Methods for the economic evaluation of obesity prevention dietary interventions in children: A systematic review and critical appraisal of the evidence

Summary Objectives We aim to describe and provide a discussion of methods used to conduct economic evaluations of dietary interventions in children and adolescents, including long‐term modelling, and to make recommendations to assist health economists in the design and reporting of such evaluations. Methods A systematic review was conducted in 11 bibliographic databases and the grey literature with searches undertaken between January 2000 and December 2021. A study was included if it (1) was an economic evaluation or modelling study of an obesity‐prevention dietary intervention and (2) targeted 2‐ to 18‐year‐olds. Results Twenty‐six studies met the inclusion criteria. Twelve studies conducted an economic evaluation alongside a clinical trial, and 14 studies modelled long‐term health and cost outcomes. Four overarching methodological challenges were identified: modelling long‐term impact of interventions, measuring and valuing health outcomes, cost inclusions and equity considerations. Conclusions Variability in methods used to predict, measure and value long‐term benefits in adulthood from short‐term clinical outcomes in childhood was evident across studies. Key recommendations to improve the design and analysis of future economic evaluations include the consideration of weight regain and diminishing intervention effects within future projections; exploration of wider intervention benefits not restricted to quality‐of‐life outcomes; and inclusion of parental or caregiver opportunity costs.


| INTRODUCTION
In 2016, the World Health Organization estimated that over 18% of 5-to 19-year-olds were affected with overweight or obesity. 1 The main cause of overweight and obesity is an imbalance between energy consumption and energy expenditure. Diets high in saturated fat and sugar lead to excess energy consumption and contribute to the prevalence and burden of obesity related diseases, including type 2 diabetes mellitus, cardiovascular disease and cancers. [2][3][4] Obesityrelated health expenditures negatively impact on limited healthcare budgets, costing the UK National Health Service alone over £5.1 billion annually. 5 Interventions that aim to improve population diet are therefore a priority for policy makers, and evidence on the economics of such interventions is becoming internationally recognized as being crucial to support effective public health policy making. [6][7][8][9] Health economic evaluations assess additional costs and benefits of an intervention against a comparator (e.g., usual practice). How this is conducted is dependent upon several factors, including the type of economic evaluation approach and whether a healthcare or societal perspective is adopted. Economic evaluations can be conducted alongside clinical trials, where costs and benefits are derived from trial data. Alternatively, clinical effectiveness data can be input into an economic model to derive long-term cost and benefit outcomes.
Where the former provides a cost-effectiveness estimate using actual trial data, the latter provides long-term projections of healthcare and societal resource use, costs and associated benefits. The way in which costs and benefits are compared between an intervention and a comparator is dependent on the evaluation framework. There are four main economic evaluation frameworks: (1) cost-minimization analysis: when different treatment options have equivalent outcomes, therefore the cheapest option is used; (2) cost-effectiveness analysis: a comparison of additional costs by additional benefits (natural units); (3) cost-utility analysis: a comparison of additional costs by additional health-related utilities (e.g., quality-adjusted life years, disability adjusted life years and health years gained); and (4) cost-benefit analysis: health and/or non-health benefits are valued in monetary terms (distinctly different to a return on investment which accounts for financial benefits only). 10 Six systematic reviews have been identified concerning the economics of childhood obesity prevention. [11][12][13][14][15][16] Most recently, Zanganeh et al. conducted a quality appraisal of the literature 16 and reviewed the methods adopted within economic evaluations of nutrition and physical activity-based interventions. However, this study was primarily descriptive in nature and did not provide a critical analysis of the methods, including strengths and limitations, adopted within studies. Oosterhoff et al. also examined key aspects in the design of economic evaluations on school-based interventions and highlighted key issues and recommendations for future economic evaluations. 14 However, such reviews have either: lacked a comprehensive search strategy, potentially compromising the inclusion of key texts 11,12,14 ; focused on a narrow population group or intervention setting 13,14 ; or focused solely on physical activity interventions. 15 There is currently a lack of consensus on the scope and content   (Table S1). 18 In addition, the grey literature was searched using broad terms: "economic evaluation," "child" and "obesity" and/or "diet.

| Exclusion criteria
Studies published before the year 2000 were excluded, to ensure the inclusion of up-to-date practices, and for pragmatic purposes, given available resources. Modelling studies of hypothetical policies were excluded as they rely on data from multiple intervention studies rather than the evaluation of a single intervention. This review focused on obesity prevention; therefore, weight loss and obesity treatment studies were excluded. Studies targeting niche population and patient groups were also excluded. Finally, studies that measured obesityrelated health conditions with no reference to obesity-prevention or dietary improvements within their aims were excluded.

| Data extraction and quality appraisal
Two data extractions tables were developed, piloted and refined. Two reviewers (S.M. and C.M.) independently extracted data and compared for completeness and accuracy. Any conflicts were discussed until agreement was met.
The Cochrane Public Health Group data extraction and assessment template form 24  Following guidance provided by the Centre for Reviews and Dissemination the BMJ 35-item checklist was used to assess the quality of economic evaluations. 27 Items designed for the critical appraisal of decision-analytic models developed for health technology assessment were embedded to cover issues relating to modelling studies. 28 These included structural assumptions, model type, time horizon, health states and cycle length. Two items from the Paediatric Quality Appraisal Questionnaire were also embedded in order to capture insights into methods for capturing parent and child impacts, including productivity and school absence. 29 One reviewer assessed the quality of all studies (S.M.) and a second reviewer (C.M.) independently validated 20%.

| Data synthesis
A narrative synthesis of the methods used by the economic evaluations was conducted. Characteristics of effectiveness and costeffectiveness studies were summarized and details concerning economic evaluation and modelling study methods were identified, compared and set within the context of the broader methods literature. Descriptions of cost-effectiveness studies, together with reported sensitivity analyses, were used to make recommendations concerning the scope and content of economic evaluations, models and key parameters. Research findings are presented based on a classification of key methodological challenges adapted from Weatherly et al. 30 Within their paper, several reviews exploring the economics of various public health interventions were investigated in which key methodological challenges were commonly identified across studies: attribution of effects; measuring and valuing outcomes; intersectoral costs and consequences; and equity considerations. were discussed between reviewers leading to a further two inclusions.
One additional paper was identified via the reference list of included studies and included in the review. 31 In total, 20 papers comprising of 19 separate studies, with one study split across two papers, 32,33 were included in the systematic review.
In the updated search strategy conducted up to December 2021, 5563 studies were initially identified, and 1336 duplicates were removed. One reviewer (S.M.) screened 4227 titles and excluded 3145 studies that were not related to the main inclusion criteria relating to obesity prevention (phase 1 screening), followed by the screening of 1082 titles and abstracts (phase 2 screening).
A final number of 27 studies were included for full text screening (phase 3 screening) whereby a second reviewer (C.M.) screened 30% of full-texts. There was 100% agreeability between the two reviewers leading to the inclusion of 7 additional studies and the exclusion of 20. No additional papers were identified from references or the grey literature.
In total, 27 papers comprising of 26 separate studies were included within this systematic review, and 46 papers were excluded overall after full-text screening. Main reasons for exclusion included: not an economic analysis (14/46), not based on a single effectiveness study, such as a hypothetical policy (13/46), not meeting criteria for population characteristics, such as age (8/46) and not an obesity prevention nutrition-based intervention (7/46). Four additional studies were excluded due to there being no intervention comparator, the study was not in the English language, the authors had no access to the paper and study data was previously reported and had been included in the initial search strategy. Figure 1 shows the pooled study selection process.
Quality appraisal outcomes are presented in Table S2 and   Table S3. There was 81% concordance in the scoring of studies between the two reviewers. None of the studies fulfilled all the quality criteria and only 19/35 items from the BMJ checklist were fulfilled by at least 80% of studies.

| Discount rates
Discounting of costs and benefits is not required in the case where an intervention lasts 1 year or less, as was the case in eight studies. 34,38,41,48,50,53,54,57 However, two studies lasting 2 years or over were not discounted. 44,45 Ten studies indicated a discount rate of 3%, 32,36,37,39,40,42,46,47,51,55 four studies indicated a discount rate of 3.5% 31,43,49,56 and one study utilized a discount rate of 5% per annum. 35 One study applied a 4% discount rate for costs and a 1.5% discount rate for benefits, per annum. 52 Though typically discount rates are selected based on country-specific recommendations, seven studies did not justify their discounting choices. 31

| Sensitivity analyses
All but three studies provided details of a sensitivity analysis. 38,44,53 Probabilistic sensitivity analysis was most often conducted within studies and seeks to explore the impact of parametric uncertainty in the model. 32,33,37,39,40,42,46,47,52,55 Though the use of probabilistic sensitivity analysis allows description of the parametric uncertainty within economic outcomes, other methods investigate uncertainty of assumptions within the analysis through the variation of one (one-way sensitivity analysis) 31,32,[34][35][36][39][40][41]43,45,[47][48][49][50][51][52][54][55][56][57] or multiple parameters (two-way or multi-way sensitivity analysis) at a time. 32,46,47,51,55 Further details of modelling methods are outlined in Table S5, and the parameters commonly investigated within sensitivity analysis are outlined in Table S6.  Table 2. The results are presented as a narrative synthesis and critical appraisal of the methods identified in the economic evaluations.

| Modelling long-term impact of interventions
Several challenges in modelling the long-term impact of interventions were identified. These include the omission of child intervention benefits when adopting lifetime horizons; the approaches used to project long-term outcomes from childhood to adulthood; and assumptions concerning the maintenance of intervention effects over time. Each of these main issues will now be discussed.
Methodological guidance commonly requires a lifetime horizon in economic analysis. This is particularly relevant in economic evaluations of obesity prevention studies, as many of the benefits of obesity prevention interventions will occur in adulthood. Nevertheless eight studies, all of which conducted economic evaluations alongside trials, based their time horizons on trial duration, which ranged from 5 weeks 48 to 28 months. 45 Whereas, modelling studies included cost and benefits over a lifetime, 31,32,36,39,40,49,52 or truncated analyses at 84, 46 65,37,55,56 or 40 47 years. Where truncated lifetime approaches were adopted, authors justified this based on a paucity of long-term outcomes data. Two studies modelled costs and benefits over a 10-year time horizon, as this was most relevant for policy makers and due to the long-term uncertainty regarding intervention effects. 42,51 One study modelled intervention costs and benefits to cover both the childhood (up to 20 years old) and adulthood years. 52 However, in most instances health outcomes and associated costs were only modelled throughout adulthood. In doing so, childhood economic benefits of interventions were often overlooked. Emerging research suggests that obesity impacts directly upon child health through early changes in metabolic risk factors 79,80 and negatively impacts on healthcare resources early on in life. 81 Failing to include childhood health outcomes risks underestimating the economic benefits of early intervention and increases levels of uncertainty when longer time horizons are considered. Moreover, some decision makers are interested in early outcomes in their own right. 82 One solution is to present economic intervention effects needs to be incorporated within models and adjusted within scenario analysis for a more accurate depiction of reality and costeffectiveness outcomes. d

Measuring and valuing health outcomes
Potential Impact Fractions (+) BMI was treated as a continuous rather than a categorical variable when considering expected disease due to changes in exposure to the risk factor by BMI unit. 32,36,40 This is a more accurate reflection of the association between BMI and diseases in comparison to methods that have used weight status to determine disease presence, 31,37,46,47,55,56 such as is the case with transition probabilities for remaining healthy, developing a weightrelated condition or death. a,d (+) Stability was assumed of all incidence and mortality rates from causes other than the diseases included in models. 40 Although this may not be representative of reality, this ensures that costs and benefits are specifically evaluated for obesity-related disease states. a • The use of BMI as a continuous outcome measure is more accurate than the use of categorical weight status to accurately reflect the associations between weight and disease. d Relative risks of disease incidence and mortality conditional on BMI (+) Due to low incidence rate data, it was assumed that BMI did not lead to many illness cases before the age of 20 years. Inclusion of illness from age 20 years is considered an improvement in comparison to studies that have investigated disease incidence during older adult years. 31,49 a (À) General population incidence rates obtained from a country not related to the study population, was frequently used with no justification. 31 46 This considers both the impacts of HRQoL of obesity and chronic disease. a,d (À) Adult based utility decrements had been applied to younger age groups. 46 HRQoL is typically more impaired within the older than younger years. 67 Though the consideration of obesity-related health impacts within the younger years is a progressive step within models, the use of adult-based data may overestimate the benefits of this. b • Where factors may be highly correlated (e.g., obesity and disease states), care should be taken when attributing utilities to weight status in case of doublecounting benefits (or lack thereof). Methods such as applying the highest disutility value between weight status and disease state may be an optimal approach to adopt. • Careful consideration needs to be taken when choosing the most appropriate utility values from the literature, including: the population describing the health state (e.g., age, sex), elicitation technique used to derive utility value, sample size and country. 56 Costs and benefits by weight status (À) Cost and benefit outcomes were based on long-term weight status categories (healthy/overweight/obese). 37,47,55 This assumes that overweight/obesity will impact all individuals equally when outcomes vary by sociodemographics. 42 Given short-term follow up of interventions, it is unlikely that any significant changes in BMI or cases of overweight/obesity avoided would have been detected to allow meaningful modelling of long-term intervention impacts. a (À) Although there is value in using BMI when assessing health risks of overweight and obesity, this is not the most reliable measure. 71 • In the face of high uncertainty within modelling outcomes, more reliable and objective methods should be adopted to measure dietary or energy intake, for example, doubly labelled water, or the use of adjustment equations for selfreported data. • Where there is a lack of data or evidence from RCTs to support long-term projections of intervention effects, alternative data sources ought to be considered. Amongst other considerations include non-experimental data, prospective studies and the application of econometric methodology. 30 • Alternative outcome measures may be better predictors of disease, other than BMI, including waist circumference, or potentially objective dietary intake. 72

Cost inclusions
Costs converted into rates (+) Converting costs into rates allows gradual costs of obesity to be factored along with the possibility that not everyone will live the same number of years, hence incurring different amounts of obesity-related costs. 33 outcomes over a selected range of time horizons up to death, allowing the impact on uncertainty to be explicitly communicated. 59,82 For example, results can be presented for 1, 5, 10, 20, and 50 years. 83 This will enable the case of investment to be presented, and will demonstrate how interventions can positively impact short-term outcomes, and avert health complications that may not present until adulthood.
Studies utilized different approaches to modelling long-term outcomes from childhood-based effectiveness data. Most commonly, literature was used to obtain childhood to adulthood body mass index (BMI) trajectories. 39,47,56 In two cases, adult obesity impacts were based directly on rates of child overweight averted in two stages, firstly at 21-29 years, then again at 40 years. This was due to a lack of single progression estimates in published data. 37,55 Such methods did not account for within-group differences (e.g., sex) that may result in variability in intervention effects (unlike regression models). 46 Alternatively, future weight was categorized based on population survey data in annual, 31,49  Maintenance of intervention effects was assumed within all basecase analyses, except one. 51 This is problematic because weight regain after weight loss is a reoccurring problem, meaning that economic outcomes may be overestimated. 84 One study used an annual depreciation rate of 2.62%, acknowledging the likelihood that intervention effects are not maintained in the long-term, which reflects clinical findings. 51 Since data on the maintenance of intervention effects within obesity prevention is currently lacking for children, adult-based estimates were adopted. To account for intervention effects degrading over time, another study used data on fruit and vegetable consumption from adolescence to young adulthood to justify a 30% lifetime extrapolation of intervention effects within sensitivity analysis. 40 Other studies examined the impact of declines in intervention effectiveness through sensitivity or scenario analysis, 31,40,46,47,49,52,54,57 allowing the assessment of parameter and structural uncertainty within the economic evaluations. These analyses led to substantial differences in cost-effectiveness outcomes in comparison to base-case scenarios. However, such assumptions were seldomly supported by evidence from longitudinal studies, with approximately half of studies justifying their choice of variables within sensitivity analysis. [32][33][34][35][36][41][42][43][49][50][51][52][54][55][56][57] Previous work has also demonstrated how incorporating an intervention decay rate can substantially affect the cost-effectiveness of an obesity intervention, 59 suggesting the importance of factoring in changes to intervention effectiveness over time.

| Measuring and valuing health outcomes
A number of methodological issues associated with measuring and valuing health outcomes were also identified. These related to the methods for associating weight status to disease incidence and mortality, methods for linking disease severity to health utility, the scope of obesity related diseases considered, the wider non-weight related potential health impacts and the use of current utility instruments.
Inclusion of disease states within models was done through incorporating Potential Impact Fractions, which calculate the proportion change in expected disease or death by change in BMI. 32,36,40 The use of a continuous risk factor (e.g., BMI) is more accurate than a categorical classification of weight status (e.g., healthy weight/overweight/ obese) when predicting disease incidence and mortality rates. 63 The number of obesity-related chronic disease states used within models also varied from four 56 to fourteen, 31,49 and commonly included diabetes, cancers, stroke, hypertension, and heart disease.
Although disease states were omitted from models, 37,55,56 potentially due to lack of available data or low incidence rates by weight status, this could exclude relatively rare conditions with a significant economic burden. More simplified models have based cost and benefit outcomes directly on long-term weight status, whereby cost of illness is associated with overweight/obesity status. 37,42,47,51,53,55 This assumes that overweight/obesity will impact health states of individuals equally, yet costs may vary by age, sex, socioeconomic status and ethnicity. [68][69][70]91,92 No study considered the wider non-weight related potential health gains from improvements in nutrition. 93 This could underestimate the potential impact of interventions in cases where recipients comply with behavioral changes that have no impact on weight outcomes. 94 Despite these studies being termed as ineffective, 31,35,43,53,56 they may have positive effects on comorbidities or non-health outcomes. [95][96][97] Two economic evaluations alongside clinical trials used utility instruments (the Health Utility Index 35

| Cost inclusions
Limitations involving the inclusion of costs were identified across studies. These comprised of the methods by which costs were included within models, the dismissal of healthcare costs associated with overweight and obesity related health states, and the exclusion of wider costs and potential cost-savings.
The costs included in an economic evaluation can have a significant impact on the results. Most models incorporated costs associated with either obesity-based or obesity-related disease costs. 37 converted obesity-related disease costs for each sex and 5-year age group into rates for the Australian population. All disease-specific rates for each sex and age group were summed to give a total obesity-related disease cost rate. Total cost rates were incorporated into lifetables at each one-year age group via extrapolation methods.
More recently published studies within this review considered medical care costs for both children and adults, 42,51,52 taking into consideration GP and specialist visits as well as a comparison of medical costs between those with healthy weight and overweight/obesity. 52 Exclusion of such costs risks the underestimation of costeffectiveness outcomes. In addition, only one study incorporated both obesity-related chronic disease cost and disease costs associated with longer years lived (independent of weight). 52 Other costs were also not considered by most models. Only three studies incorporated productivity costs by quantifying the number of lost sick days for individuals with and without obesity. 37,52,55 In addition, 65% of studies did not discuss the relevance of productivity changes to the study question. 31,[35][36][37]39,40,43,46,47,50,56,57 Considering the impact of obesity on productivity, 101 omitting these costs may lead to a large underestimation of cost-effectiveness. Moreover, preventing cases of childhood overweight/obesity may lead to a reduction in supervised healthcare visits, and consequently costsavings of opportunity costs of lost time. However, only four studies considered opportunity costs of lost time for parents and informal caregivers, 33,36,43,52 whilst others considered such inclusions within sensitivity analysis, 43,55 and one study considered school absences 52 which also holds repercussions to parent/carer workplace productivity costs through increased absenteeism. As such, societal perspectives may be better suited than healthcare perspectives, due to cross-sector cost implications.
The family unit plays an integral component within childhood obesity-prevention studies. Childhood obesity prevention interventions are likely to impact the whole household, and not just the recipient child, especially as changes in diet will likely be the result of food purchasing behaviors. This is particularly the case when interventions are not restricted to changes within the school environment, but also involve parents in their administration. [31][32][33]38,40,43,46,51,52,54,57 As such, childhood obesity prevention trials may lead to spill-over effects onto other family members, 102 accruing greater intervention benefits and cost-savings from disease prevention. 51 Changes to dietary behaviors can also hold financial repercussions to the household, given that healthier substitutions are more costly than unhealthy, energy-dense foods. [103][104][105] However, these were rarely considered within studies.

| Equity considerations
The consideration of equity is a key component for economic models of particular relevance for public health interventions. 106 Health inequalities describe differences in health status between population subgroups associated with economic or social conditions. 107 Childhood obesity is a worldwide concern that impacts those within disadvantaged groups disproportionately. 58,77,78 However, less than half of included papers considered equity within their evaluations.
Four studies compared outcomes by gender, 32,33,36,39,40 four studies When modelling the long-term impact of interventions, assuming that intervention effects are maintained from childhood through to adulthood carries a danger of over-estimating cost-effectiveness outcomes. Children and adolescents are amenable to changes from the point at which trial data is collected at childhood until adulthood.
Therefore long-term predictions of outcomes may be questionable, especially when intervention effects are known to diminish with time, 108  Results suggested considerable differences between reference intervention effects and expert elicited scenarios. 52  it would allow us to better understand the implications of much shortterm intervention research. In developing and validating models of long-term effects, researchers should explore other reliable sources of data, including commercial providers or existing registries. 112 We are currently living in an obesogenic environment. Unhealthy diets are more prevalent due to the availability, affordability and accessibility of calorie-rich foods. 113 51,115 Though this will incur additional substantial costs, the availability of such interventions will ensure that individuals will have constant exposure to obesity prevention strategies, increasing likelihood of long-term behavior change. However, adopting a life course approach may pose challenges for economic evaluation, as has previously been reported. 116 For instance, given the number of players involved in implementing a whole of system intervention, spanning across numerous sectors and implemented by both formal (e.g., school) and informal (e.g., parents) parties, tracking of cost inclusion estimates and intervention maintenance costs will be difficult and timely. Until long-term data is available, there may be uncertainty regarding suitable follow-up periods for intervention effect size estimates, alongside a suitable comparator. Data collection requirements may be burdensome for community members, and need to be feasible. 117 There is also a likelihood that intervention benefits will extend beyond child and adolescent recipients, 83 and may lead to non-weight related health outcomes. The exposure to multiple behavior change strategies may interact with one another leading to expected or unexpected consequences, which may be difficult to predict and account for. 118 As such, it has been advised that system dynamic models ought to be utilized in such scenarios to predict changes in system shifts. 116 Review findings have also highlighted the potential underestimation of cost-effectiveness outcomes due to the neglect of wider inter-  128 There has also been an increase in studies investigating childhood obesity related healthcare costs, with findings suggesting substantial medical costs as early as the first 5 years of life, 129 and greater utilization of general practitioner and specialist weight services. 52,92,130 Although the inclusion of such costs can be a laborious task, economic evaluations ought to consider cross-sectoral costs or discuss potential intervention impacts across sectors.
Decision makers have expressed that economic evidence should consider minimizing inequality alongside maximizing efficiency, 106 and called for a formal weighting of outcomes by population subgroups.
Alternative methods may include separate cost-effectiveness analyses by subgroup, however this has implications for both primary research, for example increased sample sizes to detect subgroup effects, and secondary modelling that would require subgroup specific parameter inputs. 131  Finally, combating health inequalities is core to public health interventions. It is imperative for studies to explore differences in costeffectiveness by subgroups should data permit this.