Intervention and context
The DMPs for type 2 diabetes and CHD were rolled out through the SHI system in 2002 and 2003, respectively [7]. The SHI is a system of non-profit health insurance companies insuring around 88% of the German population. In contrast to DMPs in the United States and other countries, which often focus on high risk patients, German DMPs target all patients with the disease and are characterized by a high degree of homogeneity. The key contents of the German DMPs for diabetes and CHD are enforcement of medication therapy, enhanced patient activation and self-management education, continuity of care according to current guidelines and the use of information technology systems for routine documentation/benchmarking [18]. Whereas the enrollment of patients in those programs was initially incentivized by a large risk surcharge, since 2009 physicians and the health insurance receive a flat rate premium of 125€ and 20€ per year per DMP enrollee from the central ‘German Health Fond’ [7, 19, 20]. Since their implementation, the proportion of patients enrolled in the German DMPs for type 2 diabetes and CHD rose to more than 50% of individuals with those respective indications, to around 6 million people (Fig. S1 in Additional file 1). Currently, 87 and 94% of the enrollees in the type 2 diabetes and CHD DMPs, respectively, are 56 years and older, meaning that around 5 million of the 26 million Germans (around 20% of the population) aged 56 years and older are enrolled in one of these two DMPs [8, 21].
Study design
In order to assess the effectiveness of the German DMPs we conducted a SC study [21], where we compared circulatory system-related and all-cause mortality rates before and after DMP implementation in the German elderly population with those of other European countries. We concentrated on circulatory and all-cause mortality, because, in the long-term, it can be expected that better care processes lead to better intermediate clinical outcomes and translate to longer survival. Furthermore, as most of the targeted care in the German DMPs targets cardio-metabolic care processes (blood pressure, lipids, HbA1c), we assumed that the effect of the DMPs would be strongest for circulatory mortality.
The SC study is a quasi-experimental study similar to a difference-in-differences study, but it allows for the construction of a control group in instances where there are several sites to choose from, but no clear rationale for choosing the most appropriate control site. Specifically, a SC study compares changes in areas receiving the intervention with changes in a weighted average of control areas that provides the most similar comparison, with respect to the pre-intervention outcome trend and a pre-defined set of covariates [22]. Through the estimation of this counterfactual, i.e. what would have happened had the intervention not been implemented, one can account for existing secular trends, as well as for potential changes in the outcome not associated with the intervention occurring on a larger geographical scale [18].
Data
Outcomes: We obtained data on mortality and population for the years 1998–2014 from the World Health Organization’s (WHO) Cause of Death Query (CoDQL) online platform for several European countries: Austria, Belgium, Bulgaria, Croatia, Cyprus, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, the Netherlands, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland and the United Kingdom. We excluded Greece and Hungary because valid outcome data were not available. For included countries, we extracted the overall and sex/age stratified annual numbers of deaths due to diseases of the circulatory system (ICD-10: I00-I99) and all-cause deaths. By combining this mortality data with overall and sex/age stratified annual numbers of population sizes, we calculated overall and age-stratified mortality rates.
Covariates: We further identified demographic (age, sex), economic (per capita gross domestic product (GDP), unemployment rate, health care expenditure in % of GDP), clinical (prevalence of diabetes, hypertension and obesity) and behavioral (smoking and alcohol consumption) factors that are potentially associated with circulatory and all-cause mortality and for which data was available for the respective years and countries. Data on these factors were extracted from the OECD, the World Bank, the NCD Risk Factor Collaboration and the WHO (Table S1 in Additional file 1). As data on smoking and alcohol consumption were unavailable for Croatia, Estonia, Malta and Romania for the entire study period, we excluded these countries from the analysis. Where intermittent data were missing for a covariate, we imputed these values using linear interpolation using the country’s previous and next annual value.
Statistical analyses
Main analyses
In conducting a SC study one must define an intervention site, which is exposed to the intervention after its implementation, and a synthetic control donor pool, which should not be exposed to the intervention. In our study, Germany served as the intervention site and several European countries as the synthetic control donor pool. To ensure that results were unbiased by similar interventions, we conducted a series of non-systematic searches to identify interventions implemented in European countries that may have influenced the population-level diabetes and/or CHD rates from 2003 to 2009. We first checked the Cochrane Database of Systematic Reviews and Google Scholar for systematic reviews of such interventions. Additionally, for studies evaluating DMPs cited above, we used Google Scholar to identify studies subsequently citing these studies. Studies evaluating potentially relevant interventions were identified in some countries, however, in almost all cases these evaluated small-scale interventions rather than population-level interventions. According to these literature searches, the only country that introduced a nationwide program to improve the quality of chronic care was England, in which a pay for performance scheme was introduced in 2004. As an evaluation indicated that this program improved diabetes care [23], we excluded the UK from our analyses.
The SC group was created from the donor pool based on the pre-intervention outcome trend, as well as a set of pre-defined potentially important covariates, i.e. those described above. These aspects were weighted within the donor pool to best match the pre-intervention outcome trend in Germany, and to create the post-intervention counterfactual, i.e. how the outcome trend would have continued in Germany had the DMPs not been implemented. We calculated the treatment effect of interest, the average treatment effect on the treated (ATT), by estimating counterfactuals for Germany at each time point using control group information based on a linear interactive fixed effects model that incorporates unit-specific intercepts interacted with time-varying coefficients [22]. This treatment effect is the difference between the observed series, i.e. the post-intervention outcome trend observed in Germany, and the synthetic control time series, i.e. the post-intervention counterfactual series. Given the progressive enrollment of the German DMPs discussed above, and the resulting uncertainty regarding at which point they may have affected the mortality at the population level, we modeled the intervention at three different time points: immediately after implementation in 2003, after four years of enrollment in 2006, and after seven years of enrollment in 2009. As almost 90% of DMP participants are 55 years or older we restricted the main analysis to people in the following age groups: 55–64, 65–74 and 75 and older.
Subgroup and sensitivity analyses
To explore whether any differential effects were masked by the use of an aggregated age group of 55 years and older, we conducted the SC analysis for both outcomes, circulatory and all-cause mortality, in age subgroups, including ages 55–64, 65–74 and 75 and older. Additionally, to assess whether changes in cardiovascular and all-cause mortality across all ages, not just in the elderly, were influencing the study results, we repeated the main analyses for a younger age group, 20–54 years, because no effect, or at most a minimal effect, due to the intervention would be expected in this age group. This approach is similar to a difference-in-difference-in-differences estimate in which, given a real effect, we would expect differences in mortality trends in exposed populations (people aged 56 and older) but no differences in mortality trends in unexposed populations (people aged 55 and younger).
To ensure that no single control country was driving the results related to the DMP effects (for example if an unknown and effective population-level diabetes program had been implemented in a specific country in 2003), we conducted a series of leave-one-out analyses. Specifically, we conducted the main analyses repeatedly, each time removing a single country from the control donor pool.
All data processing and analyses were conducted using R version 3.3.2. The synthetic control analyses were conducted using the Generalized Synthetic Control Method (gsynth) package [24].