Declining regional disparities in mortality in the context of persisting large inequalities in economic conditions: the case of Germany

Abstract Background Subnational regional mortality inequalities are large and appear to be mostly increasing within industrialized countries, although comparative studies across high-income countries are scarce. Germany is an important country to examine because it continues to experience considerable economic disparities between its federal states, in part resulting from its former division. Methods We analyse state-level mortality in Germany utilizing data from a newly constructed regional database based on the methodology of the Human Mortality Database. We compare time trends (1991–2015) in the German state-level standard deviation in life expectancy to that of other large, wealthy countries and examine the association between mortality and economic inequalities at the regional level. Finally, using contour-decomposition methods, we investigate the degree to which age patterns of mortality are converging across German federal states. Results Regional inequalities in life expectancy in Germany are comparatively low internationally, particularly among women, despite high state-level inequalities in economic conditions. These low regional mortality inequalities emerged 5–10 years after reunification. Mortality is converging over most ages between the longest- and shortest-living German state populations and across the former East–West political border, with the exception of an emerging East–West divergence in mortality among working-aged men. Conclusions The German example shows that large regional economic inequalities are not necessarily paralleled with large regional mortality disparities. Future research should investigate the factors that fostered the emergence of this unusual pattern in Germany.


German Regional HMD
Raw population counts for most states and years were extracted from the Genesis Online system (60). For the eastern German states (1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990) we obtained data from the state statistical offices who had reconstructed the population exposures backwards to the last East German census of 1981. These exposure data were adjusted as described in the main text. Birth counts were gathered from 1990 onward from the Genesis Online system of the German Federal Statistical Office. For the 1980s we obtained birth counts for western German states from statistical publications, and for eastern German states through data requests from the state statistical offices who had reconstructed the births backward.
Death counts originated from the statistical offices of the German states and the Federal Statistical Office of Germany. As not all states publish death statistics by single year of age up until the highest possible age, we used individual-level death register data from the Research Data Center of the Federal Statistical Office and the Statistical Offices of the German states (15) to move the open age category to ages 95 and above. In addition, we reconstructed death statistics for the eastern German states for the period 1980-1989 based on data derived from the official individual-level cause-of-death register of East Germany. The state of Berlin is separated into East Berlin and West Berlin to allow for continuous trends. A final hurdle was that it was no longer possible to distinguish East and West Berlin from official statistical data beyond 2001. We therefore used the estimation method implemented for the HMD (37) to separate births, deaths and population estimates into the two city parts.
Our calculations use the standard HMD methods protocol. Within the HMD project it is currently discussed to apply additional procedures for regional data in the HMD to more accurately measure mortality at higher ages. Thus, we cannot exclude that the life expectancy data which will be published in the German Regional HMD will slightly differ from the data which we used in this paper. After last methodological issues have been resolved and final data protection clearance has been obtained, the German Regional HMD will be made freely available online.

Remaining life expectancy at age 5 (e5) at the state level (international)
We used data from all large OECD countries with 30 million inhabitants or more. Excluded from the analysis were Korea, Mexico and Turkey because data were either unavailable or of a lower quality. We based our quality assessment on whether data at the national level were provided in the Human Mortality Database (HMD), which includes countries only when death registration is virtually complete (36). We used unaltered preexisting state-level life tables for all countries depicted in Figure 2 of the main text. European countries e5 by sex, year and state (or equivalent regional agglomeration, hereafter also described as state) were extracted for all EU countries with five or more states (Spain, France, Italy, Poland, and UK) from the abridged life tables available from Eurostat. This was done at the NUTS 1 level to make the agglomerations equivalent to the German states. French overseas territories were excluded.
As a robustness check, e5 at the NUTS 2 level was also retrieved for all European countries including Germany. Mortality data for European countries were retrieved from Eurostat using the Rpackage eurostat (65).
Per capita GDP, and all contextual variables used for the panel data analysis were obtained from the European Union Urban Data Platform (66).

The Contour Decomposition method
When analysing the convergence in life expectancy between two populations, we are often faced with the question of how much the current difference is a legacy of the past, and how much is owing to different age-specific mortality trends. The recently developed contour decomposition method (39) allows us to answer precisely this question by splitting the age contributions of a conventional between-population decomposition into components reflecting (a) initial age-specific differences relating back to some point in time, and (b) a component relating to differences in age-specific mortality trends. In other words, this allows for a direct comparison of age-specific mortality trends given different initial mortality differences.
In a sense, the method works by adjusting the mortality contributions of the past e5 difference to the new, typically lower mortality context. Specifically contour decomposition involves stepwise decomposing the age-specific mortality of two populations along an age-period contour. At each age, replacements of age-specific death rates were made in order from Baden-Württemberg in the final time period, back to Baden-Württemberg in the initial time period, then to the population of interest in the initial period and finally forward to the population of interest in the final period. After each replacement step, e5 was recalculated. The end result was two vectors of age contributions to the current e5 difference: an (averaged) trend contribution from the two populations and an initial between-population contribution. Importantly, these two vectors of age-specific contributions sum exactly to the age-specific contributions of the current e5 difference obtained by conventional decomposition-a result that would not be possible by performing and combining three separate decompositions (the trend decompositions of changes in e5 for each separate population plus one decomposition of initial differences in e5).
To reduce the number of decompositions we grouped federal states into meaningful population     2) The coefficient of variation in GDP per capita is for both sexes combined.

Online Supplementary Appendix B
In order to examine statistically the association between GDP per capita and life expectancy at the regional level we used a panel regression model with fixed effects having the following general specification:   Our regression analyses consisted of two kinds of models: (Model 1) the relationship between GDP per capita and e5, and (Model 2) the relationship between the standard deviation in GDP per capita and the standard deviation in e5. Because of data availability constraints we were not able to make proper adjustments for confounders such as unemployment. The complete regional time series for this variable were only available for a few countries and shorter time periods.
Both models were run separately by sex using the regional data for the following six countries: France, Germany, Italy, Poland, Spain, and the UK. For each model we ran several diagnostic tests, and when applicable adjusted them for major disturbances such as heteroscedasticity and serial autocorrelation. We also tested a log transformation of the independent variable (GDP) given that the relationship between life expectancy and GDP has been found to be log-linear across countries at different levels of economic development (30). The use of a log transform did not influence the results. For the sake of a straightforward interpretation of the obtained regression 13 coefficients we preferred to use the absolute value of GDP per capita. All statistical analyses were performed in Stata 14.2 SE using the command xtreg.
In Model 1 we analyzed the association between per capita GDP and e5 for European countries using region-specific fixed (time invariant) effects, with the data from the EU Urban Data Platform underlying Figure 3: The basic model included GDP and e5, and in a consistency check we replaced GDP with employee compensation as an alternative measure of economic development. The latter consists of wages and salaries and employers' social contributions, and thus might better approximate the income generated by private households. GDP, on the other hand, is the Total Gross Value Added, plus taxes less subsidies, and includes income generated both by companies and households.
The results obtained for Model 1 (Table S1) show that change in the absolute value of regional income (regional GDP per capita) is directly and consistently associated with change in regional life expectancy. For example, the regression coefficient obtained for German men implies that after adjusting major disturbances, an increase in GDP per capita by 1000 Euro is associated with an increase in e5 by 0.65 [0.61;0.69] years. Models using employee compensation as an alternative measure of economic development were largely consistent with these results.
The notable exception here was Italy, where an inverse association between GDP and e5 was found. In order to check to what extent the Italian data was influencing our results, we ran Model 1 (all countries) with and without Italy. The exclusion of Italy did not result in any notable changes.

16
In Model 2 we tested the relationship between the standard deviation (SD) in GDP per capita and the SD in life expectancy at age 5 (e5): The results (Table S2) suggest no association between SD in e5 and SD in GDP per capita among our six observed countries. This was also the case in our robustness check where we replaced the SD in GDP with the SD in employee compensation.