Center-Size as a Predictor of Weight-Loss Outcome in Multicenter Trials Including a Low-Calorie Diet




It has not been studied yet whether factors such as the number of subjects recruited by specialized centers for multicenter trials may influence weight loss during a low-calorie diet (LCD). This study aimed at determining whether the number of recruited subjects per center might predict relative weight loss. This is a post hoc analysis of an existing database: 701 obese subjects (77% women, 23% men, mean BMI: 38.9 kg/m2) were enrolled at 22 sites (4–85 subjects/site) in five countries to follow a LCD providing 800–1,000 kcal/day during 8 weeks. The main outcome measure was the percentage weight loss after the 8-week LCD. Mean weight loss differed significantly between participating centers (5.8–11.8% of the initial weight; P < 0.001). There was a significant positive correlation between relative weight loss and the number of recruited subjects per center (r = 0.38; P < 0.001). In a multiple stepwise regression analysis, the number of recruited subjects per center appeared to be the main predictive factor of weight loss (R2 = 0.07; P < 0.001). As the number of participants within each center is clustered, we applied a hierarchical model to model the average weight loss vs. the number of participants included at each center. This model allows to predict that for 10 extra patients in a center, the average weight loss would increase by 0.5%. This is the first study suggesting that the number of recruited subjects per center may impact weight loss, and could therefore be considered as a new predictor for weight loss that is independent from the individual.


A reduced calorie intake constitutes an indispensable part of any successful weight-loss management plan (1). During a reduced calorie diet, the main predictor for weight loss is the amount of energy consumed per day: thus, in the absence of changes in physical activity, a deficit of about 500 kcal per day predicts a weight loss of about 0.45 kg per week. Additionally, the macronutrient composition of the diet or the use of meal replacements (2) may play a role, even if this remains controversial. Furthermore, several studies have attempted to identify anthropometric (3) or psychological (4) predictors of weight loss. However, to the best of our knowledge, it has not been studied yet whether external factors such as the number of subjects recruited by centers participating in a multicenter weight loss trial may have an impact on weight loss. We hypothesized that the number of subjects a specific center is able to recruit for a weight-loss study will positively correlate with the weight loss these subjects will attain during the study. We based our hypothesis on the following assumptions, by taking the example of bariatric surgery—where it has been shown that the experience of a given surgeon is a predictor of better patient outcomes (5,6,7,8,9): (i) health-care practitioners working at a center that recruits higher numbers of subjects are more experienced; (ii) more experienced health-care practitioners will help their patients to achieve more important weight losses.

In order to test our hypothesis, we conducted this post hoc analysis of an existing database.

Methods and Procedures

This study is a post hoc analysis of an existing database emanating from a large, multicenter, randomized trial, which examined the efficacy and safety of topiramate vs. placebo on long-term maintenance of weight loss induced by a low-calorie diet (LCD) (10). After an initial run-in phase of 8 weeks consisting of an LCD, subjects were allocated to either topiramate or placebo for 60 weeks.

The study was carried out from August 2000 to June 2002. It was conducted in accordance with the Declaration of Helsinki and approved by Ethics Committees at all sites. All subjects provided written informed consent before enrolment.

Study population

We analyzed data from all subjects who started and successfully completed the 8-week LCD (completers analyses).

Subjects were eligible for enrolment if they were between 18 and 75 years of age and had a BMI ≥33 to <50 kg/m2 or a BMI ≥30 to <50 kg/m2 in the presence of controlled hypertension and/or dyslipidemia with a stable medication regimen. Subjects with type 1 and type 2 diabetes were ineligible unless they were newly diagnosed with diabetes by means of oral glucose tolerance test at the enrolment visit, and antidiabetic medication was not deemed necessary by the investigator. Subjects were also ineligible if they had known significant cardiovascular or renal disease, a history or family history of kidney stones, uncontrolled thyroid disease, or significant central nervous system-related or psychiatric disorders. To be eligible, a subject's weight had to have been stable for at least 3 months and smoking habits stable for at least 2 months before enrolment.

Anthropometric measurements

Each of the above mentioned visits included, among others, anthropometric measurements. Weight was measured using a standardized calibrated balance throughout the study: subjects were weighed in similar attire (light clothing) at each visit and at approximately the same time of the day. Height was measured using a wall-mounted stadiometer, with subjects not wearing shoes. Waist and hip circumference were measured with the subject standing, wearing underwear, and with the measuring tape kept horizontal. Waist circumference was measured at a level midway between the superior aspect of the iliac crests and the lower lateral margins of the ribs. Hip circumference was measured at the level of the pubic symphysis anteriorly and the greater femoral trochanters laterally.


Subjects had to lose at least 8% of their initial body weight during the LCD in order to be randomized to either topiramate or placebo. The composition of the LCD was at the discretion of individual centers; however, it had to be nutritionally balanced by providing all nutrients according to recommended dietary allowances and to contain 800–1,000 kcal/day. The diet could either be a liquid formula diet (= meal replacement), a mixed formula/solid diet, or a purely solid diet. We contacted all investigators in order to get information about the type of diet they used. Thirteen out of 22 investigators answered our request and provided partial or full information about the composition of the diet they used (Table 1).

Table 1.  Overview of relative changes in weight in each participating center
inline image

Statistical analyses

Results are shown as means (s.d.). For comparisons of means between multiple groups, a one-way ANOVA was used. Simple bivariate correlations (Pearson's correlations) were used in order to assess for (i) significant relations between number of participants per center and the attrition rate at the end of the LCD; (ii) significant relations between relative changes in weight and the number of participants per center, age, gender, type of diet (meal replacement or solid food), relative contents of carbohydrates, lipids, and proteins of the diets. We present the average weight loss per center as a forest plot, including heterogeneity statistics (χ2 test) (11) and evaluation of inconsistency across centers (I2 index) (12). As the number of participants within each center is clustered, we apply a hierarchical model to model the average weight loss vs. the number of participants included at each center. This was modeled using study-level data, applying a random-effects restricted maximum likelihood model (13). For all analyses, the significance level was determined as P < 0.05. SAS software (SAS Institute, Cary, NC) was applied for all the hierarchical model statistics, whereas SPSS package 16.0 (SPSS, Chicago, IL) was used for all the other analyses.


Twenty-two different centers recruited a total of 701 subjects, among whom 631 completed the 8 weeks on LCD (90%). Each center enrolled between 4 and 85 subjects. The number of participants per center did not correlate significantly with the attrition rate during the LCD (r = 0.06; n = 22; P = 0.11) (Figure 1). The mean weight at enrolment was 109.1 (17.1) kg, waist circumference was 113.8 (12.6) cm, and waist-hip ratio was 0.91 (0.10). Other baseline characteristics are shown in Table 1. Subjects differed significantly among centers in terms of age (P = 0.03) and BMI (P < 0.001).

Figure 1.

Correlation between number of participants recruited per site and attrition rate. CI, confidence interval.

Weight loss at the end of the LCD

Overall, according to the pooled mean value, subjects lost 10.9 (3.9) kg or 9.9 (2.9) % of their initial body weight. Mean weight loss differed significantly among centers, ranging from 5.8 to 11.8% (P < 0.001), as shown in Table 1. Figure 2 shows that the heterogeneity between centers was very high (χ2 = 206.7; P < 0.001), as also indicated by extreme inconsistency (I2 = 91.9% and a between-center variance (τ2) of 2.24). Thus, the heterogeneity statistics raise doubt about relying on a single summary estimate representing the treatment effect for all patients, in all centers. However, applying a restricted maximum likelihood model approach still supports—as anticipated—a statistically significant weight reduction 9.1% (95% CI: 8.5–9.8%; Z = 26.82; P < 0.0001).

Figure 2.

A forest plot showing the average weight loss in each center following LCD, presented in the same order as Table 1. CI, confidence interval; LCD, low-calorie diet.

Figure 3 illustrates the significant association between the number of subjects included per center and the relative change in weight presented as a variance-weighted meta-regression plot. This resulted in a significantly improved model, with less between-center variance (i.e., τ2 = 1.03), corresponding to an I2 of 42%. As shown by the line of prediction, there was a significant slope (0.05 (95% CI: 0.03–0.07); Z = 4.42; P < 0.0001), meaning that the larger the center the more average weight loss expected; i.e., the model predicts that for 10 extra patients in a center, the average weight loss would increase by 0.5% points.

Figure 3.

Meta-regression analysis modeling the hierarchical data structure: the size of the circles is proportional to the precision of the estimate used in the meta-regression. The line indicates the predicted effects (regression line). CI, confidence interval.

No other predictive variables could compete with the number of subjects included per center. However, we also identified significant correlations between the relative change in weight and gender (R2 = 0.06; P < 0.001), type of food (R2 = 0.06; P < 0.001), % of carbohydrates (R2 = 0.01; P = 0.006), and % of proteins (R2 = 0.03; P < 0.001).


To the best of our knowledge, this is the first study that evaluates the association between the number of recruited subjects per site in a multicenter trial and the relative weight loss attained after an 8-week LCD at each site. We observed significant differences between the mean relative weight loss obtained at different sites. The main predictive factor for this difference seemed to be the number of recruited subjects of the site, even if this association remained relatively modest.

We found no data in the literature showing similar findings. In fact, up to now, mainly individual predictors for weight loss have been investigated (3,4,14,15).

The main reason for the correlation between weight loss and the number of recruited subjects per center may be that principal investigators, study coordinators, and dieticians working at centers with high numbers of recruited subjects have more experience than those working at smaller centers; therefore, they are probably more efficient in treating and counseling subjects during an LCD. Here, one could establish a parallel between these findings and the importance of the learning curve in bariatric surgery. In fact, numerous reports have shown that operative times as well as the rates of postoperative complications decline when the experience of a surgeon increases (5,6,7,8,9).

Other reasons might explain the observed correlation. For example, it may be that the larger centers have a larger pool of subjects, and thus recruit the most highly motivated individuals, whereas the smallest recruiting sites maybe used up their prevalent pool of patients and thus are pulling less motivated patients.

Some factors may have biased our analysis. First, our database provides no information about the experience and background of the dieticians, as well as about the way they handled the meetings with participating subjects (individual or group meetings, for example). Second, nine centers did not provide any information about the type of diet they used, which of course limits our analysis. Third, compliance to the prescribed diet may have varied due to cultural differences that could not be assessed in this analysis.

Due to these limitations, our results should be considered as preliminary. Nonetheless, they seem significant to us and allow to raise several important questions. To which extent should study sponsors take the number of recruited subjects per center of a site into account when planning obesity trials that include a preliminary LCD? Should they seek to include fewer, but more experienced sites, so that the number of subjects per site would be greater? This would have the advantage of being less cumbersome in terms of study planning and expenses, and might lead to better weight losses, as our results suggest that the number enrolled per site is something of a proxy for the experience of the site. The disadvantages of such a strategy would be (i) to potentially create a selection bias, as the participants treated in high performing centers might not adequately represent the entire population and (ii) to potentially slow down the recruitment of participants. When planning clinical trials in obesity, sponsors should carefully weigh these advantages and disadvantages while waiting for further data that will allow to draw firmer conclusions on the relationship between number of participants enrolled per site and weight loss.


The data analysis and the writing of the article were supported by a grant from the Program for Exchange in Endocrinology Expertise. The Parker Institute: Musculoskeletal Statistics Unit is supported by an unrestricted grant from the Oak Foundation.


The author's responsibilities were as follows: C.G. was responsible for data analysis and writing the manuscript; R.C. was responsible for data analysis, interpreting data, and revising the manuscript; T.M.L. participated in writing and revising the manuscript; F.V. participated in providing the data set, participated in data analysis, and revised the manuscript; S.T. participated in writing and revising the manuscript; A.A. provided the data set, participated in elaborating the research hypothesis, participated in data analysis, and revised the manuscript. All authors had full access to all of the data in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. C.G., R.C., T.M.L., and S.T. had no personal or financial conflict of interest. At the time of submission of the paper, C.G. is an employee of Merck Serono SA, Geneva, Switzerland. F.V. is an employee and shareholder of Johnson & Johnson. A.A. was a consultant for Johnson & Johnson in relation with the development of topiramate and other antiobesity compounds between 1998 and 2004.