 
|
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Journal of the New Zealand Medical Association,
28-January-2005, Vol 118 No 1208
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Tony Blakely, Jackie Fawcett
Mortality differences by occupational class have been the
mainstay for monitoring changes over time, and differences between countries, in
social inequalities in health—both in New
Zealand1–5 and internationally,
particularly Europe.6–13 However, there
are numerous biases that may arise in such comparisons due to the collection of
occupation and the use of either ‘unlinked’ (cross-sectional) or
‘linked’ (cohort) analyses of census and mortality data.
Unlinked analyses
involve determining the number of deaths in each occupational class from
mortality data (the numerator) and the number of people in each occupational
class from census data (the denominator). Mortality rates are then calculated by
dividing the former by the latter within each occupational class. Examples of
unlinked studies include the Decennial Supplement series in England and
Wales6,14 and those by Pearce and colleagues in
New Zealand for 1974–1978, 1985–87, and
1995–97.1–5,15,16 These latter New
Zealand analyses have found a two-fold difference in mortality for the lower
occupational classes compared to the higher occupational
classes1,2,5,15,16, and possibly increasing
mortality gradients by occupational class from 1975–1977 to
1995–1997.3
Linked analyses
involve record linkage of census and mortality data, thereby creating cohort
studies of census populations followed-up for mortality. Occupational class is
assigned using the census data, negating the need to use occupation codes on
mortality data. Examples of linked studies include the UK OPCS Longitudinal
Study17 and the Scandinavian record linkage
studies.18–22
Unlinked analyses are prone to numerator-denominator bias
because of differences in the way that occupation is recorded between mortality
and census data. For example, census and mortality data may differ with regard
to: who provides the occupation (self-identified or next of kin); coding rules;
and whether current, usual or last occupation is recorded. The direction and
magnitude of numerator-denominator biases that consequently arise vary between
countries. In the UK, it is thought to be
modest.23,24
Linked analyses using current occupation to calculate
occupational class mortality gradients can underestimate the gradient due to
differential health selection out of the labour
force.25–27 This differential arises
because unwell people in lower occupational classes (e.g. labourers) are more
likely to exit the active labour force than people with the same level of poor
health in higher occupational classes (e.g. lawyers). As those not in the labour
force are excluded from occupational class analyses, differential health
selection will bias the association of occupational class with mortality.
However, there are methods to adjust for this health selection
bias.10,28
Mortality and census records have been anonymously and
probabilistically linked in the New Zealand Census-Mortality Study
(NZCMS).26,29 This linkage provides the
opportunity to measure mortality gradients by occupational class using linked
census-mortality data. However, such linked analyses may be biased by
differential health selection as occupation on the New Zealand census is
current occupation rather than last or
usual occupation. Conversely, the previous unlinked analyses that utilise usual
or last occupation on death certificates may be biased by numerator-denominator
bias.
The objectives of this paper are:
MethodsCensus
data—For this paper, analyses were limited to 25 to 64-year-old
males. Occupational class was assigned to census data using the New Zealand
Socioeconomic Index (NZSEI).30,31 The NZSEI
ranks occupations by socioeconomic position on a scale of 10 to 90. We used
cut-points of 30, 40, 50, 60, and 70 to form six occupational
classes.30 We also created a separate class for
farmers.
Record linkage and
mortality data—Mortality records were anonymously and
probabilistically linked to the 1991 census using
Automatch®32 as described
elsewhere.33,34 Briefly, 5,844 of 8,145 (71.7%)
eligible male decedents who were aged 25–64 years on 1991-census night and
died in the second and third years following census night (5 March 1991) were
successfully linked to a census record using these matching variables. Greater
than 95% of linked census and mortality records were estimated to be correct
linkages.35 Census records linked to a
mortality record where death occurred in the first year of follow-up were
discarded to militate against health selection biases, and due to inaccurate
recording of occupational codes on mortality data during 1991.
Cross-tabulating
census and mortality data occupational class (Objective 1)—To
investigate for numerator-denominator bias, we cross-tabulated the mortality
data occupational class (usual occupation) by census data occupational class
(current occupation on the census night preceding death) among the subset of
linked census-mortality records.
Analyses of the
association of occupational class with mortality—We used logistic
regression to measure occupational class mortality gradients. As death is
uncommon among 25–64 year olds over a 2-year period, the odds ratios were
close approximations of the risk ratio. Analyses were conducted separately for
25–44 and 45–64 year olds. For the ‘linked’ analyses, we
simply used census data occupational class (where non-missing) as the
independent variable. The ‘unlinked’ analyses also used the linked
data, but with the occupational class from the mortality data as the independent
variable.
We also calculated relative indices of inequality
(RIIs), details of which are available
elsewhere.36 The RII is a summary relative risk
measure that is derived by fitting a regression line to the rates of death (or
in our paper the odds ratios of death) for each occupational class on the
midpoints on a cumulative rank distribution for each occupational class. For
example, if 10% of the population are in occupational class 1 then this class is
assigned an x-value of 0.05 (i.e. the midpoint of the range 0 to 0.1).
If occupational class 2 comprised 30% of the
populations ion this example, its assigned x-value would be 0.25 (0.1 [for class
1] plus half of 0.3). The slope of this regression line is the difference in the
estimated mortality rate (or odds ratio unit in out paper) for the hypothetical
person of lowest occupational class compared to highest occupational class. The
RII is then calculated as the [slope + intercept] / [intercept]. The regression
was weighted by the inverse of the variance of the class rate.
Adjusting linked
analyses for possible linkage bias (Objective 2)—Not all eligible
mortality records could be linked to a census record, and it was therefore
possible that the odds ratios by occupational class could be biased if the
linkage success varied by occupational class. We quantified this possible
linkage bias by comparing the proportion of mortality records linked to a census
record across occupational classes (usual occupational class according to
mortality data), using log-link
regression.26,34 The relative risks from these
regressions were then used to adjust the odds ratios in the cohort analyses for
linkage bias.
For example, if the relative risk of linkage for class
6 decedents compared to class 1 decedents was 0.95, and the observed odds ratio
of death in the cohort analysis was 1.60, then the linkage bias adjusted odds
ratio was 1.6/0.95 = 1.68.
Adjusting for
exclusion of economically inactive (Objective 2)—We used a
previously published method to adjust our linked current occupational class
results to approximate what would have been observed had usual occupation been
available on the census.10,28 This method uses
external survey data of the proportion of each (usual) occupational class in
current employment and the relative risk of mortality for the non-employed
compared to the employed. The New Zealand Household Labour Force Survey (HLFS;
1991 data) collects data on current occupation and, if no current occupation,
the last occupation within five years.
Accordingly, we estimated that 2%, 2%, 5%, 11%, 12%,
and 12% of males aged 25–44 years in classes 1 to 6 respectively had no
current occupation, and likewise 10%,
13%, 15%, 21%, 25%, and 16% among 45–64 year old males. The
relative risk of mortality in the NZCMS for the non-employed compared to
employed were 2.77 and 2.35 for 25-44 and 45–64 year olds, respectively.
Comparing linked and
unlinked analyses (Objective 3)—We conducted a close approximation
of an unlinked analysis by using the census occupational class for all
non-linked census records, but using the mortality data occupational class for
the census records linked to a mortality record. As this method only allows
analyses on the actually linked mortality records, we also adjusted these
analyses for linkage bias using the same methods as above.
ResultsHow
does occupational class recording on census data differ to that on mortality
data?—Table 1 shows the cross-classification of occupational class
for 25–64 year old male decedents according to census data (using current
occupation on 1991 census night) and according to mortality data (usual or last
occupation). Neither the mortality nor the census data can be considered the
‘gold standard’. However, we have presented the data from the
perspective of the mortality data ‘usual or last’ occupational
class, and how census ‘current’ occupational class classification
performs with respect to the mortality data.
Many of the possible cross-classification cells in Table 1
have been merged to meet privacy requirements of Statistics New Zealand that all
tabular data be randomly rounded to a near multiple of three, and a minimum cell
size of six.
The main points to note from Table 1 are, firstly, more of
the 5844 decedents had an occupational class assigned on their mortality record
(4,932 or 84%) than on their corresponding census record (2,754 or 47%). Second,
the percentage of decedents on the mortality data with the same occupational
class on census data decreased by class
from 43% in class 1 to 24% in class 6. Third, decedents in lower occupational
classes on the mortality data were less likely to have had a current
occupational class recorded on census data (38% in class 6 compared to 62% in
class 1). Fourth, and offsetting the previous point, it was more common among
men with no usual occupation on mortality data to be assigned to census
‘current’ class 5 and 6 (8% and 6%, respectively) than to census
class 1 and (1% and 3%, respectively).
Despite these considerable differences in occupational class
assignment between the census and mortality data, the net result was that the
ratio of the total census count to the total mortality count for each class is
around 0.6 for most classes (final column of Table 1), although there was still
some notable variation that is described further below.
What is the association of
class with mortality using analyses of the linked census-mortality
data?—Table 2 shows the distribution of census counts and linked
census-mortality records by census ‘current’ NZSEI occupational
class, separately for 25–44 and 45–64 year olds males. The linked
deaths in Table 2 are equivalent to the deaths in the bottom row of Table
2—e.g. 240 deaths in census NZSEI class 1 from Table 1 equals 66 deaths
among 25-44 year olds plus 174 deaths among 45–64 year olds as shown in
Table 2. Odds ratios from a logistic regression adjusting for 5-year age groups
and ethnicity are also shown in Table 2. There was a mostly monotonic gradient
of increasing mortality odds moving from class 1 to 6, with the observed
mortality odds in class 6 being nearly 50% greater than in class 1 for both
25–44 and 45–64 year olds. The relative indices of inequality (RIIs)
for 25–44 and 45–64 year old males were 1.68 and 1.61 (Table 2).
That is, the expected mortality rate for the males of lowest occupational class
was 68% or 61% higher than the expected mortality rate of the males of highest
occupational class.
What is the impact of
adjusting linked analyses for linkage bias?—Figure 1 presents the
association of mortality with occupational class for a range of analyses, for
both 25–44 year olds and 45–64 year olds. RIIs for each type of
analysis are placed just beneath the x-axis. The first sets of bars for each age
group is the age and ethnicity adjusted odds ratios presented in Table 2. The
second set adjusts the former results for linkage bias. This adjustment
increases the strength of association of class with mortality by 34% for 25-44
year olds (i.e. [1.91–1.68] / [1.68–1]) and by 8% for 45–64
year olds. Therefore, there was notable linkage bias among 25–44 year
olds, but relatively minor linkage bias among 45–64 year olds.
What is the impact of
adjusting linked analyses (i.e. current occupation) for exclusion of the
economically inactive?—The third set of columns for each age-group
show the impact on the linked analyses of adjusting for economic inactivity
(Figure 1). The mortality gradient increases (in excess terms over and above a
null RII of 1.0) among 45 to 64 year olds by 42% from an RII of 1.66 to 1.94,
and among 25 to 44 year olds by 58% from an RII of 1.91 to 2.44. Therefore, it
appears that there is notable underestimation of occupational class mortality
gradients using census (current occupation) data if one is trying to estimate
the gradient according to usual occupation.
Are linked and unlinked
analyses similar?—The final sets of columns show the results for
the unlinked analyses (i.e. closely approximating just using the mortality data
as numerators and census data as the denominators). Among 45–64 year olds,
the unlinked analysis produced a gradient similar to that for linked analyses
adjusted for both linkage bias and economic inactivity. Note the higher odds
ratio for class 4 in the unlinked compared to linked analyses—this is
predicted by the lower census to mortality ratio for class 4 in Table 1. (The
45–64 year olds dominate the cross-classifications in Table 1.) The lower
relative risk for class six compared to class five in the unlinked analysis is
also consistent with the varying census to mortality ratios of 0.55 and 0.64 for
these two classes as shown in Table 1.
Among 25–44 year olds, a visual inspection of the
unlinked analysis is distorted by a (probably spurious) higher mortality rate
than might be expected in class 1—the referent group. Replotting the
histograms using class 6 as the referent group, the visual impression is of
similarity between the unlinked analysis and both the adjusted linked analyses.
The RIIs (that do not rely on treating one category as the reference group) also
confirm this interpretation, with a RII for the unlinked analysis of 2.03 being
between the RIIs of 1.91 and 2.44 for the adjusted linked analyses.
DiscussionOnly 47% of linked deaths had an
occupation recorded on census data, compared to 84% on mortality data. Whilst a
basic point, the missing data on occupational class from either mortality or
census data (but especially the latter) opens the door to substantial potential
bias in measuring differences in mortality by occupational class. The higher 84%
availability of class data on mortality data does not remove the potential for
bias in unlinked analyses, as numerator-denominator biases are likely. For
example, our cross-classification of class by census and mortality data
demonstrated that relatively fewer deaths were identified as class 4 on census
data (census to mortality ratio of 0.45) compared to other classes (ratios 0.55
to 0.64).
On the other hand, our linked analyses that used current
occupational class according to census data (i.e. occupational class only
available for 47% of those census respondents that were linked to a mortality
record) were also prone to bias. First, among 25–44 year olds, the
class-mortality association was underestimated due to linkage bias. This linkage
bias was easily overcome by adjustment.
Further, the recent development of weights on NZCMS data to
adjust for linkage bias 37 (not used in this
paper) means that linkage bias can be easily addressed in future analyses on the
linked data using occupational class or any other socioeconomic factor. More
problematic though, and more specific to occupational class, was the bias due to
excluding the economically inactive when using census occupational class.
If we are trying to estimate the
association of usual or last occupational class with mortality, our linked
analyses appear to be substantially underestimated. (Our results for the
association of current occupational class with mortality remain valid—but
most researchers would prefer the association of usual or last occupational
class to avoid health selection biases.25,27)
We used data from the HLFS on the last occupational class
within 5 years among those with no current occupation to adjust our unlinked
analyses. However, this method is far from perfect. First, the HLFS did not
provide us with an occupational class for those people with an occupation prior
to five years ago—but no occupation in the last 5 years. Second, the HLFS
is a survey with its own imprecision when it comes to point estimates in small
cells—a situation that arises in the adjustment methodology.
Nevertheless, it seems highly probable that the use of
current occupation on census data in the linked analysis has resulted in
differential health selection bias whereby unwell people from lower classes are
less likely to be able to stay at work compared to unwell people from higher
classes.25,26
At first inspection, it is reassuring that the unlinked
analyses roughly agree with the linked analyses adjust for both linkage bias and
bias due to excluding the economically inactive (Figure 1). However, a closer
inspection of Table 1 suggests that both analyses may be underestimating the
‘true’ association of occupational class with mortality.
Why? Because whether one starts from census or mortality
data, deaths from lower classes (e.g. 5 and 6) are more likely to have missing
occupation on the other data-set. The flip-side of this is that deaths with
missing occupation on one data-set are more likely to have a low occupational
class than a high occupational class on the other data-set. This may be the
result of greater rates of entry and exit into the employment among lower
socioeconomic groups (i.e. greater churning) or greater amounts of
misclassification biases. Regardless, this pattern points to a relative
under-enumeration of lower class deaths on both census and mortality data. If
this under-enumeration of deaths is not matched by similar under-enumeration of
all people (i.e. denominators), then the strength of association of occupational
class with mortality will be underestimated regardless of the method of
analysis.
From an international perspective, our findings may provide
some guidance to researchers undertaking similar analyses. Our findings also
provide some indirect support for the method of Kunst et
al10 to adjust for differential health
selection. From a New Zealand perspective, our findings suggest that linked
analyses (with no adjustment for differential health selection) are biased. But
perhaps more importantly, it is difficult to escape the conclusion that
differences in the recording of occupational class on census and mortality data
in New Zealand make measuring mortality differences by class thwart with
difficulty and potential bias.
If one assumes that biases for any particular method of
analysis are constant (or at least similar) over time, and one carefully adjusts
where possible for bias, using occupational class to monitor trends in
socioeconomic mortality gradients may be valid.
Author information:
Tony Blakely, Senior Research Fellow; Jackie Fawcett, Research Fellow,
Department of Public Health, Wellington School of Medicine and Health Sciences,
University of Otago, Wellington
Acknowledgments: The
NZCMS is conducted in collaboration with Statistics New Zealand, and the
assistance of many members of staff. The NZCMS is funded by the Health Research
Council of New Zealand, with co-funding from the Ministry of Health. June
Atkinson assisted with the preparation of tables in this paper. Peter Davis and
Neil Pearce provided comments on early drafts of this paper. Andrew Sporle
provided comments on final drafts of the paper. Alistair Woodward and Neil
Pearce supervised TB’s PhD where many of the analyses presented in this
paper were initiated.
Correspondence: Dr
Tony Blakely, Department of Public Health, Wellington School of Medicine and
Health Sciences, University of Otago, PO Box 7343, Wellington. Fax (04)
389-5319; email tblakely@wnmeds.ac.nz
Summary Statistics New
Zealand Security Statement: (The full security statement is in a
technical report to be published by the Wellington School of Medicine in
hardcopy and at http://www.wnmeds.ac.nz/nzcms-info.htm)
The (New Zealand Census Mortality Study) NZCMS is a study of
the relationship between socioeconomic factors and mortality in New Zealand,
based on the integration of anonymised population census data from Statistics
New Zealand and mortality data from the New Zealand Health Information Service.
The project was approved by Statistics New Zealand as a Data
Laboratory project under the Microdata Access Protocols in 1997. The datasets
created by the integration process are covered by the Statistics Act and can be
used for statistical purposes only. Only approved researchers who have signed
Statistics New Zealand's declaration of secrecy can access the integrated data
in the Data Laboratory.
For further information about confidentiality matters in
regard to this study please contact Statistics New Zealand.
References:
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| | Current
issue | Search journal |
Archived issues | Classifieds
| Hotline (free ads) Subscribe | Contribute | Advertise | Contact Us | Copyright | Other Journals |
|