Assessing Recidivism Risk Among Young Offenders
| DOI | 10.1375/acri.38.3.324 |
| Published date | 01 December 2005 |
| Date | 01 December 2005 |
| Author | Aldis L. Putniņš |
324 THE AUSTRALIAN AND NEW ZEALAND JOURNAL OF CRIMINOLOGY
VOLUME 38 NUMBER 3 2005 PP.324–339
Address for correspondence: Aldis L. Putni¸nˇs, Chief Clinical Psychologist, Youth &
Juvenile Justice, 2 Norton Summit Rd, Magill South Australia 5072, Australia. E-mail:
Aldis.Putnins@dfc.sa.gov.au
Assessing Recidivism Risk
Among Young Offenders
Aldis L. Putni¸nˇs
South Australian Government Department for Families and Communities, Australia
The development of a recidivism risk index for use with young offend-
ers is described.A construction sample was drawn from the first 458
incarcerated youths in South Australia approached to undertake a routine
standardised psychosocial screening. Items that met the selection criteria
for the risk index were number of prior proven offences,current age, age
at first offence, alcohol and inhalant use frequencies,and ADHD signs.The
resulting index had significant correlations (.28–.53) with 6-month
postrelease recidivism status among various assessment, age, gender and
ethnic subgroups. A correlation of .36 was obtained between the index
and 6-month postrelease recidivism status with an independent sample of
149 incarcerated youths. The results compare well with the predictive
validities reported elsewhere for other risk instruments. An important
use for the index could be to guide more intensive services toward those
who are at highest risk of reoffending.
Many decisions made by the police, judiciary, parole/review boards and various staff
involved with offenders are, at least to some extent, based on predictions of future
behaviour. Gottfredson and Gottfredson (1988) have described this as the ‘ubiqui-
tous centrality of prediction’ in the criminal justice system. A predicted outcome of
particular relevance in work with offenders is that of reoffending.
There are fundamentally two ways that decisions regarding the likelihood of
behavioural outcomes can be made, namely clinical and actuarial methods. The
clinical method is based on personal judgment. From knowledge of past outcomes,
perhaps supplemented by theory or reinforced by collective professional experience
passed on through training and interaction with colleagues, the assessor makes a
judgment about the degree of similarity with those past cases and forms an opinion
about the likely outcome of the present case. A degree of flexibility is maintained
and the assessor can take into consideration whatever information might be avail-
able, including the many nuances of individual differences that are often difficult
to fully quantify. Prognostic accuracy is, however, likely to be eroded when the
assessor’s experience is limited. This can occur if the disorder does not occur very
often, is concentrated in populations that the assessor has had little exposure to, or
the outcome is delayed so that the assessor seldom observes the full course of the
disorder. Various subjective biases and cognitive limits to information synthesis can
reduce consistency between assessors.
The actuarial method is quantitative and relies on using known correlates of the
outcome to estimate the probability of similar outcomes in the future. The advan-
tages of such methods are that they are standardised, quantifiable, have an empirical
basis and have clearly defined criteria and decision-making rules. The content,
method and application of actuarial measures can be critically reviewed and refined
much more easily than can clinical prediction because of the clarity of the proce-
dures and the consistency of their application.
Is the actuarial method applied to human behaviour superior to clinical judge-
ment? Dawes, Faust and Meehl (1989) concluded from the overwhelming weight of
evidence of about 100 comparative predictive studies that it is. This conclusion has
been reinforced by a subsequent meta-analysis of 136 studies (Grove & Meehl,
1996) and by Hanson and Bussière’s (1998) meta-analysis of sexual offender recidi-
vism studies. This is not to say that clinical observations cannot be part of the data
considered when making a prediction (such information is often very valuable), but
the prediction itself is usually more accurately made using the actuarial method.
The critical point is that the items must have demonstrated empirical relationships
with the predicted outcome.
The relationships between individual variables and recidivism tend to be weak.
This was evident in the earliest studies of parole outcome prediction (see Borden,
1928; Warner, 1923). The first validated instrument was the parole prediction
procedure reported by Ernest Burgess in 1928 and put into use in 1933 in Illinois
(Lejins, 1962). It consisted of items that had been found to distinguish between
parole violators and nonviolators. Each variable was scored as a dichotomy, given
the same weight and then summed. This method of scale construction has come to
be known as the Burgess method. Since then a number of theoretically more power-
ful techniques have been proposed or used to analyse predictive relationships and to
construct risk measures. These include multiple linear regression, logistic regression,
clustering methods, multidimensional contingency table analysis, automatic inter-
action detector analysis and an iterative classification tree procedure (Gottfredson,
1987; Steadman et al., 2000; Tarling & Perry, 1985). Despite the mathematical
sophistication of these techniques, when prediction instruments constructed using
such methods are further tested, their predictive strengths usually fall to levels that
are no better (sometimes less) than instruments based on the Burgess method (see
Challinger, 1974; Farrington & Tarling, 1985; Gagliardi, Lovell, Peterson, &
Jemelka, 2004; Gottfredson, 1987; Gottfredson & Snyder, 2002; Silver, Smith, &
Banks, 2000). It seems that the more sophisticated methods are more prone to
capitalise on chance relationships in construction samples and are thus more prone
to ‘shrinkage’ of predictive strength when applied to validation samples. At present
there is no convincing evidence that any single construction method has consis-
tently superior predictive validity. However, in their comparative study Silver et al.
(2000) found that the performance of different analytic methods depended to some
degree on the purpose of the prediction device. They suggested that if the purpose
were to categorise cases into a series of risk classes, then linear-based approaches
such as the Burgess method or logistic regression would be preferred.
325
ASSESSING RECIDIVISM RISK AMONG YOUNG OFFENDERS
THE AUSTRALIAN AND NEW ZEALAND JOURNAL OF CRIMINOLOGY
To continue reading
Request your trial