Economic evaluations of DPP-like lifestyle interventions have used different methodologies and produced inconsistent results. The first economic evaluation of DPP, which compared lifestyle interventions and pharmacotherapy (metformin) with placebo care, was performed by the DPP Research Group in 2003 [37]; it covered only the three-year intervention period without any DAM and from a societal perspective. As DPP intervention costs were very high, it was proposed that the intervention be offered on a group basis (10 people per group) instead of face to face; it was assumed that the effectiveness would be similar, so that the costs of lifestyle intervention would be reduced. As a three-year time horizon overestimates the treatment costs and underestimates the benefits of lifestyle intervention and metformin, the same research group later extrapolated the trial data into lifetime costs and benefits using a Markov model [38]. The progression of IGT to clinical onset of diabetes and from diabetes-related complications to death was assessed.

A later study [39] used the DPP data for five countries, i.e., Australia, France, Germany, Switzerland, and the UK, applying a simple three-state Markov model (i.e., alive with IGT, alive with T2DM, and deceased) over a lifetime horizon. Another study analyzed the DPP results from a different perspective [40], i.e., whether investment in a DPP intervention program by a health insurer would cut costs. The model was same as that of the DPP Research Group but used for two time periods. The conclusion was that, if the insurer and Medicare shared the DPP intervention costs (24% borne by the insurer), the insurer would recover the investment via avoided future medical care costs.

Two studies examined [41,42] screening for diabetes among overweight and obese people followed by DPP intervention, while Icks et al. [42] studied the cost-effectiveness of DPP in a real-world setting, i.e., when acceptance of and adherence to the intervention is low and the dropout rate is high. In addition to metformin, another drug, acarbose, was examined in Caro et al. [43] using a four-state Markov model of a Canadian population over ten years; this study estimated that lifestyle modification would prevent 117 cases of diabetes, while metformin and acarbose would prevent 52 and 74 new cases of diabetes, respectively. The lifetime cost and effectiveness of DPP was estimated by Eddy et al. [44] using an Archimedes model, which contains infinite health states. This is the only study estimating that DPP-like lifestyle intervention has a mere 0.1% chance of costing under 50,000 US$/QALY.

DPP or DPP-like interventions have been studied in other countries as well. IDPP was performed by Ramachandran et al. [45]; although the analysis examined only the trial period (three years), the cost-effectiveness ratio was much lower than for DPP. Galani et al. reported two studies [46,47] on lifestyle interventions for overweight and obese Swiss population groups, with assumed effectiveness taken from DPS. A seven-state Markov model over a lifetime horizon estimated that lifestyle intervention could be cost-effective depending on sex, age group, and threshold values. DPS effectiveness was applied to a Swedish population [48] using a Markov model, which included the cost of added life years.

Methodological disagreement is the main issue in DPP-like studies. The results of DPP interventions are reported as 8,800 US$/QALY or 62,600 US$/QALY depending on whether a Markov [38] or Archimedes model [44] is applied. If 50,000 US$/QALY is considered a cutoff value for cost-effectiveness, the same trial is cost-effective with one method but not the other. The disagreement stems from different model assumptions on the rates of progression to diabetes and complications [49,50]. Both authors provide arguments and counterarguments defending their assumptions [51,52].

Despite the disagreement, it was predicted that DPP-like lifestyle intervention would delay the onset of diabetes and lead to fewer complications, longer lives, and improved quality of life [50]. All the DPP/DPS/IDPP-like lifestyle interventions were cost-effective, except that reported by Eddy et al. [44], even if the intervention started with costly screening [41,42]. IDPP was much more cost-effective than similar programs in high-income countries, because the intervention cost was much lower even though staff went to participant homes to provide the intervention [45]. The cost-effectiveness ratio was in the cost-effective range when DPS was applied to a Swedish population [48], even though unrelated costs for added years of life were included, which is debatable [53].

Several methods have been used to determine the physical activity level of participants in the economic evaluation of physical activity studies. One study used metabolic equivalent (MET) minutes per week to classify participants as inactive, irregularly active, meets guidelines, and highly active [54]. Some studies used duration and intensity of physical activity, i.e., ≥2.5 hours of moderate intensity or vigorous exercise per week as marking a physically active person [55], while other studies used the self-administered physical activity questionnaire (PAQ) [56,57]. There were only two Markov models [54,55] of physical activity: the CDC MOVE Markov model was based on five diseases (i.e., coronary heart disease, ischemic stroke, T2DM, breast cancer, and colon cancer), which led to underestimated cost-effectiveness, as physical activity also affects other diseases such as depression or anxiety [58].

Indirectly measuring the level of physical activity (e.g., via PAQs) raises validity concerns [59]. The selection of participants for physical activity interventions is also a major concern, i.e., whether the control group also participates in the program or only includes motivated people; for example, one study [56] excluded highly active participants. Participation rate and adherence to physical activity are other important issues meriting attention when modeling the long-term benefits of physical activity.

Surprisingly, the ICER of physical activity interventions is much lower in Australia [55] and New Zealand than in the USA [54] or UK [56]. This could be because Roux et al. [54] analyzed the physical activity promotional intervention using a model containing considerably more health states and because the participants were older in the Munro et al. [56] study; however, all ICERs are in the cost-effective range.

An economic evaluation of ten different nutritional interventions for the Australian population was performed by Dalziel and Segal [60]. Four Markov models were developed to analyze nutritional interventions. Reduction of dietary sodium intake to reduce hypertension, eating five servings of fruit and vegetables per day, and adhering to a “Mediterranean-style” diet seemed to be cost-effective. The effect of the Mediterranean diet was estimated to reduce the 10-year CVD risk based on the Framingham risk equation. The cost-effectiveness of grains fortified with folic acid as well as vitamin supplementation with folic acid and cyanocobalamin (vitamin B12) to prevent coronary heart disease (CHD) was analyzed in one study [61] of the entire US population via the CHD Policy Model. Nutritional education was compared in Cox et al. [62], one group receiving traditional face-to-face advice, while another received self-administered video lessons. The effectiveness of the intervention was analyzed using a behavior questionnaire on changes in 10 dietary factors.

There are few economic evaluations of dietary interventions and the quality of studies is often limited, since some important economic evaluation parameters (i.e., age, risk factors, sensitivity analysis, and price year) are not clearly specified. The use of self-administered questionnaires to measure food intake raises validity issues.

Dietary interventions are very diverse, and probably not very comparable. Moreover, varying study quality makes it difficult to compare the results of the dietary interventions, but fruit and vegetable intake has the lowest cost-effectiveness ratio (46 AU$/QALY) [60] and similar results are obtained from universal fortification of grains with folic acid.

Three articles presented economic evaluations of a national policy or action plan [63–65]. The new Finnish hypertension care guidelines were compared with previous ones [64], which did not incorporate lifestyle intervention; the new guidelines were estimated to save an additional 49,000 life years. The Dutch national action plan for counteracting obesity included a community-based intervention in which 90% of the population was screened and received lifestyle advice; an intensive lifestyle program was applied to 10% of the overweight adults in a healthcare setting. The combined program was evaluated using the RIVM chronic disease model, finding that it would save 110,000 life years over 20 years [63]. Using a diabetes model [65], it was estimated that diabetes care and prevention strategies applied to the entire Australian population aged 45–74 years would prevent 53,000 new cases of diabetes over 10 years.

Three articles studied childhood obesity prevention programs [66–68]. The “traffic light diet,” physical activity reinforcement, self monitoring, and stimulus control were used as interventions in one study; their effectiveness was measured as a reduction in standardized body mass index (Z-BMI) and proportion of overweight children [67]. Another study [68] examined Planet Health, a school-based obesity prevention program focused on reducing television viewing, reducing consumption of high-fat foods, increasing fruit and vegetable consumption, and increasing moderate and vigorous physical activity. The randomized controlled trial included male and female subjects, but there were significant reductions in obesity only among female students after two school years. This intervention effect was then extrapolated up to age 64 years using a decision model measuring health benefits as QALYs. The same method was used in another study [66] of a different school-based obesity prevention program (CATCH) including male and female participants from grades three (age 8) to five (age 11). CATCH appears to be more cost-effective than the Planet Health program.

Two studies have focused solely on reduction of CVD incidence using the Framingham risk equation [69,70]. One study compared lifestyle interventions, including video and video + self-help guidelines, with routine care for Australian participants at risk of CVD [70]. In another study, community-based education to change dietary habits to reduce total cholesterol level was estimated to prevent approximately 4.5 cases of CVD every year, with the highest benefit in the 55–64-year age group, in which it would prevent 8–9 cases [69] in a population of 5,500. The WISEWOMEN project, which aimed to reduce CVD risk in older uninsured and undernourished women by means of either CVD screening + enhanced lifestyle intervention or CVD screening + minimal intervention, was evaluated in two studies [71,72]. The earlier study [72] measured effectiveness as the percentage reduction of the 10-year probability of having CVD, while the later study [71] measured it as LYG as well, resulting in a cost per LYG of US$ 4400. The effect of a single randomized controlled trial intervention of diet, exercise, and diet + exercise extrapolated on a cohort of 60-year-old healthy subjects without CVD was evaluated in one study [73]. The Markov model included 10 health states but not diabetes. The cost-effectiveness of interactive group sessions advising on nutrition and physical activity for cohabiting Australian couples was studied [74] on an intention-to-treat basis with outcomes measured on 14 variables (e.g., energy intake, fat intake, fiber intake, sodium, fruit and vegetable intakes, BMI, physical activity level, cholesterol profiles, and blood pressure). Two types of lifestyle intervention were evaluated by Jacobs et al. [75]—a community-based intervention for the general population covering many people at a relatively low cost, and a healthcare-based intervention for high-risk people covering fewer patients at a higher cost—using the RIVM chronic disease model.

Many studies lack detailed information about certain important economic evaluation parameters and on how data have been derived, for example, via meta analysis, systematic review, selective studies, or expert opinion. These parameters include cost (e.g., unit cost, total cost, intervention cost, disease cost, and productivity cost) [65,72,74], utility [63], and disability weight [65]. In childhood obesity studies [66,68], the explicit assumption requires evidence, which is missing, from epidemiological studies linking childhood overweight to adult overweight and to weight regain after two-year interventions. The same issue applies in CVD cases: a time lag after termination of lifestyle intervention in improvements of CVD risk factors. Lindholm et al. [69] considered a five-year lag, whereas

Dzator et al. [74] considered no time lag. The transferability of one country’s clinical trial data to another [65] also requires adjustment. For community-based interventions [63,69,75], special attention is required to consider the spillover effect of lifestyle interventions and the efficacy of interventions in various socioeconomic groups, i.e., the equity concern is missing from the studies. However, the main question concerns the applicability of intervention effectiveness data from clinical settings to real-world settings, particularly when national action plans or policies are being evaluated [63–65].

Interventions starting from childhood have a low cost-effectiveness ratio (900 US$/QALY) [66]. The community-based approach is also attractive, as seen in the Dutch [63,75] and Swedish studies [69]. The advantage of community-based lifestyle prevention programs is that the health gains achieved through population-based approaches often exceed those achievable by targeting specific groups in clinical or subclinical settings.

Three articles [76–78] concerning the drug sibutramine were studied by the same research group using the same model but in different country settings, i.e., Finland, Germany, the USA, the UK, and Switzerland. Lifestyle intervention was included along with the drug to treat overweight or obese people. The total number of fatal and non-fatal CHD events avoided in five years of analysis was estimated as 3.49 in Finland, 4.18 in Germany, 4.49 in Switzerland, and 1.96 in the UK per 1,000 people, while the average number of diabetes cases avoided was 3.0. In an Italian obese population, another drug, orlistat, was studied [79]. The same drug combined with a low-calorie diet for obese patients was also assessed for the whole Dutch population, estimated over a lifetime horizon using the RIVM chronic disease model [80]; the combined therapy was estimated to produce an additional 34,000 life years.

A weight-reduction drug, rimonabant, was compared with lifestyle intervention [81] using a decision tree with five arms, i.e., treatment options. Two years’ treatment with rimonabant combined with lifestyle intervention produced the most cost-effective option. Employing a Markov model, Roux et al. [82] compared four weight-loss strategies in overweight and obese women. The strategies were diet only, diet + pharmacotherapy (orlistat), diet + exercise, and diet + exercise combined with behavior therapy. The weight-loss intervention consisted of six months of intervention followed by a six-month maintenance program. Diet + exercise combined with behavior modification was the best strategy.

When statin drugs and cholesterol-lowering diets were compared with no interventions for patients with a high cholesterol level using the CHD Policy Model over a 30-year period, both interventions produced much higher cost/QALY in the primary prevention [83]. This may be because statins were expensive at the time (1997), before they were available as generics. In another study [84], four strategies for preventing or delaying diabetes, i.e., screening for early detection, screening + lifestyle intervention, screening + lifestyle + pharmacological intervention, and no screening, were compared with a hybrid model using a three-arm decision tree and seven-state Markov model. Screening followed by lifestyle intervention was the most cost-effective strategy.

All pharmacological studies included a short-term trial period (six months to one year) extrapolated to a five- or ten-year time horizon, except one that employed a lifetime horizon [80]. One issue in pharmacological interventions was weight regain after intervention end; the National Institute of Clinical Excellence (NICE) recommendation [85] on assumed uniform weight regain over a three-year period had not been taken into consideration. Furthermore, though all drugs had some side effects, the related loss of quality of life was not considered. We agree with previous review findings that the main sources of uncertainty in pharmacological interventions are weight-loss sustainability, utility gains associated with weight loss, extrapolation of long-term benefits from short-term trials, dropout rate, side effects, and bias towards the funding authority [86].

Surprisingly, all effectiveness data for drug interventions are based on literature reviews, except data from the study [79] reporting the highest ICER (75,300 €/QALY). When lifestyle interventions are compared with pharmacotherapy, lifestyle interventions are more effective, i.e., in survival years, disease-free time, and quality-adjusted life expectancy [82], indicating that lifestyle interventions are better options for preventing lifestyle diseases.

Of 46 studies, 31 employed DAM: six used decision trees, 20 used Markov models, and one used an Archimedes model (Table 6). One article used both a Markov model and a decision tree [84], one used four Markov models [60], one used an Archimedes model [44], and two used a life table approach. A decision tree is a simple visual representation of possible options and their consequences. Decision trees start with the options, each of which branches out to explore all potential health outcomes and their respective probabilities and costs. In Markov models, participants move from defined health states (Markov states) in discrete time periods (Markov cycles). Each health state incurs particular costs and health consequences [18]. It is common to use a previously developed model; all the

Dutch [63,75,80] studies used the RIVM chronic disease model adapted to the study objectives, while the CHD Policy Model [61,83] and the CDC diabetes model [38,40,41] have been used in many studies.

Philips et al. [92] emphasized several issues for good practice in modeling: model structure; data employed, i.e., inputs (costs) and outcomes (health benefits); and model consistency or validity. For Markov models, the structure generally concerns the health states included, as inclusion of costly diseases (e.g., stroke due to diabetes in diabetes progression models) might overestimate the long-term results. Another example is that the RIVM chronic disease model [63] included certain types of cancer absent from another model [46], although both models were developed for overweight and obesity.

The effectiveness data were taken from a single trial or literature review. For most models, the model population was a hypothetical cohort but, when the effectiveness of a trial was transferred to different country settings, the model population was often matched with the study population; for example, the survey population from the US National Health and Nutrition Survey [41] and the participants in the German KORA study were matched with the DPP participants [42], and similarly in three other studies [39,44,60].

Two types of cost data were included in the models: cost of intervention and cost of disease avoided. Some studies estimated the intervention cost from a clinical trial [91], some calculated the cost retrospectively [71], and some based the cost on national administrative databases [65], expert opinion [73], or even modeler opinion. The avoided disease costs were country specific if available; otherwise, the applicable data were taken from other countries. The model outcomes were QALYs, DALYs, LYG, or other measures of health. Different countries were found to have used different instruments when estimating QALYs: QWB-SA was used for the DPP models, European studies frequently used EQ-5D, while Finnish and Australian studies preferred their own instruments (i.e., 15D and Aqol). Models are subjected to internal validation (i.e., comparing model output with the data used in building model), external validation (i.e., checking whether the model output is consistent with the disease outcome and epidemiological data), and between-model validation (i.e., comparing the estimated intervention outcome with the outcomes of other models based on similar assumptions and addressing similar diseases) [92,93], as model quality depends largely on input data quality [94]. No studies provided any details on all the three types of validation. However, in modeling, it is recommended that a technical report [19], i.e., a detailed description of all assumptions and parameter values used to construct the model, be provided. Not all model studies mentioned technical reports, though some provided supplementary materials.