurement error explains this null and measurement error explains this positive finding are not symmetric claims. The first is often plausible; the second requires differential error, which is a much stronger and more specific allegation. [INFERRED (Willett - Nutritional Epidemiology 3e) — entailment from the correction asymmetry.]

Red flags

• A null reported without any validation data or attenuation discussion.
• Correction methods applied on the assumption of purely random within-person error — Willett says this “is unlikely to be true for dietary questionnaires.”
• A “gold standard” sharing an error source with the test method (e.g. the same food-composition database), which biases the correction itself — though Willett’s own verdict is that “moderate correlations in these errors do not appear to result in serious bias if they are ignored” under regression calibration. Worth checking, not disqualifying.
• Calibration that smuggles in a covariate: in one published case, “calibrated energy intake” predicted cancer risk, but the associations “were entirely accounted for by BMI” — the calibration had contributed nothing.

Willett’s counter to the strong sceptical reading

He rejects using energy under-reporting to discredit diet studies generally: total energy “is rarely of direct interest”, the analytic goal is usually to remove energy variation, and under-reporting appears roughly composition-neutral — several studies found no relation between under-reporting and energy-adjusted macronutrient shares. “Moreover, systematic biases do not hinder the capacity to find important associations.” [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 11, Implications for Food-Frequency Questionnaire Data]

Recorded as his position, not as settled: this is the field’s founder defending the field’s core instrument, and the corpus holds no independent appraisal of the claim yet. AWAITS a source not authored by a principal of the method.

Limits

• “Each method requires assumptions that are rarely perfectly satisfied.” His conclusion from that is not to abstain: “Despite these limitations, careful use of these adjustment procedures will provide better estimates… than analyses that ignore the effects of measurement error altogether.” [EXTRACTED (Willett - Nutritional Epidemiology 3e) Ch 12, Summary]
• The corrected-value (regression-calibration) method does nothing for rank-based analyses: relative risks by quintile “are not affected by this method because the relative rankings of individuals are unchanged”. Not a limit on correction generally — Willett points to a separate method for ordinal categories (Rosner et al. 1989), which this page does not hold.
• Residual bias toward the null persists even after logistic correction, though Willett judges it “small and unimportant” only “with small or modest relative risks (less than about five) and a moderate degree of measurement error” — a condition that matters on a page arguing dietary error is large.

The cross-domain contrast — and what the parameter table refuses to let us claim

[2026-07-27, WHO PA Evidence Profiles ingest] The wiki now holds a domain where exposure is device-measured, which makes a comparison possible that was not available before. The comparison was run as a parameter table first, per the blocking rule, and the table cut the claim down substantially.

Parameter	Nutrition corpus	WHO PA 2020	Same quantity?
Exposure	self-reported dietary intake (FFQ / recall)	accelerometer counts, min/day by intensity	NO — different exposure entirely
Outcome	CVD / T2D / colorectal cancer incidence (SACN); body weight (Te Morenga)	all-cause mortality	NO — different outcomes
Reported curve shape	”as more dietary fibre was consumed a greater reduction in risk was observed” — monotone, no knee (SACN); not estimable at all (Te Morenga, P=0.392)	«A non-linear, dose-response association… (p<0.02)»; maxima named	YES — both are the shape of the reported curve
Certainty on the shape	not stated; dose-response often not computed	HIGH	YES — both GRADE-family certainty on the shape
Validation against an objective standard	a doubly-labelled-water sub-study (see the quoted span below the table)	none needed — the device is the measurement	NO — asymmetric by construction

The validation row, quoted in full because the truncated form drops the comparator that makes it mean anything:

«A doubly-labelled water sub-study carried out as part of the NDNS rolling programme (Bates et al., 2014) found that reported energy intake in adults aged 16-64 years was, on average, 34% lower than total energy expenditure (TEE) measured by doubly labelled water.» [EXTRACTED (SACN - Carbohydrates and Health 2015) chunk 02]

What the table permits. Only the two YES rows are comparable, and they are properties of the estimate, not of the biology: where exposure was device-measured, this corpus’s one instance recovered a non-linear shape at HIGH certainty; where exposure was self-reported, the shapes are monotone or not estimable.

What the table forbids, and this is the part that would have been written otherwise. The tempting claim — measurement error explains why nutrition curves flatten — is not supported by this comparison, because the two domains differ in exposure, outcome, biology, range of variation and confounding structure all at once. The measurement difference is confounded with everything else that distinguishes the domains. One device-measured domain and one self-reported domain is n=1 per arm.

What survives is a mechanism with independent support, not an inference from the contrast. The attenuation claim stands on its own footing — the DLW validation substudy, and Te Morenga’s own attribution of its failed dose-response to “a considerable degree of measurement error even when using validated methods”. The PA contrast is consistent with that mechanism and does not evidence it. [INFERRED (WHO - Physical Activity Web Annex Evidence Profiles 2020; SACN - Carbohydrates and Health 2015; Te Morenga - Dietary Sugars and Body Weight 2013) — entailment from the matched-parameter rows only; no source draws the comparison, and the causal reading is explicitly declined here]

Consequence for the telos. The every reduction pays default was induced from four self-reported-exposure sources. It should be scoped to that condition rather than asserted as a general property of decision-relevant curves — knees exist in this corpus, they are simply not in the half of it that is measured by asking people. [REVIEW: re-scope the falsified knees prior in CLAUDE.md to name the measurement condition — owning op is a telos review, not this weave] -> [[Physical Activity Dose and Mortality]]

Self-critique of the section above [2026-07-27]

Run because this page went from one source to four in a single edit, and the added material makes a cross-domain comparison — the laundering surface the gate exists for.

• Over-claim? The tempting causal reading (measurement error explains the flat nutrition curves) is stated and declined in the text, with the confound named. The surviving claim is restricted to the two YES rows. Held.
• Laundered independence? No type-E is claimed. WHO PA, SACN and Te Morenga are not presented as independent routes to one conclusion — they are matched on two comparable parameters and explicitly not matched on three. No E-token, correctly.
• sources: padding? Each of the three added sources contributes a distinct quoted value to the table (SACN: the DLW figure and the fibre gradient; Te Morenga: the failed dose-response and its stated cause; WHO PA: the non-linear shape and its certainty). Dual test passes.
• The residual risk, stated rather than resolved: the comparison is n=1 per arm. A second device-measured domain, or a self-reported domain that does show a knee, would move it either way. AWAITS either.

Two fibre measures that are not interchangeable [2026-07-27, SACN chunks 08/10]

SACN reports fibre in two incompatible units in the same report, and mixing them silently overstates any shortfall:

Measure	What it counts	SACN’s use
AOAC fibre (methods 985.29 / 991.43)	the broader analytic definition	the 30 g/day dietary reference value
NSP (non-starch polysaccharide)	a narrower fraction	the intake tables (Table 3.12: ~11-15 g/day by age and sex)

The trap, recorded because the wiki nearly fell into it: reading the NSP intake figure against the AOAC target would show a much larger gap than exists. SACN gives the correct like-for-like comparison itself — «current mean intakes would be 10-11g below the dietary reference value for men and 13g below for women» — stated in AOAC terms. [EXTRACTED (SACN - Carbohydrates and Health 2015) chunk 08]

This is a measurement-definition error rather than a measurement-precision one, and it is the cheaper of the two to make: no instrument noise is involved, only two names for different quantities. The parameter-table discipline exists for exactly this — the “same quantity?” column returns no. -> [[Whole Grains Refined Grains and Pulses]]

The variance structure underneath it all — and what it does to an SD-scaled increment [2026-07-28, Willett ch.3]

This page holds measurement error as the constraint that binds hardest in this domain. Willett’s chapter 3 supplies the structure producing it, and one consequence the corpus has been quoting numbers against without holding.

The mechanism is within-person day-to-day variation, and it is large. Willett’s own summary — note the qualifier, which is his and matters:

«The daily variation in nutrient intake among free-living subjects has consistently been found to be large, although the degree of variation differs greatly according to nutrient. The ratio of within-person to between-person variation is usually substantially greater when the nutrient is expressed in relation to energy intake (energy-adjusted). Measurements of dietary intake based on a single or small number of 24-hour recalls per subject can provide an accurate estimate of the mean for a group, but the standard deviation will be greatly overestimated.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 03, abstract]

Read the two halves separately, because they fail in opposite directions. A few-day assessment gets the group mean approximately right and the spread badly wrong — inflated, because the highest-measured people caught a high day and the lowest caught a low one. Bias in the mean and bias in the SD are different defects with different consequences, and a method can be fine for one and useless for the other.

Willett’s worked demonstration (194 Boston-area women, 28 days of records each): the 90th-to-10th percentile ratio for total fat fell from 3.0 (1 day) to 2.2 (1 week) to 1.9 (4 weeks); for vitamin A from 6.4 to 3.0 to 2.5. The single-day distribution is roughly half again as wide as the four-week one. [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 03]

And he names the real-world casualty:

«National nutrition surveys have traditionally employed only a single 24-hour recall per subject; thus, many major reports of the distributions of dietary intake are extremely misleading.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 03]

The consequence for a per-SD increment — this is the decision-relevant part

An effect expressed per standard deviation of intake inherits whatever inflated the SD. If the SD comes from few-day assessment it is too wide, so “one SD” is a larger exposure contrast than the label implies, and the relative risk attached to it describes a bigger change than the reader thinks.

This lands directly on a form the corpus holds. SACN - Carbohydrates and Health 2015 standardises its cohort increments to «one standard deviation of intake» (50 g/day total sugars, 20 g/day individual sugars), and Baseline Risk and the Relative-Absolute Split records that a per-SD relative risk does not transport without its population’s SD. Willett adds the prior problem: the SD may not be right for its own population either.

[INFERRED (Willett - Nutritional Epidemiology 3e; SACN - Carbohydrates and Health 2015) — the SD-inflation mechanism is Willett's and the per-SD increments are SACN's; joining them is this page's]

Stated as a conditional, because the wiki has not established SACN’s assessment method: if the SD behind an increment was estimated from a small number of days per subject, that increment is inflated. AWAITS a targeted read of SACN’s dietary-assessment methods — the question is how many days per subject the contributing cohorts used, and it is answerable from the report. Do not treat the inflation as established for SACN specifically; the mechanism is established, the application is not.

Table 3-6 — and a precision correction to Willett’s own summary

Within- and between-person coefficients of variation, 194 women, unadjusted and calorie-adjusted:

Nutrient	Within (unadj)	Between (unadj)	Ratio	Within (adj)	Between (adj)	Ratio
Energy (kcal)	27.0	19.3	1.40	—	—	—
Protein	32.9	16.4	2.01	25.0	14.0	1.79
Total fat	38.4	22.6	1.70	27.3	14.1	1.94
Monounsaturated fat	42.5	23.6	1.80	27.8	13.1	2.12
Polyunsaturated fat	64.2	28.3	2.27	47.3	20.2	2.34
Cholesterol	62.2	23.8	2.61	61.5	24.1	2.55
Carbohydrate	29.9	26.5	1.13	18.7	13.5	1.39

[EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 03, Table 3-6 — CVs are Willett’s; the ratio column is computed here]

Willett says the ratio is usually substantially greater when energy-adjusted. On his own table it rises for four of six nutrients and FALLS for protein (2.01 -> 1.79) and cholesterol (2.61 -> 2.55). Usually is the right word and the direction is not universal — worth stating, because a reader who takes it as a rule will mis-predict for exactly the two nutrients where it reverses.

Why the ratio is the quantity that matters: it sets how many days of measurement are needed. Willett gives the formula (Beaton) n = (Z_alpha x CV_w / D)^2, and works it: estimating one person’s cholesterol intake to within 20% of their true mean, 95% of the time, at CV_w = 62%, needs about 37 days of records. That is the scale of the problem — and it is why a single 24-hour recall cannot support an individual-level exposure claim, however large the study. [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 03]

Differential error, quantified — the underreporting gradient runs with adiposity [2026-07-28, Willett ch.4]

This page holds the non-differential/differential split as a distinction. Chapter 4 supplies the worked case, with numbers, and it is the one that matters most for the diet-obesity literature.

«Estimating energy expenditure using doubly-labeled water, the AMPM method of conducting 24-hour recalls, resulted in zero, 14%, and 20% underreporting of energy intake among lean, overweight, and obese men, respectively (Moshfegh et al., 2008). Among women, the same percentages were 6%, 15%, and 21% underreporting of energy intake» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 04]

And at the extreme, from food records:

«The extent of underreporting of total energy intake varied among individuals. Obese individuals tended to underreport as much as 47% (Lichtman et al., 1992) and 53% (Lansky and Brownell, 1982), substantially more than lean individuals …» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 04 — the sentence continues into a citation list only; elision marked]

Why this is the dangerous kind of error and not just noise. Non-differential error attenuates toward the null — it makes a real effect look smaller. Error that varies with adiposity is differential with respect to a variable that is simultaneously an outcome, an exposure proxy and a confounder in most of this literature. It can therefore create, erase or reverse an association rather than merely shrink it. A measured association between any reported intake and obesity is partly a measurement artifact of exactly the size shown above, and the direction is not predictable from first principles. [INFERRED (Willett - Nutritional Epidemiology 3e) — the underreporting gradient is Willett's; the consequence for diet-obesity inference is this page's]

Consistency check against what the wiki already holds, and the full SACN sentence adds a stratum the corpus had been dropping:

«A doubly-labelled water sub-study carried out as part of the NDNS rolling programme (Bates et al., 2014) found that reported energy intake in adults aged 16-64 years was, on average, 34% lower than total energy expenditure (TEE) measured by doubly labelled water. The difference for other age groups was similar except for children aged 4-10 years where reported energy intake was 12% lower than TEE.» [EXTRACTED (SACN - Carbohydrates and Health 2015) chunk 02]

The 34% figure is adults 16-64, and children aged 4-10 sit at 12% — a stratum the wiki’s prior citations of this number omitted. So underreporting has an age gradient as well as an adiposity gradient, and quoting 34% as the misreporting figure overstates it for children by roughly threefold. Willett’s decomposition and SACN’s are the same phenomenon cut on two different variables; two sources, one mechanism, and the decompositions are what make it actionable.

The flat slope syndrome — a mechanism that flattens dose-response curves

«In general, those who consumed considerably less than the average were more likely to overreport intake, while those who ate more than the average tended to underreport, sometimes referred to as the “flat slope syndrome”» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 04]

This compresses the measured exposure range toward the mean from both ends at once, which is mechanically different from ordinary noise: it does not merely blur the ranking, it shortens the x-axis. A true gradient estimated across a compressed range yields a shallower measured slope, and the compression is worst exactly where dose-response questions are decided — in the tails.

Recorded as bearing on the corpus’s dose-response prior, NOT scored against it (adjudication is out of scope for an ingest). The wiki’s four examined curves were monotone over the studied range or not estimable. Flat-slope compression is a mechanism that could hide a knee — by pulling extreme intakes toward the middle, it removes the very observations that would locate one. It cannot manufacture a knee, only erase one, so it bears asymmetrically: it is a reason to distrust a measured absence of curvature, not a reason to distrust a measured curve. Handle for whoever scores it: the dose-response [PRIOR] in CLAUDE.md.

Records change what people eat — reactivity, and it biases TRIALS specifically

«compared with an unannounced telephone recall, the food record overestimated the difference in fat intake between the study groups by approximately 40% at 6 months (p = 0.08) and by 25% at 12 months (p = 0.62) (Buzzard et al., 1996)» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 04]

Willett is explicit that neither figure reached significance and frames them as «consistent in direction and magnitude at the two time periods, providing cause for concern» — so this is a flagged concern, not an established effect size, and must not be quoted as one.

The structural point survives the imprecision: in a diet trial the intervention arm has more reason to record carefully, so the instrument measuring adherence is itself affected by the intervention. That inflates the apparent exposure contrast, which inflates the apparent effect. -> Framing a Decision Question

The instrument that generates the epidemiology is the wrong one for checking a threshold

Chapter 4 states the division of labour plainly:

«For most epidemiologic investigations of relationships between dietary intake and other health-related factors, relative rankings of food and nutrient intakes are adequate for determination of correlations or relative risks. However, in some situations, such as when comparing nutrient intakes with specific dietary recommendations, estimates of the absolute magnitude of energy and macronutrient intakes may be required. In such cases, records or recalls are generally the methods of choice.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 04]

Read that against how the corpus’s thresholds are actually used. Guidance issues absolute cutpoints — «below 10 E%», «30 g/day», «<5% of total energy» — and the cohort evidence behind them comes overwhelmingly from food-frequency questionnaires, which this chapter positions as adequate for ranking, not for absolute magnitude. So “is my intake under the threshold?” is a question the dominant instrument is not the method of choice for, even where the threshold itself is sound.

This is a limit on APPLICATION, not on the evidence. The FFQ-based ranking that established a gradient can be perfectly adequate for that job while being the wrong tool for auditing one person against a cutpoint. The two failures the wiki should keep separate: a threshold can be poorly located (a claim about the evidence) or well located and unmeasurable in practice (a claim about the instrument). Chapter 4 speaks only to the second. [INFERRED (Willett - Nutritional Epidemiology 3e) — the ranking/magnitude division is Willett's; the application to guidance thresholds is this page's]

A validity coefficient is itself biased — correlated error inflates it [2026-07-28, Willett ch.6]

The attenuation arithmetic on this page takes a validity coefficient as its input. Chapter 6 shows that number is not a clean measurement either, and the bias has a known direction: upward.

The mechanism is a timing artifact. A validation study compares a questionnaire against a reference method — a diet record, or a biomarker. In principle their errors are independent, because the techniques differ. In practice they are not, when both are collected at the same moment:

«if we consider that true intake is the average over an extended period, for example a year, but the biomarker and the recall/record both reflect relatively short-term intake and are collected close in time, then errors due to random variation in intake over a year will similarly affect both, and create correlated error. For example, blood levels of vitamin C reflect intake over a few days, and are likely to vary over a year due to seasonal availability of foods and other factors, so if a diet record and blood level of vitamin C are collected close in time, the errors in both will be correlated and the correlation between the methods will overstate validity.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 07]

The worked case is the part that makes it usable, and it needs both halves. Collected at the same time, the diet record beat the questionnaire against blood vitamin C. Willett’s very next sentence reverses the reading: «when using a blood measure of vitamin C collected approximately a year before the diet record and food frequency questionnaire assessments, the two measures of intake were similarly correlated with the biomarker.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 07]

So the diet record’s apparent superiority was an artifact of when the blood was drawn, not a property of the instrument. Move the biomarker a year away and the two methods perform alike.

Three consequences, in decreasing certainty:

• A published validity coefficient inherits the design that produced it. Ask when the reference measure was collected relative to the questionnaire before treating the coefficient as a property of the instrument.
• The bias runs one way. Correlated error can only inflate agreement between two methods, so validity coefficients from same-visit designs are ceilings, not estimates.
• Therefore the attenuation on this page is a LOWER bound on the real attenuation. If true validity is below the published coefficient, the true effect being attenuated is larger than the correction implies. This does not change the direction of any correction — it means the corrections understate. [INFERRED (Willett - Nutritional Epidemiology 3e) — the correlated-error mechanism and the vitamin C case are Willett's; the consequence for this page's attenuation figures is this page's]

What this does NOT establish. Willett describes the artifact and one demonstration (a conference presentation plus Bingham 2008); he does not quantify how much of the published validation literature is affected. The size of the problem across the field is not established here — only that the mechanism is real, has a direction, and is «frequently not considered».

Differential error by DISEASE STATUS — and a third route to a spurious U-shape [2026-07-28, Willett ch.7]

The adiposity gradient above is one differential; chapter 7 supplies the other one, and it is the mechanism that separates two study designs.

The measured case. In a Hawaii multiethnic cohort, recall bias was disease-specific:

«In a multiethnic cohort in Hawaii (Wilkens et al., 1992), retrospectively reported intakes of total and saturated fat were overreported by colon cancer cases compared to controls, but no bias was seen for breast and prostate cancers. This difference for cancers is notable because, for colon cancer in particular, strong positive associations with total energy and saturated fat have been seen with remarkable consistency in case-control studies, whereas in prospective studies associations have tended to be inverse, sometimes significantly so» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 08]

Read the whole sentence, because the «no bias was seen for breast and prostate» half is what makes it evidence rather than assertion. The bias appeared exactly where the design-class divergence appears and was absent where it does not. That is a matched prediction, not a post-hoc story — and Willett’s immediately following sentence supplies the counter-case (Lindsted and Kuzma found no biased reporting for overall food frequency), which is why his own summary is hedged: biased findings occur «in some circumstances, but not in all studies».

The starker demonstration is Zutphen (Bloemberg 1986, 615 Dutch men): analysed prospectively, men who later became CHD cases had consumed 229 kcal/day less than those who stayed free of it; analysed from diet recalled 15 years later, no difference between cases and controls at all. Recall erased a real prospective association. Willett bounds his own example — the diagnosis-to- interview gap «could have been up to 15 years, which would rarely be the case in the usual case-control study» — so it demonstrates the mechanism at an extreme, not the typical magnitude. [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 08]

A U-shape can be manufactured by unequal PRECISION, with no bias at all

«Even if there is no systematic bias, differences in random error—that is, precision in reporting past diet—can distort the shape of dose-response relationships. For example, greater random error for cases than controls would create a U-shaped relationship if there was no true overall relationship.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 08]

This is a distinct mechanism from the two usually named, and the distinction is the point:

Route to a spurious U/J	Requires	Detectable by
Reverse causation	early disease changing intake	lag analysis, excluding early follow-up
Confounding by frailty / sick-quitter	a third variable tracking both arms	covariate adjustment, restriction
Unequal precision between groups	no bias and no confounder — only different error variance	neither of the above

The third survives the standard checks, because it is not bias and not confounding. It needs a comparison of reporting precision between groups, which validation substudies rarely stratify by case status.

Recorded as bearing on the corpus’s U/J [PRIOR — test me], NOT scored — adjudication is out of ingest scope, and the handle is the U/J prior in CLAUDE.md. What this pass establishes is narrower than a verdict: the prior names reverse causation and confounding as the artifact routes, and a third route exists that neither of its named checks would catch. Whether U-shapes in the wiki’s actual holdings arise this way is untested — the corpus has one J-shape mention in total.

The design-class consequence, and an independent warrant for a rule the corpus already holds

Willett’s verdict on the pair of designs:

«we now have substantial evidence from diet and disease relationships that have been extensively examined in both case-control and cohort studies that the results of case-control studies can frequently be seriously distorted by a combination of recall and selection bias.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 08]

SACN - Carbohydrates and Health 2015 excluded case-control studies by design — «case-control, cross-sectional and ecological studies were not considered» — on stated concerns about their limitations. Willett supplies the mechanism and the measured demonstration behind that rule.

This is a genuine independence worth naming carefully. SACN is a guidance body stating an inclusion rule; Willett is a methods textbook documenting why the rule is warranted, from cohort-nested recall- bias substudies. Different fields, different methods, neither citing the other for this claim — so the convergence is not self-stated. [E-independent] [INFERRED (Willett - Nutritional Epidemiology 3e; SACN - Carbohydrates and Health 2015) — the rule is SACN's and the mechanism is Willett's; pairing them is this page's]

Recall horizons, held for transportability judgments — quoted whole, because the qualifier is in the same sentence as the reassuring half:

«Studies conducted to date suggest that diet may be recalled with acceptable levels of error up to approximately ten years; beyond this period, greater uncertainty exists. Similarly, diet during adolescence (high school) can be adequately recalled up to about thirty years later, but validity of recall much later in life appears to be low for some, but perhaps not all, aspects of diet.»

Recall bias is worst with longer intervals and more advanced disease, so it is most dangerous for cancers typically diagnosed late. [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 08]

Not all biomarkers measure intake — the three-class taxonomy [2026-07-28, Willett ch.8]

Everything above turns on comparing a self-report against something better. Chapter 8 (Van Dam & Hunter) says which biomarkers can play that role, and most cannot.

First, the disambiguation the word biomarker hides. A biochemical indicator can be used as a measure of intake, of nutrient status, or as a predictor of disease — three different jobs. Chapter 8’s scope is the first, and it is explicit that the others are distinct. The reason they come apart:

«nutrient intake is just one determinant of nutrient status because the concentration of a nutrient in blood or tissues is typically influenced by variation in the absorption, transport and distribution, metabolism, and excretion of the nutrient.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 09]

Those intervening processes are themselves affected by genetics, other dietary intakes, smoking, physical activity and disease — so a blood concentration is a joint function of intake and individual biology. A biomarker can therefore be an excellent measure of status and a poor measure of intake at the same time, which is the confusion the single word invites. -> Surrogate Outcomes, Is the Food Category Doing Any Work

The three classes, and why only one can serve as a reference standard

«A recovery biomarker is a biomarker for which a quantitative relationship exists between values for the biomarker and dietary intake in a specific time period. For example, it is known that about 80% of nitrogen intake in the form of protein is “recovered” as nitrogen in urine, with limited variation between individuals (Bingham, 2002).» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 09]

«In addition to urinary nitrogen, urinary potassium, urinary sodium, and energy intake estimated using the doubly labeled water technique are recovery biomarkers. The 24-hour urinary excretion of sucrose plus fructose is not considered to be a recovery biomarker because only a small fraction is recovered in the urine, but it has been labeled a predictive biomarker to distinguish it from concentration biomarkers because correlations with intakes are relatively high, and biomarker values may be used to estimate absolute intakes.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 09]

Class	Relationship to intake	Can it give an ABSOLUTE intake?	Members named
Recovery	quantitative, a known fraction recovered over a defined period	yes	urinary nitrogen, urinary potassium, urinary sodium, doubly-labelled-water energy
Predictive	high correlation, small recovered fraction	yes, with modelling	24-h urinary sucrose + fructose
Concentration	correlated only	no	blood/tissue nutrient concentrations generally

This closes a loop that has been open across three chunks. The underreporting figures on this page — SACN’s 34%, Willett’s 0/14/20% adiposity gradient — are all measured against doubly-labelled water, i.e. against a recovery biomarker. The reason those studies could establish a magnitude rather than merely a correlation is that the reference method was in the one class that yields absolute intake. Where no recovery biomarker exists, misreporting can be detected but not sized. [INFERRED (Willett - Nutritional Epidemiology 3e; SACN - Carbohydrates and Health 2015) — the taxonomy is Van Dam & Hunter's and the substudy is SACN's; the observation that the validation literature's reach is bounded by the recovery class is this page's]

The consequence that is decision-relevant, and it favours one exposure the corpus holds

Sodium is measured by a recovery biomarker; almost nothing else in the corpus is. Urinary sodium gives absolute 24-hour intake, so the sodium literature rests on an exposure measure of a different class from the FFQ-based evidence behind sugars, fibre, saturated fat and grains. -> Sodium Intake and Blood Pressure

Two guards against over-reading that.

• A better exposure measure is not a better study. It removes one error source; design, confounding, and outcome measurement are untouched. The sodium page’s own finding — high certainty on blood pressure, very low on hard outcomes — is unaffected by this and is not improved by it.
• 24-hour urine has its own failure mode: it measures one day, and this page’s chunk-03 material shows a single day is a poor estimate of habitual intake. A recovery biomarker fixes the calibration problem, not the day-to-day variation problem — the two are independent, and a single 24-hour urine inherits the second in full. [INFERRED (Willett - Nutritional Epidemiology 3e) — the two-problem decomposition is this page's]

A biomarker is not automatically the safer instrument [chunk 10]

The taxonomy above says which biomarkers track intake. Chapter 8 adds a warning that runs against the intuition motivating the whole exercise:

«although biomarkers appear “objective,” the use of fatty acid biomarkers can introduce bias and confounding that is not present for studies of fatty acid intakes.» [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunk 10]

Two distinct reasons a biomarker can be worse than the questionnaire it replaces:

• Endogenous synthesis. «Biomarkers generally perform poorly for fatty acids that can be produced endogenously, including even-chained saturated and monounsaturated fatty acids» — the concentration is intake plus de novo production, and the terms cannot be separated. -> Saturated Fat Intake and Replacement
• Homeostatic regulation. Where a concentration is defended by physiology it stops tracking intake at all: 1,25(OH)2D «is not a good reflection of vitamin D nutritional status because concentrations are tightly regulated», while 25(OH)D serves. Regulation is the enemy of a biomarker of intake — and the converse holds, which is why blood carotenoids work: they «are not closely regulated by homeostatic mechanisms». [EXTRACTED (Willett - Nutritional Epidemiology 3e) chunks 10-11]

The generalisation worth carrying: a biomarker measures intake well exactly to the extent the body does not care about the quantity. Anything physiologically important enough to be regulated is, for that reason, badly measured in blood. This is a structural trade, not a technical limitation to be engineered away. [INFERRED (Willett - Nutritional Epidemiology 3e) — the inverse relation between homeostatic regulation and biomarker utility is this page's compression of Van Dam & Hunter's examples]