Table 1 compares the rates of access to HD for Maori and NZ European offenders for each of the two outcome measures (leave to apply for HD, being granted HD). Examination of the table suggests prima facie evidence of a bias against Maori: whereas 39% of Europeans were granted leave to apply for HD at the point on sentencing, only 29.1% of Maori were granted leave to apply (RR = 1.34; 95%CI 1.25-1.43). Similarly, 19.3% of Europeans were granted HD at a Parole Board hearing compared to only 10.7% of Maori (RR = 1.80; 95%CI 1.60-2.02). In both cases these differences were highly significant (p<.0001).
One explanation of these differences is that there may be third or confounding factors that could explain the apparent differences in access to HD. In particular there may be differences in the offending history or personal characteristics of Maori and European offenders that may be related to the likelihood of access to HD, and that could explain some component of the ethnic differences in access to HD. To examine this possibility the data were reanalyzed using logistic regression methods in which the association between offender ethnicity and each outcome (leave to apply for HD, granted leave for HD) was adjusted for individual characteristics, and measures of offence seriousness, offending history and related factors from the database (see Methods). As noted in Methods, to simplify the presentation of results the analysis was conducted using a data reduction approach in which a number of covariate factors were combined using a composite regression score approach. In addition, the analysis of granting HD at the Parole hearing has been based on the full sample to avoid issues of selection bias resulting from the processes relating to leave to apply for HD at the sentencing stage.
The results of this analysis are summarized in Tables 2a and 2b which show the fitted logistic regression models for leave to apply and granting HD respectively. For each outcome the tables show the fitted regression coefficients, the corresponding standard errors and tests of significance for each factor in the regression model. For comparative purposes the corresponding coefficients from the simple unadjusted model that included only ethnicity as a predictor are also shown. Examination of the Table shows:
For the outcome leave to apply for HD at sentencing, adjustment for the correlated effects of personal characteristics, measures of offending and other covariates factors explained a substantial component of the association between ethnicity and access to HD. Prior to adjustment the regression coefficient for ethnicity was B = .442 (SE = .052; p<.0001); after adjustment this was reduced to B = .192 (SE = .060; p = .001). A range of covariate factors was significant in the adjustment model including gender, age at sentencing, the risk of reincarceration index, youth offending (first sentence <20 years) and the composite score measures of offence severity and offending history.
For the outcome granting of leave by the Parole Board, adjustment for covariates also reduced the association substantially. Prior to adjustment the regression coefficient for ethnicity was B = .688 (SE = .070; p<.0001); after adjustment this was reduced to B = .348 (SE = .100; p = .0005). For this outcome significant covariate factors included: gender, age at sentencing, the risk of reincarceration index, and the composite score measures of offence severity and offending history.
For both outcomes the adjusted association between ethnicity and access to HD remained statistically significant after adjustment. However, with such a large sample even very small differences may be statistically significant, and it is probably better to examine adjusted measures of effect size, rather than focus on statistical significance per se.
To illustrate the extent of any residual bias after control for confounding factors Table 3 shows the adjusted rates of access to HD for each outcome calculated from the fitted regression models. As described in Methods, the adjusted rates have the interpretation of the hypothetical rates of access to HD that would have been observed had Maori and European offenders had an identical distribution of personal characteristics, offending history and other covariate factors. The Table also shows the adjusted relative risk (RR) of access to HD for European compared to Maori offenders derived from the adjusted rates.
In confirmation of the regression results in Table 2, The Table clearly shows that adjustment for confounding explained a substantial component of the differences in access to HD between Maori and European offenders. Prior to adjustment, the gap between Maori and European offenders in the rate of leave to apply for HD at sentencing was in the region of 10%. After adjustment this gap had reduced to 3.6%. For the granting of leave at the parole hearing the initial gap in access between Maori and European offenders was 8.6%; after adjustment this had reduced to 2.4%. Similar trends are apparent from a comparison of the unadjusted and adjusted RRs for each outcome.
These results suggest that, for this sample, for Maori to have had equivalent rates of access to HD to European offenders, there should have been approximately 145 more Maori granted leave to apply for HD at sentencing (3.6% of 4023 Maori), and approximately 96 more Maori granted HD by the Parole Board (2.4% of 4023 Maori).
The above analysis was based on a simple main effects model in which it was assumed that the impact of the covariate factors was on HD outcomes was the same for Maori and European offenders. However, it could be suggested that different processes may operate for different ethnic groups, so that the factors that predict access to HD for Maori may be different from those which operate for European offenders. This in turn could make it difficult to draw firm conclusions about the extent of ethnic biases in access to HD since different processes might operate for the two groups. To examine this issue, a series of supplementary analyses was conducted to test the equivalence of the regression models for Maori and European offenders. This analysis showed that the effects of the covariate factors were not significantly different between Maori and European offenders, and that the same model could be applied across the two groups. Thus, within the limits of the available data, it would appear that the same explanatory model was applicable to both Maori and European offenders, and hence that the main effects analysis above was probably adequate to examine the issue of bias.