National Academies Press: OpenBook

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

Chapter: 5. The Rand Inmate Survey: A Reanalysis

« Previous: 4. Co-offender Influences on Criminal Careers
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 161
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 162
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 163
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 164
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 165
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 166
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 167
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 168
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 169
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 170
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 171
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 172
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 173
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 174
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 175
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 176
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 177
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 178
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 179
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 180
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 181
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 182
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 183
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 184
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 185
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 186
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 187
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 188
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 189
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 190
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 191
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 192
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 193
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 194
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 195
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 196
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 197
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 198
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 199
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 200
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 201
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 202
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 203
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 204
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 205
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 206
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 207
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 208
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 209
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 210
Suggested Citation:"5. The Rand Inmate Survey: A Reanalysis." National Research Council. 1986. Criminal Careers and "Career Criminals,": Volume II. Washington, DC: The National Academies Press. doi: 10.17226/928.
×
Page 211

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.

The Rand Inmate Survey: A Reanalysis Christy A. Visher In 1982 the Rand Corporation released its findings from a 1978 survey of jail and prison inmates and presented provoca- tive information about the individual of- fencling patterns of criminals. Rand's "second inmate survey," as it is called, involved nearly 2,200 inmates in three states who completed detailed] question- naires about the variety and intensity of their criminal activity. Analysis of these self-report data re Christy A. Visher is research associate at the National Research Council, National Academy of Sciences. She prepared this paper while a National Research Council Fellow at the National Academy of Sciences in 1983. The data used in this paper were made available by the Inter-university Consortium for Political and Social Research in Ann Arbor, Michigan. The data were originally collected by the Rand Corporation of Santa Monica, California. Neither the original source or collectors of the data nor the consortium bear any responsibility for the analyses or interpre- tations presented here. The author would like to thank Alfred Blumstein, Jeffrey Roth, and Douglas Smith for many helpful comments and suggestions in preparing this paper and Allan Abrahamse, Jan Chaiken, Peter Green- wood, and Charles Wellford for providing addi- tional data, documentation, and assistance. ~6~ vented that the distribution of the annual number of crimes an offender commits, often referred to as lambda (A), is highly skewed. Most of the inmates in the Rand survey reported small values of A, about five crimes per year, for most crime types. Some individuals, however, committed crimes at very high frequencies more than 100 crimes per year. These results suggest that most criminals, including the majority of those who are incarcerated, actually commit few crimes. High-rate offenders make up only a small propor- tion of the inmate population, but they may account for most of the crime prob- lem. This finding makes it particularly desirable to identify them. In one ofthe Ranc! reports based on the survey data, Varieties of Criminal Behav- ior, Chaiken and Chaiken (1982a) cIassi- fiec] the surveyed inmates into 10 groups according to the combination of crimes in which they engaged. One important re- sult of their research was the identifica- tion of a single category of serious crimi- nals, whom they designated as "violent predators." These offenders engaged in assault, robbery, and (lrug dealing at very

162 high rates, but they also committed prop- erty crimes at high rates. In fact, these "violent predators" committed more bur- glaries and other thefts than the criminals who specialized in those crimes. Chaiken and Chaiken concluded that these partic- ular offenders are especially troublesome and become entrenched in a deviant life- style in their juvenile years. The extreme skewness in offending fre- quencies ant] the identification of a small group of violent predators have intensi- fied interest in iclenti*ing high-rate, seri- ous offenders. If the most serious offencI- ers can be distinguished with information about their patterns of behavior and indi- vidual characteristics, the criminal justice system could become more efficient in identifying the most appropriate cancli- dates for long periods of incarceration. The Rand study made an important con- tribution to this effort by using self- reported information from the inmate sur- vey to identify serious offenders. Chaiken and Chaiken (1982a) showed that per- sonal factors and life-styTe characteristics, including persistent drug use, certain types of juvenile criminal involvement, and unstable employment, were strongly related to a violent, predatory pattern of offending. Other types of offending groups were similarly distinguishecl by particular observable behavior patterns and demographic attributes. The researchers at Rand also went a step further and attempted to translate findings about the characteristics of high- rate offenders into a policy instrument that could be used to guicle decisions in the criminal justice system. One sug- gested approach for addressing simulta- neously the problems of prison crowding and high aggregate crime rates is to em- phasize incarceration for the particularly serious high-rate offenders and to Reemphasize it for the others. Another Rand report (Greenwood, 1982) exam- inec3 the possibilities and consequences CRIMINAL CAREERS AND CAREER CRIMINALS of using this strategy selective incapac- itation as a specific policy in sentencing convicted offenders. Using a simple scale of seven variables that correlated with high annual offending frequencies, Green- wood estimated that a particular selective incapacitation policy could reduce rob- bery rates by 20 percent without increas- ing the prison population in California. As a crime control strategy, the idea of selective treatment of some offenders is not new. The concept of"predictive sen- tencing" has a long history (for a review, see Morris anct Miller' 1985), and it un- derlies the common use of risk-factor scales in (recisions regarding parole re- lease (see Gottfrecison and Gottfredson, this volume). The Rand research has en- hancect the potential value of predictive sentencing because of the skewness of the reported distribution of A. It has also generated considerable controversy. The criticisms that have been directec! at the Greenwood report and at the Ranc! study in general have both methodologi- cal and ethical elements. Some critics argue that the analysis is methoclologi- cally flawed and that Rand's sample of prisoners is not representative of the con- victec3 offenders judges have to sentence. Others are skeptical of the truthfulness of inmates' reports concerning the crimes they hac! committed. Observers are also concerned that most of the variables in the seven-point scale are baser! on self- report rather than official data and would be much less reliable if based on official records. Still others regard the variables involved as inappropriate as a basis for sentencing in any event. These and other criticisms are reviewed in a later section of this paper. Criticism of the Rand results has been stimulated by the extensive public atten- tion the seven-point scale has received. Some state legislators introduced bills in 1982 and 1983 to implement selective incapacitation as part of new sentencing

TTIE RAND INMATE SURVEY: A REANALYSIS policies (see Blackmore and Welsh, 1983~. Some police and prosecutors may already be using the scale informally to guide their decisions. An experimental program in Illinois is testing the predic- tive accuracy of the Ranc3 scale, along with other types of guidelines, in identi- fying offenders who are likely to recicli- vate. These actions have raised serious concerns that the results of this single study, which has a number of readily identifiable technical flaws (see Cohen, 1983) ant! which has not been subjectec] to internal or external validation, couIcI be implemented widely in making decisions regarding individual liberty. Thus, an intensive review of the Rand study is necessary. This paper provides a first internal validation basect on an exten- sive reanalysis of the actual inmate re- sponses to validate the findings and test their robustness to variations in the ana- lytic procedures used. An external vaTicia- tion using different settings is also neces- sary to assess the generalizability of the Ranc] results to a new sample of inmates and to samples of convicted offenders who are not in prison, but that test is beyonc! the scope of this effort. Three interrelated objectives are cen- tral to this reanalysis. The first objective is to validate the reported estimates of A and to assess the sensitivity ofthose and other findings to the interpretation of ambigu- ous ant! incomplete survey responses, ar- bitrary choices in constructing variables, treatment of missing clata, and decisions regarding scale development. The sec- ond objective is to examine the predictive accuracy of the seven-point scale in the three states, for specific crime types, and in other subsamples. The third objective is to reevaluate the reported incapacita- tion effects in light of the reestimation of A and reconstruction of the prediction tables. Data for the reanalysis were obtained from a machine-readable, public-use tape 163 of the inmate responses, supplied by the Inter-university Consortium for Political and Social Research, which maintains a data archive for the research community. Data obtained directly from RancT pro- vided adclitional cletail on how the ana- lysts translated the survey responses into the variables used in their analyses. With the generous help of the Rand research- ers, every effort was macle to determine Rand's analytic procedures. Copies of cocking manuals and computer source codes were studied, and persons at Rancl who were familiar with the analysis were consultecI. This reanalysis is limited to two key findings in the Rand reports: the esti- mates of annual individual offending fre- quencies, A (Chaiken ant! Chaiken, 1982a), and the use of the survey data to clevelop a prerliction instrument to iclen- tify high-rate offenders (Greenwood, 19821. Robbery ant! burglary offenses are the exclusive focus because of the prom- inence they received in the Rand reports and because of their prevalence among the sampler] prisoners. The remainder of this paper is orga- nizecl into three major sections. First, the purposes and general methods of Rand's second inmate survey and specific fincI- ings reported by Chaiken and Chaiken and by Greenwood are summarized. Pub- lished critiques of the Rand studies are also reviewed in this section. In the sec- onc1 section the results of the reanalysis are presented and compared win Rand's publishecl finclings. In the final section major findings and conclusions are presented. THE SECOND RAND INMATE SURVEY The Rand Corporation's 1978 survey of inmates extended previous work at Rand on studies of incarcerated offenders. In an exploratory study Petersilia, Greenwood,

164 and Lavin (1977) conducted extensive in- terviews with 49 convicted robbers in California prisons. Rancl's "first inmate survey" (Peterson and Braiker, 1981) was a self-administered questionnaire given to 642 prison inmates in California. The findings from both studies indicated that most inmates committee! few crimes per year and that a small group reported much higher frequencies of offending. The researchers considerecl their findings preliminary because the information on individual offending frequencies was im- precise, serious offenders were overrepre- sented in the sample relative to sen- tenced offenders, and only one state was involves] in the studies. Thus, a third, more intensive research project was cle- signed. Data and Methods The sample for the second inmate sur- vey, actually the third research project, covered three states, California, Michi- gan, and Texas. The sample was drawn to represent a typical cohort of incoming inmates for those states; a weighting scheme was used in which "each inmate was given a sampling weight proportional to the inverse of the length of his prison term" (Peterson et al., 1982:541. In addi- tion, to obtain a range of severity among the conviction offenses, inmates from both prisons and jails were sampled. Re TABLE Type CRIMINAL CAREERS AND CAREER CRIMINALS placement procedures were used to re- duce the usual problems of nonresponse bias. ESee Peterson et al. (1982) for other details of the sampling design, site selec- tion, and pretesting procedures.] The inmates selected for the study were asked to complete a cletailect ques- tionnaire that elicited information about their juvenile criminal behavior, aclult criminal behavior in the period (up to 2 years) prior to the arrest leading to their current incarceration, past and recent use of illegal drugs and alcohol, as well as information concerning employment his- tory, attitudes, and demographic clata. The survey was not anonymous so that official record data, which were collected on all prison inmates, couIc3 be matched to the inmates' self-reports. More than 2,500 inmates actually completed the questionnaire, but jail respondents in Texas were excluded from the analysis because, unlike jail inmates in over states, they were predominantly sen- tenced offenders awaiting transfer to prison. The final sample consisted of 2,190 inmates. The distribution of the 2,190 prison and jail inmates from the three states is shown in Table 1. Given the focus on robbery ant! bur- glary in the reanalysis, of particular inter- est are the inmates who reported commit- ting robbery or burglary during the 1- to 2-year period before they were arrested for their conviction offense, referred to as Distribution of Sample Across States, by Type of Institution and Crime Total Survey Robbers Burglars State PrisonJail Prisonfail PrisonJail California 357437 16894 182163 Michigan 422373 15466 174112 Texas 6010 1450 2520 Total 1,380810 467160 608275 NOTE: Data for robbers and burglars were computed as part of the reanalysis. Offenders in the two groups are defined by their reports of whether they committed any robberies or burglaries during the measurement period. Some individuals are included in both groups. SOURCE: Chaiken and Chaiken (1982a:6).

THE BAND INMATE SURVEY: A REANALYSIS the "measurement period." Table 1 shows that Texas had somewhat fewer inmates who reported committing rob- bery (24 percent) than either California (33 percent) or Michigan (28 percent). Respondents who reported committing burglary were more prevalent and were distributed more evenly among the three states 43 percent, 36 percent, and 42 percent in California, Michigan, and Texas, respectively. Many potential sources of error exist in a survey of this type. The most readily apparent systematic error arises from members ofthe sample refusing to partic- ipate; those nonresponclents couIc3 well be different in their crime patterns from those who were willing to respond. To correct partially for this potential source of bias, a "replacement respondent" was selected for each sampled prison inmate prior to the survey's administration. The replacement was matched with the sam- pled inmate on several criteria, including age, record, and conviction offense. The actual response rate varied consid- erably across states and type of institu- tion. In jails in all three states, the re- sponse rate averaged 70 percent. In Michigan and California prisons the rate was 49 percent and in Texas prisons, 82 percent. Replacement respondents in all three states were asked to complete the survey, but the replacement data from Texas were not used because of the Tow number of refusals among the main sam- ple. After including the replacements, Peterson et al. (1982:viii) concluded that "no statistically significant differences were found between responding and nonresponding inmates in any Michigan or Texas prisons, in terms of age, race, record or conviction offense." In some prisons in California, Chicano inmates were less likely than others to participate. Inmates with reading problems were underrepresentecT in all three states. Over major sources of error in surveys eliciting self-reportec! information are un 765 reliable responses and nonvalid survey instruments. Researchers at Rand carried out extensive analyses of these problems (Marquis, 1981; Chaiken and Chaiken, 1982a; Peterson et al., 19821. Two design strategies were built into the survey for later use in the analysis of reliability: redundant questions were asked within the survey, and 250 respondents were retested 1 week later. Chaiken ant] Chaiken (1982a:Appendix B) relied on the first approach and developed mea- sures of the internal quality and external reliability of the survey responses. The internal checks inclucled looking for cor- rect skip patterns, consistent answers, minimal confusion, and few omitted questions. The external checks relied on comparisons between each inmate's offi- cial record and his responses to 14 self- reportecl items (e.g., conviction offense, arrest incidents, ant] prior prison terms). The two measures were strongly corre- lated in each state. Chaiken and Chaiken concluclec! that most individual characteristics and be- havior patterns, including age, race, con- viction offense, and reports of crimes committal, were unrelated to the quality and reliability of inmates' responses. About 83 percent of the inmates "passed" the internal quality test, whereas only 56 percent achieved a similar level of exter- nal reliability.) Scattered evidence sug- gested that respondents who gave consis- tent and reliable answers were less likely to report very high offending frequencies and less likely to deny committing crimes. Finally, key regression analyses were carried out with and without inconsistent 1"Failure" is defined as having more than 20 percent "bad" indicators on the external or internal reliability measures (see Chaiken and Chaiken, 1982a:9, 222-239). Other data reported suggest that the low level of external reliability is partly the result of incomplete of ficial records, especially juve- nile records (p. 229~. Inmates often reported juve- nile convictions or incarcerations that were not found in their records.

166 or unreliable respondents (42 percent of sample), anal no "meaningful differences" were found between the two analyses (Chaiken and Chaiken, 1982a:9), although the actual results were not reported. Purposes of the Rand Study The Rand survey was designed to achieve a number of purposes (see Peter- son et al., 19821. One major purpose was to gather information on incliviclual pat- terns of criminal behavior types of crimes committed, degree of speciaTiza- tion in crime types, and changes in crim- inal patterns over time. Questions were asked about juvenile criminal activity and criminal behavior during the 6 years prior to incarceration to explore hypotheses about whether offenders progress through stages of increasing crime seri ousness. A large section of the survey was de- voted to obtaining offenders' estimates of the number of times they committee] each of 10 crime types2 during the measure- ment period. Estimates of annual offencI- ing frequencies, A, have been calculated by other researchers using a variety of techniques based primarily on inferences from arrest records (e.g., Greenberg, 1975; Blumstein and Cohen, 1979), but no broad consensus has yet been reached in these estimates (for a review, see Cohen, 19831. The Rand survey was the first to use a self-report technique to ob- tain annual estimates of )< for a group of known adult offenders. The use of a self-administerec] ques- tionnaire also permitted researchers to collect richer data on offenders' personal 2The 10 crimes that were included in the ques- tionnaire were burglary, business robbery, personal robbery, assault during robbery, other assaults, then, auto theft, forgery/credit card swindles/bad checks, fraud, and drug dealing. The specific word- ing of the questions is available in Chaiken and Chaiken (1982a:19-201. CRIMINAL CAREERS AND CAREER CRIMINALS characteristics than can typically be found in official records. Extensive infor- mation was gathered on (1) criminal ex- periences at young ages; (2) use of illegal drugs as a juvenile and as an adult; (3) adult offender histories, including arrests, convictions, and incarceration; and (4) life-style characteristics, such as marital status, employment record, and geo- graphic mobility. The survey also con- tained a number of questions about atti- tudes toward crime. The researchers at Rand believed that the self-report data on personal characteristics and annual of- fencling frequencies might help to distin- guish different types of offenders and, particularly, to identify the serious "ca- reer criminals." Such information could be helpful to criminal justice agencies in making decisions regarding sentencing, parole, work release, or (lrug treatment programs. The Rand Results The findings from the Rand study ap- pear in several reports (Petersilia ant! Honig, 1980; Rolph, Chaiken, ant! Houchens, 1981; Chaiken ant] Chaiken, 1982a; Greenwood, 19821. Three results are especially relevant to policy decisions in the criminal justice system: (1) esti- mates of A and its skewed distribution (Chaiken and Chaiken, 1982a), (2) the development of an offender typology and the use of a multivariate approach to clis- tinguish among types of offenders (Chaiken and Chaiken, 1982a), and (3) the identification of high-rate offenders using self-reportecl information (Green- woo(l, 1982~. Rancl's summary statistics for the annu- alizecl incliviclual offending frequencies are shown in Table 2. The statistics for each crime type are based on only those inmates who reported committing that crime. The distribution of A, as noted, is highly skewed. More than one-half the

THE RAND INMATE SURVEY: A REANALYSIS TABLE 2 Estimates of A for Respon- clents Who Reported Committing the Crime Crime Type Median Value at the 90th Percentilea Burglary Robbery Assault Theft Forgery and credit cards Fraud All except drug dealing 5.45 232 5.00 87 2.40 13 8.59 425 4.50 5.05 14.77 206 258 605 aTen percent of the respondents who commit the crime commit it at or above the rate indicated. SOURCE: Chaiken and Chaiken (1982a:44). inmates who committed robbery or bur- glary in the measurement period dic3 so at rather Tow rates about five crimes per year. On the other hand, the worst 10 percent committee] robbery and burglary at the rate of two to four crimes per week, or 20 to 40 times as frequently as the median offender. In this highly skewed situation, the mean does not accurately represent the central tendency of such a distribution. Further, the mean is ex- tremely sensitive to the values of the few 767 offenders in the right tad] of the clistribu- tion anti, therefore, the median rate is preferable for estimates of a "typical" offencler's crime rate. The survey also providecI the data for Chaiken and Chaiken's development of an offender typology. They found that inmates could be categorized according to the combination of crimes they com- mit, such as robbery and assault or bur- glary and drug dealing. In Table 3 the medians and 90th percentile values of A are compared for offender types that in- clucle robbery or burglary as one of the defining crimes. These six types consti- tute 62 percent of the inmate sample (Chaiken and Chaiken, 1982a:271. As seen from Table 3, violent predators com- mitted robbery and burglary at very high frequencies; however, the me(lian A was 9. The most active 10 percent in this group reportecITy committed at least 516 burglaries per year, whereas the 90th per- centile of the "burglar-(leaTers" (who commit burglary and other property crimes and sell cirugs) committed 113 burglaries per year. The violent preda- tors, especially the worst 10 percent, thus appear responsible for the majority of robberies ant] burglaries committed by the inmates in the Rand survey. Realizing TABLE 3 Estimate of A for Robbery and Burglary for Six Offender Types Percent Robberya A Burglary A Offender Type of Sample Median Mean 90th Pet. Median Mean 90th Pet. Violent predatorsb 15 9 70 154 9 172 516 Robber-assaulters 8 5 50 141 5 69 315 Robber-dealers 9 4 32c 87 14 122 377 Low-level robbers 12 2 10 13 4 48 206 Burglar-dealers 10 4 42 113 Low-level burglars 8 2 36 105 Others 38 aIncludes both business robbery and personal robbery. bThose who commit robbery, assault, and drug dealing concurrently. COne outlier has been removed. Includes "mere assaulters," property and drug offenders, low-level property offenders, drug dealers, and about 13 percent who did not report committing any of the crimes studied. SOURCE: Chaiken and Chaiken (1982a:27, 219).

168 the impact of these serious offenders on the crime problem, the Rand researchers used several techniques to identify them. Using a multivariate approach Chaiken and Chaiken founcI that some self-report- ed information could distinguish violent predators from other inmates. These of- fenders were often young people with a history of serious juvenile criminal activ- ity, including initiation of delinquent be- havior before age 16, involvement in both violent and property crimes, frequent use of illegal drugs, and multiple commit- ments to state juvenile institutions. They were generally unmarriecl, unemployed, and extremely heavy drug users, often at costs exceeding $50 per day for heroin. A regression mode! using these variables explainec! 35 percent of the variance in annual offending frequencies. However, many inmates predicted to be high-rate robbers with this moclel actually reported committing no robberies at all. Greenwood (1982) independently at- tempted to identify the high-rate offend- ers with a simple, seven-point scale. He selectee] seven variables (six self-report and one official record variable available only for prison inmates-see below) that correlated fairly well with high annual robbery and burglary offending frequen- cies anct whose use might be appropriate for sentencing purposes. The resulting additive scale (variables were scorer! as 1 or O clepending on the presence or ab- sence of the attribute) could be used to identify high-rate offenders. Inmates were classified as low-rate (scoring O or 1), medium-rate (scoring 2 or 3), or high- rate (scoring 4 or more) offenders. The mean annual offending frequencies were reported to stiffer sharply across these groups. For inmates in Califomia, the respective mean As for robbery were 2.0, 10.1, and 30.8. This pattern is consistent for robbery in the other states and for burglary, but the group differentials are widest in CaTifomia. CRIMINAL CAREERS AND CAREER CRIMINALS Variables Used in Scale to Distinguish Inmates by Individual Crime Rates Convicted previously for same charge (official criminal record; prison inmates only) Incarcerated more than 50% of preceding 2 years (self-report) Convicted before age 16 (self-report) Served time in state juvenile facility (self-report) Used drugs in preceding 2 years (self-report) Used drugs as a juvenile (self-report) Employed less than 50% of preceding 2 years (self-report) Using the model of incapacitation de- velopecl by Avi-Itzhak and Shinnar (1973), Greenwood estimated the poten- tial crime control effects of increased sen- tences for the identified high-rate offend- ers. For California, he reported that a policy of sentencing predicted high-rate robbers to S-year terms and all other rob- bers to 1-year jail terms could reduce the robbery rate by a maximum of 20 percent, without increasing the prison population. Such a strategy does not work as well for burglary. (A detailed analysis of the sev- en-point scale and its use in identifying high-rate offenders is presented in a later section in conjunction with the reanalysis of the Rand data.) Criticisms of the Rand Study Because of the provocative policy im- plications of the Rand results, the inmate study has received a considerable amount of attention, and not all of it has been positive. Some researchers have raised moral objections to the mechanical use of any such scale for determining sentences. Others argue that the findings are flawed and therefore policy proposals should not be based on Rand's results. Ethical Concerns Ethical concerns emerged largely in response to the analyses presented in the Greenwood report. The Rand study has also mobilized arguments about se

THE RAND INMATE SURVEY: A REANALYSIS lective incapacitation as a sentencing phi- Tosophy and, especially, the use of ex- plicit predictions in sentencing. These debates have become quite vigorous. The issues are discussed only briefly here, however.3 One of the most frequent objections to the Greenwood report concerns the se- lection of variables for the seven-point scale. In particular, critics argue that some of the variables in the scale are past behaviors or social characteristics that cannot be changed. Employment status, drug use, and juvenile criminal history account for five of the seven variables. Retributivists and others have pointed out that using these criteria as a basis for sentencing is contrary to the widely accepted "just deserts" philosophy, whereby differences in sentences are based on the seriousness of the convic- tion offense. Greenwood (1982:Table 4.11) anticipated these criticisms and tested his scale without three of the most "objectionable" predictors Juvenile drug use, recent drug use, and recent employ- ment history). The limited scale, how- ever, was less effective in distinguishing high-rate offenders from medium-rate of- fenders compared with the full seven- variable scale. A more fundamental ethical objection has been raised to the concept of sentenc- ing offenders according to a prediction of their future behavior. Because of its ex- plicit relationship to sentencing policy, Greenwood's analysis was the recent tar- get ofthese critics. Some critics argue that this type of sentencing policy would vio- late principles of fairness and "just deserts" (von Hirsch, 1976, 1981) ant! others question whether future high-rate 3For other discussions of these topics, see Dershowitz (1973, 1974), Cohen (1983), von Hirsch (1976, 1981, 1984), Floud and Young (1981), Hinton (1982), Moore et al. (1984), and Morris and Miller (1985~. |69 offenders can be accurately iclentifiect (Blackmore anci Welsh, 1983; van Hirsch and Gottfrecison, 1984~. Any classification system is likely to misidentify offencI- ers-classifying some low-rate offenders as high-rate ("false positives") and some high-rate offenders as Tow-rate ("false negatives". The expected level of error is totally unacceptable to some (von Hirsch and Gottfredson, 1984) but considered reasonable within some definitions by others (Morris and Miller, 19851. Green- wood also raised some of these same issues in his report, but he differs from his critics in believing that these ethical (and some empirical) problems are only limi- tations on the usefulness of selective in- capacitation and not barriers to its poten- tial use. In summary, the Ranc3 reports have intensified the ethical debate about selec- tive incapacitation and predictive sen- tencing. Any resolution will involve hard choices about acceptable error rates and appropriate prediction instruments. Em- pirical information about the predictive capability of different scales may help to inform those choices for some. Empirical Concerns Empirical concerns regarding the Ranc! study cover a wide range of issues, in- cluding reliability of the inmates' re- sponses, construction of the seven-point scale, and the robustness of the incapaci- tation effects to variations in the model. The following discussion reviews pub- lishec! critiques and raises some addi- tional concerns. The examination of po- tential limitations to Rancl's results provides direction for the reanalysis that follows. First, some observers have questioned the reliability of the inmates' self-report- ecl responses, especially the data used to estimate the annual number of crimes, A, an inmate committed prior to his incarcer

170 ation (Blackmore and Welsh, 1983; von Hirsch and Gottiredson, 1984~. The many sources of error in self-report methods have been widely discussed (e.g., GoIc3, 1966; Farrington, 1973; Reiss, 19731. The Rand study presented further problems because of its sample-convicted offend- ers. Some inmates could have concealed crimes they committed, and others might have exaggerated their criminal activities, ant! these practices could contribute to the observed skewness in the reporting of offending frequencies. The Ranc3 finding that a small group of inmates reported committing hundreds, or even thousands, of robberies and burglaries a year has led critics to speculate that some respondents inflated their illegal behavior to appear "tough" or important (von Hirsch ant] Gottfredson, 19841. The opposite type of response error, concealment, is also plau- sible, especially since between 24 ant] 36 percent of all convicted robbers in the sample denied committing any robberies in the measurement period (Greenwood, 1982:Table 4.1~. The accuracy of estimates of A also depends on an assumption of stable of- fending patterns over time (Cohen, 19831. But criminals may operate erratically, committing many crimes in a short period and then ceasing their illegal activities for a while. If the "crime spurting" phenom- enon describes even a minority of Rand's inmate sample, the estimates of annual offending frequencies might well be in- flated (Cohen, 19831. Second, the criticism directed against the Greenwood scale was even more vig- orous. The variables in the seven-point scale were mostly self-report measures, and the only scale variable that was con- structed from official record information was whether the inmate had a prior con- viction for the same offense. Some critics were concerned about the availability of necessary information if predictions re- garding future criminal behavior were to CRIMINAL CAREERS AND CAREER CRIMINALS be made (Blumstein, 1983; Cohen, 19831. Of course, Greenwoocl's scale, basecl on self-reportec3 information, was only sug- gestive of the kinds of factors that may be predictive of high-rate offending. If the scale was to be used operationally, the needed information would have to come from inclependent sources, such as official records or other inquiries, like those re- flected in presentencing investigations. But the use of official records invariably involves some clecay in reliability be- cause of missing records, recording er- rors, ant! other mistakes. Data from inde- penclent sources are also likely to be incomplete and less helpful because some information, such as drug use, is not gathered consistently. Third, the treatment of missing data in the scale is another source of concern. Each of the variables in this scale was coded 1 or O to indicate the presence or absence of the attribute, and missing in- formation on any scale item was also codecl 0. However, for at least one of the scale variables- prior conviction for the same offense-the missing-data problem was systematic: official records were only available for the prison sample, and so all jail inmates were assigned a zero for this variable. In the analysis the past- convictions variable thus becomes a measure that distinguishes jail and prison inmates (Cohen, 19831. Since high-rate offenders are probably already more likely to be sentencee] to prison than to jail, this variable is more a "predictor" of who was sent to prison than of any other inmate characteristics. Missing data was a prob- lem for another variable, juvenile (lrug use; 14 percent of the respondents failed to answer the questions on this topic (Greenwood, 1982:52~. Fourth, the predictive accuracy of the seven-point scale turns out to be no better than that for other prediction instruments developed over the past 10 years. The final sample used in the prediction anal

THE RAND INMATE SURVEY: A REANALYSIS ysis was prison and jail inmates in the three states who were currently serving a sentence for a robbery or burglary convic- tion. Among inmates predicted to be high-rate offenders, only 45 percent (Cohen, 1983) actually were, according to estimates of their annual offending fre- quencies. Stated another way, 55 percent of the predicted high-rate group was in- correctly iclentifiecI. This level of"false positives" is close to the average false- positive rate (60 percent) reported in a review of other prediction studies (Monahan, 1981~. The scale floes much better among predicted Tow-rate offenc3- ers: the accuracy rate is 76 percent (Cohen, 19831. These differences are due in part to the different base rates of the two groups arbitrarily specified as the lowest 50 percent for the low-rate group and the highest 25 percent for the high- rate group. However, the data reported by Greenwood and reanalyzed by Cohen focus on overall accuracy rates for the entire analysis sample, and there has been no examination to date of whether the scaTe's predictive accuracy is consis- tent across states, crime types, and other important subgroups. Finally, another area of major concern about the results of the RancI study relates to the validity of the incapacitation effects reported by Greenwood (Blackwore and Welsh, 1983; Cohen, 1983; von Hirsch and Gottirecison, 19841. In her review of research on incapacitation, Cohen (1983) notes! that Greenwooc3's clevelopment of a prediction scale is based on retrospec- tive clata. The reporter! incapacitation ef- fects, therefore, clo not take into account the possibility that future rates of offend- ing might change (e.g., regress towarc! a mean) or that there might be a differential likelihood of terminating criminal activ- ity. Thus, the prospective accuracy of the seven-point scale in identifying high-rate offenders can only be judged with an appropriate longitudinal panel (resign. In other 171 fact, the use of retrospective data may lead to an overestimate of the crime-reduc- tion effects. Another serious validation problem is the lack of any test of the scale on an inclepen~lent sample. This is particularly important because a selective sentencing policy wouIc3 be applied to convicted of- fenclers, and predictive information in that population may be different from that in a sample of inmates (Cohen, 19831. Other research has shown that predictive accuracy for the initial sample for which a prediction scale is constructed tends to be greater than for a separate validation sam- ple (GottEredson and Got~redson, 1980; Farrington and TarTing, 1985~. Thus, re- ported reductions in aggregate robbery rates that are tied to any particular scale will diminish for new samples. However, Greenwood argues (1982:91) that, since his 0-1 prediction scale was not closely fitted to the inmate sample (as is the case with regression models), the expected shrinkage wouIc3 be less than with regres- sion weights. This characteristic of Bur- gess ((~1) scales, compared with closely weighted scales, is cliscussecI by Gottfrec3- son and GottErecison in Chapter 6. The state-specific results obtained in the crime control analysis for California and Texas also illustrate the sensitivity of the Rand findings to the population being studied. For California, it is reported that aggregate robbery rates could be reduced by 20 percent and burglary by 12 percent without any increase in prison population by using a selective sentencing strategy (Greenwood, 1982:79~. In Texas, how- ever, a similar sentencing policy would actually increase the robbery rate be- cause there are so few high-rate offenders (Greenwood, 1982:Figure 5.31. In summary, several critical reviews of the Ranc3 inmate study have raised impor- tant questions about the sensitivity of the results reported in Chaiken and Chaiken (1982a) and Greenwood (1982) to the in

172 terpretation of the survey responses, to variable construction, ant! to variations in the estimates of the parameters user! in the calculation of incapacitation effects. The reanalysis presented in the next sec- tion addresses some of these questions. REANALYSIS This reanalysis ofthe Rand inmate data involves many interconnected analyses that must be completed in a particular sequence. As the reader will quickly dis- cover, the complexity of the survey in- strument and Rand's novel use of the self-report data also complicated the rep- lication. Moreover, Rand's procedures were not always straightforward. As a re- sult, in some situations the Rand method was known but an alternative approach was chosen, for reasons that are explainer! as the reanalysis is presented. In other instances Rand's analytic procedures were followed as closely as possible. The initial task is to replicate Rand's calculations of incliviclual offending fre- quencies, A, which are the most important result in the Ranc! research, and which provide the underpinnnings for all the other major findings. Once A is recom- putect for robbery and burglary, the pre- cliction scale is reconstructed and mea- sures of predictive accuracy for various inmate subgroups (e.g., states, jail in- mates) are calculated. Finally, using the recomputed estimates of A and the pre- diction scale, the projected incapacitation effects on crime rates and prison popula- tions of selectively incarcerating robbers in California are reestimated. An impor- tant part of this reanalysis is to assess the sensitivity of the findings to changes in the model's input variables and to alter- native cut points for clefining the predicted low-, medium-, and high-rate groups. Estimating A from the Rand Data Ideally, estimates of A for a group of offenders coulcI be obtained by having a CRIMINaL CAREERS AND CAREER CRIMINALS representative sample of"active" crimi- nals keep daily logs of their criminal ac- tivities for an entire year. In the expected absence of that level of cooperation, a number of alternative methods for mea- suring unobservable behaviors have been clevelopecI. The RancT researchers chose ~' "self-report" approach, i.e., they asked inmates to answer a series of de- tailec] questions about the number of crimes they had committed in a defined "measurement period," a period of 1 to 2 years prior to the arrest that led to their current sentence. Inclivi(lual offending frequencies can be expressed as a fraction: the number of crimes committed (the numerator) di- videct by the number of years of "street time" (the denominator), which takes into account any time spent incarcerated dur- ing the measurement periocl. Chaiken and Chaiken (1982a:42) provide an exam- ple of how to calculate A for burglary for a respondent whose measurement period was 14 months, 5 months of which were spent in jail, and who reported commit- ting six burglaries during the 14-month period: the ~~sett-report A = (6 burglaries) t(12 months per year) -. (14 - 5) months] = 6 (12 . 9 months per year) = S.0 burglaries per year "Active" and "Inactive" Offenders Before turning to the computation of A, the inmate group that is the focus of the reanalysis must be defined. The relevant group consists of the inmates who re- ported committing robbery or burglary during the measurement period (see Ta- ble 11. In the analyses involving We sev- en-point scale, the relevant group is somewhat different: all inmates who re- portec! that their current incarceration was the result of a robbery (burglary) conviction, whether or not they reported committing robbery (burglary). Thus, the "convicted" group is not a subset of the

THE RAND INMATE SURVEY: A REANALYSIS TABLE 4 Percentage of Active Offenders: Inmates Convicted of Robbery or Burglary 173 Type of Involvement California Michigan Texas Convicted Robbersa Active 75.2 62.7 71.8 Inactive 21.9 34.7 28.2 Unknownb 2.8 2.7 Total N (178) (150) (117) Convicted Burglarsa Active 73.1 68.5 66.0 Inactive 25.6 31.5 34.0 Unknownb 1.3 Total N (160) (124) (203) NOTE: Active convicted robbers are those who admitted committing robbery during the measurement period, whereas "inactives" denied committing any robberies; the same distinction applies for convicted burglars. aThere were 56 inmates in the sample who were convicted of both robbery and burglary. To maintain consistency with Greenwood's (1982) procedure, they are treated in this table and subsequent analyses only as convicted robbers. bInmates' responses did not permit an unambiguous determination of whether they were active. SOURCE: Data were calculated by the author from the original survey responses. The total sample size is 932. The number of cases in subsequent tables, when the sample is all convicted robbers and burglars, varies from this number because of the omission of inactive respondents, missing data, and a small weighting factor used for Texas respondents. -~committing" group because some con- victed robbers (about one-quarter) denied committing any robberies during Me measurement period; the same distinc- tion applies to burglars. The latter group of convicted robbers and burglars are re- ferrec] to as "inactive" offenders. The dis- ~ibution of active ant] inactive offenders among convicted robbers and burglars is shown in Table 4. The percentage of convicted robbers and burglars who reportedly were inac- tive in Weir respective conviction offense types is astonishingly high. About 28 per- cent of Me convicted robbers and 30 per- cent of Me convicted burglars reported Mat they had not committed any robber- ies (or burglaries) in the past 1 to 2 years. These figures are similar to Rose re- ported by Rancl.4 A group of inactive con 40ne of the Rand reports presents comparable data indicating that the percentage of convicted robbers who are defined as active in California, Michigan, and Texas is 76, 64, and 72, respectively (Greenwood, 1982:42~. These percentages are very close to those obtained independently in the reanaly sis. For convicted burglars, however, Rand reports that 94, 91, and 88 percent, respectively, were active in the three states, rates that differ considerably from those reported in Table 4 by this author. The discrepancy is due to differences between Rand's and this author's definition of"active" and "inactive" burglars. In their Appendix A, Chaiken and Chaiken (1982a: 186-189) lay out explicit rules for determining whether a respondent is to be considered active in a particular crime type. But the rules are different for robbery and burglary because "non-bu- rglary robbery" (the definition of robbery used in most of the Rand analyses and adopted by this author) is a summary crime that combines business and personal robbery (see Chaiken and Chaiken:196~. Two vari- ables that are used in determining activity for bur- glary, CK7 and CK14 (see p. 197), have no equivalent for "non-burglary robbery." The robbery category for question CK7 was not considered specific enough to be used as an indicator of activity (J. Chaiken, 1984, personal communication). Since robbery and burglary offenses are the focus of this reanalysis, the definitions of activity and inactivity for the two crime types were made con- sistent. As a result of this decision, about 125 bur- glars (all with annual burglary rates of zero) who were defined as "active" by Rand's definition were defined as inactive in this reanalysis. This accounts for the reduced percentages of reportedly active burglars in Table 4, compared with Rand's figures.

174 YEAR BEFORE AR RESTED YEA R AR RESTED CRIMINAL CAREERS AND CAREER CRIMINALS INSTRUCTIONS FOR USING THIS CALENDAR ARE INCLUDED IN THE SURVEY Winter Spring :;nuary |Febr: | April ' Summer Fall August September October November December I May I June I July . . . January February _ Ma rch Apri I _ _ , I. May | June 1 1 August September STREET MONTHS l l ON THE CALENDAR | l . October | November | December | . , FIGURE 1 Calendar useciby all respondents in calculating street months. Source: Ebener (19831. victecI robbers (N = 42 of the 124 inac- tives) may have confused robbery with burglary since they clid report committing burglary. Of the inactive convicted bur- glars (N = 149), 11 reported committing robbery. It is also possible that all the "inactives" clid not commit those crimes or agreed to pleact guilty to that charge. Another interpretation is that these in- mates committed only one robbery (or burglary) during the measurement pe- riod, were promptly arrested, and did not count that offense in their report.5 The most plausible explanation, however, is that many of them simply field even though they knew that their official rec- ords were available to the researchers. Without any additional information, it is difficult to determine the truth in this situation; thus, in this reanalysis the esti- mate of A for the "inactive" inmates is zero, as in the Ranc3 analysis. Respondents may have been confused about whether to include the crimes they committed dur- ing their arrest month in their report, especially if they were incarcerated ifor most of that month (ei- ther prior to their arrest or after it). This sequence of questions caused many problems for some respon- dents. Chaiken and Chaiken (1982a) include a copy of the questionnaire in their report. , Determining Street Months To estimate A for the inmate sample, the Rand researchers hac] to establish a defined measurement period to facilitate the inmates' recall of their criminal activ- ity and other events before their incarcer- ation. It was also important to determine each inmate's "street time," referred to by Ranc] as "street months," which is the time when they were free to commit crime and the denominator used in caTcu lating A. Figure 1 is a copy of the calendar, completed by all responclents, that was user] to facilitate recall and determine "street months" for each inmate. Inmates were instructed to mark their arrest month in the year of their arrest with an X and draw a line through the remaining months in that year. (The months prior to an(1 inclu(ling the arrest month are called the "measurement period" and could range from 13 to 24 months.) Other months in which the respondent hac] been incarcer- ated were also markocl with an X. An in- mate's "street months" is the number of months he was on the street (not in prison or jail) an(1 able to commit crimes, which could be 1 to 24 months. Questions were then asked about

THE RAND INMATE SURVEY: A REANALYSIS events during the measurement period. This memory-recall technique was cho- sen because the amount of time inmates had spent in prison (or jail) since their measurement perioc] varier! within the sample. The distribution of inmates' (rob- bers and burglars) self-reported time served on their current sentence to the time of the survey provides an indication of recall time (see Table 5~. Three-quar- ters of the inmates had served less than 2 years at the time ofthe survey. Recall may have been more difficult for respondents who served longer, which coup] result in underestimates of A for this group. But they were presumably more serious crim- inals and may have committed more crimes than the others. Thus, the actual effect of recall time on estimates of A is confounclecI by the characteristics of the sample, but no attempt is made in this reanalysis to adjust A for any recall bias. Despite Rancl's procedures for stimu- lating recall, many inmates hac] trouble filling out their calendars (see notes about cocting decisions, Ebener, 1983:41~3~. The estimate of street months couIcl be obtained from four sources two places TABLE 5 Distribution of Self Reported Time Served on Current Sentence for Inmates Who Reporter] Committing Robbery or Burglary Time Serveda Percent Cumulative Percent 1-6 months 7-12 months 13-18 months 19-24 months 25-36 months 37-47 months 4 or more years 22.3 23.3 15.5 13.2 12.5 5.3 7.9 22.3 45.6 61.1 74.3 86.8 92.1 100.0 aMean = 20.6 months; median = 14.5 months. SOURCE: Data were calculated by the author from the original survey responses. The N is 1,052, which excludes 48 cases (4.4 percent) be- cause of missing data. 175 TABLE 6 Cross-Tabulation of Rand s Average Estimate of Street Months for Robbers and Burglars by the Difference Between Rancl's Maximum and Minimum Estimates (percentages; N= 1,235) Maximum Minimum 1-6 Average Estimate of Street Months 7-12 13-18 19-24 ~5 6-9 10 or more N 83.1 63.9 77.9 11.3 11.9 9.0 4.2 7.9 5.2 0.7 5.1 2.3 0.7 11.1 5.6 142 252 520 90.0 7.8 1.9 0.3 0.0 321 SOURCE: Data were calculated from data pro- vided by J. Chaiken, Rand Corporation. in the questionnaire, the calendar, ant! an estimate by survey editors (left blank if the responclent's answers were consistent and matched his caTenciar). Although the majority of inmates had no apparent prob- lems with this section of the survey, re- calling their activities still appears to have been a complicated cognitive task for many respondents. Estimating street months was a straightforward task when inmates were consistent in their answers. But, for the inmates who gave incomplete or ambigu- ous responses, street months had to be estimated differently. Ranc3 analysts re- lie(1 on e(litors to examine problematic questionnaires and to generate their own estimates of street months. Then, all pos- sible interpretations of the responses were used to obtain a maximum possible value and a minimum possible value for street months for each inmate (Chaiken and Chaiken, 1982a: 18~186~. An aver- age estimate was the mean of the two extremes. As seen in Table 6, for most respon- dents (78.8 percent) the street months estimate was unambiguous, and so the difference is zero. However, larger dis- crepancies between the maximum and

176 TABLE 7 Distributions of Final Estimates of Street Months for Inmates Who Reported Committing Robbery or Burglary: Rand and Reanalysis (percentages) Street Months Randa Reanalysis Less than 6 months 7-12 months 13-18 months 19-24 months Missing/unknown Total Mean 11.2 20.0 41.2 25.5 2.1 100.0 14.4 11.8 17.7 41.7 28.2 0.6 100.0 14.6 aRand's estimates were calculated from data provided by J. Chaiken, which gave Rand's esti- mates of maximum and minimum street months for each inmate. minimum estimates are more common when Rand's average estimate for street months (the mean of the two extremes) is 12 months or less. Such a value could occur only if the respondent had been off the street (most likely imprisoned) for some time in the year before the current incarceration began. That these larger discrepancies appear more often at the Tower end ofthe overall street-month dis- tribution is particularly disturbing be- cause A may be overestimated if the in- mate was actually on the street for less than 12 months. These differences be- tween Rand's minimum and maximum street months suggest that one source of unreliable data for some inmates may be in figuring months "not on the street," which the inmate was supposed to ex- clucle from street months. The Rand edi- tors and some respondents apparently of- ten disagreed on this point. Review of Ranks procedures raised concerns that the strategy of using mini- mum and maximum estimates for street months could result in misTeacling esti- mates of A, especially when the minimum was exceptionally Tow. For the reanalysis, the following set of rules was establishe for choosing a single estimate of an in CRIMINAL CAREERS AND CAREER CRIMINALS mate's street months from the four avail- able sources of informational 1. If the inmate gave consistent an- swers on the two questions, use the in- mate's response. 2. If the inmate gave conflicting an- swers on the two questions, but the Rand editor provided no corrected estimate, use the response from the second ques- tion.7 3. If the inmate gave conflicting an- swers, but the response to the second question was the same as both the editor's estimate and the estimate obtained from the calendar, use the inmate response to the seconc! question. 4. If the inmate's answers, the editor, and the information on the calendar were all in disagreement, use the editor's esti- mate. 5. If some disagreement existed be- tween the inmate's second response, his calenclar, an(1 the eclitor's estimate, but two of the estimates were in agreement, use the value on which there was some agreement. In Table 7 the estimates of street months calculates! in the reanalysis are compared with the Rand estimates, which are simply the averages of Ranc3's minimum and maximum estimates. De- spite the alternate analytic approach to 6A small residual category, which was not encom- passed by these rules, included cases in which the Rand editor found the inmate's calendar indeci- pherable (coded "unknown"), cases in which all four sources were missing (coded "missing"), and four cases that were treated as exceptions to the rules for various reasons. 7Of the responses to the two questions, Rand considered the second response (C10) less prone to error than the first response (C91. During routine checks for errors in all questionnaires, the Rand editors were more concerned about correcting er- rors in C10 than C9 (J. Chaiken, 1984, personal communication). This information was considered in establishing the coding rules for the reanalysis, which focused mainly on C10.

THE RAND INMATE SURVEY: A REANALYSIS the calculation of street months, the dis- tribution of the recomputed estimates is virtually identical to that generated by the Rand analysts. The average number of street months for inmates who reported committing robbery or burglary accorc3- ing to both estimates was 14 months. About 30 percent of the sample were "on the street" for less than a year, which means that those respondents spent some time in prison or jail during the measure- ment period. The ability to replicate Rand's esti- mates of average street months, however, should not overshadow the conceptual ant! methodological problems associated with this measure as the denominator of A. First, the section of the survey involv- ing the calendar was complicated, and respondents hacl trouble with the ques- tions and the caTenclar and gave inconsis- tent answers. Second, it was important to the Rand study design that inmates accu- rately recall their street months, but the effect of differential recall abilities is un- known. Finally, respondents with few street months may have disproportion- ately higher estimates of A, and part of that relationship may be an artifact of the way in which A is estimated for those inmates. Determining Crimes Committed The numerator of A is the number of crimes of a specific type that the inmate reported committing cluring the months he was on the street, according to the calendar he filled in. Questions were asked about 10 crime types. Chaiken ant! Chaiken (1982a:42) describe the general format of these questions: After answering "yes" that he had committed a given type of crime, say, burglary, during the measurement period, the respondent was asked to tell how many burglaries he had committed by specifying a range, either "1 to 10" or "11 or more." If the range was "1 to 10," |77 he was asked, "How many?" If the range was "11 or more," he was led through a sequence of questions about the number of months in which he committed burglary and his daily, weekly, or monthly rate of commission. Figure 2 is a copy ofthe questions used in the Rand inmate survey to determine the number of business robberies in- mates committed cluring their street months. This series of questions first dis- tinguishes between the inmates who committed 1 to 10 crimes (referred to here as low-frequency offenders) and the group who committee] 11 or more crimes (the high-frequency offenclers).8 As the format of Figure 2 shows, the high-fre- quency offenders had a much more dif- ficult cognitive task. Since the Tow- and high-frequency offenders answered sepa- rate questions, different problems arise in computing their estimates. In the re- analysis Ranks procedures for estimating "crimes did" (Chaiken and Chaiken, 1982a: 191-192) were carefully examined, and this information was used as a start- ing point for recomputing the estimates. Estimates of crimes committed for the 1-to-10 group and the 11-or-more group are discussed separately below. As shown in Figure 2, offenders who reported committing 1 to 10 business rob- beries were asked to give a specific num- ber, and most of them did so. However, 17 percent of the inmates who reported committing either robbery or burglary at a Tow frequency clic! not answer the fol- low-up question. For these inmates, Rand analysts assigned 1 as the minimum esti 8For inmates who reported committing business robbery, 72 percent were low-frequency offenders, 24 percent were high-frequency offenders, and 4 percent did not check either box. For personal robbery, the percentages were 76 and l9, respec- tively, for low- and high-frequency offenders (data were missing for 5 percent). For burglary, 66 per- cent were low-frequency offenders, 31 percent were high-frequency, and data were missing for 3 per- cent.

178 CRIMINAL CAREERS AND CAREER CRIMINALS II. 1. During the STREET MONTHS ON THE CALENDAR did you rote any businesses? That is did you hold up a store, gas station, bank, taxi or other business? YES O NOO2 2. In all, how many businesses did you rob? O 1 1 OR MOR E 3. Look at the total street months on the calendar. During how many of those months did you rob one or more businesses? Months 4. In the months when you did business robberies, how often did you usually do them? (CHECK ONE BOXJ EV ERYDAY OR How many ALMOST EVERYDAY O per day? SEVERA L Tl MES A WEEK EVERY WEEK OR ALMOST EVERY WEEK O LESS THAN EVERY WEEK go on to page 20 O1 TO10 How many? 9O on to next page ,~ Business Robberies DHow many days / a week usually? / I OHow many per week? l I How many l I per month? L I How many / / LJ per month? / I FIGURE 2 Sample page from inmate questionnaire. Source: Greenwood (1982). mate and 10 as the maximum estimate, which results in an estimate of 5.5 when the minimum and maximum values are averaged.9 - 9For these inmates, Rand calculated two esti- mates of A using the two numerators (and possibly two denominators if street months had a minimum and a maximum value). For most of the analyses, however, the analysts simply averaged the two A estimates. The possible distortions introduced by Rand's procedure of using minimum and maximum estimates rather than a single estimate are discussed in a later section. The "average" or "typical" Tow-fre- quency offender, however, clid not report committing five or six crimes during his measurement period but more often ad- mitted to only two or three crimes. Thus, the minimum-maximum strategy used by Rand was likely to produce a high esti- mate of crimes committed for inmates who gave partial answers. To avoid this potential bias in the reanalysis, an alter- native approach was taken. For inmates who gave incomplete answers, estimates from 1 to 10 were assignee] in a way that

THE RAND INMATE SURVEY: A REANALYSIS 179 TABLE 8 Average Number of Crimes Reported by Low-Frequency Offenders and Two Estimates for Inmates with Missing Information Business Personal Robbery Robbery Burglary Low-Frequency Offenders Mean N Mean N Mean N Answered follow-up questions Did not answer follow-up questions Rand estimate Alternative estimate 3.3 5.5 3.2 233 55 3.5 5.5 3.4 290 48 4.0 5.5 4.0 497 74 NOTE: Data are presented only for inmates who were "active" in (i.e., reported committing) the specific crime type. aThe follow-up question was posed to all offenders who said they committed "1 to 10" crimes (see Figure 21. The exact number of low-frequency offenders for whom Rand assigned minimum and maximum values of 1 and 10 could not be determined. The two groups with missing information, however, are essentially identical since missing data were already coded on the public-use tape of the inmates' responses. simulated the distribution of the re- sponses to the follow-up question for the larger sample.10 In Table 8 the mean for inmates who answerer! the follow-up question is com- pared with two means for inmates who failed to answer the follow-up question- one obtained using the Rand strategy and the other using the alternative procedure just clescribec3. The altemative strategy i°The distribution-matching procedure was adopted only after other alternatives were consid- ered and rejected. Assigning the mean value (from the group who did answer the follow-up question) for each inmate who left the follow-up question blank would have distorted the true distributions of "crimes-did" and A by lumping 17 percent of the low-frequency offenders at a particular value. The median was rejected for similar reasons and because it conceals variation in the distribution. (The distri- bution of crimes committed by low-frequency of- fenders looked a lot like A skewed to the right with few inmates reporting 9 or 10 crimes.) For burglary, the distribution-matching proce- dure was used along with another question, CK14D, which was originally intended as a reliability check on the number of burglaries an inmate reported. (This question had ordinal response categories: 0, 1-2, 3 5, ~10, 10 + .) If the inmate failed to answer question CK14D, the estimate of burglaries commit- ted was determined in the same manner as robbery. for handling missing ciata provides an estimate of crimes committed that more closely approximates Me responses of those who clid answer the follow-up ques- tion than does the Rand procedure. This alternative approach shouIcl lower A esti- mates, but the overall impact on the dis- tribution may not be significant, since the number of cases involved is small and the offenders are aIreacly at the low end ofthe distribution. Computing the number of crimes com- mitted by the high-frequency offenders (those who reported "11 or more" crimes) is more complicated. Chaiken and Chaiken (1982a:191-192) give a short cle- scription of their general computational strategy, which was also used in the reanalysis. The first task is to determine the number of months cluring which high- frequency offenders committed crimes. Referring again to Figure 2, in question 3 respondents were asked during how many months (of their total street months) they committed at least one robbery. Then, the respondent was supposed to check one of four categories indicating the frequency with which crimes were committed every clay or almost every

180 day, several times a week, every week or almost every week, or less than every week. A follow-up question elicited a specific number of crimes committed (crimes per day, per week, or per month). Depending on which category was incti- cated, the total number of crimes commit- ted cluring an inmate's street months was computed in one of the following ways (see Chaiken and Chaiken, 1982a:191- 192~: Crimes = cr~mes/day days/week months did 4.3 weeks/month, or = crimes/week months did 4.3 weeks/mon~, or = cr~mes/month months did. Incomplete or ambiguous responses in this set of questions were fairly common. Typical problems included checking more than one frequency category, re- porting ranges (e.g., 2 to 4 crimes/week), indicating a frequency level (day, week, or month) but not the number of crimes committed, and reporting no information about rate of criminal activity. Chaiken and Chaiken did not provide specific cle- taiTs about their treatment of missing data, but they diet report that "reasonable ranges [were] used in the calculations" and that both maximum and minimum estimates were calculated (1982a: 1911. Examining the detailed materials pro- videc! by RancT for the reanalysis clarified how missing data were handled in most cases. thin response to a request for additional informa- tion about how crimes-did was estimated, especially in ambiguous cases, fan Chaiken at Rand provided relevant portions of the computer code that was used to transform the raw data into variables. The overall strategy of the Rand analysts was to calculate minimum and maximum estimates if inmates gave incomplete, ambiguous, or conflicting responses. According to Rand's major report (Chaiken and Chaiken, 1982a: 184), "minimum and maximum es- timates are not intended to be 'worst possible' cases, but rather reasonable conclusions from the data." Information in Me computer code indicated Mat (1) CRIMINAL CAREERS AND CAREER CRIMINALS The procedures adopted in the re- analysis for dealing with missing data and other ambiguous responses were con- servative. If inmates reported ranges for their answers, the midpoint was taken as the estimate. Multiple responses from high-frequency o~enclers (answers to more than one of the frequency catego- ries) were averaged. Reasonable esti- mates were used in place of missing data only if the respondent provided at least a partial answer (i.e., checked "several times a week" but click not specify how many crimes per week). These substi- tutec] values were based on responses by similar inmates; the specific value was chosen to match the distribution of others who did provide an answer. However, if questions concerning "months-did" were left blank, the inmate was excluded from any furler calculations. Finally, follow- ing Rancl's proceclure, inmates who re- ported that they committed "11 or more" crimes but left other questions in the sequence blank were also excluded. Thus, the procedures user! in the reanalysis for calculating street months, months-dicl, and the number of crimes committed (of a specific crime type) dif- ferecI in important ways from the methods user] by Me Rand analysts. Most notably, instead of using Rancl's strategy of esti- mating minimum and maximum esti- mates for the numerator and denominator of A, an alternative, conservative estimate was developed based on the available clata. The next section suggests some im ~ ~ 1 ~ a range of values (e.g., three to five) was substituted for missing answers on questions about rate or number of crimes committed, but those ranges ap- peared to be arbitrarily chosen by the analysts; (2) if months-did was missing, estimates of minimum and maximum street months were substituted; (3) mul- tiple responses to a single question were treated as minimum and maximum estimates of crimes-did; and (4) if ranges were specified for any response, both the minimum and maximum values were used in the calculations.

THE RAND INMATE SURVEY: A REANALYSIS plications of Rancl's strategy for estimat- ing A and presents the A estimates cal- culated according to the procedures discussed in the last several sections. Calculation of A As discussed earlier, an offender's an- nual crime rate is simply the number of crimes he committed per year of"street time." Rancl's intermediate calculations led to two estimates of A a minimum (based on the minimum estimate for crimes-did clivicled by the maximum esti- mate of street months) and a maximum (based on the maximum estimate of crimes-did dividect by the minimum esti- mate of street months). However, in the Chaiken and Chaiken report the sum- mary statistics by state and all the analy- ses are computed using the average of the two estimates of A. The authors state that the minimum and maximum estimates of A, and thus the average A, are reasonable conclusions from the data, but those methods actually produce the smallest possible minimum estimate and the larg- est possible maximum estimate, which was often double or triple the minimum estimate. The average of these two ex- tremes can be very sensitive to the maxi- mum value, and this couIct account for some of the skewness in the distribution of A. In Table 9 the distributions of Rand's 181 minimum and maximum estimates of A for robbery and burglary are compared with that ofthe single estimate generated in this reanalysis. As seen in Table 9, the reanalysis esti- mates for both robbery and burglary are practically iclentical to Rancl's minimum estimates, but they diverge consiclerably from Ranks maximum estimates. At some points in the distribution, the values of A from the reanalysis are actually lower than Rand's lowest estimates. (This is probably the result of using smaller val- ues in substitutions of missing data.) Ranc3's average estimate of A, then, will be higher than the reanalysis estimate, as shown in Table 9. But Table 9 confirms one important result of the Rand survey: the distribution of A computed from re- ports by the sample of incarcerated of- fenders is highly skewed, with about 50 percent of the sample reporting fewer than five crimes per year, and the top 10 percent reporting at least 70 crimes per year. (A more detailed cumulative per- centage distribution is presented in Ap- penclix Table A.1.) Important (differences emerge when Me estimates of A for incarcerated robbers and burglars are broken down by the three states California, Michigan, and Texas See image ~ uJ. ~ he annual o~end- ing frequencies for Me active robbers in Me CaTifomia and Michigan samples are ~/ ~ ~\ ~' TABLE 9 Distribution of A: Rand Minimum and Maximum Estimates and Estimate from the Reanalysis for Inmates Who Reported Committing Robbery or Burglary Robbery Burglary Rand Rand Reanalysis Rand Rand Reanalysis Statistic Minimum Maximum (Rand Avg.) Minimum Maximum (Rand Avg.) 25th pet. 1.8 2.3 1.5 (2.0) 2.4 2.8 2.0 (2.2) 50th pet. 3.6 6.0 3.8 (5.0) 4.8 6.0 4.7 (5.5) 75th pet. 12.0 21.5 12.4 (16.0) 23.3 35.0 23.4 (300) 90th pet. 68.0 100.5 71.6 (87.0) 196.0 265.0 195.9 (232.0) Mean 40.6 62.2 43.4 75.8 118.6 79.0 SOURCE: The Rand estimates were calculated from data provided by J. Chaiken.

182 TABLE 10 Differences in Distributions of A for Inmates Who Reported Committing Robbery or Burglary, by State Statistic California Michigan Robbery 25th pet. SOth pet. 75th pet. 90th pet. Mean Burglary 25th pet. 50th pet. 15th pet. 90th pet. Mean 2.1 5.1 19.8 107.1 42.4 2.3 6.2 49.1 199.9 98.8 1.4 3.6 13.1 86.1 45.4 1.9 4.8 24.0 258.0 82.7 0.9 2.5 6.2 15.2 13.1 1.2 3.1 9.9 76.1 34.1 SOURCE: Data were computed as part of the reanalysis. similar. In contrast, inmates in Texas pris- ons who admitted committing robbery reported an average of about 13 robberies per year, about one-third the rate of the robbers in the other two states. The esti- mates of A for burglary also show a large difference between Texas burglars and those in California and Michigan. Largely the saline patterns were observed by Chaiken and Chaiken (1982a:Appendix Tables A.3 and A.6), although the esti- mates of A are lower in the reanalysis. These interstate differences in the dis- tribution of A for robbery and burglary may reflect actual variation in criminals' offending patterns in the three states. A1- ternatively, they may be a consequence of different criminal justice processes. In general, the distribution of A derived from self-reports of incarcerated offenders will not be representative of the distribution for the larger convicted population or for the general criminal population. Biases are introduced because convicted per- sons are incarcerated selectively rather than randomly, that is, more serious, high-frequency offenders are incarcer- ated in greater numbers and for longer CRIMINAL CAREERS AND CAREER CRIMINALS periods than other offenders. California appears to be especially selective, limit- ing its available prison space to serious repeat offenders, and Texas appears to operate much less selectively. Once these differences in criminal justice sys- tem practices are taken into account, the A distributions for "street offenders" in Cal- ifornia, Michigan, and Texas may be much more similar (see Spelman, 19841. In any event, the state-specific esti- mates of )< are consistent with Rand's finding of highly skewed distributions of annual offending. This pattern is weaker for Texas inmates, especially for those convicted of robbery, but the form of the distribution remains unchanged. Despite having replicated the shape of Rand's A distribution, however, the extent to which ambiguous or missing data used in calculating A (such as the estimates used in calculating street months or crimes- did) might affect the overall distribution is still a matter of concern. In particular, A estimates for high-frequency offenders might be less reliable than the estimates for those who commit crimes at a lower level because the questions asked of very active offenders were more complex. Data to address this question are pre- sented in Table 11. Inmates who reported committing rob- bery or burglary were divided into five groups for the reanalysis, depending on the types of ambiguity in the responses that were used to calculate A. The four types of ambiguity were (1) cases with ambiguous numerators (crimes-did), which included low-frequency offenders who did not specify a number between 1 and 10 for crimes committed and high- frequency offenders who gave a range for a number or rate, gave multiple answers for a single question, or gave partial an- swers; (2) cases with ambiguous denomi- nators, which meant that the respondent had problems answering the questions about street months or completing the

THE RAND INMATE SURVEY: A REANALYSIS TABLE 11 Distribution of A for Robbery and Burglary, Adjusting for Types of Response Ambiguity Cases with Cases with Cases with Unambiguous + Ambiguous +Ambiguous + Street Months Cases Numerator Denominator Less than 7 (1) (2) (3) (4) ~3 All Inmates Percent Committing Change from Crimea Column 1 to 5 (5) (6) Robbery 25th pet. 1.1 1.1 1.3 1.5 1.5 27 50th pet. 2.8 3.0 3.0 3.7 3.8 26 75th pet. 6.9 8.1 8.9 12.5 12.4 44 90th pet. 43.2 36.2 54.8 68.9 71.6 40 Mean 36.5 43.7 39.7 43.2 43.4 16 N (294) (362) (475) (548) (594) Burglary 25th pet. 1.8 1.7 1.7 1.9 2.0 10 50th pet. 4.3 4.2 4.3 4.8 4.7 8 75th pet. 17.5 16.7 20.8 23.5 23.4 25 90th pet. 158.4 156.2 180.6 191.7 195.9 19 Mean 55.7 54.7 69.0 77.7 79.0 29 N (451) (520) (682) (768) (824) aThe change from Column 4 to Column 5 represents a small group of cases for which both numerators and denominators were defined as ambiguous. calendar; (3) cases with low street months (because of a concern for the vaTiclity of A estimates for this group); ant! (4) cases with both ambiguous numerators and cle- nominators. Unambiguous cases make up the fifth group. As seen in Table ll, the estimates are affected by response ambiguity. The per- centiles (values) and especially the mean change significantly from the group of unambiguous cases, to cases with varying levels of ambiguity, to all cases (column 51. Moreover, respondents for whom the numerators or denominators are ambigu- ous and those with short street times do have higher A estimates than respondents with unambiguous responses. The values ofthe summary statistics for the unambig- uous cases are much Tower than the val- ues for all cases, as inclicatecI by the per- centage change in column 6. This pattern suggests that A for higher frequency of- fenclers is particularly susceptible to mea- surement error and problems of unreli- ability. Of course, estimates at the high end of any distribution will have a greater variance than those at the Tow end, but, because A is a ratio variable, substantial measurement error at the high end may artificially stretch the tad] of the clistribu- tion. Finally, even with the unambiguous cases, the highly skewed distribution of A persists, albeit at Tower levels. The High-Frequency Offenders The extremely high-frequency offencI- ers in the distribution of A raise a number of serious problems. First, are these few respondents telling the truth about the number of crimes they committed ? For 5 percent of the active robbers or burglars in the Ranc] sample, estimates of A ex- ceecle(1 300 (robberies or burglaries). Af- tera thorough review ofthe data, Chaiken and Chaiken (1982a:24~251) dill not find any systematic evidence that overall as- sessments of validity for respondents were related to their self-reports of crime. However, respondents who exaggerated their criminal activity are probably com- pensatecl for by respondents who under

184 stated their activity at least for the me- dian. Incleed, this pattern of response errors (both exaggeration and conceal- ment) emerged in one study of errors in self-reports of arrests (Wyner, 19801. Wyner shows that grouping the data (e.g., into low, medium, ant] high groups) re- duces the impact of these errors, but the group variances may be quite large due to the combination of underreports and overreports. Thus, response errors may inhibit one's ability to predict the actual level of criminal activity of convicted of- fenders accurately. The issue of preclic- tion is acldressed in great cletafl in the next section. Second, the small group of high-fre- quency offenders has a tremendous im- pact on the overall distribution of A. Chaiken and Chaiken used a logarithmic transformation of A in their analysis be- cause of the extreme variation in the data. Otherwise, there is very little association between A and characteristics of offend- ers. It is also clifficult to describe the distribution of A using standarcl measures of central tendency. The arithmetic mean is very sensitive to extreme values; per- centfle values are presenter] in the tables in this paper, as in the Chaiken and Chaiken report, for this reason. However, the 90th-percentfle value is almost as vol- atfle as the mean because of the wicle spread between indiviclual estimates at the high end of the distribution. Thus, the skewness ofthe distribution poses special problems for data analysis, and the results may be sensitive to the choice of an ana- Tytic strategy. Third, high-Eequency offenders may not commit crimes at a stable rate throughout the year, and, therefore, they pose special problems for measuring an- nual offending frequencies. In a separate Rand report, Rolph, Chaiken, and Houchens (1981:37) analyzed the in- mates' self-report crime ciata and found that some respondents committed crimes CRIMINAL CAREERS AND CAREER CRIMINALS in "spurts." However, the questionnaire design and the technique used to esti- mate A (especially for high-frequency of- fenders) in the Rand study assume stable monthly rates of criminal activity. Re- spondents are supposes] to estimate the number of crimes committed during their "street months" by focusing on a typical month. But some offenders appear to al- ternate between periods of high criminal activity and no activity. Or offenders may be especially active just prior to the arrest month, in which case their reports of offending would reflect this anomalous period. Are these respondents likely to compute an average monthly rate that takes into account these high and Tow periods? Or might they focus on their most active period and report the number of crimes committed during that month? A similar measurement problem exists for inmates who spent several months in prison or jail just prior to the current arrest, and thus had fewer street months, but were very active during their time on the street. For these types of inmates, self-reports of their crimes in a month probably do not reflect one-twelfth their yearly rate, and, consequently, estimates oftheir annual offending frequencies may be artificially inflated. Whether active of- fenclers in the Rand sample committed crimes in spurts during the months before their arrest is difficult to determine. How- ever, the estimates of A for inmates with differing periods of street time can be compared, as is done in Table 12 for the respondents who reported committing robbery or burglary. The data in Table 12 clearly show a negative relationship between A and street months. Respondents with short street times (less than 12 months), and especially less than 7 months, have higher estimates of A (100 or more) than respondents with longer street times. This pattern is stronger for inmates who committed robbery than for those who

A THE RAND INMATE SURVEY: A REANALYSIS TABLE 12 Cross-Tabulation of Street Months By A for Inmates Who Reported Committing Robbery and Burglary (percentages) Street Months ~5 1-6 7-12 13-18 12-24 N Percent Robbery <2 3.6 21.7 36.4 49.4 190 32.8 ~5 14.6 27.4 25.1 22.2 135 23.2 6~0 50.0 33.9 26.4 17.3 166 28.6 31-99 18.3 5.7 4.3 5.5 40 6.9 100+ 13.4 11.3 7.4 5.6 49 8.4 N (82) (106) (231) (162) 580 99.9 Burglary <2 3.2 21.3 28.8 36.4 220 26.4 ~5 18.1 20.0 26.5 32.2 215 25.9 6-30 41.4 36.2 20.4 16.4 205 24.7 31-99 10.6 5.6 6.4 7.4 57 6.9 100+ 27.6 16.9 17.8 8.2 133 16.0 N (94) (160) (343) (233) 830 99.9 committed burglary. All inmates with street time of less than 12 months spent some time in jai] or prison in the year before their arrest; therefore, those of- fenders are probably a very active group, although specific estimates of their As may be inflated. Conversely, low esti- mates of A (fewer than two crimes per year) were reported by 49.4 percent of robbers with at least 19 street months, whereas only 3.6 percent of those with less than 7 street months reported such a low rate of criminal activity. These findings confirm a suspicion that for respondents with very short street times, valid estimates of A probably can- not be attained using Rand's question- naire design. Estimates of A for respon- dents who commit crimes in "spurts" may also be misreading. As Cohen (1983) predicted, the overall impact of this prob- lem appears to be overestimates of A for some respondents. In summary, the (listribution of A for robbery and burglary remains highly skewed even after taking into account unreliable respondents, respondents with few street months (or substantial time "not on the street"), and other problems in the estimation of individual offending fre quencies. The precise incliviclual esti- mates of A, however, are highly sensitive to analytic choices in computation. Data of this type are probably best grouped into several categories, such as Tow-, meclium-, and high-rate offenders, or transformer] with a logarithmic function before analy- sis. Such procedures preserve the orcler- ing of inmates according to the frequency of their criminal activity but eliminate the need to rely on specific estimates of individual offending frequencies. Development of a Prediction Scale The profound skewness in A for Ran(l's inmate sample and the existence of a small group of extremely active criminals led the Rand analysts to clevelop several types of mo(lels to identify the high-rate offenders. Of course, a predictive mode] that prospectively iclentified likely high- rate offenders would be invaluable to criminal justice decision makers. Chaiken en cl Chaiken (1982a:Chapter 3) attempted to develop such a profile using a multivariate approach. As dis- cusse(1 earlier in the review of the Ran results, a variety of self-report measures were predictive of high robbery rates.

186 However, when inmates were divided into predicted Tow- and high-rate groups (based on the regression model), 67 per- cent of the predicted high-rate group re- ported committing fewer than 10 robber- ies per year, including 30 percent who committed no robberies (Chaiken and Chaiken:8~92~. Also, the Rand model that used only official record information (aclult prior convictions, recent arrests, juvenile convictions, and recorded drug history) produced a high percentage of false predictions. Chaiken and Chaiken (1982b) concluded that high-rate con- victed robbers cannot be adequately dis- tinguished from other convicted robbers because much information that is espe- cially predictive is not in existing official records. In particular, official records are weak predictors because high-rate of- fenders often have short official records (because they tend to be young), juvenile records are frequently incomplete, and detailed information about drug history is usually unavailable. In a separate analysis of the inmate data, Greenwood (1982) used an alternate approach in creating a prediction instru- ment. As discussed, Greenwood selected seven characteristics whose presence or absence was associated with high annual robbery and burglary rates and created a simple, seven-point additive scale. Rob- bers and burglars were classified into pre- dicted low-, medium-, and high-rate groups based on arbitrary cut points on the seven-point scale. This scale has been the subject of much discussion since the release of Greenwood's report Selective Incapacitation, but the report left many questions unanswered. The next several sections of this paper examine the empir- ical ant! conceptual relationships of the seven variables to A and to each other, issues in the development of the seven- point scale, the predictive accuracy of the scale (especially interstate differences), and the scalds utility as an air] in sentenc CRIMINAL CAREERS AND CAREER CRIMINALS ing, crime reduction, and controlling prison populations. Identifying the Seven Variables Greenwood initially iclentified 13 can- diciate predictors of high-rate offenders on the basis of prior research and possible relevance. Then, focusing only on con- victed robbers and burglars, he divided the inmates into Tow-, meclium-, ant] high-rate groups, depending on their an- nual offending frequencies (A) for robbery or burglary. The partitions, used for all offense types and states, were: below the 50th percentile (Iow), between the 50th and the 75th percentiles (medium), ant! above the 75th percentile (high). The val- ues for these percentiles differ widely among the states. For example, the 75th percentile cutoff values are 12.0, 6.2, and 3.3 (computed for the reanalysis) for Cal- ifomia, Michigan, and Texas, respec- tively. Thus, "high-rate" offenders have very different estimates of A in the three states. Greenwoocl cross-tabulated each of the 13 candidate predictors against the three frequency groups and chose seven vari- ables based on the strength of their asso- ciation win the groups (Greenwood: 49 52, 9~1071. However, the tabulations were based on the entire sample of con i2The 13 yes-no variables are (1) prior conviction for current offense (robbery or burglary), (2) incar- cerated more than 50 percent of 2 years preceding current arrest, (3) convicted before age 16, (4) juve- nile commitment to state facility, (5) heroin or bar- biturate use in 2 years preceding arrest, (6) heroin or barbiturate use as a juvenile, (7) employed less than 50 percent of preceding 2 years, (8) convicted on multiple counts, (9) prior felony convictions, (10) prior prison term, (11) more than three jobs in the preceding 2 years, (12) less than 23 years old at time of arrest, and (13) prior arrest for current offense type. Greenwood eventually selected the first seven variables for his scale. (The phrase "preceding 2 years" refers to an inmate's measurement period, which could actually range from 13 to 24 months.)

THE RAND INMATE SURVEY: A REANALYSIS TABLE 13 Means of the Seven Variables for California, Michigan, Texas, anct Combined Samples 187 Variable All States California Michigar~Texas .44 Prior conviction for current offenses Incarcerated 50% or more of preceding 2 years Convicted before age 16 Juvenile incarceration Recent adult drug use Juvenile drug use Unemployed 50% or more of preceding 2 years Sum of the seven variables N .32 .17 .32 .26 .45 .48 .54 2.63 (884) .34 .23 .43 .35 .59 .58 .60 3.12 (317) .16 .16 .28 .24 .43 .44 .62 2.33 (255) .12 .26 .20 .40 .33 .42 2.17 (312) NOTE: All variables are coded 0 or 1; thus, the means represent the proportion of inmates with the attribute. Missing data are also coded O. following the Rand procedure (Greenwood, 1982:50). The sample is all inmates who were convicted of either robbery or burglary. aThese means are slightly distorted because all jail inmates received a zero for this variable, but the Texas sample did not include jail inmates. victed robbers and burglars from all three states (Greenwooc3:51-521. This use of the entire sample obscures the possibility that mean values of the predictor vari- ables (i.e., the proportion possessing each attribute) diner across the states, perhaps as a result of criminal behavior, criminal justice system operations, or recorcI- keeping practices. Such state-specific dif- ferences will affect the distribution of scale scores, hence the optimal cut points for classification across the states. State-specific means for the seven vari- ables, shown in Table 13, indicate the magnitude of these interstate differ- ences.~3 The California sample has a i3Greenwood provides little information in his report about how the seven variables were con- s~ucted (i.e., specific survey questions used), ex- cept to say that one was coded from official record data (past conviction for same charge). The variables were independently c~ons~ucted for this reanalysis and the two procedures compared after Greenwood provided a copy of the computer code he used to create his variables. Very few differences exist be- tween the two procedures. Greenwood's overall means for the seven variables (1982:51-52) are (listed in the order given in Table 131: 0.33, 0.20, 0.33, 0.27, 0.47, 0.50, and 0.56, which are within 0.03 higher proportion of inmates who had an . . . . ear y conviction or a Juvenl e 1ncarcera- tion history in comparison with inmates in the Michigan and Texas samples. Cal- ifomia inmates also appear more likely to have a history of serious drug abuse. Fi- nally, inmates seem to have better work records in Texas than in California or Michigan. Thus, the mean of all seven variables is 2.63, baser! on the entire sam- ple. Because of interstate differences in Me indivi(lual items, however, the mean ranges from 3.12 in the CaTifomia sample to only 2.17 in Texas. Relationships Among Variables The conceptual and empirical relation- ships among the seven variables are not discussed in the Rand report. However, of Me means in Table 13. These slight differences are largely due to the change in sample size (781 to 884) because of the redefinition of active burglars. (Two typographical errors exist in Greenwood's table: variable 6 should have frequencies of 509, 255, and 167 and variable 10 should have frequen- cies of 299, 436, and 45.)

188 CRIMINAL CAREERS AND CAREER CRIMINALS TABLE 14 Inter-Item Correlations of Seven Variables in Greenwood Scale for All Convicted Robbers ancl Burglars Variable (1) (2) (3) (4) (5) (6) (1) Past convictions (2) Recent incarceration .12 (3) Early conviction .02 .18 (4) Juvenile incarceration .05 .22 .43 (5) Recent adult drug use .04 .10 .08 .00 (6) Juvenile drug use -.04 .13 .14 .07 .47 (7) Recent unemployment -.02 .09 .09 .02 .14 .12 NOTE: Variable descriptions have been abbreviated. See Appendix B for full explanation of each variable. two pairs of variables seem conceptually clependent: juvenile incarceration and conviction before age 16 and juvenile drug use and recent adult drug use. Also, according to Table 4.4 in the Greenwood report (pp. 51-52), the frequency clistri- butions for these variable pairs are very similar. Thus, one variable in each pair may provide reclunciant information. To assess the empirical relationships among all seven variables, an inter-item correla- tion matrix was calculated. The inter-item correlations, shown in Table 14, are very Tow, averaging .10, except for the two pairs of variables that overlap conceptually, and their correla- tions are significantly higher (r = .43 and .47, respectively). These two pairs of vari- ables (items 3 and 4 and items 5 and 6) may be the most dominant items in the scale, since drug use and juvenile crimi- nal history are, effectively, being in- cluded twice. A second issue is the bivariate associa- tion between each of the seven variables and high-rate offending. Greenwoocl chose a chi-square test to determine the strength of the relationships between each variable and the three offense groups. In the reanalysis, for which a slightly larger sample was used, the re- sults were similar for the chi-square tests of the seven variables. However, when the chi-square test is adjusted for missing ciata on the past-conviction variable, it is not significant at the .05 level.~4 Thus, the only official record information in Green- wood's scale is a poor predictor of high rates of robbery and burglary. Inclividually the seven variables show moderate to fairly strong associations with A (statistically significant at least at i4Greenwood tested the hypothesis that the per- centage of inmates with the attribute in the three groups would be different from the marginal distri- bution of cases (50 percent low rate, 25 percent medium rate, 25 percent high rate). The appropri- ate chi-square test is one that assumes a fixed total (the number of inmates responding "yes"), which is a one-tailed test with two degrees of freedom; x2(p = .01) = 9.2, x2(p = .05) = 5.9. (Greenwood also tested the distribution of the "no" responses, but it is not clear why this is important.) There are some indications that Greenwood did not use the appro- priate test, since under these rules and the data in the Rand report, juvenile drug use is definitely significant at the 0.01 level (1982:52~. None of the significant variables would change, but some vari- ables reported as nonsignificant may actually be significant. The problem with the chi-square test is that the null hypothesis of 50, 25, and 25 percent is not the correct comparison for the past-convictions variable (or the juvenile drug-use variable) because of sub- stantial missing data (see Greenwood, 1982:51-57), which affects the marginal distributions. The per- centa~e of cases falling into the three groups, dis- counting the missing cases, is actually 47.3, 24.3, and 28.4 percent for low-, medium-, and high-rate groups, respectively. This distribution is not signif- icantly different (at the .05 level) from 45, 28, and 30 percent, which is the reported distribution of prior convictions for robbery or burglary across the three groups (p. 51~.

THE RAND INMATE SURVEY: A REANALYSIS 189 TABLE 15 Summary Information from Eight Regressions with Estimates of A for Robbery or Burglary as the Dependent Variables and the Seven Items in Greenwood's Scale as the Independent Variables All States Independent Variable AR AB California Michigan Texas AR AB AR AB AR AB Past conviction X X (X) X Recent incarceration X X -X Early conviction X X Juvenile incarceration X X (X) Recent adult drug use X X (X) X X X X X Juvenile drug use X X X X X X Recent unemployment X X X X Adjusted R2 .12 .19 .20 .22 .08 .12 .05 .15 N (848) (311) (245) (292) NOTE: The sample is all convicted robbers or burglars; 36 cases were omitted because of missing data on A. The dependent variable is loge (Al + .5), calculated separately for robbery and burglary. An "X" indicates that the variable was significant at the .05 level in that equation; a "-X" indicates that the variable was negatively associated with individual crime rates; and (X) indicates that, while the variable was not significant in that particular equation, a di~erences-between-slopes test revealed that the coefficient was not statistically different from other states' coefficients on that variable. The variables were all coded 0 or 1, with missing data coded as 0. The variable descriptions are abbreviated; the reader should refer to the text (including Appendix B) for more information. the .05 level), but how do these factors relate to offending rates when they are combined? To assess both the indepen- dent effects of these variables and their collective impact, regression equations were estimated separately for each crime type and within each state, as well as for all data combined (a total of eight regres- sions). The seven items in the scale are the only inclependent variables, and the dependent variable is T°ge (Ai + 05), following Chaiken and Chaiken's proce- dure. (The regression coefficients for these equations are reported in Appendix Table A.2.) The results are summarized in Table 15; an "X" indicates that the variable was significant at the .05 level in that equation. The pattern of significant variables that is evident from this summary table sug- gests strong state differences, some ef- fects for offense type, and very different saliency of the seven variables in their associations with A. The equations based on data from all states mask some impor- tant findings. The interstate differences in the percentage of variance explainer! are especially striking. The equation explains 20, 8, ant! 5 percent of the variance in self-reported annual robbery rates in Cal- ifomia, Michigan, and Texas, respec- tively. The regression for burglary also fits the data better in California than in the other states. In his report Greenwooc! focuses primarily on CaTifomia robbers, but the data in Table 15 suggest the relationships in California are probably much different from the relationships in Michigan and Texas. The seven variables explain more vari- ance in A for burglary than for robbery in all three states, but part of this result may be clue to the higher variance in the reported burglary frequencies in compar- ison with the robbery frequencies. Some variables are only significant for a specific offense type (in the state-specific equa- tions). Based on the entire sample, a past conviction for current offense and convic- tion before age 16 affect only offending frequencies for burglary. In contrast, a juvenile incarceration or incarceration in

190 the 2 years preceding the current arrest is associated only with offending frequen- cies for robbery. This suggests that sepa- rate scales would be appropriate for rob- bery and burglary. Finally, at most, only three or four of the seven variables are significant in each state-specific equation. In the combined equations, however, five variables are sig- nificant, which is probably due to the larger sample size. Also, these overall equations obscure the different effects of the seven variables. For example, recent incarceration and recent unemployment are significant in only two-of the six state- specific regressions. The two juvenile criminal- history variables (conviction be- fore age 16 and incarceration in a juvenile facility) do not appear to be strong predic- tors in these equations, perhaps because of their high intercorrelation. However, at least one of the drug use variables, which are also correlated, is significant in all six state equations. To summarize, in a multivariate frame- work the proposed seven-point scale ap- pears weak. The crime-type and state- specific regressions show that, with the exception of drug use, the effects of these variables on individual offending fre- quencies are not at all robust across states or offense types. The seven variables fit the data best in California, but even there they explain only 20 percent of the re- spective variances in individual robbery and burglary rates. In contrast, in Texas only one variable is significantly related to ~ for robbery, and the R2 is just 5 percent. Further, several variables in Greenwood's scale appear to be related to A in only one state, once other factors (such as drug use) are taken into account. Missing Data Incomplete data are common in survey research, and the Rand inmate survey was no exception. The questionnaire was CRIMINAL CAREERS AND CAREER CRIMINALS long, the skip patterns were complicated, and a few questions probably were con- fusing to some inmates; hence, respon- dents left some questions blank. As dis- cussed above, Rand analysts used a variety of ways to deal with missing data for questions about number of crimes committed. Then, in their multivariate analysis, Chaiken and Chaiken (1982a:81) replaced other missing values with the state-specific means. Missing information is also a problem for some of the seven variables in Green- wood's scale (pp. 51~7~. The proportion of cases (convictec! robbers and burglars) in which information is missing for any particular variable because of skipped questions or other reasons is generally about 5 percent. However, one variable past conviction for current offense is missing for 31 percent of Greenwood's sample. As pointed out by Cohen (1983), this variable was coded from official rec- ords, which were only available for in- mates surveyed in prisons. Thus, jail in- mates account for most of the missing values for the past-convictions variable. Since missing information for at least one of the seven variables was common among the sample, those cases could not simply be excluded. Instead, Greenwood set missing values for each of the vari- ables in the scale to 0, thus combining the {` ~ ,' ~ {` · . ~ ,, T ~ . ~ no s ana t ae missing s. me exp alnect that this conservative procedure would "bias [scale] scores downward" (Green- wood:501. However, this solution is not appropriate for the past-convictions vari- able because all the jail inmates were coded O for Greenwood's analysis. The unintended effect is that the variable is transformed into a measure that distin- guishes jail and prison inmates in the sample. Such a measure is a priori a predictor of high-rate offending if the prison versus jail decision tends to result in high-rate offenders being sent to prison and others being sent to jail.

191 Number All Missing Cases THE RAND INMATE SURVEY: A REANALYSIS TABLE 16 Distribution of the Number of Missing Variables in the Seven- Point Scale Jail Inmates o 2 3 4 5 464 316 82 17 4 1 133 51 9 2 NOTE: None of the respondents had missing values on more than five variables. Of course, any attempt to reconstruct the scale faces the same problem. The distribution of the number of missing variables in the seven-point scale for con- victed robbers and burglars is shown in Table 16; the distribution for jai] inmates is also shown separately. For slightly more than one-half the sample (N = 464), none of the seven variables has missing values; 36 percent of the sample has one variable missing, ant] another 12 percent is missing two or more variables in the scale. fail inmates are only 22 percent of the sample but account for 46 percent of cases that have missing values for one or two variables, largely because of the problem with the past-convictions vari- able. The seven-point scale is particularly sensitive to missing information because it affects an inmate's maximum scale score. For example, all respondents sur- veyect in jai] (where Tow-rate offenders are presumably overrepresentec3) have a possible total score of only six, but are being compared with other respondents whose maximum score can be seven. In this case the missing clata would spuri- ously improve the apparent predictive accuracy of the scale. Other, less preva- lent cases of missing data could have other effects. The problems of missing data that are associated with the seven-point scale are not easily resolved. One "solution" is to omit the major source of error from the scale- the past-convictions variable- and to redefine the prediction scale as one with six predictors. Unfortunately, this variable is also the only official record measure in the scale, but it is not strongly related to high offending frequencies. In some ofthe following analyses, especially the tests of predictive accuracy, a six- point scale, which excludes the past- convictions variable, is tried and the sen- sitivity of the results to this change is assessed.l5 Accuracy of Scale To simplify his analysis, Greenwooc! colIapsecI the seven-point scale into three predicted offense-rate categories: low- rate offenders (scores of 0 or 1), meclium- rate offenders (scores of 2 or 3), and high- rate offenders (scores of 4 or greater).~6 One way to measure the effectiveness of this prediction scale is to compare aver- age offense rates among the three pre- clictecl groups. In Table 17 both the means reported by Ranc] and those gen i5Another alternative that would correct for all types of biases introduced by missing data (not just those related to the past-convictions variable) would be to multiply each respondent's score on the seven- point scale by the fraction: 7/(7 - number of missing variables). However, this solution produces noninteger values for respondents' scale scores (e.g., 3.5), which would make comparisons between scales difficult. i6In an effort to maintain comparability between this reanalysis of the Rand inmate survey and the Rand results, Greenwood's scale cut points (~1, 2~3, 4 or more) were used for most of the analyses involving the seven-point scale. But the cut points probably should be based on state-specific distribu- tions of scale scores for each crime type. Later in this reanalysis additional findings for California robbers are presented using cut points that equalize the marginal percentage distributions for the predicted groups with the collapsed categories of reported offense rates (50 percent-low, 25 percent me- dium, 25 percent high).

192 CRIMINAL CAREERS AND CAREER CRIMINALS TAB:t~E 17 Mean Reporter! Offense Rate and Other Statistics for Predicted Low-, Medium-, and High-Rate Offenders in the Three Sample States: Rand and a Reanalysis Predicted Reanalysisa Randb State Offense Rate Percent A2~A75 AA Robbery California Low 17 0-1.3 0.92.2 (N= 166)C Medium 35 0-4.9 8.111.0 High 48 2.5-36.0 20.830.9 Michigan Low 35 0-0.9 2.26.1 (N= 142) Medium 50 0.6-7.4 7.411.7 High 15 0-22.7 9.520.6 Texas Low 40 0-0.9 1.31.4 (N= 114) Medium 41 0~.9 3.25.4 High 18 1.6-11.8 5.97.7 Burglary California Low 25 0-2.2 7.212.6 (N= 151) Medium 41 0-18.2 33.187.6 High 34 5.5-174.6 83.2156.3 Michigan Low 24 0-2.8 15.9a71.6 (N= 113) Medium 50 0-10.6 21.834.0 High 26 0-7.0 42.2101.4 Texas Low 34 0-1.6 3.86.0 (N= 199) Medium 48 0-5.0 8.020.5 High 18 1.2-74.1 22.451.1 aThese columns give some information about the three predicted offense-rate groups: the percentage of cases in each group, the range of reported offense rates from the 25th to the 75th percentile, and the "truncated means" all offenders who reported offense rates greater than the 90th percentile have their rate set at the 90th percentile (calculated separately for each state and offense type). See Greenwood (1982:56) for other details. bTruncated means. Source: Greenwood (1982:Table ES.1). CThe N's in this table are from the reanalysis and differ from those reported by Greenwood (1982) in Table ES.1 (p. xvii) primarily because his N's include cases for which A could not be calculated. Thus, the N's in this table are not the ones from which his truncated mean offense rates were calculated. atTwo respondents in this category reported an annual crime rate of over 150 burglaries, which inflates the estimate of A for predicted low-rate burglars. crated from this reanalysis are presented. A comparison of the last two columns shows that Me overall pattern of increas- ing average As for the low-rate to the high-rate offender groups, which was re- ported by Rand, is confirmed in this reanalysis. However, We means in this reanalysis are much Tower than the RancT estimates. The large differences in the two estimates of mean offense rates re- flect the lower, recomputed estimates of A and the redefinition of active burglars.~7 i7In addition, it was learned that the estimates of A for all the analyses concerning the seven-point These results confimn Mat estimates of A, especially the mean, are very sensitive to alternate me~ocls of computation. The scale were actually Rand's maximum estimates of A, ~ , ~ _ not the average of the minimum and maximum estimates, which was used by Chaiken and Chaiken. (Greenwood provided the computer source codes that described his estimates of A.) The estimate of A computed for this reanalysis is much closer to Rand's minimum estimate (see Table 9~. Using the maximum estimate of individual offend- ing frequencies may be a conservative choice for partitioning inmates into low-, medium-, and high- rate offending groups, but it seriously inflates the three average, within-group estimates of A that are used later in Greenwood's analysis.

THE RAND INMATE SURVEY: A REANALYSIS Tower means may alter the estimated in- capacitation effects that Greenwood re- ports. Despite the apparent differentiation of predictecI groups basest on average values of A, a closer look at the distribution of A in each group reveals considerable over- lap. The A25-A7s statistic that appears in Table 17 is the range of As from the 25th to the 75th percentile for inmates pre- dicted to be in a particular group. For example, of the 80 Califomia inmates convicted of robbery who were predicted to be high-rate offenders, 25 percent re- ported fewer than 2.5 crimes per year, the midctIe 50 percent reported a crime rate between 2.5 and 36 per year, and the other 25 percent committee! more than 36 robberies in the perioc! before their ar- rest. The amount of overlap between the 25th and 75th percentiles across the three predicted groups is quite surprising. The values of A for all the medium- and high- rate groups overlap to some extent with the Tow- and meclium-rate groups, respec- tively. Moreover, in some instances pre- dictec] Tow-rate and high-rate robbers have similar rates of offending. In Michi- gan at least 25 percent of both robbers and burglars predicted to be high-rate offenders actually reported not commit- ting any robberies or burglaries at all. Thus, the seven-point scale does not a(l- equately identify which respondents are Tow-, medium-, or high-rate offenders at the state level when the distributions of A are compared across the three groups. Important interstate differences are also evident in Table 17. The scale iden- tifies high-rate robbers much better in CaTifomia than in Michigan and Texas, perhaps because in Califomia the distri- bution of A for robbery is especially skewed. The distinction between pre- clictec] high-rate and meclium-rate rob bers in Michigan and Texas is especially ~s .. . . In her critical review of Greenwood s analysis, poor. But CaTifomia s high-rate robbers Cohen (1983) pointed out that in Table 4.S of the committed more crimes than similar rob- Rand report the cut points partitioning offenders on ]93 hers in Michigan or Texas: the estimates of A at the 90th percentile for convicted robbers in the three states were 66, 29, and 13, respectively. The three average rates for precticted high-rate robbers 2().S 9.5 and 5.~follow this pattern. ~' . . ~ . ~ ~ 1 1 1 hese state cllnerences raise closets about the generaTizability of the scale as a prediction instrument for convicted! rob- bers outside the state of California. Finally, in both the robbery and the burglary analyses the predicted [ow-rate offenders are identified surprisingly well. (The exception is Tow-rate burglars in Michigan, but see footnote ~ to Table 17~. Other (lata (not presented in tabular form) show that 93 percent of the predictecl low-rate robbers and burglars (N = 255) reported committing fewer than six rob- beries or burglaries per year. These fincT- ings may be particularly relevant to the use of the scale in sentencing decisions, and they will be explored more fully below. A more common method of evaluating a prediction instrument and its cut points is to determine what fraction of respon- dents are correctly classified. In this case (he predicted offense rates are tabulated against actual offense rates, A, using three preclictec] groups (based on scale cut points) and three groups based on self- reports of crimes committed. fRecall that actual offense rates are partitioned into Tow-, medium-, and high-rate categories using the 50th (meclian) and the 75th percentile values as cut points.] In the Ranc] report Greenwood presented data of this type that compared respondents' predicted offense rates with self-reportecl offense rates structured according to these cut points. A replication of this prediction table from the reanalysis is presented in Table 18, and Greenwoocl's figures appear in a footnote to the table.

194 CRIMINAL CAREERS AND CAREER CRIMINALS TABLE 18 Distributions of Offenders by Predicted and Actual Offense Categories (percentages; N = 886) Predicted Offense Self-Reported Offense Rate Rate (Score Values) Low Medium High Total Low (0-1) 22 5 2 29 Medium (~3) 22 12 10 44 High (4-7) 7 8 12 27 Total 51 25 24 100 NOTE: The cell percentages, based on N = 781, reported by Greenwood and corrected by Cohen (1983) are low: 20, 5, 2; medium: 22, 12, 9; high: 8, 9, 13. The individual cell percentages for the two tables are very similar and differ by only one or two percentage points. Based on this reanalysis, the percent- age of respondents correctly classified by the seven-point scale is 46 percent (the sum of the cliagonal entries), which is slightly higher than Cohen's corrected figure of 45 percent for Greenwood's pre- ctiction table (see Cohen, 1983~. On the other hand, 54 to 55 percent of the con- victed robbers and burglars are miscIassi- fiec3. However, these overall rates mask differences across states and offense types, which were not discussed in the the basis of their reported offending rates resulted in a distribution of 30 percent (low rate), 42 percent (medium rate), and 28 percent (high rate). But ear- lier in the report Greenwood partitioned the sample into 50 percent (low), 25 percent (medium), and 25 percent (high) to identify variables that were related to high rates of offending. Cohen recalculated Greenwood's prediction table based on the original categories; see Table 18 for a comparison of the replicated prediction table and Cohen's corrected figures. In his report Greenwood does not explain why the cut points for actual offending categories were changed, but it was learned subsequently that it was done to equalize the marginals for predicted rates and reported rates in the table (Abrahamse, 1983, personal communication), i.e., to equate the base rate to the selection ratio. This redefinition tends to reduce the rate of false-positive errors in classifying high-rate offenders. Therefore, the accuracy of the improvement-over-chance classification implied by Greenwood's Table 4.8 combines the effect of scale accuracy with the artifactual effect of the redefini- tion. Rand report. Moreover, earlier results of this reanalysis suggest that the scale may be differentially predictive for high- and Tow-rate offenders. In Table 19 data are presenter] that a(l(lress these questions. The first two columns give accuracy rates for Green- wood's scale among predicted low- and high-rate offenclers. For Greenwood's (lata, Cohen (1983) calculated that 76 per- cent of the respondents predicted to be Tow-rate offenders reporter] Tow rates of robbery (or burglary), but only 45 percent of the respondents predicted] to be high- rate offenders actually reported high of- fencTing frequencies. As seen in Table 19, this reanalysis of the Rand data also shows greater accuracy rates among pre- lictecl low-rate offenders than high-rate offenders. However, these differences are due in part to the different base rates of the two groups 50 percent of the in- mates were defined as low-rate offenders and 25 percent were defined as high-rate offenclers. Accuracy rates also differ substantially within crime types and across states. (The data on which the figures are based are presented in Appendix Table A.3.) Pre- dictions of low offense rates for both rob- bery and burglary are more accurate in California than in the other states, but predictive accuracy for high-rate offend- ers is consistently higher in Texas. More- over, of those classified as high-rate of- fenders, 60 percent of the robbers in

TtIE RAND INMATE SURVEY: A REANALYSIS TABLE 19 Measures of Predictive Accuracy anc! Percent Relative Improvement Over Chance (RIOC) for Different Respondent Subgroups and Types of Prediction Moclels (percentage) 19S Accuracy RIOC' Subgroup or Model Low-ratea High-rateb Low High Greenwood 76 45 48 35 Reanalysise 76 45 50 31 Robbers California 93 40 86 57 Michigan 78 41 55 21 Texas 69 52 39 38 Burglars California 84 48 67 48 Michigan 74 34 44 19 Texas 68 61 33 48 Six-variable Scalef 72 47 48 27 Unambiguous Casesg 76 43 43 30 aThe percentage of respondents predicted to be low-rate offenders (scoring 0 or 1 on scale) who actually reported low rates of robbery or burglary (below the median for their state and offense type). bThe percentage of respondents predicted to be high-rate offenders (scoring 4 or more on scale) who actually reported high rates of robbery or burglary (above the 75th percentile for their state and offense type). CThese measures adjust for the difference in base rate and are calculated according to the formula provided by Loeber and Dishion (1983). See text and note 20 for details. The figures in the first two columns are based on Cohen's (1983) correction of Greenwood's data (N = 781). eThe sample is all convicted robbers and burglars (N = 886~. Prediction scale without one variable past conviction for robbery or burglary- and using same cut points. BIncludes only respondents for whom A could be unambiguously calculated, and respondents with only slight ambiguity in responses to questions about number of crimes committed (N = 568). California and 66 percent of the burglars in Michigan are incorrectly classified, compared with 39 percent of the burglars in Texas. Some ofthis interstate variation is the result of different distributions of predicted offense groups in the three states (see Table 17~. The difference be- tween this selection ratio and the base rate (reported offense rates) affects the accuracy rate in a particular table. How- ever, it is difficult to determine how much ofthe interstate differences in Table 19 is a~tifactual. Another measure of predictive accu- racy, the percent relative improvement over chance (RIOC), also indicates siz- able interstate differences. Loeber and Dishion (1983) recommend this measure because it relates improvement over chance to maximum possible accuracy, which is an artifact of the difference be- tween actual offending patterns ("base rate") and predicted patterns ("selection ratio"~.~9 Using RIOC, the accuracy ofthe it'The formula to compute the relative improve- ~nent over chance is: RlOC = Percent Total Correct - Percent Random Correct . Percent Maximum Correct - Percent Random Correct The advantage of this measure is that "percent maximum correct," which is the maximum ceiling or accuracy for a given table, adjusts for differences between the base rate and the selection ratio. This measure can be helpful when comparing the effi- ciency of prediction instruments across different

196 prediction scale across different sub- groups and models is reported in the third anct fourth columns in Table 19 in terms of how well the scale predicts Tow- and high-rate offenders. The RIOC measures show extreme variability in the predictive power of the scale across states and offense types, even after adjusting for the different distribu- tions of predictecI offense groups. In gen- eral, the predictions of robbery rates are better than the predictions of burglary rates, and Tow-rate offenders appear to be predicted! better than high-rate offenders. (The one exception is Texas, where the accuracy rate is higher for high-rate bur- glars.) The scale appears especially strong in California. On the other hancI, high-rate robbers and burglars in Michi- gan are poorly identified; the scale actcts only a 20 percent improvement over chance. In part, this variation may reflect differences in the inmate populations. However, the fixed scale cut point of 4 leacis to different selection ratios in the different states. The variations in the se- lection ratios in turn artifactually affect the measures of predictive accuracy. These issues are not explored here but are clevelopect in Volume I (Chapter 6) of the panel's report. Two modifications to the original scale and sample produced only moderate ef studies or different samples, but it must be inter- preted in light of the respective selection ratios and base rates (see Volume I, Chapter 61. For Table 19, the medium- and high-rate offenders are combined to compute RIOC for low-rate predictions, and the low- and medium-rate offenders are combined to compute RIOC for high-rate predictions. The fol- lowing example illustrates how RIOC was calcu- lated for predicted low-rate California robbers Me uppermost led cell using the marg~nals from the collapsed 3 x 3 table (data are presented in Appen- dix Table A.3): 26- [(84 x 28)/166] 28- [~84 x 28)/166] CRIMINAL CAREERS AND CAREER CRIMINALS fects on predictive power using the stan- clard measures of accuracy. (See Appen- clix Table A.3 for the complete prediction table.) First, a six-variable scale (Green- wood's scale with prior conviction for the same offense removed as a predictor vari- able) was constructed. This reduced po- tential biases related to missing data on the seventh variable, but it also removed potentially useful information. This scale appears less accurate in predicting Tow- rate offenders than the full scale and slightly more accurate in predicting high- rate offenders. However, when the differ- ences in base rate and selection ratio are controlled in the RIOC measure, the scaTe's predictive capability for high-rate offenders is appreciably Tower than the full scale, as would be expected. The second mollification, excluding cases with serious response ambiguity (see Table 11 and related text), did not significantly affect the accuracy of either Tow- or high-rate predictions. Surpris- ingly, once the base rate/selection ratio problem is taken into account, removing ambiguous survey responses reduces the ability of the scale to identify Tow-rate offenclers, but it does not affect the accu- racy of the high-rate predictions. Thus, the accuracy of Greenwood's seven-point scale cannot be aclequately assessed from data in a "pooled" predic- tion table. The substantial differences in the measures of predictive accuracy across states and offense type indicate that the scale does not uniformly identify high-rate robbers and burglars. Predictive accuracy for high-rate offenders, accord- ing to the RIOC measure, is best for California robbers and worst for Michi- gan robbers and burglars. Convicted rob- bers (and burglars, but to a lesser extent) in California prisons and jails appear to be quite different from other respondents in the Rand inmate sample. This reanaly- sis thus far has shown that robbers and burglars in California, when compared = 86 percent

TtIE RAND INMATE SURVEY: A REANALYSIS with other subgroups, reported the high- est As (see Table lO), the most extensive juvenile records (see Table 13), and the greatest involvement in multiple drug use both as juveniles and adults (see Ta- ble 131. Moreover, in regression analyses the seven-point scale explainer] more variance in self-reported robbery and bur- glary frequencies for California inmates than for respondents in other states. Thus, the scale is more sensitive to the at- tributes of high-rate offenders in CaTifor- nia than to the attributes of offenders elsewhere. Finally, the scale appears to be able to predict Tow-rate offenders more consis- tently, and those predictions are more accurate than predictions of high-rate of- fending. It is possible that Greenwood's scale and other similar prediction instru- ments could be used to identify the least active criminals since the absence of the seven scale attributes seems to coincide with Tow rates of criminal activity under six crimes per year. The data presenter! here on the predictive accuracy of one proposed prediction instrument suggest that prediction tools in the criminal jus- tice system could play a role in deciding who should be sent to prison and for how Tong and also who should not be sent to prison. However, these results may be a consequence of particular characteristics of the inmates surveyed from the prisons and jails in California, Michigan, and Texas. The results of this reanalysis do suggest that any prediction instrument may re- quire some changes for the particular characteristics of the population. The considerable differences in the predictive capability of one prediction instrument reported here indicate that prediction scales developed with one population should be tested extensively before they are applied to different populations. Fur- ther, the scale proposed by Greenwood can only diminish in its predictive power 797 when applied prospectively to popula- tions of all convicted offenders, whether incarcerated or not. In view of the poten- tial value of prediction instruments ant] the sizable error rates presently associ- ated with them, more research in devel- oping prediction instruments appears warranted. Incapacitative Effects Using a Prediction Scale The primary objective of the Green- wood report was to identify high-rate of- fenders in the Rand sample and to deter- mine whether targeted or "selective" incarceration could lead to decreases in crime, decreases in the prison population, or both. In this last phase of the reanaly- sis, Greenwood's procedures are used to estimate these incapacitation effects for robbers in California, and the sensitivity of his results to alternate estimates of A and to the reconstruction of his prediction scale are assessed. The Incapacitative Elect of Incarceration Incarcerating convicted offenders not only punishes offenders for their criminal behavior but also prevents them from committing crimes in the community. The reduction in crime directly attribut- able to incarceration is referred to as "incapacitative effect." Calculating this effect requires information about criminal justice system operations and criminal behavior. In particular, one needs to know the expected time spent in prison for a crime and the value of A for active offenders. The longer the expected incar- ceration time per crime- which is a func- tion of the probabilities of arrest (qj, con- viction and incarceration (J), and average sentence length (S) and the larger the average crime rate (A), the greater the estimated incapacitative effect. As a first

198 approximation, these relationships can be mathematically expressed as 1 ~ - . 1 + AqJS This model was developed by Avi-Itzhak and Shinnar (1973) and expanded in other papers (Shinnar and Shinnar, 1975~. Cohen (1978, 1983) provides an excellent discussion ofthe incapacitation literature, including a comprehensive review of pre- vious research that has estimated the incapacitative eject of incarceration. Greenwood suggested that the extreme variation in A that had been observed in the Band data warranted (1isaggregating this model to develop estimates for of- fenders with low, meclium, and high val- ues of A.20 Greenwood ( 1982:xiii) re- ported that with this revised moclel, the amount of crime prevented by any given incarceration level can be increased if we lengthen the terms of those in the high-rate group and shorten the terms of those in the low-rate groups.... this type of sentencing policy fis called] "selective incapacitation." Thus, (lisaggregation might produce much greater incapacitative ejects than those estimated from a model based on an average A with all offenders treated ho- mogeneously. Specifically, Greenwood proposed classifying offenders into three groups based on their predicted offending rate, which is cleterminec3 by their scores on the seven-point scale. To analyze the ef- fects of selective incapacitation, three pa- rameters of the basic mo(le] are allowed to vary across the o~encler groups: indi- vidual offending frequencies (A), proba- bility of incarceration given conviction 20Marsh and Singer (cited in Cohen, 1983) origi- nally demonstrated, with hypothetical data, that larger reductions in crime might be possible if A was assumed to vary in the criminal population. Cohen (1978) also discussed the statistical underpinnings of such a revised model. CRIMINAL CAREERS AND CAREER CRIMINALS d), and the average time served for those incarcerated (S). Then, one can calculate the expected amount of crime contrib- ute(1 (ancl prison space usecl) by the Tow-, medium-, and high-rate groups under the current sentencing policy and contrast that with the expected amount of crime (and prison space) under a selective sen- tencing policy for example, one that sends high-rate offenders to prison for Tong terms and all other Fencers to jail for shorter terms. Several critical assumptions underlie this moclel ancl Greenwood's application of it to the Rand inmate data (see Cohen, 1983; Blackmore and Welsh, 1983~. Of particular concern here are the accuracy of average estimates of A, the distribution of inmates and the offender population across the three o~ense-rate groups, and the stability and continuity of A over time. This reanalysis explicitly tests the sensi- tivity of Greenwood's results to variations in estimates of A and to assumptions about the total offender population. The impli- cations of interstate differences in esti- mating incapacitative effects are also acI- lressec3. A Selective Incapacitation Mode! To estimate the proposer! mode] of incapacitative effects, Greenwood neec3- ecl information about how offenders in the three states are currently sentenced. He had this information for California only and thus most of his analyses are focused on California robbers and bur- glars.2i This reestimation of the inca 'iGreenwood also presents some estimates for Texas robbers and burglars using the California values for the probability of arrest, conviction, and incarceration. Since no jail inmates were included in the Texas sample, it is unclear how Greenwood arrived at estimates of the prison and jail popula- tions in Texas. But more importantly, it is inappro- priate to use California's sentencing policy as a benchmark for estimating the potential incapacita

THE RAND INMATE SURVEY: A REANALYSIS pacitation moclel used by Greenwood also focuses on the effect of selective sentencing policies on robbery rates ant] on the prison population in California. In adclition, an attempt was made to repli- cate Greenwood's results for CaTifomia robbers using his published values for all variables in the moclel (GreenwoocI:77, 10~118). Greenwood tested the model with a highly selective sentencing policy that would double the "expected sentence length" from approximately 4 years to 8 years for pre(licted high-rate robbers22 and would send all other robbers to jail for 1-year terms (Greenwood:79~. He re- ports that a large incapacitative effect could be achieved in California: a 20 per- cent reduction in the robbery rate without any increase in the prison population. But this conclusion is not supported by other data Greenwood presents. The graph that is supposed to depict this relationship is live effects of a selective sentencing policy in Texas. As shown earlier, California and Texas inmates are very different in their individual offending frequen- cies and in their values on the predictor variables (especially juvenile criminal history and use of ille- gal drugs), which suggests that in these two states the sentencing policies or the offender populations are not at all alike. Thus, Greenwood's estimates of incapacitative effects in Texas using the California sentencing parameters are likely to be significantly In error. 22The prison term assigned by the judge after conviction is different from the "expected sentence length," which is used in calculating incapacitative effects. Although state policies differ, a convicted offender usually serves only one-half to two-thirds of his sentence because of reductions for "good behavior." Thus, California robbers who are pre- dicted to be high-rate offenders reported an ex- pected sentence length of about 4 years but were probably given prison sentences of 6 to 8 years. (Official information on expected date of release was not available for all California inmates; therefore, the self-report measure was used as a substitute.) Increasing the expected sentence length to 8 years actually means that the prison sentence for robbery for this high-rate group would have to be 12 to 16 years. 199 slightly at odds with the text (Green- wood:78, plot 6). More important, there are other difficulties with the data that underlie it. In an earlier attempt to recal- culate Greenwood's results, Cohen (1983) was unable to replicate the results precisely using the data Greenwood re- ports. According to Cohen's calculations, the maximum incapacitative effect (with 8-year expected terms for high-rate of- fenders) is a 13 percent reduction in crime with an 8 percent decrease in the prison population. The recalculation of Greenwood's findings concerning the po- tential incapacitative effects of his selec- tive sentencing policy by this author, also using Greenwood's published numbers, produced the same results found by Cohen (see Appendix C). Thus, this replication and that of Cohen confirm that some reduction in the Cali- fornia robbery rate might be possible by selectively imprisoning the preclicted high-rate offenders but that the maximum potential using the hypothetical sentenc- ing policy is about 13 percent, not 20 percent as Greenwood reports. Any devi- ation from the assumptions of the mode} will probably Tower this estimate still further. In fact, Cohen shows that the 13 percent effect is sensitive to some of the input values used in the model, particu- larly the stability of A (1983:Figure 3) and the distribution of offenders across crime- rate categories (1983:Figure 41. In previ- ous sections of this paper, Tower average estimates of A were reported for CaTifor- nia robbers: 0.9 (Iow rate), 8.1 (medium rate), and 20.8 (high rate), compared with the 2.0, 10.1, and 30.8 reported by Green- wood. Slightly different values were also computed for several other parameters in the incapacitation model.23 The alterna ~3In this reanalysis, the distribution by offense- rate group for prison and jail inmates in California was slightly different from that reported by Green

200 CRIMINAL CAREERS AND CAREER CRIMINALS 100 95 a) ._ a) Q 90 f r \ \ a) ~-. . . o , a) 85 a) ~ Q ~ of o ~ c, 80 a' ~ Q ~ x ~ . . . . . Using variable values from reanalysis Using variable values from Greenwood ( 1982) - Using 1 2-year sentence length for high-rate group 85 90 95 100 105 Number of Offenders I Incarcerated for Robbery as Percent of Current Level for Robbery FIGURE 3 Comparison of reanalysis estimates of incapacitative effects for highly selective sentencing policy with effects reported in Greenwood (19821. wood (in parentheses), which also altered some of the other parameter values: Pre dicted Group Jail Rate Sample Low 20 (24) Medium 19 (14) High 14 (14) Estimated Total Prison Incarcerated Sample Population 14 (17) 2,865 (3,480) 45 (43) 4,942 (4,401) 66 (66) 5,942 (6,099) Estimated Average Sentence Length (months) 46.7 (49 5) 56.3 (53~3) 48.9 (50.6) The reconstruction of the scale changed the classi- fication of some offenders in the three groups, which probably accounts for the different average ex- pected sentence lengths. The full set of revised parameters for Greenwood's incapacitation model is available in Appendix Table A.4. live estimates of these and other parame- ters for the moclel were used in the reanalysis of the incapacitative effects of selective sentencing policies in CaTifor- nia. The full set of revised parameters for the incapacitation model is given in Ap- pendix Table A.4. Figure 3 shows two estimates of the potential incapacitative effects of a highly selective sentencing policy for convicted robbers in California; the dashed line represents a corrected interpretation of Greenwood's data and the solid line rep- resents the reanalysis. The reduction in

THE RAND INMATE SURVEY: A REANALYSIS the robbery rate reported by Greenwood is relatively unaffected by using Tower estimates of A. This reanalysis and Cohen's (1983) replication of Green- wood's data both indicate the possibility of about a 13 percent reduction in rob- bery. But the prison population would remain essentially unchanged using esti- mates of the input variables obtained from the reanalysis.24 Further reductions in the robbery rate beyond 13 percent can only be achieved by increasing the expected sentence length. The end points of the dashed and solid lines in Figure 3 and those in Greenwood's report are based on the hy- pothetical sentencing policy of 1-year jail terms for Tow- and medium-rate offenders and about 8-year expected terms for high- rate offenders. Thus, any extension to a lower crime level actually involves a change in the sentencing policy. To see if Greenwood's finding of a possible 20 per- cent reduction in robbery could be achieved, Greenwood's hypothetical pol- icy was revised and the average time served for high-rate offenders was in- creased by a factor of 3, to slightly over 12 years. This modification is represented by dotted lines in Figure 3. These data reveal that with extremely stiffexpected sentences for high-rate rob- bers (actual prison sentences would prob- ably be 16 to 24 years), the robbery rate 24The model that is used to estimate these inca- pacitative effects is based on a series of calculations involving a wide range of magnitudes; therefore rounding and truncation error (for example, using 2 decimal places instead of 4 or 5) may slightly alter the estimates of changes in crime rates and prison populations. Details of the model and the interme- diate calculations can be found in Cohen (1984a) and Volume I (Chapter 5) of the panel's report. The projections reported here for incapacitative effects are conditional on assumptions stated in those sources, and actual effects are likely to differ from these projections because the assumptions may be violated in ways that are discussed later in the paper. 20] couIct be reduced by only 18 percent. Moreover, the prison population might have to be increased (according to the reanalysis estimates) to accommodate the longer sentence lengths. But more impor- tant, a sentencing policy Mat gives 1-year jail teens to most convicted robbers arid sentences a small group of predicted high-rate offenders (which inclucles an error rate of at least 50 percent, according to the prediction tables presented earlier) to about 20 years would represent ex- ~eme disparity in sentencing. Incapacitative effects for Michigan were not estimated in Me Greenwood report because the necessary data on cur- rent sentencing policies were not avail- able. The data for Michigan robbers were obtained for the reanalysis, and the incapacitative effects that wouIc] be ex- pected under Greenwoocl's model were computed.25 The results were quite dif- ferent from hose for CaTifomia. With S- year sentence lengths for predicted high- rate robbers and 1-year jai] terms for all over robbers, the robbery rate in Michi- gan would increase by 33 percent, but Me prison population would decrease by 25The data on current sentencing policies in Michigan are taken from the official records of a large sample of Michigan arrestees (Blumstein and Cohen, 1984, personal communication) and state- level summary data supplied by the Michigan State Police. These sources gave nearly identical esti- mates of the parameters needed for the incapacita- tion analysis based on Michigan robbers and they were averaged to arrive at the following estimates: conviction rate .44; number of robbery arrests in 1977~,281; probability of incarceration given con- viction .86; prison commitment rate .86; jail com- mitment rate-.05. These parameters were substi- tuted into Table B.4 (Greenwood, 1982:112) to estimate current numbers of robbers in Michigan prisons and jails. Then, those estimates and data on the 150 convicted robbers in the Michigan subsam- ple were used to generate a table similar to Table B.6 (p. llS) for Michigan. Further details about estimating the potential incapacitative effects of a selective sentencing policy for robbers in Michigan are available from this author.

202 nearly 50 percent. The hypothetical pol- icy is clearly not satisfactory in Michigan because incarcerated high-rate offenders, as cleaner] by a minimum score of four on the seven-point scale, are apparently a very small group in Michigan prisons and jails, compared with CaTifomia. More- over, all convicted robbers in Michigan are aIreacly serving Tong prison terms (an average of 5 years) and few robbers are sentenced to jail. Thus, in Michigan a policy that reserves Tong prison sentences for only the small group of predicted high-rate offenders actually would in- crease the crime rate and reduce the prison population. This would occur be- cause most robbers (those cleaned as low- anc! meclium-rate) would spend a smaller portion of their offending careers in prison or jail under this policy than under Michigan's current policy ant] would have more "free time" in which to com . . mat more crimes. The crime rate was also increased when Greenwood applied his incapacita- tion moclel ant! selective sentencing pol- icy to the Texas robbers and burglars (GreenwoocI:79~811. In Texas preclicted high-rate offenders, using the seven-point scale, were also a small group; conse- quently, Greenwooc3's selective sentenc- ing policy would reduce the prison pop- ulation but would not reduce the robbery or burglary rate. Finally, one important parameter of the original incapacitation moclel was omit- ted from the Greenwood version- of- fender's career length. Other analyses of the Rand data reveal that when career length is included in the model for CaTi- fomia, estimates of crime reduction Mat could be achieved by a selective sentenc- ing policy drop to about 5 to 10 percent (Cohen, 1984a; Spelman, 19841. In a re- cent report on the duration of criminal careers, Blumstein, Cohen, and Hsieh (1982:55) estimated that the maximum mean resi(lual career length for robbery CRIMINAL CAREERS AND CAREER CRIMINALS (the number of years left in a criminal career at any given age) is only 7 years. Therefore, many of the targeted high-rate offenders wouIcl likely have ended their careers before the ens! of their 8-year prison term anyway, in which case the projected reductions in crime would be overstated. Thus, these and other analy- ses suggest that, under the best assump- tions, significant reductions in crime can- not be easily achieved by identifying the high-rate offenders and targeting them for long prison terms. Selecting Scale Cut Points One of the fundamental parameters of Greenwood's calculations is the choice of cut points on the seven-point scale that (refines the distribution of incarcerated ofl~enclers across the three preclictecl oF fence-rate groups. The cut points are used to estimate the total offender population in California and the probability of prison (versus jail) for convicted offenders in each group. In Greenwood's incapacita- tion analysis, the Tow-, medium-, and high-rate groups are clefinecl by the scores ~1, 2<, and 4-7 on the seven-point scale derived from the survey data. The distri- bution of these scores within the inmate sample is used, along with information about California's current sentencing pol- icy for robbers, to estimate the total an- nual jail and prison population in CaTifor- nia. Using these methods, predicted high-rate offenders turned out to be 43 percent of the incarcerated robber popu- lation (Greenwood:771. Greenwoocl introcluced his scale, how- ever, as a device for identifying a rela- tively small group of high-rate offend- ers specifically, the most active 25 percent of the convicted robbers, accor(l- ing to their self-reports. (Chaiken and Chaiken, 1982a, chose the top 20 per- cent.) As Loeber and Dishion (1983) noted generally, this excess of selection

THE RAND INMATE S URVEY: A REANALYSI S ratio (43 percent) over base rate (25 per- cent) guarantees a false-positive rate of at least 18 percent. To maximize predictive efficiency, the seiecuon ratio (me per- centage of respondents predicted to be high-rate offenders) should be equivalent to the 25 percent base rate (the percent- age of respondents defined as high-rate according to their reported crime rates). Therefore, in the reanalysis the model was reestimated to assess the sensitivity of the results to alternative scale cut points for Califomia robbers-. Adjusting the cut points to equalize approximately the selection ratio and base rate and substituting the lower aver- age estimates of A for the three groups dramatically altered the potential crime reductions associated with Greenwood's hypothetical selective sentencing policy. The cut points of the predictor scale were changed to ~2 (low). 3 - (me- dium), and ~7 (high).26 The values of A . . , . ~, 26There are actually two closely related issues: one is substantive and the other is technical. The cut point decision is also a policy issue-how much error in prediction is acceptable and how are pre- dicted high-rate offenders to be defined (e.g., having four or five of seven attributes), given the character- istics of a specific offender population. The techni- cal issue relates to the estimated distribution of offenders across the three offense-rate groups. In Greenwood's model (and this reanalysis), this esti- mate is dependent on the cut points because the distribution of the three groups defined by the cut points in the sample is used, in conjunction with other parameters, to estimate the distribution of the three groups in the general offender population. These estimates were necessarily based on the small number of convicted robbers in the California inmate sample (N = 178~. Changing the cut points of the prediction scale reduces the high-rate group to 22 percent of the estimated total incarcerated population, which is closer to the 25 percent figure that Greenwood initially thought would be appropriate. Once dif- ferent proportions of "street time" among the three offender groups (most for the low-rate group and least for the high-rate group) are taken into account, Greenwood's model estimated that about 13 percent of the total population of robbers in California are 203 for the newly defined groups indicate less differentiation between the medium- and high-rate groups-2.2 (Iow), 16.9 (me- dium), and 20.8 (high), compared with 0.9, 8.1, and 20.8 using the other scale cut points. Surprisingly, changing the cut points did not alter the average high rate A, which highlights the difficulty of dis- tinguishing between medium-rate and high-rate offenders with the prediction scale. With the alternative scale cut points and resulting changes in the model's pa- rameter values, the California robbery rate would actually increase about 6 per- cent under Greenwood's selective sen- tencing policy, although the imprisoned population would decrease about 20 per- cent. As with the Michigan and Texas estimates discussed earlier, the h,vpothet- ical increase in the crime rate and the reduction in prison population would oc- cur because the large majority of robbers would spend a smaller portion of their careers incarcerated, under the assump- tions of this revised model, and so would be free to commit crime. Thus, it appears clear from these anal- yses that the potential incapacitative ef- fects derived from a model that assumes a selective sentencing structure are sensi- tive to the choice of scale cut points and high-rate offenders (see Greenwood:771; the reestimation here of the model with the revised scale cut points makes the explicit assumption that fewer California robbers (i.e., only 6 percent) are high-rate offenders. Of course, since it is impossible to know how many "active" robbers actually exist in any state or how they are distributed across low-, medium-, and high-rate groups, these numbers must be estimated. But using Greenwood's method could distort the estimates! number of offenders in each group if, for example, the incarcerated population contained an unusually large group of predicted high-rate of- fenders, as was the case in California. An alternate method would be to estimate the total offender population using seven groups (one for each score value on the scale) rather than the three groups arbitrarily defined by the cut points.

204 the nature of the offender population. The appropriate cut points on the predic- tion scale may depend on the definition of "high-rate" offender, which could dif- fer across states. The most active 10 per- cent of the robbers in the Texas sample each reported committing at least 15 crimes per year, but the top 10 percent of the California robbers reported 100 or more crimes a year (see Table 10~. Chaiken and Chaiken (1984:223) suggest that the low rates of robbery reported by inmates in Texas compared with inmates in Michigan and California could reflect unmeasured aspects of the environment on patterns of criminality. California o~- cials may be more willing to tolerate some forms of criminal behavior than their counterparts in Texas. The probable interstate differences in offender populations, criminal justice sys- tem practices, and projections of incapac- itative effects highlight the need for cus- tomizing the development of prediction rules, the selection of cut points, and the implementation of selective sentencing policies within each jurisdiction. Factors specific to the local situation should be consi(lered before any prediction instru- ment is adopted, even one having some degree of accuracy. Moreover, cut points for decision rules may also be influenced by local values as to the relative costs of the criminal behavior and the sanctions being imposed according to the rule (Blumstein, Farrington, and Moitra, 1985; Morris and Miller, 1985~. In any choice of cut points, the lower the cutoff defining the high-rate offenders, the greater the risk of incorrectly classifying some of- fenders in this group. SUMMARY AND CONCLUSIONS The single most important contribution of Ranc3's second inmate survey is the highlighting of the extreme skewness of CRIMINAL CAREERS AND CAREER CRIMINALS the distribution of A for a sample of known serious criminals. Although the technique used to elicit this information an(1 the Rand sample of incarcerated of- fenders may have introcluced errors into these estimates, the Rand study has sig- nificantly advanced our unclerstanding of individual patterns of criminal behavior. Although some minor differences exist in the precise numbers in the distribution, this reanalysis of the Rand data confirms that the distribution of A is highly skewed at least for the offenders sam- plecl from the prisons and jails of Califor- nia, Michigan, and Texas. Half the of- fenders report committing no more than five crimes a year, while a small but very important group may commit several hundred crimes a year. The estimates of A for robbery ant] burglary, however, are sensitive to choices in computation, such as the inter- pretation of ambiguous survey responses, the treatment of missing data, and the computation of the length of responclents' "street time." Moreover, the veracity of some respondents, particularly the large group of convicted robbers and burglars who clenied committing any robberies or burglaries and the few respondents whose reports implied annual rates of 1,000 or more robberies or burglaries, may be affecting the observed distribu- tion of A. Another problem is obtaining accurate annualized rates for those re- sponclents who are incarcerated for long portions of the observation period and who have intensive, but short, street time, or for those who commit crime sporadi- cally. Changes in the clesign of the Rand questionnaire or some analytic adjust- ments to the estimates of annual offend- ing rates may be necessary to provide more valid estimates of crime rates for such respondents. Finally, A varies con- siderably across the three state samples and further research is needed to deter

THE RAND INMATE SURVEY: A REANALYSIS mine whether this variation is- due to differences in the states' offender popula- tions or is a consequence of different selectivity arising from the criminal jus- tice processes in these states. The Rand finding that has received the greatest public attention is also the one about which the most questions are raised in this reanalysis: the Greenwood formulation of a particular scale for iclen- tifying high-rate offenders. A fundamen- tal problem relates to how well this iclen- tification can be accomplished in an operational setting and how well the Rand report demonstrates the feasibility of cloing so. Although the scale certainly does bet- ter than chance in all the jurisdictions examined, one would expect improve- ment from any scale that invoked the predictors it did and that was fitted to the sample data. There is no indication that Greenwood's scale would perform any better, even in California, than any other scale that has been used operationally. The relative improvement over chance varied consiclerably across the three states; the best performance was ob- served for California (57 percent for rob- bery and 43 percent for burglary) and the worst for Michigan (21 percent for rob- bery and 19 percent for burglary). The prediction scale also seems to work some- what better in identifying Tow-rate of- fenders than the high-rate offenders at whom it was targeted, even adjusting for the higher prevalence of low-rate offenc3- ers in the population. These results em- phasize the importance of each jurisdic- tion's developing and vaTiclating its own scale rather than simply applying the sev- en-point prediction instrument devel- oped by Greenwood or any other instru ment. If one could identify the high-rate of 205 tainly make it possible to reduce crime by selectively incarcerating those high-rate offenders. This reanalysis of the Ranct data found that Greenwood overesti- mated the anticipated reduction in the California robbery rate. Using a seven- item scale and a sentencing policy that would double sentence lengths for high- rate offenders, the most favorable effect achieved in the reanalysis was a reduc- lion of about 13 percent. However, the scale used to identify high-rate offenders is more sensitive to the attributes of those offenders in California than to the at- tributes of high-rate offenders elsewhere. If the same sentencing policy and predic- tion scale were appliecl in Michigan and Texas, the crime rate would probably increase because of differences in current criminal justice practices and offender populations in the three states. More importantly, even in California, the assumptions necessary to make the calculation inflate the estimate of inca- pacitation effects. The estimate of a 13 percent reduction in crime with a selec- tive sentencing policy, which has been demonstraterl only with California clata, will clecline further if any ofthe following obtain: 1. Predictive power decreases as the model is applied to any new population ("shrinkage") and especially to a popula- tion of all convicted offenders rather than prisoners; 2. The comprehensive self-report data used in the Rand analyses are replaced by less complete official records of the pre dictor variables; 3. The reports of A gathered retrospec tively in the Ranc] survey fait to persist into the future, especially after the longer periods of incarceration impliecl by the selective incapacitation policy; fenders prospectively, the extreme skew- 4. The criminal justice system limits ness in the distribution of A should cer- the proposecl policy through judicial dis

206 cretion or other adaptive responses in ways Mat reduce We disparity that arises from a sentence of 8 years for predicted high-rate offenders compared win 1 year for other convicted persons. Thus, future research is needed to identify characteristics of high-rate of- fenders and how those characteristics vary across offender populations. Re- search is also needed to develop and test locally appropriate, prediction-based se- lection rules to distinguish high-rate of- fenders from other offenders using oper- ationally available data. On the basis of this reanalysis, much more realistic esti- mates of the true operational effective- ness of a prediction instrument are needed before the current enthusiasm about the estimated reduction in crime through selective incapacitation is war- rantecI. APPENDS A: SUPPLEMENTAL TABLES TABLE A. 1 Comparison of the Cumulative Percentage Distribution of Rand Estimates of A with Esti- mates Produced from a Reanalysis of the Rand Data Robbery Burglary ~Rand Reanalysis Rand Reanalysis A < 1 13.1 16.8 9.4 < 2 24.9 33.0 20.8 < 3 35.3 41.8 32.7 < 4 43.8 52.9 40.1 < 5 49.4 56.1 48.2 < 10 65.8 71.9 62.1 < 20 77.5 81.2 70.3 < 30 82.7 85.4 75.0 < 40 85.1 87.4 77.3 <50 86.6 87.7 78.8 < 100 90.9 92.1 82.4 2 100 99.9 99.9 100.0 SOURCE: Chaiken and Chaiken robbery; 203, burglary). 12.1 26.5 35.8 46.0 52.2 65.7 73.1 77.0 79.1 80.4 83.9 99.9 (1982a:206, Variable CRIMINAL CAREERS AND CAREER CRIMINALS TABLE A.2 Regressions of Estimates of A on the Seven Variables in Greenwood's Scale: Inmates Convicted of Robbery or Burglary Unstandardized Coefficients Robbery Burglary I. Three States (N = 848) Past conviction Recent incarceration Early conviction Juvenile incarceration Recent drug use Juvenile drug use Recent unemployment Constant Adjusted R2 II. California (N = 311) Past conviction Recent incarceration Early conviction Juvenile incarceration Recent drug use Juvenile drug use Recent unemployment Constant Adjusted R2 III. Michigan (N = 245) Past conviction Recent incarceration Early conviction Juvenile incarceration Recent drug use Juvenile drug use Recent unemployment Constant Adjusted R2 IV. Texas (N = 292) Past conviction Recent incarceration Early conviction Juvenile incarceration Recent drug use Juvenile drug use Recent unemployment Constant Adjusted R2 -.01 .43a .11 .4oa .46a .48a .33a -.39 .12 -.01 l.l8a .12 .SOa .31 .62a .22 -.41 .20 .15 _ .71a -.01 .46 .s7a .39 .6la -.30 .08 -.14 .14 .19 -.11 .39a .29 .10 -.27 .05 .46a .28 .44a .18 gga .7la .47a -.37 .12 .58a .16 .32 .35 .soa .68a .27 -.54 .22 .44 .41 .64a .16 .78a .8oa .40 -.34 .12 .4oa .19 .39 -.005 .6sa . 7oa .62 -.22 NOTE: Missing data for the independent varia- bles were coded as 0. In the regression including all three states, 36 cases were excluded because of missing data for the dependent variable. For simi- lar reasons, 6 cases were excluded in California, 10 cases were excluded in Michigan, and 20 cases were excluded in Texas. ap < .05.

THE RAND INMATE SURVEY: A REANALYSIS 207 TABLE A.3 Frequency Distribution of Offenders by Predicted ant] Self-Reported Offense Rates for Specific Subgroups Self-Reported Offense Rate Predicted Offense Rate Low Medium High Reanalysis with All Cases (N = 886) Low193 4122 Medium193 10988 High65 67108 California Robbers (N = 166) Low26 20 Medium35 139 High23 2632 Michigan Robbers (N = 142) Low38 6 Medium25 2422 High8 5 Texas Robbers (N = 114,, Low32 113 Medium22 1114 High3 711 California Burglars (N = 151) Low31 42 Medium35 1611 High10 1725 Michigan Burglars (N = 113) Low20 3~ Medium29 1711 High12 71C Texas Burglars (N = 200) Low46 15. Medium47 2821 High9 522 Only Cases with Unambiguous Responses (N = 568) Low149 3414 Medium126 7259 High29 3649 Six-Variable Scale (N = 886, Low231 6030 Medium178 103102 High43 5487

208 CRIMINAL CAREERS AND CAREER CRIMINALS TABLE A.4 Parameters for the Incapacitation Model According to Estimates from the Reanalysis Predicted Offense Rate Parameter SymbolLowMediumHigh Total Number of offenders N95,50016,47310,611 Average annual offense rate A.98.120.8 Probability of arrest and convic tion q.030303 Probability of incarceration given conviction J.86.86.86 Probability of prison given incar ceration p.12.27.46 Average jail term in years s1.01.01.0 Average prison term in years S3.8924.6924.075 Average time served in years S1.3471.9972.415 Incarcerated population R2,8654,9425,942 13,749 Fraction of time free r197.70.44 Total crime C83,37293,40297,112 273,886 APPENDIX B.: DESCRIPTION OF QUESTIONS USED TO CONSTRUCT THE SEVEN VARIABLES The following information provides an overview of how the seven variables se- lectec3 for the scale used in the Green- wood report were constructed. All vari- ables are coded either 1 (yes) or 0 (no or missing). Short variable labels are user] in the following descriptions. 1. Prior Conviction Official records for most prison inmates contained information on the number of past convictions for several crime types. This variable was coclecl 0 if a convicted robber (or burglar) had no prior convic- tions in his record for robbery (or bur- glary) and 1 if one or more of the defining convictions were in his record. 2. Incarcerated Before Arrest A question on the survey asked inmates to indicate the months that they were in jail or in prison on their calendars. The percentage of possible "street time spent imprisoned was calculated, and in- mates with more than 50 percent were coded 1. (Greenwood user] Rand's mini mum estimate of street months for his calculations; thus, more inmates were coded 1 for this variable in his analysis than in the reanalysis.) 3. Convicted Before Age 16 The survey asked, "How old were you when you were first convicted of a crim- inal offense (an adult or juvenile convic- tion, other than a traffic violation)?" In- mates who reported a first conviction at age 15 or younger were coiled 1. 4. Juvenile Incarceration The survey asked, "Were you ever sent to a statewide orfe~leral juvenile institu- tion?" Inmates who respondecl "yes" were codecl 1. 5. Recent Drug Use The survey askocl, "During the months when you were using drugs, how often would you say you usually used each of the drugs listed below?" The drug types were: heroin/methaclone, barbiturates/ clowners/"reds," and amphetamines/up- pers/"whites"; the response categories were: did not use at all, few times a month, few times a week, every clay, or more than once a clay. This variable was

THE RAND INMATE SURVEY A REANALYSIS coded 1 if inmates responded that they used heroin or barbiturates at all. 209 offenders and 8.43 years in prison for high-rate robbers can be calculated as follows: 6. Juvenile Drug Use The survey asked about drug use of the Low Medium High Total following types before age 18: marijua na, LSD/psyche(lelics/cocaine, uppers/ downers, heroin; frequency levels were: often, sometimes, just once or twice, never. According to Greenwood's defini tion only the uppers/clowners and heroin responses are relevant, but Rand's com puter code indicates that the LSD/ psychedelics/cocaine category was also included. In the reanalysis this variable was coded as "yes" if inmates used her oin or uppers/clowners either sometimes or often as juveniles, and "no" otherwise. 7. Unemployed Before Arrest The survey askect, "During how many of the street months on the caTenclar cticI you work?" The percentage of street months spent working was calculated, and inmates who worked less than 50 percent of the time were cocled 1. APPENDIX C: CALCULATION OF POTENTIAL INCAPACITATIVE EFFECTS USING DATA REPORTED BY GREENWOOD Greenwood (1982:74) found the maxi mum incapacitative effects using the fol Towing hypothetical policy: "low- and mectium-rate offenders are sentenced to jai! and high-rate offenders are sentenced to prison for terms of increasing length." He also states that"none of . . . the sen tence lengths for high-rate offenders Lis] increased by more than a factor of 2" (p. 791. This means that the end point in his graph (Figure 5.1) represents an expected sentence length of 8.43 years (twice the 50.6 months reported in Table 5.1, p. 77~. Using this information and the data in Table 5.1, the maximum incapacitative effect for a sentencing policy that assumes l-year jail terms for low- and medium-rate Number of Offendersa 49,714 11,895 9,028 c . ''b Time Free (71) Incarcerated PopulationC (R) Total Crimed (C) .95 .79.13 2,486 2,4987,854 12,838 94,457 94,91036,148 225,515 aThese estimates are reported in Greenwood (p. 77) with the exception of the low-rate offender estimate, which was corrected by Cohen and con- firmed by Abrahamse (1984, personal communica- tion). bUsing equation for ~ reported in Greenwood (p. 7s) ~ = 1/[1 + (2.0)(0.03)(0.86)(1)] = 0.95 EM = 1/[ 1 + ( 10.1)(0.03)(0.86)(1)] = 0.79 AH = 1/[1 + (30.8)(0.03)(0.861(8.43)] = 0.13 CUsing equation for Ri reported in Greenwood (p. 7s) RL = 49,714(1 - 0.95) RM = 11,895(1 - 0.79) RH = 9,028(1 - 0.13) dUsing equation for Ci reported in Greenwood (P- 75): CL = 49,714(0.95)(2.0) CM = 11,895(0.79)(10.1) CH = 9,028(0.13)(30.8) Under the current sentencing policy, the estimated incarcerated population is 13,930 (Table 5.1, p. 77) and the estimated number of robberies is 259,917 (corrected figure; Abrahamse, 1984, personal communica- tion). Percent decrease in incarceration: 1 - (12,838/13,930) = 8 percent Percent decrease in robbery: 1-(225,515/259,917) = 13 percent REFERENCES Avi-Itzhak, B., and Shinnar, J. 1973 Quantitative models in crime control. Jour- nal of Criminal Justice 1 :18~217. Blackmore, J., and Welsh, J. 1983 Selective incapacitation: sentencing accord

210 ing to risk. Crime and Delinquency 29(0cto- ber).504-528. Blumstein, A. 1983 Selective incapacitation as a means of crime control. American Behavioral Scientist 27(1):87-108. Blumstein, A., and Cohen, J. 1979 Estimation of individual crime rates from arrest records. Journal of Criminal Law and Criminology 70(4):561~85. Blumstein, A., Cohen, J., and Hsieh, P. 1982 The Duration of Adult Criminal Careers. Unpublished final report for the National Institute of Justice. Grant No. 79 NI-AX- 0099. Urban Systems Institute, Carnegie- Mellon University. Blumstein, A., Farrington, D., and Moitra, S. 1985 Delinquency careers: innocents, Resisters, and persisters. Pp. 187-200 in M. Tonry and N. Morris, eds., Crime and Justice: An An- nual Review of Research, Volume 6. Chi- cago, Ill.: University of Chicago Press. Chaiken, J., and Chaiken, M. 1982a Varieties of Criminal Behavior. Prepared for the National Institute of Justice, U.S. Department of Justice. Report R-2814-NIJ. Santa Monica, Calif.: Rand Corporation. Chaiken, J., and Chaiken, M., with Peterson, J. 1982b Varieties of Criminal Behavior: Summary and Policy Implications. Prepared for the National Institute of Justice, U.S. Depart- ment of Justice. Report R-2814/1-1015. Santa Monica, Calif.: Rand Corporation. Chaiken, M., and Chaiken, J. 1984 Offender types and public policy. Crime and Delinquency 30~2~:19~226. Cohen, J. 1978 The incapacitative effect of imprisonment: a critical review of the literature. Pp. 187-243 in A. Blumstein, J. Cohen, and D. Nagin, eds., Deterrence and Incapacitation: Esti- mating the Effects of Criminal Sanctions on Crime Rates. National Research Council. Washington, D.C.: National Academy Press. 1983 Incapacitation as a strategy for crime control: possibilities and pitfalls. Pp. 1~4 in M. Tonry and N. Morris, eds., Crime andJus- tice: An Annual Review of Research' Volume 5. Chicago, Ill.: University of Chicago Press. 1984a Career Length and the Selective Incapacita- tion Model. Paper presented at the 1984 annual meeting of the American Society of Criminology, November, Cincinnati, Ohio. 1984b Selective incapacitation: an assessment. Il- linois Law Review 1984~2):25~290. Dershowitz, A. 1973 Preventive confinement: a suggested frame CRIMINAL CAREERS AND CAREER CRIMINALS work for constitutional analysis. Texas Law Review 51:1277-1324. 1974 The origins of preventive confinement in Anglo-American law. Part I: the English ex- perience. University of Cincinnati Law Re- view 43:1~0. Ebener, P. 1983 Codebook for Self-Report Data from the 1978 Rand Survey of Prison and Jail In- mates. Prepared for the National Institute of Justice, U.S. Department of Justice. Report N-2016-NIJ. Santa Monica, Calif.: Rand Cor- poration. Farrington, D. 1973 Self-reports of deviant behavior: predictive and stable? Journal of Criminal Law and Criminology 64:9~110. Farrington, D., and Tarling, R., eds. 1985 Prediction in Criminology. Albany, N.Y.: SUNY Press. Floud, I., and Young, W. 1981 Dangerousness and Criminal Justice. Lon- don: Heinemann Educational Books. Gold, M. 1966 Undetected delinquent behavior. Journal of Research in Crime and Delinquency 13:27~6. Gottfredson, M., and Gotttredson, D. 1980 Decisionmaking in Criminal Justice: Toward the Rational Exercise of Discretion. Cambridge, Mass.: Ballinger. Greenberg, D. 1975 The incapacitative effect of imprisonment: some estimates. Law and Society Review 9:541-580. Greenwood, P., win Abrahamse, A. 1982 Selective Incapacitation. Report prepared for the National Institute of Justice, U.S. Department of Justice. Report R-2815-NIJ. Santa Monica, Calif.: Rand Corporation. Hinton, J., ed. 1982 Dangerousness: Problems of Assessment and Prediction. London: Allen and Unwin. Loeber, R., and Dishion, T. 1983 Early predictors of male delinquency: a re- view. Psychological Bulletin 94:6~99. Marquis, K., win Ebener, P. 1981 QualitgofPrisoner Self-reports: Arrest and Conviction Response Errors. Prepared for the National Institute of Justice, U.S. Depart- ment of Justice. Report R-2637-DOJ. Santa Monica, Calif.: Rand Corporation. Monahan, I. 1981 Predicting Violent Behavior: An Assessment of Clinical Techniques. Beverly Hills, Calif.: Sage Publications. Moore, M., Estrich, S., McGillis, D., and Spelman, W. 1984 Dangerous Offenders: The Elusive Target of -

THE FlAND INMATE SURVEY: A REANALYSIS Justice. Cambridge, Mass.: Harvard Univer- sity Press. Morris, N., and Miller, M. 1985 Predictions of Dangerousness. Pp. 1~0 in M. Tonry and N. Morris, eds., Crime andJustice: An Annual Review of Research, Volume 6. Chicago, Ill.: University of Chicago Press. Petersilia, J., and lIonig, P., with Hubay, C. 1980 The Prison Experience of Career Criminals. Prepared for the National Institute of Justice, U.S. Department of Justice. Report R2511- DOJ. Santa Monica, Calif.: Rand Corporation. Petersilia, J., Greenwood, P., and Lavin, M. 1977 Criminal Careers of Habitual Felons. Report R-2144-DOJ. Santa Monica, Calif.: Rand Corporation. Peterson, M., and Braiker, H., with Polich, S. 1981 Who Commits Crime? Cambridge, Mass.: Oelgeschlager, Gunn and Hain Publishers. Peterson, M., Chaiken, J., Ebener, P., and Honig, P. 1982 Survey of Prison and Jail Inmates: Back- ground and Method. Prepared for the Na- tional Institute of Justice, U.S. Department of Justice. Report N-1635-NIJ. Santa Monica, Calif.: Rand Corporation. Reiss, A. 1973 Surveys of Self-Reported Delicts. Unpub- lished paper. Yale University, New Haven, Conn. Rolph, J., Chaiken, J., and Houchens, R. 1981 Methods for Estimating Crime Rates of In 2IZ dividuals. Prepared for the National Institute of Justice, U.S. Department of Justice. Re- port R-2730-NIJ. Santa Monica, Calif.: Rand Corporation. Shinnar, R., and Shinnar, S. 1975 The effect of the criminal justice system on the control of crime: a quantitative approach. Law and Society Review 9:581~11. Spelman, B. 1984 A Sensitivity Analysis of the Rand Inmate Surveys. Paper presented at the 1984 annual meeting of the American Society of Crimi- nology, November, Cincinnati, Ohio. von Hirsch, A. 1976 Doing Justice: The Choice of Punishments. New York: Hill and Wang. 1981 Desert and previous convictions in sen- tencing. Minnesota Law Review 65(4):591- 634. 1984 The ethics of selective incapacitation: obser- vations on the contemporary debate. Crime and Delinquency 30(2~:17~194. von Hirsch, A., and Gottfredson, D. 1984 Selective incapacitation: some queries on research design and equity. New York Uni- versity Review of Law and Social Change 1241~:11~1. Wyner, G. 1980 Response errors in self-reported number of arrests. Sociological Methods and Research 9(3~: 161-177.

Next: 6. Accuracy of Prediction Models »
Criminal Careers and "Career Criminals,": Volume II Get This Book
×
Buy Paperback | $110.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

Volume II takes an in-depth look at the various aspects of criminal careers, including the relationship of alcohol and drug abuse to criminal careers, co-offending influences on criminal careers, issues in the measurement of criminal careers, accuracy of prediction models, and ethical issues in the use of criminal career information in making decisions about offenders.

  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!