The main explanation for widening inequalities in health is that they are the result of the accumulated effects of social disadvantage over time [7,18] and there is some evidence to support this [7,12,15,19,20]. However, the weight of evidence suggests that such inequalities narrow in old age. House et al. [6,14,21] suggest this is a result of the 'universality of biological frailty'; morbidity becomes compressed among affluent groups until late in life and hence inequalities are reduced at older ages. More broadly, however, selective mortality is believed to be the main cause of narrowing inequalities: disadvantaged people die younger leaving relatively robust survivors and hence a reduced gap between affluent and disadvantaged groups in mortality [5,22] and morbidity at older ages [7,9-11]. A few studies have tried to investigate the effect of selective mortality directly by imputing health data for decedents [10,11,17] or comparing the results of different kinds of models that included and excluded decedents and other dropouts until the point they died or dropped out [8,13]. Kim and Durden [15] combined both approaches and concluded that excluding people who dropped out underestimated education and income differences in depression, but that this effect diminished with age. Lynch [8] suggested that these patterns may be complicated by cohort effects i.e. inequalities widen with age but increase with younger cohorts, producing an artifactual appearance of convergence if age is modelled without adjustment for cohort. A further criticism of this debate is that many analyses use static measures of socioeconomic status (SES) from one (often distal) point in time, for example, baseline measures of income, occupation [10,18] or education [8,9,11,13,16]. Using more proximal measures of social class has been shown to dilute the effect of (SES) on health [19].

Table ​Table11 also shows the extent of socioeconomic change experienced by each of the cohorts during the course of the study. Unsurprisingly, the social class distribution of the oldest cohort (who age from 56 to 76 during the study) has only modestly changed. However, for the 1970s cohort, while nearly 40% of them came from a household headed by a manual worker when they were aged 15, only 17% of them were in a manual household at age 36. The 1950s cohort experienced more modest changes in their household class during the study; with just over a third being in a manual class at baseline when they were 36, and just over a quarter being in a manual class 20 years later when they were 56/7.

The final two columns show the data used in this paper (person-waves); the first column includes all waves in which respondents were alive, participated in the study and had valid data for all variables, whereas the final column includes extra person-waves for each wave in which a respondent had not participated because they had died.

Figure 1a, b show the growth curves for the cubic age function (from the fixed part of the models) with 95% confidence intervals (the shaded area), for the simple (binary) variable and for the continuous health outcome, which takes account of severity, for people in manual and non-manual classes at baseline. Below this, in Figure 1c, d, the 'health gap' between those in manual and non-manual classes at baseline is illustrated by showing the absolute difference in their predicted values calculated every 10 years. For both ways of measuring health, the estimates of self-assessed health for those in manual and non-manual social classes at baseline diverge around the age of 30. The confidence intervals no longer overlap (indicating a clear inequality) from around 40 years of age, after which the gap widens until the early 60s and then narrows with the confidence intervals over-lapping again around the age of 75. There also appears to be a cohort effect with an absolute increase in the level of reporting poor health for each of the older two cohorts, but testing interactions between social class and cohort suggested that there were no cohort differences in the inequalities experienced.

Figure ​Figure22 shows the effect of using a time-varying measure of social class. Inequalities emerge earlier, in the late 20s; the confidence intervals separate in the mid-30s and then continue to widen until the age of 65 for the simple (binary) outcome (Figure ​(Figure2a)2a) and into old age for the continuous measure, which captures severity (Figure ​(Figure2b).2b). However, there was a significant class, age and cohort interaction in this model, which does create some modest convergence at the oldest ages (not shown). Comparing Figure 1c, d with Figure 2c, d shows that the change in health inequalities is mainly the result of a reduction in inequalities from the mid-40s to mid-60s once more proximal measures of social class are employed.

The effect of expanding the health outcomes to include death is shown in Figure ​Figure33 for baseline social class. Not surprisingly, including decedents changes the shape of the health trajectory dramatically, with steep increases in the levels of poor health in both of the older cohorts, particularly for those in manual classes, and hence a larger gap between them and the non-manual group as people age. For the binary measure (Figure 3a, c), which makes a simple comparison between those reporting excellent/good health against those reporting fair or poor health or being dead, there appears to be a reduction in the health gap after the age of 65, although this remains statistically significant at the 95% level. Moreover, given the high probabilities of poor health (i.e. > 0.8) at these ages this may be due to the shape of the logit model imposing a ceiling on this group. For the continuous variable (Figure 3b, d), which has five categories ranging from 1 (excellent) to 5 (dead), capturing both the severity and the prevalence of health problems, inequalities in health continue to increase with age throughout the range studied.

Figure ​Figure44 shows the models for time-varying social class with the measures of health status incorporating death. The pattern is similar to that for baseline class, with much steeper increases in poor health or death in general, particularly for those from manual classes, as people age. Convergence after the age of 65 years is only seen for the binary outcome measure but the gap remains statistically significant (Figure 4a, c). When using the continuous measure of health, which is sensitive to severity, inequalities continue to widen into old age (Figure 4b, d). As before using recent measures of social class reduces the magnitude of the inequalities between manual and non-manual respondents in middle age, although the effect persists into old age for the continuous measure of health. There was a significant interaction between gender, age and class, in this last model only, which suggests a slightly steeper widening of health inequalities for women as they age (not shown).

A number of other studies also find that disadvantaged groups become ill at younger ages in relation to physical health [12], disease conditions [9] and self-assessed health,[19] but that having widened in middle age inequalities narrow again at older ages [9,13,14]. A few studies [7,12] have contradictory results depending on the health outcome and socioeconomic measure being considered. The studies that explicitly considered mortality and other selection effects [8,10,15,17] generally found that inequalities widened until older ages, although they did begin to narrow again. The only other study to investigate the effect of using time-varying class also found that this reduced inequalities in working age [19].

This paper has extended the age range and time period over which the issue of widening or converging inequalities in health has been considered in a single study. It directly examined the shape of age-health trajectories and unlike much of the existing literature, which has used linear or quadratic functions, found a cubic shape performed best, consistent with the only other study that tested for this [16]. Employing a cubic function allows greater flexibility in the shape of the health trajectory as people age. This suggests that there are improvements in self-assessed health in adolescence and early adulthood and steep increases in reporting poor health at older ages. The changes in prevalence with age may reflect both changes in actual health status and changes in respondents' perception of their health. For example, it has been shown that younger people associate being healthy with physical fitness and having healthy behaviours while older people tend to focus on physical functioning and mental well-being [33].

While many studies have speculated about the effects of survival bias, only a few have investigated it directly. We go further by including the available data on people who died and investigating what the pattern of social inequalities in health would look like if decedents were not excluded from the analyses after their death. We also use a measure of time-varying SES. This is the first study to take both issues into account at the same time.

There are a number of limitations to this analysis. First, while the Twenty-07 Study covers ages 15/6 to 75/6, it cannot investigate inequalities at very old ages; this is a limitation in much of the literature, reflecting the length and scope of many longitudinal studies. Secondly, the study was originally set in a predominantly urban area with a wide range of health experiences but generally poor health [24], and as such the results may not be generalisable to other kinds of places. For example, in the late 1980s, when the study began, there were relatively few people of minority ethnic backgrounds living in Clydeside. Whilst the situation in Clydeside is different today, the study may not reflect the experiences of a multicultural society. Similarly, there were few rural areas in the study and so it is difficult to know whether results would generalise to such places. Thirdly, this paper has made modest attempts to explore cohort effects on health inequalities as people age, but the three cohorts only had partial overlap in their ages so these conclusions are limited. Fourthly, we have employed head of household social class based on occupation for our SES measure. However, using occupation to categorise people can be more problematic for some groups of the population than others. For example, the last occupation of those who have retired or are not working for other reasons may not reflect their current circumstances [34]. Also, while for some respondents their social class reflects both their household socioeconomic circumstances and their own occupational exposures, for others it is purely a measure of household socioeconomic status. Repeating these analyses with other measures of socioeconomic status would help to validate the results beyond this specific measure. Finally, only 15% of all respondents, and 37% of the 1930s cohort, had died by the final wave. This is a small proportion of the sample as a whole, but a significant proportion of the oldest cohort. Nevertheless, the effects of survival bias may change as more respondents die. However, since dying young is more prevalent among those from manual classes this analysis is likely to have incorporated the most significant biases caused by unequal mortality at younger ages. Noymer [35] criticises Beckett [11] and other studies for imputing health data for decedents arguing that they are likely to differ significantly from survivors and hence the approach is invalid. However, although we appreciate the crudeness of this approach, we have not tried to estimate the health of decedents as if they had stayed alive, but to extend our health measures to include the ultimate poor health state - death [23].

There are two ways in which this analysis might underestimate the extent of social inequalities in health. First, it is based on self-assessed health, which has been shown to be a good predicator of mortality [23] and morbidity, [36] but the way people answer the question may change with age and period [37], and there is conflicting evidence about whether people's answers vary by their socioeconomic background. Some studies have found no [38], others small [39] and others significant [40] differences in the association between self-assessed health and mortality by socioeconomic status. In addition, there is some evidence that more affluent groups over-reported poor self-assessed health when compared to more objective measures of health status [41]. Overall, therefore, these studies suggest that investigating health inequalities based on self-assessed health measures may under report their magnitude. Secondly, those respondents who are still alive but dropped out of the study are only included until they drop out. Such people may have different health trajectories to those who remain in the study, however, as Table ​Table11 shows they tend to be in poorer health and from manual social classes, which suggests excluding them will again underestimate the inequalities gap between manual and non-manual classes.