Socioeconomic indicators of health inequalities and female mortality: a nested cohort study within the United Kingdom Collaborative Trial of Ovarian Cancer Screening (UKCTOCS)

This is one of the largest studies examining the association between individual-level education and area-level indicators of socioeconomic position and all-cause and cause-specific mortality that we are aware of. Our analysis of over 50,000 older women residing in the UK between 2001–12 found both IMD rank and education to be positively associated with all-cause mortality with a clear dose response relationship. IMD rank remained positively associated with all-cause mortality after taking into account a wide range of potential confounders such as age, smoking, alcohol consumption, BMI, co-morbidities and mutually adjusting for education. There was a clear gradient even within the context of a screening trial with a significant ‘healthy volunteer effect’ [26] with the hazard ratio for all-cause mortality for those living in the most deprived areas (1.42) being higher than reported in previous studies. The survival curves (Figures 2 and 3) suggest that neither educational nor IMD differences in survival converge at later years of the study (when populations get older). We also found a statistically significant inverse trend with CVD mortality and all cancer mortality after adjusting for individual-level education. The results support increasing evidence that place of residence of individuals affects their health, [16,17,32,33] and that the IMD scale can be an effective tool to assess this effect.

A key strength of our study is that unlike previous studies which only examined education (individual-level variable) or IMD rank (area-level variable) as indicators of deprivation, we used both and therefore were able to account for individual social circumstances and area-level deprivation. We were able to adjust for a large number of individual-level factors such as BMI, comorbidity, smoking and alcohol use unlike analyses using national statistics. The scale of this analysis of 50,029 women was significantly larger than comparable previous studies with sample sizes ranging from 1,811–31,264 [13,25]. The fact that IMD was obtained from an external source while education was collected through postal questionnaire from trial participants and the large number of clusters (6954) ensured that the choice of method accounting for clustering of data did not substantially influence the finding, allowing a separation between the effects of IMD and education. An additional methodological strength was low proportion of missing data (8.3%). Sensitivity analysis was undertaken with imputation of the missing results as well as inclusion of the entire cohort (79,006) including those who had only completed the baseline questionnaire (24,467) with no change in results.

Certain design characteristics of our screening trial need to be considered when interpreting the current findings. While almost 1 million women were randomly invited from ‘age/sex’ registers of 27 health authorities [27], those that participated may not be entirely representative of the women living in England. Health authorities in the North East and South West England were not included as there were no trial centres in these regions. Our cohort seems to be somewhat different from the general population with 97.8% white participants compared to 85.5% of the UK population [34]. Alcohol consumption and smoking were also much higher in our cohort. Alcohol consumption (>7 units per week) was 40.4% compared to 19% (in women aged 45–64) or 6% (aged over 65) and smoking was 44.3% versus 37% in the UK population (all ages) [35,36]. The IMD scores of the invitees were higher than the national average, possibly due to an over-representation of urban centres [26]. This indicates that the 50,029 (17.3%) of those eligible and willing to participate, and included in this nested cohort study, were less deprived which could potentially bias the study. A significant ‘healthy volunteer effect’ has been reported on this cohort [26] which could have led to an underestimation of the impact of area-level deprivation on mortality. In addition, underestimation might have also resulted from inclusion of the variables in the models as confounders when they were at least in part mediators. For example, deprivation may well impact on mortality, in part, through greater access to fast food, alcohol and tobacco outlets, encouraging poor health choices, which in turn cause comorbidity and ultimately lead to higher death rates. Education was the only available indicator of personal deprivation in UKCTOCS. There is controversy whether education alone without income data is a sound indicator of individual deprivation [14,37,38].

The use of the Index of Multiple Deprivation as a predictor of health has been debated. However, the English IMD 2007 used in this analysis has been developed over the years to a high standard. It has seven complex and multifaceted domains - income and employment each contributing to 22.5% of the IMD score; barriers to housing/services, crime and living environment 9.3% each; and education and health deprivation/disability 13.5% each. While the ‘health deprivation/disability’ domain is based on years of potential life lost, comparative illness and disability ratio, measures of acute morbidity and the proportion of adults under 60 suffering from mood/anxiety disorders [39] and could contribute to the association between IMD rank and mortality, as is clear from above, it only contributes to a tenth of the score. It would have been interesting and useful to look at each domain separately but linkage to this data was not possible. Full multilevel analysis is probably the most appropriate method to answer the question of whether there are associations between area-level deprivation and mortality over and above individual-level measures. However, multilevel Cox regression using multiply imputed data was not possible in the current version of the statistical software available to the authors. We accounted for potential geographical clustering of the data by calculating standard errors using a cluster-robust estimator. We believe that this method, together with the large number of clusters, ensured that clustering did not substantially influence the results and we were able to separate between the effects of IMD and education.

Our study’s results show a strong association between area-level deprivation expressed by IMD and all-cause mortality which is in line with previous findings [40,41]. Multiple domains of IMD might represent several possible explanations and mechanisms that can link social disadvantage and mortality including differences in wealth and income, risky health behaviours, access to healthcare and treatment, accumulation of social disadvantage in different stages of lifecourse or psychosocial factors such as control, autonomy or trust [42]. IMD rank showed a dose response relationship on subgroup analysis of cancer and cardiovascular disease mortality, with the magnitude of effect increasing as deprivation score increased. There is much evidence to support the concept that socioeconomic status [43-45] and area-level measures of deprivation [46,47] are related to cancer survival. The association with cardiovascular mortality was not significant although there was a clear gradient in keeping with other reports [48-50]. The non-significant results may be due to the relatively low number of cause-specific deaths. In our cohort (median age 60) there were approximately three cancer deaths to every CVD death, in keeping with recent national statistics which report cancer to be the leading cause of death in those aged below 85 [51,52].

While education showed a clear social gradient with all-cause mortality in keeping with previous literature, [13,14,37,38,53,54] when mutually adjusted for IMD rank, a reduced, non-significant effect was seen (although p for trend remained significant probably due to the large sample size). This is in contrast to most other studies where education was found to be a significant predictor of mortality. However, our findings in older women are in keeping with those of Steenland et al. [19] who found that in multivariable analysis of a large US sample, education remained an important predictor only among men but not women. Others have also reported a weak association of education with mortality among women [55-57]. Variations in education level could result in variations in mortality through both access to higher paying occupations and through health literacy.