Cost-effectiveness of bariatric surgery and non-surgical weight management programmes for adults with severe obesity: a decision analysis model

Objectives To determine the most cost-effective weight management programmes (WMPs) for adults, in England with severe obesity (BMI ≥ 35 kg/m2), who are more at risk of obesity related diseases. Methods An economic evaluation of five different WMPs: 1) low intensity (WMP1); 2) very low calorie diets (VLCD) added to WMP1; 3) moderate intensity (WMP2); 4) high intensity (Look AHEAD); and 5) Roux-en-Y gastric bypass (RYGB) surgery, all compared to a baseline scenario representing no WMP. We also compare a VLCD added to WMP1 vs. WMP1 alone. A microsimulation decision analysis model was used to extrapolate the impact of changes in BMI, obtained from a systematic review and meta-analysis of randomised controlled trials (RCTs) of WMPs and bariatric surgery, on long-term risks of obesity related disease, costs, quality adjusted life years (QALYs) and incremental cost-effectiveness ratios (ICERs) measured as incremental cost per QALY gained over a 30-year time horizon from a UK National Health Service (NHS) perspective. Sensitivity analyses explored the impact of long-term weight regain assumptions on results. Results RYGB was the most costly intervention but also generated the lowest incidence of obesity related disease and hence the highest QALY gains. Base case ICERs for WMP1, a VLCD added to WMP1, WMP2, Look AHEAD, and RYGB compared to no WMP were £557, £6628, £1540, £23,725 and £10,126 per QALY gained respectively. Adding a VLCD to WMP1 generated an ICER of over £121,000 per QALY compared to WMP1 alone. Sensitivity analysis found that all ICERs were sensitive to the modelled base case, five year post intervention cessation, weight regain assumption. Conclusions RYGB surgery was the most effective and cost-effective use of scarce NHS funding resources. However, where fixed healthcare budgets or patient preferences exclude surgery as an option, a standard 12 week behavioural WMP (WMP1) was the next most cost-effective intervention.


Data inputs Epidemiological data Population data
Demographic data was collected for England, Scotland, Wales and Northern Ireland. Information was collected on the age and sex distribution of the population, and the distribution of deaths by age and sex.
The data were processed as text files, in a format suitable for inclusion in the microsimulation programme. The data sources were as follows

Disease data
A number of obesity-related diseases were modelled (see Table 2). The list of diseases modelled for obesity was determined after a review of the literature conducted for the WRAP study (5)( Table 2).  (7) Non-terminal Non-terminal Zheng

Incidence, Prevalence, Mortality data by disease Breast cancer
Prevalence data was not available on breast cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required. Coronary heart disease (CHD)

Colorectal cancer
Prevalence data was not available on colorectal cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Endometrial cancer
Prevalence data was not available on endometrial cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Renal cancer
Prevalence data was not available on Renal cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Oesophageal cancer
Prevalence data was not available on Oesophageal cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Ovarian cancer
Prevalence data was not available on Ovarian cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Pancreatic cancer
Prevalence data was not available on Pancreatic cancer data, but the model does not require the input of prevalence, only of incidence, so this parameter was not required.

Survival data
Survival statistics for CHD and Stroke were not identified in the literature. We modelled these using prevalence and mortality data,see Module two: Microsimulation model section Approximating missing disease statistics for methods. RR is the given relative risk, RR' is the adjusted relative risk and A is the adjustment multiplier).             Primary care is often the primary point of contact of someone seeking care. GP visits are the main source, but studies, when available, include other types of services offered by most of the GP practices. These include nurse visits, home visits, phone/email/fax consultations.

CHD
 Prescription costs are usually estimated as the volume times the costs of primary care prescription.
 Inpatient costs are the total costs of treating a patient at hospital for a specific diagnosis (episode). They include day cases, elective and emergency admissions.
 Outpatient costs capture the costs of visits to specialists.
We have not included in our different costs indirect costs such as the loss of income when hospitalised. More information about the costs used in each paper can be found in Health Economic Review excel workbook. 1 We have been advised that Programme Budgeting data does not fully capture the actual health care expenditures, in particular for social care costs. Thus costs from the literature were preferred where data exists.

Summary of identified costs
All the costs were adjusted using prevalence when necessary to represent the total cost per type of care and per disease group for England.
For the microsimulation model, we need the cost per case, which is the total cost divided by the prevalence in 2016. Therefore this figure is not necessarily equal to the unit cost as patients use different combinations and quantities of care.
Relevant sources were collated using a systematic literature review in PubMed, and completed using Google searches. We searched for peer-reviewed articles using PubMed. We also used Google to identify reports from other sources. We focused exclusively on costs based on English or UK data.
While multiple studies from the search results were considered, the most relevant, recent studies were used for the final cost estimates. We rarely had the choice between two references, but in this case our selection criteria were the transparency of the method to estimate the costs with a preference for bottom-up approaches, 2 the clarity of the methodology and definitions, the source of data with a preference for national representative samples, and the years for which the costs were reported. All costs were adjusted for inflation using the CCEMG-EPPI-Centre cost Converter and divided by the prevalence in order to have a "cost per case".(1) 2 A bottom-up approach, in contrast to a top-down approach, reflects the actual needs. It quantifies each resource required to provide the services or treatments to care for patients with a specific condition, multiplied by the input costs. A top-down approach allocates a total figure (e.g. the NHS programme budgeting cost) to different services as such is less likely to capture the actual spending associated with a specific disease.

Limitations
The main overall limitation of using costs from the literature is that the estimation methods vary significantly from one paper to another, and the inputs for each category of care are slightly different for each condition. When the estimates come from a bottom-up approach, the costs are likely to underestimate the true costs as the possible missing components are set to zero. When the estimates come from a top-down approach, the allocation rule is often not clear and it is hard to know how comparable they are to the true cost. Yet, the order of magnitude is likely to represent the true costs for the NHS, as the overall costs are broken down into parts. Furthermore, the authors often argue that their method is conservative and that the estimated costs represent lower-bound estimates.

Utility weights
All utility weightings for use in QALY calculations were obtained from Sullivan et al's 2011 Catalogue of EQ-5D scores for the United Kingdom and NICE (20) Males and females were allocated the same EQ-5D score, as this is not specified by gender in the publication. The diseases were mapped onto conditions listed in the publication using matching, or closest matching ICD9 and Clinical Classification Categories.

UKHF microsimulation methodology Microsimulation framework
Our simulation consists of two modules. The first module calculates the predictions of risk factor trends over time based on data from rolling cross-sectional studies. The second module performs the microsimulation of a virtual population, generated with demographic characteristics matching those of the observed data. The health trajectory of each individual from the population is simulated over time allowing them to contract, survive or die from a set of diseases or injuries related to the analysed risk factors. The detailed description of the two modules is presented below.

Microsimulation Module one: Predictions of overweight/obesity over time
BMI was analysed within the model as risk factors (RF), as described in Table 6.

Multinomial logistic regression for each risk factor
Measured data is extracted from the survey data set. They consist of sets of probabilities with their variances. Each set represents the probabilities of individuals of normal weight, overweight, obesity class I & class II and obesity class III at specific time values (i.e., the year of the survey). For any particular time the sum of these probabilities is unity.
Each data point is treated as a normally distributed random variable; together they are a set of N groups (number of years) of K probabilities      Module two: Microsimulation model

Microsimulation initialisation: birth, disease and death models
Simulated people are generated with the correct demographic statistics in the simulation's start-year. In this year women are stochastically allocated the number and years of birth of their childrenthese are generated from known fertility and mother's age at birth statistics (valid in the start-year). If a woman has children then those children are generated as members of the simulation in the appropriate birth year.
The microsimulation is provided with a list of BMI-related diseases. These diseases used the best available incidence, mortality, survival, relative risk and prevalence statistics (by age and gender This following section provides an overview of the main assumptions of the model.

Population models
Populations are implemented as instances of the TPopulation C++ class. The TPopulation class is created from a population (*.ppl) file. Usually a simulation will use only one population but it can 46 simultaneously process multiple populations (for example, different ethnicities within a national population).

Population Editor
The Population Editor Allows editing and testing of TPopulation objects. The population is created in the start-year and propagated forwards in time. An example population pyramid which can be used when initialising the model is shown in Figure 1 shows the population distribution for England in 2016 used in the initialisation of the model.

Figure 1 Population pyramid for England in 2016
People within the model can die from specific diseases or from other causes. A disease file is created within the program to represent deaths from other causes. The following distributions are required by the population editor (Table 7).

Deaths from modelled diseases
The simulation models any number of specified diseases some of which may be fatal. In the start year the simulation's death model uses the diseases' own mortality statistics to adjust the probabilities of death by age and gender. In the start year the net effect is to maintain the same probability of death by age and gender as before; in subsequent years, however, the rates at which people die from modelled diseases will change as modelled risk factors change.

Continuous risk factors
In the case of a continuous RF, for each discrete distribution there is a continuous counterpart. Let key requirement for these sets of longitudinal trajectories is that they reproduce the cross-sectional distribution of RF categories for any year with available data. The method adopted here and in our earlier work is based on the assumption that person's RF value changes throughout their lives in such a way that they always have the same associated percentile rank. As they age, individuals move from one age group to another and their RF value changes so that they have the same percentile rank but of a different RF distribution. Crucially it meets the important condition that the cross-sectional RF distributions obtained by simulation match the RF distributions of the observed data.
The above procedure can be explained using the example of the NO2 distribution. The NO2 distributions are known for the population stratified by sex and age for all years of the simulation (by extrapolation of fitted model, see equation (0.1)). A person who is in age group and who grows ten years older will at some time move into the next age group ′ and will have a BMI that was described first by the distribution   Where 1 F  is the inverse of the cumulative distribution function of  , which we model with a continuous uniform distribution within the RF categories (see Table 6). Equation (0.7) guarantees that the transformation taking the random variable  to '  ensures the correct cross-sectional distribution at time ' t .
The microsimulation first generates individuals from the RF distributions of the set D and, once generated, grows the individual's RF in a way that is also determined by the set D . It is possible to implement equation (0.7) as a suitably fast algorithm. Remember that different probabilities will apply to different age and gender groups. Typically the data might be divided into 10 year age groups.