Why Popular Diets Don't Last: Keto, Paleo, Fasting

In 2007, Christopher Gardner at Stanford published the A TO Z Weight Loss Study in JAMA. Three hundred and eleven women were randomised to one of four named diets: Atkins, Zone, LEARN, or Ornish. Each had a published book, a public reputation, and a substantial fan base. The study followed participants for twelve months under standard conditions. The result was the kind of finding that publishers prefer to bury: the four diets produced statistically similar long-term weight loss, in a range of roughly 2.6 to 4.7 kg at one year. Atkins came out slightly ahead, but the difference between the best and worst arms was smaller than the variation within each arm.

Gardner extended this work in 2018 with DIETFITS, a 609-person trial comparing healthy low-fat and healthy low-carbohydrate diets across twelve months, published in JAMA. Again, the headline result was a non-finding: average weight loss differed by less than a kilogram between the two arms. Some low-fat participants did beautifully. Some failed. Some low-carb participants did beautifully. Some failed. The macronutrient strategy was not the variable that distinguished outcomes.

Twenty years of these trials, with different research groups and different diet permutations, keep arriving at the same answer. The diet labels matter less than the diet literature suggests they should.

What the head-to-head trials actually find

Frank Sacks at Harvard ran POUNDS Lost in 2009, published in the New England Journal of Medicine. The trial randomised 811 adults to one of four diets varying in fat, protein, and carbohydrate content. At two years, the four groups had lost statistically equivalent amounts of weight: approximately 4 kg on average, with wide individual variation. The macronutrient composition assigned to each participant had essentially no detectable effect on long-term outcomes. Adherence — measured by attendance at group sessions — did.

Krista Varady's group at the University of Illinois Chicago tested intermittent fasting against caloric restriction in a 12-month trial published in JAMA Internal Medicine in 2017. The intermittent-fasting arm did not lose more weight than the daily-restriction arm. The dropout rate was higher. The cardiometabolic improvements were comparable. The study did not invalidate intermittent fasting as a tool; it did invalidate the idea that the eating schedule was an independent driver of weight outcomes once total intake was held similar.

The PREDIMED trial, published in 2013 and revised in 2018 in The New England Journal of Medicine, examined the Mediterranean diet versus a control low-fat diet across a median 4.8 years in over 7,000 high-cardiovascular-risk adults. The Mediterranean arms reduced cardiovascular events meaningfully. Weight loss between arms differed only modestly. The trial is widely cited as evidence for the Mediterranean pattern's cardiovascular benefits — which are real — but the weight-loss differential was not the headline.

Why all the diets produce similar curves

The mechanism is straightforward. Every named diet that produces weight loss does so by creating a caloric deficit, whether or not its proponents acknowledge or measure calories. Low-carb diets reduce intake by removing a major source of palatable, energy-dense foods. Low-fat diets do the same with a different category. Intermittent fasting reduces the eating window, which reduces opportunities to consume calories. Paleo eliminates processed foods and most refined carbohydrates. Each diet's mechanism converges on intake reduction, regardless of the framing.

The body's response to that deficit, then, follows the same trajectory regardless of the macronutrient mix. Adaptive thermogenesis kicks in. Ghrelin rises. Satiety hormones drop. The biological pressure toward regain emerges. The Sumithran 2011 hormonal data are not specific to any particular dietary protocol. They are specific to the state of sustained caloric restriction itself.

The Dansinger trial: adherence beats macronutrients

If one study captures the pattern most cleanly, it is Michael Dansinger's 2005 trial in JAMA. His group at Tufts randomised 160 overweight adults with at least one cardiac risk factor to four of the most heavily marketed programmes of the era: Atkins, Zone, Weight Watchers, and Ornish. These are not minor variants of one another. Atkins is very low carbohydrate; Ornish is very low fat; Zone and Weight Watchers sit between them. If the macronutrient framing mattered, this design should have exposed it.

It did not. At one year, mean weight loss ranged from 2.1 kg (Atkins) to 3.3 kg (Ornish) — modest, and statistically indistinguishable across the four arms. The more telling numbers were the dropout rates: roughly half of the Atkins and Ornish participants abandoned their assigned diet before the year was out. The diets that demanded the most dramatic departure from a habitual British or Western diet were the ones people were least able to sustain.

Dansinger then did the arithmetic that named-diet marketing tends to avoid. He correlated weight loss with self-reported adherence, and separately with diet type. Adherence correlated with weight loss at r = 0.60 — a strong relationship. Diet type correlated at r = 0.07 — essentially nothing. Put plainly: knowing how faithfully someone followed their plan told you a great deal about their result; knowing which plan they were assigned told you almost nothing. This is the quantitative backbone beneath the softer claim that "the best diet is the one you can stick to."

The adherence problem that no diet solves

The variable that consistently does predict long-term outcomes is adherence — how closely participants follow the diet they were assigned. In POUNDS Lost, attendance at group sessions correlated more strongly with weight loss than any macronutrient variable. In DIETFITS, the participants who self-reported high adherence in either arm lost substantially more than the participants who self-reported low adherence in either arm. The diet labels were essentially noise on top of the adherence signal.

It is worth being precise about what "adherence" is, because the word is too often treated as a synonym for willpower. The trials suggest it is closer to compatibility — a fit between the diet and the person's appetite, food environment, and tolerances. We have argued elsewhere that appetite is regulated by biology rather than governed by willpower, and the diet-comparison literature is consistent with exactly that reading. The participants who "adhered" were rarely the most disciplined; they were the ones whose assigned pattern happened not to fight their hunger signalling. When a diet works against those signals, adherence decays no matter how motivated the dieter begins.

Adherence, in turn, is not equally distributed. People who find a given dietary pattern compatible with their food preferences, schedule, social environment, and biological tolerances tend to adhere to it. People who find it incompatible drift away. The matching of person to pattern matters enormously. The pattern itself, in the aggregate trial data, matters less than the marketing suggests.

This is partly why head-to-head trials have such large within-arm variation. Some low-carb participants are perfectly suited to the structure and lose 10% of body weight. Some low-carb participants find the restriction unsustainable and lose nothing. The same is true for every other dietary approach. The aggregate average obscures these realities.

The role of food reward in adherence collapse

One of the consistent findings in long-term diet trials is that adherence decays over time, and the decay is not random. The foods participants find themselves drifting back toward are predictable: ultra-processed, hyperpalatable, energy-dense. Kevin Hall's NIH metabolic-ward study in 2019, published in Cell Metabolism, demonstrated that when participants were offered ad libitum access to ultra-processed versus minimally-processed food in tightly matched calorie, protein, fat, sugar, fibre, and sodium conditions, they ate approximately 500 kcal more per day on the ultra-processed arm. The food category drives intake beyond what nutritional matching would predict.

Any diet that eliminates these foods will reduce intake initially. The question is whether the person can keep eliminating them across the food environment they actually live in. For most participants in most trials, the answer eventually becomes no. The drift back toward palatable, accessible food is the route by which diet adherence collapses across all named protocols.

What each approach gets right

The trials are not arguments that all diets are equivalent in every respect — only that they tend to produce similar weight outcomes over twelve to twenty-four months in unselected populations. Each approach has features that suit specific people:

Ketogenic and low-carbohydrate diets reduce glycaemic variability and can be particularly useful for type 2 diabetes management. The short-term satiety effect of higher protein and fat can suit people who struggle with frequent hunger on higher-carbohydrate diets. The Diabetes Prevention Program and some glycaemic-control trials have produced strong outcomes on lower-carb arms.

Mediterranean and largely whole-food patterns have the strongest cardiovascular outcome data and are typically the most sustainable across years. PREDIMED's hard-endpoint findings on cardiovascular events are among the most robust in the nutritional epidemiology literature.

Intermittent fasting suits people whose food environment makes calorie tracking exhausting and who do better with binary rules (eat / don't eat now) than with quantitative ones. The eating-window approach removes some categories of decision fatigue.

None of these is a wrong answer. The trial data simply indicates that the choice between them is less consequential than the marketing suggests, and that the underlying mechanism — sustained caloric deficit against a biology that opposes it — is shared.

Why the adherence problem points beyond diet

Charles Billington at the University of Minnesota and others have argued that the consistent five-year regain patterns across all major dietary interventions point to a structural limitation: the interventions are addressed to behaviour, and the regain pressure is biological. The hormonal adaptation that follows weight loss does not care which diet produced the loss. The pressure to regain is independent of the macronutrient strategy that produced the deficit.

This reframes the entire diet-selection question. If the dominant predictor of long-term success is not the macronutrient ratio but whether the post-loss biology can be held in check, then the most consequential decision is not which diet to start but how to defend the result once the weight is lost. We have set out the evidence on that separately, but the short version is that regain can be substantially mitigated — by sustained support, resistance training to preserve lean mass, and, increasingly, by pharmacology that blunts the hunger rebound. None of those levers is a property of the diet itself. They sit alongside it, which is precisely why head-to-head trials of diets keep failing to separate the diets.

This is the framing in which GLP-1 receptor agonists are best understood. The medications do not replace dietary patterns; many patients still benefit from a Mediterranean-style or higher-protein pattern. What they change is the biological pressure that makes adherence to any pattern so hard to sustain. The reduced hunger and reduced food reward on semaglutide or tirzepatide make adherence less effortful, which is why the long-term weight curves on these medications look different from the long-term curves on any diet in the comparative literature.

The trial data on diet choice has been remarkably consistent for twenty years. Picking the right diet is, statistically, less important than people are led to believe. Addressing the biology that drives adherence collapse is, statistically, more important than the diet literature has historically acknowledged. For patients navigating food choices on a GLP-1, this reframe is the practical implication: the diet pattern is a layer on top of the medication, not a substitute for it, and the comparative trials are clear that the layer's specific shape is largely a matter of personal compatibility.