A means of accurately predicting whether an individual will re-offend would have a number of applications. Firstly it would provide an objective way of deciding which offenders should be granted parole. Secondly it would allow judgements to be made on the effectiveness of different prison or probation regimes. The government has recognised the potential value of an accurate predictive tool, and has financed a number of scientific attempts to design one. These Home Office studies are the most comprehensive and sophisticated British attempts to determine just how predictable re-offending is, and are detailed later in this essay.
There are two rival methods of predicting future behaviour, although they are not mutually exclusive. Broadly these are the 'clinical' method, and the 'statistical' method. The statistical method is in its essence objective. A statistician scores a subject on a range of measures, and then by referring to the percentage of previous subjects with the same scores who have acted in a certain way, arrives at a prediction of a future action for that individual.
For example a statistician may note that prison inmates with a long criminal record are more likely to re-offend than those with no previous convictions, he can then make a prediction based on this data for any individual inmate. The clinical method, in contrast, is largely subjective. It involves a prediction by qualified professionals who have an intimate knowledge of the individual's history, based on their own experience and training. For example a prison psychiatrist may give his opinion on the likelihood that an individual inmate will re-offend, basing his opinion on whatever data he thinks applicable.
Which of these methods is the most accurate is the subject of considerable controversy, and different methods tend to predominate in different fields. In his seminal work (1) Meehl (1966) examines the problem, and reviews the literature comparing the accuracy of the two methods. A number of studies comparing clinical and statistical predictions of re-offending are highlighted by Meehl, these include Schiedt (1936), Burgess (1942), Borden (1928), and Hamlin (1934). I will not examine these studies in depth. Suffice it to say that in all cases the statistical predictions proved more accurate than the clinical ones.
These results appear to have been heeded because in almost all recent studies attempting to predict re-offending, the statistical method has been used, although several attempts have been made to combine the two methods (see Simon (1971), 1964 study). 1 – Meehl, P. E. 1954. Clinical v. Statistical Prediction: a theoretical analysis and review of the evidence. University of Minnesota Press, Minneapolis. Ohlin (1951)(1) carried out what has been described as 'one of the most thorough prediction studies which have been done'(2).
Ohlin based his predictor on 12 items which had demonstrated predictive power. These were not all objective; one item was social type, in which offenders were classified under categories such as 'erring citizen', 'farmer', 'ne'er-do-well', or 'floater'. In this particular study Ohlin's sample consisted of 4,941 prisoners paroled from Illinois penitentiaries. Despite the relatively unsophisticated nature of the study, Ohlin achieved quite good results, with a Mean Cost Rating (M. C. R. )(3) of 0. 36 on validation. His results are illustrated in fig. 1. Ohlin concluded that,
"As prediction methods find more use and our experience increases, the refinement of prediction instruments and the increase of prediction accuracy can be expected to develop rapidly. " Fig. 1. Results of Ohlin's 1951 study, data taken from p. 130 of Ohlin (1951) 1 – Ohlin, S. E. 1951. Selection for Parole. Russell Sage Foundation, New York. 2 – Simon, F. E. 1971. Prediction Methods in Criminology. H. M. Stationary Office, London. 3 – The Mean Cost Rating is a statistic that describes the effectiveness of a predictor, taking into account the weights of the various items.
In the 1950's and 60's the government carried out a number of studies, as part of a review of the effectiveness of the probation service in preventing re-offending. One study that formed part of the Home Office Probation Research Project was aimed at designing an accurate statistical predictor of re-offending. This was to allow a balanced comparison of the effectiveness of different probation regimes. The study was carried out by Simon (1971). Simon based his instrument on 62 objective items of information that described the probationer's personal or criminal history.
All these items could be found in government or case records, although not all had any apparent relevance to the likelihood of re-offending. The criterion for failure was re-conviction of a Standard List offence within three years of the probation order commencing. Simon's sample consisted of 539 young men between the ages of 18 and 21, who all began probation orders in 1958. Simon carried out 17 predictive analyses on the data using a variety of methods and selections of the 62 variables. Most of these analyses produced a predictor of moderate power on construction, but almost all shrank considerably on validation.
The most successful predictive analysis was a Predictive Attribute Analysis which produced a predictor with an M. C. R. of 0. 45 on construction, which shrank to 0. 25 on validation. When the predictor was applied to a fresh set of data in 1964 it achieved an M. C. R. of 0. 18 on construction, that shrank to 0. 9 on validation. These results are considerably lower than one would expect following a review of earlier studies. Vold (1931) recorded Correlation Coefficients of over 0. 2 for several individual items, including criminal record, marital status, social type, and previous work record.
Vold's most successful predictor achieved an M. C. R. of 0. 71 on validation. This would appear exceptional, but many others have designed predictors that have achieved M. C. R. 's of between 0. 3 and 0. 4. Ohlin (1951) produced a predictor that achieved 0. 36 on validation, and this score was matched by Glaser (1962). Simon concluded that his poor results were largely the result of a sample that was not large enough to identify genuine trends, and reject co-incidental relationships, and that as such considerable shrinkage occurred on validation, this seems to be the case.
For example one item, the age of offender at the time of a break or disturbance in his domestic arrangements, showed a predictive relationship of P<. 001 on construction, this relationship vanished to P>. 05 on validation. The sample size certainly seems small in comparison with earlier studies. Vold (1931) used a sample of 1,192, Glaser based his study on a sample of 2,637, and Ohlin examined data from 17,097 offenders. Simon concludes pessimistically about the potential usefulness of prediction studies, saying, "the study shared the general fate of criminological prediction studies…
although small groups of good or bad risks could be distinguished, for many of the cases little discrimination is achieved. " The Home Office again had cause to need an efficient prediction instrument for re-offending when it introduced parole in 1968. On this occasion the study was carried out by Nuttal et al, and is detailed in Nuttal et al (1977)(1). It would appear the C. P. Nuttal and his team learnt the lessons of Simon's failure. They based their predictor on seventeen items that were easily accessible from administrative records, and that had consistently shown predictive power in previous studies.
Their sample was of 2,276 male prisoners released from prison in the first six months of 1965. The criterion for failure was re-conviction within two years of release. The results of this study were very encouraging, with the predictor achieving an M. C. R. of 0. 5 on construction that expanded to 0. 557 on validation, see fig. 2. As Nuttal concludes, this study appears to show that, "by using a large sample… an efficient instrument could be devised for predicting the likelihood of re-conviction among medium and long-term prisoners. "