Statistical methods
2.1 Development of the data-sets
The two major aims of this research were to determine the statistical variables that influence sentencing decisions and to examine whether changes in sentencing practice have occurred since the Report of the Penal Policy Review Committee was published in 1982. Therefore, information was needed for a large sample of court cases from several years, including data on the offence(s) committed, the sentence imposed, and the demographics and previous criminal history of the offender.
Due to the large size of criminal history data-sets, not all years could be studied. Four evenly-spaced years were chosen for the analysis - 1983 (before the Criminal Justice Act 1985), 1987 (before the Children, Young Persons, and Their Families Act 1989), 1991 (before the Criminal Justice Amendment Act 1993), and 1995. 1995 was selected rather than the most recent year so that analysis could also be conducted (as part of a separate research project) on recidivism rates with a two-year follow-up period. The trends in sentencing have been relatively stable since 1995 (Figure 1.1), so 1995 should be indicative of current sentencing practice.
Every offender who had one or more proved cases involving an imprisonable offence in the selected years (1983, 1987, 1991, 1995) was included in the data-set. The analysis was restricted to imprisonable offences, as the aim was to compare sentencing for offences for which the full range of sentences are available. This excludes offences for which the maximum penalty is a fine (the majority of these being minor traffic and minor disorder offences). A few other offences which have undergone major changes since 1983 were also excluded (mainly minor traffic offences and offences against a variety of Acts relating to dog control, broadcasting, tax etc).
The relevant cases for each offender were selected using the charge-based data-sets developed by the Ministry of Justice from the Case-Monitoring sub-system of the Law Enforcement System computer (see Spier 1998 for more information on this database). Cases were formed from the charge-based data by selecting the most serious sentence imposed for all charges finalised on the same day. This sentence and the offence that led to it are termed the 'current sentence' and 'current offence' for each offender in each year. The number of charges finalised in the current case, the seriousness of the current offence, the plea, and the demographic details of the offender were also taken from this data source.
The criminal and traffic histories were then requested for each offender, as identified by the personal identifier code recorded in the charge-based data. The histories include records for any known aliases linked to this identifier. The criminal and traffic histories identify all past proved charges for each offender, including the details of the sentence imposed, the offence, and the date the charge was finalised. These data were summarised into the required criminal history variables, as set out in section 2.2.2, and merged with the data on the current case.
The final data-sets, with one summary record per offender for each separate case in each year, represent a very large sample of offenders (63,011 in 1983, 76,467 in 1987, 77,278 in 1991 and 81,651 in 1995).
2.2 Selection of explanatory variables
Explanatory variables are the statistical variables tested in the logistic regression models to see which factors best explain the variation in sentences imposed between individual offenders. The explanatory variables were chosen from the factors known to influence sentencing, as detailed in Hall (1998), but excluding those factors which could not be measured using the available computerised statistics. The potential influences on sentencing are outlined in section 2.2.1. The statistical variables available for this research are listed in section 2.2.2, along with the sub-categories selected for each variable.
2.2.1 Potential influences on sentencing
Seriousness and type of offence
According to Hall (1998, page B/111): 'The most significant factor in sentencing involves an assessment of the 'gravity of the offence''... meaning both 'the gravity of the particular crime' (e.g. sexual violation is more serious than indecent assault) and 'the particular set of acts or omissions proved against the offender in the particular case' (e.g. the use of violence or a weapon, the degree of injury done or risked, the culpability of the offender, whether there was premeditation or provocation, and the vulnerability of and the degree of impact on the victim).
The seriousness of the offence can be included in the statistical analysis using the average seriousness score for each offence type, as measured by the Ministry of Justice seriousness scale. The specific circumstances of the case cannot be included in the statistical model.
The number of separate charges proved against the offender is also likely to be important. For example, a person being sentenced for ten burglaries, prosecuted as one case, is likely to receive a greater sentence than a person convicted of one burglary.
The specific type of offence may also be relevant, given the following Criminal Justice Act 1985 guidelines on sentencing:
- violent offenders should be imprisoned except in special circumstances (violent offenders are those convicted of an offence punishable by two or more years of imprisonment and who either (i) used serious violence or caused serious danger to the safety of another person, or (ii) used violence or caused danger and have a previous conviction within the previous two years for a similar offence punishable by two or more years imprisonment) (s5)
- people convicted of property offences punishable by seven years imprisonment or less should not be imprisoned, except in special circumstances (s6)
- custodial sentences may be imposed where the offender is unlikely to comply with other sentences (s9). [Therefore, if the current offence is a breach of another sentence, imprisonment may be more likely.]
As having a previous conviction for a violent offence can influence the sentencing decision, this was also included in the model.
Previous offending
The relevance of previous offending to sentencing is summarised in Sentencing Policy and Guidance: A Discussion Paper (page 100; Ministry of Justice 1997):
'In New Zealand case law there are two distinct yet related concepts regarding previous offending. New Zealand courts have clearly considered the absence of a criminal record as a mitigating factor, particularly where the person can demonstrate living in the community for many years with a 'blameless record', thus LEFTing a reduction in sentence. There are two sorts of reasons suggested for this approach. One is that persons who have not previously offended may be acting 'out of character' and that such people have built up credit (by establishing good character) which reduces their blameworthiness for offending. It follows that as convictions accumulate the credit diminishes. Where there has been earlier offending, credit may still be given (is built up again) for a subsequent significant period of law-abiding behaviour by way of reducing the effect of those previous convictions. The second is that someone offending for the first time may not have been fully aware of the seriousness of what they were doing, and is unlikely to infringe again in the same way if given a second chance in the form of a lesser sentence.'
'The converse of the above is that New Zealand courts deal with previous offending as a potentially aggravating factor at sentencing. In so doing, the courts have been careful to stipulate that this should not amount to punishing the offender again for previous convictions, and that the sentence must remain in proportion to the seriousness of the current offence(s). The Court of Appeal has stated that previous convictions may be taken into consideration in determining the character of the offender, as part of the process of fixing the sentence for a particular offence committed by a person of that character. Previous offending has also been held to be relevant to determining how offenders may respond to a particular sentence.'
As Hall (1998, page B/195) notes: 'A further aggravating factor is reoffending within a short time of previous offending...'.
Previous offending was included in the statistical models in a number of ways in an attempt to separate the potentially different effects of the number, seriousness, and frequency of past offences and the offender's response to the sentences imposed. Variables tested in the model included the total number of previous proved cases, the accumulated seriousness of the previous cases, the rate of conviction, the gap between the current and most recent offence, the number and type of previous sentences imposed, the most recent sentence, and previous convictions for a breach of a sentence. As noted above, previous convictions for a violent offence were also included in the model.
Circumstances of the offender
Hall (1998) sets out a number of characteristics that may be taken into account. Many of these could not be included in the analysis as they are not collected in the statistical data. Such factors include 'medical problems, personality disorders, emotional difficulties, financial difficulties, depression, overwork, marital and family problems' (Hall, 1998 page B/161) and the behaviour and circumstances of the offender subsequent to the offence (page B/199-2).
Of the behavioural characteristics, only the offender's plea could be taken into account. A plea of guilty may reduce the sentence depending on the circumstances of the case (Hall, 1998 page B/223-6). Personal characteristics that can be included in the statistical analysis are the age, sex and ethnicity of the offender.
Age is considered important as an indicator of rehabilitative potential (Hall 1998; page B/165-1). A youthful offender may therefore receive a lesser sentence than a more mature offender and, in particular, imprisonment sentences are to be avoided for young offenders where possible. The Criminal Justice Act 1985 states that prison sentences should not be imposed on a person under the age of 16, except for a purely indictable offence (s8). Periodic detention cannot be imposed on offenders aged under 15 years. The Youth Court cannot impose prison or community-based sentences.
In contrast, gender is not a factor in sentencing (Hall 1998, page B173-4):
'Difference in gender is not in and of itself a justification for discriminating between offenders; a female offender should not expect to be treated more favourably than a male.'
Likewise, people of different ethnic or cultural groups must be treated equally, although as Hall notes (Hall 1998, page B/172-173):
'the Court should impose the penalty which reflects matters of mitigation arising from the offender's background and personal situation, and which recognises the structure and operation of the society within which he or she lives, in particular the degree to which the offender's cultural or ethnic heritage predominates and any problems of a cross-cultural nature that may have been experienced' ... 'alternative means of rehabilitation which appropriately take account of different cultural or ethnic values should be utilised.'
2.2.2 Description of explanatory variables
Logistic regression models determine which statistical factors (explanatory variables) best explain the probability of an event. For the sentencing logistic models this corresponds to the probability of receiving a given sentence. In each case the sentence is coded as a dichotomous variable. For example, when examining the probability of imprisonment, the sentence is coded as either '1' (prison sentence) or '0' (any other sentence).
The explanatory variables, or factors tested for their influence on sentencing decisions, were also coded into dichotomous variables or discrete groups, even though continuous variables may be tested in logistic models. This was done to allow for the expected non-linear effects of the continuous variables on the intermediate sanctions. For example, while it is reasonable to expect the probability of imprisonment to increase as the seriousness of the offence increases, the same would not be expected for community-based sentences like periodic detention. Instead, the probability of periodic detention is likely to be lower for both low and high seriousness offences than for offences of moderate seriousness, as a monetary penalty is more likely for offences of very low seriousness and imprisonment for those of very high seriousness.
Some variables, such as the previous offending history, can be quantified in a number of ways. For example, the number of previous proved cases, the time that has passed since the last case, the rate of offending (convictions per year), and the seriousness of previous cases may all be important factors in determining an appropriate sentence.
The categories or sub-groupings of each variable were chosen on the basis of the approximate turning points in the relationships between the variables and the probability of each sentence type, as highlighted in the single-variable analyses in Chapters 4-10. For example, offences with seriousness scores of less than 20 are much more likely to receive a monetary penalty than more serious offences (section 10.2), while offence with a seriousness score of over 365 are more often associated with prison sentences (section 4.2).
For each variable, one category was excluded from the model to act as a reference level to compare other categories to. For example, for the ethnicity variable, the Māori and Pacific categories were entered into the model, with a combination group of all other ethnic groups (Pakeha/Other) serving as the reference group. For the continuous variables (e.g. offence seriousness), the category least likely to result in imprisonment (e.g. the lowest seriousness level) was selected as the reference level.
The explanatory variables tested in the models cover a range of potential influences on sentencing, as discussed in the previous section:
- Current seriousness - the seriousness score of the major offence in the current case (i.e. the seriousness score of the offence resulting in the most serious penalty), coded into five groups: >0-1 (reference level), >1-20, >20-180, >180-365, >365. The seriousness score does not measure the actual seriousness of the specific case, but rather the average seriousness for the type of offence. As explained in Spier (1998, p10):
'The seriousness score assigned to each offence is the average number of days of imprisonment imposed on every offender convicted of that offence between 1990 and 1994, where the average is taken over both imprisoned and non-imprisoned offenders. Suppose, for example, that between 1990 and 1994 there were 100 cases of offenders convicted of a particular offence. Of these cases, 50 resulted in a custodial sentence, and the average length of the custodial sentences imposed on these offenders was 30 days. The seriousness score for this offence is (30 * 50/100), or 15.'
The reference group for the seriousness variable comprises offences of very low seriousness (>0-1), which are mainly minor traffic offences, with some property damage offences and other minor offences such as disorderly behaviour. Offences with a seriousness score of >1-20 include offences such as theft, cannabis use, common assault and breach of periodic detention. Offences with a seriousness score of >20-180 include offences such as driving while disqualified, burglary, fraud, and male assaults female ('domestic violence'). Offences with a seriousness score of >180-365 include some more serious assaults and indecent assault, robbery and some drug dealing and serious property offences. Offences with a seriousness score of over 365 (i.e. offences resulting in average prison sentences of more than a year) are mainly serious types of violent offences such as aggravated robbery, sexual violation and assaults causing injury.
- Current offence type - the offence receiving the most serious sentence in the current case (imprisonable offences only), coded into nine groups, with property offences acting as the reference group:
- Serious offences against the person: homicide, kidnapping, serious assaults, sexual offences, robbery, and threats to kill or cause grievous harm. Note that 'offences against the person' are mainly violent offences, but also include some offences not necessarily involving violence (e.g. incest). Serious offences against the person were separated out as serious types of violent offences are given special recognition in the legislation (see section 2.2.1). Not all the offences in this group are in the 'Serious Violent' category given special recognition for parole purposes.
- 'Domestic violence': this offence group includes the offence categories most often linked with domestic violence (i.e. male assaults female, assault on a child, common assault - domestic), but excludes any domestic violence offences that are charged under other codes (e.g. grievous assault). Domestic violence was separated out as the growth of prosecutions for this offence are strongly associated with the increased use of supervision in recent years.
- Minor offences against the person: less serious offences against the person such as common assault and threats or intimidation.
- Property offences: burglary, theft, vehicle conversion, receiving stolen goods, fraud and property damage.
- Drug offences.
- Breach of periodic detention: this groups covers all breaches of community-based sentences, but the name of the group reflects the fact that 95% of the imprisonable offences in this group are breach of periodic detention, with the remainder being breach of parole. Other breaches of community-based sentences are either non-imprisonable (breach of community service and supervision) or are not offences (community programme). Breaching a community-based sentence may influence the likelihood of receiving a further community-based or prison sentence.
- Other offences against justice: breach of bail (i.e. failure to appear), escaping custody, breach of a non-molestation or protection order, and other offences against the administration of justice.
- Disorder/other offences: a wide definition of offences against good order, including disorderly behaviour, trespassing, resisting or obstructing an officer, unlawful assembly, possession of a weapon, offences under the Arms Act, and any miscellaneous offences not covered by other categories.
- Traffic offences: all imprisonable traffic offences, including driving causing death or injury, driving while disqualified, and driving under the influence of alcohol or drugs.
- Current charges - the number of charges proved in the current case, coded into three groups: 1 charge (the reference group), 2-4 charges, 5+ charges.
- Plea - recorded guilty pleas (coded as 1) were compared to all other pleas (coded 0). The recording of the plea variable is imperfect as changes of plea from not guilty to guilty during the course of the case may not always be recorded. Therefore, this variable measures the minimum effect of a guilty plea.
- Previous cases - the number of previous cases proved against the offender, coded into four groups: no previous cases (first offender) and 1-3, 4-10, 11+ previous cases. First offenders were the reference group. Only proved cases are recorded in the criminal history data. A 'first offender' by this definition may have had previous contacts with the Police (e.g. they may have received a caution or warning) or other court cases where the offender admits guilt but the case is not formally proved (e.g. offenders who successfully complete Police Diversion and youth offenders dealt with by a Family Group Conference). All previous proved cases are included (both imprisonable and non-imprisonable offences).
- Accumulated seriousness of previous cases - a weighted sum of the seriousness scores of the most recent previous cases (up to a maximum of seven previous cases), with the seriousness of the most recent cases weighted most heavily (the seriousness of the most recent case plus 0.9 times the seriousness of the next most recent case plus 0.8 times the seriousness of the third most recent case and so on).1 If a person has no previous cases this variable is set at zero. Coded into four groups: 0-10 (the reference group), >10-60, >60-180, >180.
The seriousness of the most recent case was also tested separately, but was not found to add significantly to the analysis.
- Time since the most recent case - this was included in the analysis as a long gap between offences can be considered as a mitigating factor. The time was measured between the current and last conviction date or the estimated date of release if the last sentence was a prison sentence. For people with no previous convictions this variable is set to the total number of years in which a conviction could have occurred (current age in years minus 13 years). For a person convicted while already serving a prison sentence, the variable is set at zero. Coded into four groups: one month or less, more than a month to one year, >1-4 years, >4 years (the reference group).
- Rate of conviction - the total number of charges ever proved against the offender divided by the number of years over which proved offending has occurred (the time between the first proved case and the current case). For an offender with court cases only within a single year this variable is equal to the total number of proved charges in that year. Coded into three groups: 2 or fewer charges per year (the reference group), >2-8 charges per year, >8 charges per year.
- Previous offence types - the offence type of the most recent case prior to the current case, coded into six groups: offence against the person, property offence, drug offence, offence against justice, traffic offence, other offence. These variables were not found to add significantly to the analysis during initial tests, so were dropped from the final analysis. Two previous violent offence variables were also tested (any previous violent offence and any previous serious violent offence), but these were also dropped from the final analysis.
- Previous breach - whether the offender has ever breached any community-based sentence (periodic detention, community service, supervision, parole) prior to the current case.
- Most recent sentence - the sentence received in the most recent case prior to the current case, coded into seven groups: prison, periodic detention, community programme, community service, supervision, monetary penalty (the reference group) and other sentences (including no sentence).
- Previous sentences - whether the offender has ever received a previous sentence of each of seven types: prison, periodic detention, community programme, community service, supervision, monetary penalty (the reference group) and other sentences (including no sentence). To avoid overlap between variables, the most recent sentence is excluded. For example, if the most recent sentence was a prison sentence, the previous prison sentence variable indicates whether there have also been other prison sentences in the past.
- Age - age in years on the date the current case was finalised, in four groups: <17, 17-19, 20-29, 30+. The small number of offenders with an unknown date of birth were excluded from the data. The 20-29 age group was chosen as the reference group as special consideration may be give to both younger and older offenders. Children aged under 10 years cannot be prosecuted and children aged 10-13 can only be prosecuted for murder or manslaughter. Most offenders aged under 17 years are dealt with by Family Group Conferences rather than by the courts, and so do not appear in the data used in this study. Of those who do go to court, many are dealt with solely by the Youth Court, which cannot impose prison or community-based sentences.
- Sex - females were coded as 1, males (and the 0.1% of people of unrecorded sex) were coded as 0.
- Ethnicity - coded into three groups: Māori, Pacific and any other ethnic group (the reference group). The 22% of people of unknown ethnicity were included in the reference group. Coding unknowns (some of whom will be Māori or Pacific peoples) this way means that any ethnicity-related effects will be slightly less significant than they would otherwise have been. This was seen as preferable to excluding all unknowns from the data, as this would have significantly biased the offence distribution, given that 95% of the people with an unrecorded ethnicity are traffic offenders.
This variable list is by no means exhaustive of the potential factors influencing sentencing. Factors that could not be quantified from the current statistical data collections could not be included. For example, many of the aggravating and mitigating factors outlined in the previous section could not be quantified. Also, better results may have been achieved by more extensive testing of different transformations of the variables given above. In particular, the interaction effects of different combinations of these variables were not extensively tested. (A potentially key interaction - that between previous cases and current seriousness - was tested, but was not found to add a great deal to the significance of the models.)
2.3 Fitting and testing of the logistic regression models
Logistic regression models determine which statistical factors (explanatory variables) best explain the probability of an event. For the sentencing logistic models this corresponds to the probability (statistical risk) of receiving a given sentence. Separate models were developed for each of the sentences examined: imprisonment, suspended prison sentences, periodic detention, community service, community programme, supervision and monetary penalties. In each case the outcome is coded as a dichotomous variable. For example, when examining the probability of imprisonment, the sentence is coded as either '1' (prison sentence) or '0' (any other sentence).
The statistical software procedure SAS-Logistic 2 was used to fit all the models. The variables that had a significant influence on predicting the probability of each sentence type were determined by a stepwise elimination procedure, with the entry and exit criteria set at a significance level of 0.05. That is, only variables which had a significant effect at the 95% confidence level were included in the final model. The significance of each model was estimated using the log likelihood ratio statistic, which tests the internal fit, or closeness of fit of the model to the data used to construct the model. The significance of the residuals (the error terms) was also tested.
Because a good fit may be achieved by chance if enough variables are tested, the models were also tested on unseen data. That is, each model was developed on one set of data and the fit of the model was tested on another equivalent set of data. In this initial test phase, models were fitted for each sentence type to half of the 1995 data (the 'fitted data') and tested on the other half of the 1995 data (the 'test data'). These data-sets were selected by taking cases finalised in alternate months, so that any change in sentencing practice within 1995 would not bias the results.
The general predictive accuracy of each model over the range of predicted probabilities was then tested by plotting the actual proportion of offenders receiving each sentence against the predicted proportion at different levels of predicted sentence use. Good general predictive accuracy is achieved if the actual values are close to the predicted values for different levels of probability. For example, if a group of offenders has a 50% predicted probability of imprisonment according to the model, do close to half of them actually receive a prison sentence? The accuracy of predictions was also tested by checking the predicted probability against the actual percentage of people receiving each sentence across the full range of each variable.
If the fit to the test data was reasonable, a final 1995 model was estimated for each sentence type using all the 1995 data. The most significant explanatory variables in this model (those with the highest Chi-square value) indicate the statistical factors most strongly influencing current sentencing practice. The magnitude of the independent effect of each variable is given by the odds ratio, the 'relative risk' or amount by which that variable increases or decreases the probability of receiving the sentence being examined relative to the reference group.
If sentencing practice with respect to these statistical variables has not changed over time, the model derived from the 1995 data should fit the data from other years equally well. That is, a person with similar statistical characteristics should have a similar probability of receiving any given sentence in different years if sentencing practice has not changed (assuming there have not been major changes in the average value within sub-categories of continuous variables).
The 1995 model was used to calculate the probability of receiving each sentence in each of the other years (1983, 1987, 1991) and this predicted probability was compared to the actual proportion of offenders receiving each sentence in these years. The fit of the 1995 model to each year was compared graphically by plotting the actual probability against the predicted probability for each year. In addition, the effect of specific variables on changes in sentencing practice were assessed by comparing the actual and predicted probabilities for 1995 with each of the other years for each variable.
The same methods were used to analyse the probability of receiving no sentence (Appendix I) and the probability of reconviction (Appendix II).
Footnotes
1. The specification of this variable came from research on recidivism by the Psychological Service, Department of Corrections. I am grateful to Leon Bakker for sharing the findings of this research.
2. SAS is an integrated system of software providing data management, analysis and presentation tools from the SAS Institute, Cary, NC, USA.
