National Academies Press: OpenBook

Criminal Careers and "Career Criminals,": Volume I (1986)

Chapter: 6. Use of Criminal Career Information in Criminal Justice Decision Making

« Previous: 5. Crime Control Strategies Using Criminal Career Knowledge
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 155
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 156
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 157
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 158
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 159
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 160
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 161
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 162
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 163
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 164
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 165
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 166
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 167
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 168
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 169
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 170
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 171
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 172
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 173
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 174
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 175
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 176
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 177
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 178
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 179
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 180
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 181
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 182
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 183
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 184
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 185
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 186
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 187
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 188
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 189
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 190
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 191
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 192
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 193
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 194
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 195
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 196
Suggested Citation:"6. Use of Criminal Career Information in Criminal Justice Decision Making." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume I. Washington, DC: The National Academies Press. doi: 10.17226/922.
×
Page 197

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

6 Use of Criminal Career Information in Criminal Justice Decision Making INTRODUCTION All stages of decision making within the criminal justice system may use infor- mation regarding an incliviclual's criminal career: a police officer in deciding whether or not to arrest a suspect, a pros- ecutor in cleciding whether to pursue a case, a magistrate in setting pretrial re- lease conditions, a sentencing judge in weighing community safety when ciecicI- ing on a punishment, and a parole board in making a release decision. In all of these decisions, there are multiple, often competing, objectives, of which one re- flects a concern for public safety: that concern involves at least an implicit at- tempt to assess an indiviclual's criminal career. The dimensions of criminal ca- reers cannot be measurer! directly, but decision makers often attempt to draw inferences about them from observable characteristics such as prior record, age, employment status, or drug clependency. The resulting characterization of an of- fender reflects a decision maker's own experience with offenders in the context of her or his personal background and 155 theories of behavior. Thus, these con- structed portraits of offenders are likely to be basest on varying information and to slider considerably across decision makers. At each decision stage, the characteris- tics used to assess a suspect are con- strained in part by the type of information at hand and by the time avaflable to gather adclitional information. Aciclitional time may permit successive decision makers to assemble a more complete pic- ture. However, decision makers are rarely able to generate detafled informa- tion about the most operationally relevant criminal career dimensions: the fre- quency of offending and the time remain- ing in the career. Rather, they use the available information to roughly cate- gorize inclivicluals on the basis of a presumed likelihood of future criminal activity by converting an indivicluaT's observable attributes into some general risk classification, such as high, low, or medium risk. This process may be ex- plicit, as in the formal scoring methods used by parole boards for release deci- sions and by prosecutors for case assign

156 meets to criminal career units, or implicit, anc! perhaps only as a secondary objec · . · , · ~ Eve, In a JUC .ge s sentencing ctec~s~ons. For most decisions, a risk classification contributes to the selection of one of the available choices low bone! or high boncI; a sentence to prison, jail, or proba- tion; release on parole or clenial but it is almost never the only consideration. The complexity of these other considerations, such as retribution in sentences or strength of the evidence in prosecutorial decisions, makes it particularly clifficult to isolate the role of criminal career informa- tion in the clecision. In aciclition, in all these decisions, ethical considerations limit the ways in which a risk cIassifica- tion can be used in a decision. The nature of those limits varies consiclerably across the different stages of the criminal justice system; for example, the use of risk cIas- sifications at parole has long been ac- ceptec] anc] is wiclely practiced, while such classifications at sentencing are more controversial. There is also consiclerable debate about the kinds of variables that shouIc! legitimately enter any formal risk cIassifi- cation. The most legally relevant vari- able-the current offense- is an impor- tant consideration, as is information about prior adult convictions. Other defendant characteristics that might have predictive power, but that are more controversial, include variables like prior arrests (espe- cially if not followed by convictions), or socioeconomic factors such as employ- ment status. The most controversial vari- ables are so-called ascribed characteris- tics, such as sex or race, over which the offender has no control. Previous chapters state of knowledge about criminal careers ant! the means by which the criminal justice system might try to intervene to modify those careers. In this chapter we summarized the CRIMINAL CAREERS AND CAREER CRIMINALS examine ways in which criminal career information is used, both formally and informally, in decision making and the means by which better information might be provided to facilitate those uses. We first examine the various ways in which criminal career information is combined with other considerations in decisions at each stage of criminal processing. Such information is most relevant in decisions about whether to hold a person in con- finement. We next examine methodological ap- proaches for the development of predic- tion scales. In some jurisdictions for some stages of the criminal justice system, the use of risk classifications invokes explicit prediction methods in which predictor variables are selected, weighted, and combined to generate a precliction score. This score is used to form a classification rule: individuals with prediction scores above a specified cut point are designated "high risk" and those with scores below that cut point are designated "low risk." Any such classification rule inevitably re- sults in some errors among predicted high risks who are not (false positives) and predicted Tow risks who are not (false negatives). Use of prediction in criminal justice decision making raises some ethi- cal questions: the weight to give to pre- dicted future offending, "unacceptable" predictor variables, and the selection of the cut point, which dictates the mix of false positive and false negative errors. Although the choices of weights, vari- ables, ant! cut points must be macle lo- cally, we attempt to identify the critical issues to be considered in making those choices. Finally, since all prediction scales are dependent on the quality ant! complete- ness of the data available, we examine some issues in the use of individual rec- ords, and particularly juvenile recorcls,

USE OF CRIMINAL CAREER lNFORMATlON lN DECISION MAKING which might contribute appreciably to improved classification. CRIMINAL CAREER PERSPECTIVES IN CRIMINAL JUSTICE DECISION MAKING A substantial bocly of research, summa- rized in a paper prepared for the pane! (Gottfredson and Got~redson, Volume II), has been directed at establishing the determinants of each major criminal jus- tice decision: arrest, prosecution, pretrial release, sentencing, and parole.) The dis- cussion in this section examines the de- gree to which those decisions selectively target individuals with serious criminal careers, i.e., those with the highest aver- age values of offending frequency (A), especially for serious crimes, and the longest residual careers. Chanters 2, 3, ant! 4 highlight the relevance of the in- stant offense, records of the prior adult and juvenile criminal activity, drug use, and employment history as indicators of more serious criminal careers; this chap- ter emphasizes the role of those factors in criminal justice decision making.2 iWith respect to research on decisions on arrest, see, for example, Piliavin and Briar (1964), LaFave (1965), Black and Reiss (1970), Reiss (1971), and Smith and Visher (1981~; on prosecution choices, see Forst and Brosi (1977), Forst, Lucianovic, and Cox (1977), Jacoby, Mellon, and Smith (1982), Feeney (1983~; on pretrial release see Ebbeson and Konecni (1975), Bynum (1976), Bock and Frazier (1977), Roth and Wice (1978), Goldkamp (1979), Goldkamp and Gottiredson (1984~; on sentencing, see major reviews by Blumstein et al. (1983), Hagan and Bumiller (1983), Garber, Klepper, and Nagin (1983), and Klepper, Nagin, and Tierney (1983~; on parole, see major reviews by Schuessler (1954), Mannheim and Wilkins (1955), and D.M. Gottfredson, Wilkins, and Hoffman (1978~. 2The literature contains research on many other factors, ranging from psychological tests to defend- ant demeanor in the courtroom, that we did not review and are not covered in this discussion. ~- 7 157 Police Decisions Faced with the decision to arrest a suspect, police officers must weigh the evidence of"probable cause" (LaFave, 1965) along with considerations of main- taining public order and protecting other people. Information about the criminal careers of suspects is most useful in ad- vancing the latter objective. Except when an arrest occurs at the end of an investi- gation, such information has typically not been available at the time of the ar- rest decision. However, this situation is changing as telecommunications equip- ment and advanced record retrieval systems allow officers on the scene to check a suspect's wanted or warrant sta- tus. Studies of police officers' arrest deci- sions report that the dominant variable accounting for a police decision to arrest is the seriousness of the alleger] offense (see, e.g., Black, 1971; Sherman, 1980; Smith, 19821. Over factors that increase the probability of arrest are hostile behav- ior by the suspect, victims' preferences for arrest, a stranger-to-stranger relation- ship between the victim and the suspect, the officer's prior knowledge of the par- ties involved, and low socioeconomic neighborhood (see, e.g., Piliavin and Briar, 1964; Frieclrich, 1977; Smith, 19861. Some evidence suggests that in encounters involving interpersonal dis- putes, suspects who appear to be under the influence of alcohol are more likely to be arrested (Smith and Klein, 1984~; however, the effect of suspects' drug use on arrest decisions has not been exam- ined. Programs have been established to structure police decision making in order to focus additional resources on offenders with serious criminal careers. This can be accomplished by targeting prearrest in

158 vestigation on such offenders, by postar- rest investigation to assist in prosecuting them, and by placing high priority on serving warrants on them (Gay and Bow- ers, 19851. Of the three approaches, only prearrest targeting leaves the choice of targets to the police; postarrest priorities are generally establisher! by statute or by prosecutors. As one example, prearrest targeting was a major theme of the Repeat Offender Project (ROP) of the Washington, D.C., Metropolitan Police Department. A sne- cial ROP unit was established in 1982 to target and apprehend offenders believed to be committing five or more FBI index offenses per week (an annual frequency of at least 260) or to be trafficking in stolen property. Initially, a committee of senior officers selectee] ROP targets without any explicit rule-on the basis of informa- tion developed through informants and investigations, as well as available police records; other offenders became ROP tar- gets because of outstanding warrants or arrest opportunities that arose during in- vestigations of other targets. This selec- tion process led to the targeting of of- fenders who were older than average and had longer-than-average records of previ- ous adult inclex arrests. Subsequently, the ROP unit adopted an explicit rule that gave greatest weight to verified infor- mants' information, but also incorporated variables related to serious criminal ca- reers, such as "criminal history" (not cle- finecl further), narcotics addiction, pend- ing cases, and status as a parolee, probationer, or unemployed! person (Mar- tin, 19841. This approach exemplifies a philosophy expressed in other locations using prearrest targeting, that official- record indicators of the career should be supplemented by additional information on the current level of activity, which may be obtained through informants or other investigative methods (Gay ant] Bowers, 19851. CRIMINAL CAREERS AND CAREER CRIMINALS Prosecution The extent to which prosecution prac- tices successfully target offenders with serious criminal careers is related to how much prosecution efforts are focused on such offenders and whether such efforts result in higher conviction rates or longer sentences. Information on this issue de- rives largely from experience from career criminal units (CCUs), many of which were established with federal funding to give special attention to "career crimi- nals," but with each local prosecutor's office cleaning its target populations. However, other evidence is available from statistical analyses of variables asso- ciatecI with measures of prosecution ef- fort. Decision Making in the Absence of Career Criminal Units Since obtaining a conviction is a pri- mary prosecution goal, it is not surprising that the availability of evidence has been found to be the principal determinant of prosecution effort (see review in Rhocles, 1984~. Other determinants are the seri- ousness of the current charge and the relationship between the defendant and victim. The degree to which prosecution is focused on defendants with serious criminal histories is far less clear. Several stuclies have found evidence that charge reflections or other indicators of plea bar- gaining occur less frequently in cases involving defendants with extensive ar- rest histories (Bernstein et al., 1977; McDonald, 1978; lacoby, Mellon, and Smith, 1982~. Empirical results on alloca- tion of effort vary, depencling on the mea- sure of prosecution effort being analyzer! (e.g., time from arrest to (lisposition, plea bargaining indicators, attorney hours).3 sin an econometric analysis of 5,717 felony cases, Forst and Brosi (1977) found no relationship be

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING The influence of a juvenile arrest or adjudication record on adult prosecution decisions is largely unexplored. The bi- furcation of juvenile and adult records certainly inhibits prosecutors from ob- taining and using juvenile records at early decision points. Greenwood, Petersflia, and Zimring (1980) report that in a na- tional survey, 60 percent of prosecutors responded that juvenile police records were "rarely or never" avaflable at the time of filing, while a majority (74 per- cent) reported that adult records are "al- ways or usually" available. Boland and Wilson (1978) hypothesized that because of the lack of information about juvenile activity, the aclult criminal justice system makes no distinction between the first adult arrest of a chronic juvenile offender and the arrest of a true first offender. A test of the Boland-Wflson hypothesis re- quires a comparison between a jurisclic- tion with integrated juvenile and adult records and a comparable jurisdiction with more limited sharing of information. No adequate test has yet been carried out. One effort in this direction was uncler- taken by Greenwood, Abrahamse, and Zimring (1984) in comparing prosecutors' tween prior record and time from arrest to disposi- tion, which was the authors' proxy for effort. In part, this lack of relationship is appropriate because arrest history is included as a factor in their strength-of- evidence measure, which was found to positively influence prosecutor efforts. On the basis of an experimental study of 855 prosecutors, Jacoby, Mellon, and Smith (1982) report that defendants' criminality (as measured by prior convictions, ar- rests, and parole or wanted status) was related to the priority assigned to a case, once it had been ac- cepted for prosecution. Hausner, Mullin, and Moorer (1982), analyzing a large sample of cases presented to U.S. attorneys, found that cases involv- ing repeat offenders were more likely to be ac- cepted than cases involving first offenders. But the analysis did not control for strength of evidence, and the acceptance decision patterns may not apply outside the federal system because the declined cases were also eligible for prosecution in state courts. 159 decisions in Seattle, which has an inte- gratec] record! system, with those in Los Angeles and Las Vegas, which do not; however, there were not enough relevant cases in Seattle for satisfactory analysis. In that study, Greenwood, Abrahamse, and Zimring (1984) argue that the avail- abflity of the juvenile record! cloes not guarantee that its severity will influence case disposition. There have been few studies of the influences on prosecution efforts of other variables relater! to serious criminal ca- reers, such as drug use and employment status, and the evidence is not consis- tent.4 Career Criminal Unit Evaluations Career Criminal Units were estab- lished in more than 100 prosecutors' of- fices during the 1970s. Their caseloads were generally restricted by case selec- tion criteria that were left to the discre- tion of the local jurisdictions. One evalu- ation found that CCU cases received more attention: CCU convictions were found to consume between five and seven times as many attorney hours as other convictions; smaller but still sizable differentials in attorney hours were also observed for pleas and dismissals (Rhodes, 19801. Two CCU evaluations (Springer and Phillips, no..; Chelimsky and Dahmann, 1981) provide information on whether CCUs selected cases involv- ing defendants with the most serious criminal careers and on whether the extra resources led to higher conviction rates or 40ne analysis of prosecution decision making in 1,196 burglary cases found the probability of case acceptance to be positively related to the existence of a prior record, but for defendants with prior records, a prior drug-related charge made prosecu- tion less likely in the instant case. However, be- cause the cases arose from a special burglary pro- gram, these conclusions may not generalize to other settings (see Rhodes, 19841.

160 longer sentences for the targeted defen- dants. The four CCUs evaluated by Chelim- sky and Dahmann showed marked varia- tion in formal selection criteria. Defini- tions of"career criminals" differed in terms of current charge thresholds (e.g., robberies only, felonies only, or any charge); attention to prior arrests and con- victions; and "status indicators," such as pretrial or parole release following an earlier case. While the evaluators claimed that it was impossible to say "with any certainty how closely the group of indi- viduals prosecuted by these programs represented the ideal career criminal group" (1981:72), the selection criteria used by the four CCUs were in fact ~en- erally consistent with indicators of seri- ous criminal careers, although the selec- tion might have been improved by also considering drug involvement. Evaluating the impact of CCUs on case processing is more problematic. On the basis of their reviews of 15 evaluations, Springer and Phillips (n.d.) conclude that prosecution by CCUs was more success- ful than routine prosecution in terms of pretrial detention rates, conviction rates, dismissal rates, and incarceration rates. In contrast, Chelimsky and Dahmann (1981) found no effect of CCUs on these rates, but concurred with the Springer and Phil- lips finding that the units had increased the seriousness of the charges for which convictions were obtained. The differ- ences in conclusions may have arisen because only one of the four units evalu- ated by Chelimsky and Dahmann was in a large urban office. Had they evaluated other jurisdictions whose regular units were under a greater press of caseloads, Chelimsky and Dahmann might have found more substantial effects of CCUs on case outcomes. Results of CCU evaluations should be treated carefully because of two common methodological problems in the research CRIMINAL CAREERS AND CAREER CRIMINALS to date. First, pretest and pastiest periods were typically of only 1 year, so that evaluations were based on only the earli- est cases processed by the units, which may not have been representative of later performance. Second, the comparison group designs used in CCU evaluations cannot completely control for bias in case selection. In maintaining CCU caseloads at desired levels, the formal selection rules were frequently tightened or re- laxed through unwritten informal revi- sions. These shifts may have invoked strength of evidence, perhaps by choos- ing important but weak marginal cases for which extra prosecutorial attention was likely to affect case outcome (thereby bi- asing evaluation results against the unit) or by skimming the best cases to improve the unit's success rate. In either case, informal practices could have distorted the evaluations of CCU effectiveness. Thus, the evidence is still ambiguous on the question of whether resource alloca- tion has a measurable effect on case out- come rates: CCUs do appear to target high-rate serious offenders, but their ef- fect on case outcomes is still unproven. Pretrial Release While the primary consideration in pre- trial release decisions is ensuring appear- ance for trial, community safety during the period until trial is also often an im- portant concern. Indeed, "preventive de- tention" statutes in some jurisdictions permit a judge to focus explicitly on the risk to the community in detaining a sus- pect prior to trial. No assessments are available of the effects of preventive de- tention laws in routine use. (See, how- ever, Bases and MacDonald, 1972, for an early evaluation, and Gottlieb, 1985, for a preliminary report of an evaluation in progress.) Even in the absence of preven- tive detention laws, however, judges can and do consider current criminal justice

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING status, prior record, and other factors re- lated to criminal careers in their pretrial ~ · ~ cleclslons. Analyses of the pretrial release deci- sion have generally considered the role of seriousness of the current charge, indica- tors ofties to the community, and eviden- tiary strength. Charge seriousness domi- nates the decision, measured in terms of either release of the defendant on per- sonal recognizance or the amount of fi- nancial bond demanded (Gottfredson and Gottiredson, Volume II; Rhodes, 19841. The influence of charge seriousness re- flects joint concern over the incentive for nonappearance at a trial that might result in a serious sentence as well as concern about harm to the community in the event of repetition of the crime during the pre- trial period. Community-ties indicators (e.g., local residence of family) are in- tended to predict appearance at trial; however, they have not been consistently found to influence the decision to release a defendant on personal recognizance, and they seem not to affect the amount of bait or conditions of release in serious cases when the defendant is not released on recognizance (Goldkamp and Gott- fredson, 1981a,b). Adult criminal history emerges as com- parable in importance to seriousness of the current charge in determining pretrial release decisions (M.R. Gottiredson, 1974; Bynum, 1976; Roth anc] Wice, 1978; Goldkamp, 1979; Goldkamp and Gottfred- son, 1981a, b; Nagel, 1983; Stryker, Nagel, and Hagan, 19831. Two related measures have been found significant: indicators of the overall adult record (measured in such teas as total prior arrests or number of felony convictions in the preceding 5 years) and indicators of current criminal justice status (e.g., pend- ing cases at the time of arrest). Two other variables related to serious criminal careers~rug use and employ- ment status have been studied. For fel ]6] ony cases in urban courts, indicators of drug use have been found not to influ- ence either the decision to release on recognizance or the bond amount set for defendants not released immediately (Roth and Wice, 1978; Goldkamp and Gottfredson, 19841. The most recent study of the federal system also found no evidence that drug use influences pretrial release decisions (Rhodes, 19851. In at least two jurisdictions, urinalysis tests for current drug involvement are made at the time of arrest. However, analyses are not yet available ofthe relationship of the test results to pretrial release decisions and to clefenclants' pretrial conduct. Employ- ment status, however, has been consis- tently found to influence decisions on pretrial release on recognizance. Em- ployecI clefendants are more likely than unemployed defendants to be released on personal recognizance, controlling for charge seriousness and prior record (see, e.g., Roth and Wice, 1978; GolUkamp and Gottfrecison, 1981a). This finding may re- flect both a recognition of their commu- nity ties as well as a reluctance to impose an extralegal penalty (namely, loss of job ant] income) before trial. Sentencing Determinants of sentencing were re- cently reviewed by a pane} of the Na- tional Research Council, which con- cluded that "using a variety of indicators, offense seriousness and offender's prior record have emerges] as the key determi- nants of sentences" (Blumstein et al., 1983:83~. Despite measurement prob- lems that tend to cause underestimates of its effect, prior record consistently emerges as one of the strongest effects on sentence, second only to the current charge (Blumstein et al., 1983:83~7; Got~recison and Gottfredson Volume II). A review prepared for this panel (Rhocles, 1984) i(lentified three studies,

162 conducted in support of sentencing guidelines development, that have also found drug use to be an aggravating factor at sentencing for certain charges: sex crimes, fraud, weapons charges, and forg- ery. These offenses are neither so serious that incarceration is nearly automatic nor so minor that it would be inappropriate. With such intermediate offenses, there is more room for the influence of factors such as drug use that are not related to the current charge and prior record. Sentencing is the decision point at which the potential effect of juvenile record availability receives the greatest attention, and analyses have provided some fragmentary evidence suggesting that routine availability of juvenile rec- ords might increase the sentences of young adult offenders with extensive ju- venile records. Greenwood, Abrahamse, and Zimring (1984) compared sentencing practices for young adults in three juris- dictions: Las Vegas, with no disclosure of juvenile records before the presentence investigation; Los Angeles, with frequent pretrial information sharing between the two systems; and Seattle, with one of the few integrated repositories containing both juvenile and adult records. While their study was limited, the researchers found that, among young adult defen- dants, those with more extensive juvenile records were more likely than others to be sent to a state prison in all three sites (1984:531. They attribute the finding in part to informal record sharing, and they also point out that even in the absence of information, a sentencing judge may pre- sume the existence of a median juvenile record, in which case presentation of a minimal record would be treated as a mitigating factor. Uncertainty concerning the influence of juvenile records on sentencing in the adult system makes clear the need for more careful comparisons of individual case decisions in a "two-track" system (in CRIMINAL CAREERS AND CAREER CRIMINALS which juvenile records are shielded from adult authorities) with decisions in a sim- ilar "one-track" system with shared infor- mation. The most comprehensive analy- sis of decision making in a one-track system was based on data from London, where juvenile court information is shared with adult authorities. That study concluded Mat the number of juvenile convictions does influence sentence se- verity at the first adult conviction (Langan and Farrington, 1983~. While the relation- ship between expected incarceration time and age is confounded by the dif- ferent institutional options available at different ages, the probability that a con- viction leads to incarceration increases monotonically with the number of prior convictions over the entire age range of 10 to 24. Replication and comparison of this analysis in comparable American set- tings for example, Seattle and another jurisdiction in the state of Washington with separate juvenile and adult reposito- ries would be an important step in as- certaining the effect of a single repository for both juvenile and adult records in informing the sentencing decision. Parole Release In their release decisions, parole au- thorities consider retributive concerns, support for maintaining institutional dis- cipline, and facilitating inmate reentry into society. The parole decision also in- vokes prediction of future offenses, fre- quently by means of formal guidelines (HofEnan and Beck, 1974; HofEnan, 19831. Analyses of parole decisions have gen- erally measured the decision outcome ei- ther in terms of time served in prison or a binary release/no release outcome at the time of first eligibility for parole. Neither of these measures is entirely satisfactory. "Time served in prison" confounds the parole decision with the sentence, espe

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING cially the minimum sentence; the binary indicator fails to reflect the severity of punishment imposed. Furthermore, nei- ther indicator captures aspects like the term and special conditions of parole. Despite the difficulty of analyzing pa- role decisions, two variables emerge as dominant influences: the charge that led to the incarceration and the adult criminal history (Scott, 1974; D.M. GottEredson et al., 1978; Elion and Megargee, 1979~. Not surprisingly, these variables are also the primary bases of parole guidelines used to structure decisions in the federal sys- tem and in many states. Other indicators of serious criminal ca- reers also play a role in parole decision making. Surveys of board members (Parker, 1972) and analyses of cases (Elton and Megargee, 1979) suggest that knowledge of juvenile offenses ant! other life events incorporated in presentence reports play an influential although not 1 · · 1 ~ 1 1 1 1 1 163 Conclusions Criminal justice decisions arrest, pre trial release, prosecution, sentencing, and paroIc are macie with different ranges of discretion, different objectives, and dif ferent information available at the time of the decision. Despite the differences in objectives, each decision includes some assessment of whether the subject is pur suing a serious criminal career ant] hence is a candidate for confinement. A key question is whether these assessments do in fact leacI to more severe sanctions for offenders whose careers involve the most frequent and serious offending over the longest durations. A review of relevant research shows that seriousness of the current charge and prior adult record two factors closely re lated to serious criminal careers~omi nate the decision-making process. Em ployment status only influences decisions Decisive role. mono use nas Seen con- regarding pretrial release. The impact of a sistently found to be a predictor, although juvenile criminal history on decision a weak one, of parole denials (see making has been thoroughly explored GottEredson and Gottfredson, Volume II). only in sentencing; even though items in A history of drug abuse is considered a -' ~ ' ' ~ ' negative indicator at parole hearings, ac cording to both a survey of parole board members (Parker, 1972) and a content analysis of parole board hearings (Carroll and Payne, 1977a, b), but it rarely emerges as a dominant influence on the decision to release (e.g., Scott, 1974; D.M. Gottfrecison et al., 19781. Misconduct during incarceration in- formation not available at sentencing emerges as a strong factor affecting parole decisions (see Gottfredson and GottEred- son, Volume II). Its influence on parole decisions presumably reflects its value in predicting recidivism as well as its moti- vational value in maintaining institu- tional discipline (Glueck and Glueck, 1930; Mannheim and Wilkins, 1955; CarIson, 1973; GottEredson and Aclams, 19801. the juvenile history are related to adult criminal behavior, legal restrictions ant] the inaccessibility of juvenile history in- formation have minimized its use in most criminal justice decisions. Drug use, es- pecially concurrent use of narcotics and other drugs, which is strongly associated with frequent and serious criminal activ- ity, has not been found to be an important determinant of criminal justice decisions. ISSUES IN PREDICTION-BASED CLASSIFICATION Introduction While a variety of considerations are weighed at each stage of criminal justice decision making, the prediction of future offending is a consideration common to all stages. Most often these predictions

164 i.e., based solely on the are "clinical," decision maker's judgment and previous experience. For more than 60 years, ef- forts have been made to improve the quality of these predictions and to sys- tematize their use by means of statisti- cally derived, precliction-basec! cIassifica- tion rules. Such empirically based rules require clevelopment of scales that relate prediction variables to estimates of future offending and transform every offender's scale score into a risk classification level, such as high risk or Tow risk. The risk classification of each offender is then con- sidered as one factor in making individual decisions. Prediction-based classification rules have been developed for use in pretrial release (GoIc~kamp and Gottfred- son, 1981a), in assignment to a career criminal unit for prosecution (Rhodes et al., 1982), at sentencing (Greenwood, 1982), for institutional classification (D.M. Gottfrecison and Bonds, 1961), and for parole release (Hart, 1923; Mannheim and Wilkins, 1955; Hoffman and Beck, 19741. General arguments for the use of such rules have included enhancing the visi- bility of the decision process, improving the equity of decisions, and providing routine feedback on the outcomes of de- cisions. But perhaps most important, it has been argued that using such rules informs decision makers of systematic patterns occurring beyond their own ex- perience and so can improve the accuracy of their predictions. Interest in predic- tion-based rules has recently been in- creased by attention to selective incapac- itation policies, especially by claims that selective incapacitation can reduce crime, prison populations, or both. Increased pub! c concern about crime, coupled with the pressure of growing prison populations, has intensified the search for more efficient ways to use lim- itec! criminal justice resources. New lim- itations on the discretion of parole boards , ,, CRIMINAL CAREERS AND CAREER CRIMINALS have encouraged the consideration of fu- ture risk at earlier decision points, espe- cially at sentencing. Finally, newly de- veloped prediction scales have attracted attention as candidates for incorporation in prediction-based classification rules. In order to address the technical ant! ethical issues in precliction-based cIassi- fication rules, it is important to first out- line the process by which they are devel- oped. Three basic steps are involved in deriving a scale that assesses the likely future course of an inclividuaT's criminal career: first, assessing the ethical limita- tions on the role of prediction in decision making and on the choice of the predic- tors; second, combining appropriate risk factors to form a prediction scale and using that scale to define a classification rule; and third, validating the perform- ance of the rule in terms of its cIassifica tion errors. Many jurisdictions may not have the necessary data or are otherwise unable to clevelop their own prediction scale and accompanying classification rule, and so have consiclered the transfer of an exist- ing prediction-based classification rule for their own jurisdiction. However, transferring such a rule outside the juris- cliction in which it was developed creates accuracy problems that must be consid- ered carefully. These problems in the transfer of classification rules are clis- cussed at the end of this section. Ethical Issues Opinions differ on the proper role in criminal justice decision making of pre- clictions of future criminality. In one view, because a convicted offender is aIreacly vulnerable to state intervention within statutory limits, use of predictive criteria following conviction, especially in deciding whether to release on parole, is widely accepted by many people. From another view, because of the symbolic

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING import and strong just-deserts component of sentences, many people find the use of prediction at sentencing to be acceptable only within narrow limits, if at all. Some people would permit the use of preclic- tion only in certain preconviction deci- sions, such as selection for pretrial release on personal recognizance or, possibly, for intensive prosecution. Predictive considerations have long played an implicit though informal role in criminal justice decision making. Explicit prediction-based classification rules have been widely used for several decades as a basis for structuring decisions framed in terms of reducing punishment, such as parole release and pretrial release; few ethical objections have been raised against those uses. More recently, explicit rules have been advocated and used as a basis for imposing punishment and in- creasing the risk of punishment: at sen- tencing (Greenwood, 1982), in selection for intensive prosecution (Forst et al., 1982), and in targeting for proactive po- lice investigation (see Martin, 1984; Moore et al., 1984, for discussions). In addition to their presumed advan- tage in effectively targeting crime control resources on worst-risk offenders, explicit rules can make the exercise of discretion more precise and fair. By formalizing the decision criteria, rule-bound decisions ar- ticulate the bases for discretionary deci- sions and make those decisions more pre- dictable. At the same time, when rules serve only as guiclelines, the discretion to diverge from them is retained, but usually carries some burden for providing rea 165 offenders as high risk is acknowledged as a greater concern as the sanction becomes more severe and as the preclicted crimi nal behavior becomes less serious If. Monahan, 1981; Morris ant! Miller, 1985~. For example, imposing life sentences un der the predictive presumption of habit ual offender laws is generally less accept able than briefly detaining the writers of threatening letters to the White House during a presidential public appearance. For any classification rule, the ratio of false-positive to false-negative error rates can be adjuster! by setting more or less stringent threshold scores for classifying a subject as high risk. But the ability to simultaneously reduce both rates re quires improved knowledge about corre lates of the criminal career. Informing policy makers of the estimated error rates is one essential step in enabling them to decide what role prediction-based cIassi fications should play. The choice may well be different for different crimes and sanctions, for different stages of the crim inal justice process, and in different jurisdictions. Ethical Considerations in Selecting Cantlidate Risk Factors Ethical considerations limit the choice of variables that can be used in predic tion-based classification rules. Serious ness of the instant charge appears to have nearly universal acceptance on ethical grounds and is also most useful from a crime control perspective. Charge seri ousness occupies a central ethical place sons for the divergence. in the just-cleserts theory of punishment Several ethical considerations are in- (von Hirsch, 1985:10), and its power to volved in using prediction-basec] cIassifi- enhance incapacitative effects has been cation rules in decision making. One con- clemonstratec] empirically, using data sideration is the relative importance from several jurisdictions (see Chapter 5; given to the predicted incremental dan- Cohen, 1984a). Except for charge serious ger to the public compared with the cost ness, virtually all other variables having to the defendant of the contemplated ad- predictive content have been the subject ditional sanction. Inaccurately classifying of ethical objections to their use in cIas

166 sification rules. The main concerns are addressed in terms of three general con- siderations: relationship to blameworthi- ness; logical and empirical relationship to the criterion variable; and congruence with fundamental social values. Blameworthiness. As noted by Moore (Volume II), among defendant character- istics having predictive power, those that are more blameworthy are more widely accepted as ethical bases for increasing sanctions. In assessing blameworthiness, it is useful to distinguish between harm and culpability. Harm is measured in terms of specific consequences to the vic- tim. The offender's culpability is dimin- ished to the extent that the harm occurred! through chance, negligence, diminishecl mental capacity, or ignorance, rather than through specific intent, or that the crime was motivated by victim provocation, du- ress, or a matter of conscience (von Hirsch, 1985:64-741. In terms of blame- worthiness, these circumstances of the instant crime moderate the charge itself as a basis for decisions (1985:7~761. Other individual characteristics that are statistically associated with criminal ca- reers differ in terms of ethical acceptabil- ity, in part because they involve different levels of blameworthiness. Since chil- dren are considered less responsible than adults, for example, a given juvenile record is widely considered less blame- wor~y than a similar adult record. Evi- dence of prior illegal drug use may be considered less appropriate than evi- dence of previous violent or property crimes because drug use itself imposes no harm on others. Objections to the use of prior unemployment have invoked both the lack of harm and the offender's imperfect control over employment sta- tus. On grounds oftheir lack of blamewor- thiness, unalterable characteristics such as race and sex are consiclered inadmissi- ble as preclictors, and would be even if they were found to be correlates! with the frequency of serious offending. CRIMINAL CAREERS AND CAREER CRIMINALS Relationship to the Criterion Variable. In gauging the admissibility of various defendant characteristics as bases for de- cisions, the classification variable should have theoretical and empirical relation- ships to the criterion variable. Moore (Volume II) emphasizes the need for the- or.etical linkages. Historically, a strong theoretical linkage to a particular crite- rion variable has sometimes been substi- t-uted for empirical evidence in justifying a particular variable as a basis for deci- sions: for example, the theoretical con- nection between indicators of community ties and appearance at trial led to their use in pretrial release guidelines long before empirical evidence concerning He relationship was developed (Ares, Rankin, and Sturz, 19631. Measures of an offender's prior crimi- nal career are widely considered appro- priate predictors of future behavior on theoretical and empirical grounds. Among available measures, an arrest record samples a larger fraction of inci- dents in a criminal career than does a conviction record; hence, it is preferable on grounds of empirical standing, but is less preferable in terms of the value of its proven certainty. Both drug use and em- ployment status have multiple theoretical relationships to frequency of offending, but there is stronger empirical evidence of a connection to drug use than to em- ployment. Concern for Equal Protection. It is widely agreed that classification rules should not affront basic social values, such as concern for equal protection of all citizens under the law. Historically, equal protection concerns have been especially acute win respect to personal character- istics such as race or religion. There is no empirical evidence that these variables have important predictive power with re- spect to criterion variables of concern to criminal justice decision makers. But even if predictive power were shown, these variables would be widely viewed

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING as totally inappropriate bases for offender classifications. While clisapproval of such variables is virtually universal, there are responsible differences of opinion with respect to the legitimacy of their correlates. Moore et al. (1984:73) argue that "interest in justice counsels restraint in the use of . . . vari- ables associated with class and ethnic background, inclucling wealth, religion, race, and national origin;" In contrast, while Morris and Miller (1985:47) ad- monish against use of"race or any super- ficial substitute" in making preclictions, they would permit incorporation of"fa- cially neutral factors such as education, housing, ant] employment which reflect racial inequalities." However, even though the races differ appreciably on these factors in the general population, it is not clear that these race differences prevail among active offenders. In sum, assessing any particular cIassi- fication variable requires careful balanc- ing of all three values discusser] above, along with the variable's contribution to crime control. The choices of admissible predictor variables may thus be different across jurisdictions using prediction- basec! classification rules. However, with respect to race and ethnicity, it is worth reiterating their general inappropriate- ness as predictor variables because they have no relationship to blameworthiness, there is no demonstrated theoretical or empirical basis for linking them to perti- nent criterion variables, and their use affronts concern for equal protection. Developing Classification Rules Defining the Criterion Variable Predictions in criminal justice cleci- sions are usually concerned with avoid- ing some undesired behavior, such as crimes, parole violations, or failures to appear at trial. The criterion variable is the specific measure of unclesired behav- ior, and may inclucle, for example, crimes ]67 of all types, only violent crimes, or only crimes of specific types such as robbery or burglary. Research on prediction has tradition- ally relies! on dichotomous criterion vari- ables measuring recidivism i.e., rear- rest, reconviction, or parole revocation within some follow-up period. The fol- Tow-up periods are sometimes uniform in length (usually between 1 and 5 years) for all sample members and sometimes vary among the inclividuals being followed (e.g., the length of each inclividual's pre- trial or parole periocI). Comparisons across subsamples are based on the frac- tion of each group that recidivates before the end of the observation period. For scale clevelopment, there are three important shortcomings of dichotomous recidivism indicators: they are not suffi- ciently sensitive to variations in the un- derlying continuous variable, A; they fait to distinguish acloquately between of- fenders whose careers have terminated and those who remain active but avoid arrest; and they measure individual crime rate as observed through an official filter, such as arrest or parole revocation, and so incorporate factors that may be associates] with the official decision rather than with the offender's behavior. Because of the insensitivity of dichoto- mous recidivism indicators, preliminary conclusions about recidivism probabili- ties based on short follow-up periods have sometimes been substantially re- vised in light of experience during a longer follow-up period. Maltz ( 1984:73) presents an example of a recidivism rate differential between two subsarnples that was large and statistically significant at 6 months, had nearly disappeared by 24 months, and appeared likely to be re- versec! by 36 months. Such an outcome would occur if one sample contained of- fenclers with a higher average A but a shorter average resiclual criminal career length than Me other: a relatively high proportion of active offenders in the first sample wouIc! be rearrested shortly after

168 release, but a higher proportion would also drop out of offending before ever being rearrested. These differences between samples would be more adequately reflected by the use of continuous criterion variables that more reliably capture criminal activ- ity over time. For samples of released offenders, one such continuous variable is elapsed time (t) between release and the first recidivism event as measured in official records. In any given sample of released offenders followed for some ob- servation period, some will continue criminal activity (and may eventually be rearrested) while others will cease crimi- nal activity during the observation pe- riod. In measuring recidivism and devel- oping predictive scales for classifying offenders, it is important to distinguish the two groups of released offenders. The criminal justice system is primarily con- cerned with offenders who remain active at high offending frequencies. Various failure-rate models have been developed that use the statistical distribution of t to estimate both ,u, the annual arrest rate for active offenders, and 6, the dropout rate, reflected in the fraction of the sample whose criminal careers terminate before the end of the observation period (e.g., Maltz and McCleary, 1977; Harris and Moitra, 1978; Bloom, 19791. The earliest failure-rate models as- sumed that all released offenders re- mained active in their criminal careers and estimated their assumed constant ar- rest rate, ,u (Stollmack and Harris, 19741. By ignoring the termination of some crim- inal careers, this assumption thus led to underestimates of the true value of ,u for those who did remain active. Later refine- ments to the basic model permitted joint estimation of the values of ~ and ,u (Maltz and McCleary, 1977) and of different val- ues of ~ and ,u for different subsets of a sample (Carr-Hill and Carr-Hill, 1972; Greenberg, 1978; Harris, Kaylan, and CRIMINAL CAREERS AND CAREER CRIMINALS Maltz, 19811. By permitting heterogene- ity in ,u and 6, these models could parti- tion observed changes in estimates off during the follow-up period into effects due to early attrition of high-rate offend- ers (who are quickly rearrested and re- moved from the samples at risk soon after release) and dilution effects due to of- fenders whose careers have terminated, who constitute an increasing fraction of the remaining sample over time. Another approach to defining the crite- rion variable, used by the Rand Corpora- tion with inmate samples (Chaiken and Chaiken, 1982a), invokes self-reported values of A rather than the recidivism measures more commonly used. This ap- proach has the advantage of accounting for all the crimes by active offenders rather than only those filtered through the arrest or other criminal justice decisions leading to designation of a recidivist event. This reliance on crimes committed reflects differential rates of actual offend- ing and thus avoids distortions associated with differences in vulnerability to cap- ture by the criminal justice system. Since offenders' self-reports are retrospective, however, the measure cannot capture postrelease differences in career termina- tion pattems, which can be reflected in failure-rate analyses. Selecting and Weighting Variables The process of selecting and weighting variables to form a prediction scale in- volves a series of steps. The initial task is to select some set of candidate predic- tors variables on which data are avail- able and for which a connection to crim- inal careers might be expected on the basis of theory or empirical research. Among all the candidate predictors that could be used in the scale, some will be rejected because of the poor quality of available data or their redundancy with other candidate variables. From the re

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING maining variables, others may be elimi- natec! through exploratory statistical anal- yses that reveal them to have little predictive power. Once a set of canclidate predictors has been selected, weights are assigned. This section describes the ma- jor features of techniques that have been used to develop weights. Copas and TarI- ing (Volume II) and Farrington ant! TarI- ing (1985) contain more complete cliscus- sions. Burgess Scores. The Burgess score method was one of the first classification techniques applied in criminal justice (Burgess, 1928; OhTin, 19511. To use it, the predictor variables must all be coder! dichotomously (e.g., present = 1, absent = 0) and defined so that presence of a factor is statistically associated with the high-risk category. Thus, this technique may discard useful information that can- not be captures! in a simple yes/no vari- able (e.g., frequency of drug use). The analyst incorporates as predictors those variables that most effectively cliscrimi- nate between high-risk ant! low-risk members of the sample. An individuals Burgess score, which is the number of predictors present, is used to classify the inclividual. Burgess scores place equal weight on each predictor variable in the constructed scale, a feature that is appro- priate only if each predictor conveys the same amount of information about the criterion variable and if all the predictors are mutually independent (Copes and TarTing, Volume II). In early applications, the Burgess score was user! to preclict a dichotomous criterion variable, parole vi- olation. One can also apply the technique to classify offenders on the basis of a continuous criterion variable, such as A, that has been recoiled categorically, high-, medium-, or Tow-A, as illustrates! in Greenwood (19821. The Burgess technique was originally justified more for its simplicity than for its |69 statistical properties. The subsequent emergence of more sophisticated ap- proaches to weighting the predictor vari- ables, such as multiple regression and discriminant analysis, lee! to clismissal of the Burgess scale and its variants as "in- aclequate" (Wilkins and McNaughton- Smith, 19641. Nevertheless, Burgess scores continue to be used in cIassifica- tion, primarily because their accuracy in repeated samples has usually been fount! comparable with that of scales clerivec3 using methods of multivariate statistical analysis (Ohlin and Duncan, 1949; D.M. Gottfrecison ant! BalIarcl, 1965; Warcl, 1968; Simon, 1971; Van Alstyne ant] Gottfreclson, 19781. Least-Squares Methocls. The 1950s saw the first application of least-squares estimation of a linear predictor function to criminal justice classification (Kirby, 1954; Mannheim and Wilkins, 19551. In contrast to Burgess scales, least-squares estimates place more weight on predictor variables that provide more information about the criterion variable, and they do not require that information be thrown away by dichotomizing continuous crite- rion or predictor variables. When linear prediction was recognized as inappropri- ate for analyzing the (dichotomous crite- rion variables then common in criminal justice prediction research (Palmer and CarIson, 1976), attention shiPced to alter- native multivariate methods such as cliscriminant analysis, probit analysis, ant! logistic regression analysis. A scale esti- matec3 by using least-squares weights de- scribes the construction sample more ac- curately (uncler a minimum-squared-error criterion) than floes a Burgess scale or any other weighted sum ofthe same predictor variables. However, even in construction samples, the improvements over the Bur- gess scales are small in actual practice (see discussion in Farrington and TarTing, 1985; see also D. M. Gottiredson and

170 Ballard, 1965; Ward, 1968; Challinger, 1974; GottEredson and GottEredson, 1980; Nuffield, 1982~. One problem with least-squares weights is their sensitivity to outliers in a construction sample. Because outliers are unlikely to recur in repeated samples, the accuracy of least-squares scales estimated on a construction sample containing out- liers deteriorates, or shrinks, by a greater proportion than the accuracy of Burgess scales when both are applied to other samples from the same population. The magnitude of shrinkage is proportionally greater for techniques that rely more heavily on the data and less heavily on the analyst's prior specification.5 Ongoing work with "bounded-influence" regres- sion (Krasker and Welsch, 1982; Ruppert, 1985 - which uses a weighting function to bound the influence of outliers on the regression estimates-and with "pre- shrunk" regression estimators (Copes, 1983) shows promise for reducing the shrinkage of predictive accuracy due to outliers. Special problems emerge when the correctly specified prediction equation gives a significant weight to one or more variables that have been rejected as ethi- cally inadmissible for decision making CRIMINAL CAREERS AND CAREER CRIMINALS employed offenders) is deemed inadmis- sible but is found to be correlated with A. Suppose also that unemployment is more common among offenders having long arrest records. Then estimation of a least- squares equation that includes arrest record but omits employment status would fail to produce a prediction scale that is neutral with respect to the inad- missible employment status variable. Rather, the effect of unemployment will be reflected in the least-squares weights estimated for admissible correlated pre- dictors. In this example, a long arrest record would lead, on average, to an overprediction of the effect of record lend alone on A because the weight on (see Fisher and Kadane, 19831. Suppose, for example, that employment status (coded as 1 for unemployed and 0 for 5The formulas for expected shrinkage that are reported by Copas and Tarling (Volume II) imply that statistical shrinkage increases with the number of parameters estimated in the construction sample: e.g., the number of coefficients in a multiple regres- sion equation and the number of subgroup means in an interactive analysis such as Automated Interac- tion Detection (AID) (Sonquist, Baker, and Morgan, 1973~. Moreover, when a "search" strategy is used to incorporate only the parameter estimates that are significantly different from zero, shrinkage is pro- portional to the number of parameters computed rather than to the number incorporated. _ is _ - ~,7 record length would also reflect influ- ences of its omitted correlate, employ- ment status. To develop an employment- neutral prediction scale, one would have to include employment status in the esti- mation equation, estimate least-squares weights for all predictors, and then omit the inadmissible employment-status vari- able by setting its weight to zero in com- puting values of the scale for individuals. Failure-Rate Analysis. The primary use of failure-rate analysis has been to evaluate the effectiveness of programs in reducing recidivism. But recently it has also been used to assess the validity of classifications based on a predictive scale. Schmidt and Witte (1980) and MaTtz (1984) used failure-rate analysis to esti- mate the influence of covariates of t, the time to failure, that could serve as predic- tors. In principle, by partitioning a sam- ple into categories defined in terms of candidate predictors and comparing the recidivism experiences of the subsam- ples, the techniques could be used to select the most powerful predictor vari- ables (Maltz, 1984:1311. However, a more direct approach is to

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING use failure-rate models, such as the pro- portional hazards model of Cox (1972), that make even less restrictive assump- tions about the arrest frequency rate, ,u. Barton and Turnbull (1981) have appliecI a mode! developer! by Kaplan and Meter (1958) to investigate age and education effects on elapsed time to rearrest. Fail- ure-rate moclels have been applied only rarely to criminal justice prediction prob- lems (Barton, 1978; Barton and Turnbull, 1979, 1981), and so their properties and the extra statistical power they can pro- vide in separately predicting frequency rates for active offenders and dropout rates are not yet widely unclerstood and appreciated. Defining the Classification Rule After a precliction scale has been devel- oped, it can be used to compute a scale score for every offender. This score is related to a prediction about the criterion variable for that offender. Commonly, scales are definecI so that higher scores are associated with higher risk in terms of the criterion variable. In a typical appli- cation a cut point score k is chosen, and a classification rule is clefinect that takes the form: Classify the offender as "high risk" if his scale score is equal to or greater than k, and classify the offender as "low risk" otherwise. Offenders classified as high risk are recommended for incarceration or other intensive treatment, and offenders cIassi- fied as low risk are recommended for re- lease or other less intensive treatment. Because no prediction system can be perfect, two types of classification errors occur. A false-positive error occurs when an offender classified as high risk would be a low-risk offender if given the oppor- tunity. Ethical opposition to prediction- based classification rules has tended to focus on these errors, which lead to rec 117] ommendations for"unnecessary~ ceration or other more intensive treat- ment. A false-negative error occurs when an offender classified as low risk is actu- ally a high-risk offender. False-negative errors impose a cost on society in the form of decreased community safety. Deci- sions that are based on predictions reflect explicit or implicit assumptions about the relative costs of the two types of errors. The numbers of each type of error cle- pend on both the predictive power of the scale and the choice of the cut point. For any given scale, the number of false- positive errors can be reduced by making the cut point higher, but only at the cost of more false-negative errors. Conversely, lowering the cut point will increase the number of false-positive errors and de- crease the number of false-negative errors. Simultaneously reducing errors of both types requires development of a new scale with greater predictive accuracy. The relationship between choice of cut point and classification errors is illus- tratecl in Table 6-1, which displays of- fencler counts, error rates, ant! accuracy measures associated with each possible cut point of the seven-factor scale devel- oped by Greenwood (1982) for classifying incarcerated robbers ant] burglars. The rule specified by Greenwood classified all inmates with scores of 4 or more as high risk, and thus candidates for ex- tenciecI sentences. Uncler that rule, 27.2 percent of all offenders in the construc- tion sample would be classified as high risk, a selection ratio of 27.2 percent. Under that rule, 54.7 percent of the in- mates cIassifie(1 as high risk were truly low risk (i.e., low and medium A in Greenwood's terms), a false-positive rate of 54.7 percent. The false-negative rate associated with that rule was 16.9 per- cent. A more restrictive classification rule, with the cut point raised to 6, for ,, . 1ncar

172 CRIMINAL CAREERS AND CAREER CRIMINALS TABLE 6-1 Classification Performance of Rand Scale, by Cut Point Value, for Convicted Robbers and Burglars (three states) Selection Number of Ratio for False Cut Point Ski Offenders High Positive (Class = High High- Low- Rate Rate if Score 2 k) Rate Rate False- Relative Number of Errors Negative Improvement Rate False False Over Chance (percent) (percent) (percent) Positives Negatives Total (RIOC)a 0467 100.0 75.4 0.0 668 0 668 Indeterminate 117167 91.9 73.7 5.6 601 4 605 0.750 233180 71.2 68.8 8.2 434 21 455 0.661 355122 47.2 60.8 11.5 254 54 308 0.532 44877 27.2 54.7 16.9 132 109 241 0.312 54136 13.1 47.4 20.4 55 157 212 0.371 61118 4.4 48.7 23.4 19 198 217 0.356 791 1.1 10.0 31.3 1 209 210 0.867 Total218668 - - aSee text for discussion of RIOC. SOURCE: Data from Visher (Volume II). example, would react to a Tower false- positive rate, only 48.7 percent, but a higher false-negative rate, 23.4 percent.6 Regardless of the predictive power of the scale, the minimum number of errors is constrained by the relationship be- tween the selection ratio and the base rate, the fraction of offenders who are truly high risk in terms of the criterion variable. If the selection ratio exceeds the base rate, the number of offenders cIassi- fiec! as high risk will exceed the number who are actually high risk, and so some false-positive errors must occur. Simi- larly, if the base rate exceeds the selec- tion ratio, some false-negative errors must occur. The actual error rates observed will depend on the predictive accuracy of 6In reporting classification accuracy, Greenwood (1982:59, Table 4.8) invokes a slightly different definition of high-A offenders than the one used elsewhere in his report. Under the redefinition of high A, the base rate is 28 rather than 25 percent. Values in Table 6-1 are derived from computations by Visher (Volume II), and because Visher retained the base rate of 25 percent and computed A using a different procedure from that of Greenwood, values in Table 6-1 for a cut point of 4 imply slightly different values of accuracy measures from those implied by Greenwood (1982:Table 4.8~. the scale, but they can never be smaller than the minimum number of errors that result from the relationship between the base rate an(1 the selection ratio. Only when a cut point is chosen so that the selection ratio is equal to the base rate is it possible for both error rates to be zero. The choice of cut point, however, de- pends on actual classification accuracy and on the relative concern for the dif- ferent kinds of errors. The error rate most commonly considerecl is the false-pos- itive rate, the ratio of false-positive errors to the number of positive (high-risk) cIas- sifications, but it falls to reflect the social cost of false-negative errors. Conversely, the false-negative rate, the ratio of false- negative errors to the number of negative predictions, ignores the social cost of false-positive errors. Moreover, because these measures are rates rather than counts, they fall to convey the full impact of changes in cut points on classifications. Returning to Table 6-1 as an example, a shift of the cut point from 4 to 6 produces only a moderate decrease in the rate of false-positive errors (from 54.7 percent to 48.7 percent) and a moderate increase in the rate offalse-negative errors (from 16.9

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING percent to 23.4 percent). But because the shift greatly reduces the selection ratio for high-rate offenders, it affects the error counts far more dramatically than the er- ror rates: the number of false-positive errors drops from 132 to only l9, and the number of false-negative errors increases from 109 to 198. An accuracy measure that reflects both types of errors but assigns them equal cost is the total error rate, the fraction of all classifications that are either false pos- itives or false negatives. A shift in the cut point that increases the selection ratio will increase false-positive errors anct cte- crease false-negative errors, but its effect on the total error rate cannot in general be anticipated in advance. Since the total error rate is simply the sum of the two kinds of errors, its use implies that they have equal social costs. Assignment of equal costs, however, is not appropriate in most situations. The costs depend on the severity of the intervention to be imposed on offenders classified as high risk and the gravity of the harm one is trying to prevent. For example, the U.S. Secret Service uses risk-prediction pro- files to identify potential presidential as- sassins who are followed and nerhans detainee! for a few hours' questioning dur- ing a local presidential appearance. False-positive errors are of far less con- cern in this application than they are in a selective incapacitation policy like the one analyzed by Greenwood (1982), where convicted burglars classified as high risk were to be incarcerated for 8 years compared with 1 year for other convicted burglars. Another problem with all three accu- racy measures is that they fad! to provicle an appropriate reference base. Ideally, one would like to compare the accuracy of the classification rule with the accuracy achieved by decision makers in actual practice. The latter is generally unknown because decision makers do not articulate 173 their classifications, so it is necessary to define other reference points. Accuracy achieved by any rule is sometimes mea- sured as improvement over random accu- racy, the accuracy that would be achieved if classifications were made randomly us- ing the same selection ratio as in the rule (Duncan et al., 1953~. Because random accuracy is influenced by the difference between the selection ratio and the base rate, it is also desirable to measure accu- racy achieved by the rule relative to the maximum accuracy that could be achieved under the given base rate and selection ratio. Loeber and Dishion (1983) have de- finec! a measure called relative improve- ment over chance (RIOC), which relates the accuracy achieved to maximum accu- racy and random accuracy. Their measure is defined by Achieved accuracy RIOC - random accuracy Maximum accuracy - random accuracy Formulas are provided by Copas and Tarling (Volume II) for computing maxi- mum accuracy and random accuracy. The value of RIOC is zero for a classification rule that fails to improve on random accu- racy and unity for a rule that achieves maximum accuracy given the current se- lection ratio and base rate. While RIOC does provide one more measure for assessing classification accu- racy, it, like the total error rate, gives equal weight to false-positive and false- negative errors. Moreover, for any predic- tion rule, its value is quite sensitive to the difference between the selection ratio and the base rate. As illustrated in Table 6-1, RIOC is highest at the extreme cut points of 1 and 7, and lowest at the cut point of 4, where the selection ratio is most nearly equal to the base rate. There- fore, RIOC is not a helpful measure for setting cut points. Furthermore, in using

174 RIOC to choose among alternative scales for a particular application, care must be exercised to ensure that the values of RIOC reported for the altematives are under similar base rates and selection ratios. Finally, like other accuracy mea- sures, it provides no comparison between accuracy using the rule ant] accuracy un- der existing practice. A measure of the predictive accuracy of the underlying scale that is independent of the particular classification rule is the mean cost rating (MCR): the average change in accuracy achieved at each suc- cessively higher scale score over the en- tire range of scores to 7 for the Rand scale illustrated in Table 6-1, for exam- ple.7 The value of MCR is increaser] by true-positive classifications and de- creased by false-negative classifications over the entire range of the scale. There- fore, while a high value of MCR will identify a prediction scale with generally high accuracy over a range of cIassifica- tion rules, it may not identify the preclic- tion scale that leacis to the highest cIassi- fication accuracy under any particular base rate and selection ratio. In short, no classification accuracy mea- sure or set of measures provides sufficient basis for defining the best classification rule associated with a given prediction scale that reflects all relevant concems. Rather, alternative scales and cut points have to be evaluated on the basis of social consensus concerning the relative impor- tance of false-positive and false-negative errors. This consensus will depend in 7The formula for MCR is given as follows. Let 1 K Mean Cost = 2 ~ (Ck + Ck-1)(uk - Uk-I), k= 1 where Ck is the proportion of negatives incorrectly rejected at score k, and Uk is the proportion of positives correctly predicted at score k. Then MCR = 1 - 2 x (mean cost), a transformation to obtain an index in the interval [0,1l, with higher values corre- sponding to improved performance. CRIMINAL CAREERS AND CAREER CRIMINALS part on the severity of sanctions to be imposer! on high-risk offenders and the gravity of the risk one is trying to prevent. A decision-theoretic approach to the choice of a classification rule, in which the relative cost of false-positive and false-negative errors is expressed as a "civil-libertarian ratio," is illustrated in Blumstein, Farrington, and Moitra (1985~. - Validating the Predictive Relationship Before implementing a prediction scale that has been cleveloped by using data from one construction sample, it is essen- tial to validate the scaTe's accuracy in predicting the criterion variable for other samples. A careful validation shouIc3 in- clude four elements: · a previously developed prediction scale and classification rule and measures of classification accuracy in the construc- tion sample at alternative cut points; · a validation sample Mat represents the population to which the rule will be applied, and that contains no observa- tions that were included in the construc- tion sample; · a criterion variable that adequately captures the behavior of interest; and · a follow-up period that is sufficiently Tong to observe variations in the criterion variable and to assess the accuracy of the classification rule. The following discussion illustrates the importance of these aspects of the vaTida- tion process by describing some typical violations and the problems that can re- sult from inadequate validation of the prediction scale. A common problem in validating a pre- diction-based classification rule is that the validation sample is selected to rep- resent the same population as the con- struction sample, rather than the popula- tion to which the rule will be applied. In one approach, called "sample fraction

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING 17S ation," members of a single available offenders awaiting sentence. For exam- sample are randomly assignee! to con- ple, incarcerated samples will overrepre- struction and validation subsamples. sent offenders who have committed espe- .' ~ ' ~ - ~ cially serious crimes and those with wan tins approach, accuracy Is expected to decrease only slightly in the validation sample. This shrinkage effect is due solely to sampling variation; the expected magnitude of shrinkage can be computed using formulas reported by Copas and TarTing (Volume II). Unfortunately, measures of shrinkage obtained in this way provide no informa- tion about furler deterioration in accu- racy that would be expected when the prediction scale is applied to other popu- lations whose characteristics may be clif- ferent from those in the construction sam- ple. A frequent cause of this difference is selection bias in obtaining the construc- tion and validation samples. Although all scales are intende<l for application to of- fenders at some decision point (such as sentencing), scale development is almost always based on selectecI samples for which only one of the possible decisions has been made. For example, in studies of the use of precliction-based cIassifica- tion rules at sentencing, with recidivism as the criterion variable, analyses of offi- cial records have used samples of re- leased offenders because only they are free to be rearrested. An analysis baser! on offender self-reports involved only in- carcerated offenders because they were accessible for interviews (Greenwood, 19821. Because the factors that influence incarceration and release decisions are also related to the behavior being pre- dieted, this selective sampling creates the potential for selection bias, which must be addressed in the development of pre- diction-basec3 classification rules. Because of selection bias, both the pre- dictor weights and the classification accu- racy measures developed in a construc- tion sample made up entirely of releasees or entirely of incarcerated inmates will be inappropriate for samples of convicted especially long arrest records. The few inmates with short arrest records are likely to have committee! more serious crimes. In such a sample, an analysis that failed to take adequate account of selec- tion bias could produce a scale for pre- clicting serious offending that mistakenly assigns a negative weight to the length of an arrest record. The best way to avoid! distortions aris- ing from selection bias is to use samples that adequately represent the population to which a scale will be applied. In the case of prediction scales for use at sen- tencing, for example, an appropriate sam- ple would include all those who are sentenced both incarcerated and unin- carceratec! offenders. The absence of eas- ily accessible and uniform data on all sentencee! offenders which are now scattered among local courts, probation offices, jails, state correction facilities, and parole offices has greatly inhibited anal- ysis of appropriate samples. If selected samples must be used, a variety of statistical techniques are avail- able to estimate and reduce the impact of selection bias. If data exist to permit anal- ysis of both the release decision ant! the subsequent behavior of a sample of releasees, including all variables relevant to both outcomes, an "enclogenous selec- tion" model (Heckman, 1979) can be used to correct for selection bias in clevel- oping predictor weights. The estimated weights then provide unbiased predic- tions of the behavior of offenders about whom the release decision is to be made, rather than only of releaser! offenders. Heckman's technique has been applied in conjunction with logistic regression analysis to find predictors of the pretrial release decision and the probability of pretrial rearrest (Rhodes, 1985~. Even

176 though the Rhodes analysis incorporated more than 40 candidate predictors of the release decision and of pretrial rearrest, it is still possible that selection bias arising from unobserved factors (e.g., subjective decisions based on juclges' prior experi- ence or unusual family circumstances of a defendant) might persist. Because Burgess scales are less depen- dent on the construction data than scales developed with more sophisticated meth- ods, the accuracy of classification scales based on Burgess methods is relatively unaffected by selection bias. Conse- quently, even though the more sophisti- cated approaches outperform Burgess scales in construction samples, Burgess scales incorporating variables with widely acknowledged predictive validity continue to be developed ant! to perform reasonably well in a variety of samples. As more is learned about characteristics that differentiate convictees who are re- leased from those who are incarcerated, data on those characteristics can be col- lected and included in the analysis. The application of moclels such as Heckman's to data incorporating those characteristics will lead to improved predictive scales cleveloped from selected samples. Even a correctly developed cIassifica- tion rule can be expected to change over time. Introduction of a scale may change behavior patterns. For example, offenders may avoid seeking drug treatment to pre- vent creation of a record of drug abuse that would increase their scale score, or prosecutors aware of the weight on prior convictions may insist on retaining multi- ple charges in plea bargaining. Such changes will cause deterioration of pre- dictive accuracy. In addition, the behav- iors of offenders and criminal justice au- thorities may change over time for reasons independent of the scale, for ex- ample, as the age composition of offend- ers varies or the resource constraints of criminal justice agencies are altered. CRIMINAL CAREERS AND CAREER CRIMINALS For both reasons, periodic validation and recalibration of the prediction-based rule are integral parts of the implementa- tion process. In validation, the accuracy of an existing prediction-based cIassifica- tion rule is measured in an entirely new sample. The deterioration of accuracy may become sufficiently large to warrant recatibration~reestimation of predictor variable weights and reconsideration of the choice of cut point. The accuracy of the recalibrated rule should then be ver- ified by using another independent sam- ple. The distinction between these two steps is not always maintained. For exam- ple, early "validations" of the 1980 Iowa Risk Assessment System involved recali- bration through the incorporation of new "special risk factors" found to have pre- dictive power in the validation sample. Then, the combined construction and val- idation samples were used to compute accuracy measures (Statistical Analysis Center, 19841. Because of this noninde- pendent approach, the reported accuracy measures from the validation were over- stated. Issues in Transferring Classification Rules The cost and difficulty of developing a prediction-based classification rule have led many jurisdictions to consider instead transfer of an existing rule. However, there are a number of reasons for pro- ceeding very cautiously with such trans- fers and for expecting the classification accuracy of a rule to deteriorate signifi- cantly when it is applied in a new juris- diction. Even if offenders with similar characteristics have similar criminal ca- reers in different jurisdictions, differ- ences in criminal justice system behav- ior-as measured by the decisions of police, prosecutors, and judges may lead to crossdurisdictional differences in the behavior of offenders at any given

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING stage in the process. For example, as discussed above, selection bias within a single jurisdiction will lead to differences between samples of released and incar- cerated offenders. And because of cross- jurisdictional differences in the decision processes that lead to clifferent types of offenders reaching different stages of the criminal justice process, samples of incar- cerated offenders may differ across juris- clictions even if offenders on the street do not. The impact of crossdurisdictional ef- fects on estimated relationships involving predictor variables and criminal behavior was documenter! in Visher's (Volume II) reanalysis ofthe second Rand inmate sur- vey, which sampled the same offender types- incarcerated robbers and bur- glars in California, Michigan, ant! Texas. All inmate responses from the three states were combined in the con- struction of Greenwooc3's seven-factor scale for selective incapacitation. But as reported by Visher (Volume II), the mean offending frequency (A) for preclicted high-rate offenders in the Texas sample was only about one-third that of the Cal- ifornia and Michigan samples. It is possi- ble that this difference arose because Texas offenders behave differently from those in the other states. But it is more likely that Texas criminal justice policies lead to incarceration of a larger share of less serious offenders because of differen- tial selection effects. These differences lead one to expect cross-state variation in the observed predictive relationships, and indeed Visher (Volume II:Table 15) reports differences across states in the factors that show significant predictive power and in the overall goodness-of-fit between scale items and values of A. In a comparison of felony case process- ing in six large urban jurisdictions, case attrition rates (through nonacceptance at screening, nolle prosequi, or clismissal) ranged from 40 to 76 percent (Brosi, I77 1979:141. Such variations limit the trans- fer ability of classification scales across jurisdictions anct mandate at least local validation and recalibration of any such scale on a sample selected from the new jurisdiction. A further potential impediment to suc- cessful transfer of a predictive scale arises from crossjurisdictional variations in the meaning, completeness, and quality of data on both the predictor and criterion variables. Such variations can arise from definitional differences (e.g., a purse snatcher may be charged with robbery or larceny); from differences in the accuracy with which predictor variables are veri- fied (for example, employment status as cletermined during a presentence investi- gation is likely to be more accurate than employment status determined immedi- ately after arrest); or from crossjuris- dictional differences in definitions of predictor variables (such as the character- ization of previous charges anc! disposi- tions as felonies or misdemeanors). In considering the transfer of a predic- tive scale, it is important to recognize that the accuracy of a scale developed to pre- dict one type of behavior tends to deteri- orate when it is used to predict some other behavior. Although this tendency is clear in principle, it is frequently over- looked in practice. For example, Fisch- er's (1984) accuracy comparison of several scales on a common data base ranks them in order of their ability to predict the criterion variable for the Iowa Risk As- sessment System instrument a serious- ness-weighted index of a broad range of crime types. While the ranking presented there reflects accuracy in predicting the Iowa criterion variable, it should not be assumed to apply with respect to other criterion variables, such as likelihood of parole violation or rearrest for a specific crime. The principle would apply, for example, to jurisdictions attempting to implement Greenwood's (1982) seven

178 factor scale to predict rearrest for any of a broad range of crimes, despite the fact that it was developed to predict indivicI- ual crime rates for robbery and burglary for incarcerated offenders. For all these reasons, comparative anal- ysis of differences in case attrition and in variable definition shouIc] prececle trans- fer of a statistical precliction crevice across jurisdictions. Evidence of substantial crossdurisdictional variation warrants caution in transferring any predictive de- vice, even if the jurisdictions are other · . ~ wise slm1 ar. EXPLICIT CLASSIFICATION SCALES Recognition of empirical relationships between frequency, duration, and seri- ousness of offending, on one hand, and observable offencler characteristics on the other, has led to development of ex- plicit classification scales based on these relationships for use in criminal justice decisions. This section reviews four cIas- sification scales receiving extensive cur- rent attention: · the seven-factor scale developed by Greenwood using data from the Rand Corporation inmate survey, and de- scribed in a report that popularized the term "selective incapacitation" (Green- wood, 19821; · the Salient Factor Score used in fed- eral parole guiclelines, perhaps the long- est-lived system in current use (Hoffinan and Beck, 1974; Hoffinan, 19831; · a risk assessment crevice for parole release in Iowa, which has claimed dra- matic improvements in predictive accu- racy over other approaches (Fischer, 1983, 19841; and · a scale developecl by INSLAW, Inc., for use by U.S. attorneys in assigning federal cases for special attention in ca- reer criminal units (Rhodes et al., 19821. All these scales share a common con- cern for distinguishing offenders in terms CRIMINaL CAREERS AND CAREER CRIMINALS of criminal career attributes, and the scales are empirically derives! from ob- served statistical relationships between reaclily obtainable offender attributes and criminal career dimensions. In each case, the scale is intended for use in identify- ing offenders who, on the basis of their criminal careers, pose the greatest risk for continued serious offending. This review of the scales focuses primarily on their classification accuracy and on the ade- quacy of the variables used in each scale. We do not acIdress the process by which the variables were selected and weighted, primarily because that process is not well documented in the available literature. Rand Inmate Survey Scale Growing interest in pursuing selective incapacitation as a sentencing policy prompted Greenwood (1982) to devise a method for distinguishing which offend- ers commit offenses at high rates. The scale is unique in its use of offending frequency (A) as the criterion variable, rather than officially recorded events, such as arrest, which enter the more com- mon recidivism measure. Thus it intro uced two important changes: (1) a focus on crimes rather than arrests and (2) a concern for differences in frequency of offending-a continuous variable-rather than a simple binary recidivism measure. However, in contrast to a recidivism mea- sure that combines the effects of fre- quency and career termination, the Rand scale focuses only on frequency. Using the scale as intencled for determining the length of incarceration implies an assumption that there is no career termi- nation, that offending at the estimated frequency will continue inclefinitely. Using data from a self-report study of 2,190 incarcerated offenders in Califor- nia, Michigan, and Texas, Greenwood se- lectecl seven variables that individually classified inmates reasonably well into

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING high, medium, and low rates of offending for robbery and burglary and that he deemed appropriate for sentencing pur- poses. The variables, shown in Table 6-2, were each scored either O or 1 to form an additive Burgess-type scale that results in a prediction score between O and 7 for each inmate. Inmates were then cIassi- fied as committing crimes at a low rate (scoring O or 1), medium rate (scoring 2 or 3), or high rate (scoring 4 or more). Em- pirical, ethical, ancI operational concerns regarding the Rand Inmate Survev Scale are reviewed extensively elsewhere (Visher, Volume II; see also Cohen, 1983, 1984b; Spelman, 1984; van Hirsch ant] Gottfrectson, 19841. Some of the results of these reanalyses and critiques of the Rand data are highlighted here, espe- cially the work of Visher, which was com- missioned by the panel. The principal utility of the scale for achieving future crime reduction de- pends on the accuracy of the scale in prospectively identifying offenders who will commit crimes at high rates in the future. The existing research, as Green- wood states, is limiter! to retrospective data on past offending and no validation of the instrument on future offending was carrier! out. Without explicit prospective validation, the scale's utility for prospec- tively identifying future high-rate offend- ers depends on the retrospective accu- racy of the scale and on the extent to which past high-rate offending continues into the future. The analysis shows that the retrospec- tive accuracy of the seven-variable cIassi- fication rules is no better than other sim- ilar instruments (see Cohen, 1983; Visher, Volume II:Table 191. Specifi- cally, even within the sample used to construct the scale, 55 percent of the classified high-rate group (27 percent of the total sample) were false positives who did not commit crimes at high rates. Sim- ilar 50 to 60 percent false-positive errors (typically measured by no new arrests, 179 TABLE 6-2 Variables Used in Rand Scale to Distinguish Inmates by Individual Crime Rates Variablea Source Official criminal recordsb 1. Prior conviction for same charge (robbery or bur- glary) 2. Incarcerated more than 50 percent of 2 years 3. Convicted before age 16 4. Served time in state juve Self-report Self-report Self-report nile facility 5. Drug use in preceding 2 years 6. Drug use as a juvenile Self-report 7. Employed less than 50 per- Self-report cent of preceding 2 years Self-report aAll variables are scored as 0 or 1 depending on the presence or absence of the attribute. bData available only for prison, not jail, in- mates. parole violations, or revocations cluring a clesignated follow-up period) have been found in the predicted] worst-risk groups for various statistically based prediction devices intended! to assess the risk of recidivism for use in parole release deci · c> slons.o Furler examination of the scale's ac 8For example, in a 1976 validation of the Federal Parole Commission's Salient Factor Score, 44 per- cent of the "poor risks" (20.3 percent of the sample) were rearrested or had a parole violation warrant issued in a 1-year follow-up after release on parole (Hoffman and Beck, 1980~. In a validation of a slightly revised Salient Factor Score (Hoffman, 1983), 49 percent of the predicted poor risks (23.9 percent of the sample) recidivated during a 2-year follow-up (i.e., were recommitted to prison, had outstanding parole violation warrant issued, or were killed while committing a crime). J. Monahan (1981:103) reports that a similar risk-assessment device developed by the Michigan Department of Corrections found a 40 percent recidivism rate (i.e., arrest for a new violent crime while on parole) in their "very-high-risk" group (4.7 percent of the sam- ple). M.R. Gottfredson, Mitchell-Herzfeld, and Flanagan (1982) report a 49.6 percent readmission rate to state correctional institutions in a 5-year follow-up for their high-risk group (16.3 percent of their sample of 1972 releasees from correctional institutions in a northeastern state).

180 curacy revealed important cross-state variation in its predictive power. False- positive rates for high-rate robbers were 60 percent in California and Michigan and 48 percent in Texas. Focusing on high-rate offenders, especially robbers, these false-positive rates reflect a 57 per- cent relative improvement over chance (as measured by RIOC) in California, but only 21 percent and 38 percent in Mich- igan and Texas, respectively (Visher, Vol- ume II:Table 191. The Ranc! scale has not yet been vaTi- dated on samples other than the one used to construct the scale. Because of random variations in attributes across samples drawn from the same population, statisti- cal shrinkage will slightly recluce the ac- curacy observed in the construction sam- ple. Even larger reductions in accuracy can be expected if the scale is applier] to different populations. In comparison with a population of all convicted offenders, the more selected subpopulation of in- mates used to develop the scale may differ in important ways in their attributes ant! in the relationship of those attributes to high-rate offending. If so, the applica- tion of a scale developed on inmates to a broacler population would reduce cIassi- fication accuracy still furler. Also, be- cause the scale captures only frequency and not termination of careers, its pro- spective accuracy will deteriorate further as careers encl. Thus, even the modest classification accuracy of the scale, re- flectec! in a 55 percent false-positive rate for high-rate robbers in the construction sample, is likely to overstate the accuracy that would be obtained in actual imple- mentation of the scale. Some ethical concerns also exist about the variables that make up the scale. The variables (indicated in Table 6-2) invoked are consistent with previous empirical research, and few people question the use of prior adult criminal record. But measures ofthe juvenile recorcl, drug use, and recent employment may be ques CRIMINAL CAREERS AND CAREER CRIMINALS tioned on grounds of blameworthiness. The controversy surrounding these vari- ables may lead to their exclusion in im- plementing the classification scale. Such exclusion of statistically useful variables may further diminish the accuracy of the amencled classification scale, thereby di- minishing the crime-reduction effects of implementing the scale. In anticipation of such possible amendments, it is advis- able-as was done in Greenwood's anal- ysis of the inmate scats- to examine the accuracy of alternative versions of the scale under various exclusion conditions. While Greenwood (1982:61-66) com- pares different scale versions in terms of differences in mean crime rates for scale- identified offender groups, no data on individual error rates are reported. These error rates were subsequently derived in the panel's analysis of the Rand inmate data and are reported in Table 6-3. Elimination of potentially con- troversial variables relating to drug use, employment, and juvenile criminal record does not substantially affect the false-positive rate among preclicted high- rate inmates, nor does it reduce the over- all predictive accuracy of the scale. Adult criminal record variables alone perform nearly as well on these measures as clid the full seven-factor scale. Elimination of variables from the scale, however, does significantly reduce the accuracy of the scale in identifying high-rate offenders; the percent of true high-rate offenders who are missed by the scale increases from 50 to 87 percent when the scale is limited to only adult criminal record vari- ables (Modified Scale A). The accuracy of the scale also varies with the cut point that is used to desig- nate hi~h-rate offenders (see Table 6-1~. In Greenwoocl (1982), all inmates with scores of 4 or more are designated as high-rate offenders. As shown earlier, this choice of a cut point comes closest to matching the scalds selection ratio (27.2 percent predicted to be high rate) to the

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING TABLE 6-3 Changes in the Error Rates of Rand Inmate Scale as Predictor Variables Are Eliminated ]8] Error Rates for High-Rate Offenders Scale False Positives Among Predicted High-Rate Offenders (percent) False Negatives: Actual High- Rate Offenders not Correctly Identified (percent) Total Inmates Predicted Correctly (percent) Original scale 7 factors 55 50 46 Modified scale Ba 60 81 41 Modified scale Ab 60 87 45 NOTE: Accuracy is assessed on the sample of inmates convicted of robbery or burglary in the three states surveyed, the same sample used to construct the scale. aOnly variables relating to prior criminal record are included; drug use and employment variables are excluded (see Table 6-2~. bOnly the variables relating to adult criminal record are included; juvenile record, drug use, and employment variables are excluded (see Table 6-21. defined base rate of 25 percent high-rate offenders. At this cut point, 50 percent of high-rate offenders are correctly identi- fied by the scale ant] there are 55 percent false positives among predicted high-rate offenders. As is found when prediction variables are eliminatecI from the scale (Table 6-3), altering the cut point for designating high-rate offenders substantially changes the accuracy with which high-rate offend- ers are identifiecI. When a score of 6 or more is required, only 9.2 percent (20 of 218) of all high-rate offenders are cor- rectly identified (Table 6-11. This accu- racy in identifying high-rate offenders in- creases to 75.2 percent (164 of 218) when the cut point is Towered to 3. The false- positive rate, by contrast, is far less sensi- tive to changes in the cut point over much of its range (see Table 6-11. As the cut point is lowered, more inmates are desig- nated as high-rate offenders on the basis of their scale score. For example, the false-positive rate increases from 48.7 percent at a cut point of 6 to 60.S percent at a cut point of 3 or more. There are also questions about the availability of the scale variables for op- erational use. In Greenwood's analysis, all of the variables, except past convic- tions for the same charge, are based on self-reports by inmates. Implementation of the scale at sentencing would have to depend on official records rather than self-reports as the source of data for scale variables. For some ofthe scale variables, especially those relating to juvenile record, drug use, and employment, this requirement may be more than records systems can routinely deliver. Salient Factor Score Most scales for classifying prisoners for parole release have focused on recidi- vism as the criterion variable. Parole guidelines were first put into use in the federal system in 1972 (for a review, see Gottfredson and Got~redson, Volume II). The current guidelines (last revised in 1981) have two components that are pro- vided to parole commissioners as aids for their decisions: the seriousness of the commitment offense and an empirically derived assessment of future recidivism risk, based on six offender characteristics. The latter is known as the Salient Factor Score (SFS 81 since 1981) and is repro- cluced as Table 64. The higher the score,

182 TABLE 6-4 Salient Factor Score (SFS 81) CRIMINAL CAREERS AND CAREER CRIMINALS A. PRIOR CONVICTIONS/ADJUDICATIONS (ADULT OR JUVENILE) ........ None = 3 One = 2 Two or three = 1 . Four or more = 0 B. PRIOR COMMITMENTS OF MORE THAN THIRTY DAYS (ADULT OR JUVENILE) None = 2 One or two = 1 Three or more = 0 C. AGE AT CURRENT OFFENSE/PRIOR COMMITMENTS ...... Age at commencement of the current offense: 26 years of age or more....................................... 20-25 years of age ..................... = 2*** = 1*** 19 years of age or less = 0 ***F.X(?F`PTION If five or more prior commitments of more than thirty days (adult or juvenile), place an "x" here and score this item... D. RECENT COMMITMENT FREE PERIOD (THREE YEARS) .......... No prior commitment of more than thirty days (adult or juvenile), or released to the community from last such commitment at least three years prior to the commencement of the current offense .......... 0therwise = 0 E. PROBATION/PAROLE/CONFINEMENT/ESCAPE STATUS VIOLATOR THIS TIME . Neither on probation, parole, confinement, or escape status at the time of the current offense; nor committed as a probation, parole, confinement, or escape status violator this time = 1 0therwise = 0 ~= ~ ALIT /~T A TU ^U DU OTIS STIR . . . . ...... F. r1 ~n~ll~/~` 1^ l 1:. ~1 ~ ~1~ ............. No history of heroin or opiate dependence ..... Otherwise .................................... TOTAL SCORE .................................. - SOURCE: Hoffman (1983:Appendix A). Reprinted with permission from Journal of CriminalJustice' Volume 11, Hoffman, Peter B., Screening for risk: a revised salient factor score (SFS 81), ~ 1983, Pergamon Press, Ltd. the higher is the likelihood of no recom- mitment (see Hoffinan, 1983; [anus, 19851. Since initial implementation of federal parole guidelines, the Salient Factor Score has been revised and vaTiciatecl pro- spectively on several new samples (see Hoffman, 1983; Hoffinan and Beck, 1976, 1980; HofEnan, Stone-Meierhoefer, and Beck, 19781. Two measures of predictive power-point-biserial correlation, mean cost rating show that for all versions, the score and the four risk categories are at the high end of the accuracy range re- portec] in other parole recidivism studies (Gottfredson and Gottfredson, 19801. In addition, these accuracy measures have been very stable over several validation samples. Hoffinan (1983) reports that for construction and vali(lation samples, re- spectively, 46 and 49 percent of parolees classified as poor risks by SFS 81 were considered failures (primarily because of reincarceration) at the end of a 2-year follow-up period. This accuracy is typical of statistical classification scales that as- sess offending risk, 50 to 55 percent of those classified as poor risks appearing as false positives. It is noteworthy that the SFS shows little evidence of deteriora- tion from the construction sample of 197~1972 prison releasees to the valida- tion sample of 1978 releasees. Indeed, the false-positive rate among poor risks cleclines from 54 percent in the construc- tion sample to 51 percent in the vaTida

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING tion sample ant! the false-negative rate remains unchanged at 24 percent. Con- trasting offenders classified as poor risks to the combination of all other offenders, the relative improvement in accuracy over that obtained by chance (RIOC) also increases slightly from 24 percent in the construction sample to 28 percent in the validation sample. In another validation study of SFS 81, using a broacler criterion of rearrest in a 3-year follow-up period for the same val- iciation sample, the scale performed bet- ter: 66 percent of the poor-risk offenders incurred an arrest, compared with 22 per- cent of the very good risks, for a 34 per- cent false-positive rate ([anus, 1985; see also Hoffman, Stone-Meierhoefer, and Beck, 19781. In moving from Hoffiman's (1983) recidivism criterion of reincarcera- tion in a 2-year follow-up to the broader criterion of rearrest used by Janus (1985), the recidivism base rate increased from 30 to 44 percent, and so some reduction in the false-positive rate would be expected. Similarly, the RIOC in identifying the poor risks (compared with all others) also improved somewhat, increasing from 28 percent to 38 percent in the vaTiciation sample. The increase in the base rate is undoubtedly responsible for some of the improvement in predictive accuracy, but there may also be some gain in predictive accuracy because the scale is applied to rearrest rather than reincarceration. In addition, the study reports that the SFS appears to ctistinguish among releasees in the four predicted risk groups according to their arrest frequency rates, ,u: those classified as fair or poor risks were rearrested sooner and at higher average rates than releasees classified as very good or good risks. As was demon- strated in Chapter 5 (Table 5-5), however, the differences in ,u across different risk groups are much smaller when only the frequencies for active offenders (i.e., those with at least one arrest in the fol ~3 low-up period) are compared. The dif- ferent risk groups also clisplayed little difference in offense seriousness ([anus, 1985: 122-123~. The SFS thus appears to discriminate most effectively between persisters and those who do not incur another arrest within 3 years (which in- cludes mainly offenders who have termi- nated their criminal careers) and is less effective in distinguishing frequency rates for active offenders. This finding still leaves a challenge to finch adequate predictors of differential frequency rates among those who do persist in their crim- inal careers after release. The SFS appears to be a simple, stable, and fairly reliable prediction-based cIas- sification scheme, which is also reason- ably accurate using rearrest in a 3-year follow-up as the criterion variable. The latest version, SFS 81, emphasizes crimi- nal history items; it reflects concerns about cliff?iculties in recording some vari- ables and omits measures of employment stability, family ties, and high school ed- ucation, which were included in earlier versions. Heroin or opiate dependence and age at current offense are the only remaining items that might be challengecl on ethical grounds. Comparisons of alter- native versions of the SFS on the same sample (Table 6-5) show no detectable deterioration in accuracy associated with these changes in scale variables (Hoff- man, 1983~. The percentage of favorable outcomes for different risk levels, the mean cost rating, and the RIOC change hardly at all for the more restricted ver- sions of the scale. Iowa Risk Assessment Instrument Another classification instrument for assessing parole recidivism risk has re- cently been cleveloped in Iowa (Chi, 1983; Fischer, 1983, 19841. All of the information about this instrument ap- pears in unpublished reports and re

184 TABLE 6-5 Changes in the Accuracy of the Salient Factor Score as Predictor Variables Are Eliminateda Percent Favorable Outcomeb Risk Category SFS 81C SFS 76 Very good Good Fair Poor All cases Mean Cost Rat- ing (MCR) Relative Im- provement Over Chance (RIOCi 88 (N = 735)e 75 (N = 502) 61 (N = 542) 51 (N = 560) 70 (N = 2,339) 89 (N = 642) 73 (N = 662) 64 (N = 497) 50 (N = 538) 70 (N = 2,339) .40 .28 .29 aAccuracy is assessed on an independent vali- dation sample of inmates released from federal institutions in 1978. bFavorable outcomes are defined as no commit- ment of 60 days or more for a new offense, no return to prison as a violator, no parole violation warrant outstanding, and not killed while commit- ting a criminal act. CPredictor variables reflect prior criminal rec- ord, offender's age, and drug use (see Table 6-4). fin addition to variables reflecting prior crimi- nal record, age, and drug use, the earlier scale included an employment stability measure. eThe number of cases in each risk level is reported in parentheses. fThe RIOC measure was computed from data reported by Hoffinan (1983) using a selection rule that treats "poor risks" as high-risk and aggregates all other risk-levels. SOURCE: Data from Hoffman (1983:Tables 1 and 2~. search documents. Moreover, the existing documentation concerning the construc- tion of the scale, validation attempts, and accuracy leaves many questions unan- swered. Early reports claimed that the Iowa classification system was more than 80 percent accurate in identifying poor parole risks. In response to those dra- matic reports, considerable momentum has developed for disseminating and rep- licating the scale widely. Because of the widespread interest in the Iowa scale and concerns about its development and val CRIMINAL CAREERS AND CAREER CRIMINALS idation, the pane! asked that the Iowa risk assessment instrument be included in the review of the accuracy of classification scales it commissioned (see GottEredson and Gottirecison, Volume II). The approach used to develop the ini- tial scale presented by Iowa is not pub- licly documented. Personal communica- tions with the developers revealed that they used interactive methods (using logic similar to Automatic Interaction De- tection (AID); see Sonquist, Baker, and Morgan, 1973) to develop the scale. Sta- tistically, such methods use up large numbers of degrees of freedom in fitting the mode] very closely to the construction sample data. Furthermore, the "valida- tion" results presented were actually based on a contaminated validation sam- ple that combined cases from the original construction sample with a much smaller subsequent sample. As a result of these and other criticisms, Iowa developed a revised scale, shown in Table 6-6. This latest version of the Iowa offender risk assessment instrument (Fischer, 1984; hereafter referred to as the 1984 version) primarily includes criminal history items: current offense type; disposition if cur- rent offense involves escape, jailbreak, or flight; total "street time" (time not incar- cerated for a felony) since age 14; prior arrest history for violent offenses; and prior juvenile and adult convictions and incarcerations. The prior record variables are weighted by offense seriousness and time since a prior event. A detailed sub- stance abuse score also figures into the classification. According to this informa- tion, each offender is assigned to risk categories for two separate risk assess- ments: general/safety risk and violence risk. The revised scale is conceptually very similar to the Salient Factor Score, although with an appreciably larger num- ber of variables to be scored. The scale has been validated prospec- tively using follow-up recidivism data. The contamination of the validation sam- ples with construction sample data that

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING TABLE 6-6 The Iowa Offender Risk Assessment Scale: 1984 Version A. CURRENT OFFENSE SCORE Robbery, Larceny from Person, Aggravated Burglary, Arson Murder, Manslaughter, Kidnapping, Rape, Sodomy Burglary, Selling Narcotics, Motor Vehicle Theft, Forgery, Bad Checks, Fraud Aggravated Assault, Extortion, Armed with Intent, Conspiracy to Commit Violent Felony, Larceny, Stolen Property Vandalism, Weapons, Conspiracy to Commit Nonviolent Felony None of Above B. PRIOR VIOLENCE SCORE Sum of prior violent felony charges weighted by seriousness and time back to arrest: 91+ 11-90 0-10 C. STREET TIME SCORE Total time free on street (i.e., not incarcerated for a felony) since age 14: 0-6 years 6-11 years 11-14 years 14+ years D. CRIMINAL HISTORY SCORE Sum of prior felonies resulting in conviction or incarceration weighted by seriousness, time back to arrest, and total street time: 140+ 41-139 16-40 0-15 E. CURRENT ESCAPE SCORE If current incident for prison escape, jailbreak, or flight: Convicted Arrested/Charged Only Not Above SUBSTANCE ABUSE SCORE History of PCP Use, Non-Opiate Injections, Sniffing Volatile Substances History of Opiate Addiction History of Heavy Hallucinogen Use History of Drug Problem History of Opiate or Hallucinogen Use, or Alcohol Problem No History of Above G. SERIOUS OFFENDER CLASSIFICATION (Used with Violence Risk Score) 185 General Violence Risk Risk 2 3 1 ~ 2 1 1 o 4 2 1 o o s 3 O O 3 3 2 2 1 O O 6 6 3 5 1 O O 1 3 4 2 O O 7 5 4 4 3 4 2 O O Current Conviction for Violent Felony, or Escape/Jailbreak/Flight; Prior Yes Conviction for Felony Against Person in last 5 years of Street Time; Prior Violence Score 35+; Substance Abuse Score 7 None of Above Factors TOTAL SCORE X-Score = A + B + C Y-Score= D + E + F SOURCE: Fischer (19841. No

186 distorted the validation of earlier versions of the Iowa scale does not appear to be a problem in the 1984 validation (Statistical Analysis Center, 19841. Fischer (1984:23) reports that of 1,000 cases of 197~79 releasees available for development of the 1984 version, 186 were reserved for use as a validation sample and that sepa- rate outcome measures are reported for these construction and validation sam- ples. A further problem of inadequate controls for varying times at risk, found in the evaluation of the earlier scales, also appears to have been remedied by use of a 4-year follow-up period for all members ofthe validation sample ofthe 1984 scale. As noted above, early versions of the scale reported by the Iowa Statistical Analysis Center claimed accuracy levels greater than any other known criminal risk assessment scale, but the reanalysis of the outcome data by Gottfredson and GottEredson (Volume II) concluded that the accuracy claims for the early scales were highly inflatecI.9 The evaluation of the 1984 version is much more careful. On the basis of follow-up data on rearrest and reincarceration reported in Fischer (1984:24-25), the RIOC contrasting the combined poor and very poor risks to all others is calculated to be 54.1 percent in the construction sample and 58.8 percent in the validation sample. These improve- ments in accuracy are associated with 29 and 38 percent false-positive rates in the two samples, respectively. In compari- son, as mentioner! earlier, the RIOC for predicted high-rate offenders, using Greenwood's seven-factor scale, aver- aged 31 percent for the total sample and varied widely among states and crime 9After a careful study of the empirical underpin- nings of the mean cost rating measure used in the Towa valiclation. Gottfredson and Gottfredson (Vol ume II) concluded that the MCR as calculated by the Iowa researchers is inflated compared win standard MCR measures reported for over scales and is "essentially meaningless." CRIMINAL CAREERS AND CAREER CRIMINALS types: 57 percent for high-rate California robbers but only 19 percent for Michigan burglars. These comparisons among dif- ferent prediction instruments suggest that the latest Iowa classification system may be somewhat more accurate than other prediction devices. GottEredson and Gottfredson (Volume II) attribute this im- proved accuracy to the general strategy of relying on reasonably homogeneous subsamples in developing intermediate classification devices that are then com- bined into a final scale. The 1984 Iowa scale differs in substan- tial ways from the earlier versions. In particular, several variables that might be ethically objectionable have been elimi- nated from the scale. Among the ex- cluded variables are marital status, em- ployment status, job skill level, and age at first arrest; and a previous heavy empha- sis on juvenile record (rather than adult criminal record) was also eliminated. The 1984 scale relies heavily on current of- fense, recent criminal record, and sub- stance abuse. Removing the controversial variables apparently led to no reduction in scale accuracy (see Gottfredson and Gottfredson, Volume II). Proposed INSLAW Scale At the prosecution stage, formal selec- tion rules are widely used for assignment of cases to career criminal units for spe- cial attention. One such rule developed at INSLAW for CCUs in U.S. attorney's of- fices has strong empirical underpinnings (Rhodes et al., 19821; see Table 6-7. Based on a follow-up of 1,708 offenders after their release from federal custody, a nine-factor scale was developed that cIas- sified offenders on federal probation or parole in terms of the time to first rearrest within a 60-month period. Because the scale was developed on a sample of re- leased offenders rather than on an ar- restee sample and was neither validated

USE OF CRIMINaL CAREER INFORMATION IN DECISION MAKING TABLE 6-7 Proposed INSLAW Scale for Selecting Career Criminals for Special Prosecution \7ariable Points 187 Heavy Use of Alcohol Heroin Use Age at Time of Instant Arrest Less than 22 23-27 28~2 33~7 38-42 43+ Length of Criminal Career 0-5 years 6-10 11-15 16-20 21+ Arrests During Last Five Years Crimes of Violence Crimes Against Property Sale of Drugs Other Offenses Longest Time Served, Single Term 1-5 months 6-12 13-24 25~6 37-48 49+ Number Probation Sentences Instant Offense Was Crime of Violence* Instant Offense Was Crime Labeled "Other"** Critical Value to Label an Offender As a "Career Criminal": +47 points +5 +10 +21 +14 +7 o -7 -14 +1 +2 +3 +4 +4 per arrest +3 per arrest +4 per arrest +2 per arrest +4 +9 +18 +27 +36 +45 + 1.5 per sentence +7 -18 *Violent crimes include homicide, assault, robbery, sexual assault, and kidnapping. **Other crimes include military violations, probation and parole violations, weapons and all others except arson, burglary, larceny, auto theft, fraud, forgery, drug sales or possession, and violent crimes. SOURCE: Rhodes et al. (1982:Table Van. On an independent sample nor imple- mentecI in practice, no assessment can be made about its accuracy beyonc] the con- struction sample. Relying on time to rearrest as the de- pendent variable, the INSLAW scale is intended to distinguish among offenders in terms of their frequency rates- which are reciprocally related to the time to rearrest. On the basis of a statistical mode} that fit time to rearrest to offender at- tributes, including prior record, 200 "ca- reer criminals" were iclentifiecI. Again, by relying on predicted! time to rearrest as a basis for estimating inclividual arrest rates, combined with estimates of a ho- mogeneous risk of arrest per crime by offense type, sharp differences in esti- mated individual offense rates, A, were calculated. Individuals iclentified by the scale as career criminals were each esti- matec! to commit an average of 38 nondrug crimes per year, compared with estimated rates of only 4 nondrug crimes per year for other offenders. There are no published results on the

188 accuracy of these estimates of A that are based on actual rearrest experiences oh- pated. served in the two offender risk groups. The only measures of accuracy reported, baser! on arrests observed during the fol- Tow-up period, are error rates. The false- positive rate was 15 percent among scale- identified career criminals and the false-negative rate was 36 percent among all other offenders; the RIOC is 74 per- cent, and the scale appears quite accurate in distinguishing among offenders in the construction sample in terms of their risk of future arrests. Some deterioration in this accuracy can be expected as the scale is applier! to independent validation samples, especially if the scale, which was developed using released offenders on probation and parole, is applied to the intended target population of ar- restees. The INSLAW scale presented in Table 6-7 invokes a number of offender charac- teristics in ways consistent with what is known about the relationship of those characteristics to the frequency, serious- ness, and duration of offending. Drug and heavy alcohol use are treated as aggravat- ing factors, as is early age at onset (in the form of elapsed career length). Age is presented as an aggravating factor until age 32 and as a mitigating factor after age 38, but over the entire age range, the older the suspect, the better the score. Both frequency and seriousness of prior arrests are aggravating factors. Because of its consistency with available knowledge of criminal careers, the INSLAW scale shows promise; it should be validated on other samples, perhaps leading to exper- imental implementation. Because some scale variables may be controversial e.g., substance abuse and reliance on past arrests rather than convictions- such analyses should include assessments of changes in accuracy under alternative scale formulations; in light of experience with other scales examined here, large CRIMINAL CAREERS AND CAREER CRIMINALS reductions in accuracy are not antici Summary Some recent classification scales rely- ing on similar predictor variables re- flecting primarily prior criminal record and drug use- have displayed greater ac- curacy in identifying worst-risk offenders than earlier scales. When criteria of rear- rest in follow-up periods of at least 3 years are used, false-positive rates among of- fenders classified as worst-risks are under 30 percent in construction samples and under 40 percent in validation samples (see especially Rhodes et al., 1982; Fischer, 1984; and Janus, 1985 in Table 6-~. This difference from the 50 to 60 percent false-positive rates that have tra- ditionally characterized statistical cIassifi- cations of poor-risk offenders is due in part to reliance on broader criterion vari- ables, particularly recidivism measures based on arrests rather than recommit- ment to prison, and measures using longer follow-up periods that increase the likelihood of detecting recidivism. The resulting higher base rates contribute to greater classification accuracy, especially in relation to low selection ratios. For example, the INSLAW scale, with its high base rate and low selection ratio, results in a low false-positive rate (15 percent) but a high false-negative rate (36 percent). However, to the extent that the gains in accuracy are beyond those ex- pected merely from the increased base rates, the broader criterion variables may yield a more sensitive measure of releasees' offending behavior. The Rand scale displays the least im- provement from earlier 50-60 percent false-positive rates. As the only scale based on self-reports, it is unique in at- tempting to predict actual crimes commit- ted rather than arrests. There may be more measurement error in individual

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING TABLE 6-S Accuracy of Several Explicit Classification Scales Assessing Offending Risk ]89 Iowa Assessment INSLAW Rand Inmate Salient Factor Score 81 Scale Scale Scale Hoffman Janus Fisher Rhodes et al. Greenwood Scale Attribute (1983) (1985) (1984) (1982) (1982) Construction sample Federal prisoners Iowa state prisoners Federal pro- Prison and jail released in released in 1976-80 bationers inmates in 1971-72 (N = 814) and parol- California, (N = 3,955) ees re- Michigan, leased from and Texas custody in in 1978 1970-71 (N = 886) (N= 1,708) Validation sample Federal prisoners (N = 2,186) Iowa state prisoners None None released in first released in 1976-80 half of 1978 (N = 186) (N = 2,339) Recidivism criterion Reincarceration Rearrest Rearrest or reincarcer- Time to rear- Individual fre ation for serious rest for seri- quency crime ous crime rates for robbery and burglarya Length of follow-tip 2 3 4 5 None (1- to 2 (years) year retro spective ob servation period) Base rate (%)b Construction sample 31 N.R. 37 42 25 Validation sample 30 44 26 N.V. N.V. Selection ratio (%)c Construction sample 34 N.R. 33 12 27 Validation sample 24 23 30 N.V. N.V. False-positive rate (%) Construction sample 54 N.R. 29 15 55 (3_69)d (0 - 64) (0 - 58) (3_75) Validation sample 51 34 38 N.V. N.V. (0-70) (0-56) (4-74) False-negative rate (%) Construction sample 24 N.R. 20 36 17 (0~1) (4~7) (30-42) (0-25) Validation sample 24 38 11 N.V. N.V. (6-30) (21-44) (0-26) Relative Improvement Over Chance (RIOC)e Construction sample .24 N.R. .54 .74 .31 Validation sample .28 .38 .59 N.V. N.V. NOTE: N.R. = not reported; N.V. = not validated. aThe criterion variable in the Rand Inmate Scale is individual frequency rates of committing crimes, and not a traditional binary recidivism variable (e.g., rearrested or not during a follow-up period). bPercent unfavorable outcomes found in the total sample. CPercent predicted to be high risks in the total sample. dThe minimum and maximum possible values of the false-positive (FP) and false-negative (FN) rates appear in parentheses. These are determined from the selection ratio (SR) and base rate (BR) as follows: FP c 1 - BR; FN c BR; FP 2 SR - BR when SR 2 BR, = 0 otherwise; FN 2 BR - SR when BR 2 SR, = 0 otherwise. eThe Relative Improvement Over Chance (RIOC) compares actual classification accuracy to the maximum accuracy possible and the random accuracy associated with the selection ratio and base rates in sample (see text): RIOC [(1-FP) SR + (1 - FN) (1 - SR)] - [(SR) (BR) + (1 - SR) (1 - BR)] [MIN(SR, BR) + MIN(1 - SR, 1 - BR)] - [(SR)(BR) + (1 - SR) (1 - BR)]

190 crime rates estimated from self-reports than in criterion measures based on offl- cially recorded arrests, and this increased error in the dependent variable would reduce the accuracy of predictions. The Rand scale is also only a preliminary attempt at calibrating a classification scale: simple bivariate relationships are used to identify potential predictor vari- ables, and Burgess weights (scored either O or 1) are assigned to the selected vari- ables. This approach contrasts with the extensive multivariate analysis, which in- volves more flexible characterization of the contributions by individual predictor variables. These analyses are reflected in the more complex weighting schemes found in other scales. The recent classification scales are also noteworthy because when independent validation samples representing similar populations have been used, as with the federal parole system's Salient Factor Score and the 1984 Iowa risk assessment scale, there is little evidence of substan- tial statistical shrinkage in their accuracy. This result is due in part to the use of reasonably large samples in scale devel- opment more than 800 individuals- which reduces the impact of statistical shrinkage due to sampling variation. Also, validation studies indicate that offender populations appear to be reasonably sta- ble over time: there was no detectable reduction in accuracy between applica- tions of the Salient Factor Score to a sample of federal prisoners released dur- ing 1970-1972 and another sample re- leased in 1978. The various classification scales also show no evidence of important reduc- tions in scale accuracy when various con- troversial variables like employment sta- tus are removed from them. It thus appears that accommodating just-deserts concerns by avoiding ethically objection- able classification variables will not seri- ously impair classification accuracy. CRIMINAL CAREERS AND CAREER CRIMINALS ADULT, JWENILE, AND NON-JUSTICE SYSTEM RECORDS AND THEIR INTEGRATION Attention to certain offender character- istics whether achieved through formal rules or infollllal practice-can, to some extent, improve the accuracy of offender classification in terms of criminal career dimensions. Incorporating these cIassifi- cations into decisions requires that deci- sion makers receive timely and accurate information about predictor variables, and doing so will require improvements in criminal justice and other record- keeping systems. We focus here on adult criminal histories, the record of prior adult arrests and dispositions; juvenile records, the records of police contacts and juvenile court adjudications; and other records maintained outside the justice system on such matters as drug use, employment history, school perform- ance, and social history. These record categories differ in terms of how accessi- ble they are to criminal justice decision makers. Adult Records In most jurisdictions, the arrest and fingerprinting of an individual normally initiates both an update of the local arrest record and retrieval of the record for use by the prosecutor at screening and for use by the judge at the initial court appear- ance shortly after arrest, when charges are filed and pretrial release conditions are set. In some jurisdictions, local authori- ties do not receive the entire history of arrests in other jurisdictions for several weeks, when arrest records are received from repositories in state identification bureaus and the FBI. These delays can lead to the use of incomplete prior record information at the important stages of ini- tial screening by the prosecutor, charg- ing, and pretrial release decisions. How

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING ever, many states have established statewide computerized criminal history files, and many metropolitan areas have established regional systems combining arrest histories from participating juris- clictions. From these networks, records of prior arrests by participating jurisdictions can be retrieved quickly ant! linked to the proper individual using demographic identifiers and fingerprint classification. An arrest is the result of a law enforce- ment action based on the evidentiary standard of"probable cause"; it is not "guilt beyond a reasonable doubt." Therefore, it is argued that, under the latter standard, only prior convictions should be used as an indicator of previous criminal activity and that records of ar- rests not followed by a record of a convic- tion should not be used. Advocates of this position point to potential biases in deci- sions to make and record arrests and to the possibility that use of arrest records in decision making may encourage more fi~- ing of charges or filing on multiple counts on the basis of weaker evidence, at least for offenders who are considered espe- cially persistent. Furthermore, advocates of the position argue that police arrest records receive no independent scrutiny and are therefore more likely to contain undetected inaccuracies than are court conviction records, which are open to the public. In contrast, advocates of current prac- tice argue that exclusion of arrest data until disposition data become routinely available would, in some jurisdictions, preclude access to useful information for years. Moreover, some advocates con- sider an arrest record to be a more sensi- tive and vaTic! measure of a criminal ca- reer than a conviction record because the former more fully reflects variations in the rate of activity, because a failure to convict is not necessarily indicative of factual innocence, and because plea bar- gaining can distort the meaning of convic 191 tion records. Conviction records are also vulnerable to distortions when their in- fluence on decisions is altered: for exam- ple, when the number of prior convic- tions became an explicit basis for sentencing under guidelines introduced in Minnesota, prosecutors seemed less willing to dismiss charges than they pre- viously had been (Minnesota Sentencing Guidelines Commission, 1984:811. It is-generally acknowledged that dis- position records are less readily available than arrest records. Those who advocate a requirement for disposition data argue that the requirement would encourage police, prosecutors, and courts to under- take the costly cooperative effort needed to improve the availability of those data. They argue that information on prior case dispositions and sentences is useful be- cause it describes an offencler's previous sanctioning experience, which permits adjustment for incarceration time in esti- mating the frequency of previous arrests while free, a measure that has been found to predict future arrest frequency (Barrett ant! Lofaso, 1985~. However, while con- viction information can always be re- trieved through a search of written court records, only in a few jurisdictions- where courts, prosecutors, and police op- erate an integrated criminal justice infor- mation system is it routinely available in time for screening, charging, and pre- trial release decisions. Data on actual time served in jail or prison following arrest or conviction are even less gener- ally available, since corrections records on individuals are linked to police and court records in only a few states. And data on dispositions other than convic- tions are not systematically recorded. Because knowledge of previous case (lispositions ant] sanctions can contribute to a current decision, court disposition and corrections data should be integrated with arrest records. Even before full inte- gration is accomplished, several steps

192 should be taken to enhance the validity of arrest records, which would benefit both decision makers and those who study criminal careers. These steps include pro- moting standards for the completeness of arrest records; encouraging crossduris- ~ i c t i o n a ~ s t a n c ~ a r ~ i z a t i o n o f n o m e n c ~ a t u r e that describes aggravating or mitigating attributes of the alleged offense as well as its legal category; and quickly noting ar- rests that are unfounded, either by police officials or by the screening prosecutor. Regardless of the quality of arrest rec- ords, opinions will continue to differ about their appropriate use in decision making. However, decisions about their use should be influenced by attention both to the ethical issues discussed ear- lier in this chapter and to the value of arrest-history information in predicting the future course of a criminal career. Juvenile Records CRIMINAL CAREERS AND CAREER CRIMINALS these restrictions mean that the arrest histories available for formal adult pro- ceeclings usually do not include records of juvenile contacts with police and, usu- ally, court dispositions that occurred be- fore the defendant reached legal adult- hood. Police information is sometimes shared informally with prosecutors (for whom it may affect charging decisions), but in a national survey, 60 percent of responding prosecutors reporter! that po- lice records of juveniles were "rarely or never" provided at the time charges were filed (Greenwood, Petersilia, and Zim- ring, 1980~. Juvenile court records may legally be used in adult courts in most states for postadjudication decisions, such as dispo- sition and classification. In many states they may be used for predisposition deci- sions such as pretrial (retention. How- ever, the sharing of juvenile court records with adult authorities is operationally hindered by statutory prohibitions against storing juvenile court records in In considering the use of juvenile rec- ~ ~ ordsas decision adds for adult defendants, the same repositories as adult records. it is useful to distinguish between police Such records are routinely retrieved dur records of contacts with juveniles and ingpresentenceinvestigations,whichoc juvenile court records of referrals and curIong after arrest. They can usually be dispositions. According to a recent report obtainers although with some difficulty, prepared for the Bureau of Justice Statis tics (SEARCH Group, Inc., 1982:29 34), police agencies have broad legal discre tion in recording contacts with juveniles. But other criminal justice agencies en counter great difficulty in obtaining this information about individuals prior to their conviction in adult court. Routine exchange is hinclered by laws restricting the fingerprinting of juveniles, a restric tion that complicates subsequent record retrieval and unambiguous matching. Generally, fingerprints of juveniles may be taken only following contact concern ing a serious offense, must be stored sep arately from adults' fingerprints, and must eventually be sealed or purged under various conditions. In most jurisdictions, by police agencies for investigative prear- rest use; but time constraints and the absence of positive fingerprint identifica- tion generally prevent access to them in time for early postarrest decisions. Under these circumstances the disposition infor- mation in juvenile court records is even less likely to be available for use shortly after an arrest than is the record of local police. The completeness of juvenile record repositories themselves is affected by the sealing and purging of the records. Many states have adopted statutes that specify conditions under which juvenile records must be seabeds (i.e., restricted from dis- closure outside the repository except pur- suant to a court order) or purged (i.e.,

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING destroyed). Motivated by concern over stigmatization of adults for their clelin- quent activities as juveniles, these stat- utes generally authorize record sealing within a few years after a person reaches adult age and purging several years later if the subject is not arrested as an adult and petitions the court requesting these actions (SEARCH Group, Inc., 1982:4 511. In addition to legal requirements for sealing and purging individual records, some repositories have begun more gen- eral purging as a means of reducing the cost of storing juvenile records. In combination, the prohibitions against merged juvenile and adult rec- ords, the failure to routinely include juve- nile court data in police record systems, and the sealing and purging of juvenile records create a situation in most jurisclic- tions in which criminal justice authorities frequently make their decisions with no information about police contacts with juveniles. Even when police contact in- formation is available, it is usually not accompanied by information about the disposition following the contact. These conditions, particularly the purging of ju- venile records, have also hampered the efforts of researchers to study criminal careers over the ages of transition from juvenile status to adulthood. Restrictions on the storage and dissem- ination of juvenile records resulted from the orientation of juvenile courts toward rehabilitation rather than punishment and from the belief that juveniles are less responsible than adults for acts that would be labeled criminal by the adult justice system. Some people believe that the fingerprinting of juveniles, a key to accurate identification, may create a "criminal self-image" that may weaken self-imposec! restraints against future crime. It is also argued that dissemination of records outside the juvenile court sys- tem stigmatizes a juvenile with the desig- nation "clelinquent," which may lessen |93 opportunities for legitimate employment, thereby increasing the likelihood of sub- sequent offending. The prospect of hav- ing juvenile records purged after an ar- rest-free period is advocated as an incentive for juveniles to terminate crim- inal careers before adulthood. Others have argued that the ambiguity of nota- tion that is characteristic of juvenile court records makes them very Circuit to inter- pret and summarize (e.g., Greenwood, Abrahamse, and Zimring, 19841. Finally, in pursuing their objective of rehabilita- tion, juvenile courts collect "social data" on such episodes as misconduct in school, mental health status, family his- tory, and other topics that are not only especially sensitive to an individual but may also involve other parties whose pri- vacy would be compromised by dissemi- nation. It is argued that routine sharing of these records could disrupt continued co- operation by the agencies that now pro- vide such background information to the juvenile court. For those who advocate the use of ju- venile records, the challenge is to re- spond to these concerns by designing systems and procedures that inform adult justice system decision makers more fully about juvenile delinquent careers with- out undermining the rehabilitative goal of juvenile courts. Maintaining accessible fingerprint records of juveniles arrested for serious crimes-crimes that would be felonies if committed by an adult may be warranted as a means of eliminating cases of mistaken identity. Using finger- prints to identify persons uniquely, sys- tems could! be developed for the routine transfer of juvenile disposition data to police recorcls. Augmenting police rec- orcis wig data on juvenile dispositions could both improve the fairness with which those records are subsequently used and eliminate the need for access to the juvenile court files with their sensi- tive social data. Juvenile arrest and dispo

194 sition histories could be maintained in a separate repository, with no clisclosure permitted for uses outside the criminal justice system. Adult justice system agen- cies would gain access to the juvenfle record] at the time of a person's first seri- ous criminal involvement as an adult. In defining criteria for access, different juris- dictions might choose different thresh- olds of crime seriousness and different stages of criminal justice processing (e.g., arrest, indictment, or conviction). Requir- ing a high level of seriousness would reflect a desire to avoid opening a juve- nile record in connection with only a minor offense. Precluding access until after conviction would reflect a belief in the principle that a juvenfle record should not influence decisions in the adult system that are made before guilt is established with legal certainty. Alterna- tively, permitting prosecutors and judges to see a juvenile record when making charging, pretrial release, and plea bar- gaining decisions would allow them to become more fully informed of the public safety risks associated with their deci- sions. Different jurisdictions will weigh these concerns differently and so may choose different thresholds and stages. One choice that seems reasonable would be the first adult arrest for a felony. A juvenile record would be appended to an adult record only if the adult arrest leads to a conviction. The juvenile record, with appropriate safeguards against dissemina- tion, would be retained in the juvenfle repository at least as Tong as it retains value for decision making in the aclult criminal justice system. When juvenfle records are no longer operationally useful, they should be pre- served in an otherwise inaccessible way for research purposes. Research on a number of important questions has been hindered by the bifurcation of juvenfle and adult record systems. As highlighted in Chapter 3, particular problems have CRIMINAL CAREERS AND CAREER CRIMINALS been the misleading conclusion of clesis- tance-"false Resistance" from offend- ing that is implied by the truncation of juvenfle court records when indivicluals "graduate" to a(lult jurisdiction ant! the inability to inclucle the record before age 18 in career-length estimates for adult samples. The bifurcation has hampered research on such key questions as the effect of juvenfle justice interventions on adult criminal careers and the influence of information about juvenfle careers on the processing of cases involving young adults. There would be considerable re- search value in linked records of juvenile and adult arrests and dispositions. Record purging preclucles such research. There- fore, while access to juvenile records should be carefully controlled to protect in(livid~uals' identities, those records should be stored as a basis for research. Other Records Some information that is useful in pre- dicting the course of criminal careers is not available in either the adult or juve- nile justice system records. For example, Chaiken and Chaiken (1984:211) found that official-record information was not sufficient to distinguish the offenders they designated as "violent predators" from other offenders. Despite violent predators' self-reports of frequent, violent delinquency as juveniles, their juvenile records either did not exist or were found to be indistinguishable from those of other offenders. Greenwood (1982) pointed to two specific characteristics that distinguish high-rate offenders a history of drug use and significant periods of unemployment. Data on these character- istics are not routinely inclu(led in arrest histories or court disposition files, al- though they sometimes appear with other social data in juvenile court records or adult presentence investigation. Presentence investigations will un

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAEUNG doubtedly continue to be the primary vehicle for the collection of information on drug use, employment status, and fam- iTy and social history. Although maintain- ing such data in permanent criminal jus- tice records might improve the quality of preconviction decisions, that benefit might be at least partly offset by other consequences. For example, routine dis- closure of drug treatment agency records to justice agencies for predictive use might discourage users from seeking treatment or disrupt the therapeutic rela- tionship between users and counselors. Urinalysis at the time of arrest is an alter- native means of determining drug use, and test results could be appended to the arrest record for use following subse quent arrests. This procedure would as- ~~ sist efforts to provide police and prosecu tors with information about an arrestee's history of drug use and would avoid prob lems associated with inflation sharing between treatment agencies and justice agencies. In addition to data on employment his- tory, substance abuse, and other social history, presentence reports in most juris- dictions also include the investigators' sentencing recommendation. To the ex- tent that these recommendations have in- fluence, it is important that their underly- ing predictive assumptions be based on facts. Presentence investigators should tee informed of base rates of characteristics (such as substance abuse or truancy that 195 cies for those decisions. They do so more often on the basis oftheir own knowledge and experience than on the basis of statis tical analysis of information about repre sentative offenders. Because of the grow ing body of data and research on offenders, there are considerable oppor tunities for expanding ant! improving the bases for decision making in terms of predictors of specific offenders' careers. Exploiting such opportunities does not preclude the pursuit of other criminal justice objectives, such as assuring ap pearance at trial or imposing deserved punishment. Decision makers can com bine prediction-based classifications with aclditional information- ties to the com munity, victims' preferences, seriousness ot tne crime, or amenability to various forms of correctional treatment- that is pertinent to those other objectives. In concert with these other objectives, at tempts to reduce crime through incapac itation of offenders could be improved by familiarity with the accumulating knowI edge about offenders' criminal careers. That knowledge would inform decision makers of typical criminal careers and of the indicators of high- and low-risk of fenders. It is likely that the research knowleclge will confirm some of their current beliefs and practices. But it is also likely that some factors they believe to be important as predictors of offending will be found not to be so, while other factors they may not view as salient will be are often used to distinguish among of- shown to be important. fenders. They should also be in foe of Because decision makers' predictions correlations between those characteris- about offenders' criminal careers are not tics and the frequency of serious offend ~ng. CONCLUSIONS Criminal justice decision makers use knowledge about criminal careers both in making specific decisions affecting indi vidual offenders and in establishing poli _ routinely recorded, it is impossible to measure directly the improvement that could be achieved by giving decision makers statistical predictions about the offenders who come before them. Indi- rect evidence suggests that with available statistical scales, gains in crime control efficiency through selective incapacita- tion would be modest at best a ~10

196 percent reduction in robberies by adults, for example, with an increase of 1(~20 percent in the number of convicted rob- bers who are incarcerated. Similar crime control through general increases in in- carceration would require substantially higher increases in prison populations. The gains from selective incapacitation are limited in part because decision mak- ers already invoke many of the offender characteristics that figure prominently in the scales. As criminal career research progresses and the quality of available data improves, the crime-control effi- ciency of selective incapacitation can be expected to improve somewhat. As progress continues and new predic- tion scales emerge, there will be a ten- dency to take scales developed in one jurisdiction and use them in another. to use scales developed for one stage of the criminal justice process at another stage, and to take scales developed at one time and use them at a later time. It is impor- tant to recognize that such transfer must be done with great caution and with val- idation and recalibration in each new set- ting. The characteristics of convicted of- fenders at the sentencing stage in one state may be considerably different from those in another state. Hence, the factors that distinguish high-risk offenders in one state may well be different from those in the other state. Similarly, the factors that distinguish high-risk prisoners who have already been filtered through multiple stages of the criminal justice system (probably through several occurrences of recidivism) are likely to be different from those that distinguish high-risk arrestees being considered for pretrial release. Fur- thermore, since there may be changes over time in both the environment asso- ciated with offending and the selection processes within the criminal justice sys- tem, predictors even within a particular jurisdiction must be periodically recali- brated. CRIMINAL CAREERS AND CAREER CRIMINALS Since prediction scales are potentially vulnerable to variation by jurisdiction, by stage of the criminal justice system, and over time, successful validation in several jurisdictions can increase confidence in scale transferability by identifying key variables that seem to have consistently high predictive value, by helping to re- fine data collection forms and variable definitions, and by developing initial can- didate scales that can be adapted in any new setting. It is crucial, however, that every scale be tested, validated, and recalibrated at each decision stage and jurisdiction in which it is to be used. In choosing offender characteristics to be incorporated in prediction scales, it is important to consider the ethical appro- priateness of any candidate variable. In general, the seriousness of the instant offense is the most acceptable predictor, followed closely by a record of prior con- victions. Any other characteristics that are considered for use should have theoreti- cal import and validity, reflect blamewor- thiness, bear a strong predictive relation- ship to a criminal career, and should take account of other social concerns such as equal protection. Other ethical issues in- volve concern about errors, especially when prediction is used to increase indi- vidual punishment, but also when it is used to decrease some individuals' pun- ishment. In all such prediction, ethical questions must be addressed carefully and must be resolved locally. In the use of prediction scales, the validity of data on the predictor variables is a crucial concern. Records available to the criminal justice system raise many questions involving the use of such data. First, some variables used as predictors- such as drug use or juvenile convic- tions may be poorly or ambiguously re- corded in the information sources avail- able. If such data are to be used, information systems will have to improve the accuracy of their records, avoiding

USE OF CRIMINAL CAREER INFORMATION IN DECISION MAKING errors both of omission ant! of commis- sion. In considering the variables and records to be used as predictors, it is also important to anticipate reactions as a con- sequence of such uses. If information about drug involvement is used as a pre- dictor of offending-as current informa- tion suggests it should-one must con- sider whether such use would inhibit individuals from participating in drug treatment programs. Addressing these concerns may require continuing to re- strict the source of information about cer- tain variables: for example, to obtain cer- tain information only from criminal justice sources rather than from therapeu- tic agencies. Juvenile records are of particular con- cern because of the bifurcation between the adult and the juvenile criminal justice systems. Records of juvenile adjudica- tions are typically unavailable to the adult criminal justice system, presumably to avoid lifetime stigmatization as a result of some minor juvenile escapades. While that principle is certainly reasonable for individuals whose juvenile involvement 197 is indeed minor and especially for those who do not persist into an adult criminal career, the bifurcation floes not seem rea- sonable for juveniles whose delinquency careers are serious and who persist into serious adult offending. Thus, while juve- nile records should continue to be pro- tected from general public access, the adult criminal justice system should have access to juvenile records of at least those offenders arrested as adults on a felony charge. There is a clear capability in the cur- rent state of the art to develop predictors that are better than chance, but it is still unclear how much better than current practice statistically assisted prediction can be. One can expect currently avail- able scales to produce only limited im- provements to existing decision practices in terms of crime control. It is reasonable to expect, however, that future research will improve criminal justice decisions, at a minimum by highlighting additional salient predictor variables and by point- ing to variables that are often used but are incleed weak.

Next: 7. An Agenda for Future Research »
Criminal Careers and "Career Criminals,": Volume I Get This Book
×
Buy Paperback | $114.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

By focusing attention on individuals rather than on aggregates, this book takes a novel approach to studying criminal behavior. It develops a framework for collecting information about individual criminal careers and their parameters, reviews existing knowledge about criminal career dimensions, presents models of offending patterns, and describes how criminal career information can be used to develop and refine criminal justice policies. In addition, an agenda for future research on criminal careers is presented.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!