Nucleus of the food-addiction cluster. “Food addiction” is a behavioural / reward
construct: the claim that some foods trigger addiction-like loss of control, measured by the
Yale Food Addiction Scale (YFAS) — a 25-item self-report scale mapping the DSM-IV
substance-dependence criteria onto eating (Gearhardt 2009). This page opens the construct as a
domain node, on three sources: Schulte 2015 (moderate-tier attribute survey), Gordon 2018
(gold-tier construct-validity SR), and Pursey 2014 (gold-tier prevalence SR/MA, added 2026-08-30).
The construct is highly controversial, and the page holds two separate evidence states: its
construct validity is gold-SR-supported-but-controversial (Gordon), while its bearing on any
patient-important outcome stays insufficient-evidence — neither benefit nor harm on what to
eat. The three sources answer three different descriptive/appraisal questions and none touches
the outcome question: Schulte the attribute question (which food properties predict a high
addictive-like-eating score), Gordon the validity question (does the construct hold against
substance-use-disorder criteria), Pursey the prevalence question (how common is a YFAS diagnosis,
and in whom).
(inferred from Schulte et al., 2015)
What the YFAS measures — a self-report symptom count, not a diagnosis of a disease
- YFAS operationalizes addictive-like eating on the DSM-IV substance-dependence criteria (loss of control, continued use despite harm, inability to cut down, tolerance, withdrawal). A “food addiction” diagnosis = >=3 symptoms + clinically significant impairment/distress. In Schulte’s samples this was 6.7% (Study 1 undergrads) and 10.2% (Study 2 MTurk adults) (Schulte et al., 2015). YFAS symptom count correlated with BMI (r ~ 0.21) but not gender.
- The instrument is the exposure’s weak point (streetlight guard). YFAS is a self-report perception of one’s own eating, not measured intake, not observed behaviour, and not a measured biological addictive response. Schulte’s own limitation: «the current findings are limited to participants’ reports of whether certain foods are perceived to be associated with addictive-like eating behavior» (Schulte et al., 2015), and «we did not collect observational data to assess the frequency that these foods were consumed» (Schulte et al., 2015). So a “high YFAS score for chocolate” is a reported perception, several inferential steps from a demonstrated addiction.
Which food attributes predict a high score — processing, fat, glycemic load
Schulte’s finding (35 foods varying in processing/fat/GL; Study 1 forced-choice ranking n=120, Study 2 hierarchical linear model of 1-7 problem-ratings n=384): not all foods are equally implicated, and three attributes predict the addictive-like-eating rating, all large effects. The coefficients below are the ones Schulte reports for each attribute in the model it was retained in — the d’s are not all from a single joint model (the collinearity structure below forces separate models), so read down the column, not across it.
| Attribute | Effect size d (in the model it was retained in) | Coefficient | Note |
|---|---|---|---|
| Processing (highly-processed vs not) | d = 1.444, p < 0.001 | γ10 = 0.653 | the single most influential attribute; top-10 ranked foods all highly processed |
| Fat (grams) | d = 1.581, p < 0.001 | γ10 = 0.025 | fat’s coefficient in the FINAL model (co-modeled with GL); predicts regardless of who rates |
| Glycemic load | d = 0.923, p < 0.001 | γ20 = 0.021 | in the FINAL model (with fat); larger than sugar (0.814) or net-carb (0.657) — the rate facet, not just the quantity |
- Two collinearity constraints shape which attributes share a model — so no d in the table above is directly comparable to another. (i) Processing and GL correlate r = 0.756 (p < 0.01), so processing is fitted in its own model and GL in a separate one; (ii) fat and sodium correlate r = .623 (p < 0.001), barring them from one model. The three tabled attributes are not independent levers — they are facets of the same highly processed profile. (Schulte et al., 2015)
- Fat was chosen over sodium for the final model — a separate screening step, its own numbers. Assessed independently (the r = .623 collinearity barred co-modeling), fat’s standalone effect beat sodium’s, so fat — not sodium — was carried into the fat+GL model: «We assessed fat and sodium independently, and though both were significant level-one predictors, we determined that fat had a larger effect size than sodium (fat: d = 1.853, p < 0.001; sodium: d = 1.223, p < 0.001). Thus, fat was utilized in the second model.» So fat carries two source d’s for two model contexts — d = 1.853 assessed alone against sodium, d = 1.581 once co-modeled with GL — and they do not conflict. (Schulte et al., 2015)
- GL beats sugar and net-carb — so it is «not just the quantity of refined carbohydrates … but the rapid speed in which they are absorbed» that predicts the rating, the paper’s pharmacokinetic (dose + rate-of-absorption) analogy to drugs of abuse. (Schulte et al., 2015)
- Fat vs GL differ in individual-difference profile (the one within-source refinement). GL’s effect is moderated by YFAS symptom count (higher-symptom individuals report more problems with high-GL foods; γ21 = 0.003, d = 0.297); the FAT effect is moderated by no individual difference. Schulte reads this as fat tracking a general tendency to overeat rather than an addiction-specific process — consistent with the animal literature (opiate-like withdrawal on sucrose removal but not fat). So within the construct, sugar/GL is the more “addiction-like” facet and fat the more “general overconsumption” facet. (Schulte et al., 2015)
The paper’s engineering definition of the implicated foods: «foods that have been designed to be particularly rewarding through the addition of fat and/or refined carbohydrates» (Schulte et al., 2015) — and it proposes the construct might be «more appropriately titled ‘highly processed food addiction’» (Schulte et al., 2015), i.e. narrowing scope, not validating the addiction claim.
Distinct from Hyper-Palatable Foods — same neighbourhood, different construct [type-B]
The food-addiction construct and Hyper-Palatable Foods point at overlapping foods (fat + refined carbohydrate, engineered reward) but are different objects, and a claim about one does not transfer to the other:
- Hyper-Palatable Foods is a nutrient-COMPOSITION definition (Fazzino’s fat/sugar/sodium/carb threshold combinations, read off a nutrition label) — it is agnostic about behaviour; it says nothing about whether anyone loses control.
- Food addiction is a behavioural / reward construct — a person’s YFAS-measured loss of control — measured by a self-report scale, agnostic about the food’s exact composition (it uses processing/fat/GL as predictors, not a defining threshold).
- Schulte’s contribution is REWARD-MECHANISM evidence about a channel HPF’s composition definition leaves open (processing + fat + GL predict addictive-like eating), but the two must stay separate: a food being HPF by composition does not make it “addictive”, and a food scoring high on YFAS is not thereby shown to meet an HPF threshold. (inferred from Schulte et al., 2015)
The construct-validity appraisal — a gold SR that generally supports it, on a weak bar [gold SR, construct-validity only]
Gordon 2018 is the first systematic review of food addiction not limited to YFAS- or body-weight-based definitions — «To our knowledge, this is the first systematic review on food addiction that was not limited to definitions based on the YFAS or body weight status» (Gordon et al., 2018). It organizes the evidence against eight addiction characteristics — «(a) neurobiological changes, (b) preoccupation with the substance, (c) impaired control, (d) social impairments, (e) risky use, (f) tolerance/withdrawal, (g) chronicity of the condition, and (h) relapse» (Gordon et al., 2018) — and concludes the literature «generally support[s] the validity of food addiction as a diagnostic construct, particularly as it relates to foods high in added sweeteners and refined ingredients» (Gordon et al., 2018). It opens the abstract by naming its own subject «a highly controversial subject» (Gordon et al., 2018).
Read the bar it clears, not just the verdict — the support is vote-counted, lopsided, and
animal-heavy. The SR counts how many studies back each criterion, not a pooled effect on any
one — «each primary criterion having support from at least one study … though some sub-criteria
have not yet been studied»
(Gordon et al., 2018). The distribution
is extremely uneven: «Brain reward dysfunction and impaired control were supported by the largest
number of studies (n = 21 and n = 12, respectively); whereas risky use was supported by the fewest
(n = 1)»
(Gordon et al., 2018). So a criterion
supported can rest on a single study, and the construct’s weight
sits almost entirely on the two easiest-to-measure neurobiological/control facets. The pooled
magnitude that would size this as a lever is exactly what an SR of this heterogeneous
animal+human literature structurally cannot yet produce — a G (needs aggregation) gap, not a
finding.
The SR self-flags the biases that most threaten this exact verdict. «Limitations include that our search was limited to two electronic databases and only included studies published in English, and that animal studies limit generalizability to humans … The study question may also have produced biased results, as researchers interested in evaluating the validity of the food addiction construct may be more inclined (consciously or not) to observe and report confirmatory results» (Gordon et al., 2018). A confirmation-bias self-warning on a construct-validity SR is a strong reason to hold the generally supports verdict at arm’s length under symmetric standards — the direction of such a bias is not random, it inflates the very conclusion drawn.
Substance over behavior; and food-specificity that resists a clean processing story. Gordon reads the symptom profile as fitting substance-use disorder better than behavioral addiction — «symptoms appear to better fit criteria for substance use disorder than behavioral addiction» (Gordon et al., 2018), the «substance (highly-palatable food) component may be more salient … than the behavior (eating)» (Gordon et al., 2018). On which foods: «data from human studies suggest that the combination of sweet and fat is more commonly associated with addictive symptoms than sugar alone» (strong sugar support is animal-only) (Gordon et al., 2018). But the SR itself flags an anomaly against a tidy processing rule, citing Schulte’s own data: «nuts (typically considered a whole food, without added sugars) were rated more addictive on average than granola bars (typically processed, with added sugars and fats)» (Gordon et al., 2018).
NOT independent-E — Gordon’s evidence base includes Schulte. The nuts>granola anomaly above is
Gordon quoting Schulte [13] — the same paper this page’s attribute section rests on — so Gordon and
Schulte are not independent backings converging on the construct; they share the Gearhardt/YFAS
lineage and, literally, the same study. Their agreement raises coverage, not confidence: it is
shared-lineage coherence, not type-E robustness. No [E-independent] is claimed.
(inferred from Gordon et al., 2018)
Two separate evidence states — the crux. Gordon moves the construct-validity question, and only that. (inferred from Gordon et al., 2018)
- Construct validity: gold-SR-supported-but-controversial. A gold-tier SR now concludes the construct generally holds against SUD criteria — a real upgrade from one survey assumes the frame. But it holds on a weak base (vote-counting, one-study criteria, animal-heavy, author-flagged confirmation bias, non-independent from Schulte), and the SR calls its own subject controversial. So: plausibly a valid construct per SR, on contested and bias-exposed evidence — not a settled fact.
- Patient-important outcomes: still insufficient-evidence — unchanged. Gordon appraises construct validity, not what the construct does to anyone. It carries no weight-gain, T2D, or mortality pathway, and no addiction-specific outcome distinct from ordinary overconsumption. A gold SR on the construct does not license any decision-change about what to eat that was not already reachable from the ordinary highly processed foods drive overconsumption reasoning. The outcome G-gap below stays fully open.
Prevalence — a pooled 19.9%, but read the sample it is drawn from [gold SR/MA, prevalence only]
Pursey 2014 is the prevalence SR/MA: 28 articles describing 25 studies, 196,211 participants — «A total of 196,211 participants were examined across reviewed studies ranging from one to 134,175 participants. Participants were predominantly female, with six studies investigating females exclusively … and an additional nine studies investigating a population with >70% female participants» (Pursey et al., 2014). The headline pooled figure and its subgroup pattern:
- Pooled YFAS food-addiction diagnosis prevalence: 19.9% (weighted mean across adult population samples). Doubled in overweight/obese vs healthy-BMI samples, roughly doubled in females vs males, and higher in the older stratum: «Using meta-analysis, the weighted mean prevalence of FA diagnosis in adult population samples was 19.9%. Meta-analysis indicated that FA prevalence was double that in overweight/obese population samples compared to those of a healthy BMI (24.9% and 11.1% respectively) and in females compared to males (12.2% and 6.4% respectively). FA prevalence was also higher in adults older than 35 years compared to adults younger than 35 years (22.2% and 17.0% respectively).» (Pursey et al., 2014)
| Subgroup | YFAS FA prevalence | Contrast |
|---|---|---|
| All adult samples (pooled) | 19.9% | the headline weighted mean |
| Overweight/obese | 24.9% | vs healthy BMI 11.1% (~2x) |
| Female | 12.2% | vs male 6.4% (~2x) |
| Age > 35y | 22.2% | vs < 35y 17.0% |
| Clinical (symptom count) | 4.0 symptoms | vs non-clinical 1.7 (>2x) |
The 19.9% is NOT a general-population prevalence — the source says so itself (transportability caveat, load-bearing). The pooled pool is predominantly overweight/obese, female, and clinically-recruited, and Pursey draws the inflation conclusion directly: «When meta-analyzed by clinical status, clinical populations endorsed more than double the number of symptoms compared to non-clinical populations (4.0 and 1.7 symptoms respectively). However, it should be noted that the population samples in the included studies were predominantly comprised of overweight/obese females recruited from clinical settings. Hence, the prevalence of YFAS FA diagnosis and the average symptom scores are likely higher compared to a nationally representative general population sample due to the characteristics of included participants.» (Pursey et al., 2014) So the subgroup gradient (OW/OB, female, clinical -> higher) is also the composition of the pooled pool — the 19.9% sits toward the high end of what a general population would show, not at its center.
- Prevalence is sample-dependent — the numbers do not converge on one rate, and that is the point . Schulte’s own YFAS diagnosis rates on this very page — 6.7% (undergrads) and 10.2% (MTurk adults) — sit well below Pursey’s clinically-weighted 19.9%, and below even Pursey’s healthy-BMI 11.1%. A YFAS prevalence is a property of who was sampled, not a fixed population fact; quoting 19.9% bare would launder a clinically-enriched figure into a population claim. The honest read is a range whose level tracks BMI, sex, age and recruitment setting. (inferred from Pursey et al., 2014)
- Prevalence is descriptive — it moves NEITHER evidence state. How common a YFAS diagnosis is says nothing about whether the construct is valid (Gordon’s question) and nothing about whether it changes any patient-important outcome. A high prevalence of a self-report symptom count is not evidence that the underlying construct is real or that it harms anyone — the outcome state below stays insufficient-evidence, untouched. (inferred from Pursey et al., 2014)
- NOT independent-E. Pursey 2014 shares the Gearhardt/YFAS lineage with Schulte and Gordon (Gordon
2018 cites Pursey as ref [34]); all three rest on the same instrument. Pursey adds a different
facet (prevalence) to the same construct — a banked-C cluster extension, not a type-E
independent convergence. No
[E-independent]is claimed. (inferred from Pursey et al., 2014)
Saturation. With Pursey the construct-description cell — attribute prediction (Schulte), construct validity (Gordon), prevalence (Pursey) — is saturated across three mutually non-independent YFAS-lineage sources: a fourth YFAS-based descriptive source would add neither an independent perspective nor a quality upgrade. What stays unsaturated / never opened is the outcome cell (a food-addiction/HPF-exposure -> patient-important-outcome SR), which no held source touches — see the G-gaps below.
Evidence state — contested construct, no outcome evidence [insufficient evidence / contested]
Schulte 2015 is a single cross-sectional survey with self-report YFAS; it carries the attribute finding, not the construct’s validity and not any patient-important outcome. Gordon 2018 supplies the construct-validity layer and Pursey 2014 the prevalence layer (both added 2026-08-30) — see the sections above — but both leave the outcome state exactly where it was: insufficient-evidence. Validity and prevalence are descriptive/appraisal facets of the construct; neither is a patient-important outcome, and neither licenses a decision-change about what to eat that was not already reachable from the ordinary highly processed foods drive overconsumption reasoning.
- The whole “person x substance” addiction frame is assumed by the paper, not tested — Schulte studies which attributes predict a YFAS score given the frame. Whether “food addiction” is a valid distinct construct (vs a re-description of binge eating, vs a label for high palatability + overconsumption) is a live open question, not settled here.
- The authors are consistently hedged: «preliminary evidence», «may share characteristics with drugs of abuse», “food addiction” in scare-quotes throughout. The finding is that highly processed foods «appear to be particularly associated with “food addiction”» (Schulte et al., 2015) — an association with a self-report construct, NOT a demonstration that these foods ARE addictive.
- Symmetric-standards / fad-bar: an engineered to be addictive claim gets the same bar as any fad — being controversial, discussed, or intuitively compelling is not a pass. On present evidence the construct is insufficient-evidence, and no decision (what to eat, for whom) turns on it that is not already reachable from the ordinary highly processed foods drive overconsumption reasoning. The animal-model backdrop (sugar bingeing -> tolerance/withdrawal) is rodent, not human — directional at most (transportability caveat).
(inferred from Schulte et al., 2015)
Gaps and open questions
- Construct-validity is adjudicated by a gold SR (Gordon 2018) and prevalence by a gold SR/MA (Pursey 2014, landed 2026-08-30) — see the construct-validity and prevalence sections above. Gordon appraised 52 studies across 35 articles against eight addiction characteristics and generally supports the construct, on the weak/animal-heavy/bias-flagged base detailed above; Pursey pooled 25 studies (196,211 participants) to a clinically-weighted 19.9% prevalence that its own authors caveat as higher than a general population would show. Schulte, Gordon and Pursey all share the Gearhardt/YFAS lineage (Gordon’s evidence base literally includes Schulte [13]; Gordon cites Pursey as ref [34]) -> NOT independent-E (a shared author/instrument/study defeats independence); their agreement is shared-lineage coherence, not independent convergence. The construct-description cell (attribute / validity / prevalence) is now saturated; the outcome cell below is not.
- G-gap — no food-addiction (or HPF) -> patient-important-outcome SR exists. There is no held pooled analysis linking a YFAS-defined addictive-eating exposure (or the implicated food attributes) to weight gain, T2D, or mortality as an addiction-specific pathway distinct from ordinary overconsumption. Until one exists the construct cannot enter the Layer 1 - Ranking Interventions for a Stratum hierarchy as a sized lever. — an addictive-eating/HPF-exposure -> hard-outcome analysis.
- G-gap — no measured addictive response. Schulte measures reported perceptions; whether the hallmark addictive processes (tolerance, withdrawal, neural reward dysregulation) actually fire in humans to these foods is unstudied here — the authors call for future work measuring biological responses and observing eating behaviour directly.
- Not a joined issue with the overconsumption mechanism (yet). Whether “addiction” adds anything over the eating-rate / energy-density / palatability channels that already explain UPF overconsumption (-> Ultra-Processed Food and Health Outcomes) is an open question, not a filed tension — the construct and the mechanistic channels have not been tested against each other.
Self-critique [run 2026-08-30, Pursey 2014 prevalence weave, before commit]
- Over-claim check (the prevalence trap): the 19.9% is reported and immediately bounded — the transportability caveat is Pursey’s own next sentence (predominantly OW/OB females from clinical settings; likely higher than a nationally representative general population), and the page states the number sits toward the high end of a general population, contrasts it with Schulte’s own 6.7%/10.2%, and frames prevalence as sample-dependent (report a range, not a point). No bare 19.9% of people are food-addicted is asserted. The number-trace holds: every figure (19.9, 24.9/11.1, 12.2/6.4, 22.2/17.0, 4.0/1.7, 196,211, 25) is inside a cited Pursey quote.
- Evidence-state discipline: prevalence is filed as a descriptive fact that moves neither evidence state — not validity (Gordon), not any patient-important outcome. The page explicitly keeps the outcome state at insufficient-evidence after Pursey, blocking the it’s common, therefore it matters inflation.
- The beyond-summary move stays the two-evidence-states separation (A/C): Pursey extends the construct-description cluster with a third facet (prevalence) alongside attribute (Schulte) and validity (Gordon) — a banked-C cluster extension. It does not clash with the others (no D) and is not independent of them (no E).
- Independence: no
[E-independent]written. Pursey shares the Gearhardt/YFAS lineage with Schulte and Gordon (Gordon cites Pursey as ref [34]); all three rest on the same instrument. Agreement is shared-lineage coherence, flagged in the prevalence and gaps sections. - Saturation verdict recorded: the construct-description cell is saturated across three non-independent YFAS sources; the outcome cell was never opened and stays a G-gap. This is the stopping rule made explicit — a fourth YFAS-descriptive source adds neither perspective nor upgrade.
- Counter-passage read: Pursey’s discussion (chunk 02) and results/Table 4 (chunk 01) were read end-to-end; the source’s own general-population caveat is reported rather than suppressed, which is what makes the transportability point EXTRACTED, not merely INFERRED.
- Confidence: stays
low— a gold prevalence SR/MA describes how common the self-report construct is in clinically-enriched samples; it adds no independent backing, no pooled outcome magnitude, and no patient-important-outcome evidence. Descriptive breadth does not lift a page whose outcome state is still insufficient.