2.5 Findings from Reconviction/Re-imprisonment Analysis

The extent to which ethnically-based bias may be a factor in judicial sentencing decisions can also be inferred from the Department’s own risk assessment methodology (“RoC*RoI” 1). A statistically-based formula based on the conviction and sentencing outcomes of many tens of thousands of offenders, RoC*RoI was introduced to enable estimation of individual offenders’ likelihood of being reconvicted or re-imprisoned. It was developed through exhaustively analysing the relative contribution of a range of offender-related factors in influencing reconviction and sentencing outcomes.

To achieve this capability, the developers used the reconviction histories of tens of thousands of offenders who were convicted of at least one criminal offence in the 1980s. The characteristics of those who were reconvicted within a five-year period were then closely examined. A wide range of possible “predictors” of subsequent reconviction was selected, and then either singly, or in combination with others, run through sophisticated statistical analyses to measure the extent of their predictive power. This extremely laborious, trial-and-error process occurred over many months, and progressively served to identify a set of specific factors that either singly, or in various combinations, could be used to predict the likelihood of future outcomes.

To illustrate the process of factor selection, well-known correlates of risk in convicted offenders include age (younger), gender (male) and number of previous convictions (higher number). Each of these is strongly associated with higher risk of reconviction. These factors showed up as important predictors in the RoC*RoI model, and were incorporated into the final design. However, more subtle factors were also found to be predictive: for example, length of time spent in custody proved to be strongly predictive. Further, unique combinations of factors were similarly important: for example, the effect of offender age was shown to be mediated by the amount of time spent in custody. A 22 year old offender who had already spent five years in custody might have the same (high) risk as an 18 year old who had spent only one year in prison.

Having completed the statistical analyses of factors, the final algorithm has subsequently been tested against new sets of reconviction data, and shown to be accurate. More recently, the reconviction and re-imprisonment rates of offenders released from prison or commencing new community sentences in 2003/04 were analysed with reference to offenders’ risk scores, and almost perfect correlation was found: very low rates of reconviction or re-imprisonment were found amongst offenders with very low scores, and very high rates amongst high-score offenders (see Table 10, p. 40 of the Department of Corrections 2004/05 Annual Report).

The developers of RoC*RoI examined ethnicity as a potential predictor of reconviction, and initially found it to contribute an additional, though small, degree of predictive power. This was demonstrated when, for testing purposes, ethnicity was set to “0” in the model, but overall accuracy reduced by just 2%. However, the model was subsequently updated, using more recent reconviction history data, to ensure that risk scores for reconviction and re-imprisonment were aligned to current conviction and sentencing frequencies. When this was done, it was found that removing ethnicity entirely from the model could be done without any measurable loss in accuracy.

Findings from this recent re-calibration of RoC*RoI have important implications for the current question of whether disproportionate representation of Māori within criminal justice statistics is reflective of bias. On the basis of the RoC*RoI data it can reasonably be deduced that, following a reconviction, the likelihood that an offender will be sentenced to a term of imprisonment, rather than a community-based sentence, is not significantly affected by the offender’s ethnicity. That is, in the case of offenders of different ethnicity, where all else is equal (for example, seriousness of current offence, number of previous convictions, number of previous prison terms, etc), equivalent sentencing outcomes can be expected.