14 November 2018
Cara B Ebbeling, principal investigator, and associate professor,
Henry A Feldman, principal biostatistician, and associate professor,
Gloria L Klein, study director, Julia M W Wong, associate study director, and instructor, Lisa Bielak, nutrition research manager, Sarah K Steltz, data and quality manager, Patricia K Luoto, professor, Robert R Wolfe, professor, William W Wong, professor, David S Ludwig, principal investigator, and professor
Introduction
Evidence from animal and human studies shows that biological factors strongly influence body weight. With weight loss, hunger increases and energy expenditure decreases—physiological adaptations that defend against long term weight change. Genetic factors are known to affect body weight, explaining some of the variance in body mass index (BMI) among people.
However, genetic factors cannot explain why the average person today, compared with 40 years ago, seems to be “defending” a much higher body weight.
According to the carbohydrate-insulin model of obesity, the increased ratio of insulin to glucagon concentrations after consumption of a meal with a high glycemic load directs metabolic fuels away from oxidation and toward storage in adipose tissue.
This physiological state is hypothesized to increase hunger and food cravings, lower energy expenditure, and predispose to weight gain, especially among those with inherently high insulin secretion. The carbohydrate-insulin model offers a physiological mechanism for understanding why obesity rates have increased since the 1970s in the United States, as dietary fats were replaced with high glycemic load foods, including refined grains and added sugars.
This model has been challenged, primarily owing to lack of evidence from controlled feeding studies. A recent meta-analysis reported no meaningful difference in energy expenditure between low carbohydrate and low fat diets.
The studies included in that analysis, however, were short term (mostly <2 weeks), whereas the process of adapting to a low carbohydrate, high fat diet seems to take at least two or three weeks. For this reason, transient effects of macronutrients cannot be distinguished from long term effects on the basis of existing evidence.
We compared the effects of diets varying in carbohydrate to fat ratio on energy expenditure during weight loss maintenance through 20 weeks.
Methods
The study protocol has been previously published.19 We collected data on the campus of Framingham State University, Massachusetts, between August 2014 and May 2017. For implementing controlled feeding protocols with free living participants, we established a partnership with Sodexo, the food service contractor at Framingham State University.
For the final year of the study, a satellite feeding site was established at Assabet Valley Regional Technical High School (Marlborough, MA). The study was known as the Framingham State Food Study, or (FS)2.
Design
We carried out a randomized controlled trial with run-in and test phases (fig 1). During the run-in phase, energy intake was restricted to promote 12% (within 2%) weight loss over 9-10 weeks. We randomly assigned participants who achieved the target weight loss to high, moderate, or low carbohydrate test diets for a 20 week test phase.
During the test phase, participants’ energy intake was adjusted periodically to maintain weight loss within 2 kg of the level achieved before randomization. Participants were asked to weigh themselves daily using calibrated Wi-Fi scales (Withings, Cambridge, MA) during both phases.
Study outcomes were assessed at several time points: pre-weight loss, start of trial (weeks −2 to 0, before randomization), midpoint of test phase (weeks 8 to 10), and end of test phase (weeks 18 to 20), as summarized in figure 1 and supplemental eTable 1.
Participants
Adults aged 18 to 65 years, with a BMI (weight (kg)/(height (m)2) of 25 or higher and body weight less than 160 kg, were screened for participation before pre-weight loss assessments. Supplemental eTable 2 presents additional eligibility criteria.
For each of three cohorts, recruitment occurred during the spring semester before the respective academic year (August to May) of study participation. Participants provided written informed consent at the time of enrolment.
The stipend for participation was $3280 (£2559; €2880) over the course of the study, and meals were valued at $3220, for total compensation of $6500. (See supplemental methods for details on implementation of randomization.)
Dietary interventions
During the run-in phase, the macronutrient composition of the run-in diet was 45% of total energy from carbohydrate, 30% from fat, and 25% from protein. The target macronutrient composition of the run-in diet reflects ranges considered acceptable by the Institute of Medicine, with protein at the upper end of the range to enhance satiety during weight loss.
We determined individual energy needs on the basis of resting requirements, estimated using a regression equation2324 and multiplied by a physical activity factor of 1.5 (which corresponds to a light activity lifestyle). Energy intake was restricted to 60% of estimated needs. The research team monitored participants’ body weight and adjusted the amounts of food when necessary to achieve the target weight loss.
At the end of the run-in phase, we adjusted energy intake to stabilize body weight on the basis of the recent rate of weight loss for each participant: energy intake during weight loss (kcal/d)+(rate of weight loss (kg/day)×7700 kcal/kg) (1 kcal=4.18 kJ=0.00418 MJ). During the test phase, high, moderate, and low carbohydrate diets varied in carbohydrate (60%, 40%, and 20% of total energy, respectively) and fat (20%, 40%, and 60%, respectively), with protein fixed at 20% (table 1).
We controlled for protein, in view of its higher thermic effect, to provide a more specific test of the carbohydrate-insulin model. The relative amounts of added sugar (15% of total carbohydrate), saturated fat (35% of total fat), and sodium (3000 mg/2000 kcal) were held constant across diets.
Based on regression of body weight (g) on time (days), a slope of 15 g or more each day over 14 days indicated the need to adjust energy intake to achieve weight stability within 2 kg of the start of trial weight. (See supplemental methods for details on menu development, quality control, and strategies to promote adherence.)
Study outcomes
Prespecified outcomes included energy expenditure, measures of physical activity, and metabolic hormones. To test for effect modification predicted by the carbohydrate-insulin model we assessed insulin secretion (insulin concentration 30 minutes after oral glucose) at pre-weight loss.
Staff masked to dietary group assignment collected data on outcomes. Total energy expenditure (primary outcome) was assessed using the doubly labeled water method. Participants provided two pre-dose spot urine samples on separate days and seven post-dose samples at regular intervals over an assessment period of 14 days. Isotopic enrichments of urine samples were measured in duplicate using gas isotope ratio mass spectrometry.
The equation of Ravussin et al was used to calculate total energy expenditure from carbon dioxide production (rCO2), with food quotient as a proxy for respiratory quotient. We expressed total energy expenditure in kcal per kg body weight, then normalized this to average start of trial body weight (82 kg) for analysis and reporting.
This approach takes into account small changes in body weight that might occur during the test phase, within our definition of weight loss maintenance (within 2 kg of the start of the trial weight), and thereby improve precision.
Some investigators discourage adjustment of total energy expenditure for weight because of confounding that would arise from individual differences in relations between total energy expenditure and body weight, body composition, and metabolically active mass.
However, this problem, inherent to cross sectional comparisons between people, would not apply to the within individual comparisons over several months in our study, especially during weight loss maintenance when these relations would not change in any meaningful way.
We also examined absolute total energy expenditure expressed as kcal/d, with and without body weight included as a covariate, and we obtained similar results. (See supplemental methods for details on measurement of body weight, resting energy expenditure by indirect calorimetry, energy intake, physical activity by accelerometry, skeletal muscle work efficiency by cycle ergometry, oral glucose tolerance testing, and assays of blood samples.)
Statistical analysis
Sample size calculations were based on data from a preliminary study. The target of 135 completers provided 80% power, with 5% type I error, to detect a difference of 237 kcal/d in total energy expenditure change between one diet group and the other two diet groups.
This difference is smaller than the effect detected in the previous study and is consistent with a predicted effect of 50 kcal/d per 10% decrease in the contribution of carbohydrate to total energy intake.
Before unmasking of diet group assignment, the primary outcome measure, total energy expenditure, was derived from a non-linear decay model fitted jointly to urinary disappearance curves of stable oxygen and hydrogen isotopes after oral administration of the doubly labeled water. We used the jackknife technique to smooth the parameter estimates and discarded a small number of incomplete or poorly fitting curves, deviant data points, and implausible values.
The prespecified analytic framework for the primary outcome was repeated measures analysis of variance spanning three time points (start of trial, midpoint of test phase, and end of test phase), with diet assignment as a three level independent variable (high, moderate, low carbohydrate).
The value at pre-weight loss, rather than start of trial, was originally specified in the registry as the basis for calculating change scores, but this error was corrected in an amendment to the institutional review board protocol, before unmasking diet group assignment. (See protocol amendment history in supplement for details.)
The main model was unadjusted except for design factors (study site, cohort, and enrolment wave). A fully adjusted model for the primary outcome also included demographic characteristics (sex, ethnicity, race, and age); pre-weight loss values for BMI, percentage lean mass, and total energy expenditure; and weight loss from pre-weight loss to start of trial.
An unstructured covariance matrix provided maximum flexibility in modeling correlation within participants over time. From parameters of the fitted model, taking account of all data, we constructed the mean test phase change in total energy expenditure for each diet (covariate adjusted change between start of trial and midpoint of the test phase and end of the test phase, the latter two averaged) and tested the hypothesis that this change was uniform across diets, using a two degrees of freedom F test with a P value threshold for significance of 0.05.
When this hypothesis was rejected, the principle of closed testing permitted us to make the three pairwise comparisons of the different macronutrient diets with critical P value 0.05 while preserving a maximum 5% type I error rate for the set of four potential comparisons (one overall and three pairwise).
The high versus low carbohydrate diet comparison was equivalent to a test for linear trend across the three diets according to their equally spaced carbohydrate content.
To test for effect modification, we divided the sample into thirds of pre-weight loss insulin secretion, fasting glucose, and fasting insulin; added appropriate interaction terms to the repeated measures model; and constructed contrasts to test for linear trend across thirds for the between diet differences in change during the test phase.
Secondary outcomes (resting energy expenditure, physical activity, and the metabolic hormones ghrelin and leptin) were analyzed similarly to total energy expenditure. For analysis, we log transformed the concentrations of the hormones and triglycerides. For reporting, we retransformed the adjusted mean and standard error to the original units (exp(mean log) ±exp(mean log)×(exp(SE log)–1)), and changes were expressed in percentage units (100%×(exp(change in log)–1)).
Analysis was performed on the full intention-to-treat sample and a per protocol subset comprising those participants who maintained weight loss within 2 kg of the start of trial weight during the test phase, the latter potentially providing a more precise effect estimate. After each analysis, we examined residual patterns to detect outliers or other departures from assumptions of the statistical model.
Recognizing that estimates of food quotient introduce some imprecision when calculating total energy expenditure, due in part to uncertainty in estimates of metabolizable energy, we conducted sensitivity analyses to determine how plausible errors in food quotient could influence results.
To test for selective dropout, we compared pre-weight loss characteristics of participants who completed the end of the test phase assessment with those who did not. To fully assess the influence of missing data (dropouts and unusable data points), we performed an inverse probability weighted version of the primary analysis, constructing a logistic model for missingness and employing the fitted probabilities to assign weights in the primary analysis. We used SAS software version 9.4 for all computations (SAS Institute, Cary, NC).
Missing data and quality of fit
Two randomized participants were excluded from all analyses: one developed hypothyroidism and one provided unreliable data for doubly labeled water at the start of the trial and then withdrew before notification of diet assignment.
Of 486 potential total energy expenditure values for use in the primary repeated measures analysis (162 participants×three time points), 457 were available (94%); for the per protocol analysis, 337 of 360 (94%) were available. The missing values were attributable to 24 missed doubly labeled water studies (nine during the midpoint of the test phase, and 15 at the end of the test phase) and five studies that yielded non-convergent curve fits or implausible parameters (one at the start of the trial, three during the midpoint of the test phase, and one at the end of the test phase).
Neither the intention-to-treat nor the per protocol findings changed materially when we applied inverse probability weighting to compensate for the missing data. For secondary outcomes, the percentage of non-missing values varied between 94% (resting energy expenditure, physical activity, 459 of 486) and 95% (hormones, 460 of 486).
Residual patterns showed a satisfactory fit to the repeated measures model in all cases, with no extreme outliers or pathological distributions.
Patient and public involvement
No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results.
Study participants received a written summary of their clinically relevant results. We plan to invite study participants to Framingham State University for an oral presentation of findings after publication of the primary outcome. Information may be disseminated to the general public via any media coverage of study findings.
Results
Participants
Of 1685 people screened, we enrolled 234 participants for the run-in phase (fig 2). Of these, 164 achieved 12% (within 2%) weight loss and were randomly assigned to one of three macronutrient diets for the test phase, comprising high (n=54), moderate (n=53), or low (n=57) levels of carbohydrate. Table 2 presents the characteristics of the randomized sample at the pre-weight loss time point.
Each stratification factor in the randomization was balanced across the three diet groups according to Fisher’s exact test (P≥0.28). Among the 162 participants included in the intention-to-treat analysis, primary outcome data were available for 161 (99%) at the start of the trial, 150 (93%) at the midpoint of the test phase, and 146 (90%) at the end of the test phase.
