METHODS

Record linkage of census and mortality records

This paper uses a cohort study of all New Zealanders formed by record linkage of 1991 census records to 1991–94 mortality records.^24,25 A total of 1 654 314 people aged 25–64 years completed the 1991 census. Of the 19 128 mortality records for people aged 25–64 years on census night and dying within three years after the 1991 census, 14 322 (74.9%) were linked back to a census record.

Individual level variables

Individual level variables from the 1991 census included sex, age, ethnicity (Maori (indigenous population), Pacific, non-Maori non-Pacific), equivalised household income and rurality. We calculated the equivalised household income by first assigning to each census income category the mean income for the equivalent category from the 1991 New Zealand Household Economic Survey, then totalled the income of all adults usually resident in the household. Equivalisation was achieved by dividing the total household income by the square root of the number of people in the household. For example, a total household income of $50 000 for four people became an equivalised household income of $25 000. (No household income could be calculated if a usually resident adult was not home on census night, or one or more adults did not specify their income).

Regional level variables

Three different sets of regions were analysed: 14 regional councils, 35 health funding authority sub-regions (hereafter called sub-regions), and 73 territorial authorities. Sub-regions are aggregations of territorial authorities that approximate hospital catchment areas. Sub-regions were used for most analyses as: they have been used previously (with slight modification) in an ecological study of the association of income inequality with mortality¹⁷; the variation in population size between territorial authorities was too large (range 760 to 315 668 persons (all ages), 25th centile of 14 433 and 75th centile of 51 607); and there were only 14 regional councils. The sub-regions ranged in population size from 13 280 (North Canterbury) to 315 668 (Central Auckland), with a median of 74 587, 25th centile of 41 958, and 75th centile of 131 537. For each of the regions, the average household income and Gini coefficient (measure of income inequality) were calculated using 1991 census household income data. At the sub-region level, the 25th, 50th, and 75th centiles of the Gini coefficient were 0.334, 0.338, and 0.351, and of the average household income were $20 019, $21 545, and $23 842.

Analyses

The analyses in this paper excluded deaths in the first six months to reduce the possibility of health selection effects and were restricted to those census records with complete data, resulting in a sample size of 1.4 million 25–64 year olds (688 857 men and 702 261women; 84.1% of all 25–64 year old census respondents). Most exclusions were attributable to missing household income data, and extensive sensitivity analyses published elsewhere suggested little resultant selection bias.²⁶ Analyses were not conducted for children due to lower statistical power and were not conducted for 65 year olds and over due to convergence of household incomes. Four causes of death were also examined (cancer, cardiovascular, unintentional injury, and suicide) on the basis that any mechanism of action for income inequality would probably vary by cause of death (for example, social cohesion and fragmentation might be more important for suicide). There is also the possibility of varying time lags by cause of death. All individual level variables were treated as categorical variables in the analyses, but the two regional variables were analysed both as quartile and continuous variables. For models including the Gini as a continuous variable, we report the odds ratio for a 0.01 increase in the Gini. Analyses were conducted in SAS v8.0 using logistic regression on site at Statistics New Zealand under strict privacy requirements (see SNZ Security Statement). Regression models used exact data, but all frequency output presented in tables in this paper are random rounded to a near multiple of three as per SNZ protocol. As the outcome of death among 25–64 year olds followed up for three years is relatively rare, the odds ratio is a good approximation of the risk ratio.

Three sensitivity analyses about the all cause mortality results were conducted. Firstly, the main model at the sub-regional level was run at the territorial authority and regional council levels. Secondly, we adjusted for variation in the record linkage success by the regional categories of income inequality (that is, adjustment for linkage bias). Log-link regression models were used including only the eligible mortality records, with: the dependent variable being whether the mortality record was linked to a census record; and independent variables being age, ethnicity, rurality, socioeconomic deprivation (measured by a census derived index for small areas of about 100 people²⁷), regional average income (quartiles) and regional income inequality (quartiles). The relative risks of linkage from these models were then used to adjust the observed odds ratios in the cohort analyses. For example, if the relative risk of linkage for deaths in the highest quartile of income inequality compared with the lowest quartile was 0.95 (that is, 5% less likely to be linked), and the observed odds ratio of death in the cohort analysis was 1.05 (that is, a 5% observed excess mortality risk for people living in high income inequality areas), then the adjusted odds ratio was 1.05/0.95=1.11. The third sensitivity analysis was to rerun the main logistic regression models with random intercepts for each region, using the SAS supplied macro Glimmix in SAS v8.0. (We choose not to routinely use random effects models for our analyses for two reasons. Firstly, random effects models arose from clustered survey sampling work (for example, children within classrooms within schools within areas—our analyses are for a complete cohort. Secondly, while there was probably some residual intraclass correlation on unmeasured variables within regions following control for covariates, the same could be said of all the independent variables. For example, even controlling for age and deprivation, there will be residual correlation of unmeasured factors (for example, income) with ethnicity. Thus, it seems inappropriate to just specify random intercepts for the regional variable.)

RESULTS

Table 1 shows the distribution of census respondents by the variables of interest.

All cause mortality

Controlling for age, ethnicity, rurality, and equivalised household income, the quartile odds ratios for income inequality and average income showed no association with all cause mortality (table 1). Modelling the Gini and average income as continuous variables (but with no other change to the model) the odds ratio of mortality for a 0.01 increase in the Gini was 1.007 (95% confidence intervals 0.989 to 1.024) for men and 1.004 (0.983 to 1.026) for women.

While we found no overall association of income inequality with all cause mortality, it was possible that the association of household income with mortality might be stronger in higher income inequality sub-regions (that is, cross level effect modification). We divided the cohort into those living in high (top two quartiles) and low (bottom two quartiles) income inequality sub-regions. There was no difference in the association of household income with all cause mortality between these two sub-populations (results available from authors).

Figure 1 shows the odds ratio of all cause mortality for an increase in the Gini of 0.01 for models that include the Gini as a continuous variable. The first model includes just age and rurality as covariates, and sequential models introduce sub-region average income (also as a continuous variable), household income, and ethnicity. For both sexes there is an approximate 1% increase in mortality risk for a 0.01 increase in the Gini when controlling for only age and rurality (model 1 in fig 1). Additional control for average sub-region income (model 2) results in odds ratios of 1.021 (95% confidence intervals 1.004 to 1.039) and 1.027 (1.006 to 1.049) for a 0.01 increase in the Gini for men and women, respectively. This excess risk changes little after control for household income, but essentially disappears after control for ethnicity at the individual level with odds ratios becoming 1.007 (0.989 to 1.024) and 1.004 (0.983 to 1.026), respectively.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Odds ratio of all cause mortality (95% confidence intervals) for an 0.01 increase in the Gini for sequential models at the sub-regional level.

Sensitivity analyses for all cause mortality

Using the model specification shown in table 1 there was little evidence of an association of income inequality with all cause mortality for both sexes at either the territorial authority or regional council levels (table 2). (Because of slight variation in geographical boundaries, 30 417 census respondents that had missing values for sub-region actually had a value for territorial authority and regional council.)

The two models shown in table 1 were repeated using random intercepts for each of the sub-regions. The quartile odds ratios were relatively unchanged and the confidence intervals increased in size by 50% to 60% (results available from authors).

Regarding possible linkage bias, female (but not male) deaths in higher income inequality sub-regions tended to be slightly less likely to be linked to a census record than deaths in the lowest quartile of income inequality (results available from authors). Adjusting for this modest linkage bias (as shown in the final column in table 1) did not substantially change the findings. Likewise, there was no substantial impact of adjusting regional council and territorial authority analyses for linkage bias (final column for each sex in table 2).

Cause specific mortality

Table 3 shows the income inequality odds ratios for cancer, cardiovascular disease, unintentional injury, and suicide deaths. The findings were null for all male causes of death and null for female cardiovascular deaths. Regarding women, the quartile odds ratio suggested some positive association for injury and suicide deaths and (possibly) some negative association for cancer deaths. However, these latter findings were based on comparatively few deaths and much of the apparent association might have been attributable to comparatively low mortality in the reference category. Notably, the 95% confidence intervals for all the continuous Gini odds ratios in table 3 included 1.0. Adjusting the cause specific results for linkage bias had no notable impact on the results other than perhaps strengthening the suggestion of a null finding for female cancer deaths (table 3).

DISCUSSION

We found no convincing association of income inequality with all cause mortality within New Zealand. A previous New Zealand ecological study of income inequality and all cause mortality by the same sub-region regions used in our multi-level study found some association in the expected direction.¹⁷ Based on the results presented in figure 1 it seems this previous ecological finding was attributable to confounding by the varying ethnic group composition of regions in New Zealand. Maori and Pacific people tend to have lower incomes than non-Maori²⁸ and higher mortality rates,^29,31 meaning that regions with high proportions of Maori and Pacific people will have higher mortality rates. In parallel, as Maori and Pacific people are minority ethnic groups overall in New Zealand, regions with higher proportions of Maori and Pacific people will tend to have a more dispersed range of incomes resulting in higher income inequality. Consequently, ecological analyses that do not adjust for ethnicity will observe a spurious association of income inequality and mortality due to (uncontrolled) confounding for ethnicity at the individual level. Indeed, repeating the previously published age and sex standardised ecological analyses,¹⁷ but additionally standardised for ethnicity, removes the association of regional income inequality with mortality consistent with the analyses in this paper (results available on request from O’Dea).

Could controlling for ethnicity at the individual level be a form of over-control? For this to be the case either people modify their self identified ethnicity consequent on regional income inequality, or Maori (non-Maori) tend to migrate to (out of) high income inequality regions. We believe these two possibilities are implausible. Instead, confounding by ethnicity at the individual level of the crude association of income inequality with mortality is the most probable explanation. Our study reinforces the importance of ethnic inequalities in health in New Zealand. In this context, these ethnic inequalities also give rise to important regional inequalities in health. There are good reasons why we might not expect an association of income inequality with adult mortality within New Zealand. Firstly, New Zealand has comparatively little regional variation in education, health, welfare and other systems that might plausibly explain any neo-material basis to the association of income inequality with health.^21,32 Secondly, and related to the first reason, there is probably less variation in income inequality by region than in other countries because of the same taxation and government systems applying throughout in New Zealand. While it is difficult to make comparisons of income inequality between countries it seems likely that both the range of income inequality (the exposure) within New Zealand is narrower than for US counties and metropolitan areas and the level of regional income inequality is less.¹⁴ These two reasons may also be possible explanations for the lack of association of regional income inequality with mortality in Canada.¹⁸

It is possible that our null study may unduly dismiss the association within New Zealand. For cardiovascular disease and tobacco related cancers it would take time for any contextual effect of income inequality to influence risk behaviours and (patho)physiology.³³ Future analyses in New Zealand should examine the possibility of time lags by modelling income inequality measures some time before death. Also, given the large increase in income inequality in New Zealand in the late 1980s, following up mortality in the late 1990s and beyond based on early 1990s income inequality measures seems sensible. However, elsewhere we have found that investigating income inequality time lags is difficult because of relative stability of regional rankings by income inequality over time and individual mobility over time between regions. Of note, if psychosocial mechanisms underlie any association of income inequality with health, then one might still expect to see an association of income inequality with suicide (plus or minus injury) in our study—there was possible evidence of this for women, but it was not statistically significant.

Our study does not provide evidence against the possible importance of income inequality at the country level. For example, New Zealand has experienced a rapid increase in income and other inequalities in the past 20 years. It is, however, a moot point whether this has been detrimental to the overall health status of New Zealanders.

Our findings, and other recently published studies,^19,20,34 suggest that the study of the association of income inequality with health within other countries should be approached cautiously. Firstly, any apparent association from ecological analysis may be attributable to individual level confounding, and not necessarily just by individual level income as might be expected mathematically.³⁵ In our study ethnicity was confounding the association of income inequality with mortality. Secondly, there are many limitations to using within country analyses to test the income inequality hypothesis including unmeasured confounding by region, inadequate allowance for time lags, and inadequate variation of income inequality by regions within countries.^33,36 Regarding the association of income inequality with health at the state level in the US, while it often remains after controlling for individual level confounding,^9,11,33 it does not seem to be generalisable to regional analyses within other countries. This lack of generalisability may either be because the US association is spurious because of ecological level confounding,³⁷ or because the US has both large enough within country variation of income inequality and high enough levels of income inequality for the association to be observable.¹⁸