As research on the question of specialization and diversity in offending has grown, so too have discussions about how it should best be studied and measured from a methodological perspective. The question of specialization is one of a few areas related to criminal career research that has generated a fair deal of disagreement among criminologists. This debate has often played out in arguments about how best to measure and analyze specialized or diverse patterns of offending. Consequently, several different measurement and analytic approaches have been proposed and utilized in the study of this property of the criminal career.
Some early specialization research relied on the use of transition matrices, which are aggregate measures indicating how often an offender “switches” from crime to crime over the course of a criminal career (Blumstein, Cohen, Das, & Moitra, 1988). Such studies provide a sense of the average sample probability that individuals would switch offenses or maintain the same type across a series of arrests. Although one can examine a number of transitions, this method has been criticized as it may give too much weight to consecutive offenses at the expense of a broader sense of offending patterns. For example, if a substantial proportion of an offender’s arrests are for the same crime type but they are not ordered, this approach might still suggest a generalized offending pattern. Alternatively, an offender with a smaller proportion of arrests for the same crime, occurring consecutively, would demonstrate a greater degree of specialization on that measure.
Farrington (1986; Farrington et al., 1988) introduced a single summary measure drawn from transition matrices called the forward specialization coefficient (FSC) (see also Paternoster et al., 1998). This measure extends the traditional transition matrix approach to include observed and expected frequencies in an offense array so that the pattern identified in the data can be benchmarked against what one would expect if there were no relationship between consecutive offenses. Using these basic elements, one can then calculate an estimate that reflects cases where there is no specialization in offending (i.e., the observed values perfectly match those expected when there is no relationship between consecutive offenses) ranging to those with complete specialization (i.e., offense two is the same as offense one in all cases). While this approach provides a concise summary, it has also been criticized because its interpretation is somewhat murky and it lacks statistical properties that would open it up to further analysis (Britt, 1996).
Another criticism of both the transition matrix and specialization coefficient approach is that they are aggregate measures of offending patterns. As such, their interpretations are based on the offense patterns at the group level, not for individual offenders. As a result, attempts to make statements about individual patterns of specialization or versatility may lead to mistakes in the attribution of results to individuals versus groups (Piquero et al., 1999). This may be problematic since the criminal career framework (Blumstein et al., 1986), which in part prompted renewed interest in these questions, generally focuses on individual patterns of offending, not on aggregate crime rates. So, the transition matrix, and those measures derived from an aggregate array of offenses, may get away from the foundation of criminal careers research, which tends to seek out more individual-centered explanations for the nature and etiology of offending.
Partly as a result of this limitation, some have used the “diversity index,” which provides an estimate of the likelihood that any two offenses selected from an individual’s offending profile will differ (Mazerolle et al., 2000). The total number of offense types included in the measure creates an upper boundary on its value, and the lower value lies at zero. A diversity index value of zero indicates that the offender engages in complete specialization (i.e., all offenses come from the same category). As the value of the index increases toward its upper limit, the offender is identified as having committed a greater variety of offenses (i.e., covers more categories). One of the key features of the diversity index is that it measures offense specialization at the individual level (Mazerolle et al., 2000). So, relative to other measures, the diversity index approach is a closer fit to criminal career research and life course criminology. Also, unlike transition matrices, the order of offenses does not figure into its calculation, so the same offense type need not occur successively to denote specialization. Rather, it considers the pattern of offenses in the particular time period as a whole and then makes an assessment regarding their mix (or lack thereof). It can also be incorporated in further statistical analysis of individuals. Still, this measure may be influenced by the number and type of offenses included in its calculation and also the observed time window (Sullivan et al., 2006).
B. Analytic Approaches
Osgood and Schreck (2007) outlined a number of characteristics necessary for the appropriate measurement and analysis of specialization in offending. According to these authors, key facets in the measurement of specialization include a focus on the type of crimes committed rather than their ordering, definitions of specialization at the individual level, separation of specialization from the more general frequency of offending, and consideration of an individual’s pattern of offending in relation to the prevalence with which particular offenses are committed in the sample as a whole. These suggestions were drawn from a review of the methodological difficulties apparent in previous studies of specialization.
Working from this foundation, Osgood and Schreck (2007) used a statistical modeling approach that nests individual offense patterns within individuals to determine whether or not they engage in more offenses of a particular type (violent vs. nonviolent) than would be expected by chance alone. In this approach, specialization is captured by an individual’s relative balance of violent and nonviolent offenses. This method also allows researchers to account for other factors, such as the offender’s overall frequency of criminal activity, which may be important in the study of specialization.
Another method of assessing specialization, adopted in some inquiries, involves statistically identifying individuals who fall into relatively distinct categories of offending. One statistical-modeling approach that can provide some evidence of both prevalence and type of offending patterns is latent class analysis. This method draws on observed offender response patterns to conditionally place individuals into classes based on specific offending types. In this procedure, an offender is asked (or his or her record might be checked) whether he or she has committed any of an array of offenses, and a statistical model is then used to place the offender in a class based on those responses. For example, Britt (1994) found that individuals in the Seattle Youth Study were best represented by two groups, which largely reflected those who committed delinquent acts and those who did not. This suggests a lack of specialization, providing support for general theories of offending. More recently, Francis et al. (2004) undertook a similar type of analysis with a wider array of offense categories and found distinct classes such as shoplifting offenders and fraud/general theft offenders. Where measures like the diversity index and forward specialization coefficient demonstrate the level of observed specialization, these latent class models can help researchers to determine the nature of the offense type clusters that may be present in the data. These models also offer extensions that allow researchers to discern whether and how offenders may move or transition across offending types over time. Knowledge from such an approach can help in understanding findings of specialization at fixed points in time in the context of versatility over the longer criminal career.
C. Time Window
While several studies have shifted away from aggregate, population-level measures of offending specialization, they still tend to aggregate offending careers across a number of years. Sometimes studies pool offenses across more than a decade’s worth of time. Therefore, despite the use of measures that tap into individualized offense patterns, the findings that emerge from studies of specialization must be considered in relation to the time frame in which the data are viewed. One recent study, for example, found that the observed level of specialization shifted gradually as the time window was lengthened (Sullivan et al., 2006). This point is also partly supported by Osgood and Schreck’s (2007) finding of lower stability in individuals’ likelihood of violent specialization as the range of time in focus grew longer.
As mentioned above, operational definitions concerning specialization vary from study to study in terms of the use of time frames and also the order of offenses. Clearly, the degree to which the length of offending career is aggregated or disaggregated in a particular study may impact the degree of specialization or diversity observed. It might also be necessary to use longer time intervals in order to obtain enough volume in individual offending to make an assessment about the presence or level of specialization. As a result, it is important to contextualize findings of specialization and diversity in their observed time window and consider whether a full portrait of the criminal career is desired or a segmented view of some portion of that career is more relevant. Differences in the time window studied will inherently impact findings, so it is important to take note of how specialization is being assessed in a particular study so that it can be properly considered in relation to theory and practice.
D. Data Source
Offending has traditionally been measured in two main ways: self-reports and official records. Both have identified strengths and weaknesses for understanding the general prevalence of offending, which also extend to the study of criminal specialization. While official records may allow for the ordering of particular offenses and facilitate transition matrix-based analysis of offense patterns (Farrington et al., 1988), they also have shortcomings in terms of understanding the full scope of offending. It is well understood that official records capture only a fraction of the overall offenses committed by an individual. In the case of the study of specialization, this has important implications as it may distort the observed pattern of criminal activity toward more serious offense types and not fully capture the array of crimes engaged in by individuals. For example, Lynam et al. (2004) compared levels of specialization in violent offenses across official records and self-reports. Specifically, they examined the observed distribution of violent acts relative to what would be expected based on the overall distribution of offenses to determine whether their values differed, thus indicating specialization. They found evidence of violent specialization in self-reports, but not in official records.
Osgood and Schreck (2007) suggest that self-report information is preferred to the use of official records in studying specialization because it provides more depth of coverage of individual offense patterns and is less susceptible to biases of the type mentioned above. Still, there may be some problems with the use of self-reported offending as a foundation for understanding specialization. Bursik (1980) notes, for example, that self-report data may carry a specific form of risk in the context of such research. He argues that an offender who commits fewer offenses is likely to have better recall regarding their precise nature than a high-rate offender. This results in estimates of offending patterns that may be systematically biased. Clearly, for both official records and self-report methods, any potential systematic inclusion/exclusion of particular offense types might distort findings related to the question of specialization. So, as with offending prevalence in general, it is important to consider the source of the data in making sense of estimates of offending diversity and specialization.
E. Underlying Classification of Offenses
In addition to the methodological considerations noted above, the categorization of offenses that underlies a particular measure of specialization has important implications for research findings. Miethe et al. (2007) point out that the potential for finding specialization will decrease as the array of offenses included in its calculation increases. This owes to the fact that finer classification of offense types inherently reduces the likelihood that the same offense(s) would be observed on multiple occasions. Often, researchers investigate specialization based on some configuration of the three overarching crime types— violence, property, and drugs (see Mazerolle et al., 2000). Studies relying on broad “violence,” “property,” and “other” categorizations of crime might generate less observed diversity than those that fully distinguish among offenses that fall within these wider umbrellas (e.g., auto theft, larceny, and fraud for “property”). Also, the manner in which offenses might fit together empirically may be different from initial researcher and practitioner perceptions about which types are similar.