Predictors of weight loss outcomes in obesity care: results of the national ACTION study

This cross-sectional study, sponsored by Novo Nordisk Inc. and approved by an Institutional Review Board [16], was conducted in 2015 (from October 29th to November 12th) among US adults age ≥ 18 with obesity (defined as self-reported body mass index (BMI) of 30 kg/m² or higher). PwO responded to an online survey which assessed obesity-related attitudes and behaviors and was developed based on literature review and qualitative research [17]. Respondent-level weighting was applied [18] to ensure the sample was representative of the U.S. population [19]. Except for participant characteristics, the data presented in this paper are weighted unless otherwise specified. A detailed methodological description has been previously published [15]. Five-point end-anchored scales assessed agreement, where “1” meant “do not agree at all,” and “5” meant “completely agree.” Responses of “4” or “5” were coded and reported as “agree” unless otherwise noted.

A self-reported weight loss success outcome was defined within the survey instrument explicitly. A multivariate logistic regression model assessed variation in “weight loss success” (dependent variable) defined as: 1) Weight loss history: at least 10% weight loss in the previous 3 years; and 2) Success: weight loss at the time of survey response that was maintained for at least 1 year (by respondent self-report).

Bivariate associations between the outcome of interest and 140 possible independent variables were assessed and grouped into 3 domains: demographic, attitudinal, and behavioral. Independent variables with large bivariate effect sizes and significant practical implications were identified. Variance inflation factors were used to assess the degree of multicollinearity present among the remaining 32 independent variables; no variables were removed at this step.

To achieve a parsimonious model, a purely statistical approach was used to reduce the inputs. A Bayesian variable selection approach was used to overcome the biases and shortcomings of stepwise variable selection. A logistic regression model was estimated using the nine remaining independent variables; two were removed for non-statistical significance and high degree of correlation with other independent variables (feelings after most recent discussion of weight with HCP: supported; barriers to initiating a weight loss effort: my lack of motivation, respectively). Seven characteristics taken from demographic, attitudinal and behavioral survey domains were included in the model. Six of the seven variables included in the final model were statistically significant at the 5% level of significance; discussing weight with a diabetes educator was not statistically significant but increased the stability of the model.

A sensitivity analysis was conducted to determine the independent impact each variable in the model has on the average chance of being successful in losing weight and keeping it off for 1 year. To conduct the sensitivity analysis, we implemented the following procedure for each independent variable in the model: 1) for categorical variables, a 1000-iteration bootstrap sample (a technique of randomly sampling the data and iteratively calculating the statistics to generate more accurate population estimates) was performed to calculate the mean predicted probability of sustained weight loss success for incremental improvements and deteriorations of 1 to 100%; 2) for continuous variables, the mean predicted probability of sustained weight loss success was calculated for incremental improvements and deteriorations of 1 to 100%.

Model results are described in Table 2. The odds of sustained weight loss success are compared to a base case PwO who weighed 248 pounds 1 year ago and 221 pounds 10 years ago, and who has been formally diagnosed with obesity, reports their HCP not often recognizing the PwO’s previous weight management efforts, does not agree that he/she is motivated to lose weight, does not agree that his/her lack of motivation is a barrier to initiating a weight loss effort, and has not discussed or would not want to discuss weight with a diabetes educator.

The model’s statistical fit was assessed by using the model diagnostics of accuracy (percentage of the time the model accurately classifies a PwO as successful in achieving sustained weight loss), specificity (percentage of failed weight loss attempts accurately predicted by the model), and sensitivity (percentage of successful weight loss attempts accurately predicted by the model). Individual observations were assessed for their degree of leverage/influence and residuals were assessed via a studentized residuals vs. fitted probability plot. Higher leverage points were dispersed throughout the fitted probability space – these points were not overly influencing prediction to one class or the other. The points with the highest residual error were all cases of successful weight loss and occurred at the lower end of the fitted probability scale (i.e. 0.0–0.2) indicating that a low probability threshold should be used for classification of successful weight loss. A threshold of 0.11 was identified as the boundary that best-balanced model accuracy (70%), sensitivity (68%), and specificity (70%). Two-thousand (2000) splits of the data into 80% training and 20% test showed stable model performance at this threshold with values as shown in Table 2.

Self-reported weight 1- and 10- years prior were both significantly associated with sustained weight loss success, controlling for beginning and ending weight (1 and 10 years prior, respectively); i.e., the model controlled for variation associated with varying weights. On average, each additional pound of weight 1 year ago predicted a decrease in the success odds by ~ 1%; however, each additional pound of weight 10 years ago increased the odds of success by a factor of 1.02 (increasing the odds by ~ 2%).

Not formally being diagnosed with obesity was associated with a decrease in PwO’s odds of sustained weight loss success by a factor of 0.64. PwO who felt their lack of motivation was a barrier to weight loss had lower odds of reporting success, even after controlling for other factors such as their changes in weight and attitudes towards initiating weight loss.

Discussion with a diabetes educator (or willingness to discuss) was associated with increased odds of sustained weight loss success by a factor of 1.22. HCPs’ recognition of PwO’s previous weight loss efforts and PwO’s self-reported motivation to lose weight had the greatest impact on the odds of success, increasing them by a factor of 1.99 and 1.81, respectively.

To further illustrate how the odds of successful sustained weight loss may be manifested at the individual level, we explored several hypothetical scenarios (Table 3).

The presence of PwOs' motivation in the absence of a formal obesity diagnosis or recognition of previous weight management efforts (“Inactive HCP; Motivated PwO”) was associated with a predicted 19% decrease in the rate of sustained weight loss success compared with the population average. A PwO who was unmotivated to lose weight but receives a formal diagnosis, has a discussion with a diabetes educator, and receives recognition from an HCP (“Active HCP, Unmotivated PwO”) is predicted to have a greater rate of success, with a 6% increase over the population average prediction. Finally, PwO who receive recognition of previous weight loss efforts from HCPs and report feeling motivated (“Encouraging HCP; Motivated PwO”) are more likely to have success, with a 98% increased likelihood compared with the population average.

In the sensitivity analysis described above, we found that increasing obesity diagnoses rates and discussions with a diabetes educator have the least independent influence on the predicted success rate across the sample. Increasing the rate of diagnosis or discussions with a diabetes educator by 50% results in an absolute change of < 1% point from the baseline prediction of sustained weight loss success of 11.5% (baseline prediction is the model’s estimated sustained weight loss success at the means for each independent variable). A decrease in the prevalence of PwO’s self-reported motivation or a reduction in the prevalence of motivation as a barrier by 50% results in an absolute change of 1.4% points and 1.1% points from the baseline prediction, respectively. Increasing the proportion of PwO whose previous weight management effort is recognized by HCPs by 50% improves the average chance of success by 1.8% points.

Although the absolute point changes in the average predictive probabilities are modest, these translate to substantial population level increases (number of PwO) relative to the average baseline rates of weight loss success (Fig. 1). Increasing the rate at which a PwO’s previous weight management efforts are recognized by an HCP by 25% is predicted to result in an 8% improvement in the rate of success. The same level of increase in a PwO’s self-reported motivation to lose weight is associated with a 6% improvement in the weight loss success rate.

Looking outside the model, we further sought to understand if weight loss discussions and attitudes towards them are associated with the BMI of PwO. Evaluating PwO by obesity class (Class I: BMI 30- < 35, Class II: BMI 35- < 40, Class III: BMI ≥ 40) [20], we found significant differences in attitudes and behaviors. PwO with Class III Obesity were significantly more likely (30%, 95 CI [26–35%]) than those with Class I (21%, 95 CI [18–24%], p < .001) or Class II Obesity (22%, 95 CI [18–25%], p < .001) to report having sought support from their HCP regarding weight management after they have tried losing weight on their own with no success. PwO with Class I or Class II Obesity (39.1%, 95 CI [35–43%] and 39%, 95 CI [35–43%], respectively) were significantly more likely than those with Class III Obesity (31%, 95 CI [27–35%]) to report not seeking support from their HCP (p = .002 and p = .003, respectively). Conversations with HCPs differed by obesity class; when asked who typically brings up weight during the appointment, PwO with Class I Obesity were significantly more likely to say they, rather than their HCP, typically started the conversation about their weight (51%, 95 CI [47–55%] than those with Class II (45%, 95 CI [41–50%, p = 0.026] and Class III Obesity (43%, 95 CI [39–47%], p = .002. Those with Class II and III Obesity were significantly more likely than those with Class I Obesity to report negative feelings of “embarrassed,” “discouraged,” and “blamed” after their most recent discussion with their HCP. When asked how their HCP could better support them in achieving a healthy weight, PwO with Class II and Class III Obesity were significantly more likely than those with Class I Obesity to cite “be more understanding of the challenges of losing weight,” “be more understanding of the challenges with living with obesity,” and “refer me to a dietitian.”