British Medical Journal 
© British Medical Journal 1996.

----------------------------------------------
Volume 313(7066)             9 November 1996             pp 1177-1180
----------------------------------------------

Do socioeconomic differences in mortality persist after retirement? 25 Year
follow up of civil servants from the first Whitehall study
[Paper]
Marmot, M. G.;  Shipley, Martin J.
Department of Epidemiology and Public Health, University College London Medical
School, London WC1E 6BT (M G Marmot, professor, Martin J Shipley, senior
lecturer in medical statistics).
Correspondence to: Professor Marmot.
BMJ 1996;313:1177-80

----------------------------------------------

Outline

Abstract
Introduction
Subjects and methods
Results
Discussion

DECLINE IN PREDICTIVE POWER OF GRADE

Key messages
REFERENCES

Graphics

Table 1
Table 2
Figure 1

----------------------------------------------

Abstract

Objective: -To assess the risk of death associated with work based and non-work
based measures of socioeconomic status before and after retirement age.

Design: -Follow up study of mortality in relation to employment grade and car
ownership over 25 years.

Setting: -The first Whitehall study.

: -18 133 male civil servants aged 40-69 years who attended a screening
examination between 1967 and 1970.

Main outcome measure: -Death.

: -Grade of employment was a strong predictor of mortality before retirement.
For men dying at ages 40-64 the lowest employment grade had 3.12 times the
mortality of the highest grade (95 percent confidence interval 2.4 to 4.1).
After retirement the ability of grade to predict mortality declined (rate ratio
1.86; 1.6 to 2.2). A non-work based measure of socioeconomic status (car
ownership) predicted mortality less well than employment grade before retirement
but its ability to predict mortality declined less after retirement. Using a
relative index of inequality that was sensitive to the distribution among
socioeconomic groups showed employment grade and car ownership to have
independent associations with mortality that were of equal magnitude after
retirement. The absolute difference in death rates between the lowest and
highest employment grades increased with age from 12.9 per 1000 person years at
ages 40-64 to 38.3 per 1000 at ages 70-89.

Conclusions: -Socioeconomic differences in mortality persist beyond retirement
age and in magnitude increase with age. Social differentials in mortality based
on an occupational status measure seem to decrease to a greater degree after
retirement than those based on a non-work measure. This suggests that alongside
other socioeconomic factors work itself may play an important part in generating
social inequalities in health in men of working age.

----------------------------------------------

Introduction

Do social inequalities in death rates that are apparent at younger ages persist
into old age? It might be imagined that as the cumulative probability of death
for each person approaches 100 percent social and other predictors of differences
in death rates would discriminate less well. In Britain much of the analysis of
social class differences in adults has used the registrar general's classification
based on occupation and has therefore been confined to people of working age.
Other social classifications show differentials in mortality continuing beyond
age 65, but differences are reduced. [1-3] In the United States the shorter life
expectancy of people with less education continues but the black-white
differential in mortality (largely social in origin) is reversed after age 75.
[4]

A major question with occupation based social class differentials in mortality
is the extent to which resulting mortality differences are due to occupation
itself or to broader social differentials. [5] If due to work, social differences
in mortality should narrow after retirement; if due to other factors associated
with socioeconomic position, the narrowing may be less. We examined this in data
from the 25 year follow up of British civil servants in the first Whitehall
study. By using a work based measure of status (employment grade) the Whitehall
study showed an inverse social gradient in mortality that was steeper than that
observed nationally with the registrar general's classification of social class.
[6] A non-work based measure (car ownership) added to the predictive power. [7]
By using both work based and non-work based socioeconomic measures and analysing
mortality differentials before and after usual retirement age we can examine the
persistence of social gradients in mortality into older ages and the specific
contribution of a work-based measure.

Subjects and methods

In the Whitehall study 19 019 men aged 40-69 years attended the initial
screening between September 1967 and January 1970. Men were classified into four
employment grades: administrative, professional and executive, clerical, and
other. "Other" included messengers and other unskilled manual workers. For some
analyses we grouped the four employment grades as high (administrative and
professional and executive) and low (clerical and other). For 886 men from the
diplomatic service and the British Council employment grading was different.
These men were excluded from analysis.

Data regarding car ownership were available from answers to the question, "Do
you own a car?" The questionnaires used in the study were modified during the
study, and data on car ownership were collected only from the 70 percent of men
seen during the middle period of the survey. Other studies have used access to a
car rather than ownership of a car as a classification of social class. We
expect there to be almost perfect agreement between these two definitions, as
the Whitehall study men were middle aged and most likely the head of household.

Records from 99.3 percent of the remaining 18 133 men were identified and
flagged at the National Health Service Central Registry, which notified us of
all deaths up to the end of January 1995. A total of 18 001 men were therefore
followed up for a minimum of 25 years. Each man's follow up period was
partitioned into single years of observation. For each individual year of follow
up a new record was created, consisting of each man's current age at risk
together with his employment grade and car ownership details and length of
follow up during that year. Deaths were allocated to the appropriate current age
category. This expanded dataset was used as the basis for all analyses of
mortality. As comparatively few deaths occurred in men aged over 90 analyses
were restricted to ages 40-89. Death rates were calculated by using person years
at risk as outlined above.

The groups before and after retirement spanned age ranges of 25 and 20 years
respectively. We therefore adjusted for the residual confounding effects of age
by using five year age groups. Men at lower risk of death had, on average,
longer follow up. Hence when comparing employment grades or men with and without
a car the lower risk subjects had, on average, longer follow up. Over the follow
up period death rates in England and Wales declined. In all analyses, therefore,
estimated rate ratios for employment grade and car ownership were adjusted for
length of follow up to avoid these effects being overestimated.

The proportions of men in the lower employment grades and without a car differed
greatly according to age at entry into the study. These proportions also
differed for given ages at entry. When comparing the effects of employment grade
and car ownership the rate ratios for these effects may be expected to be larger
for classifications which produce smaller, more extreme groups at the margins of
the population. We overcame this problem by using the relative index of
inequality, based on employment grade and car ownership. [8] For each measure
the socioeconomic position of each group within each five year age band is
assigned a value between zero and one according to the proportion of subjects in
the population who are above the midpoint of that group. Values of zero and one
therefore correspond to subjects who are at the top and bottom of the socioeconomic
hierarchy. Rate ratios estimated by this method therefore show the relative
death rates for subjects at the bottom of the socioeconomic hierarchy compared
with those at the top. The relative index of inequality has the additional
benefit that it enables employment grade categorised into four levels to be
compared directly with car ownership categorised in two levels.

Adjusted rate ratios, their confidence intervals, and tests for linear trend in
the rate ratios were computed by Poisson regression, fitted by using the
statistical package GLIM. [9]

Results

(Table 1) shows (by age at death) death rates per 1000 person years for the four
employment grades together with the rate ratios for the lowest versus the
highest grades and those for the bottom versus the top of the socioeconomic
hierarchy by using the relative index of inequality based on grade. For men
dying at 40-64 years of age (before retirement in most cases) there was a strong
gradient in death rates across the four employment grades, corresponding to a
rate ratio of 3.06 (95 percent confidence interval 2.5 to 3.7) based on the
relative index of inequality. After retirement the relative increase across the
grades was not so steep and the corresponding rate ratio declined to 1.77 (1.6
to 2.0). Men aged 65-69 years included some in work and some retired, and we
therefore kept them separate in the analysis. The grade gradient for this group
fell between those for the pre-retirement and post-retirement groups.

----------------------------------------------
Table 1. All cause mortality per 1000 person years (No of deaths), attributable
risk, and rate ratio by employment grade and age at death
----------------------------------------------

The lower rate ratio after retirement than before retirement was consistent more
with a change after retirement than simply with a weakening of the social
differential with age. Before retirement there was no change with age: the rate
ratios (by using the relative index of inequality based on grade) were 3.06 at
ages 45-49, 2.66 at 50-54, 3.64 at 55-59, and 2.94 at 60-64 (test for trend: P
greater than 0.5). Follow up before age 45 was not used in this comparison as
only eight deaths occurred in this age range. After retirement there was a small
suggestion of a decline with age, though the trend was not significant (P =
0.43): rate ratios were 1.81 at ages 70-74, 1.83 at 75-79, 1.74 at 80-84, and
1.45 at 85-89.

Given the rise in death rates with age the absolute difference in death rates
(attributable risk) between grades was greater after retirement. For example,
the lowest employment grade had a death rate higher by 12.9 per 1000 person
years at ages 40-64, increasing to 38.3 at ages 70-89 Table 1. Overall, 30
percent of men in the administrative grade died during the 25 year follow up
compared with 69 percent of men in the lowest grade.

(Table 2) shows the death rates according to car ownership and age at death, the
rate ratios for car ownership, and the rate ratios based on using the relative
index of inequality. Before retirement the rate ratio for non-owners of cars
versus car owners was much less than that between the "other" and administrative
grades (rate ratios 1.57 v 3.12 respectively; P less than 0.001). These two
grade levels were more extreme socioeconomic groups in the population compared
with the car ownership groups. However, using the relative index of inequality
reduced but did not remove this difference (rate ratios 2.41 v 3.06 respectively;
P = 0.13). After retirement the rate ratios based on car ownership declined,
though the trend was not as large as for employment grade (comparison of trends:
P = 0.17) and the predictive effects of car ownership and grade were equal but
still highly significant.

----------------------------------------------
Table 2. All cause mortality per 1000 person years (No of deaths) and rate ratio
by car ownership and age at death
----------------------------------------------

(Figure 1) shows the rate ratios for the combined effects of grade (bottom two
grades versus top two grades) and car ownership by age at death. The two
measures of socioeconomic status show independent effects. The last columns in
Table 1 and Table 2 show that these independent effects were highly significant
both before and after retirement--that is, the confidence intervals for the rate
ratios do not overlap 1.0.

----------------------------------------------
Figure 1. Rate ratios (logarithmic scale) for all cause mortality by grade and
car ownership according to age at death
----------------------------------------------

As with grade the absolute difference in death rates between non-owners of cars
and car owners was greater after retirement.

Discussion

Among men aged 40-69 years at entry to the Whitehall study two different
socioeconomics measures continued to be associated with differences in death
rates over a 25 year follow up period to age 89. Lack of ownership of a car was
associated with a 57 percent higher mortality before retirement and 34 percent
higher mortality after retirement. Grade of employment was inversely associated
with mortality, with a 212 percent higher mortality in the lowest grade compared
with the highest grade up to age 64. The relative difference was less after age
69 (when most men would have retired) than up to age 64 (before retirement for
most men). Even with the decrease in relative mortality after retirement there
was still an 86 percent increase in mortality among men in the lowest grade
compared with the highest.

Declining social differences in mortality ratios with increasing age may be
expected if there is selective removal of sick people at younger ages leaving a
survivor population at older ages. [4] Other analyses of these data show that
this was unlikely to be the case in our study, with only a small percentage of
the decline in the grade rate ratios possibly due to a survivor effect (P J van
de Mheen et al, unpublished observations). An alternative explanation might be
that subjects at older ages had on average a shorter time between examination
and death. However, we adjusted for length of follow up, which made little
impact on the declining rate ratios for grade with age.

In detailed analyses of the Office of Population Censuses and Surveys longitudinal
study that followed up a 1 percent sample of the 1971 census population of
England and Wales, Fox et al showed that social differences in mortality at
older ages could not easily be attributed to sick people being downwardly mobile
socially. [10] They also argued that if sickness influenced social position
rather than the reverse this effect would be most evident in the first five
years of mortality follow up after determination of social position. In our
study 91 percent of the deaths occurred five or more years after the baseline
examination at which grade was determined.

Grade of employment was a powerful predictor of mortality and morbidity in both
the Whitehall and Whitehall II studies. [6,11] Though grade is a measure of
position in the occupational hierarchy, it reflects more than experience of
work. It relates to educational and social background, status, self esteem, and
income and associated living conditions. The Whitehall II study measured these
directly in an attempt to sort out which may be most important in generating the
social gradient in ill health.

The present analyses suggest two different ways of approaching the question of
the contribution of work: examining the degree to which a work based and a
non-work based measure of socioeconomic position predict mortality; and
examining mortality before and after presumed retirement age. Car ownership was
chosen as a non-work based socioeconomic measure because it was shown to predict
mortality in the Whitehall study [7] and is closely related to car access, which
predicted mortality in the Office of Population Censuses and Surveys longitudinal
study. [12] In 1971, when the longitudinal study began, about half the
households in Britain had access to a car. In general these were the wealthier
households. Car ownership therefore served well as a general socioeconomic
measure.

DECLINE IN PREDICTIVE POWER OF GRADE

Pre-retirement, employment grade (the work based measure), and the non-work
measure independently predicted mortality. After retirement the ability of the
general social measure to predict declined less than the predictive power of the
work based measure. This decline in predictive power of employment grade seems
more compatible with an effect of retirement than of diminution with age given
the lack of an age grade interaction before retirement and the shallow
diminution with age of the grade association after retirement.

This comparison of grade and car ownership needed to take into account the fact
that there were four employment grades but only two car ownership groups. Other
things being equal, comparing smaller extremes of a distribution will give
apparently greater predictive power than comparing two halves. We therefore used
the relative index of inequality, which allows direct comparison between the
predictive power of the two measures by using the slope of the relation with
mortality independent of the size of the groups. The ability of grade to predict
independent of the effect of "car" declined significantly after retirement,
which was not the case for car ownership independent of the effect of grade.
However, the difference between the slopes was not significant and, though
plausible, it is perhaps an overinterpretation to conclude that the effect of
work on mortality but not that of other factors associated with socioeconomic
status declines after retirement.

There is clear evidence in these data of a non-work effect. Adjusting the
association of car ownership with mortality for the effect of employment grade
is probably a good way to remove the effects of work. It is therefore likely
that the association between car ownership and mortality after grade adjustment
represents a minimum estimate of the effect on mortality of factors other than
those associated with work. The reverse is not the case. Adjusting the
grade-mortality association for "car" does not mean that the residual association
is all due to work, though some of it is likely to be. The problem in separating
the effect of work from other influences associated with social position is that
grade is a guide to work and social circumstances outside work, but not all of
these non-work effects are summed up in the car ownership measure.

There are two conclusions to be drawn from these analyses. Firstly, important
socioeconomic differences in mortality persist beyond retirement age, at least
up to age 89. On an absolute scale these differences increase with age.
Secondly, social differentials in mortality based on an occupational status
measure decrease after retirement whereas those based on a non-work measure seem
to decline less. This suggests that, alongside other socioeconomic factors, work
itself may play an important part in generating social inequalities in health in
men of working age. We cannot conclude from these data that improving socioeconomic
circumstances of older people would necessarily reduce social differentials in
mortality. There are, however, other reasons for believing that this is a
desirable policy objective.

Key messages

-Socioeconomic differences in mortality persist up to age 89

-Relative differences in mortality between low and high employment grades are
less after retirement, suggesting the importance of work in generating
inequalities in health

-A non-work based measure of socioeconomic status (after adjustment for
employment grade) continues to predict relative differences in mortality after
retirement

-Absolute differences in mortality between less and more advantaged groups
increase at older ages

Funding: MGM is a Medical Research Council research professor; MJS is supported
by the British Heart Foundation.

Conflict of interest: None.

(Accepted 10 September 1996)

REFERENCES

1. Goldblatt P, ed. 1971-1981 Longitudinal study. Mortality and social
organisation. London: HMSO, 1990. 

2. Kaplan GA, Seeman TE, Cohen RD, Knudsen LP, Guralnik J. Mortality among the
elderly in the Alameda Country study: behavioral and demographic risk factors.
Am J Public Health 1987;77:307-12. [Medline Link] 

3. Phillimore P, Beattie A, Townsend P. Widening inequality of health in
northern England, 1981-91. BMJ 1994;308:1125-8. [Fulltext Link] [Medline Link]
[BIOSIS Previews Link] 

4. Guralnik JM, Land KC, Blazer D, Fillenbaum GG, Branch LG. Educational status
and active life expectancy among older blacks and whites. N Engl J Med
1993;329:110-6. [Fulltext Link] [BIOSIS Previews Link] 

5. Fox AJ, Adelstein AM. Occupational mortality: work or way of life? J
Epidemiol Community Health 1978;32:73-8. [Medline Link] 

6. Marmot MG, Shipley MJ, Rose G. Inequalities in death--specific explanations
of a general pattern. Lancet 1984;i:1003-6. [Medline Link] 

7. Davey Smith G, Shipley MJ, Rose G. Magnitude and causes of socioeconomic
differentials in mortality: further evidence from the Whitehall study. J
Epidemiol Community Health 1990;44:265-70. [Medline Link] 

8. Kunst AE, Mackenbach JP. The size of mortality differences associated with
education level in nine industrialized countries. Am J Public Health 1994;84:932-7.
[Medline Link] [BIOSIS Previews Link] 

9. Numerical Algorithms Group. GLIM system release 3.77 manual. Oxford: NAG,
1987. 

10. Fox J, Goldblatt P, Jones D. Social class mortality differentials: artifact,
selection or life circumstances? In: Goldblatt P, ed. Longitudinal study.
Mortality and social organisation. London: HMSO, 1990:99-108. 

11. Marmot MG, Davey Smith G, Stansfeld S, Patel C, North F, Head J, et al.
Health inequalities among British civil servants: the Whitehall II study. Lancet
1991;337:1387-93. [Medline Link] 

12. Goldblatt P. Mortality and alternative social classifications. In: Goldblatt
P, ed. Longitudinal study. Mortality and social organisation. London: HMSO,
1990:163-92. 

----------------------------------------------