Ministry of Justice, New Zealand

Introduction

The actual effects of changes in sentencing practice will depend in part on differences in recidivism (reconviction) rates between sentence types. In particular, one of the effects of increasing the use of community-based sentences for less serious cases is that subsequent sentences imposed on the same offenders are also more likely to be community-based sentences or prison sentences. This potential for sentence escalation would have less of an impact if community-based sentences have lower recidivism rates than monetary penalties. That is, if a smaller proportion of people sentenced to community-based sentences reoffend, then sentence escalation will apply to a smaller proportion of people. Indeed, one of the main aims of community-based sentences is to reduce reoffending.

A full analysis of the factors that influence recidivism rates is too large a topic to cover in detail in this report. However, because of its relevance to the sentencing issue, a summary of results from a simple model is included here.

Methods

As for the sentencing analysis, logistic regression modelling was used to determine the factors influencing the recidivism rate. The aim of the analysis was to see if different sentences have any effect on recidivism rates once the effect of other factors (such as an offenders previous criminal history) are taken into account.

Recidivism was measured as a simple dichotomous variable - whether or not the offender was reconvicted for any other offence in the two years following the conviction date of their 1991 proved case or, for offenders who served a prison sentence as a result of their 1991 case, whether they were reconvicted within two years of their estimated release date.

This very simple measure of recidivism does not take into account other important measures of the success of a sentence either from the recidivism perspective (e.g. the frequency and seriousness of reoffending) or from other perspectives (e.g. skills learnt, reparation made to the victim or community, or changes in attitude or behaviour).

The 1991 data was used to maximise the inclusion of offenders serving longer prison sentences. The full two-year follow-up was possible for all but those prison inmates serving very long sentences, with estimated release dates after mid-1996. This group comprised less than 0.1% of the total 1991 sample. Prison release dates for prisoners eligible for parole were estimated on the basis of the average proportion of the sentence served before parole (Spier 1995).

The logistic modelling methods used to develop the sentence models (Chapter 2) were also used for the recidivism model. The same explanatory variables were used, to simplify comparisons to the models developed in the body of this report, with one exception. The 'most recent previous sentence' variables were excluded and the 'previous sentence' variables modified to include the most recent sentence. For example, if the most recent sentence prior to the 1991 case was a periodic detention sentence, or any other previous sentence was periodic detention, then the 'previous periodic detention sentence' variable would be set at one.

Two models were developed. The first model excluded the 1991 sentence type and instead used the model developed to predict the recidivism rate that would be expected for each sentence type, taking into account the type of offender who is likely to receive each sentence. The second model included the sentence imposed for the 1991 case as a potential explanatory variable, to test whether sentence type is a significant predictor of recidivism.

Results

Both logistic regression models achieved a highly significant overall fit to the data, as indicated by log likelihood ratios significant at the 0.0001 level of probability. Neither of the models had a significant residual (unexplained variation) term.

Figure II.1: Plot of the predicted recidivism rate versus the actual proportion reconvicted in the two years following cases finalised in 1991

The first model, developed without the current (1991) sentence variables, was used to calculate the predicted recidivism rate for each sentence type based only on the criminal history and demographic characteristics of the offender and the type and seriousness of the current offence. The model achieved a high level of predictive accuracy. That is, when the actual recidivism rates are plotted against the rate predicted by the model, the results are close to the ideal line (Figure II.1). The results were equally good when the 1991 data-set was divided in half, with one half used to develop the model and the other half used to test the fit on unseen data (not shown).

The percentage of offenders reconvicted within two years of their 1991 case (or estimated date of release from prison) differs between sentence types (Table II.1). Offenders sentenced to imprisonment or periodic detention had the highest actual reconviction rates, at 82% and 77% of offenders reconvicted respectively. However, these sentences also had the highest predicted reconviction rates, indicating that the characteristics of the offenders receiving these sentences were also the characteristics of offenders more likely to be reconvicted (see below).

Similarly, while offenders receiving community service or a monetary penalty or no sentence had lower reconviction rates, this would be predicted on the basis of their characteristics. Supervision and community programme had intermediate levels of both actual and predicted reconviction rates.

Table II.1: Actual percentage of offenders reconvicted within two years compared to the percentage predicted using the logistic regression model, by sentence type

In general terms, the predicted reconviction rates were similar to the actual rates, indicating that the sentence type does not have a major independent effect on recidivism. The community-based sentences, except community programme, had slightly higher actual reconviction rates than would be expected on the basis of the statistical characteristics of the offender that determined the predicted rates. Conversely, prison, monetary penalties and no sentence had slightly lower actual reconviction rates than would be expected.

As these differences are quite small, a second logistic model was developed, with sentence type as one of the potential explanatory variables, to test the explicit effect of sentence type on recidivism.

The results of the second model indicate that the most important variables for predicting recidivism are the criminal history and demographic group of the offender (Table II.2), with less significant effects from the type and seriousness of offence and the type of sentence served.

The two most important variables are the length of time between the 1991 case and the case prior to that and the total number of previous cases. The longer the gap between the current and previous case, the lower the odds of being reconvicted. For example, offenders whose previous conviction occurred less than a year prior to the current case are more than four times as likely to be reconvicted compared to offenders who haven't had a conviction within the last four years.

Similarly, the greater the number of previous cases, the higher risk of reconviction. A high rate of conviction (a large number of proved charges per year) is also associated with an elevated risk of reconviction.

Age is also a very significant factor for predicting recidivism. Offenders aged under 20, and especially those aged under 17, are more likely to be reconvicted, while older offenders are less likely to be reconvicted, relative to 20 to 29 year olds. Females are less likely than males to be reconvicted, even when other characteristics of criminal history and offending are taken into account. Māori and Pacific peoples also have a higher risk of reconviction.

The seriousness of the current offence, which was a key factor in determining sentencing, has relatively little impact on recidivism. In fact, people convicted of more serious offences (those with seriousness scores of more than 180 and especially more than 365) had a lower relative risk of reconviction. Offenders who have committed an offence against the person (e.g. a violent offence) and traffic offenders have a lower probability of reconviction than property offenders. The number of charges in the current case and the plea do not appear to be related to the probability of reconviction.

Offenders sentenced to periodic detention, supervision and community service appear to have a higher probability of reconviction than offenders receiving a monetary penalty. However, imprisonment and community programme do not appear to have a significantly higher recidivism rate, once the characteristics of offenders given these sentences are taken into account.

Previous community-based and imprisonment sentences also appear to increase the risk of recidivism.

Table II.2: Logistic regression model of the probability of reconviction, 1995

Note: An odds ratio of >1.0 indicates a high relative risk (i.e. more likely to receive this sentence than the reference group). The most significant variable (highest Wald Chi-square, lowest probability P), is rank '1'.

These findings don't necessarily indicate that the community-based sentences are less effective at preventing reoffending. The results may, for example, indicate that people receiving these sentences have other characteristics that increase reoffending that could not be measured statistically. On the other hand, a wide range of variables were taken into account that should be broadly indicative of the characteristics of the offender, such as previous criminal history, yet the sentence type was still relevant over and above these factors. At the least, these findings call into question the assumption that community-based sentences have positive benefits in terms of reducing recidivism.

If community-based sentences do not reduce recidivism rates, especially relative to monetary penalties, then the feedback effect of sentence escalation on the number of correctional sentences served (i.e. an increase in the use of community-based sentences leading to a further increase in the use of community-based sentences and imprisonment) will not be offset be a reduction in numbers due to lower recidivism rates. In fact, if recidivism rates for community-based sentences, and periodic detention in particular, are higher than for monetary penalties, then the feedback effect is likely to be intensified.