|Home | About | Journals | Submit | Contact Us | Français|
Clinical trials of obesity treatments have been limited by substantial dropout. Participant-level variables do not reliably predict attrition, and study-level variables have not yet been examined. We searched MEDLINE and identified 24 large randomized controlled trials of weight loss medications. These trials were comprised of 23 placebo and 32 drug groups. Two authors independently extracted the following for each treatment group: treatment received; design characteristics (inclusion of a lead-in period, selection of participants with weight-related comorbidities, study location, and number of study visits); sample characteristics (sample size, % female, and mean baseline age and BMI); and attrition (total, adverse event [AE] related, and non-AE-related) at 1 year. The primary outcome was total attrition, which was significantly related to treatment (i.e., 34.9%, 28.6%, 28.3%, and 35.1% in placebo, orlistat, sibutramine, and rimonabant groups, respectively, p < .0001). In adjusted multivariable models, total attrition was significantly lower in groups that completed a pre-randomization lead-in period than in those that did not (29.1% vs. 39.9%, p < .01). Gender also was significantly related to total attrition; groups with more women had higher dropout (p < .01). The pattern was similar for predicting non-AE-related attrition. Findings suggest ways to design studies that maximize retention.
Attrition is the bane of clinical trials. Nowhere is this problem more apparent than in trials of weight loss medications, as underscored by two recent commentaries.1-2 Although there are multiple analytic methods for addressing the problem of missing data, none is a replacement for measuring the weight of all randomized participants, and each has significant shortcomings.
Completers-only analyses, for instance, can result in an inflated estimate of efficacy because lack of weight loss may be a reason for attrition. Thus, most large trials report results from intent-to-treat (ITT) analyses, with the last observation carried forward (LOCF) in the case of missing data. LOCF imputation, however, does not account for the likelihood of regaining lost weight after discontinuing treatment and can overestimate weight loss efficacy.2-3 A more conservative imputation method – a modification of LOCF – assumes a constant rate of weight regain (e.g., 0.3 kg/month) and figures it into the imputed value.4-5 This procedure can lead to an over- or underestimation of efficacy, depending on the timing of, and reason for, dropout. Perhaps the most stringent imputation procedure is known as baseline observation carried forward (BOCF), which assumes no change from randomization for participants who drop out of the study.1 However, given the high rate of attrition from many trials of pharmacological weight loss agents, this imputation method may underestimate a drug’s effect. Model-based imputation methods assume that missing data are due to random factors and, thus, are also significantly flawed.1
Limiting attrition is, therefore, an important consideration in obtaining reliable and valid study results. If persons at greatest risk of dropping out can be identified, they may be targeted for proactive retention efforts. Although individual studies have identified some demographic and behavioral characteristics that predict attrition, results have been equivocal.6-9 An alternative approach would be to design studies in ways to maximize retention. It is unknown, however, whether study design characteristics – for example, including a pre-randomization lead-in period or targeting a certain patient population – can limit attrition. A systematic review and synthesis of the research is needed.
The present investigation summarized rates of attrition from large randomized controlled trials of weight loss drugs with adult participants and at least 1 year’s duration. The primary objective was to identify treatment, design, and sample characteristics that were associated with attrition from all causes. In secondary analyses, the same variables were examined as predictors of attrition that was either related or unrelated to adverse events. Adherence to treatment is a separate (but related) issue that is infrequently reported in clinical trials and, thus, is not addressed in the present study.
Published studies were identified using the MEDLINE electronic database. Trade and generic names of frequently studied weight loss medications (including agents used off-label for weight loss) were entered as search terms. These terms were intersected with terms related to obesity and weight loss. Only human studies, published in English, before September 2007 were returned. (The detailed search strategy is available from the first author upon request.) This search yielded 1,777 published papers, including original research articles, reviews, case studies, and editorials.
Only randomized controlled trials were considered for inclusion in the present study. Additional inclusion criteria were: 1) participants were overweight or obese adults (age ≥ 18 years); 2) the intervention included administration of orlistat, sibutramine, rimonabant, phentermine, fenfluramine, the phentermine-fenfluramine combination, dex-fenfluramine, topiramate, or bupropion; 3) at least 100 participants were randomized into each study condition; 4) change in weight (expressed in kg or % of initial weight) was reported; 5) attrition was reported; and 6) study duration was at least 1 year post-randomization. One author (EAG) examined study abstracts and/or texts to determine whether studies met inclusion criteria, and another author (ANF) reviewed the first’s determinations to verify eligibility.
Twenty-nine studies, published from 1989 to 2006, met all inclusion criteria. The study drugs included orlistat, sibutramine, rimonabant, topiramate, and dex-fenfluramine. We decided to exclude the dex-fenfluramine and topiramate trials, because there was only one study of each.10-11 Additionally, we excluded two orlistat studies and one sibutramine trial because they tested the drugs’ efficacy for weight loss maintenance, not weight loss.12-14 Thus, 24 randomized controlled trials of orlistat, sibutramine, or rimonabant were included in the analyses described below.15-39 (Although rimonabant has been withdrawn from the market, we retained the rimonabant studies because they were well designed and they met all inclusion criteria for our investigation.) All but one study was placebo-controlled; the excepted study varied the calorie target of the accompanying diet plan.36 For another investigation, data were extracted from two separate papers.30-31
Two authors (ANF and MLB) independently reviewed each paper in detail to extract data for each treatment condition across the 24 included studies. The independent variables of interest were treatment (i.e., placebo, orlistat, sibutramine, or rimonabant) and various study design and sample characteristics. Study design characteristics included: 1) location (North America, Europe, or multiple/other continents); 2) whether participants were selected based on the presence of a weight-related comorbidity (e.g., type 2 diabetes); 3) presence or absence of a pre-randomization lead-in (i.e. “run-in”) period; 4) number of post-randomization study visits; and 5) number of post-randomization visits in which lifestyle counseling was provided (i.e., “lifestyle visits”). Sample characteristics included mean age and body mass index (BMI) at baseline, as well as sample size and the percentage of female participants in each study group. Sample characteristics were based on the ITT population of each study group, for whom those characteristics were reported. (Most included studies defined the ITT population as those participants who took at least one dose of study medication and returned for at least one visit after randomization. Although this definition of ITT is clearly not ideal, we have elected to use it to remain consistent with the language of the included trials.)
The dependent variables extracted pertained to study attrition at 1 year after randomization. The primary outcome variable was total attrition from all causes, which was calculated as number of dropouts at 1 year divided by number of randomized participants. Attrition related to adverse events (i.e., AE-related attrition), when reported, was extracted separately. From total and AE-related attrition, we calculated attrition not related to adverse events (i.e., non-AE-related attrition). AE-related and non-AE-related attrition were secondary outcomes.
A summary of each included study is provided in tabular form. Means and standard errors of all continuous variables and frequencies of categorical variables are provided in the results section. In all inferential analyses, the unit of analysis was the treatment group.
To model the fact that each study contributed two or three groups to these analyses, mixed models were fit to: 1) compare baseline sample characteristics (i.e., % female, age, and BMI) between treatments (i.e., placebo, orlistat, sibutramine, and rimonabant); and 2) determine relationships of treatment, design, and sample characteristics to total, AE-related, and non-AE-related attrition. All mixed models included one level of random effect to adjust for study variability and at least one fixed effect to adjust for sample size. Where treatment was not the independent variable of interest, analyses included a second fixed effect to adjust for treatment. Least square means ± standard errors were derived from the mixed models and are presented for all analyses.
Continuous characteristics (% female, age, weight, BMI, and number of postrandomization study and lifestyle visits) were first included in the mixed models in continuous form to assess significant relationships with attrition. To characterize important relationships, β ± standard errors are presented for continuous variables that were significantly associated with attrition (p < .05) or those that showed a trend (p < .10). Additionally, those variables were dichotomized using median split, and the least square means ± standard errors of attrition for the lower and upper 50th percentile of these continuous variables are presented.
The final multivariable mixed models presented are a result of two modeling steps. The first step consisted of a mixed model that simultaneously included, in addition to adjustment for sample size and treatment group, all variables that were found to be significantly (p < .05) or marginally (p < .10) associated with the attrition outcome in univariable analyses. The final multivariable model included only those variables that were significantly associated (p < .05) with the outcome in the initial multivariable model. All analyses were conducted using the statistical software package, SAS 9.1 (SAS Institute, Cary, NC, USA).
Of the 24 studies included, 16, 4, and 4 were trials of orlistat, sibutramine, and rimonabant, respectively. Seven were conducted in North America, 12 in Europe, and 5 in multiple or other continents. In 11 studies, all participants were selected for the presence of a weight-related comorbidity. Eighteen studies included a lead-in period, which, in every instance, included single-blind administration of a placebo (i.e., the participant was blinded) for 2 to 5 weeks prior to randomization. The mean ± standard error length of the lead-in period of 2.5 ± 0.4 weeks, including six studies with no lead-in period. The mean number of post-randomization study visits during year 1 was 13.5 ± 0.9. The number of post-randomization lifestyle visits (i.e., visits in which diet/exercise counseling was provided) could be extracted from 12 studies. These trials included a mean of 11.8 ± 1.9 such visits in year 1. A summary of study characteristics is shown in Table 1.
In total, 18,918 participants were included in 23 placebo groups and 32 active drug groups. Across treatment groups, the mean age of participants was 47.9 ± 0.8 years, and BMI was 35.5 ± 0.2 kg/m2. On average, 69.4 ± 1.9% of participants were women. Table 2 displays sample characteristics for groups that received placebo, orlistat, sibutramine, and rimonabant, regardless of dose.
The mean total attrition rate across the 55 treatment groups was 32.8 ± 1.6%. In the 48 groups from which AE-related attrition and non-AE-related attrition could be extracted, means were 7.8 ± 0.6% and 26.7 ± 1.5%, respectively.
Results from univariable analyses are summarized in Table 3. All models were adjusted as described above.
The treatment participants received was significantly related to all attrition outcomes. Total attrition was significantly higher in groups that received placebo than in those that received orlistat and sibutramine. Rimonabant groups differed only from orlistat groups; rimonabant groups had significantly greater total attrition.
Placebo, orlistat, and sibutramine groups all had similar rates of AE-related attrition. AE-related attrition was significantly higher in rimonabant groups than placebo and orlistat groups.
Non-AE-related attrition was significantly higher in placebo groups, compared with orlistat and sibutramine groups, which did not differ from each other. Rimonabant groups did not differ from any other groups with respect to non-AE-related attrition.
There was significantly lower total and non-AE-related attrition in study groups that included a lead-in period than in those that did not. Presence or absence of a lead-in period was not associated with AE-related attrition.
Selection of participants based on the presence of a weight-related comorbidity was related to marginally lower total attrition and significantly lower non-AE-related attrition. AE-related attrition did not differ by whether or not a weight-related comorbidity was an inclusion criterion.
The location in which the study was conducted was marginally associated with total attrition. Location, however, was unrelated to both AE-related and non-AE-related attrition.
None of the 1-year attrition outcomes was related to the number of post-randomization visits included in the first year. Furthermore, no attrition variables were related to the number of visits in which lifestyle counseling was provided.
The percentage of women in the study group (measured continuously) was significantly related to total attrition (β = 0.33 ± 0.13) and non-AE-related attrition (β = 0.37 ± 0.13). To further characterize these relationships, the percentage of women was dichotomized using the median split procedure (median = 77.8%) into low (mean = 57.3%) and high (mean = 82.0%) groups. Subsequent comparisons showed that low-percentage female groups had significantly less total attrition (28.9 ± 2.6% vs. 34.3 + 2.5%) and non-AE attrition (22.0 + 2.8% vs. 28.2 + 2.6%) than high-percentage female groups (ps < .05).
Mean age of study groups, as a continuous variable, was unrelated to all attrition outcomes. Thus, we did not perform median splits to test low vs. high groups.
Although mean BMI at baseline (measured continuously) was not significantly related to total attrition, it was associated with both AE-related (β = - 0.86 ± 0.40) and non-AE-related attrition (β = 2.51 ± 1.19). However, after using the median split procedure (median = 35.7 kg/m2) to categorize BMI into low (mean = 34.3 kg/m2) and high (mean = 36.8 kg/m2), we found no significant differences between low- and high-BMI groups in AE-related attrition (9.1 ± 0.9 vs. 7.3 ± 1.0, p = .18) or non-AE attrition (23.3 ± 2.9 vs. 28.0 ± 3.1, p = .23).
All predictors that were significantly or marginally related to total or non-AE-related attrition in the univariable analyses were entered simultaneously into an initial multivariable model for that type of attrition. Both models were adjusted for treatment and sample size (which also were the standard covariates included in the univariable models). Final multivariable models, adjusted for the same covariates, included predictors that significantly predicted attrition in the initial multivariable models. Results are summarized in Table 4. Multivariable models for AE-related attrition were unnecessary because mean BMI was the only significant univariate predictor, other than treatment.
The initial multivariable model for total attrition included presence/absence of a lead-in period, whether a comorbidity was an inclusion criterion, study location, and percentage of women in the sample. The final multivariable model suggests that, when adjusting for all measured factors related to total attrition, the presence of a lead-in period and having a smaller percentage of women in the sample were significantly related to lower total attrition.
The initial multivariable model for non-AE-related attrition included presence/absence of a lead-in period, whether a comorbidity was an inclusion criterion, percentage of women in the sample, and mean sample BMI. As was true of total attrition, the final multivariable model for non-AE-related attrition indicated that groups that had a lead-in period and those that had a lower percentage of women had significantly lower non-AE-related attrition after adjustment for all other relevant predictors.
This research represents the first attempt to characterize factors related to attrition across a variety of randomized controlled trials of weight loss drugs. Previous studies have focused exclusively on participants’ psychosocial or demographic characteristics within individual trials and have yielded largely mixed results.
In our analysis of group-level and study-level variables, we found that treatment was strongly related to total attrition. Dropout rates were highest among groups treated with placebo or rimonabant, and lowest among those treated with orlistat or sibutramine. Adverse events appear to have contributed to the equivalence between placebo and rimonabant groups with respect to total attrition. Rimonabant groups had 5% greater AE-related attrition and 5% lower non-AE-related attrition than placebo groups. We suspect that the relationship between treatment and total attrition was, in part, a function of weight loss; participants who lost lest weight may have been more likely to drop out. Unfortunately, however, we could not control for weight loss in our analyses for two reasons: 1) LOCF weight loss data are confounded by attrition; and 2) multicollinearity between weight loss and treatment was too high to include both predictors in the same analyses.
Inclusion of a lead-in period was a study design feature that was significantly related to lower total attrition in both univariable and fully-adjusted multivariable analyses. These lead-in periods consisted of single-blind administration of a placebo for 2 to 5 weeks. The majority of lead-ins also included a low-calorie diet prescription. In many studies, however, the only stated criterion to continue on to randomization was participants’ achieving ~70% medication (i.e., placebo) adherence, as determined by unused doses returned to the clinic. Regardless of its length, nature, or requirements, the inclusion of a lead-in period allows for more rigorous evaluation of potential participants than is possible with a telephone screening or a single on-site visit. Candidates with suboptimal motivation, commitment, or availability can be excluded prior to randomization. The disadvantage of this approach is that it limits researchers’ ability to generalize findings to less adherent individuals.
A lead-in period also presents candidates a truer picture of study participation (i.e., requirements) than the informed consent document can provide. For example, the Look AHEAD study – a large randomized trial of lifestyle modification in adults with type 2 diabetes – included an intensive screening process that required candidates to keep detailed records of food intake and physical activity for 2 weeks.40 This lead-in period simulated an important aspect of study participation (i.e., self-monitoring of energy intake and expenditure). Approximately 5% of Look AHEAD candidates failed this behavioral criterion,41 which may well have reduced attrition by a similar amount.
The selection of participants with a weight-related comorbidity was marginally related to lower attrition in univariable analyses but was not related in more fully adjusted models. Attrition may have been somewhat lower in participants with weight-related comorbidities because of their greater perceived urgency for weight loss and lifestyle change, or because of stronger support from family members and health care providers. Concern about the health consequences of obesity may have increased their threshold for side-effects and other potential hassles related to study participation, such as attending clinic visits. However, the effect of this study design feature was not strong enough to predict attrition when adjusting for lead-in and gender composition of the sample.
Of the sample characteristics we examined, the percentage of females treated was the only significant predictor of total attrition in fully adjusted models. This relationship appeared primarily attributable to non-AE-related attrition because gender composition of the sample was not associated with AE-related attrition. Although gender was often a constant or was not examined in previous studies that sought to identify predictors of attrition,8-9 at least one investigation found that women were more likely than men to drop out of a weight loss trial.7 The present findings suggest that targeting males for recruitment in weight loss studies may not only enhance generalizability of findings, but also improve retention.
We were surprised that neither the number of study visits, nor the number of visits in which lifestyle modification counseling was provided, was related to attrition. We expected that more frequent contact with the clinic and more intensive lifestyle counseling would facilitate retention, as these factors were associated with greater weight loss in some pharmacotherapy trials.4,42 Perhaps the costs of visiting the clinic (e.g., time, effort, money) balanced whatever benefits the increased contact may have afforded. Additionally, the lack of association between study visits and attrition may be attributed to limited variability in the former. Nineteen of the 24 included studies had at least monthly visits and only one averaged more than two visits per month. We also note that most reports provided a limited description of the adjunctive lifestyle modification offered, which precluded our considering the quality of the intervention in our analyses. Simple prescriptive advice was coded as equivalent to protocol-driven behavioral or dietary counseling. We suspect that the use of empathic staff (e.g., dietitians) who are trained and supervised in the delivery of a sound, structured intervention, may reduce attrition from weight loss trials. Several large studies that included a lifestyle modification component and followed this approach (e.g., Look AHEAD, the Diabetes Prevention Program, PREMIER) retained well over 90% of participants at 1 year and beyond.41,43-47
The chief liability of our report is that several factors that may have accounted for significant variance in attrition (e.g., socioeconomic status, psychosocial stress, interactions with study staff, compensation for participation, efforts to increase retention) were not measured or reported in most of the studies we examined. Thus, we could not include all potentially relevant predictors in our analyses. Additionally, our investigation did not address the issue of adherence, which, like participant retention, is critical to evaluating the efficacy of an intervention. Unfortunately, clinical trials rarely report adherence variables, such as percentage of medication doses taken or visits attended.
Although our use of group-level data reduced power and may have masked significant findings, this approach can be considered a strength of our investigation. By examining group-level data, we were able to look beyond individual demographic or psychosocial predictors of attrition to examine the roles of study design features and treatment. We also examined AE-related and non-AE-related attrition separately, under the assumption that the latter is more modifiable than the former. This separation not only emphasized that non-AE-related attrition (e.g., lost to follow-up, withdrawn consent) is the greater contributor to total attrition but also that non-AE-related attrition is more responsive to study design features than is AE-related attrition.
Future studies may seek to replicate the present findings in randomized controlled trials of medications used to treat other medical conditions. Given the resources and effort that sponsors, clinicians, and patients all invest in clinical research, reducing attrition is in the interest of all parties. Moreover, the analytic and interpretive problems posed by attrition underscore that maximizing retention of participants is in the best interest of clinical science.
This work was supported in part by a Merck grant to Fabricatore and by National Institutes of Health (NIH) grants K23 DK070777 and K24 DK065018 to Fabricatore and Wadden, respectively. Fabricatore has served as a consultant to Pfizer and has received research support (for this study) from Merck. Wadden serves on the Advisory Boards of Merck, Novo Nordisk, and Orexigen and has received research support from Merck, Orexigen, Pfizer, and Vivus. Erondu, Heymsfield, and Nguyen are employed by Merck. The other authors have no potential conflicts to declare.