  |
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Journal of the New Zealand Medical Association,
11-December-2009, Vol 122 No 1307
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Does mortality vary between Pacific groups in
New Zealand? Estimating Samoan, Cook Island Māori, Tongan, and Niuean
mortality rates using hierarchical Bayesian modelling
Tony Blakely, Ken Richardson, Jim Young, Paul Callister,
Robert Didham
Pacific people have mortality rates intermediary between
Europeans and Māori when considered as a single
group.1 But the Pacific population is
heterogeneous, and health status (including mortality rates) is likely to vary
between specific Pacific groups (e.g. Samoan, Tongan, Cook Island Māori,
and Niuean).
First, there are demographic and migration variations
between the Pacific ethnic groups that might generate variations in health
status. For example, migration histories to New Zealand vary. Cook Islanders and
Niueans have been New Zealand citizens since 1901 and Tokelauans since 1916.
Migration of these ethnic groups to New Zealand thus has a longer history and
may have been less health-selective than other Pacific ethnic groups. With the
increasing need for unskilled labour in the 1960s and 1970s, immigration from
the Pacific grew. A Treaty of Friendship was signed with the Samoan Government
in 1962, and the Western Samoan Quota scheme was established to facilitate
migration from Samoa.
In 1945, the Pacific population in New Zealand was just over
2000 people, Samoans being the largest group. In recent times, the Pacific
population was 202,233 in 1996, rising to 231,801 in 2001, and increasing
further to 265,974 in March 2006.2 The Cook
Island Māori and Niuean population in New Zealand are now substantially
higher than their respective ‘home’ population (4 and 14 times
greater, respectively, at the 2006 Census) and more than 70% are New
Zealand-born. In contrast, the ‘home’ Samoan population is larger
than the New Zealand population, and 60% of those residing in New Zealand are
born New Zealand. The highest overseas-born proportion is seen amongst Tongans.
Table 1 shows a range of sociodemographic characteristics
for the four largest Pacific ethnic groups, Samoan, Cook Island Māori,
Tongan, and Niuean. Cook Island Māori and Niuean are least likely to
self-identify with a single ethnic group. Cook Island Māori have the lowest
rate of formal qualifications, and Tongan the lowest median income.
Second, health-related risk factors vary between Pacific
ethnic groups. At the 2006 Census, 38% of the Cook Island Māori population
smoked, compared to 28% and 29% for Samoan and Tongan (Table 1). Sundborn et al
(2008) have recently published estimates of a range of risk factors for Samoan,
Tongan, Niuean, and Cook Island Māori participating in the Diabetes Heart
and Health Study.3
Being a workforce survey, the population was considerably
older than the census population. The smoking prevalence also differed from
census data, raising concerns about representativeness of the Pacific
population. Nevertheless, the study found substantial variation in health risk
factors between different Pacific groups.
The aim of this paper, therefore, is to present mortality
rates for the four largest Pacific ethnic groups living in New Zealand, using
linked census-mortality data for 2001–04, thereby bypassing the problems
of non-comparable collection of data for Pacific ethnic groups between census
and mortality data (i.e. numerator-denominator bias). However, mortality is an
uncommon outcome, and Pacific populations are small and young. Thus, given the
small number of deaths in some Pacific strata, we have used hierarchical
Bayesian (HB) regression modelling that allows smoothing of posterior mortality
rates by shrinkage towards a prior covariate
structure.4
Table
1. Demographic data for Samoan, Cook Island Māori,
Tongan, and Niuean living in New Zealand at the 2006 census (and 2001 census for
selected variables)
(a) % of adults in fully or partly owned home
† ‘Sole any-Pacific’ mean the person
did not self-identify as any non-Pacific groups, but might have self identified
as two or more specific Pacific groups (e.g. Samoan and Niuean).
‡ ‘Sole specific-Pacific’ means the
person only identified one Pacific ethnic group—and nothing else (e.g.
Samoan only)
# Only 2006 reported due to either missing data for
2001 or little change from 2001.
Source of much of the 2006 data, Pacific Profiles:
2006, http://www.stats.govt.nz/analytical-reports/pacific-profiles-2006/default.htm
MethodsData—Linked census-mortality
data from the 2001–04 cohort of the New Zealand Census-Mortality Study
(NZCMS) were used; details of the linkage, weighting for incomplete linkage of
mortality data to census data, and variables specifications can be found
elsewhere.5 Data was restricted to 380,000
person-years for 0–74 year olds with non-missing data on ethnicity, sex,
age, and equivalised household income (62% of all eligible person years).
Analyses were conducted on aggregated data for the 240 strata formed by
cross-classifying ethnicity (4 groups) by sex (dichotomous) by age (5 groups) by
income (tertiles) by natality (New Zealand-born versus overseas-born).
The ethnicity variable was classified using a
“total count” definition.6,p13 For
example, all of the following people would be categorised as Niuean:
self-identified Niuean only; self-identified Niuean and Samoan; self-identified
Niuean and New Zealand European; and self identified Niuean and Māori.
Based on these examples, using the total count method, the Niuean and Samoan
person would be counted in both Niuean and Samoan groups in this paper. (The
contribution of Pacific self-identification to health status of European and
Māori responses are not directly addressed in the paper.)
For the HB modelling, five age-groups were used:
0–14 years, 15–34 years, 35–44 years, 45–64 years, and
65–74 years, centred at the 35–44 age group, and scaled so that each
unit increase in scaled age corresponded to an actual increase of 10 years.
Thus, the above age ranges are represented by their end-points which, after
centring and scaling become elements from the set {-3, -1, 0, 2, 3}. To allow
for the non-linear increase in mortality with age, a linear spline for age with
knots at the 35–44 and 45–64 age groups was included in the prior
mean [equation (3), below].
We used equivalised household income, categorised into
tertiles within strata of sex and age-group, and based on income for the New
Zealand Pacific population. Income ranks were median centred and scaled (divided
by 10). Thus, income ranks (coded as 1, 2, 3) are transformed to (-0.1, 0, 0.1)
and the income effect was assumed to be log linear.
Hierarchical
Bayesian Poisson regression—We largely follow the HB methods used
previously in the NZCMS by Young et al (2006).4
In brief, the method was as follows. Assuming death is a Poisson process such
that for Pacific ethnicity
 and
stratum 
with deaths dij, mortality rate
λij, and person-years at risk
Pij, and using the notation
 to represent a random
variable x distributed as D with mean a and variance
b, a three-level Poisson model was defined by:The mortality rate
λij had a gamma distribution with
mean µij and variance
µij/ζ2,
and the prior mean µij had a
structure that depended on covariates Xi
and parameters βj through a
log-link function. Second-level parameters,
βj (the regression
“hyper-parameters”) and ζ (the mortality rate variance or
“shape” hyper-parameter) were assigned independent prior
distributions (“hyper-priors”) at the third level of the hierarchy.
Extending the Young et al (2006) model to
allow for variation by Pacific ethnicity, the regression hyper-parameter vector
was partitioned as
 to
allow the intercepts (β0j)
to vary by Pacific ethnicity. A standard approach was adopted for
βj, with uniform prior
distributions for each component.The prior covariate structure influences the mean of
the posterior rate, but the degree of influence depends on the overall support
for the prior covariate structure in the data, as well as on how much local
information is available. Given the structure of the model defined by equations
(1) and (2), the conditional posterior distribution for the mortality rate is
also gamma with mean
where
 is the observed mortality rate in the ith stratum of
the jth ethnicity,  and
Thus, the conditional mean for
λij is a weighted average of the
prior mean µij and the observed
mortality rate (yij). The
Bij, which lie between zero and one, are
known as shrinkages because larger values shrink the conditional posterior mean
mortality rates towards the prior mean. The gamma shape parameter
ζ provides a measure of the
influence of the prior mean. The relatively uninformative uniform shrinkage
prior of Christiansen and Morris (1997) was adopted for
ζ.7,8
A priori, following previous NZCMS
work1 4, we expected interaction of age and
income, and sex and income as predictors of the mortality rate. Thus the
components of the regression hyper-parameters in equation (3) for the most
complex prior model were
To allow comparison across the four ethnic groups,
posterior mortality rates were directly standardised to the total Pacific
population distribution using three combinations of sex, age, income and
natality (country of birth; CoB). The first model (model 1) computed posterior
mortality rates for each age, sex, ethnicity stratum using underlying data
stratified by the same variables. Standardisation was by sex and age. Model 2
computed posterior rates for each age, sex, natality, and ethnicity stratum,
using underlying data stratified by the same variables, and standardised by sex,
age, and natality. Model 3 (the “full” model) added income as
well.
All analyses and plots were done using the R
environment (http://www.r-project.org)
for statistical computation v2.2.0 available from the Comprehensive R archive
Network (CRAN) website (http://cran.r-project.org) or SAS v8.2 (SAS
Institute Inc., Cary, North Carolina). All HB analyses used WinBugs 1.4,
available from (http://www.mrc-bsu.cam.ac.uk/bugs/winbugs/contents.shtml),
and the R2WinBUGS package version 2.0-4.
Sensitivity
analyses—We examined the likely impact of two possible biases:
return migration and census under-enumeration. Methods were straightforward and
are best described along with the results are presented below.
ResultsPosterior mortality rates standardised for sex and age
(model 1), plus natality (model 2) plus income (model 3), are shown in Table 2.
The relative positions of the four ethnic groups did not alter greatly with this
sequential posterior standardisation process. For the final fully standardised
model, all-cause mortality rate ratios compared to Samoan were: 1.21 (95%
credible interval 1.05 to 1.42) for Cook Island Māori; 0.93 (0.77 to 1.10)
for Tongan; and 1.07 (0.88 to 1.29) for Niuean.
Table 2. Posterior
all-cause mortality rates (per 100,000) and rate ratios (95% credibility
intervals) by Pacific groups from models extended to include natality and
household income
Model 1 = data stratified by
ethnic group, sex and age only; sex and age included as independent variables in
prior model; posterior rates directly standardised to the NZ Pacific population
using a total definition of ethnicity; Model 2 = model 1, but extended to be
additionally stratified by natality; additionally including natality as
independent variable; posterior rates standardised as for model 1; Model 3 =
model 2, but extended similarly for household income.
Figure
1. Posterior all-cause and cause-specific mortality rates
(per 100,000; 95% credible intervals) by Pacific groups for the fully adjusted
model (model 3)
 Table 3 and Figure 1 show the
all-cause and cause-specific mortality rates and rate ratios for the fully
adjusted model. Cardiovascular disease (CVD) mortality rate ratios, compared to
Samoan, were: 1.66 (1.26 to 2.13) for Cook Island Māori; 1.11 (0.72 to
1.58) for Niuean; and 0.86 (0.58 to 1.20) for Tongan. That is, CVD mortality was
nearly twice as high among Cook Island Māori compared to Tongan.
Moderating these strong CVD differences were weaker (and
perhaps opposing) non-statistically significant differences in cancer. For
injury/suicide mortality combined, Niuean people had a rate ratio of 0.53 (0.17
to 0.99) compared to Samoan people. Injury/suicide mortality rates for Samoan,
Cook Island Māori, and Tongan groups were similar.
Checks of posterior predictive distributions of mortality
rates against empirical estimates produced no evidence of model
lack-of-fit.
Sensitivity analyses: return migration and census
under-enumeration—The results of basic sensitivity analyses about
the sex and age-adjusted (only) HB mortality rates for 0–74 year olds are
shown in Table 4. The sensitivity analyses are crude.
For example, input parameters (e.g. percentage returning to Pacific country) are
applied to overall rates, not by strata of sex, age and so on. We consider
return migration and census under-enumeration in this paper, but other possible
systematic biases are considered
elsewhere.9
Return migration might occur when people are unwell, and
they decide to return to their home country (country of birth, one would assume)
to die. We set 8% as the best estimate of underestimation of Pacific deaths due
to return migration, and 4% and 12% as low and high scenarios (see elsewhere for
justification9). When such bias is the same for
all four Pacific ethnic groups, the rates vary but rate ratios do not (Table 5),
meaning no bias in the rate ratios estimated in this project. However, would
return migration when terminally ill be the same for the four Pacific groups in
this project? We view return migration as more likely among Tongan and Samoan
people, because: there are still substantial Samoan and Tongan populations (and
health care facilities) in the home Island to act as ‘pull factors’
for return migration; and lower proportions of Samoan and Tongan people are born
in New Zealand, probably predicting greater return migration for these groups
when terminally ill.
Scenario A (10% of terminally ill Samoan and Tongan
people returning home to die, and 5% of Niuean and Cook Island Māori
returning home to die when terminally ill) is our best estimate of such
differential bias. For all cause-mortality, this would reduce the Cook Island:
Samoan excess rate ratio by about a quarter from 1.21 to 1.15, but would only
reduce the CVD rate ratio by about 13% from 1.63 to 1.55. The two remaining
differential return migration scenarios (B = 12% Samoan and 3% Cook Island
Māori; C = 25% Samoan; 0% Cook Island Māori) are both in our view
unlikely, if not implausible. Thus we conclude differential return migrant bias
is an unlikely cause of differences between Pacific-specific ethnic groups in
CVD mortality.
Census under-enumeration of Pacific people is likely, and we
selected a 4% undercount as an overall estimate, and 2% and 6% as low and high
scenarios about this estimated census undercount (see elsewhere for
justification9). With respect to the
calculation of mortality rates, these percentages cause an overestimate of the
mortality rate.
Pacific mortality rates adjusted for such overestimation,
when nondifferential across the four Pacific groups, have no impact on the rate
ratios compared to Samoan (Table 5)—just the rates themselves. Is it
plausible that census undercounting might vary between the four Pacific groups,
and hence overestimation of mortality rates might vary by Pacific group? As
Niuean and Cook Island Māori have automatic New Zealand citizenship, there
might be less willingness among Samoan and Tongan population to be enumerated
due to the history of action against ‘over-stayers’.
We posited a 2% census undercount for Niuean and Cook Island
Māori, and 6% for Samoan and Tongan people. However, this only serves to
widen the mortality gap between Cook Island Māori and Samoan people (Table
4). Therefore, we conclude that differential census under enumeration across the
four Pacific groups in our study is very unlikely (if not impossible) to cause
the observed differences between mortality rates for Pacific-specific ethnic
groups.
DiscussionTo our knowledge, this project is the first in New Zealand
to show clear (and marked in the case of CVD) differences in mortality have been
demonstrated between Pacific ethnic groups. This finding is important for
further health research and policy. At present, policy approaches to Pacific
health disparities often assume that they are the same across ethnic subgroups.
Our results demonstrate that this is not the case. There is
parallel emerging evidence from the New Zealand Ministry of Health and
University of Auckland on differences between Pacific groups in youth health
status (work in progress) and some indicators relating to adult health status
(Health and Disability Intelligence, Ministry of Health, work in progress), as
well as published evidence of diversity among Pacific ethnic groups for mental
health.10
We have also demonstrated the added value of using
hierarchical Bayesian methods for sparse data problems. We anticipate that the
use of such methods will accelerate in the future both within New Zealand and
internationally, for a range of research questions including the monitoring and
understanding of the health of ethnic minorities.
We present quantitative sensitivity analyses in this paper
about return migration to the Pacific when terminally ill, as well as census
under enumeration of Pacific people, and conclude that neither bias could
plausibly give rise to the observed variations in mortality—especially the
elevated CVD mortality among Cook Island Māori.
Migration to New Zealand due to ill health should also not
alter the findings in this paper, as we had data on usual residency on the 2001
census cohort and years in New Zealand on the mortality data that (assuming
reasonable data accuracy) allowed us to exclude non-residents and ineligible
deaths. We have also conducted basic comparative analyses on the 1996-99 NZCMS
cohort9; Cook Island Māori also had an
elevated CVD mortality rate in this period.
Explanations for the elevated CVD mortality risk of Cook
Island Māori relative to the other Pacific ethnic groups include: (1)
lesser degrees or greater waning of health-selective migration, and (2) greater
degree of uptake of adverse risk factors (as evidenced, for example, by their
higher smoking prevalence). Against these hypotheses is that inclusion of
natality in our models had little effect.
In summary, Cook Island Māori have notably elevated CVD
mortality compared to Samoan, Tongan and Niuean ethnic groups, and Niueans have
lower injury and suicide mortality. No obvious explanation for these differing
mortality risks is evident from our analysis. Researchers should, if at all
possible, try to analyse and interpret results separately for specific ethnic
groups within the Pacific grouping. This additional analysis should not only aim
to explain the causes and contexts of these differences but also should
assist in the development of policies that overcome these within-Pacific group
disparities.
Policymakers and their advisors need to be aware of
differences in health status and risks between Pacific ethnic groups, and
consider Pacific ethnic group-specific policies and programmes where relevant,
alongside pan-Pacific approaches.
Author
information: Tony Blakely1, Research
Professor; Ken Richardson1, Senior Research
Fellow; Jim Young2, Honorary Senior Research
Fellow; Paul Callister3, Associate Professor;
Robert Didham4, Demographer
Acknowledgements: We
acknowledge the expert advice of Patrick Graham and June Atkinson on statistical
and data issues, respectively. Rachel Foster provided assistance with final
editing of the paper. This study was funded as part of the Official Statistics
Research programme. The New Zealand Census-Mortality Study (NZCMS) was initially
funded by the Health Research Council of New Zealand, and receives ongoing
funding from the Ministry of Health. Access to the NZCMS data used in this study
was provided by Statistics New Zealand under conditions designed to give effect
to the security and confidentiality provisions of the Statistics Act 1975.
Correspondence:
Tony Blakely, Department of Public Health, Wellington School of
Medicine and Health Sciences, University of Otago, PO Box 7343, Wellington,
New Zealand. Fax: +64 (0)4 3895319; email: tony.blakely@otago.ac.nz
References:
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| | Current
issue | Search journal |
Archived issues | Classifieds
| Hotline (free ads) Subscribe | Contribute | Advertise | Contact Us | Copyright | Other Journals |
|