Weight loss interventions on health‐related quality of life in those with moderate to severe obesity: Findings from an individual patient data meta‐analysis of randomized trials

The relationship between BMI and health‐related quality of life (HRQoL) critically affects regulatory approval of interventions for weight loss, but evidence of the association is inconsistent. A higher standard of evidence than that available was sought with an IPD meta‐analysis of 10,884 people enrolled in five randomized controlled trials of intentional weight loss interventions. Cross‐sectional and longitudinal associations of BMI and HRQoL were estimated in mixed effects models specifying a latent variable for HRQoL. Spline regressions captured nonlinear associations across the range of BMI. In cross‐sectional spline regressions, BMI was not associated with HRQoL for people with a BMI < 30 kg/m2 but was for those with a higher BMI. In longitudinal spline regressions, decreases in BMI were positively associated with HRQoL for people with a BMI ≥ 25 kg/m2. The impact of change in BMI was larger for people with higher BMIs than for those with BMIs under 30 kg/m2. Lower BMI and decreases in BMI were related to higher HRQoL for people living with obesity but not in the population without excess weight. HRQoL gains from weight loss are greater for more severe obesity. Commissioners should use these estimates for future decision making.

Across the world, health economic appraisals of behavioral interventions, pharmacotherapeutic interventions, and bariatric surgery have estimated the impact of the intervention on quality of life by assuming that cross-sectional differences between people of different BMIs represent the effect of changing weight on quality of life. [10][11][12][13][14][15] Studies have used estimates that differ markedly. Different values could alter decisions on whether to commission these interventions. In the United Kingdom, recent appraisal of liraglutide by the National Institute for Health and Care Excellence (NICE) used values from a cross-sectional, observational study. 7 Here, a nonlinear function of BMI was specified. HRQoL, expressed in utility, increased with BMI up to an inflection point (around BMI = 25 kg/m 2 ); HRQoL declined thereafter. This is in stark contrast to prior approaches in the United Kingdom and internationally that have assumed a single value across the range of BMI. Another issue is that gains in HRQoL from weight loss are often assumed to be sustained, which may not be the case, particularly after nonsurgical interventions. More robust estimates of the HRQoL-BMI relationship and the sustainability over time would enable consistent commissioning decisions.
In this paper, we estimate the association between BMI and HRQoL using longitudinal data from five large RCTs of behavioral weight loss interventions in an individual participant data (IPD) meta-analysis, where weight loss can be assumed to be intentional and not arising from disease. We assess whether the association varied over the distribution of BMI and whether the strength of association was different for the physical and mental domains of HRQoL. The latter in particular is of interest given the lack of existing evidence on this issue.

| Trial identification
Studies were identified using existing systematic reviews in obesity, 2,16 reference mining of systematic reviews, and consultation with experts in the field.

| Trial eligibility
(1) Adults (18 years or older) with overweight or obesity at baseline (BMI 25 kg/m 2 or above) and enrolled in a behavioral weight loss trial, indicating intention to lose weight. For trials that used behavioral and pharmacotherapeutic arms, only data from the behavioral arms were analyzed. Pharmacotherapeutic trials were excluded to avoid any additional impact on individuals' quality of life, for example, through side-effects. Surgical trials were excluded due to the higher weight of patients undergoing surgery, the substantially larger reduction associated with the intervention and potential for other effects on quality of life not directly related to weight loss. (2) Mean change in weight in the intervention groups ≥2 kg. (3) The main outcomes, participants' BMI and HRQoL, collected at least twice. (4) Trials of more than 1,000 participants were enrolled to ensure that each trial has at least some bearing on the overall relationship of interest. Details of the five trials included-Look AHEAD, TOHP, DPP, WRAP, and DioGENES-are presented below and in detail elsewhere. [17][18][19][20][21] 1 Diets with High or Low Protein Content and Glycemic Index for Weight-Loss Maintenance (DioGENES) The population studied was overweight and obese adults in EU countries that had recently lost at least 8% of their bodyweight. A total of 1209 adults participated in the trial. Interventions were ad libitum diets in a two-by-two factorial design (low/high protein, low/high glycemic index); the control was a diet based on countries' general guidance without reference to glycemic index. The outcome was weight regain. The intervention lasted for a 26-week period; quality of life was collected up to this point. The HRQoL indicator used the obesity-specific Impact of Weight on Quality Of Life (IWQOL).

Diabetes Prevention Programme (DPP)
The population studied was adults in the United States with elevated fasting and post-load plasma glucose. A total of 3234 adults were enrolled. Interventions were a weight loss-focused lifestyle intervention and metformin (we did not include this arm-see eligibility criteria); the control was a placebo. The outcome was the incidence of diabetes. Mean follow-up was 2.8 years. The HRQoL indicator used was the SF-36.

Look AHEAD
The population studied was adults in the United States being between 55 and 76 years of age, being diagnosed with type II diabetes, and being overweight. A total of 5145 patients participated in the trial. Interventions were intensive lifestyle interventions (increased physical activity and reduced calorie intake); the control was diabetes support and education. Outcomes were adverse effects including death and AMI, stroke, or angina. Mean follow-up was 9.6 years. HRQoL indicators included SF-36, feelings thermometer, HUI2, and HUI3.

Trials of Hypertension Prevention (TOHP)
The population studied was adults in the United States being between 30 and 54 years of age, with diastolic blood pressure between 80 and 89. A total of 2182 patients participated in the trial.
Interventions were lifestyle interventions (weight reduction, sodium reduction, and stress management); the control was non-intervention.
In addition, four nutritional supplements were compared, double-blinded, to a placebo. Outcomes were changes in blood pressure. Follow-up was 18 months. The HRQoL indicator used was the general well-being scale. 5 Extended and standard duration weight-loss program referrals for adults in primary care (WRAP) The population studied was adults in the United Kingdom with BMI over 28 kg m À2 in primary care. Interventions were 12-and 52-week weight-management programs; control was brief advice and self-help material. The primary outcome was weight after 12 months of follow-up. Mean follow-up was 1.5 years. HRQoL indicators included the 3-level EuroQol Five Dimension (EQ-5D) and the EuroQol visual analogue scale (EQ-VAS).

| Outcomes
The main outcomes were indicators of HRQoL. These were the SF-36 pcs (physical component summary) and SF-36 mcs (mental component summary), Feelings Thermometer, IWQOL, EQ-5D, EQVAS, and general well-being schedule. All HRQoL outcomes are composite scores except for the EQVAS and the Feelings Thermometer. A description of these indicators is provided in Appendix A. The trials varied in the indicators of HRQoL collected, with no common measurement. LookAHEAD collected four indicators, WRAP two, and the remaining three trials collected one indicator each.

| Statistical analyses
HRQoL was treated as a latent variable within a system of equations.
Measurement equations related HRQoL, as an independent variable, to the various indicators of HRQoL across the trials, each of which was separately treated as a dependent variable. Structural equations were specified with the latent variable for HRQoL as the dependent variable. Appendix B presents the rationale for this and presents the models. Mixed effects models, based on a pre-registered analysis plan, were developed. 22 First, HRQoL was regressed on BMI, the group (within-individual) mean of BMI, individual characteristics, trial fixed effects, and time-from-baseline fixed effects. (A model with intervention arm fixed effects was tested, but the main coefficient of interest was consistent; the reported specification was retained to make use of data of all individuals in all arms, all of whose BMI/HRQoL varied.) A random effect for individuals was also specified. Look AHEAD had longer follow up than all of the other trials. We therefore specified time fixed effects only for time periods in which multiple trials observed participants, to avoid attributing time series HRQoL variation in Look AHEAD to other trials. Second, a spline regression was estimated. The range of BMI was partitioned into five regions, each corresponding to one of five BMI categories: healthy weight, overweight, obesity category I, obesity category II, and obesity category III (underweight was omitted as there were no underweight participants enrolled in these trials). We examined the HRQoL-BMI relationship not only at different starting levels of BMI but also at different starting HRQoL levels (see Appendix H). In cross-sectional analyses, all available observations of all individuals in all time periods were analyzed. Next, two similar regressions were estimated, but here, HRQoL was regressed on period-to-period BMI change to examine the relationship between HRQoL and longitudinal variation in BMI.

| Revisiting previous commissioning decisions
Our fitted model was used to repeat the cost-effectiveness of orlistat that informed NICE commissioning 12 using our updated HRQoL-BMI estimate. The VAS scale was used to derive utility as per the original study. All other metrics were taken from the original paper and held constant.

Levels of missing HRQoL and BMI data varied across trials
(Appendix E). An initial imputation exercise was conducted on each of the data sets separately to gauge the impact of missing data on the estimated relationship of interest. In all cases, the multiply imputed regression results were very similar to those without imputation (Appendix E). Characteristics of individuals at baseline are presented in Table 1.

| Descriptive statistics
A total of 8881 individuals were analyzed at baseline. For DPP, data were available at baseline, but only in an aggregated form, which is not compatible with this analysis. Mean BMI at baseline was 33.3 kg/m 2 . There were slightly more females than males (53%) with a mean age of 53 years. Individuals were predominantly white (76%) with around 13% black; this is broadly in line with census data from the trials' countries.

| Cross-sectional associations of HRQoL and BMI
In a linear model, BMI and HRQoL were negatively associated. A one-unit lower BMI was associated with a 0.13-standard deviation unit higher HRQoL (BMI = À0.13; 95% CI: À0.14 to À0.12) (

| Associations of change in BMI with HRQoL
In a model examining BMI change and HRQoL, there was a negative association. A one unit decrease in BMI was associated with a 0.09-standard deviation unit higher HRQoL (change in BMI = À0.09; 95% CI: À0.10 to À0.08) ( Table 4). In a delta spline regression, there was no evidence of association between HRQoL and changes in BMI when BMI was below 25. At higher BMI, there were significant inverse associations. The association was modest in BMI range 25-29.9 kg/m 2 at À0.09 (95% CI: À0.13 to À0.04) and was progressively stronger in each 5-point higher grouping (Table 4). Above a BMI of 40 kg/m 2 , a one unit BMI increase was associated with a À0.15 lower HRQoL; (95% CI: À0.23 to À0.07).
Here, a one-unit loss in BMI would relate to higher HRQoL on its

| Reconciling past commissioning decisions with updated BMI-HRQoL estimates
To demonstrate an application of our results, we revisit a costeffectiveness evaluation of orlistat. The previous cost (£) per qualityadjusted life year (QALY) was £24,430.50 (Table 5). Two results from this study were then applied to the same calculation: cross-sectional results from the linear regression (cost per QALY: £80,499.86) and results from the delta regression (cost per QALY: £60,517.02).
Hence, using the results of our study may have a sizable impact on whether weight reduction interventions such as orlistat are deemed cost-effective, particularly in countries such as England where thresholds of £20,000-£30,000 applied by decision making bodies such as NICE.

| Nonlinear HRQoL and QALYs
Nonlinear HRQoL change from spline regressions was converted into QALYs to demonstrate the heterogeneity of HRQoL across the range of BMI; QALYs ranged from 0 to 0.427 (Table 6). In both cases, the mean change in QALYs was different to that in each of the categories underscoring the importance of this heterogeneity. Stark differences between the linear model and the delta model reinforce the difference between BMI change and crosssectional differences in BMI: cross-sectional models may be overestimating QALY gains.

| DISCUSSION
In a cross-sectional analysis of individual participant data from five intentional weight loss trials, HRQoL was negatively related to BMI, primarily because of an inverse association in those with BMI > 30, with no evidence of an association in the BMI range under 30 kg/m 2 .
In longitudinal analysis, HRQoL was negatively related to change in BMI in people with a BMI > 25 kg/m 2 . The coefficients in the crosssectional analyses for people with a high BMI were larger than those for change in BMI. The mean of the HRQoL indicators is close to those at the population mean. However, we were able to exploit BMI variation to examine how the HRQoL-BMI relationship varied across the range of BMI.
We used IPD from large randomized controlled trials in several countries to perform a meta-analysis to provide a robust synthesis of the available evidence. We were able to exploit both the crosssectional and longitudinal aspects to provide robust evidence and new insights into the relationship between BMI and HRQoL. A natural limitation of the pooled data is that variables are not collected consistently across trials. We sought to overcome that and exploit the precision of pooled data through the use of latent variable modelling, but if the association between quality of life and each measure genuinely differs, this will only be partially successful.
Incident or prevalent disease related to BMI could partly be responsible for the differences by BMI, but given the relatively short However, the latent variable is not without potential theoretical limitations. 23 We have assumed within this approach that HRQoL is causing the indicators of it, but the opposite is of course possible. Note: All of the indicators' ranges were 0-100, save for the EQ-5D, which had a range of less than 0 to 1 and General Well-Being Schedule that ranged from 0 to 110. NB -to replicate analyses of past commissioning decisions, the EQVAS was transformed to a utility as per Foxcroft et al. 12 NB -aside from the EQVAS, HRQoL scales are analyzed in their raw form without having preference tariffs applied.
While our findings generally confirm the direction of effect in previous estimates, we provide new data on how BMI change appears to influence HRQoL, and importantly, we show that the effect of this depends upon starting BMI. The association is stronger for physical HRQoL than mental HRQoL. Prior evidence of the association between mental HRQoL and BMI is mixed, including mainly negative results, but some reports of positive associations or non-significant estimates. 3,8,24 One meta-analysis reported higher mental HRQoL in overweight than for healthy weight. 8 These inconsistent findings may be because the magnitude of the association is T A B L E 3 Cross-sectional estimates of associations between 1 kg/m 2 difference in BMI and standard deviation units of HRQoL from the structural equations from the linear model and spline regression Abbreviations: Est, estimate; 95% CI, 95% confidence interval. Note: HRQoL was in turn related to each of its indicators (see Table 2 for ranges of individual HRQoL indicators). Abbreviations: Est, estimate; 95% CI, 95% confidence interval. Note: HRQoL was in turn related to each of its indicators (see Table 2 for ranges of individual HRQoL indicators).
T A B L E 5 NICE cost-effectiveness of orlistat using the original value and using our estimates Note: "Utility gain" is the gain in utility from a one unit decrease in BMI. This is derived from the EQVAS measure which is converted to a utility score based on Hakim et al. 13 "BMI change" is the total BMI change in the trial arm. "Change in QALYs" multiplies the first two columns. "QALY change in the trial arm" computes the QALY changes per 100 respondents in each trial arm. "QALYs/100" takes the difference in QALY changes between treatment and control. "Cost/100" is the cost of the treatment per 100 respondents. "Cost per QALY" is the cost per QALY of the intervention from the original study and using our revised estimates. Figures that are not computed in this paper can be found in the original study. 12 small, and many studies are underpowered to detect effects. We found no evidence of any adverse effects of intentional weight loss on mental health, even after weight regain, providing some reassur-

CONFLICT OF INTEREST
There are no conflicts of interest to declare. Structural equations: Where HRQoL it is a latent variable of health-related quality of life. are estimated for 5 time periods as one is set to zero to avoid linear dependency. α i is an individual-specific random effect, assumed to be iid normally distributed, with zero mean and variance, σ 2 α ; that is, For spline regressions, BMI is partitioned into regions of its range, Next, model (1) was re-specified to consider the relation of longitudinal changes in BMI and HRQoL. The period-to-period BMI changes are used in the linear model. We refer to this model as the delta model to refer to the change in BMI.
Finally, the longitudinal analogue of model (2) is specified. The period-to-period BMI changes are used in the spline regression. This is termed delta spline regression, Where ΔBMI_c ijt = BMI_c ijt À BMI_c ijt À 1 .
In all specifications, models are fully specified with all possible variables. Models are then iteratively refined by removing non- FT score ijt ¼ ζ ft score HRQOL ijt þ ε ijt,ft score ðB7Þ where indicators of HRQoL, for example, SF À 36(pcs), are treated as dependent variables to be explained by the latent variable of HRQoL.
Then, the ζ capture the association between the indicators of HRQoL Note: Three models are compared for each trial: OLS, fixed effects (FE) and spline regressions (spline). For each model, parameter estimates from the model on the raw data (raw) are presented next to those on the imputed data (MI) with robust t-ratios in parentheses. À0.015 (À4.14) Note: Three models are compared for each trial: OLS, fixed effects (FE) and spline regressions (spline). For each model, parameter estimates from the model on the raw data (raw) are presented next to those on the imputed data (MI) with robust t-ratios in parentheses.