What Would Count as Evidence for a Metabolic Theory?

Explanations of weight regulation are easy to find and hard to evaluate. One says the key is a hormone, another a type of fat, another the degree of food processing, another the brain’s control of appetite. Each comes with a plausible mechanism and a few striking stories. Before this site weighs any particular theory, it has to settle a more basic question: what would count as evidence for one, and what would count against it?

The question and why it comes first

If you decide what counts as evidence only after seeing the results, almost anything can be made to fit. Deciding in advance is what keeps an investigation honest, and it is the standard this site tries to hold itself to.

A theory, in the sense used here, is a proposed explanation that links a cause to an outcome through a mechanism. It should say what happens, roughly how much, to whom, and under what conditions. And it has to be able to lose. If an idea can explain weight gain, weight loss and no change equally well, it is not explaining much.

Two things are often blurred:

• A mechanism: a pathway by which something could affect body weight.
• A demonstrated outcome: evidence that it does, in people, by an amount that matters.

A worked test case: two models of obesity

The clearest recent example of theories being put to the test is the argument between two models.

The carbohydrate–insulin model proposes that the hormonal response to a high-glycaemic-load diet drives fat storage, and that this fat storage then drives hunger and overeating. Its proponents describe this as a reversal of the usual causal direction, and argue that it offers testable hypotheses [1].

The energy balance model, as restated by its defenders, holds that the brain regulates body weight, mostly below conscious awareness, through signals that control food intake in response to the body’s needs and the food environment. Its authors argue that the latest version of the carbohydrate–insulin model dropped a formerly central claim and is better seen as a special case of their model [2]. The carbohydrate–insulin camp replied that the energy balance model has its own weaknesses that make it hard to test [3].

Arguments about framing can go on forever. What moves the debate is predictions.

Discriminating predictions, and what happened

Each row below is a prediction the carbohydrate–insulin model makes more strongly than its rival, with the human evidence that tested it.

Prediction	Test	Result
Swapping carbohydrate for fat at equal calories increases energy expenditure and fat loss	17 men on a metabolic ward, 4 weeks each diet	Small rise in expenditure (57 kcal/day in a respiratory chamber); body-fat loss slowed on the ketogenic diet [4]
Same prediction, pooled	Meta-analysis of 32 controlled feeding studies	Expenditure (26 kcal/day) and fat loss slightly greater on lower-fat diets [5]
Low-carbohydrate diets raise expenditure during weight-loss maintenance	164 adults randomised for 20 weeks	Total expenditure 209 kcal/day higher on low- vs high-carbohydrate (intention to treat), larger in people with high insulin secretion [6]
Low-carbohydrate diets reduce how much people choose to eat	20 inpatients, 2 weeks each diet, eating freely	The low-fat diet led to 689 kcal/day less intake than the low-carbohydrate diet [7]

Read together, the rows do not tell a one-sided story, and the details matter.

The measurement dispute. The 20-week trial measured expenditure with doubly labelled water, a method that requires an assumption about how the body is using fuel. When Hall and colleagues compared doubly labelled water against chamber and body-composition data from their own ward study, they concluded that calculations which ignore diet-specific effects on that assumption can wrongly suggest that low-carbohydrate diets substantially raise expenditure [8]. That is a challenge to the instrument, not just the result. Settling it needs studies that measure expenditure several ways at once.

The duration dispute. The ad libitum trial lasted two weeks per diet. Its authors concluded that the results were inconsistent with the carbohydrate–insulin model [7]. Proponents of the model responded that short ward trials risk unfounded conclusions about chronic effects, that metabolic adaptation can take weeks to months, and called for feeding studies of at least a month and efficacy trials of at least a year [3]. That is a fair methodological point, and also a prediction: if adaptation is the explanation, longer trials should show the effect emerging.

Who ran the studies. Several of the tests above come from one NIH group whose lead investigator also led the energy balance restatement [2] [4] [7]. The 20-week trial comes from the carbohydrate–insulin model’s leading proponents, and one of its authors declared royalties from books recommending a low-glycaemic-load diet [6]. Neither fact makes a result wrong. Both are reasons to want replication by groups with no stake in either model.

Notice what each side is doing. Both accept that the question is empirical. They disagree about which measurements and durations count. That is exactly the conversation worth having before anyone runs the next study.

Types of evidence, and what each can show

Different kinds of evidence answer different questions. None is useless, and none is enough alone.

• Human intervention studies. People are assigned to conditions, ideally at random. These are the best tool for testing cause and effect, but they are often short, can struggle with adherence, and may study narrow groups. Tightly controlled ward studies and free-living trials answer different questions.
• Human observational studies. These follow what people already do. They cover large groups over long periods but struggle with confounding, imprecise diet measurement and reverse causation.
• Animal studies. These allow control over diet, genetics and environment. Translating them has its own pitfalls, covered in what a mouse study can tell us about human weight regulation.
• Cellular and biochemical studies. These can show that a mechanism exists. Whether it matters for a whole person is a separate question.
• Anecdotes and self-experiments. Valuable for generating hypotheses, but they rarely isolate a cause, especially when several things change together. See the problem with changing five things at once.

Formal evidence-grading systems encode similar logic. Under GRADE, evidence from randomised trials starts as high certainty and observational evidence as low. Certainty is then marked down for study limitations, inconsistent results, indirectness, imprecision and reporting bias, and observational evidence can be marked up for very large effects or a dose–response relationship [9].

Convergence, not a single knockout

Confidence rarely rests on one study. It grows when different types of evidence point the same way despite different weaknesses; when independent groups replicate a finding; when larger exposures produce larger effects; and when a theory’s surprising predictions hold up, not just its obvious ones.

A working vocabulary for confidence

Instead of numerical scores, this site uses labels with written definitions:

• Well supported: several types of evidence converge, including replicated human intervention data, with no strong unexplained counterevidence.
• Plausible but unproven: a sensible mechanism and some supportive data, but key human tests are missing, small or inconsistent.
• Speculative: consistent with some observations, but not yet tested in a way that could have shown it to be wrong.
• Weakened: its predictions have been tested and largely not borne out, or an alternative explains the same observations better.

Theories with very different evidence get different labels. Presenting them side by side does not imply they are equally credible.

Applying the labels to the test case

• “At equal calories, cutting carbohydrate produces meaningfully more fat loss.” Weakened. Controlled feeding studies, including a pooled analysis of 32, have not shown it [4] [5].
• “Low-carbohydrate diets raise energy expenditure during long-term weight maintenance.” Plausible but unproven. One sizeable trial found it [6]; the measurement method is disputed [8]; it needs replication with multiple methods.
• “Low-carbohydrate diets reduce how much people eat.” Plausible but unproven, with counterevidence. A two-week inpatient comparison found the opposite [7]; whether longer exposure changes that is untested in the way proponents propose [3].

What would change these assessments: a feeding study of a month or more, with intake fixed and expenditure measured by chamber and doubly labelled water together, showing greater fat loss on a low-carbohydrate diet would move the first claim back toward “plausible.” Independent replication of the 20-week expenditure result with agreement across methods would move the second toward “well supported.”

Questions to ask of any metabolic claim

• What exactly does it predict, and how large an effect?
• What result would count against it?
• Which kinds of evidence support it, and which are missing?
• Does a rival explanation predict the same thing?
• Has the discriminating test been done, and do both sides accept the measurements?

Speculation, clearly labelled: I expect many of the most interesting questions in this field to sit for a long time in the “plausible but unproven” category, because the human studies needed to settle them are long, expensive and hard to run. I also suspect the final answer to the carbohydrate debate will be “it depends on the person” rather than a clean win for either model. That is a hunch, not a finding. When the evidence stays unsettled, I will say so rather than round the assessment up.

Before you change your diet based on any of this. Large shifts in diet composition have real physiological effects: one participant in the 2021 inpatient trial withdrew because of low blood sugar on the low-carbohydrate diet [7]. If you take medicine for diabetes or blood pressure, are pregnant, have kidney disease, or have a history of disordered eating, talk to your clinician before making major changes. Nothing here is a recommendation to follow any particular diet.