Elsevier

Social Science & Medicine

Volume 69, Issue 8, October 2009, Pages 1252-1260
Social Science & Medicine

Do ethnic and socio-economic inequalities in mortality vary by region in New Zealand? An application of hierarchical Bayesian modelling

https://doi.org/10.1016/j.socscimed.2009.07.036Get rights and content

Abstract

We hypothesised that ethnic and socio-economic inequality in mortality might vary by region in New Zealand. Linked 2001–2004 census-mortality data were stratified by region (District Health Boards or DHBs), sex, age and ethnic groups, and income quintiles. To accommodate data sparseness, and to achieve accurate estimates of DHB-specific mortality rates and rate ratios by ethnicity and income, we used hierarchical Bayesian methods. To aid presentation of results, we used posterior mortality rates from the models to calculate directly standardised rates and rate ratios, with credible intervals. Māori-European/Other mortality rate ratios were often similar across DHBs, but Waitemata and Canterbury DHBs (both predominantly urban areas with low Māori population) had significantly lower rate ratios. In contrast, Bay of Plenty and Waikato DHBs (heterogeneous by both ethnicity and socio-economic position) had significantly higher rate ratios. There was little variation in mortality inequalities by income across DHBs. Examining the underlying rates for ethnic and income groups separately, there were significant variations across DHBs, but these were often correlated such that the ethnic or income rate ratio was similar across DHBs. The application of hierarchical Bayesian allowed more definitive conclusions than routine empirical methods when comparing small populations such as social groups across regions. The range of hierarchical Bayesian estimates of Māori mortality and Māori:European rate ratios across regions was considerably narrower than empirical standardisation estimates.

Introduction

Cross-national and time comparisons of the magnitude of health inequalities are of etiological and policy interest. For example, varying inequalities in mortality by socio-economic position across countries may point to different diets, tobacco consumption and health care access by social group (Huisman et al., 2005, Mackenbach et al., 2003, Mackenbach et al., 2008).

There is a large literature documenting inequalities in mortality by socio-economic position and race/ethnicity at the country-level, including for New Zealand (Blakely et al., 2006, Blakely et al., 2008, Howden-Chapman and School, 2000, Pearce et al., 2002, Tobias et al., in press).

There is also a literature describing mortality variations between regions within a single country, including New Zealand. For example, Pearce (2007) finds that variations in suicide between regions among men aged 14–44 increased during the period 1980–2001. Similarly, Pearce and Dorling (2006) and Pearce, Dorling, Wheeler, Barnett, and Rigby (2006) argue for increasing between-region variations in mortality and life expectancy between 1980 and 2001 using a relative index of inequality for age–sex standardised DHB mortality rates, and a slope index of inequality of male and female life expectancy, respectively.

Related approaches can be found in the small domain estimation literature (Congdon, 2003), where the focus is typically on combining information from small areas to make population inferences for the total domain. For example, Leyland, Langford, Rasbash, and Goldstein (2000) use spatial multi-level models to predict mortality caused by neoplasms and circulatory disease in 143 postcode sectors in the Greater Glasgow Health Board, Scotland. In another example, estimates of all-cancer mortality rates for white males in 798 US health service areas are provided by Nandram, Sedransk, and Pickle (1999) using a hierarchical Bayesian model.

However, there is little published work comparing social inequalities in mortality between regions of a given country. One exception is a comparison of empirical age adjusted social class mortality rates (1991–1993) by region within the UK (Uren, Fitzpatrick, Reid, & Goldblatt, 2001). Marked regional variations in mortality for men aged 20–64 within social classes were found across the UK, with rates generally higher in all social classes in Scotland, Northern Ireland and the north of England. Relative inequalities in male mortality between the highest and lowest social class were greatest in Northern Ireland and North East England (rate ratios of 5.2 and 4.3, respectively – actual ratios not given in source reference, but read from figures), and least in East, South East, South West and London regions of England (rate ratios approximately 2.2–2.4). Social class differences in adult self-reported health, based on empirical estimates of age-standardised rates of poor health, have also been found to vary by region in the UK (25–64 year olds, 2001 census data), but now London was one of the two highest relative inequality regions along with Scotland (Doran, Drever, & Whitehead, 2004). Both studies focus on regional variations in socio-economic inequalities, and use empirical methods to estimate standardised rates and rate ratios. This leaves open the question as to whether other dimensions of spatial inequality are important, and whether routine methods are adequate to investigate them.

A priori, there were grounds for expecting some regional variations in inequalities in mortality between Māori and European/Other. Colonisation of New Zealand began nearly 200 years ago. There were marked variations across New Zealand in the extent and timing of European–Māori contact (Belich, 1996, King, 2004). Further, there is a strong tribal (iwi) and regional basis to Māori culture – although many Māori recognise ancestral ties to several iwi, and/or often live outside the traditional boundaries of those iwi. There was some regional variation in European settler patterns with respect to country of origin, timing and main economic activity. More recent migration and immigration patterns are likely to significantly impact these historical processes. Regarding socio-economic inequalities in mortality, we did not expect much regional variation. Local governments have some autonomy of local amenities, but the main policy decisions are made centrally and implemented equivalently across New Zealand. For example, one social welfare and taxation system applies to all regions of New Zealand.

The linked census-mortality data in the New Zealand Census-Mortality Study (NZCMS: Blakely et al., 2000, Fawcett et al., 2008) provide much useful information for quantifying such inequalities, and allow estimates of geographical variations in mortality and mortality inequalities using a more richly stratified dataset than have been possible in previous analyses. Even so, low person-years at risk and (particularly) deaths in some age, sex, ethnicity, income, and DHB strata are problematic for comparisons of mortality inequalities across regions. Hierarchical Bayesian models provide an approach to pooling information without forcing associations (e.g., between mortality rates and income) to be the same across strata. Thus, they give some protection against model misspecification while nevertheless permitting smoothing of crude rates. Hierarchical models have been shown to outperform classical regression in predictive accuracy (Gelman, 2006). Bayesian “shrinkage” estimators (discussed below) have good variance reduction properties, particularly when sample size is small (Best et al., 2005, Greenland, 2008). The utility of hierarchical Bayesian regression techniques for the analysis of social variations in health outcomes appears not to have been widely recognised.

Why might regional variations in social inequalities in mortality be important? First, the determinants of health inequalities are not necessarily the same as the determinants of health (Blakely, 2008, CSDH, 2008, Krieger, 2008). For example, policy variations by region might widen or narrow health inequalities compared to other regions in the same country. Second, knowing about regional variations in health and health inequalities should assist with targeting future policies and programmes, as is already the case with knowledge of average health status by region. Third, some nations, including New Zealand, fund their public health systems at the regional level using socio-demographic predictors of public health and health care need, in addition to (or instead of) actual measures of health status. Therefore, it is of interest to regional health authorities whether regional variations in health (and inequalities in health) are accounted for by the variables used in such funding formulae. Finally, identification of regional variability may provide important clues about modifiable drivers of health inequalities.

Accordingly, the major goal of this paper is to determine whether mortality inequalities by ethnic and socio-economic position vary across regions in New Zealand. As the regional variable we use 21 District Health Boards (DHBs), described later. However, comparing mortality rate differences between social groups within regions of a small country like New Zealand leads to problems of sparse data. For example, the number of Māori (indigenous population of New Zealand) deaths in some DHBs is small. The problem of low stratum person-time at risk becomes worse if the analyses also have to account for sex, age and one (or more) measures of socio-economic position. Consequently, a secondary goal of this paper is to demonstrate the use of hierarchical Bayesian methods that accommodate issues of sparse data, and provide reliable estimates of uncertainties, better than the ‘routine’ quantitative methods used in previous studies. To emphasise this point, we also examine the difference that hierarchical Bayesian methods make to final results. Third, to understand social inequalities in mortality, it is important to first understand regional variations in group-specific mortality rates, and the contribution socio-demographic factors make to those variations. Thus, the three objectives of this paper, and subsidiary research questions, are:

  • 1.

    Demonstrating social group differences in mortality across regions within New Zealand:

    • a.

      Do mortality inequalities by ethnicity vary across regions?

    • b.

      Do mortality inequalities by income vary across regions?

  • 2.

    Demonstrating the utility of hierarchical Bayesian methods:

    • a.

      Does the application of hierarchical Bayesian methods substantially alter findings, compared with simpler empirical methods?

  • 3.

    Exploring mortality rate differences across regions:

    • a.

      What is the variation in overall, Māori, European/Other, low- and high-income mortality rates by region?

    • b.

      How much of the sex and age adjusted mortality rate differences by region is explained by ethnicity and income?

We are not aware of previous empirical research in New Zealand on our main objective of demonstrating social group inequalities in mortality by region. Additionally, whilst hierarchical Bayesian models have, in our view, clear theoretical advantages in the face of sparse data, they come at considerable cost in terms of complexity, computing time and analytical expertise. Thus, we also compare hierarchical Bayesian results with routine direct standardisation results to determine what impact these more sophisticated analyses have on the final results. Regarding the third objective on variations in mortality rates between regions we present regional mortality rates adjusted for potential cofounders together with reliable estimates of uncertainty as a necessary first step, in our view, to interpreting social variations in mortality inequalities across regions.

Section snippets

Data

Linkage of 2001 census data to 2001–2004 mortality data in the NZCMS is described in detail elsewhere (Fawcett et al., 2008). Briefly, 81.5% of eligible mortality records (all ages) in the three years after the 2001 census were linked back to a 2001 census record (67,146 linked pairs). We estimated that over 97% of these linked pairs were correct linkages (Blakely & Salmond, 2002). Mortality records were less likely to be linked to a census record if any of the following conditions held: aged

Results

Table 1 provides a breakdown of actual person-years and deaths by DHB, and person-years by sex, income and ethnicity. A significant variation in population size across DHBs is evident in the 10-fold range in total person-years.

Posterior estimates of the prior model coefficients β0 (the intercept) and βeth (the ethnicity coefficient) from equation (3) suggested some variability by DHB (Fig. 2). A similar plot (not shown) for the income coefficient (βinc) showed much less evidence for

Main findings

Using a hierarchical Bayesian regression approach to model mortality rates in the 2001–2004 NZCMS cohort we found evidence for variability in ethnic disparities in mortality across regions within New Zealand. Ethnic mortality disparities, in rate ratio terms, were higher in Bay of Plenty and Waikato (both mixed metropolitan–rural areas with reasonably large Māori populations), and lower in Waitemata and Canterbury (both largely metropolitan populations). We did not find substantial evidence of

References (38)

  • T. Blakely

    Iconography and commission on the social determinants of health (and health inequity)

    British Medical Journal

    (2008)
  • T. Blakely et al.

    Income and mortality: the shape of the association and confounding New Zealand census-mortality study, 1981–1999

    International Journal of Epidemiology

    (2004)
  • T. Blakely et al.

    Probabilistic record linkage and a method to calculate the positive predictive value

    International Journal of Epidemiology

    (2002)
  • T. Blakely et al.

    Inequalities in mortality during and after restructuring of the New Zealand economy: repeated cohort studies

    British Medical Journal

    (2008)
  • C.L. Christiansen et al.

    Hierarchical Poisson regression modeling

    Journal of the American Statistical Association

    (1997)
  • P. Congdon

    Applied Bayesian modelling

    (2003)
  • CSDH

    Closing the gap in a generation: Health equity through action on the social determinants of health

    (2008)
  • M.J. Daniels

    A prior for the variance in hierarchical models

    La Revue Canadienne de Statistique

    (1999)
  • T. Doran et al.

    Is there a north–south divide in social class inequalities in health in Great Britain? Cross sectional study using data from the 2001 census

    British Medical Journal

    (2004)
  • Cited by (9)

    • Deprived yet healthy: Neighbourhood-level resilience in New Zealand

      2013, Social Science and Medicine
      Citation Excerpt :

      Geographical and social inequalities in health have been identified in many countries, including New Zealand (Pearce, Witten, & Bartie, 2006). Socioeconomically deprived places tend to have less favourable health outcomes and health-related behaviours (Chan et al., 2008; Haynes, Pearce, & Barnett, 2008; Richardson, Blakely, Young, Graham, & Tobias, 2009; Tobias & Cheung, 2003). Evidence also suggests that the gradient in health is growing (Pearce & Dorling, 2009; Wilkinson, 2005).

    • An ecological study of the relationship between socioeconomic isolation and mental health in the most deprived areas in Auckland, New Zealand

      2013, Health and Place
      Citation Excerpt :

      Thus, to compound the direct mental health impacts of living in deprived or risky neighbourhoods, there may also be psychosocial impacts (Yen and Kaplan, 1999) as groups feel deprived relative to other reference groups in society (Muntaner et al., 2004). Since patterns of neighbourhood deprivation are often geographically uneven (Richardson et al., 2009), there are instances where deprived areas are surrounded by relatively affluent areas. This could be termed ‘socioeconomic isolation’, and may be more common in urban areas (see Fig. 1 for an example).

    View all citing articles on Scopus

    The New Zealand Census-Mortality Study (NZCMS) was funded by the Health Research Council of New Zealand, and now receives ongoing funding from the Ministry of Health. The NZCMS is conducted in collaboration with Statistics New Zealand and within the confines of the Statistics Act 1975. June Atkinson provided assistance with data extraction. Helpful comments on drafts were received from Jamie Pearce. Access to the 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. The results presented in this study are the work of the authors, not Statistics New Zealand.

    View full text