Summary of results
Educational differences in age at death were substantial in all European countries but contributed only a small fraction to the total individual lifespan variation: 0.6% to 2.7% in Western Europe and 1% to 11% in Central and Eastern Europe. Less-educated groups not only had shorter mean lifespans but also had greater between-individual variation in lifespan. The gap in between-individual lifespan variation between Western Europe and Central and Eastern Europe was more evident among the least-educated groups-the tertiary-educated groups had more similar lifespan distributions in all countries.
Evaluation of data and methods
One concern is whether, given our limited number of subgroups, we are fully capturing the educational gradient in mortality. When possible we used four subgroups, but in some countries we were restricted to three subgroups, and in others (e.g., Switzerland) the vast majority of the population fell into only two subgroups. This might have resulted in a lower than actual between-group component. To be sure that the different number of subgroups was not biasing our intercountry comparisons of the contribution of between-group inequality, we also ran the analysis for all countries with educational groups one and two combined. The reduction from four to three subgroups decreased the between-group component by 15% on average (results not shown). Using three subgroups altered the country rankings only slightly, with no rank changes for females and Poland and Lithuania trading places among males, when it came to examining the overall contribution of between-group variation to the total variation in age at death. Although more subgroups would increase the BG component, so long as we are capturing most of the linear educational gradient in mortality, we do not expect this effect to be large. Even if the between-group component were to double, it would still only explain a small fraction of individual level lifespan variation.
Education is not the only component of socioeconomic status. British studies have shown, for instance, that adding car ownership or housing tenure introduced health gradients within broad occupational categories [24–26]. Presumably, other measures of socioeconomic status would explain some of the within-educational group lifespan variation that we have found here. Although the high correlation between socioeconomic variables suggests that education would be capturing a large portion of the total socioeconomic inequality, different indicators of socioeconomic position are at times associated with different health outcomes [27–32]. Thus we expect our results to be robust for capturing the extent of the contribution of educational inequalities to lifespan variation but to underestimate the full extent of all socioeconomic inequalities.
Another concern is whether the nature of unlinked studies may introduce a numerator/denominator bias. Authorized informants may state a different educational status for the deceased than was recorded in the population census. If the deceased are reported as having a higher than attained educational level ("promoting the dead"), this would lead to overestimating mortality among the highest-educated groups . However, a record linkage study for Lithuania found that unlinked estimates overestimated mortality in lower educational groups and underestimated mortality in the highest-educated groups, particularly for females . We were able to compare our unlinked estimates with these linked Lithuanian data  (see Additional file 1). We found that the range in the average age at death between the highest- and least-educated groups was less in the linked data by 22% for males and by 34% for females. This had the effect of substantially decreasing the between-group component. As a result, the contribution of educational inequalities in age at death decreased from 7.8% to 5.0% for males and from 6.9% to 2.7% for females. While the overestimation is certainly substantial, the results from the linked data confirm a larger between-group contribution in Lithuania as compared to most Western European countries. Such a bias is also likely for Estonia, given that the two countries are post-Soviet Baltic countries that experienced similar reforms to the educational system and exhibited similar trends in unlinked age and education-specific mortality. It is more difficult to determine the direction and magnitude of bias in the Czech Republic and in Poland. The Lithuanian results are likely to be context specific and should not be generalized to other countries.
Finally, there could be problems of comparability between countries given the different study years. The unlinked studies of Central and Eastern Europe take place around the year 2000, which is on average five years later than the longitudinal census-linked studies that followed subjects for the 10-year period between the 1990 and 2000 round of censuses. Alongside changing educational compositions in the population, during this period relative inequalities in mortality between educational groups increased throughout Europe [36, 37]. Some studies found that the magnitude of this widening was even greater in Central and Eastern European countries [38, 39]. Thus, if we had had data for these countries for periods comparable to the longitudinal studies, we might have found smaller differences in the between-group inequality component between Eastern and Western European countries.
Comparisons to other studies
To the best of our knowledge, we are the first to decompose individual-level variation in age at death into its between- and within-group components using Theil's index. The contribution of the between-group component that we observed is similar to American estimates made by Tuljapurkar , calculated by approximating the variance decomposition that we presented in Additional file 1. Morbidity researchers have decomposed the Gini coefficient or the related Health Concentration Index to determine the degree to which subgroup variation in age-standardized levels of health could be explained by socioeconomic status, a different but related question [41–43]. In these studies they found a much higher contribution from the socioeconomic component than we did. Yet it is difficult to make a direct comparison here: the distribution of age-standardized levels of health, in which many individuals self-report perfect health, differs considerably from the distribution of ages at death.
Should a 1% to 11% contribution from between-group differences to the total between-individual variation in age at death be considered a large or a small amount? It is important to recognize that between-individual variation arises from many different sources, including genetic, behavioral factors, environmental conditions, and chance. These factors may in part be associated with educational level and thus vary between educational groups, but there is likely to be even more variation on many of these factors within educational groups.
We are not the first to point out that between-group differences in life expectancy account for little of the total between-individual variation. Doblhammer found that a lifespan difference of nearly half a year by month of birth explained just over 0.01% of the total variation in age at death . In an additional analysis, we applied Theil's decomposition method to calculate the contribution of between-sex differences to total variation in age at death, using data from all countries of the Human Mortality Database for the year 2005. We found that the between-group component explained between 1.6% (England and Wales) and 9.9% (Russia) of total lifespan variation (results not shown). It would be interesting to run this type of analysis for risk factors such as smoking. It is also likely that lifespan variation within smoking and nonsmoking groups is larger than average differences in lifespan between the two groups. Thus we would expect a relatively small contribution from the smoking-related between-group component despite a 10-year difference in life expectancy between smokers and nonsmokers .
Hence it is not that between-group educational differences in mortality are small, it is more that the magnitude of all interindividual lifespan variation is tremendous. Even the large five-year advantage in life expectancy held by the highest-educated Swedish males over their least-educated counterparts acted mostly to shift the whole death distribution to higher ages (Figure 1). It did not alter the shape of the two distributions, which remained largely overlapped, owing to the much greater within-group variation.
In addition to putting inequalities in mortality between socioeconomic groups within a broader perspective, our analysis leads to some new insights into the nature of these inequalities. Educational subgroups differ not only in their mean length of life but also in the spread around that mean: the shorter life expectancy of less-educated groups concurs with a much greater variation in age at death as compared to higher-educated groups. While this inverse relationship is predicted by the compression of mortality theory , empirically life expectancy has been shown to be a poor predictor of lifespan variation at the macro level since the 1960s for distributions conditional upon surviving childhood [19–21, 47–50]. Why mortality compression differs by educational group warrants further investigation [5, 7]. Also, the larger educational inequalities in mortality in some Central and Eastern European countries can be seen to arise from the larger between-individual variation in age at death within their less-educated groups. This suggests that reduction of socioeconomic inequalities in mortality might primarily require a reduction of variability in age at death. This may require better protection of people with higher vulnerability, e.g., because of smaller personal resources or less favorable living conditions. It also requires a concerted effort to tackling causes of death that dominate at young ages, such as injuries and neoplasms . The results of our analysis support the idea that a main function of modern welfare states is to provide such protection against the vicissitudes of life .
Returning to the debate introduced in the introduction of this paper, it seems that individual-level variations and group-level inequalities should not be seen as competing perspectives but as interrelated phenomena. The one is embedded in the other. Our analysis illustrates the suggestion by Gakidou et al. that within-group differences are themselves interesting and substantial and a necessary complement to research into between-group inequalities . But simply measuring the sum of between-group and within-group differences, which was proposed by the WHO report as an alternative measure of health inequalities, cannot replace a specific focus on measuring inequality along socioeconomic lines or any other grouping of interest such as gender, ethnicity, region, or lifestyle.
Although socioeconomic differences in mortality are but one of many factors determining when individuals die, they are often seen to be among the most important and inequitable. This is because socioeconomic inequalities are at least partly avoidable, and because they follow from inequalities in the distribution of socioeconomic resources, which themselves are often seen to be unjust . Even if they contribute only a small fraction of all between-individual variations in lifespan, they are a legitimate concern for public health. What this study adds is that tackling inequalities in mortality between socioeconomic groups can also be approached through reducing variation in age at death among less-educated people by providing protection to the vulnerable.