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.
4 Methodological Issues in Criminal Career Research This chapter examines some method- ological issues that affect research on criminal careers: of particular concern are various aspects of the observational ap- proaches user! to obtain data and the es- timation techniques applier] to those data. This chapter draws heavily on Cohen (Appendix B), who provides a more cle- tailed review of these issues. On the basis of our examination, we propose various suggestions for improver] research strate- gies to reduce potential biases arising from sampling anct measurement prob- lems. OBSERVATIONAL APPROACHES: SELF-REPORTS AND OFFICL\L RECORDS The two main observational ap- proaches for obtaining data on individual criminal careers-self-reports and official records of contacts with the criminal jus- tice system-invoke longitudinal data for individuals. Each approach is vuInera iData on crimes committed could also be ob- tained from reports by victims, from direct observa 96 ble to various sources of error that may limit the accuracy of the derived esti- mates of criminal career dimensions. Self-Reports The sources of distortion in self-report surveys inclucle problems in design of survey instruments, response errors, and analytic problems in inferring career di- ~mensions from questionnaire responses. The role of analytic problems was iTIus- tratec! in the Chapter 3 discussion of al- ternating spurts and quiescent periods in offending: recognition that offending fre- quencies during spurts are unlikely to persist over periods as Tong as a year led to downward revisions of as much as 25 percent in estimates of A. Response errors may arise from prob tion by researchers, and from information provided by informants. Although such data might provide rich information on the nature of the offense and sometimes on the attributes of an offender, they are usually linked to particular crime events and not to individual offenders. Thus, it is often difficult if not impossible to use such data to trace the develop- ment of criminal careers for individual offenders.
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH lems in questionnaire design or aciminis- tration procedures, from intentional mis- representation by respondents, or from respondents' errors in the classification or recall of events.2 An example of question- naire design effects comes from the Rand inmate surveys: the open-ended queries about A in the first survey may have yieldecI imprecise results for hi~h-fre- quency offenclers, but the alternative ap- proach used in the second survey a complex series of closecI-ended ques- tions apparently increased the rate of ambiguous, unusable responses (see Chapter 3 for cletaiTs). One source of misclassification is re- sponclent uncertainty about which events are to be counted as arrests or police con- tacts. The ambiguity for respondents may increase as the number of interactions with police increases and respondents are less able to distinguish which of the interac- tions involve an official charge or notation in police records. There is also ambiguity in classifying self-report items into official- record crime categories: discrepancies may arise because distinctions in the official categories reflect considerations of strength of evidence, criminal intent, ant! serious- ness of the outcome, which are probably not considered] in the sel£report categories. For example, a self-reported burglary may be noted in the offender's official record as a burglary, a larceny, or possession of stolen property, depencling on the circumstances of the arrest. Memory recall is likely to be affected by the saliency of the events (with more salient events more likely to be remem- bered) and the recency of the events. Two factors potentially affecting the sa- liency of crimes are the seriousness and the frequency with which they are com- mitted. In general, more serious crimes- with the greater risks they pose both for 2This discussion of errors in self-reports draws heavily on material presented in Weis (Volume II). 97 victims ant! for offenders are expecter! to be more salient. The more frequently that crimes are committed, however, the less salient any one ofthem is likely to be, ant] so there will probably be errors in counting the total number of crimes com- mitted. These sources of memory errors are likely to be of greatest concern for high-rate, serious offenders. Since recall is usually best for events that occurred most recently, memory problems are likely to increase with longer recall periods and with greater intervals between the recall period ant] the survey ciate. These problems are most typical of surveys that request reports of lifetime frequencies. To reduce those problems, most self-report surveys now limit the recall period, for example, to the year preceding the interview. But even this period is subject to potentially seri- ous recall errors for frequent ant! Tow- saliency events. Further reducing the re- call period to less than a year, however, could jeopardize precision in estimates of the number of reported criminal events, especially for more serious offense types that occur infrequently. Defining the recall period in terms of the interview ciate (e.g., during the year preceding the interview)-although it will enhance the recency of recalled events may make the clata especially vulnerable to bouncing errors: events that occurred before the designated recall period may be mistakenly attributed (or "telescoped") to the recall period. Bound- ing errors may be reduced by specifying a recall period with more salient bound- aries for respondents, such as the calen- dar year or "age-year" (i.e., time between birthdays). A more effective but far more costly solution to the problem oftelescop- ing is to administer two surveys, one at the start of a recall period and the other at the end of that period; events that are reported in both surveys are then re- moved from responses for the bounded
98 period. Because of the extra cost ant] time required for this approach, however, bounding ot thiS sort is not usually done for self-report surveys. Even if a researcher exercises great care, there will be ambiguous responses that present cli£iculty in analyzing the data. For example, on the basis of a reanalysis of the data from the seconc! Ranc3 survey, Visher (Volume II:Table 11) indicates that 35 to 40 percent of the responses by inmates active in robbery or burglary were ambiguous. The Rand re- Official Records searchers tried to deal with that amb~gu ity by computing minimum and maxi mum estimates for each respondent (Chaiken and Chaiken, 1982a). Visher adopted an alternative strategy for deal ing with ambiguous responses: rather than cleveloping extreme estimates for all respondents, she formulated rules for cle riving a single reasonable estimate for each indiviclual and excluclecl from the analysis those few for whom no reason able estimate couIc3 be computed. For example, individuals who indicated that they committed 1 to 10 crimes but did not report the exact number were assigned a single value in the range from 1 to 10 to match the distribution of the responses by the unambiguous respondents. She used similar estimating strategies for other am biguous responses. As indicated in Ta bles 3-3 and 34 (in Chapter 3), Visher's estimates are much closer to Rand's orig inal minimum estimates than to the max 1 1e r.1 . . ~. CRIMINAL CAREERS AND CAREER CRIMINALS ally finds the reliability ant! validity of responses reasonably insensitive to these various administration conditions (Weis, Volume II). Rather, the response errors that are found in self-reports are due pri- marily to the saliency, frequency, and timing of criminal activities ant! to the structure of the survey items; further cle- velopment of survey instruments to bet- ter acIdress these aspects might recluce their effects significantly. mums. Various structural features in the acI- ministration of self-report surveys may potentially bias or limit the validity of responses. Such features include cliffer- ences between responses to self-admin- isterec! questionnaires and interviews, between anonymous and nonanonymous surveys, and the effects of differences in interviewer attributes. It is reassuring that the research that has examined these effects in self-reports of offending gener Official records are also vulnerable to important, but very different, errors that affect the accuracy of estimates of individ- ual offencling. The main sources of error are the extent to which officially recorded criminal events are limited only to crimes that result in arrests or convictions, are unreliably recorclecI, and are selective in being more likely to recorc! arrests or convictions for one population subgroup or crime type than for another. There are two main structural sources of recording errors: misclassification of events and nonrecorcling. Classification errors can result from clif- ferences among local agencies in their classification of offense types, as in the ambiguity over whether a purse snatch is a larceny or a robbery. These cIassifica- tion differences can also occur when To- cally recorclecl criminal events, based on the crime categories found in local stat- utes, are transformed to some other crime classification scheme in centrally main- tained official records. Within a single jurisdiction, classification is likely to be fairly consistent, but inconsistencies may be introduced in records that reflect cIas- sifications from multiple jurisdictions or . . . . ~ · · ~1 In comparisons across Jur~scl~ct~ons. ~ nus, a high estimate of A for some crime type in one jurisdiction might reflect a differ- ence in classification rather than a (liffer- ence in actual offending.
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH Nonrecording may occur because the event does not meet reporting standards, such as the requirement for a fingerprint or disposition data, which may not be available (Michigan State Police, 1983a, b). Nonrecorcling is also likely to vary across jurisdictions: jurisdictions that rely especially heavily on state criminal his- tory files are likely to be more thorough in their reporting. Differences in the strength of administrative ties between local jurisdictions and the agencies main- taining central records may also affect recording. Nonrecording of some events obvi- ously understates the number of arrests by sample members and thus contributes to underestimates of arrest frequencies. However, nonrecording can also lead to overestimates when some arrestees are missing from the arrest history data. Since each subsequent arrest increases the chance that a record will be created, low- rate offenders with their smaller expected number of arrests are likely to be dispro- portionately missing from the arrest his- tory ciata, contributing to overestimates of arrest frequencies. But however complete the recording may be, officially recorded arrests still account tor only a small portion of all crimes committed. In addition to the crimes clirectly associated with an arrest (or a conviction), offenders usually com- mit other crimes that do not result in arrest. Arrest records can be used to infer the volume of unobserved crimes com- mitted. Such inferences require estimates of the probability of arrest for a crime q and assumptions about the nature of that process, along with direct estimates of individual arrest rates (,u) based on the number of recorded arrests and the length of time that an offender is free and so at risk of arrest. For individuals with arrest rate ,u and probability of arrest for a crime q, A is equal to ~q. One approach to estimating the proba 99 bility of arrest for a crime was proposed by Blumstein and Cohen (19791. The es- timate relies on readily available aggre- gate data and starts with the ratio of the reported per capita arrest rate, A, to the reported per capita crime rate, C, in a jurisdiction. The ratio A/C is then ad- justed by the rate at which victims report crimes to the police, r, to reflect total crimes, including those not reported to the police. Data on reporting rates are available from annual victimization sur- veys, and the adjustment should use re- porting rates for the jurisdictions being studied whenever possible. The estimate of q is further adjusted by the average number of offenders per crime incident, O. also available from victimization sur- veys. This correction adjusts for the fact that the estimated risk of arrest per crime for any inclividual offender is overstated when the arrest data include arrests of several different offenders who are in- volvec! in the same crime incident but the crime data do not. The resulting estimate of the probability of arrest for a crime, q, is (A/O)/(C/r). There are various potential sources of error in estimates of q and thus in the associated estimates of crimes actually committed. Reported numbers of arrests and crimes are subject to nonrecording, as discussed above. Both victims' reporting rates and the number of multiple offend- ers per crime are vulnerable to recall ant] other response errors by the respondents to victimization surveys. The multiple- offenders factor is also subject to biases arising from systematic differences be- tween crimes for which victims know the number of offenders (and report it in vic- timization surveys) and those for which the number of offenders is not known or reported. The multiple-offenclers factors estimated from the victimization surveys will overstate the number of offencler- crime incidents (C x O) and thus result in underestimates of q if the number of of
100 fenders is more likely to be known in crime incidents involving multiple of- fenders. Because of possible errors in compo- nents involved in the estimation of q, it is useful to perform sensitivity analyses to assess the impact of those errors on esti- mates of A. Based on available empirical estimates of the various component val- ues a range of A/C from .1 to .3 (Federal Bureau of Investigation Uniform Crime Reports, annual), values of O in the range of 1.2S to 2.50 (Reiss, 1980b), and values of r in the range of .25 to .75 (Bureau of Justice Statistics, 1982a, bathe values of q calculated from (A/O)/(Clr) range from .010 to .180. This broad range of estimates of q can be combined with a typical an- nual value of ,u of .2 (Table 3-1) to gener- ate estimates of A (A = u/q) ranging from 1 to 20 crimes committed annually. Be- cause of the sensitivity of these A esti- mates to variations in the factors compris- ing q, better estimates are needed of the range of error in the components of q. Even if the average value of q is esti- mated accurately, estimates of both indi- vidual offending frequency (A) and partic- ipation rates (~1) may be biased by interactions between A and q at the indi- vidual level. If individual crime rates and arrest risks are negatively related to one another, with high-rate offenders less leery to ne arrested tor eacn crime, appli- cation of a homogeneous q to all offenders results in an underestimate of A. Corre- spondingly, A wfll be overestimated if A and q are positively related. If A and q vary systematically with the attributes of offenders, failure to adequately represent that variation wfl] also distort the esti- mates of A associated with different of- fender subgroups. Without adequate con- trols for variations in q, any estimate of A derived from official records wfl! con- found individual differences in A with clifferential police practices reflected in q. The limited research avaflable to date ... ~. ~. ~ r ~ CRIMINAL CAREERS AND CAREER CRIMINALS generally tails to one systematic substan- tial variations in q with differences in A or in the demographic attributes of offend- ers (see Cohen, Appenclix B). Further research on this relationship should have high priority. However, while stfl! pre- liminary, the available results suggest that the errors may be small in estimates of average values of A that are derived from arrest records made on the assump- tion that A and q are independent. Even in the face of heterogeneity in q, estimates of individual crime rates would be unbiased if the arrest risk for a crime varied independently of individual crime rates. Nevertheless, the potential for bias in estimating dimensions of the crime process from official data of the observed arrest process highlights the importance of empirically investigating the nature of any variation in q, and especially in find- ing any systematic variation in q with changes in individual crime rates or with attributes of offenders. To date, homoge- neous qs have been used because official records do not provide the information necessary to indicate how arrest risks vary across different offenders. Such informa- tion, however, can be derived by combin- ing self-reports of individual crime and arrest experiences. .~ ~ .~ ~ ~ . . Potential for Synthesis of Observational Methods Self-reports and official records are cur- rently the best avaflable methods for ob- taining longitudinal data on individual criminal careers. Because of the funda- mental differences between them and their sources of error, the two approaches are often posed as competing alternatives. This conflict between the methods has been fueled by apparently substantial dif- ferences in conclusions based on the two data sources: most notably, early findings based on official-record data showed im- portant differences in criminal participa
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH tion by social class and race that were not supported by data from self-reports. For the most part, these early differences are resolved when appropriate controls are inclucled to ensure that the two data sources are comparable in sample compo- sition and in the seriousness threshold used for criminal activities. A more constructive approach is to view the altemative methods as comple- men~ and to search for mutually bene- ficial ways of using them. The discussion above highlighted some of the more im- portant sources of error in the two ap- proaches: because their sources of error are very different, however, it is possible to use the two approaches in concert to ameliorate some of the error problems in each. For example, one can compare es- timates of individual offending patterns that are derived independently from the two methods. If the estimates are similar, some confirmation of their accuracy is provided. Since it is unlikely that the two approaches would consistently result in the same wrong estimates, similar find- ings from independent analyses would suggest that the errors in each approach ~^ homogeneous q. Self-report data provide I01 Self-reports of crimes committed are sometimes based on samples of inmates in order to increase the number of respon- dents who have a sufficient number of criminal events. While providing a usable number of reported events for analysis, such samples are not representative ofthe general offending population (see next section on sampling). In these samples, official records can be used to estimate various statistics (e.g., the probabilities of arrest for a crime, of conviction following arrest, and of incarceration following con- viction, and the average time served once incarcerated) that describe the selection process that led to the respondents' incar- ceration. These statistics, which deter- mine the chance that an offender with particular characteristics will be found among inmates, can be used to weight observations in the inmate sample in or- der to provide estimates applicable to offenders who are not incarcerated. Data from self-reports can also be used to address one ofthe main sources of error in analyses of official records the over- simplified characterization of the arrest process implied by the assumption of are not grossly distorting. l he search tor such convergent validity between results is one way of using analyses of self- reports and official records in concert. The two approaches can also be used together at intermediate levels in re- search to try to deal explicitly with vari- ous sources of error. Official records of arrests or convictions are already used in combination with self-reports of these same officially recorded events as a means of validating the accuracy of self-reports. Official records might also be used to help reduce response errors by invoking events in the official record during self- report interviews, both as a means of triggering recall of events and time neri- ods for respondents and of reducing re- spondent inclinations to intentionally misrepresent their criminal activities. an opportunity to directly link self-re- ported crimes to both self-reported ant] official-record arrests for specific indivicl- uals with known attributes. With these data, it is possible to explicitly examine the variability in q both with indiviclual crime rates and with various attributes of offenders. The patterns of variation that are found would provide a basis for intro- ducing heterogeneous estimates of q into analyses of official-record clata. SAMPLING ISSUES Analyses of criminal careers involve some important choices on the appropri- ate sample design for generating offender data. Central to problems of sample cle- sign are tradeoffs between the represen
102 tativeness of offenders and obtaining a sufficient number of active offenders for analysis. The best sample choice in any study varies with the career dimension being measured. At one extreme, a random sample of people can be drawn from the general population and their criminal careers fol- lowed through self-reports or official- record data. But since arrests and crimes occur relatively infrequently in the gen- eral population, the number of crime and arrest events will be low, especially for the more serious offense types. There- fore, samples from the general population are of value primarily for studying partic- ipation in offending. Only for very minor crime types, like truancy and smoking by juveniles, which occur in large numbers, do such samples permit analysis of active offenders' careers. Stratified sample designs can increase CRIMINAL CAREERS AND CAREER CRIMINALS all community-based samples exclude of- fenders who are incarcerated at the time of sampling, who tend to be dispropor- tionately high-rate offenders. If high-rate offenders are more likely to drop out of school, they will be underrepresented in school-based samples. High-rate offend- ers who are free in the community are hard for survey researchers to locate be- cause they are likely to be "on the run" or otherwise trying to avoid detection; those with highly transient living arrangements tend to be missed in household-based samples. And high-rate offenders who are located may be more likely than others to be uncooperative, refusing to participate in research. The principal alternative to general population samples is samples drawn from populations of presumed offenders, such as arrestees, convictees, or incarcer - ~- - v ated people. The choice of a definition for the yield of active offenders by oversam- sampling offenders involves a tradeoffbe pling high-yield subpopulations, but at tween the degree of certainty about of the cost of a reduced number of offenders fender status and the degree to which the drawn from low-yield subpopulations. sample is representative of offenders Therefore, stratification increases the pre- more generally. Sampling from convicted cision of estimates of the career dimen- offenders greatly reduces any uncertainty signs of active offenders in the high-yield about the actual criminal involvement of subpopulation and usually in the aggre- sample members. However, convicted of gate population, but decreases substan- fenders in prison, for example. are not tially the precision of estimates for the low-yield subpopulation. For example, oversampling teenaged males from low income neighborhoods will improve the precision of A estimates for Hat subpop ulation, but with a substantial loss of pre cision in A estimates for young adults, females, and residents of high-income neighborhoods. Regardless of stratification, samples drawn from community-based sampling frames like schools and households are more likely to miss offenders than nonof- fenders, and this problem is likely to be most severe for high-rate offenders, lead- ing to their disproportionate underrepre- sentation in these samples. For example, likely to be representative of all offend- ers: they are presumably the most seri- ous, the oldest, and perhaps the most inept at avoiding detection. The broadest sampling base among of- ficially detected offenders is arrestees. But selecting a sample of arrestees in- volves potential errors of commission, since some falsely arrested persons are wrongly included among active offend- ers. Selecting convictees, by contrast, is more likely to involve errors of omission because active offenders in the arrestee population who are not convicted (and often not prosecuted; see below) are ex- cluded from the sample. In dealing with dispositions for specific individuals, of
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH course, the presumption of innocence makes errors of commission unaccept- able. In dealing with the empirical char- acterization of criminal careers, however, there must be a relative weighing of these two types of errors. Fundamental to this consideration is some assessment of the factors contributing to the lack of a con- viction after an arrest. Empirical examinations of the reasons for nonconviction suggest that nonconvic- tion is by no mean synonymous with innocence (Forst, Lucianovic, and Cox, 1977; Vera Institute of Justice, 1977; Brosi, 1979; Boland et al., 1983; Feeney, Dill, and Weir, 19831. The vast majority of nonconvictions are the result of diver- sions from adult criminal courts (to juve- nile court or to pretrial diversion pro- grams) and dismissals, rather than the result of acquittals. The reasons for dis- missal frequently have little to do with the guilt or innocence of the defendant. Instead, many cases are dismissed be- cause of noncooperation by witnesses (of- ten arising from the existence of a prior relationship between the offender and victim), inadmissibility of critical evi- dence, and the lesser importance of the case compared with other cases. Manv less serious cases are also diverted prior to trial. In view of the predominantly procedural reasons that many arrests do not result in a conviction, the errors of commission associated with truly inno- cent arrestees appear to be far less fre- quent than the errors of omission that wouIcI occur if the more stringent stan- dard of conviction were user] as a basis for sampling offenders. An important consideration in using samples of identified active offenders is the clegree to which those samples are biased toward high-rate offenders. High- rate offenders wouIc} be overrepresented among persons who have at least one detected event (a crime, arrest, convic- tion, or incarceration) during a sampling 103 period, since they are more likely to incur the sampling event. This bias is greatest when the sampling period is short com- pared with the mean time between events.3 These biases can be compen- satecl analytically by accounting for the differential sampling probabilities associ- ated with offenders having different event rates. Alternatively, one can in- clude only offenders whose first recorcled event occurs in a sampling period that is short comparer} with career length. This sample of starting offenders mirrors the distribution of A among offenders. There is also a greater concentration of more serious offense types in convictee and inmate samples than in other of- fender samples. Processing cases through the criminal justice system typically in- volves increased filtering and selectivity as many less serious cases are dropped. Thus, the further into the criminal justice system that samples are drawn, the less likely samples are to be representative of street offenders generally. From the per- spective of ensuring greater representa- tiveness, samples of self-reported offencl- ers and arrestees are better than samples of inmates or even of convictees. It is possible to correct for biases aris- ing from the sampling process. The cor- rection involves reweighting the sample to reflect the differences in the sampling probabilities of sample members. Such a weighting scheme wouIcl give greater weight to offender types who are uncler- represented in a sample and less weight to those who are overrepresented. Suc- cessful correction of sample biases using 3If events follow a Poisson process, an offender with individual event rate ,u has a probability of being sampled during a period of length ~ of Ps = ~ - e-~. If ~ is very short compared to Ilk, then Ps is approximately At, and so higher-rate offenders are seriously overrepresented. If ~ is very long com- pared with 1/,u, then ps is close to unity, and the sample is reasonably representative of the offender sample.
104 this approach requires an adequate char- acterization of the sampling process and reasonable estimates of the sampling probabilities. General population and offender-basec! sampling strategies, with their different flaws, are each suited for clifferent pur- poses. General population samples, which include offenders and nonoffend- ers, are more appropriate for estimating participation rates. The undercount of high-rate offenders in such samples will understate participation, but this bias is not likely to be substantial because of the small numbers of high-rate offenders among total offenders. However, general population samples are inefficient for es- timating frequency rates of active offend- ers because of the low yield of active offenders in such samples. The ineffi- ciency is aggravated by the underrepre- sentation of high-rate offenders in those samples. Samples of arrestees or inmates are better suites! for estimating A, but corrections are required to adjust for the overrepresentation of high-rate and more serious offenders in these samples. USE OF COHORT AND CROSS-SECTIONAL DATA Research on criminal careers involves a study of the variation in an individual's criminal activity during his life, inclucling the ages at initiation and termination and the pattern of offending between those two points. Thus, all such research is inherently longitudinal. There are many ways in which such longitudinal research can be pursued. The most obvious wouIc! be to identify a cohort at birth ant! to follow that cohort prospectively for a Tong enough period of time to inclucle the termination of most criminal careers of cohort members. The most important disadvantage of this ap- proach is the Tong time required to develop results. In addition to the consid CRIMINAL CAREERS AND CAREER CRIMINALS erable cost involved, the historical envi- ronment in which the cohort is observed may no longer be relevant at the time the results become available. Thus, for exam- ple, a cohort that reached maturity before the sharp rise in drug use ofthe late 1960s would yield no information on the influ- ence of drug use on involvement in other criminal activity. A different approach to longitudinal re- search is a retrospective Tongituclinal de- sign. This approach avoids the Tong delay associated with the prospective stu(ly by defining a cohort and reconstructing its prior criminal involvement. This is the approach pursued by Wolfgang, Figlio, and Sellin (1972), who first defined a cohort all boys born in 1945 and resict- ing in Philaclelphia from ages 10 to 10 and then retrospectively collected their records of police contacts. In the absence of longitudinal data, annual cross-sectional data can be used to synthesize a cohort by examining varia- tions across age within a year as a proxy for longitudinal age variations of a cohort. However, the two may not be equivalent. If there are important cohort effects, then those cohort effects will be confounded with age effects in a synthesized cohort. If there is a positive association between career length and A, for example, then a cross-section cohort will display a larger average A and more high-rate offenders than a natural cohort; a negative associa- tion will lead to the opposite effects. Fur- thermore, a cross-section design pre- cludes examining temporal sequences within individuals, which is a main fea- ture of longitudinal cohort designs. A major problem with single-cohort de- signs is that age effects are inextricably confouncled with historical effects. Thus, a cohort that happened to reach the high- crime ages of the mid-teens at a time of consi(lerable social turmoil wouIcl display an amplified age effect in involvement in crime compared with another cohort that
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH reacher] those ages at a time of social tranquility. Analysis of a single cohort -C:7- would not be able to isolate these effects. One way to overcome this problem is by drawing multiple cohorts and obtaining longitudinal data on them. A variant of this approach involves identifying a cross-section sample of the population (thus representing multiple cohorts) and collecting longitudinal data-either pro- spectively or retrospectively~n them. Because of resiclential migration, how- ever, the members of the cross-section sample will cliffer from the birth cohorts within the jurisdiction studied. Also, if the cross-section sample is drawn from an arrestee population (or some other of- fencler sample), then the older members of the cross-section sample wit! overrep- resent individuals who have longer crim- inal careers, and any estimates of relation- ships with career length will be biased; for example, factors positively associates! with career length will be overrepresent- ed. The relative strengths ant! weaknesses of the various design aspects consiclered here suggest that an appropriate compro- mise involves drawing samples from mul- tiple jurisdictions ant! cleveloping multi- ple overlapping cohorts, each observed for a limited time through certain key developmental age periods. For each of these periods, prospective longitudinal data on criminal activity and other related events should be collected on cohort members. Those data can and shouIc! be augmented by retrospective longituclinal data whenever available. PROBLEMS OF CONFOUNDED EFFECTS A number of factors may confound es- timates of criminal career dimensions. In some cases, there are interactions among various dimensions, which result in pos- sible distortions in the separate estimates 105 for each dimension. The possibility of career termination cLuring a follow-up pe- riocI, for example, distorts estimates of A for offenders who remain active, since some offenders will ens! their criminal activity during the follow-up period. If A is calculated by assuming that all offend- ers are active throughout the follow-up period, failure to account for this short- ened duration for some offenders will result in a downward! bias in the estimate of A. Many criminal career analyses focus on changes in offender behavior as their criminal careers progress. For example, such studies include analyses of trencis in offense seriousness or in A as offenders age or accumulate arrests. When such trends are established they are often at- tributec! to developmental changes as of- fenders mature, to growth in criminality as the career unfoIcts, or to consequences of offenders' interactions with the crimi- nal justice system. Another interpretation of observed trencis that is rarely invoked, however, involves none of these causal explanations but derives from offender heterogeneity. Uncler this interpretation, different offender groups wilt display clif- ferential persistence in their criminal ca- reers. In such a situation, the more per- sistent groups and their characteristics increasingly dominate samples of offend- ers who are observer! at later stages of criminal careers. To the extent that of- fencler heterogeneity is a factor in gener- ating the observed trends and is not ade- quately controllecl in the analysis ofthem, the changing composition of the offender population over the course of careers will be incorrectly interpreted as changes in the behavior of offenders. Measurements of career dimensions of- ten slider across studies and are some- times characterizes] as presenting con- flicting information. In many cases, however, the differences in measure- ments are attributable to (differences in
106 the scope of offenses consiclerecI or to the composition of the population studied. For example, participation rates wflT be higher for all offenses than for violent offenses, for cumulative lifetime partici- pation (BL) than for participation by age 18 (Bit), and for samples of mates alone than for samples of mates and females. Therefore, any reporting of criminal ca- reer measurements must indicate the ba- sis of the measurements. Variations in exposure time can also affect measurements of criminal career dimensions. For example, if an individual initiates or terminates a career midway through an observation period, then the estimate of his offending frequency, when distributed over the entire period, would be only halfhis true rate cluring his active periocl. A similar distortion could occur in analyzing the effects of covari- ates on criminal career dimensions. Con- sider, for example, the relationships be- tween precursor behaviors, such as alcohol or marijuana use, ant! criminal participation: alcohol use generally be- gins at an earlier age than marijuana use, and thus alcohol users have a longer ex- posure time within any independently established observation period than clo marijuana users. Thus, even if both sub- stances had the same influence in initiat- ing delinquent careers, more of the alco- hoT users would have had the opportunity to begin offending within the observation period than would marijuana users. Iso- lating the relative influence of these two covariates on participation requires ade- quate controls for the differences in times at risk (Robins and Wish, 19771. Identifying the covariates of criminal careers is especially important both for improving theory on the causes of incli- vidual criminality and for distinguishing among offenders for various policy pur- poses. The proportional hazards method! is a statistical technique that permits si- multaneous control for variations across CRIMINAL CAREERS AND CAREER CRIMINALS inclividuals in covariates of criminal ca- reers and for variations in exposure times. Its primary application in criminology has been to data on time to recidivism (Barton and Turnbull, 19811. By relying on time to a first recidivist event as the clependent variable, however, these models cannot distinguish the separate effects of covari- ate s on the career dimensions of fre- quency and termination. Maltz (1984) explores one approach to disentangling the relationship of inde- pendent covariates to separate career dimensions. He proposes a model that partitions recidivism between the proba- bility of ever recidivating (which is re- lated to career termination) and the fail- ure rate of recidivists (a direct measure of offending frequency for active offenders). To examine the role of covariates on those career dimensions, the data are parti- tioned into groups that are reasonably homogeneous with respect to the covari- ates of interest, end the two dimensions of recidivism are estimated separately for each group. In an illustrative analysis of the effects of one covariatc age at re- leasc the estimated probability of ever recidivating decreases as age at release increases in three of four jurisdictions examined, but there appears to be no effect of age at release on failure rates for those who do recidivate (Maltz, 1984: 131-1331. This approach, however, is still preliminary, and considerable develop- ment and testing are required to identify the statistical properties of the technique. The problem of identifying and control- ling for the effects of independent covariates on the various career dimen- sions remains an important area for fur- ther research development. EXPLICIT MODELS OF OFFENDING Virtually all estimates of criminal ca- reer dimensions invoke some kind of im- plicit model of individual offending. Be
METHODOLOGICAL ISSUES IN CRIMINAL CAREER RESEARCH cause of the inherent clifficulties in obtaining direct observations of crimes committed by individual offenders, esti- mates of career dimensions rest on other observable data, like arrests and self- reported crimes, as indirect indicators of the underlying crime process. The vari- ous estimation strategies that are applier! to these data rest funclamentally on moc3- els that characterize both inclividual of- fending and the processes that give rise to the observable data. The accuracy of the estimates of criminal careers that emerge clepencis on the adequacy of the assump- tions in the models, which are usually unstated. Because the available observable data are only indirect indicators of actual crimes committed, improving the preci- sion of measurement of those data is only one part of needed work; estimates for the underlying, but unobserved, crime proc- ess must also be improved. This second! step requires explicit models that link the unobserved crime process with the ob- servec] data. With explicit models, the adequacy of estimates can be assesses] in terms of the reasonableness of the as- sumptions ant] the sensitivity of results to those assumptions. Models of incliviclual offending have moved from treatments of offending based on traditional aggregate measures such as per capita crime rates and recidi- vism rates to more detailed characteriza- tions that partition offending levels among the various aspects of a criminal career. The initial models of criminal ca- reers have relied on a number of simpli- fying assumptions, principally that indi- vidual careers are stationary over time ant! homogeneous across offenders. This simplest characterization underlies most currently available estimates of career cli- mensions. More recent developments have begun to enrich the basic moclel to better accommodate the complexities of real careers. One issue of concern has ]07 been possible nonstationarities in offend- ing frequencies during individual ca- reers. Two forms of nonstationarity have been addressed in recent research: spurts in criminal activity as offenders move be- tween active and quiescent periods and changes in frequencies as offenders age. Two approaches to addressing spurts in criminal activity cluring a career are avail- able. In reanalyzing the ciata from the second Rand inmate survey, Chaiken and Rolph (1985) fount! evidence that periods of criminal activity cluring the observa- tion period tended to be clustered near the arrest that led to the current incarcer- ation. The observation periods for respon- clents with short street times are thus only slightly longer than the periods of spurts in activity. When offending rates during these periods are treated as if they ap- plied to both active and quiescent periods, they lead to overestimates of of- fenders' average annual frequency, A. Chaiken ant] Rolph propose a mode! to reflect this mixture of active and quies- cent periods during a career and acldust indiviclual frequencies downward to re- flect this mixture. Lehoczky (Volume II) proposes an alternative mocle! to accom- modate spurts in criminal activity: indi- viduals alternate between active ant! qui- escent periods, and they commit crimes en cl are arrested only during the active periods. After each arrest, an offender faces a possible transition to move into a quiescent period (with probability a) or to continue in the active state (with proba- bflity 1 - a). In the Lehoczky model, sanc- tions may also have an inhibiting effect on future crimes as each arrest triggers a pos- sible move to a quiescent periocl. Other mocle} refinements address pos- sible changes in A with age, like those observer! in aggregate per capita arrest rates, which increase rapidly into the late teens and then decrease steadily for older ages (see Figure 1-21. Similar clecTines with age for adults have been observed in
108 recidivism rates and in frequency rates for broac! aggregate offense categories, such as "all offenses," or "all inclex of- fenses" (see, for example, Peterson and Braiker, 1980~. Two different approaches to mocleling these age effects are presented in papers commissioner] by the panel. Flinn (Vol- ume II) models indiviclual allocations of time to criminal activity as a rational choice baser! on the net expected returns from legitimate ant! criminal activities. In this model, declines in criminal activity with age result when wage rates from legitimate activity increase as inclividuals accumulate more work experience and when the expecter! cost per crime, mea- surec3 by expected time spent incarcer- ated, increases with the number of prior incarcerations. The model by Lehoczky (Volume II) captures the distinction be- tween aging effects for inctiviclual crime types and for the aggregate of several crime types. An individual offender is mocleled as having multiple careers, one for each crime type. The careers in each crime type are separate and operate incle- pendently of one another. For any single crime type i, an individual's frequency rate (Ai) is fixed during his career in that crime type, ant! that crime-specific career terminates with some probability ,ll after each crime committed. In this formula- tion, the total value of A for an inclivid- ual reflecting the sum over all crime typesWeclines with age as active crime types are gradually eliminated. The Lehoczky mode] also incorporates other refinements to the model of crimi- nal careers, including variation across of- fen(lers in their career (limensions. It also CRIMINAL CAREERS AND CAREER CRIMINALS permits covariates of the dimensions re- flecting both fixed background character- istics (such as sex, race, juvenile record, and age at first encounter with the crimi- nal justice system) and dynamic attributes (such as drug use and employment sta- tus). While the moclel has not been ap- pliec3` to data, various techniques for esti- mating the model parameters from empirical clata have been proposed. The mode! developments by Flinn (Volume II), Lehoczky (Volume II), and Chaiken and Rolph (1985) represent con- ceptual advances over the simple model of criminal careers representec! in Figure 1-2 that underlies most available esti- mates of the various career dimensions. As mo(lels of criminal careers are ex- tenclec! to better reflect the underlying behavioral and observational processes, the estimates derived from those models should be more valid. Application of the enricher! models to improve estimates of the distributions of career dimensions is stfl] necessary. It wfl! also be useful to compare the resulting estimates with sim- flar estimates derived using the simple model. Such comparisons will permit an assessment of the error introduced by the assumptions in the simpler moclels (e.g., homogeneous frequency rates across of- fenders or the use of a single uniform arrest probability per crime across offend- ers). Such comparisons may indicate that the assumptions of the simple moclel are reasonable approximations that yield sat- isfactory estimates and that adequately account for important effects, or more de- taflec3 sensitivity analyses may inclicate which assumptions of the more elaborate models are most important.
