Prediction of Crime and Recidivism
PREDICTION OF CRIME AND RECIDIVISM
In the Protagoras, Plato states that "he who undertakes to punish with reason does not avenge himself for past offence, since he cannot make what was done as though it had not come to pass; he looks rather to the future, and aims at preventing that particular person . . . from doing wrong again" (p. 139). Twenty-four hundred years later, preventing crime by predicting who is likely to commit it and intervening in their lives to deflect the prediction is ubiquitous in the legal system. Decisions as to who should go to prison (sentencing) and when they should be let out (parole) are in substantial part predictive decisions. Assessments of the probability of future crime influence the judicial choice of whether to grant release of an offender on bail pending trial and whether to treat a juvenile as a juvenile or to waive him or her to adult court. The U.S. Supreme Court has stated that it was permissible for a state to make the imposition of the death penalty contingent upon a prediction that a murderer, unless executed, would be likely to offend again. "It is, of course, not easy to predict future behavior," Justice John Paul Stevens wrote. "The fact that such a determination is difficult, however, does not mean that it cannot be made" ( Jurek v. Texas, 428 U.S. 262 (1976)) (see Monahan and Walker for a review of caselaw in this area).
This prevention-through-prediction strategy can take the form of changing the people who are predicted to be criminal, for example, by subjecting them to treatment in order to lower the probability that they will commit a crime. Alternatively, prevention can take the form of isolating those who are predicted to be criminal by incapacitating them in an institution, so as to deprive them of potential victims. Reviewing the history of prediction in Anglo-American law, Alan Dershowitz concluded that "the preventive confinement of dangerous persons. . .who are thought likely to cause serious injury in the future has always been practiced, to some degree, by every society in history regardless of the jurisprudential rhetoric employed. . . . Moreover, it is likely that some forms of preventive confinement will continue to be practiced by every society" (p. 57).
Preventive confinement certainly is increasingly practiced in the United States. In Kansas v. Hendricks, 521 U.S. 346 (1997), the Supreme Court upheld a civil means of lengthening the detention of certain criminal offenders scheduled for release from prison. Kansas' Sexually Violent Predator Act established procedures for commitment to mental hospitals of persons who, while they do not have a "major mental disorder" (such as schizophrenia), do have a "mental abnormality" (such as, in Hendricks's case, the personality disorder of "pedophilia") which makes them "likely to engage in predatory acts of sexual violence" (p. 357). A "mental abnormality" was defined in the act as a "congenital or acquired condition affecting the emotional or volitional capacity which predisposes the person to commit sexually violent offenses in a degree constituting such person a menace to the health and safety of others" (p. 352). The Court stated:
A finding of dangerousness, standing alone, is ordinarily not a sufficient ground upon which to justify indefinite involuntary commitment. We have sustained civil commitment statutes when they have coupled proof of dangerousness with the proof of some additional factor, such as a "mental illness" or "mental abnormality." These added statutory requirements serve to limit involuntary civil confinement to those who suffer from a volitional impairment rendering them dangerous beyond their control. . . . The precommitment requirement of a "mental abnormality" or "personality disorder" is consistent with the requirements of these other statutes that we have upheld in that it narrows the class of persons eligible for confinement to those who are unable to control their dangerousness. (p. 358)
It was not until the twentieth century that attempts were made to systematize the crime prediction process. In 1928, E. W. Burgess examined the official records of several thousand former inmates of Illinois prisons and identified numerous factors, such as prior criminal record and age at release, that were associated with the commission of crime on parole. This "experience table," as he called it (which would now be termed "statistical" or "actuarial" prediction), was then used to assess the suitability of other offenders for parole release. "Thus began a criminological research tradition characterized by the production of increasingly sophisticated instruments for predicting criminal behavior. . . . Indeed, it may be said that most later work has been largely a refinement and elaboration of Burgess' basic method" (American Justice Institute, p. 7).
Prediction by means of statistical tables, however, has been only one of two approaches used in forecasting the occurrence of crime. The other approach is generally known as "clinical" prediction. This method involves experts examining an offender and rendering an opinion based upon their subjective weighing of the factors they believe relevant to the commission of future crime. Only since the 1970s have there been scientific attempts to evaluate the accuracy of clinical predictions.
This entry will selectively review the empirical literature on the prediction of criminal behavior that has been published since the report of the President's Commission on Law Enforcement and Administration of Justice in 1967. (Reviews of studies published between Burgess's work in 1928 and the report of the President's Commission can be found in Gottfredson.) This article will also place emphasis on the prediction of violent forms of criminal behavior, rather than of property offenses. First, however, it is necessary to review briefly four key concepts in predictive decision-making.
Predictor and criterion variables
The prediction process requires that a person be assessed twice. At Time One, he or she is placed into certain categories that are believed, for whatever reason, to relate to the behavior being predicted. If one is interested in predicting how well a person will do in college, the categories might be grades in high school, letters from teachers (rated, for example, as "very good," "good," or "poor"), and the quality of the essay written for the application (perhaps scored on a scale of 1 through 10). These are all predictor variables —categories consisting of different levels that are presumed to be relevant to what is being predicted. For criminal behavior, the predictor variables might include frequency of past criminal acts, age, or degree of impulse control.
At some specified later point, Time Two, another assessment of the person is performed to ascertain whether he or she has or has not done what was predicted. This entails assessing the person on one or more criterion variables. For predicting success in college, the criterion variables might be college grades, class rank, or whether or not the person obtained a job in the field that person wanted (scored simply as "yes" or "no"). For criminal behavior, the criterion variables may include self-report, either arrest or conviction for certain crimes, or involuntary commitment to a mental hospital as a person dangerous to others.
Outcome of positive and negative predictions
There are four statistical outcomes that can occur when one is faced with making a prediction of any kind of future behavior. One can either predict that the behavior, in this case crime, will take place or that it will not take place. At the end of some specified period, one observes whether the predicted behavior actually has taken place or has not taken place.
If one predicts that crime will take place and later finds that this has indeed happened, the prediction is called a true positive. One has made a positive prediction and it has turned out to be correct, or true. Similarly, if one predicts that crime will not take place and it in fact does not, the prediction is called a true negative, since one has made a negative prediction of crime and it turned out to be true. These, of course, are the two outcomes that one wishes to maximize in making predictions.
There are also two kinds of mistakes that can be made. If one predicts that crime will take place and it does not, the outcome is called a false positive. A positive prediction was made and it turned out to be incorrect, or false. In practice, this kind of mistake usually means that a person has been unnecessarily detained to prevent a crime that would not have taken place in any event. If one predicts that violence will not take place and it does, the outcome is called a false negative. In practice, this kind of mistake often means that someone who is not detained, or who is released from detention, commits a criminal act in the community. Obviously, predictors of violence try to minimize these two outcomes.
Decision rules. Decision rules involve choosing a "cutting score" on some predictive scale, above which one predicts, for the purpose of intervention, that an event will happen. A cutting score is simply a particular point on some objective or subjective scale. When one sets a thermostat at 68°, for example, one is establishing a cutting score for the operation of a heating unit. When the temperature drops below 68° the heat comes on, and when it goes above 68° the heat goes off. The "beyond a reasonable doubt" standard of proof in the criminal law is a cutting score for the degree of certainty that a juror must have in order to vote for conviction. Conviction is to take place only if doubt is "unreasonable." In the context of parole prediction, one could state that if a prisoner has a higher than X probability of recidivism, he or she should be denied parole for a given period.
The term base rate refers to the proportion of individuals in the group being examined who can be expected to engage in violent criminality. It is the average, or "chance," rate that prediction seeks to improve upon. Prediction schemes can be evaluated either in terms of how well they differentiate true and false positives or in terms of how much they improve on the base rates. Thus, in the Michigan parole prediction study discussed below, the base rate for violent recidivism among all persons released from prison was 10 percent. A prediction scale was devised that could identify one subgroup of which 40 percent committed a violent crime after release. This device, therefore, improved on the base rate by a factor of four, although 60 percent of the individuals predicted to be violent were still false positives.
Here we consider four studies representative of the best of actuarial prediction. Ernst Wenk, James Robison, and Gerald Smith in 1972 reviewed several massive studies on the prediction of violent crime undertaken in the California Department of Corrections. One study, begun in 1965, attempted to develop a "violence prediction scale" to aid in parole decision-making. The predictive items employed included commitment offense, number of prior commitments, heroin use, and length of imprisonment. When validated against discovered acts of actual violent crime by parolees, the scale was able to identify a small class of offenders (less than 3 percent of the total) of whom 14 percent could be expected to be violent. The probability of violence for this class was nearly three times greater than that for parolees in general, only 5 percent of whom, by the same criteria, could be expected to be violent. However, 86 percent of those identified as potentially violent were not, in fact, discovered to have committed a violent crime while on parole.
The State of Michigan Department of Corrections in 1978 introduced an actuarial prediction device, the Assaultive Risk Screening Sheet, for use in program assignment and parole decision-making. Data on 350 variables were collected for over two thousand male inmates released on parole for an average of fourteen months in 1971. Statistical analyses were performed on the data for half the subjects to derive an actuarial table relating to arrest for a new violent crime while on parole. The resulting factors were then applied to the other half of the subjects in order to validate the predictive accuracy of the scale. The six items in the table were: "crime description fits robbery, sex assault, or murder," "serious institutional misconduct," "first arrest before 15th birthday," "reported juvenile felony," "crime description fits any assaultive felony," and "ever married." Using combinations of these items it was possible to place the offenders into five discrete categories: very low risk (2.0 percent recidivism), low risk (6.3 percent), middle risk (11.8 percent), high risk (20.7 percent), and very high risk (40.0 percent).
A noteworthy advance in the development of actuarial risk assessment to predict violence in the community was reported by Quinsey, Harris, Rice, and Cormier. A sample of over six hundred men who were either treated or administered a pretrial assessment at a maximum security forensic hospital in Canada served as subjects. All had been charged with a serious criminal offense. A wide variety of predictive variables were coded from institutional files. The criterion variable was any new criminal charge for a violent offense, or return to the institution for an act that would otherwise have resulted in such a charge. The average time at risk after release was almost seven years. Twelve variables were identified for inclusion in the final statistical prediction instrument, including an offender's score on the Hare Psychopathy Checklist, alcohol abuse, and elementary school maladjustment. If the scores on this instrument were dichotomized into "high" and "low," the results indicated that 55 percent of the "high scoring" subjects committed violent recidivism, compared with 19 percent of the "low scoring" group.
Along these lines, a major meta-analysis of actuarial risk factors for crime and violence among mentally disordered offenders, (Bonta, Law, and Hanson) found those risk factors to be remarkably similar to well-known risk factors among the general offender population:
Criminal history, antisocial personality, substance abuse, and family dysfunction are important for mentally disordered offenders as they are for general offenders. In fact, the results support the theoretical perspective that the major correlates of crime are the same, regardless of race, gender, class, and the presence or absence of mental illness (p. 139).
Finally, Steadman and colleagues studied actuarial risk assessment among a sample of men and women discharged into the community from acute psychiatric facilities. Using 134 risk factors measured in the hospital, they were able to classify approximately three-quarters of the patients into one of two risk categories. "High violence risk" patients were defined as being at least twice as likely as the average patient to commit a violent act within the first twenty weeks following hospital discharge. "Low violence risk" patients were defined as being at most half as likely as the average patient to commit a violent act within the first twenty weeks following hospital discharge. Since 18.7 percent of all patients committed at least one violent act toward another during this period, this meant that high violence risk patients had at least a 37 percent likelihood of being violent and low violence risk patients had at most a 9 percent likelihood of being violent. The actual rate of violence observed in the high risk group was 44 percent and in the low risk group was 4 percent.
Despite its long history and obvious advantage of economy and reproducibility, a statistical approach is used in only a minority of predictive decision-making points in the criminal justice system. Primary reliance is placed on the use of intuitive human judgment in many situations calling for a prediction of future crime.
Many studies have attempted to validate the ability of psychiatrists and psychologists to predict violent behavior. Here we consider three of the best ones. Harry Kozol, Richard Boucher, and Ralph Garofalo conducted a ten-year study involving almost six hundred male offenders, most of whom had been convicted of violent sex crimes. At the Massachusetts Center for the Diagnosis and Treatment of Dangerous Persons, each offender was examined independently by at least two psychiatrists, two psychologists, and a social worker. These clinical examinations, along with a full psychological test battery and "a meticulous reconstruction of the life history elicited from multiple sources—the patient himself, his family, friends, neighbors, teachers, employers, and court, correctional and mental hospital record" (p. 383), formed the database for their predictions.
Of the 592 patients admitted to their facility for diagnostic observation, 435 were released. Kozol and his associates recommended the release of 386 as nondangerous and opposed the release of 49 as dangerous, with the court deciding otherwise. During the five-year follow-up period, 8 percent of those predicted not to be dangerous became recidivists by committing a serious assaultive act, and 35 percent of those predicted to be dangerous committed such an act.
In 1966, the Supreme Court held that Johnnie Baxstrom had been denied equal protection of the law by being detained beyond his maximum sentence in an institution for the criminally insane without the benefit of a new hearing to determine his current dangerousness (Baxstrom v. Herold, 383 U.S. 107 (1966)). The ruling resulted in the transfer of nearly one thousand persons reputed to be some of the most "dangerous" mental patients in the state of New York from hospitals for the criminally insane to civil mental hospitals. It also provided an excellent opportunity for naturalist research on the validity of the psychiatric predictions of dangerousness upon which the extended detentions were based.
In their classic 1974 study of careers of the criminally insane, Henry Steadman and Joseph Cocozza found that the level of violence experienced in the civil mental hospitals was much less than had been feared, that the civil hospitals adapted well to the massive transfer of patients, and that the Baxstrom patients received the same treatment as the civil patients. Only 20 percent of the Baxstrom patients were assaultive to persons in the civil hospital or the community at any time during the four years after their transfer. Furthermore, only 3 percent were sufficiently dangerous to be returned to a hospital for the criminally insane during a four-year period after the decision. The researchers followed 121 Baxstrom patients who had been released into the community, that is, discharged from both the criminal and civil mental hospitals. During an average of two and a half years of freedom, only 9 of the 121 patients (7.5 percent) were convicted of a crime, and only one of those convictions was for a violent act.
Lidz, Mulvey, and Gardner (1993) took as their subjects not prisoners but rather male and female patients being examined in the acute psychiatric emergency room of a large civil hospital. Psychiatrists and nurses were asked to assess potential patient violence to others over the next six-month period. Violence was measured by official records, by patient self-report, and by the report of a collateral informant in the community (e.g., a family member). Patients who elicited professional concern regarding future violence were found to be significantly more likely to be violent after discharge (53 percent) than were patients who had not elicited such concern (36 percent). The accuracy of clinical prediction did not vary as a function of the patient's age or race. The accuracy of clinicians' predictions of male violence substantially exceeded chance levels, both for patients with and without a prior history of violent behavior. In contrast, the accuracy of clinicians' predictions of female violence did not differ from chance. While the actual rate of violent incidents among released female patients (49 percent) was higher than the rate among released male patients (42 percent), the clinicians had predicted that only 22 percent of the women would be violent, compared with predicting that 45 percent of the men would commit a violent act. The inaccuracy of clinicians at predicting violence among women appeared to be a function of the clinicians' serious underestimation of the base rate of violence among mentally disordered women (perhaps due to an inappropriate extrapolation from the great gender differences in rates of violence among persons without mental disorder).
Conclusions and implications
In no sense do the data on the prediction of criminal behavior compel their own policy implications. Given that the level of predictive validity revealed in the research has at least in the case of violent crime been rather modest, one could use the data to argue for across-the-board reductions in the length of institutionalization of prisoners: since society cannot be sure who will do harm, it should detain no one. Alternatively, and with equal fervor and logic, one could use the same data to argue for across-the-board increases in the length of institutionalization: since society cannot be sure which offenders will be nonviolent, it should keep them all in. Whether one uses the data in support of the first or the second of these implications will depend upon how one assesses and weighs the various costs and benefits associated with each, or upon the nonutilitarian principles for punishment that one adopts. In regard to the former approach, the principal impediment to developing straightforward cost-benefit ratios for predictive decision-making is the lack of a common scale along which to order both costs and benefits. For example, how are "years in a prison" to be compared with rapes, robberies, murders, or assaults prevented? John Monahan and David Wexler (p. 38) have argued in this regard that when a behavioral scientist predicts that a person will be "dangerous" to the extent that state intervention is needed, that scientist is making three separable assertions:
- The individual being examined has certain characteristics.
- These characteristics are associated with a certain probability of violent behavior.
- The probability of violent behavior is sufficiently great to justify preventive intervention.
The first two of these assertions, Monahan and Wexler hold, are professional judgments within the expertise of the behavioral sciences—judgments that can, of course, be challenged in court. The third is a social-policy statement that must be arrived at through the political process, and upon which the behavioral scientist should have no more say than any other citizen. What the behavioral scientist should do, they argue, is to present and defend an estimate of the probability that the individual will engage in criminal behavior. Judges and legislators, however, should decide whether this probability of criminal behavior is sufficient to justify preventive interventions because they are the appropriate persons to weigh competing claims among social values in a democratic society.
Barbara Underwood has asserted that one cannot evaluate the usefulness of prediction from a policy perspective other than in the context of the feasible alternatives to prediction as a basis for making decisions. In the sentencing and parole context, the principal alternative to making decisions on the basis of prediction is making them on retributive grounds. There have been numerous proposals, based in part upon dissatisfaction with the research findings reviewed above, to abandon prediction altogether and limit criminal disposition to consideration of "just deserts" for the crime committed (von Hirsch). The chief difficulty here, however, lies in the assessment of what constitutes "just deserts" for a given criminal behavior. Although the relative ranking of deserved punishments for given crimes is reliable (everyone agrees that murder deserves more punishment than jaywalking), the absolute punishment to be "justly" ascribed is determinable by social consensus only within a broad range. If for no other reason than the lack of any workable alternative, the prediction of criminal behavior is likely to remain an essential aspect of the criminal justice system.
See also Crime Causation: Biological Theories; Crime Causation: Psychological Theories; Crime Causation: Sociological Theories; Criminology: Modern Controversies; Incapacitation; Probation and Parole: History, Goals, and Decision-Making; Probation and Parole: Procedural Protection; Probation and Parole: Supervision; Scientific Evidence; Sentencing: Guidelines; Sentencing: Presentence Report: Violence.
American Justice Institute, with the National Council on Crime and Delinquency. Sentencing and Parole Release. Classification Instruments for Criminal Justice Decisions, vol. 4. Washington, D.C.: U.S. Department of Justice, National Institute of Corrections, 1979.
Bonta, James; Law, Moira; and Hanson, Karl. "The Prediction of Criminal and Violent Recidivism among Mentally Disordered Offenders: A Meta-Analysis." Psychological Bulletin 123 (1998): 123–142.
Burgess, Ernest W. et al. The Working of the Indeterminate Sentencing Law in the Parole System in Illinois. Springfield: Illinois Parole Board, 1928.
Dershowitz, Alan M. "The Origins of Preventive Confinement in Anglo-American Law: The English Experience." University of Cincinnati Law Review 43 (1974): 1–60.
Gottfredson, Don M. "Assessment and Prediction Methods in Crime and Delinquency." Task Force Report. Juvenile Delinquency and Youth Crime. Washington, D.C.: President's Commission on Law Enforcement and Administration of Justice, Task Force on Juvenile Delinquency, 1967.
Kozol, Harry L.; Boucher, Richard J.; and Garofalo, Ralph E. "The Diagnosis and Treatment of Dangerousness." Crime and Delinquency 18 (1972): 371–392.
Lidz, Charles; Mulvey, Edward; and Gardner, William. "The Accuracy of Predictions of Violence to Others." Journal of the American Medical Association 269 (1993): 1007–1011.
Monahan, John. The Clinical Prediction of Violent Behavior. Rockville, Md.: U.S. Department of Health and Human Services, Public Health Service, Alcohol, Drug Abuse, and Mental Health Administration, National Institute of Mental Health, 1981.
——. "Clinical and actuarial predictions of violence." In Modern Scientific Evidence: The Law and Science of Expert Testimony. Edited by D. Faigman, D. Kaye, M. Saks, and J. Sanders. St. Paul, Minn.: West Publishing Company, 1997. Pages 300–318.
Monahan, John, and Walker, Laurens. Social Science in Law: Cases and Materials. Westbury, N.Y.: Foundation Press, 1998.
Monahan, John, and Wexler, David B. "A Definite Maybe: Proof and Probability in Civil Commitment." Law and Human Behavior 2 (1978): 37–42.
Plato. Protagoras. Translated by W. R. M. Lamb. Loeb Classical Library, vol. 4. London: Heinemann, 1925.
President's Commission on Law Enforcement and Administration of Justice. The Challenge of Crime in a Free Society. Washington, D.C.: The Commission, 1967.
Quinsey, Vernon; Harris, Grant; Rice, Marnie; and Cormier, Catherine. Violent Offenders: Appraising and Managing Risk. Washington, D.C.: American Psychological Association, 1998.
State of Michigan, Department of Corrections. Information on Michigan Department of Corrections' Risk Screening. Lansing, Mich.: The Department, 1978.
Steadman, Henry J. "A New Look at Recidivism among Patuxent Inmates." Bulletin of the American Academy of Psychiatry and the Law 5 (1977): 200–209.
Steadman, Henry J., and Cocozza, Joseph J. Careers of the Criminally Insane: Excessive Social Control of Deviance. Lexington, Mass.: Heath, Lexington Books, 1974.
Steadman, Henry; Silver, Eric; Monahan, John; Appelbaum, Paul; Robbins, Pamela; Mulvey, Edward; Grisso, Thomas; Roth, Loren; and Banks, Steven. "A Classification Tree Approach to the Development of Actuarial Violence Risk Assessment Tools." Law and Human Behavior 24 (2000): 83–100.
Underwood, Barbara D. "Law and the Crystal Ball: Predicting Behavior with Statistical Inference and Individualized Judgment." Yale Law Journal 88 (1979): 1408–1448.
von Hirsch, Andrew. Doing Justice. The Choice of Punishments. New York: Hill & Wang, 1976.
Wenk, Ernst A.; Robison, James O.; and Smith, Gerald W. "Can Violence Be Predicted?" Crime and Delinquency 18 (1972): 393–402.