How to Read a Regain Curve
Post-cessation GLP-1RA research asks a deceptively simple question: how much weight comes back after treatment ends, and when? To answer it, the percentage must be tied to the right denominator and to a time course. This chapter builds that reading method so that a trial result or pooled estimate can be understood as a trajectory rather than mistaken for either a return to baseline or a personal forecast.
The distinction matters because weight-loss studies often place several percentages close together. One may describe the change from the starting weight, another the amount regained after treatment, and a third the portion of the original loss that returned. These are related measurements, but they answer different questions. A reliable reading keeps them separate before interpreting what the curve might mean for treatment durability.
The denominator is part of the result
There are three different quantities in a withdrawal study. The first is the change from a person’s starting weight. The second is the amount lost during treatment, measured in kilograms or as a percentage of starting body weight. The third is the share of that treatment loss that returns after cessation. The systematic review uses the third quantity as its primary outcome:
percentage weight regain = weight regained after cessation ÷ weight lost during treatment
This is not a technical detail. It determines what the headline means. Consider a hypothetical person who begins at 120 kilograms, loses 18 kilograms during treatment, and later regains 9 kilograms. The person has regained 50% of the weight lost, because 9 is half of 18. Yet the final weight is 111 kilograms, still 9 kilograms—or 7.5%—below the starting point. “Half the treatment loss returned” and “the person remained below baseline” describe the same outcome from different angles.
The same regain fraction can also represent very different amounts. Regaining 50% after a 6-kilogram loss means 3 kilograms; regaining 50% after a 20-kilogram loss means 10 kilograms. Normalizing by treatment loss helps researchers compare studies whose participants lost different amounts, but it does not erase those underlying differences. Nor does it indicate whether the regain happened steadily, in a short burst, or mostly near the end of follow-up.
Reports sometimes combine these measures in ways that are easy to misread. A percentage change in body weight from baseline is not automatically a percentage of treatment loss. If a participant loses 17.3% of baseline weight and later gains 11.6 percentage points relative to baseline, the latter divided by the former is about two-thirds. That is why the STEP 1 extension can report an 11.6-point regain, while describing the result as roughly two-thirds of the prior loss. Keeping the units and denominators visible prevents the figures from sounding contradictory.
The word “baseline” also needs a fixed reference point. In a trial, it usually means the measurement taken before the randomized treatment period. A later percentage may instead be calculated from the weight at the end of treatment. If those reference points are silently switched, a reader can mistake a change during withdrawal for the total change since treatment began. The safest interpretation asks two questions at once: compared with which weight, and expressed in which units?
A normalized measure is therefore a comparison tool, not a complete description of a person’s health. It says how much of an earlier change was reversed. It does not tell us whether the remaining loss came from fat or lean mass, whether blood pressure improved, or whether the same person would have followed the same path with continued lifestyle support. Those questions require other outcomes and a timeline.
The definition also explains why a study with very little initial loss can produce an unstable or misleading regain percentage. If the denominator is small, a modest absolute change can look very large. In its broader review, the authors treated studies with less than 3 kilograms of treatment loss as less reliable for interpreting the regain fraction. One apparent result above 100% was likely influenced by participants switching to insulin, which can affect weight independently of the withdrawal being studied. The percentage is informative only when the context around it is retained.
From a single percentage to a time-based curve
The Cambridge systematic review modeled regain as an exponential recovery process:
R(t) = A(1 − e^(-kt))
Here, R(t) is the percentage of treatment loss regained at time t after cessation. A is the estimated maximum share of the original loss that the curve approaches. k is the recovery-rate constant: a larger value means the curve rises more quickly. The form of the equation matters because it describes a changing rate. The expected movement is larger earlier and gradually slows as the trajectory approaches its estimated upper level.
This is easier to understand as a gap-closing model. At the moment treatment stops, the gap between observed regain and the model’s upper level is large, so the predicted change can be relatively fast. As regain accumulates, the remaining gap becomes smaller, and the predicted change slows. The model does not say that weight must move in smooth weekly increments. It summarizes an average shape across observations that were taken at different times.
The model was fitted to six randomized controlled trials with a total of 3,236 participants. It estimated a maximum regain of 75.3% of treatment loss, with a 95% confidence interval from 68.9% to 81.6%. The estimated rate constant was 0.0302 per week, and the corresponding half-life was about 23 weeks, with a confidence interval from 17.3 to 34.3 weeks. In plain language, the parameters summarize two different ideas: how far the curve is projected to go and how quickly it moves toward that level. They should not be collapsed into one claim that everyone regains three-quarters of their loss on a fixed schedule.
The half-life is attached to the model’s approach toward its estimated upper level. It is not a promise that a person will regain half of their original loss in exactly 23 weeks. In the fitted equation, it describes the time scale over which the remaining gap to the asymptote is reduced by half. This distinction is subtle but useful: the parameter describes the curve’s rate, while the actual percentage at a particular visit depends on the curve’s starting point, its upper level, and the time observed.
The model estimated about 60% of the original treatment loss regained at 52 weeks, equivalent to roughly 79% of its estimated maximum. Imagine, again only as an illustration, that the original treatment loss was 20 kilograms. A 75% regain would be about 15 kilograms, leaving 5 kilograms of net loss. That arithmetic explains how a large reversal can coexist with a lower weight than before treatment. It is not a prediction for a particular person; it is a way to translate the model’s denominator into a concrete scale.
The curve’s shape is as important as its endpoint. A straight line would imply the same amount of regain in every equal interval of time. The exponential form instead suggests that the early post-cessation period may contain a faster change, followed by a progressively slower approach. This distinction matters for interpretation and for follow-up. A measurement at four or twelve weeks may capture a different phase from one taken at a year. Two studies that collect data at different times can appear to disagree even when their trajectories are broadly compatible.
The diagram below separates the part of the trajectory supported by follow-up and fitting from the part that remains a projection. The fitted curve can be used within the available evidence window to summarize the observed pattern. Once it extends beyond that window, the shape is an assumption about what may happen next, not a direct measurement.
Precision does not remove uncertainty
The numbers in the model can look more authoritative than the evidence base warrants. The review found 48 relevant studies, but only six met the stricter criteria used for trajectory modeling: randomized design, at least 100 participants in the treatment arm, substantial on-treatment loss, and enough post-cessation observations to estimate a curve. The selection improves comparability, but it also means that the fitted result is not a summary of every person or every study included in the review.
The 48-study review and the six-study model answer related but different questions. The broader collection is useful for asking whether rebound appears across drugs and settings. The narrower model is useful for asking whether a common mathematical shape can summarize better-documented trajectories. More studies do not automatically make a curve more reliable if their measurements are sparse or their designs differ too much. Conversely, a strict model sample can produce a clearer estimate while leaving out groups or circumstances that matter in practice.
The available observations ended at 52 weeks after cessation. The model’s estimate of a 75.3% maximum regain therefore extends beyond the longest modeled window. No large-scale trial in the review had directly measured what happened beyond one year. A predicted plateau below the pre-treatment weight is scientifically interesting because it suggests that some population-average benefit might persist, but it remains an extrapolation. A later study could show that the curve continues upward, levels off at a different point, or differs substantially across groups.
Confidence intervals communicate uncertainty around the fitted parameters, but they do not solve every other uncertainty in the evidence. They do not account for a population that was never included, a maintenance intervention that was not measured, or a difference in treatment duration that the model could not isolate. A narrow interval around an estimate can coexist with uncertainty about whether the estimate applies outside the study conditions. Statistical precision and clinical applicability are related, not interchangeable.
The broader evidence also shows why a single curve should not be mistaken for a universal timetable. In studies with more than 3 kilograms of initial treatment loss, regain at around 12 weeks ranged from approximately 10% to 80% of the loss. That range is wide because the studies differed in treatment, duration, population, support, and follow-up. It is a description of the evidence landscape, not a menu of likely outcomes from which an individual can select a number.
There is a further distinction between what the model estimates and what it explains. The curve describes the pattern of change; it does not prove why the change occurred. It cannot by itself separate the return of appetite, physiological responses to weight loss, changes in behavior, environmental pressures, or the loss of structured counseling. Nor does it establish that the same curve applies to people with different comorbidities, smaller treatment losses, different doses, or a planned maintenance strategy.
For that reason, the most defensible reading has several layers. First, the direction of change is consistent across the supplied review: most studies report some weight regain after GLP-1RA cessation. Second, the timing is not instantaneous or uniform; the modeled population pattern is faster early and slower later. Third, the 52-week estimate is supported by the available modeled follow-up, while the long-term maximum is projected. Finally, the curve is a population summary whose value is greatest when its denominator, data window, and uncertainty remain attached.
The practical payoff of reading the curve this way is not a formula to apply without context. It is a better question. Instead of asking only, “What percentage will come back?” a careful interpretation asks: percentage of what, measured when, in which population, after what treatment and support, and observed or extrapolated? Those questions turn a dramatic headline into evidence that can be compared fairly. They also prepare us to examine a concrete example: what did one year without semaglutide actually look like for participants in the STEP 1 extension?