Microsimulation model design

The microsimulation model was developed to forecast future chronic disease burden, longevity and direct economic costs in 27 EU countries and the UK until the year 2050, as well as the extent to which future changes in epidemiological trends and specific policies can modify these outcomes. This modelling platform, initially based on previous OECD work [13], was further improved as part of the FRESHER project [14]. The model uses case-based microsimulation to create representative synthetic life histories from birth to death, and relies on detailed epidemiological and demographic information from various sources. The model also uses prevalence-based direct cost estimates as an input into the model to forecast incidence-based health expenditures associated with various scenarios/policy interventions, from the health system perspective. More detailed information on data and the modelling assumptions can be found at a technical report posted online [14,15], the accompanying paper [16], as well as in the S6 Appendix in S1 File.

Diseases modelled

In general, attempt has always been made to estimate the direct medical costs for the following NCDs, although in some countries we evaluated costs only for a subset of these diseases due to the limited data (for more information, see S1 Appendix in S1 File):

• Ischemic stroke
• Haemorrhagic stroke
• Myocardial infarction and chronic heart disease
• Cancers
• Diabetes
• Chronic kidney disease (CKD)
• Chronic Obstructive Pulmonary Disease (COPD)
• Cirrhosis
• Depression
• Neurologic disorders (including Alzheimer disease and dementia)
• Injuries (including intentional and unintentional injuries and self-harm)

We selected these diseases due to their large contribution to the total NCD burden, as well as because many of them have common aetiologies and potentially modifiable risk factors such as obesity and harmful alcohol consumption [17,18], and therefore are likely to be amenable to the scenario assumptions. In addition, since injuries (including intentional, unintentional and self-harm) are significantly linked with alcohol use disorders [19], representing about 10% of the total disease burden caused by the alcohol use [20], the cost of injuries was modelled as well. Additional information on how the injury costs were estimated is provided in S2 Appendix in S1 File.

Estimation strategy

The microsimulation model requires individual-level total medical costs, conditional on age, gender and disease status as an input. In a standard bottom-up approach to cost estimation, patient-level units of health care attributable to a specific disease are multiplied by a price for this unit, and then summed up. However, in practice administrative coding is often not very informative about the underlying reason for incurring health expenditures. For example, falls can occur excessively in people suffering from Alzheimer’s disease, but health expenditures may be recorded under alternative coding, thus giving an incomplete picture of the health expenditure burden facing this group. Alternatively (and also incorrectly), all the recorded costs for a person with a disease can be attributed to this disease. To overcome this limitation, we estimated costs using regressions, as the mean marginal difference of the predicted medical costs with a disease variable switched on or off (i.e., using between subjects comparison). This approach is commonly used to estimate incremental costs for select diseases and risk factors [21–23]. In our study, the outcome variable is the total cost of hospital and ambulatory care (including costs of drugs) at the patient level.

We also account for the possibility that having two or more diseases may lead to costs that are greater than a simple sum of costs for separate diseases [24]. For more information about how this super-additivity was modelled, see the S1 Appendix in S1 File.

Finally, we estimated age- and gender-specific residual medical costs for people who had no model- defined diagnosed diseases (but could have other diagnosed conditions, or were otherwise healthy). Taking into account such costs is important because reducing risk factor prevalence may still lead to accruing costs for unrelated medical conditions, either because people will live longer, or because they can be affected by unrelated competing disease risks. For more details about the estimation approach, see the S1 Appendix in S1 File.

Data sources and cost extrapolation

Although average healthcare spending per capita was accessible for all European countries, precise cost-of-illness estimates were not available, or even calculable, for a great majority of countries. A cost extrapolation model was therefore used to generalize the incremental costs of our 10 chronic conditions from the three “anchor countries” where the data was available -France, Estonia and the Netherlands- to the remaining EU-28 countries.

Our extrapolation methodology assumes that the annual treatment cost differentials between countries are time-invariant, and mostly driven by the differences in two components: costs per unit of treatment received, as well as population-level intensity of the treatments provided. To estimate these differentials, we used the OECD data [25] on the inpatient curative and rehabilitative care spending per capita, outpatient curative and rehabilitative care spending per capita as well as medical goods spending per capita. To ensure comparison and to smooth over temporary data variability, we used this data averaged over 3 years: 2012, 2013 and 2014, expressed in current prices.

Given a significant amount of cost-based extrapolations and uncertainties associated with future modelled scenarios, it was decided to present results not separately for each country, but aggregated over three regions: Southern Europe, Central/Eastern Europe and Northern Europe. There is no consensus on how such a grouping should be done, therefore members of the Fresher consortium decided to follow simple geographical criteria at the beginning of the project. First, Southern Europe is composed of EU27 countries bordering the Mediterranean Sea, as well as Portugal. Central/Eastern Europe counties include those which joined the EU since 2004, except Slovenia and Croatia (which belong in the Southern Europe group). The remaining countries belong in Northern Europe group. The final groupings were as follows:

1. Southern Europe (France, Italy, Spain, Portugal, Cyprus, Greece, Croatia, Malta, Slovenia);
2. Central/Eastern Europe (Estonia, Bulgaria, Poland, Romania, Slovakia, Hungary, Latvia and Lithuania);
3. Northern Europe (the Netherlands, Austria, Belgium, Czech Republic, Denmark, Finland, Germany, Ireland, Luxembourg, Sweden and United Kingdom).

This setup is not supposed to indicate that countries belonging to one group are always more similar to each other than to countries in another group. For example, healthcare system in France can have much more in common with Belgium than with Croatia. Nevertheless, deciding on group composition is likely to involve at least some degree of arbitrariness, and we tried to minimize it by using simple geographical proximity criteria. In any case, grouping decisions are not likely to affect the quality of the extrapolation so long as the differentials in costs between countries are taken into account, as we describe in the S1 Appendix in S1 File.

Disease identification

In source databases, we had to identify the occurrence of NCDs through a medical diagnosis and/or an associated treatment. Disease definitions were standardized across countries, with disease identification based on criteria such as International Classification of Diseases (ICD) diagnostic codes [26] and/or codes from the drug prescription databases. Specifically, in France, disease identification was done using ICD—version 10 codes; hospital discharge data as well as drug discharge Anatomical Therapeutic Chemical (ATC) codes, using Huber’s classification method [27]. In Estonia, identification was done using ICD-10 codes provided in the Estonian Health Insurance Fund (EHIF) dataset. In the Netherlands, identification of NCDs was only possible using ATC codes of medications prescribed to patients outside of hospitals recorded, which precluded reliable identification of diseases, such as CKD and cirrhosis. Therefore, for the Netherlands, we borrowed these later costs from France (only those missing). For more details about country-specific diseases identification, see the S1 Appendix in S1 File.

Model scenarios

We compare the outcomes under three scenarios for the evolution of the following key risk factors: smoking; obesity rates; alcohol use; physical inactivity and high blood pressure. Our BAU scenario is identical with the “the rich get healthier” scenario described in more detail in the accompanying paper [16]. Briefly, a group of 90 health experts were asked to predict the future evolution of five key risk factors based on the narrative description of four different scenarios of the future, previously identified after consulting a large number of experts on societal trends likely to impact the future evolution of NCDs. The “the rich get healthier” projection, in which it was assumed that the societal inequalities will continue to widen, and mostly market-based solutions to societal and environmental problems will be offered, was chosen as the BAU, as it was viewed as the most likely to represent the future based on the continuation of the current trends.

To implement the best (worst) case scenarios, we first calculated the spread between the Europe-wide average rate for a given risk factor in 2015, and the lowest (highest) rate for that risk factor. For example, if the average obesity rate in Europe in 2015 was 20%, and the highest observed obesity rate that year was 30%, then we assume that every country will see its obesity rate linearly increase so that in 2050 it’s greater by 10 percentage points than its current rate. This approach was implemented to ensure comparability with the other scenarios described in the accompanying paper [16].

Table 1 presents the average risk factor prevalence in 2015, as well as their evolution under the three scenarios. See S3 Appendix in S1 File for a graphical depiction of the changes in risk factor prevalence over time.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Risk factor prevalence (%) in 2015 (actual) and in 2050 (predicted).

https://doi.org/10.1371/journal.pone.0238565.t001

Fig 1 depicts an example of a model-implemented link between an increase in one risk factor-the BMI- and an increased risk of developing related conditions such as heart disease or diabetes. Each additional disease arising out of an increased BMI is then associated with the disease-specific cost, or in the case of multiple conditions, with the cost for multiple diseases, which have been estimated for all 28 countries. A total cost saving attributable to the BCP can likewise be estimated.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Conceptual framework for high BMI attributable direct costs.

https://doi.org/10.1371/journal.pone.0238565.g001

Limitations

This analysis only considers the impact of scenarios on direct medical costs. However, it is well known that improvements in health are likely to lead to improvements in other economic outcomes [34]. For example, alcohol-related economic burden may also include direct non-medical costs, e.g. relating to property damage, or “indirect” costs arising from the lost or underutilized labour market resources. Simply put, indirect costs arise when no money physically changes hands, but some resources are still lost/underutilized. We do not estimate such costs in this paper, but they can contribute substantially to the economic impact of the modelled scenarios, often surpassing the direct burden [35]. In addition, although direct costs do contribute to the GDP, in their absence, the money could be invested in other areas, with potentially higher returns, and we do not take this potential “opportunity cost” aspect into account. In the same vein, we also do not take into account the potential cost of policies needed to obtain the variations in risk factors for the two alternate scenarios, which are likely to be very country-dependent. We also do not take into account the potential impact of other variables on the evolution of costs in the future, e.g. the impact of medical technology, which may well differ depending on the scenarios, or possible changes in prices of drugs (e.g. when patents expire). Another limitation is that most available data on the relationships between the modelled risk factors and diseases are from observational studies and therefore causality cannot usually be taken for granted, although the most recent literature we used generally tries to establish reliable causal impacts rather than simple associations. Finally, there were some limitations related to the disease identification in some countries or assumptions behind the cost extrapolation approach.

Policy implications

All in all, our modelling exercise is generating three important messages. Firstly, unless there is an exceptional effort to reduce risk factor prevalence in Europe which goes beyond the BCP assumption taken in this paper, future health expenditures in this region will be dominated by the population ageing trend, and therefore will continue to grow rapidly. Nevertheless, various population-level preventative interventions may still be cost-effective of even cost-saving, and better value for money than treating a chronic disease once it has developed [36]. Indeed, with the current health system focus in many countries on curative interventions, there is significant room for shifting the focus towards prevention [37], and various innovative intervention packages may be developed to reduce the prevalence of NCDs yet further [38]. In addition, there should be a search for more sustainable modes of care in the future, including greater attention paid to improving rehabilitative and supportive care to keep the elderly independent and well-functioning for as long as possible. Secondly, by taking into account the cost of unrelated competing NCDs as well as co-morbidities, and showing that additionality and super-additionality is significant for a subset of the modelled conditions, the model provides results that not only improve on existing models available to date, but also constitute immediate policy relevance. Super-additionality in disease costs highlights the importance of targeting not only single diseases, but also overall comorbidities, with appropriate policy interventions. The model can provide further guidance about which co-morbidity reduction might result in the highest cost savings. Thirdly, as shown in the accompanying paper [16], although the expected positive effect of an improved risk factor profile is relatively modest for the financial sustainability of health systems, the impact on epidemiological burden as measured by changes in disease incidence, life expectancy and premature mortality is potentially very significant. Likewise, evidence suggests that economic returns to higher labour market productivity may significantly exceed those from healthcare system savings [38]. As the methodological approach we use in this paper does not value this impact in monetary terms (which may include both nonmarket value of improved health, as well as potential improvements in productivity and/or labour force participation), the full economic and welfare benefit from the more favourable risk factor profile is likely to be much greater than the estimated health expenditure saving may suggest.