The prediction of criminal recidivism: the implication of sampling in prognostic models

Eur Arch Psychiatry Clin Neurosci. 2007 Apr;257(3):129-34. doi: 10.1007/s00406-006-0678-y. Epub 2006 Sep 14.


Background: Instruments based on actuarial forensic risk assessment models are sensitive to the calibration sample, and the inclusion criteria for the subjects of a study population will influence the features of the resulting model. If the same instrument is used in populations that are not part of the calibration sample, the discriminative validity of the instrument is jeopardized; thus the definition of the calibration sample is very important. The aim of this study was to examine how sensitive prognostic models are to the calibration sample.

Method: Two samples (N = 773) of offenders sentenced to at least 10 months in prison for a violent or sexual offense were used in this study. The "sanction sample" (recruited during August 2000, N = 515) consisted of all violent and sexual offenders actively administrated by the Criminal Justice System of Zurich, Switzerland. The "verdict sample" (recruited over two years, N = 258) included all offenders convicted in the Canton of Zurich during a two-year period. Both samples were unbiased, since all subjects that met the study criteria were included. In the first analysis, differences between the two samples were shown with respect to socio-demographic, criminological, and psychiatric variables using bivariate logistic regressions. In the second analysis, recidivism was estimated separately for both samples, using a logistic regression model as a function of a set of psychiatric, socio-demographic and criminological variables.

Results: Bivariate logistic regression showed that different risk factors for recidivism existed for both samples.

Conclusion: Forensic risk assessment models are very sensitive to the calibration sample. There is strong evidence that, even when index-offenses and the socio-cultural background are the same, risk factors for recidivism differ depending on the stage of the judicial process in which the subjects are (e.g. whether a subject is indicted, on conditional release, on parole, or no longer under the supervision of a parole board). Unfortunately, none of the currently available actuarial risk assessment instruments that have been validated in European countries consider the different stages of the judiciary process.

MeSH terms

  • Adult
  • Crime / statistics & numerical data*
  • Humans
  • Logistic Models
  • Male
  • Models, Statistical
  • Pilot Projects
  • Prognosis
  • Recurrence
  • Risk Assessment
  • Sampling Studies
  • Sex Offenses / statistics & numerical data
  • Socioeconomic Factors
  • Switzerland / epidemiology
  • Violence / statistics & numerical data