National Academies Press: OpenBook

Understanding and Preventing Violence, Volume 4: Consequences and Control (1994)

Chapter: Predicting Violent Behavior and Classifying Violent Offenders

« Previous: Violence and Intentional Injuries: Criminal Justice and Public Health Perspectives on an Urgent National Problem
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Predicting Violent Behavior and Classifying Violent Offenders

Jan Chaiken Marcia Chaiken and William Rhodes

INTRODUCTION

This paper discusses the classification of individuals as violent persons and the prediction of individual acts of violence. It is based on a review of research reports that implicitly or explicitly define violence as physically harmful behavior carried out by an individual and directed against others. Thus we exclude research on such topics as collective violence (e.g., riots and wars), self-injury (e.g., suicide), and psychological violence (e.g., verbal aggression).

We further focus on research whose explanatory factors were individual characteristics, thereby excluding studies of subcultural factors, ecological factors (such as density of the population), and situational factors (such as availability of firearms). Although such factors are pertinent for predicting the occurrence of violence, this paper focuses on predicting parameters of individual criminal careers, in particular:

  • the prevalence of violent persons in a study population (e.g., the percent of juveniles in the United States who have ever committed an act of violence) and the likelihood that any given person will be violent;

Jan Chaiken and William Rhodes were with Abt Associates in Cambridge, Massachusetts. Marcia Chaiken was at LINC, Lincoln, Massachusetts.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
  • the rate at which violent persons commit violent acts (the number of violent acts committed by a person each year);

  • the persistence (or duration) of violent persons' histories of committing violent acts (e.g., length in years from the commission of the first act of violence to the last act of violence); and

  • the seriousness (or harmfulness) of the violent acts committed by individuals (e.g., extent of physical harm inflicted).

Our review seeks to transcend the particular foci of the source studies (which typically concentrate on particular types of victims or forms of injury, such as spouse abuse or rape) and instead summarizes commonalities in the methods and findings of studies about violence. Considerable empirical evidence suggests that violent people frequently engage in a range of violent and other types of antisocial acts. Children who throw heavy or sharp objects at their parents are likely to hit their siblings or peers and to lie, set fires, and be truant from school (Lewis and Balla, 1976); prison inmates who are ''violent predators" (they committed robbery and assault and dealt drugs prior to incarceration) are just as likely to have an arrest history including rape as are fellow inmates who are actually serving time for rape (Chaiken and Chaiken, 1982, and unpublished analysis of the same data). Thus, a focus on a narrow range of violent acts would fail to identify what is commonly perceived as violence.

Furthermore, the characteristics of victims and the forms of injury are less important for classification and prediction than are characteristics and past behavior of individuals. Although any behavioral outcome is dependent on an individual's response to the environment, including his or her access to specific classes of victims, certain biological, psychological, and social characteristics of individuals dramatically increase or decrease the probability that they will engage in specific forms of behavior, independent of environmental factors. This concept has been supported by such studies as Glaser (1964), Hare (1979), Irwin (1970), Mann et al. (1976), McCord and McCord (1959), and Robins and Wish (1977), and has guided our thinking in this review.

To carry out the review, it was useful to distinguish between classification and prediction, even in situations where this distinction is unclear in the source document. The following section explains the basis for this distinction and summarizes those methodological issues that classification and prediction have in common; it also contains a brief summary of the correlates of violence. In the following section, we turn attention to classification. There we review the purposes of classification, types of

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

classifications, and congruent findings across studies. We then focus on prediction studies and summarize our findings about predicting violence. The last two sections present the implications for research and public policy.

TECHNICAL ASPECTS OF PREDICTION AND CLASSIFICATION

DISTINCTIONS BETWEEN PREDICTION AND CLASSIFICATION

No definitive boundaries separate classification from prediction, and they cannot be clearly distinguished by their methodology or purposes. In fact, some researchers who know they are involved in classification nonetheless state that their major purpose is prediction. For this review, however, we found that a useful and practical separation can be made by defining a prediction study as one whose underlying analysis fundamentally requires longitudinal data—information about the same subjects' behavior at two or more points in time.

For a work to be considered as prediction research here, the form or coefficients of the equations, models, or procedures for separating persons into subgroups must have been initially determined by a comparison of data at two different times. For this reason, an analysis that divides violent persons into groups by using cross-sectional data is considered to be classification, not prediction, even if the work was based on a theoretical expectation that the groups would behave differently in the future. Similarly, studies that examine subsequent data to determine outcomes of groups previously classified by use of cross-sectional data are considered classification studies (e.g., the work of Milner et al., 1984, discussed under "classification" below), as are studies in which the data collection instrument was tested for stability over time by using test-retest methods.

We also considered studies called "postdiction" by their authors (e.g., see Chaiken and Chaiken, 1990b) as classification research. In postdiction studies, data are collected describing various stages in a person's life or criminal career, but the dependent variables being estimated in the statistical analysis (e.g., individual crime commission rates during the previous year) describe behavior that is contemporaneous with some or all of the data items used as independent variables. In short, the behavior being estimated does not lie in the future with respect to the "predictors." A typical purpose of postdiction research is to devise ways

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

of using officially or routinely recorded data for estimating concurrent behavior that cannot be routinely measured.

On balance we felt it is helpful to limit the term "prediction" to future-oriented research. It is immaterial whether the data for the independent variables (predictors) are collected retrospectively at one time or longitudinally.

Classification research tends to focus on dividing individuals into distinctive subgroups. In traditional classification research (e.g., Gibbons, 1975), all persons being classified are to fall in a group defined by the research, and no one is to be classified as belonging to several groups. In others, a residual group of unclassified individuals is permitted,1 or persons are classified along several dimensions rather than in a single group (for example, as in the Diagnostic and Statistical Manual of Mental Disorders, Third Edition—Revised (DSM-III-R), American Psychiatric Association, 1987, discussed below).

By contrast, prediction research infrequently focuses on mutually exclusive subgroups, instead concentrating on estimating probabilities of future occurrences. In this framework, some or all persons could have probabilities greater than zero for several (or all) of the potential outcomes.

The purposes for undertaking classification are varied and may or may not be forward looking, whereas prediction is always forward looking by our definition. In the context of violence research, classification may be undertaken to estimate the prevalence of violent persons, or categories of violent persons, in specific populations; to construct typologies that assist in understanding personal and social characteristics of categories of violent and nonviolent persons; to learn more about causes, correlates, and stability associated with categories of violent persons; to diagnose individuals for purposes of planning treatment; and to assign individuals to groups for purposes of case management.2

The purposes of prediction may be similar, but prediction involves future behavior. Predictions of future violence may be made to determine whether subjects pose a risk to the community when released from criminal justice or mental health restraints; to investigate the causal relationships between events at two or more points in time; or to project future demands on criminal justice and health care resources.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

MEASURES OF QUALITY

The standards applicable to judging the quality of classification and prediction research are basic textbook material in the psychometric literature, but they appear to have been applied unevenly, sporadically, or incompletely in criminal justice contexts. This section discusses three important measures of quality: validity, reliability, and accuracy.3

Validity

Validity generally means that a variable, test, or system of equations actually measures or predicts the theoretical construct it purports to. Three dimensions of validity are considered important in the literature: content validity, construct validity, and criterion validity (Corcoran and Fischer, 1987).

Content Validity Content validity determines whether the independently measured items are a representative sample of the content area to be covered by the instrument or data collection activity. There would be little need to introduce this limited concept of "validity" except as a reminder that truly shoddy classification and prediction research is occasionally carried out. Content validity is ascertained subjectively, either by examining the items and judging if they appear to represent the content ("face validity") or by examining the procedures used in the original research to select the independent variables ("logical content validity"). Examples of procedures that may be used to quantify the validity of self-reported data are measuring consistency among responses to logically identical or reversed questions, examining the extent or pattern of missing (blank) responses, developing "lie" scales (sets of questions whose responses can be combined into an index of untruthfulness), and specifically asking respondents whether or not they are telling the truth.

Construct Validity Construct validity entails determining that the system's measured variables, or combinations of them, correlate with meaningfully related observable variables or actual behaviors, whereas dissimilar observables are not associated with scores, subgroupings, or data items in the same way. In research on violence, construct validity is often highly problematic. The variables that purport to represent the occurrence or extent of

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

violence itself, or causative factors related to violence, are in many cases ill-defined or not representative of the intended behavior.

The reactions of other people and the institutional procedures that intervene between an occurrence of violence and a recorded instance of violence are often complex and extended in time, so that records related to violence are often difficult for researchers to interpret in the context of construct validity. For example, a past history of arrests for violent crimes is not synonymous with being a violent person, nor are recorded instances of "aggression," however defined. When histories of arrests for violent crimes are studied without addressing the question of what they represent, or by indirection leaving the implication that the data indicate whether or not the person is violent, construct validity is violated. Similarly, short follow-up periods for collecting data about persons predicted to be violent or not violent may not permit obtaining valid measures of the construct of interest.

Inadequate construct validity also occurs in independent variables used in violence research. For example, a recorded history of psychiatric treatment does not signal or characterize any clearly articulable past pattern of behavior, so its relationship to contemporaneous or future violence, if any, is difficult at best to interpret. Even many variables that superficially appear to have good construct validity because they involve "hard scientific tests" may on closer inspection be inadequate. For example a positive urine test for opiates is not a valid measure of addiction. A more valid but "softer" measure of addiction is a police officer's notation that an arrestee possessed "works," had recent track marks and ulcers, and showed signs of withdrawal irritability.

Criterion Validity Criterion validity deals with the existence of a relationship between test scores, subgroups, or independent variables and actual behavior, as represented by other measurements or observations (Golden et al., 1984). Criterion validity can be measured concurrently or predictively, whether or not the underlying study itself was predictive. Concurrent criterion validity involves showing that there is a relationship with an alternative method of measuring the same characteristic of interest at the same time, whereas predictive validity attempts to show a relationship with the behavior of interest at a future time. Both classification and prediction studies may be examined for their predictive validity.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Reliability

Reliability refers to the method's trustworthiness, as indicated by stability over time or among different groups, or by consistency in application by different researchers or in different contexts. Theoretically, reliability quantifies the degree to which a constructed test overlaps a perfect measure of the characteristic of interest (Golden et al., 1984).

Measures available for estimating the reliability of classification or prediction research include internal consistency, interrater reliability, and test-retest reliability.

Internal Consistency Measuring internal consistency differs from a consistency check for validity of responses to logically equivalent or reversed items, mentioned above. Coefficient alpha (Cronbach, 1951), a common measure of internal consistency, is based on the average correlations among items purporting to be related to the same theoretical construct. An alpha coefficient exceeding .80 is generally deemed to show that a measurement device is internally consistent. The Kuder-Richardson 20 formula (KR20) (Kuder and Richardson, 1937) is a similar statistical measure designed for dichotomous items.4

Interrater Reliability Interrater reliability is ascertained by having multiple researchers or examiners score or code the identical set of raw data or the same observed behavior. The correlations among corresponding items are then calculated, with a typical standard of acceptance being average correlations exceeding .80 and preferably above .90.

Test-Retest Reliability Test-retest reliability is determined by asking respondents the same questions twice in a single administration of a questionnaire or on two different occasions separated by days or weeks, and calculating correlations between corresponding items.5

Accuracy

Accuracy refers to the discriminating power of the method: the magnitude of the distinctions among subgroups or the proportion of a subgroup actually displaying the outcome predicted for them. A prediction equation is valid if there is some statistically significant correlation between the predictors and the actual outcomes

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

of interest, but it is accurate if there is a strong or utilizable relationship, for example if a sizable proportion of persons predicted to commit violent acts turn out in fact to commit such acts, whereas a sizable proportion of the others turn out not to commit violence.

A considerable body of literature on criminal justice prediction research bemoans the typically low levels of accuracy achieved by prediction models and the lack of standardized or commonly accepted statistics for comparing the relative quality of different prediction instruments or methodologies, especially when applied to different populations. See Gottfredson and Gottfredson (1988b:252) for a good review of this literature as related to violent criminal behavior, together with an extensive bibliography; they state that "reviewing the literature concerning the prediction of dangerousness and the propensity for violence shows that there is little evidence supporting our ability to make these predictions with acceptable accuracy. The prediction of violence is exceptionally difficult, and no one seems to have done it well."

A common observation in recent literature is that predictions of violence are especially difficult due to a low base rate, namely, that in any naturally occurring population only a small proportion of individuals will commit acts of violence. However, a low base rate does not place limits on the ultimately achievable levels of prediction accuracy, which evidently can be as high as 100 percent. To illustrate the independence between base rate and prediction accuracy, consider the case of Tay-Sachs disease, whose incidence is well under 1 percent in the general population but whose occurrence can be predicted with 100 percent accuracy based on the results of appropriate tests. The persons who will develop Tay-Sachs disease (absent an intervening alternative cause of death) can all be predicted by physical tests for the absence of a specific vital enzyme, and anyone who has the enzyme will not develop the disease.6 Of course, before the role of this enzyme was discovered, prediction of future development of Tay-Sachs disease appeared to be formidably difficult. However, the disease's low base rate was neither an obstacle to developing an accurate prediction instrument nor a bar to practical application of the prediction methodology.

A low base rate does, however, present a limitation on the usefulness of inaccurate predictions. If the base rate is low, the default "ignorant" hypothesis that no one in the study group will commit any acts of violence can be very close to the truth, but of course useless. Similarly, predictions of very low or very high

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

probabilities can be accurate but useless. For example, an exactly correct prediction that 97 percent of a specific subgroup of individuals will be nonviolent is not very helpful if the base rate for nonviolence in the entire study group is 96 percent.

In the case of perfect predictions, the selection rate (proportion of the population predicted to have the characteristic in question) is the same as the base rate, but in the realm of violence classification and prediction, typical instruments have selection rates that differ from the base rate. Although often the original researcher can arrange for the selection rate to equal the base rate in the study sample, there is no guarantee that the two will be the same in subsequent applications. Comparing and making sense out of the accuracy levels of two instruments or methodologies that have different selection rates have proved problematic.

Loeber and Dishion (1983) developed the Relative Improvement Over Chance (RIOC) statistic to help permit such comparisons; RIOC calibrates a prediction's improvement in accuracy over random accuracy with respect to the constrained range between maximum accuracy and random accuracy. Farrington and Loeber (1989) present simplified formulas for calculating RIOC and its variance. The formula for IOC used in this chapter is defined in a footnote to Table 1.

CLASSIFICATION AND PREDICTION METHODS

Formal and Informal Methods

Almost all people routinely classify and predict the behavior of others on a day-to-day basis. Classification of violent behavior, as with any form of behavior, is an ongoing social process (Mead, 1934) learned early in life (Kagan, 1982). It depends on one's experience in anticipating future behavior, or behavior in related situations, based on observable acts (Weinstein, 1969) such as nonverbal gestures (Lindesmith and Strauss, 1968) and the individuals' appearance (Goffman, 1959, 1963). Some people classify persons as violent or predict their future violence based on informal processes, whereas others use highly structured statistical procedures.

Informal processes of classification occur commonly. For example, people walking in a city street at night are likely to plan their routes based on their understanding of types of violent offenders and estimates of their prevalence in certain neighborhoods; they will react to strangers in the area by crossing the street or walking faster, based on their own informal classification of the

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

people they observe and their anticipation of possible violence from such people.

The sections that follow describe formal procedures for classifying violent offenders as developed in specific research studies. However, it should be recognized that in practice the results of even the most rigorous classification and prediction methods are interpreted in the context of the possible political, economic, and social consequences of classification decisions.7 The informal, often hidden, methods used for classification depend as much on the organizational setting as on the characteristics of the individuals undergoing classification.

Violent persons are classified by various agencies in the criminal justice system and the public mental health system, and by private psychiatric/psychological service providers. The populations served by these classifiers often overlap, but their classification methods are typically quite different, as are the factors they consider important in regard to violent behavior. A 17-year-old male who has been convicted of rape may, for example, be sentenced to prison, to a psychiatric facility, or to probation with a stipulation of private psychological treatment or counseling in a mental health agency. Depending on where he is sentenced, he will encounter different formal and informal classification methods, and the practitioners who classify him will differ regarding the consequences of making wrong predictions and the risks they are willing accept in making classification or prediction errors.

Selection Bias

A difficulty commonly experienced in quantitatively based classification or prediction studies of violence has been the lack of a clear understanding of the selection process that produced the group under study. In many instances, the research study sample is chosen for reasons of convenience and availability to the researchers. Examples of populations of this type include incarcerated offenders, persons in treatment, or juvenile court referrals. The sequences of events through which individuals came to be members of the study population may be so remote in time and place from the research project that there is little realistic possibility of obtaining descriptions of the processes or statistical probabilities associated with selection for the study population.

Whenever there are no clear and replicable rules for selection of the study population, subsequent difficulties arise in interpreting and implementing the results of the research. Often it is not

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

possible to specify the appropriate groups to which the prediction equations, test score cutoffs, or other classification schemes will apply. For example, the category "convicted male felons" in one state is not necessarily closely similar to the same category in another state.

Often, a research strategy is proposed in which tests, score cutoffs, or other methods are to be developed with one population and then subsequently tested on similar populations to demonstrate their robustness. However, the results to date of classification and prediction research on violence do not give much comfort to those who hope for the presence of robust findings. Instead, the sequential research strategy frequently demonstrates that the classification techniques applied to another group yield entirely different (unstable) categories, or that the classification or prediction equations are highly inaccurate when applied to different study populations.

Determining the Importance of Potential Predictive Factors

A related difficulty is that the original study population, unknown to the researchers, may differ from other populations in regard to its variance on variables strongly correlated with violence. Such variables will not emerge in the analysis as being significantly associated with violence, even if they are measured correctly in the study population. In addition, unmeasured variables pertinent to the selection process may also be predictors of violence. Such variables tend to be forgotten in interpreting the results of the analysis, especially if the researchers never had a genuine opportunity to observe or measure the variables in question. Two-stage prediction studies, in which the probability of selection is estimated first and then the outcome is estimated, can help reduce these types of ambiguities in research results. However, there are not many studies in the existing violence prediction and classification literature that control for potential selection biases in this way.

The problem of classification and prediction instruments not transporting well from one setting to another is compounded by researchers' well-intentioned efforts to maximize prediction accuracy within a construct sample. Stepwise computing algorithms and test-retest procedures used on the same set of data will typically produce predictions that maximize sample-specific correlations; such predictions are prone to shrinkage (reduction in prediction accuracy) when used in a second sample that does not

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

reflect the same sample-specific correlations. Shrinkage tends to be worse in small than in large samples, a notable effect, because samples used in classification and prediction are typically small.

CORRELATES AND CAUSES OF VIOLENCE

Correlates of violence have been reported in an extensive literature reviewed by Weiner (1989) and Loeber and Stouthamer-Loeber (1987). These reviews point out that variables statistically associated with participation in violence (ever having committed a violent act) are not necessary the same as variables statistically associated with persistently committing violent acts over a relatively long period of time (the duration of the ''violent career") or with the number of violent crimes committed during a specified calendar period (crime commission rate or lambda).

Participation in violence is associated with broad social categories such as sex, race, age, and socioeconomic status, but among persons who have participated in violent behavior, other factors have a stronger association with their crime commission rates and persistence. Conversely, correlates of persistence and rates of committing violent crimes—among those who engage in violent behavior—may not be correlated with participation in violent behavior. For example, many people use drugs and many have drinking problems, but most of them are not violent. However, among people who have committed violence, drug users and those with drinking problems are more likely to repeat violence, and among those who have committed a violent act and use drugs, those who frequently use heroin or other opiates are more likely than others to commit violent acts at high rates (Shaffer et al., 1984; Chaiken and Chaiken, 1982, 1987, 1990a; Johnson et al., 1985; Chaiken and Johnson, 1988; Nurco et al., 1988).

To further complicate matters, it should be noted that even the strongest correlates of violent behavior are not necessarily causes of violent behavior. Many correlates are imperfect measures of the underlying causal factors, whereas others are correlated but may actually occur after the commission of violent acts. For example, heroin users often decrease their heroin consumption when funds are low and increase consumption when the proceeds from robberies or other criminal activities are high (Johnson et al., 1985). Also, many violent, drug-using juveniles were delinquent before they started using drugs (see Chaiken and Chaiken, 1990a, for a summary of results showing this relationship). Moreover, participation in drug dealing appears to have a stronger causal

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

relationship to violent acts than drug use—the correlation between frequent drug use and committing violent crimes at high rates may be in large part due to users' involvement in the systemic violence of the drug trade or participation in the symbolic violence of the urban drug culture (Goldstein, 1989; Chaiken and Chaiken, 1990a; Altschuler and Brounstein, 1989).

Similarly, the existence of causal relationships associated with other correlates of violent behavior has been questioned, including such strong correlates as alcohol (Collins, 1989), mental illness (Menzies and Webster, 1989), and child abuse (Widom, 1989a,b). The absence of proved causal relationships does not mean that these correlates are useless for classifying and predicting violent behavior. The sections that follow illustrate that noncausal correlates can provide bases for meaningful utilitarian classification and prediction. However, unlike Tay-Sachs disease, the current state of knowledge is not even close to having an adequate understanding of the causes of violence that would be needed to classify violent offenders and to predict violence with great accuracy.

CLASSIFICATION

PURPOSES AND POPULATIONS CLASSIFIED

The "art" of classifying violent offenders is reminiscent of the traditional story of the classification of the elephant by a team of blind people—each one measuring a different part and variously describing the ear, trunk, leg, tail, or torso. Not only have researchers and practitioners attempted to describe different dimensions of the same "elephant," they have also examined elephants in herds of different age mixes, sex mixes, and settings.

Populations studied for classification of violence have included nationally representative samples of adults (Gelles and Straus, 1988) and youth (Elliott et al., 1989), populations of children (Loeber and Stouthamer-Loeber, 1987), youth and adults in "high-risk" areas (Goldstein, 1989; Fagan et al., 1986; Fagan and Weis, 1990; Simcha-Fagan and Schwartz, 1986; Williams and Kornblum, 1985), clinical populations (Lewis and Balla, 1976), defendants (Chaiken and Chaiken, 1987, 1990b), institutionalized populations of convicted offenders (e.g., Megargee, 1977; Chaiken and Chaiken, 1982), and institutionalized populations of psychiatric patients (Toch and Adams, 1989; Steadman, 1987).

Researchers and practitioners carrying out these classifications have been drawn from such diverse fields as biology, mathematics,

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

neurology, psychology, psychiatry, sociology, and statistics; their measurement techniques have included interviews, surveys, physical examinations, psychiatric examinations, ethnographic methods, and secondary analyses of data originally recorded by criminal justice and clinical practitioners. The variables they have measured have included self-reports of violent and other "deviant" behavior; reports by significant others (teachers, spouses, peers) of violent and deviant behavior; arrests, convictions, and incarcerations for violent behaviors; brain and other neurological abnormalities; and socioeconomic variables such as age, race, sex, education, and employment.

Researchers and practitioners who have different technical backgrounds or disciplinary orientations regularly disagree about the appropriateness of others' classifications, but this paper does not attempt to summarize all these variants and arguments. Rather, a description is given of the major types of methods and measures and of congruent results that have emerged from using them. The examples chosen for presentation are not meant to be comprehensive but are illustrative of methods that have experienced continued application in research or practice.

TYPES OF CLASSIFICATION

Psychological Tests—Rapid Assessments

A large number of psychological tests provide rapid diagnoses for a spectrum of psychological abnormalities (Corcoran and Fischer, 1987) by using simple univariate scaling techniques with demonstrated high reliability and validity. A few of them identify persons with violent personalities, severe psychological problems, and persistent patterns of violence. Some examples follow.

State-Trait Anger Scale (Charles Spielberger and Perry London) Respondents are asked to rank 15 specific statements as applicable to their own general feelings (e.g., "I have a fiery temper") and 15 statements about their own feeling at that moment (e.g., "I am mad"). Scaled responses are used to differentiate between anger as a state (situational response) and anger as a trait (frequent anger over time), and to differentiate between angry temperament and angry reactions. These scales, which have been tested in normal populations, have excellent reliability (.87-.93, depending on the population) and concurrent correlations with other measures of neuroticism, psychoticism, and anxiety.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Problem Solving Inventory (P. Paul Heppner) Based on 35 items about individuals' general responses to problems (e.g., "I make snap judgments and later regret them"), groups with different levels and forms of psychological disturbance are distinguished. Constructed and retested using populations of students and nonstudents, blacks and whites, a spectrum of adult age groups, and "normal" and institutionalized (alcoholic) subjects, tests of internal consistency have resulted in alphas of scales ranging from .72 to .85, and alphas of .90 for total measures.

Index of Spouse Abuse (Walter W. Hudson) Based on respondents' ranking on a five-point scale of the frequency of events described in 30 items assigned weights (e.g., "My partner belittles me"—weighted 1; "My partner becomes abusive when he drinks"—weighted 44), both physical abuse and nonphysical abuse inflicted on a woman spouse have been measured. Tested in both clinical and nonclinical samples, the instrument has good internal consistency (alpha = .90 to .94 for physical scale, .91 to .97 for nonphysical scale), with almost no measurement error, and has high correlation with factors believed to be associated with abuse.

Psychological Tests—Multivariate Scaling Techniques

Some standard psychological tests include measures of psychological or mental status associated with violent behavior, whereas other tests have been constructed specifically to classify violent persons. In clinical or correctional settings, standard tests often supplement violence-specific tests. For example, children who have displayed violent behavior may score differently from other children on portions of the Wechsler Intelligence Scale for Children, the Bender Gestalt Test, the Rorschach, or the Thematic Apperception Test. Observations of children's ability to concentrate on standard tests have also provided indicators of fluctuating attention spans that are symptomatic of neurological disorders and psychoses (Lewis and Balla, 1976).

Other standard psychological tests are administered specifically for diagnosing abnormal psychological states, including several syndromes characterized by violent behavior.

The Minnesota Multiphasic Personality Inventory (MMPI) The MMPI is one of the oldest tests still in use for classifying persons with various psychological problems. Its applicability in criminal justice contexts is enhanced by the development of a set of MMPI

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

profiles that help classify prison inmates. Published in 1940, the MMPI was developed empirically for clinically classifying patients according to syndromes, primarily to determine treatment for a wide spectrum of behaviors (Hathaway and McKinley, 1940). The MMPI questionnaire consists of short declarative sentences in the first person singular, such as "I generally am in good health." Subjects indicate whether or not the declarations apply to them.

In all, 504 items cover 25 topics ranging from general health to phobias. After administration, yes or no responses to specific subsets of items are counted to obtain an individual's raw cumulative score on several scales. The raw scores on each scale are transformed into standard scores with a mean of 50 and a standard deviation of 10 and plotted on a standard profile sheet in which the heights of scores appear as "elevations." The profile sheet is interpreted by a psychologist who examines the highest elevation on all scales and the patterns of elevations on the scales. Scores over 70 generally are thought to be clinically significant (Butcher and Keller, 1984).

The inmate typology, which has widespread use in correctional institutions (Clements, 1986), originally was developed by using MMPI profiles of a sample of youthful offenders in a federal prison (Megaree, 1977; Megargee and Bohn, 1979). Statistical cluster analysis of the MMPI profiles was used to define 10 types of inmates. The groups were given arbitrary labels corresponding to the letters of the alphabet: Able, Baker, Charlie, Delta, Easy, Foxtrot, George, How, Item, and Jupiter.

By examining written institutional staff reports about inmates classified according to this MMPI typology, Megargee and Bohn (1979) determined the model characteristics of individuals belonging to the various MMPI types. For example, Item (the most prevalent type in general inmate populations) is described as having these modal characteristics: "stable, effectively functioning well adjusted group with minimal problems, few authority conflicts." The most prevalent type in a population of psychiatrically disturbed inmates, How, has different modal characteristics: "unstable, agitated, disturbed, 'mental health cases.' Function ineffectively in all areas and have extensive needs."

The MMPI types were also analyzed to see whether they were distinguishable in terms of the extent of violence displayed by their members in the crimes for which they were convicted. The criterion measure was a researcher's ranking of violence as described in each person's presentence investigation report (PSI) for the crime of conviction. However, the types that scored the highest

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

on violence in the conviction offense were a residual unclassified group (Uncle), whose modal characteristics could not be specified, and Jupiter, a type with such low prevalence that it is ordinarily excluded in other tests of the validity and reliability of the typology.

Thus the classification does not appear to yield valid indicators for distinguishing inmates who were violent in situations outside penal institutions. Nonetheless, the typology reportedly has been useful for reducing institutional violence by segregating types that are predatory from types that are prone to victimization. For example, violence within prisons appears to be reduced by segregating Deltas (amoral, hedonistic, egocentric, bright and manipulative, poor relations with peers and authorities, impulsive, sensation seeking leads to frequent infractions) from easily victimized inmates (Megargee et al., 1988).

The utility of the typology is limited by its complexity. Computer-programmed analyses of the MMPI scores were found to classify only two-thirds of the cases into the 10 types; the remaining classifications require clinical judgments for which interrater reliability is poor.

Attraction to Sexual Aggression (ASA) Scale The ASA scale (Malamuth, 1989) instrument asks for respondents' rankings of attitudes and opinions about 13 sexual acts ranging from "necking" to rape and pedophilia. Respondents state whether or not they have ever thought about trying the activity and provide opinions on the attractiveness of the 13 acts, the percentage of males and the percentage of females who would find the acts sexually arousing, the extent to which the respondent would find the acts sexually arousing, and the likelihood of engaging in the acts if there were no negative social repercussions. The dichotomous responses of whether or not the respondent ever thought of trying the activities are combined into the ASA scale by using multivariate scaling techniques; based on the other rankings, several different scales are constructed.

A shorter version of the ASA instrument refers to six sexual behaviors. The reliability and validity of both the long and the short versions of the instrument appear to be acceptable (Malamuth, 1989). Comparisons with other measures of sexual aggression (discussed below) indicate that this scale, applied to sexually experienced men, will distinguish those who are high on sexual aggression and, applied to sexually inexperienced men, has some utility for assessing their potential future risk for sexual aggression.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Child Abuse Potential (CAP) Inventory The CAP was developed as a rapid assessment instrument for screening parents who abuse their children. The instrument's 160 items enter into construction of several scales, including a 77-item abuse scale. The instrument is self-administered, and respondents are asked to agree or disagree with statements. In addition to the diagnostic scales, items are also used to construct a lie scale. By using anonymous responses from a sample of 122 parents identified by caseworkers as physical child abusers and referred to an at-risk parent-child program for treatment, in comparison with responses from a matched control group of parents who had been referred for other agency services, the scale was found to correctly classify 82.7 percent of the abusers and 88.2 percent of the nonabusers when a 215-point cutoff was the basis for classification. The rates of correct classification were found to be higher when respondents who had a high lie scale score were omitted from classification (Milner et al., 1986). By using a lower cutoff score (166), overall classification rates were improved and false positives almost eliminated; however, the number of false negatives increased (Milner and Wimberley, 1980).

Milner and colleagues (1984) followed up 190 of 200 at-risk parents (10 refused) who had one child under 6 months and no prior history of abuse. Participants' child rearing behavior was monitored by the program staff. The analysis was limited to parent abuse as reported by program treatment staff within six months of program selection. Eleven reports of abuse were received for the 190 participants.8 The authors report statistically significant correlations between the CAP score and the occurrence of abuse.

To summarize the findings (Milner et al., 1984): Inspection of the abuse scores indicated that 100 percent of the confirmed abuse cases had scores above the cutoff score for abuse potential reported in the CAP Inventory Manual (Milner, 1980). The 11 reported abusers, however, represent only 10.7 percent of the 103 at-risk subjects who scored above the cutoff score for abuse (Milner, 1980:881). As the authors recognized, it is difficult to know how to interpret these findings. These subjects were in treatment, so the treatment may have reduced the incidence of child abuse. Furthermore, some subjects dropped out of treatment, and other clients failed to attend all counseling sessions, but the data did not distinguish them (Milner et al., 1984).

The relationship between CAP and other measures of psychological

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

problems has been studied (Milner et al., 1988; Robertson and Milner, 1985). Although the instrument appears to be valid and reliable for identifying physically abusive parents needing treatment, a study of the practical uses of the instrument revealed that it was being applied for inappropriate purposes such as differentiating between physically abusive parents and neglectful parents (Milner, 1989).

Psychiatric Classifications

In addition to considering psychological and psychosocial dimensions of violent behaviors, psychiatric classifications also assess organic disorders. For example, DSM-III-R recommends classification on three dimensions: clinical symptoms and conditions that are the focus of attention or treatment but are not attributable to a mental disorder (e.g., adult antisocial behavior); developmental and personality disorders; and physical disorders and conditions. For research and for some clinical settings, two additional dimensions are recommended for evaluation: severity of psychosocial stressors and global assessment of functioning. For each axis, practitioners are urged to provide assessments of their confidence in the evaluation. Some DSM-III-R classes are residual categories to be used after other diagnoses are ruled out. For example, intermittent explosive disorder, characterized by violent episodes, can only be used after ruling out "psychotic disorders, Organic Personality Syndrome, Antisocial and Borderline Personality Disorders, Conduct Disorder, or intoxication with a psychoactive substance" (American Psychiatric Association, 1987:321).

Although a primary diagnosis is recommended, the DSM-III-R classification provides for the occurrence of multiple, not necessarily discrete, disorders, and the editors (Spitzer and Williams, 1987) stress that disorders rather than people are being classified. The disorders are arranged hierarchically, with organic disorders given precedence over other types of disorders (if the organic disorder is responsible for initiating and maintaining the disturbance) and more pervasive disorders given precedence over less pervasive ones (e.g., schizophrenia over dysthymia).

The categories available in the DSM-III-R for classifying outbursts of aggression or rage include:

  • Organic personality syndrome usually due to structural damage to the brain (e.g., neoplasms and head trauma)

  • Dementia (disturbance of higher cortical functioning)

  • Several categories of intoxication, including alcohol, amphetamines,

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
  • cocaine, inhalants, and sedatives, hypnotics, or anxiolytics

  • Nicotine withdrawal

  • Late luteal phase dysphoric disorder (women) occurring a few days before or after onset of menstruation

  • Mental retardation

  • Conduct disorder (children) and antisocial personality disorder (adults)

  • Posttraumatic stress syndrome.

Other categories available for classifying violent behavior include sexual sadism. Reportedly, fewer than 10 percent of rapists are classified in this category, which is categorized by fantasies, frequently beginning in childhood, and possibly including a history of sadistic acts increasing in severity over time. Mood disorders are a residual category for classifying violent behavior not due to any other physical or mental disorder. Injurious behavior to the self or others may also be present in categories that are defined by other symptoms of elevated activity.

Categories in the DSM-III-R were revised from early instruments (the first instrument was published almost 40 years ago) on the basis of clinical experience with early instruments; research conducted to test the validity and reliability of classification; and advancement of knowledge in psychology, medicine, and social psychology.

American Psychiatric Association members instrumental in constructing the current version of the instrument strongly caution that use of the classification scheme is appropriate only if it is a first step in diagnosis, carried out by a trained clinician who is sensitive to cultural differences, and used in clinical or research settings. In particular, they caution against use of this instrument for legal decisions, especially in light of the acknowledged noncomprehensiveness of the classification categories.

Correctional Classification

The primary purposes of correctional classification are to determine the level of security necessary for an inmate and the level of his or her needs for particular services or programs within the institution. Level of security required is usually based on both public risk and institutional risk. Public risk is commonly assessed on the basis of information about the conviction crime, prior convictions for violent crimes, and history of institutional escapes. Institutional risk is assessed from information about prior institutional and custodial behaviors, psychological and mental

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

status, and the simultaneous institutionalization of other inmates known to have a history of conflicting relations with the person undergoing classification (e.g., rival gang members).

In most states, information for risk classification is derived from presentence investigation reports, "rap sheets," and semistructured interviews with the inmates. Classification typically is the clinical judgment of a correctional staff member, guided by the information obtained from these sources. Federal prisons and other prisons with a psychologist on staff may use standard psychological tests such as the MMPI (described above) or other multivariate scaled psychological tests; however, most jails and many prisons do not have the resources for extensive psychological or psychiatric testing of inmates for classification purposes (Clements, 1986).

Clinical Predictions

Clinical predictions are classifications carried out to identify individuals at relatively high risk of committing a specific form of violent behavior, whether or not the behavior has previously occurred. Their primary purpose generally is to permit adequate supervision or treatment that will prevent the anticipated violence. Classification can be based on a wide range of information, including factors additional to those specifically related to the individuals for whom risk is being assessed—for example, information about potential victims.

The clinical criteria developed by Ayoub and Jacewitz (Ayoub et al., 1983) to identify parents "at risk" of child abuse illustrate the kinds of factors used in clinical predictions. These include the following:

  • Biological alerts: premature infants, "difficult" infants, infants with complicating medical problems, mothers with illness, recent history of sibling illness, child's physical characteristics not meeting parental expectations.

  • Psychological alerts: parental childhood history, parental emotional/social isolation, history of past emotional difficulties, intellectual limitations, parental substance abuse or addiction.

  • Social alerts: poor living conditions, financial difficulties, unemployment, mobility, lack of transportation, history of violent and/or illegal activities, medical conditions resulting from medical/nutritional neglect.

  • Interactional alerts: marital difficulties, family difficulties, parental/infant attachment difficulties.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Official Record Data

Juvenile Arrests Perhaps the best known classification of violent offenders—that used in the Philadelphia birth cohort studies—was based on the numbers and types of police contacts accrued by 10,000 boys born in Philadelphia in 1945 and by 28,000 boys and girls born in 1958 (Wolfgang et al., 1972; Tracy et al., 1990). The findings of these seminal studies are now considered axiomatic:

  • Nonwhite youths are more likely to be arrested for violent crimes than white youths: nonwhites in the 1945 cohort were arrested 15 times more often than whites for violent crime; those in the 1958 cohort, 7 times more often.

  • A small subgroup of chronic offenders who were arrested for five or more offenses of any type constituted only 18 percent of the 1945 cohort but were arrested for 71 percent of the cohort's recorded murders, 73 percent of rapes, 82 percent of robberies, and 70 percent of aggravated assaults. The 1958 cohort has a smaller group of chronic offenders, but they had more violent arrests than the 1945 cohort's chronic offenders—independent of their race.

A more recent study using juvenile arrests for classification compared patterns of violent arrests among a sample of 908 children identified in court as neglected or abused, with those of a matched control group of 667 subjects (Rivera and Widom, 1991). Essentially the same racial patterns were found as in the earlier cohort studies. Blacks were much more likely than whites to be arrested for crimes of violence both as juveniles and as adults. Whereas the abused/neglected group of blacks was significantly more likely than the black control group to be arrested at some time for a violent crime (22 versus 13%), the black control group was twice as likely as the white abused/neglected group to be arrested for a violent crime, either as a juvenile or as an adult (13 versus 6.4%).

In this study, a large percentage of those who were violent early appeared to drop out of crime, but those who were persistently violent accumulated a relatively large number of arrests for a variety of crimes. Among those ever arrested for a violent crime, 17 percent had both a juvenile arrest and an adult arrest; one-third of all subjects who were arrested for a violent crime as juveniles were not arrested as adults. Those who were arrested for violence as both juveniles and adults had on the average been arrested 13.1 times for a variety of crimes (including 3.55 arrests for violent crimes); those who were arrested as juveniles for a

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

violent crime but appeared to desist from committing violent crimes as adults had on the average only 2.32 arrests (including violent and nonviolent crimes).

Juvenile Referrals to Court A study based on the juvenile court records of 69,504 people born in Maricopa county, Arizona, and in Utah between 1962 and 1965 (Snyder, 1988) found that approximately one-third of the sample was referred to juvenile court at least once. Of these, 5 percent were charged with a violent index crime (3% aggravated assault and 2% robbery; less than 1% for aggravated rape or murder). Less than 1 percent of the sample referred to court was referred only for violent offenses, and among those referred four or more times, not one individual was referred only for violent offenses. However, although only 41 percent of referred youth had been referred a second time, more than 50 percent of those whose first referral was for robbery were returned.

Official Records of Adult Defendants in Serious Crimes A study by Chaiken and Chaiken (1987, 1990b) examined official record data available to prosecutors in two jurisdictions to learn which items of information most accurately identified offenders as high rate (committing crimes frequently) and dangerous (committing violent crimes). Although much of the information usually available to prosecutors was found useful for identifying high-rate dangerous offenders, other commonly used information proved misleading or ineffective for purposes of identification (Chaiken and Chaiken, 1987, 1990b). The research compared official record information collected from 452 case folders with defendants' self-reports of crimes including robberies and assaults. The results indicated that although some existing guidelines for identifying high-rate dangerous offenders are valid and useful, more accurate use of official data entails a two-stage screening process. In the first (but not very accurate) stage, defendants are tentatively classified as high rate or not high rate; in the second stage, those who pass the first-stage screen are classified as high-rate dangerous or not. All the studied indicators, taken together, were only weakly associated with high-rate offending. A selection rule based on an attempt to predict high-rate offending was found to have very few false positives (less than 2 percent of low-rate offenders in the sample would be classified as high-rate). However, the selection rule would have many false negatives; it would not identify most defendants who are actually high rate.

Once the offenders who passed the first-stage screen had been

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

identified, the subset of high-rate dangerous offenders was characterized by criteria that were based on the offender's conduct in the instant crime or earlier crimes. These were found to be much more powerful than personal characteristics (e.g., age at first arrest, race, employment). Although numerous other valid indicators of dangerousness were found, five official-record items were together statistically nearly equal in value to using all valid indicators of high-rate dangerous offending in the study.

Official Records of Prisoners Aside from the MMPI profiles described above, other information collected in prisoners' case records has been used by researchers to construct typologies of different types of offenders. Numerous studies have shown that there are major differences between official-record information and prisoners' self-reports; neither source can be considered highly reliable (Marquis and Ebener, 1981; Chaiken and Chaiken, 1982). More important, typologies constructed by using self-reports cannot be replicated by using only information about arrests and convictions recorded on rap sheets for the same individuals.

Although rap sheet information, by itself, does not appear to be useful in classifying prisoners, richer information collected by criminal justice and mental health agencies does seem to provide valid and reliable data for classifying violent prisoners. For example, data obtained from public psychiatric institutions and from presentence investigation reports has been used to classify prisoners in a typology that demonstrates clear differences between inmates with histories of mental health problems and those with none (Toch and Adams, 1989).

For inmates with mental health problems, differences appeared among those with substance abuse histories, those with psychiatric histories, and those with both. Inmates with psychiatric histories were more likely than other inmates to be more serious frenzied violent offenders in terms of their conviction crimes, their chronicity of being arrested for violent crimes, and the extent of injury inflicted in their conviction crime. Inmates with substance abuse histories were more likely than other inmates to be ineffectual offenders who could not recall details of their crimes and behaved in a manner likely to lead to their arrest.

Based on cluster analyses, the largest subtype among the inmates with ''pure" mental health problems consisted of "chronic disturbed exploders" who characteristically had consistent and chronic histories of extreme and uncontrolled violence. Among inmates with drug abuse treatment histories, the largest cluster

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

consists of "drug exploders": addict inveterate recidivists with long histories of violence and other criminal involvements. Typically, this is a "career criminal with a dossier of arrests dating to adolescence … who does well in prison … and invariably recidivates, graduating from less serious to more serious offenses" (Toch and Adams, 1989). Among the inmates with neither substance abuse nor psychiatric histories, the largest subtype was the ''early career robbers": young men who had been convicted of violent offenses in the past, started committing robberies at an early age, and were likely to be imprisoned in the past and on parole or probation at the time they were arrested. The types constructed in this study are congruent with those constructed by using self-reports of inmates (discussed below). For examples, drug exploders may be analogous to violent predators, and chronic disturbed exploders may be analogous to the high-rate "mere" assaulters discussed below.

Crime Event/Criminal Profile Data To assist its agents in identifying unapprehended violent offenders, the Federal Bureau of Investigation (FBI) typology of violent persons has been constructed to distinguish between two types of suspects in violent crimes. Based on analysis of past cases, it distinguishes between disorganized asocial offenders and organized nonsocial offenders (Holmes, 1989). The former type is below average in terms of intelligence, education, and social and employment skills, and is nocturnal and a loner. The latter type is highly intelligent, sexually adequate, charming, and socially and occupationally mobile. The crime scene is used to provide evidence about the type of violent offender involved in the case, for example, whether the event appeared to be planned, restraints were used, the crime scene was controlled or chaotic, and the weapon was removed or left behind. The resulting offender profile is used to assess actions that will be taken by the offender after the crime (e.g., the disorganized offender is more likely to attend the funeral) and the reaction during questioning if apprehended.

Similarly, based on analysis of information collected about mass murderers (defined as persons who have killed four or more people), profiles of three types have been constructed (Levin and Fox, 1991; Fox and Levin, 1994). One type commits numerous murders for economic gain, usually as adjunct to robbery, and may commit the murders either serially or in a simultaneous slaughter. The other types commit multiple murders that are not incidental to robbery.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

The massacre murderer who simultaneously kills more than four people has been described as a person who does not fit any well-defined psychiatric class, although some rare cases involve serious mental disorders such as paranoia, in which victims are believed to be conspiring or voices are imagined to have suggested the crime. Massacre murderers usually have difficulty coping with failures in their own lives—often committing suicide after the event. They attribute their own misfortunes, including loss of jobs or marital difficulties, to other people. They tend to be white males between 30 and 40 who use a gun; they usually are well trained in use of firearms and have done recreational shooting under stress in the past. Typically they are loners with weak support systems who by themselves kill victims they know—family members or coworkers. Very few kill at random.

Serial murderers, on the other hand, kill to fulfill sexual fantasies, for fun, or to feel important. The motives are expressive and not, like the massacre murderer, instrumental. They rarely are recognized or diagnosed as having a serious mental disorder but frequently could be classified as having an antisocial or narcissistic personality by using DSM-III-R classifications. For the most part, they are methodical planners. They rarely use firearms and are most likely to commit strong-arm murders; stabbings are more common than gunshots but still are rare. The serial murderers predominantly are white males and vary in age, although the modal age is around 35. Their victims are usually strangers.

Self-Reports

Self-Reports Collected Through Surveys of National Samples Self-reports of violent offenses based on national samples have helped confirm many findings about the classification of violent offenders based initially on small samples or criminal justice statistics. For example, a national sample of 2,146 individual family members was interviewed in person in 1976 and asked about the frequency of violent acts between various family members ranging from shoving to using a knife or firing a gun. Among these respondents, 75 percent of parents reported striking their child at least once during the child's lifetime, whereas only 3 percent could be considered repeatedly violent, reporting that they "kick, bite, or punch their child at least once each year" (Gelles and Straus, 1988:103). Surveys based on national samples of youth have also confirmed that only a small percentage of teenage boys can be classified as persistent violent offenders, and these boys are likely

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

to commit many violent and nonviolent delinquent acts (Elliott et al., 1989)

Self-Reports Based on Surveys of Offender Populations Self-reports of violent and nonviolent criminal behavior have provided fruitful sources of classification for a number of researchers who have studied the interrelationship between dimensions of violent behavior (rates of committing crimes, persistence, and combinations of crimes committed) and correlates of these different dimensions. Although the reliability and validity of self-reports from offenders are never close to perfect, they have been adequate to demonstrate strong relationships such as those between drug dealing and committing violent crimes (Chaiken and Chaiken, 1982; Nurco et al., 1988). For example, Chaiken and Chaiken (1982) classified self-reports of approximately 2,000 inmates in prisons and jails in three states and self-reports from approximately 500 defendants in two states (Chaiken and Chaiken, 1987, 1990b) according to public perception of the seriousness of crimes they had committed, with the most serious categories specifically committing violent crimes.

The types appear to differ along several dimensions of seriousness. Inmates and defendants who were classified in the most serious groups, according to the crimes they reported committing, were also on the average more serious in terms of substance abuse, irregular employment, juvenile involvement in serious crimes, early age of onset of criminal behavior, and social instability (Chaiken and Chaiken, 1982; Chaiken, 1990). In addition, the statistical association between the seriousness level of the classification and the estimated annual rates at which respondents committed specific crimes was very strong. The classification was replicated by using self-reports from unincarcerated offenders and again showed that offenders classified as violent based on types of crimes they committed were also high-rate offenders (Johnson et al., 1985).

Epidemiologic Studies

The primary purpose of recently proposed epidemiologic studies of violence is to classify groups who have a high risk of interpersonal violence, rather than violent individuals. Based on surveillance data such as the FBI Uniform Crime Reports (UCR) and the Bureau of Justice Statistics' victimization survey data, high-risk groups could be identified through simple descriptive statistics of risk factors (Mercy and O'Carroll, 1988). For example,

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

based on current knowledge of correlates of interpersonal violence, groups at high risk could be characterized as those in which alcohol consumption is high, illicitly purchased handguns and drugs are readily available, and income level is low.

Observational Studies

Observational studies have been carried out in both natural and laboratory settings. Laboratory studies have for the most part used as a measure of aggression, the willingness of subjects to apply painful stimuli to others (in many studies, unknown to the subjects, the "painful" stimuli were in fact simulated). Studies of male sexual aggressiveness have used penile tumescence as a measure of sexual arousal in reaction to violence.

Studies by Malamuth (1986, 1988, 1989) explored the relationship between scores on a number of dimensions of violent behaviors and psychological states. The measures used included delivery of aversive noises to female and male (confederate) subjects (the outcome variable), penile tumescence in reaction to stories of rape and consensual intercourse, scores on several rapid assessment tests including the Acceptance of Interpersonal Violence Scale and the Sex-Role Stereotyping Scale; scores on several subscales of multivariate tests including a dominance subscale and the Psychoticism Scale of the Eysenck Personality Questionnaire and self-reports of a variety of violent acts actually committed.

With the exception of sex-role stereotyping, all these measures were significantly correlated with the measure of aggression toward female subjects but not toward male subjects. However, when used in combination, other measures proved to be more powerful discriminators than penile tumescence in response to stories about rape. Based on these findings, Malamuth (1988:490) concluded that "men who are aggressive toward women (e.g., rapists) are likely to commit other aggressive acts as well" and that rape and other forms of aggression may be related to common underlying factors.

Reports of persons who observe children in the normal course of events are often used in research concerning classification of violent children. For example, a review of prediction of juvenile behavior (Loeber and Stouthamer-Loeber, 1987) indicates that 9 out of 11 longitudinal studies of delinquency used teacher ratings of aggression; one used both teacher and peer ratings. Reports of teachers are used frequently in court assessments of children. Data gathered on teacher's report forms and direct observation forms,

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

as well as separate interviews with each parent, are integral to several standardized clinical (and forensic) assessments of children (McConaughy et al., 1988). In multiaxial approaches such as the DSM-III-R, such information can be used for classification on several axes. Moreover, interviews with the parents of children being assessed often reveal important observational information about the parents themselves.

CONGRUENT FINDINGS ABOUT INDIVIDUALS CLASSIFIED AS VIOLENT PERSONS

Although practitioners and researchers often disagree about classifying persons who have committed relatively few violent acts, congruent findings have merged across disciplines about offenders who are relatively extreme on all dimensions of violent behavior—the rates at which they commit violent acts, the seriousness of the acts they commit, and their persistence. This section summarizes the congruent findings.

  1. Although many people may at some time in their lives commit a more or less violent act, the majority of people do not repeatedly commit violent acts. Age, sex, and race are important variables in differentiating between people who have and have not committed violent acts; among people who have already committed violent acts, age, sex, and race are less helpful in differentiating those who commit many violent acts from those who commit few.

  2. Persons who are repeatedly violent typically also demonstrate other forms of antisocial or self-destructive behavior, often starting in early childhood, even more frequently than they demonstrate violence toward others. The frequency and seriousness of these other forms of socially undesirable and self-destructive behaviors are indicative of the frequency and seriousness of their violent behavior.

  3. Not surprisingly, persons who are repeatedly violent are very visible to those with whom they frequently interact: peers, family members, and (for children and youth) teachers. However, perhaps because violence is episodic, practitioners who have short-term contact with people are not likely to be able to identify the violent ones based solely on personal observations.

  4. Persons who commit violent acts over long periods of time (persistent violence) are not necessarily the same people who commit numerous violent acts in relatively short periods of time (high-rate violence). Similarly, persons who cause or attempt to cause

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

serious injuries to others may be neither persistent nor high rate. However, among violent persons, a subgroup has been identified by numerous researchers and practitioners as persistent, high-rate, and dangerous. In any population not selected for these characteristics, very few people would fall into this category, and even among incarcerated populations there are relatively few persistent, high-rate dangerous offenders.

  1. Violent persons may be classified into groups having different probabilities of specific behavior, but they can almost never be definitively classified as to their individual behavior. One reason that no perfect classification scheme exists is that available measures of violence are themselves imperfect. Self-reports are subject to both overreporting, resulting in false positives in classifications based on such data, and underreporting, resulting in false negatives. Information collected by the criminal justice system on arrests and convictions also does not provide reliable data for classifying violent persons. For various reasons, victims are reluctant to report persons who have abused or threatened them with violence to the police, so that many people who have committed several violent offenses are unknown to the criminal justice system and would not be correctly classified (false negatives). Conversely, many people who have committed violent crimes are either inept or uninterested in trying to conceal their crimes. These offenders are likely to be arrested for nearly every crime they commit, and classifications based on official record data are likely to falsely classify them as high-rate offenders (false positives). Physical, psychological, and social characteristics that are disproportionately likely to occur among specific types of violent persons nonetheless also occur among nonviolent persons and among persons who manifest other types of violence.

  2. Independent of the theoretical basis of classification and the specific measures used, classifications that explicitly or implicitly measure scaled multivariate abnormality appear to be more reliable and valid for distinguishing the most serious violent offenders from less serious offenders than classifications that depend on univariate measures or measures that do not capture associated dimensions of seriousness.

  3. Rather than being additive, there is usually collinearity between measures found useful for classifying violent offenders.

  4. The most important factor that appears to decrease the reliability of classification over time is desistance. Persons who are classified as displaying a particular form of violence may, for

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

reasons that are not as predictable, later stop committing all kinds of violent or even antisocial acts.

  1. Depending on the context, it may be desirable to reduce false positive rates in a classification system at the expense of increasing false negative rates, or alternatively to reduce or virtually eliminate false negative rates by increasing false positive rates.

  2. Independent of caveats provided by researchers and practitioners who develop classification methods, other researchers and practitioners are likely to apply readily accessible methods to achieve inappropriate objectives. For example, methods developed for diagnosis are inappropriately used for prediction.

PREDICTION

This section provides a review of selected empirical studies of predicting violent behavior. The selection was derived from articles brought to our attention by the staff and members of the Panel on the Understanding and Control of Violent Behavior, participants at the panel's sessions, a computer literature search,9 and communications with researchers recognized for their contributions to the study of violence. With some exceptions, we limited our review to papers that have been published between 1979 and 1989. Earlier works are included in the reviews by Monahan (1981) and Klassen and O'Connor (1989a).

The studies reviewed here are all multivariate predictions, as that term was defined earlier. First, we provide a methodological critique; then we summarize results from the 19 studies that are listed in Table 1.

METHODOLOGICAL CRITIQUE

Most predictions of violence are based on occurrence models, failure-time models, or rate models. Occurrence models predict the probability that a violent event will occur within a specified period of time, failure-time models predict the length of time until occurrence of a violent event, and rate models predict the number of violent events to be committed within a specified time period. In any of these three approaches, the major methodological problems that arise often involve censoring events, sample selection bias, or both.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

TABLE 1 Summary of Studies Predicting Violence

Key to Table:

Authors (date)

Construction

Size and composition of the construction sample

Validation

Size and composition of the validation sample

Criterion

Definition of the criterion variable

Follow-up

Length of the follow-up period

Fail rate

Percentage engaging in violence according to the criterion variable

Accuracy

Accuracy of prediction using the validation sample, when available, otherwise using the construction sample; reported as false positives, false negatives, and improvement over chance (IOC)a

Significant variables

Variables that were reported as being statistically significant in the prediction

Belenko, Chin, and Fagin (1989)

Construction

3,139 defendants arrested for "crack" cocaine drug law violations and 3,204 arrested for "other" cocaine drug law violations

Validation

None

Criterion

Arrest for a violent crime

Follow-up

2 years

Fail rate

Mean arrest rate of 0.17 for crack drug law violators and 0.09 for other cocaine drug law violators.

Accuracy

64% true positives, 37% false positives

Significant variables

Age, race, total prior arrests, prior arrest for violent offense, arrest for crack rather than cocaine drug law

Black and Spinks (1985)

Construction

125 mentally disordered offenders

Validation

None

Criterion

Occurrence of an assault

Follow-up

5 years

Fail rate

13 of 125 (10%)

Accuracy

44% false positives, 4% false negatives, 10% IOC

Significant variables

Prior convictions, prior psychiatric admissions for violent crimes, type of presenting offense

Cocozza and Steadman (1974)

Construction

98 middle-aged people who had been continuously hospitalized as criminally insane for an average of 14 years

Validation

None

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Criterion

Violent assaults based on arrests and hospital reports

Follow-up

Not specified

Fail rate

14 of 98 (14%)

Accuracy

69% false positives, 5% false negatives, -4% IOCb

Significant variables

Age, juvenile record, prior arrests, prior convictions for violent crime, severity of presenting offense

Cook and Nagin (1979)

Construction

4,154 individuals who had been arrested for a crime of violence (murder, rape, assault and robbery)

Validation

None

Criterion

Arrest for a crime of violence

Follow-up

3 years

Fail rate

14% of those charged with murder, 23% of those charged with assault, 29% of those charged with rape, and 38% of those charged with robbery

Accuracy

Not reported

Significant variables

Various, depending on criterion variable, including age and prior arrests for crimes of violence

Dembo and Colleagues (1991)

Construction

201 boys detained in a regional detention center

Validation

None

Criterion

Index of violent recidivism

Follow-up

Variable, up to 1 year

Fail rate

76%

Accuracy

r2 = .23

Significant variables

Age, abuse of alcohol, number of prior self-reported crimes against persons

Farrington (1989)

Construction

411 London males

Validation

None

Criterion

Various: self-reports and reports from secondary sources

Follow-up

From age 8 until age 32

Fail rate

134 were identified as aggressive by teachers, 119 self-identified as violent during teenage years, 140 admitted getting into a fight during the five years prior to age 32, and 50 were convicted of a violent crime between the ages of 10 and 32

Accuracy

r = .47 for adolescent aggression, .49 for teenage violence, .44 for adult violence, and .37 for convictions for violence

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Significant variables

Various, depending on criterion variable, including poor school adjustment, general tendency toward delinquent behavior, family background, miscellaneous indicators of risk taking, being nervous and withdrawn, and size and intelligence

Garrison (1984)

Construction

100 male children treated in a psychiatric treatment facility

Validation

None

Criterion

Intensive physical attack on other persons

Follow-up

5 hours per week for 2 years

Fail rate

44% of 1,038 aggressive incidents were violent acts

Accuracy

Reported as "fairly low" and only marginally better than chance

Significant variables

Age, victim status, events precipitating

Holland, Holt, and Beckett (1982)

Construction

198 adult offenders placed on probation

Validation

None

Criterion

Arrest for armed robbery, aggravated assault, forcible rape, or homicide

Follow-up

32 months

Fail rate

22 of 198 (11%)

Accuracy

IOC = 1%c

Significant variables

Age and prior convictions for nonviolent crimes

Howell and Pugliesi (1988)

Construction

930 married and cohabitating men

Validation

None

Criterion

Minor or severe violence against a spouse

Follow-up

1 year

Fail rate

177 of 930 (19%)

Accuracy

Not reported

Significant

 

variables

Age, occupational status (blue collar versus white collar), unemployment, parents were violent

Klassen and O'Connor (1988a-c)

Construction

239 male inpatients considered to be potentially violent

Validation

None

Criterion

Arrest for a violent crime or readmission to the mental health center for an act of violence

Follow-up

Up to 6 months

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Fail Rate

46 of 239 (29%)

Accuracy

41% false positive, 6% false negative, 18% IOC, r2 = .34

Significant variables

Age, prior criminal record, prior arrests for violent crimes, number of prior violent incidents, assault as reason for hospitalization, family interactions, ongoing social relationships, assaultive when drinking, suicide attempts

Klassen and O'Connor (1989b)

Construction

251 male inpatients considered to be potentially violent

Validation

265 male inpatients considered to be potentially violent

Criterion

Arrest for a violent crime or readmission to the mental health center for an act of violence

Follow-up

Up to 12 months

Fail rate

74 of 251 (29%) in the construction sample

Accuracy

48% false positive, 17% false negative, 13% IOC, 0.36 RIOC

Significant variables

Early family quality, current intimate relationships, prior arrest history, admission history, assault in the presenting problem

Malamuth (1986)

Construction

155 male volunteers

Validation

None

Criterion

Self-report scale of sexual aggression

Follow-up

Not applicable

Fail rate

Not applicable

Accuracy

r2 = .45

Significant variables

Dominance as a sexual motive, hostility toward women, attitudes facilitating violence, sexual experience, and sexual arousal in response to observed rape

Menzies, Webster, and Sepajak (1985)

Construction

211 patients at a pretrial forensic clinic

Validation

None

Criterion

11-point scale of violence

Follow-up

Variable, up to 2 years

Fail rate

Not reported

Accuracy

r2 = .12

Significant variables

Factor scores: tolerance, capacity for empathy, capacity for change, and hostility

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Moss, Johnson, and Hosford (1984)

Construction

96 inmates in a federal prison

Validation

None

Criterion

(1) Participation in a major riot and (2) arrest for a violent crime

Follow-up

10 years for the arrest criterion variable

Fail rate

44 of 79 were arrested for a crime of violence (56%)

Accuracy

Variables used did not predict arrest

Significant variables

None

Rhodes (1985)

Construction

1,711 offenders released from federal prison

Validation

None

Criterion

Arrest for a crime of violence

Follow-up

5 years

Fail rate

Homicide (1.8%), kidnapping (0.3%), rape (1.1%), robbery (6.7%), and assault (7.7%)

Accuracy

Not reported

Significant variables

Age, race, gender, prior criminal record, drug use, violent offense led to conviction

Steadman and Morrissey (1982)

Construction

257 males indicted for a felony and found incompetent to stand trial

Validation

282 males who had competency hearings prior to indictment; 250 males who had been committed involuntarily

Criterion

(1) Arrest for a violent crime or rehospitalization for assaultive behavior and (2) assaultive behavior while a patient

Follow-up

Not specified

Fail rate

Construction: 28 of 154 (in the community) and 100 of 256 (in the hospital); validation 1:39 of 227 and 85 of 282; validation 2:22 of 117 and 11 of 147

Accuracyd

Validation 1:100% false positives, 18% false negatives, and -21% IOCe (in the community) and 65% false positives, 27% false negatives, and 5% IOC (in the hospital); validation 2:100% false positives, 19% false negatives, and -19% IOCf (in the community) and 90% false positives, 6% false negatives, and -3% IOC (in the hospital)

Significant variables

For hospital assaultiveness: age, race, alcohol problems, juvenile record; for community assaultiveness: prior arrests for violent crimes and age at first hospitalization

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Virkkunen and Colleagues (1989a, b)

Construction

36 men who attempted manslaughter and 22 who committed arson

Validation

None

Criterion

New violent offense or arson as identified by police reports

Follow-up

Variable, averaging 3 years

Fail rate

13 of 58 (22%)

Accuracy

25% false positive, 14% false negative, 0.19 IOCg

Significant variables

Blood glucose nadir and CSF 5-HIAA

Weiner (no date)

Construction

1,355 Philadelphia youths who had police contacts between their tenth and eighteenth birthdays

Validation

Holdout sample

Criterion

Arrest for a UCR violent offense

Follow-up

Variable length

Fail rate

Multiple, reported for different arrest transitions and for different subsets of the sample

Accuracy

Reported as very low

Significant variables

Race

Widom (1989a, b)

Construction

908 subjects who had been abused or neglected as children and a matched sample of 667 who had not been abused or neglected

Validation

None

Criterion

Officially recorded (juvenile or adult) crimes of violence

Follow-up

Variable length

Fail rate

9.8%

Accuracy

Reported as low

Significant variables

Age, race, gender, physical abuse as a child, neglect as a child

a To compute predictions attributable to chance, the percentage of failures in the total sample is multiplied by the number of failures in the total sample. The product is added to the percentage of successes in the sample multiplied by the number of successes in the sample. The sum is then divided by the total sample size. To compute predictions attributable to statistical analysis, the number of correct predictions (success and failure) is divided by the size of the sample. The IOC equals the percentage of correct predictions attributed to statistical analysis minus the percentage of correct predictions attributed to chance. The IOC is sensitive to the criterion used to predict a failure (i.e., the cutoff criterion). When computing IOC, we assumed that the cutoff should be set so that the predicted

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

number of failures should equal the actual number of failures. However, published results did not always allow us to follow this rule, and alternative rules are identified with footnotes.

b Although only 14% of the subjects were violent, Cocozza and Steadman (1974) predicted that 37% would be violent, an assumption that likely accounts for the negative IOC. We were unable to compute IOC using alternative assumptions.

c The prediction was based on the number of prior nonviolent convictions only, because Holland et al. (1982) do not report regression results using violent and nonviolent convictions as independent variables with sufficient detail to compute summary statistics. The negative value for IOC appears to result from a nonlinear relationship between prior convictions and violence during the follow-up period, because we used the highest category ''10 or more nonviolent convictions" as the cutoff.

d Statistics were reported by the authors.

e Only 4 of 227 were predicted to be violent. In fact, 39 were violent. The IOC is practically a product of predicting almost all subjects to succeed in the community.

f Only 1 of 117 were predicted to be violent. In fact, 22 were violent. The IOC is practically a product of predicting almost all subjects to succeed in the community.

Censoring Events

All empirical prediction studies examine the occurrence of violent acts during a specified follow-up period. Those follow-up periods are frequently "censored"; that is, the length of a person's follow-up period may be truncated for reasons not directly related to violence (e.g., commitment to prison for a nonviolent crime, commitment to a mental health facility, moving to another area, or death). If the probability of a violent act occurring increases with the length of time at risk to commit violent acts, controlling for each person's time at risk would seem to be important when predicting violence. Yet many studies fail to introduce adequate controls, either because the researchers did not make full use of their data or because their data are inadequate to support the necessary analysis.

The degree to which failure to adjust for censoring may affect predictions of violence varies across studies. Among the studies listed in Table 1, we note a failure to adequately adjust for censoring to be a potential problem in Weiner (no date), Klassen and O'Connor (1988a-c), Steadman and Morrissey (1982), Black and Spinks (1985), and Belenko et al. (1989).

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Sample Selection Bias

One reason for predicting violence is to inform policy makers and decision makers of the likely consequences of failing to take actions to restrain individuals with violent tendencies. As examples, a psychiatrist might advise a judge about the likelihood of violence if a patient were to be released today as opposed to one year from today, and a statistician might inform a parole board of the probability of violence if an individual were to be released after five years rather than after ten years of confinement. Such predictions of violence are made conditional on the absence of continued social restraints.10

Estimating prediction equations is complicated by the fact that observations of violence made during a follow-up period are seldom untainted by social interventions into the lives of those who are being studied. As illustrations, in the criminal justice arena, individuals perceived to have a high risk of committing future violence are often subjected to more extended periods of confinement and more intensive community supervision than are individuals thought to be less violent; in the mental health field, those who appear violent may be detained longer in hospitals and receive more intensive aftercare than do those who do not appear to be violent. The data at an analyst's disposal have necessarily been "selected" by social processes that attempt to mitigate the very behavior that the analyst is attempting to predict. Unless analysts take steps to compensate for these practical responses to violence, their predictions may say less about human behavior than about processes of social control.11

Analyzing data to distinguish between unfettered human behavior and consequences of social control is no easy task, and there is no absolute standard to judge whether an analyst has done an adequate job. Nevertheless, ignoring the problem by simply asserting that one's analysis pertains to violent behavior, without considering or mentioning the ways in which the study subjects may have been restrained from violent acts, is almost certain to muddle predictions of violence. Incorporating in models some variables that distinguish between violent behavior and social responses is justified, even if data are uncertain and the model can merely test the predictions' sensitivity to underlying assumptions.

Some of the studies in Table 1 that are mentioned above as suffering from data censoring are also subject to selection bias: Steadman and Morrissey (1982), and Black and Spinks (1985). Selection bias problems arise also in the study by Dembo et al. (1991).

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Statistical Limitations

Aside from data censoring and sample selection bias, which raise methodological suspicion about the substance of studies' conclusions about violence, many prediction studies evidence other statistical weaknesses such as use of nonrigorous or nonoptimal methods. Four of these are discussed here.

Dichotomous Dependent Variables When researchers estimate occurrence models, the dependent variable is dichotomous, usually coded one (1) when violence occurred and zero (0) otherwise. Discriminate analysis and ordinary least squares (OLS) regression are the most frequently used methods to analyze such data, despite the fact that these tools lack strong theoretical justification in these contexts. Our own preference is for more rigorous methods such as probit, logit, or log-linear contingency tables analysis. However, because discriminate analysis and OLS regression have proved to be robust in other, similar applications, we doubt that they have any important effect on the conclusions of prediction studies.

Dependent Variables With Lower Limits We are less sanguine about using OLS regression when the dependent variable has a lower limit, such as occurs with the number of violent events during a follow-up period. Given that OLS will yield biased parameter estimates under such conditions, our preference is for regression techniques that are suitable for censored and truncated dependent variables such as Tobit analysis (see Maddala, 1983) and for "countable" dependent variables such as Poisson regressions (see Maddala, 1983; Holden, 1985; King, 1988).

Judging how much findings would be affected by substituting more suitable estimation techniques for OLS regression is speculative, of course, and alternative techniques have their own limitations. Nevertheless, use of alternative estimating techniques can be valuable for checking the sensitivity of reported results to the statistical procedures used in reaching those results.

Stepwise Regression Stepwise regression and related techniques, such as all-subsets regression analysis, can obscure causal relationships when independent variables are strongly correlated. It can also produce predictions that display greater shrinkage (reduction in predictive power when applied to an independent validation sample) than predictions that are not based on stepwise estimation

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

(Copas, 1983; Copas and Tarling, 1986). This problem can be especially acute when—as is the case with many studies of predicting violence—sample sizes are small. Nevertheless, many researchers have used stepwise regression to fit their models. Given the small samples available to most analysts, stepwise regression accentuates the risk of discarding variables that are predictive of future violence, while retaining variables that have only spurious correlations with future violence and possibly overstating the power of the prediction.

Efficient Use of Data

Most studies that we reviewed were based on occurrence models. Although there is nothing wrong with a model that simply tries to predict the occurrence or nonoccurrence of a violent event during a follow-up period, it is not as powerful as methods that take into account the timing of violent events, especially when data sets include timing information. Employing survival techniques would have made fuller use of these data and perhaps would have provided a more precise, less time-dependent prediction of violent recidivism (Kalbfleish and Prentice, 1980; Maltz, 1984; Schmidt and Witte, 1988). In addition, many data bases appear to allow for records of multiple instances of violence, but the researchers focused only on the first instance of violence. Again, there is nothing wrong with such a focus, but additional information might be tapped by using panel data techniques (Heckman and Singer, 1985), Poisson regression models, or other techniques that take advantage of the information inherent in the occurrence of multiple events.

Although all predictions have inherent problems of shrinkage when moving from a prediction to a validation sample, some techniques can minimize the extent of shrinkage. One is to avoid using small samples or large numbers of predictor variables compared to the size of the sample. Another is to use a different criterion variable in the construction sample than is pertinent for the validation sample.

PREDICTING VIOLENT BEHAVIOR

The discussion of studies' substantive and statistical limitations in the previous section is not intended to condemn the existing body of literature. Although studies of the prediction of violence could be methodologically improved, the studies we reviewed

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

are sufficient to permit meaningful interpretations. This section summarizes the content of the reviewed studies, and the sections that follow attempt to draw a cumulative picture of what variables seem to predict violence and how well those variables predict violent behavior. Given the disparity in study methods used, it was not possible to compare results from one study directly with results from another.

This section is organized around the three general empirical approaches mentioned earlier: occurrence models, failure-time models, and rate models.

Occurrence Models

As previously described, occurrence models use various statistical methods to estimate whether a violent event will or will not occur during a specified period and do not distinguish the occurrence of two or more violent events from the occurrence of a single violent event.

Klassen and O'Connor In a series of related studies, Klassen and O'Connor (1988a-c) first analyzed violence during a six-month follow-up of a group of 239 men who had been admitted as inpatients to an urban community mental health center and were considered to be potentially violent. The criterion measure was an arrest for a violent crime (simple assault, aggravated assault, arson, robbery, rape, or homicide) or readmission for violence to the mental health center during the five-month follow-up. To be counted as recidivism, the violence that led to readmission must have been judged by the researchers to have been criminal. Because most studies of violent recidivism are limited to criminal justice data, it is noteworthy that in the Klassen and O'Connor data about 60 percent of incidents recorded as violent recidivism were actually readmissions to the mental health center.

The authors were concerned specifically with short-term prediction, which—following Monahan (1984)—they argued should be the primary focus of a clinician. Given this interest in short-term prediction, the authors were especially attentive to the subject's family ties at the time of release from the hospital.

They found that several variables were correlated with future violence. Some of these were criminal record variables: number of arrests for disturbing the peace, number of arrests for violent crimes during the last year, number of violent incidents during the last year, and assault as a reason for the hospital admission.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Other variables were descriptive of the patient's family interaction during childhood: injured by a sibling before age 15, father died before age 15, parents did not provide well for needs, parents had physical fights with each other, and dissatisfaction with siblings. Still other variables characterized ongoing social relationships: never married, dissatisfaction with extended family, how often sees mother, how long ago last sexual intercourse, lives with parent, and how long ago last relationship with a woman. In addition, information about suicide attempts in the presenting problem, age, being assaultive when drinking, and abstract reasoning score strengthened the predictions of violence.

Of the 239 subjects, 46 were violent. The total correct classification was 85 percent; chance prediction would have yielded a rate of 19 percent. The false positive rate was 41 percent; the false negative rate was 6 percent.

Substantial shrinkage can be anticipated because the sample was small, the number of predictors was large, and stepwise procedures were used to estimate the model. Using a "standard formula" for preshrinking their estimates, the authors predicted a decline from .34 to .21 in the model's r2.

In an unpublished study, Klassen and O'Connor (no date) used all-subsets regression to reanalyze the above data after the follow-up period had been extended to one year. In this reanalysis, 251 subjects had been living in the community for at least three months. Of these 251, 74 had been "violent."

The authors report that the multiple r was .45, with a correct classification of 74 percent, a false positive rate of 45 percent, and a false negative rate of 19 percent. By chance, the false positive rate would be 30 percent and the false negative rate would be 70 percent, so the predictions were considerably better than chance.

Building on these findings, Klassen and O'Connor (1989b) returned to the data still again, this time with a validation sample of 265 men who were in the community for at least three months during a one-year follow-up. Eliminating from consideration the variables that had inconsistent parameters in the two regressions in their earlier validation on 100 subjects, as well as variables that lacked a significant bivariate correlation with future violence, Klassen and O'Connor used all-subsets regression to estimate a new prediction instrument. On the basis of this analysis, they retained five scales as predictor variables: early family quality, current intimate relationships, arrest history, admission history, and assault in the presenting problem.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Applying the predictions from the construction sample to the validation sample, Klassen and O'Connor (1989b:78) report:

A total of 75.8 percent of the subjects were correctly classified… By chance, 63 percent would be correctly classified overall, so these results represent a 13 percent improvement over chance. … Calculation of [RIOC] yielded a value of .36. The false positive rate was 47.6 percent and the false negative rate was 16.8 percent. Of the violent group, 49.3 percent were correctly identified as violent. By chance, only 25 percent of the violent would be identified.

The researchers also report that, contrary to their expectations, variables that characterize ongoing social relationships played a minor role in the predictions. Among them, only family satisfaction was statistically significant.

Steadman and Morrissey (1982) Steadman and Morrissey had data from three groups: (1) 257 persons who were indicted for a felony and found incompetent to stand trial; (2) 282 individuals who were unindicted for their felony charges and who required a psychiatric determination of competency to stand trial; and (3) 250 males who had been committed involuntarily (civil commitment) to one of six state mental hospitals.

The group of 257 persons was used to develop two prediction instruments that were tested by using data from the other two groups. There were two criterion variables: (1) rearrest for violent crime or rehospitalization for assaultive behavior, and (2) assaultive behavior while a patient, defined as a physical attack, not in self-defense, against another. Possible prediction variables included data about sociodemographic factors, criminal history, and mental hospitalizations. Discriminate analysis was used as the estimation technique.

Steadman and Morrisey (1982:477) report:

The discriminate analysis made it clear that there were major differences in the composition of the two functions for in-hospital and community assaultiveness. The equation predicting hospital assaultiveness contained four significant variables: (1) nonwhites were more likely to be violent than whites; (2) age at first mental hospitalization (the younger, the more likely to be violent); (3) history of alcohol problems before the index hospitalization (those with such a history were less violent); and (4) history of adjudication as a juvenile (the fewer adjudications, the more likely to be violent).

The equation predicting community assaultiveness included two variables: (1) number of prior arrests for violent crimes and (2)

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

age at first mental hospitalization (again, the younger the more likely to be violent); however race and alcohol problems were not found to be statistically important. In both equations, there was no significant discrimination associated with other social, demographic, mental hospitalization, or criminal history variables.

Three of Steadman and Morrissey's findings stand out. One is that the models did not predict well. In the construction sample, the prediction of hospital assaultiveness improved 15 percent over chance (39 to 54%) and the prediction of community assaultiveness improved 32 percent over chance (18 to 50%). In the first validation sample the improvements were 5 and -21 percent; in the second, 3 and -19 percent.

Second, "no single equation can predict both types of behavior and … one must specify in advance the setting for which the prediction is offered" (Steadman and Morrisey, 1982:477). Finally, "… whether or not the person was assaultive while hospitalized had no predictive value for determining his subsequent assaultiveness in the community following release" (Steadman and Morrisey, 1982:483).

Cocozza and Steadman (1974) Cocozza and Steadman examined the postrelease behavior of 98 Baxter patients, reported to be "a group of middle-aged people who had been continuously institutionalized in hospitals for the criminally insane for an average of 14 years." They defined violent recidivism as "acts involving violent assaultiveness against persons" based on information in arrest and hospitalization reports. Each source, arrests and hospitalizations, accounted for about half the instances.

Cocozza and Steadman found that violence could be predicted using two factors: the Legal Dangerousness Scale (LDS; a composite of juvenile record, previous arrests, previous convictions for violent crimes, and the severity of the offense that resulted in confinement at Baxter) and age. Results are reproduced and presented in Table 2.

Shrinkage should not be great with this model because the prediction instrument was developed by using a different criterion variable than that used in the validation. The original prediction instrument (the LDS scale) was constructed by using all arrests as the criterion variable, and development of the original instrument did not take into account violent acts that resulted in hospitalization.

Cocozza and Steadman make several points. One is that violence is a rare event even among these patients who were institutionalized

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

TABLE 2 Dangerous Behavior of Released Patients by Combined Measures of Age and Legal Dangerousness Scale (LDS) Score (N = 98)a

 

Dangerous Behavior

No Dangerous Behavior

Combined Measures

N

Percent

N

Percent

Less than 50 years old and LDS score of 5 or more

11

30.6

25

69.4

50 years or older and/or LDS score of less than 5

3

4.8

59

95.2

a p < .001.

because of violence. A second is that the probability of a false positive is .70. The third is that this release cohort is older, and thus less likely to commit violent acts, than are other typical groups of hospitalized offenders.

Belenko, Chin, and Fagan (1989) This study used a discriminate analysis to predict a rearrest for violent crimes among 3,139 defendants who were arrested for crimes related to "crack" cocaine between August 1986 and October 1986, and among 3,204 defendants who were arrested for crimes related to cocaine between 1983 and 1984. The follow-up period was two years.

Belenko and colleagues were most interested in determining whether offenders who were involved with crack cocaine were more violent than offenders who were involved with noncrack cocaine. Their analysis seems to demonstrate that crack-involved offenders are more violent. However, according to the authors, they were unable to determine from these data which of Goldstein's (1989) explanations12 for the relationship between drugs and violence accounted for these patterns.

After eliminating variables that were "highly correlated" and other variables that were not statistically significant, the authors report the following variables in addition to crack involvement as statistically significant: age and prior drug sales arrests (negative); male, prior violent arrests; total prior arrests; black; and Hispanic (all positive).

The authors correctly identified 64 percent of the violent recidivists. The false positive rate was 37 percent. True and false negatives are not reported.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Virkkunen and Colleagues (1989) The Virkkunen et al. (1989a) sample consisted of 58 Finnish men, 36 of whom had attempted or committed manslaughter and 22 of whom had committed arson. The researchers classified these crimes as impulsive or nonimpulsive (24 of the manslaughter crimes and all of the arson crimes were classified as impulsive). Data about each subject included age, Wechsler IQ, DSM-III scores, and abuse of alcohol. The researchers were especially interested in the predictive power of the concentration of 5-hydroxyindoleacetic acid (5-HIAA) in the cerebrospinal fluid (CSF) and of a low blood glucose nadir.13 These variables were measured prior to the subject's release.

The follow-up period lasted 35.6 ± 18.0 months. (No explanation was provided for the variable length of the follow-up.) The Finnish Criminal Register was used to identify crimes during this period. Based on police reports, court documents, and hospital records, recidivism was defined as a new violent offense or arson.

A stepwise linear discriminate analysis was used to distinguish between recidivists (13 of 58) and nonrecidivists. The blood glucose nadir entered the model first, followed by the CSF 5-HIAA. (The researchers found no correlation between repeated violence and age or IQ, nor did these variables enter the analysis.) Using the model including both physiological variables, the group predicted to be violent had six recidivists and two nonrecidivists. The group predicted to be nonviolent had 42 nonrecidivists and 10 recidivists. The authors do not report their model's accuracy when the discriminate function is used to predict 13 recidivists (rather than the total number observed in the sample).

Virkkunen et al. (1989b:603) conclude that "the psychobiological variables as such or in combination with the behavioral variables had more predictive power for the outcome than any combination of behavioral variables." Also, consistent with the low blood glucose nadir, the researchers observed that without exception these offenders committed their crimes while under the influence of alcohol.

Black and Spinks (1985) Black and Spinks analyzed recidivism for assault during a five-year follow-up period for 125 men who were discharged into the community from Broadmore hospital (England) between 1960 and 1965. Broadmore is one of five hospitals in Great Britain for mentally disordered offenders. "In general, offender-patients have committed the kinds of offenses from which it is deemed the public need protection, and the hospital order reflects the acceptance by the court that there is some psychiatric

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

TABLE 3 Recidivism for Assault During Five-year Follow-Up

Predicted Probability of Failure

Success (number)

Failure (number)

Failure Rate (percent)

0 to .3

105

4

3.7

.3 to .5

6

3

33.3

.5 to .7

1

1

50.0

.7 to 1.0

0

5

100.0

 

SOURCE: Black and Spinks (1985).

disorder that needs treatment. Hospital orders are in effect 'indeterminate sentences,' in that the patients remain in hospital until they are thought to be fit for discharge'' (Black and Spinks, 1985:177).

The dependent variable was any occurrence of an assault, as known to the hospital's research unit. The researchers used a stepwise OLS regression equation to select variables for inclusion in a final logistic regression. The regressions are not reported, but the researchers indicated that assaults could be predicted from the type of offense, age at discharge, and the MMPI scales F (a measure of neurosis) and Ex (a measure of extroversion).

Table 3 summarizes results. The instrument's predictive power appears to be considerable. However, the model was not applied to a validation sample.

Howell and Pugliesi (1988) Howell and Pugliesi analyzed self-reports of violence against a spouse by a national stratified random sample of married and cohabitating men. In the total sample, 19 percent of the men admitted "either minor or severe violence in the past year against their spouse." The authors analyzed the cases of 763 employed men and, separately, the cases of 960 men, some of whom were employed and the rest of whom were unemployed.

Variables used in this study were occupational group (blue or white collar), self-reported economic strain, age (39 or younger versus 40 or older), parental modeling (no violent model and violent model), and employment status.

Log-linear contingency tables were used to analyze these data. After conducting the analysis on employed men only, Howell and Pugliesi (1988:23) report that "having a blue collar as opposed to a white collar increases the odds of reporting spousal violence by a factor of 1.61. … Younger males (under 40) are almost three

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

times (2.84) more likely to report instances of violence than are older men … and those who did not observe parental violence are 0.57 times as likely as those who have observed parental violence to engage in spousal violence." After including the unemployed in the analysis, the researchers report that "for those in the younger age group, being unemployed increases the odds of reporting violence by a factor of 18.61. For those in the older age group, being unemployed increases the odds of reporting violence by a factor of 2.95."

Moss, Johnson, and Hosford (1984) The researchers tested the ability of the Megargee inmate classification system (previously described in the section on classification) to predict institutional violence and violent recidivism among 96 black inmates of a federal prison in 1973. Institutional violence was limited to "participation in a major riot that occurred in 1973." Recidivism appears to have been recorded for a 10-year period, during which "44 were found to have been arrested for violent crimes and 26 for nonviolent crimes." The researchers found that the Megargee classification could not distinguish between recidivists arrested for violent and for nonviolent crimes.

As noted earlier, Megargee's inmate classification system was developed to classify inmates for management purposes, not to predict violence. It is not surprising that in this study, violence—whether measured in the institution or in the community—was not correlated with the inmate's classification. However, the outcome variables used in the study are imperfect measures of violence; riot participation is not a typical form of institutional violence, and the most active violent offenders are also those most likely to commit nonviolent crimes. Therefore, the findings do not definitively allow us to answer the question of whether the inmate classification scheme predicted violent behavior.

Moss and his colleagues cited four other unpublished studies that tested the inmate classification system's ability to predict violence within prison. According to Moss et al. (1984:227), these studies found "the system to be ineffective in predicting institutional adjustment both within a medium security setting and a penitentiary."

It is worth noting in this context that others have found that the inmate classification system does not seem to predict violence. In a review, Kennedy (1986:172) concludes:

Findings of more recent studies … designed to test the efficacy of the Megargee topology as a predictor of inmate adjustments

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

and violence question its use for this purpose, however. Baum (1981) divided subjects into violence-prone and nonviolence prone groups based on their Megargee protocols. It was found that the violence-prone topology group did not commit proportionately more violent acts during confinement than did the nonviolence-prone group. It appeared that an inmate's past violent behavior (based on conviction offense) was a better predictor of violent behavior in prison than the Megargee topology. Similarly, Louscher et al. (1983) found that the Megargee topology groups were not effective in predicting which chronic and high-risk maximum security inmates … would be antisocial or aggressive during incarceration.

Cook and Nagin (1979) Using data from the Prosecutor's Management Information System (PROMIS) in Washington, D.C., Cook and Nagin constructed an analysis file consisting of 4,154 individuals who had been arrested for a crime of violence (murder, rape, assault, or robbery) during 1973. Also, using the prosecutor's information system, the researchers recorded rearrests for each of these individuals during a three-year follow-up. The researchers were unable to determine which of the individuals were incarcerated during the follow-up period.

Probit analysis was used to estimate the probability of being rearrested for a crime of violence. Four regressions were estimated, one for each of the four types of charges (murder, rape, robbery, and assault) at the initial arrest. The explanatory variables were type of weapon (gun, other, and none); prior arrests for violent crimes (none, 1, 2-3, and 4 or more); age (20 or younger, 21 through 29, and older); and some control variables intended to mitigate the problems of missing information about time in prison during the follow-up period.

For robbery and assault at the initial arrest, findings were statistically significant that rearrest for a violent crime increased with the number of prior arrests for crimes of violence and decreased with age. The prior arrest effect was statistically significant for rape; age mattered for neither rape14 nor murder; the prior arrest effect was not significant for murder. The weapon type did not help to predict recidivism in any of the four regressions.

When murder was the first arrest, the offender had a .14 probability of an arrest during the three-year follow-up. Comparable figures for the other initial arrest types were .23 for assault, .29 for rape, .38 for robbery, and—for purposes of comparison with nonviolent initial offenses—.26 for burglary and .19 for weapon possession. Cook and Nagin indicate that these three-year probabilities

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

understate the rates at which these offenders would have been arrested for violent crimes had incarcerations during the follow-up period been known and factored into the calculations.

Holland, Holt, and Beckett (1982) Holland et al. (1982) recorded the official police records of 198 adult male offenders who were released on probation. The follow-up period lasted 32 months. The dependent variable was success (no arrests); nonviolent failure (crimes against property, public order, and technical violations of supervision); and violent failure (armed robbery, aggravated assault, forcible rape, and homicide). Independent variables were the number of prior convictions for crimes of violence, the number of prior convictions for other crimes, and age.

The authors report that rearrest for a crime of violence could be predicted from prior arrests for nonviolent crimes and from the offender's age. Prior arrests for crimes of violence were not useful in predicting violence.

Widom (1989) Widom (1989b) developed a data file of 908 subjects who had been physically or sexually abused or neglected as children and 667 matched subjects who had not been abused or neglected. Abused and neglected children were identified through juvenile court and adult criminal court records. Matched subject records were identified through birth records and school records.

Physical abuse was defined as "knowingly and willfully inflicted unnecessarily severe corporal punishment" or "unnecessary physical suffering." Sexual abuse was inferred from the nature of the charge against the parent(s). Neglect meant that the child had to have "no proper parent care or guardianship, to be destitute, homeless, or to be living in a physically dangerous environment" (p. 244).

Among other dependent variables, Widom included for each subject the number of arrests for a crime of violence as recorded in juvenile court records, probation department records, and adult criminal history records. The follow-up period for collecting these data varied; about 10 percent of the subjects were under 20 at the time of follow-up and about 5 percent were over 30.

A logistic regression showed that blacks were more violent than whites, that those who had been physically abused as children were more violent than those who had not been abused physically, and that those who had been neglected as children were more violent than those who had not been neglected. Sexual abuse did not seem to predict violent behavior (Z score of 1.20). The probability

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

of violence increased with age, but as Widom recognized, this effect might be attributed to the potentially longer follow-up of older subjects.

Widom did not present measures of association for her logistic regressions.15 However, as she notes, only 11 percent of those children who were abused or neglected had a subsequent violent criminal record. Furthermore, the effect of being abused or neglected does not seem to be large. About 8 percent of those who had not been abused or neglected were also violent; thus, improvement over chance prediction based on knowledge of childhood abuse alone is negligible.

Failure Models

Failure-time models typically attempt to explain the length of time from some point M until the occurrence of a violent crime. When M corresponds to the occurrence of the nth violent crime, the analyst attempts to explain the time between the nth and (n+1)st violent crimes. When M corresponds to the occurrence of any crime, the analysis attempts to explain the time from that crime to a violent crime. When M corresponds to release from prison, the analyst attempts to explain the time until a violent crime once the offender is no longer restrained. When M corresponds to birth, the analyst addresses the onset of a violent criminal career.

Whatever the criterion variable, the model is typically:

where λ is a parameter (such as the mean) of a specific distribution, ß is a row vector of parameters, and X is a column vector of explanatory variables. The problem is to estimate ß and any other parameters that may be required given the distributional assumptions.

In the violence-prediction literature to date, failure-time models have been used to estimate single transitions, that is, the occurrence of the first violent crime after M. Methods are available for jointly estimating the timing between multiple violent acts (see Heckman and Singer, 1985; Flinn, 1986), but our search did not reveal any studies that used such panel techniques when predicting violence.

Rhodes (1989) Rhodes predicted violent behavior among a random

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

sample of 1,711 offenders who were released from federal prisons during 1979 and whose subsequent recidivism (rearrest or parole revocation) was monitored and recorded for five years. The Rhodes' model is a split-population (see Maltz, 1984; Schmidt and Witte, 1988), competing events survival model with an adjustment for selection bias.

Time until recidivism for a violent event is modeled as a latent variable distributed as log-normal, a form that is presented by Kalbfleish and Prentice (1980) and has been used by several researchers to analyze criminal recidivism (Rhodes, 1986; Schmidt and Witte, 1988). The competing events assumption permits right-hand censoring to occur for two reasons: the end of the follow-up period or incarceration for a nonviolent crime prior to any occurrence of a violent crime. In the latter case, violent recidivism is precluded by the offender's imprisonment for a nonviolent crime.

The model's correction for selection bias is particular to the nature of the data in this study. The data, based on a random sample of offenders released from prison during 1979, are intended conceptually to represent offenders entering federal prisons. However, most convicted federal offenders (roughly 70 percent) are sentenced to probation, so a study of violent recidivism based on a prison release cohort might provide a distorted representation of recidivism among federal offenders in general. Rhodes used maximum-likelihood procedures suggested by Heckman (1979) to adjust for the selection bias.

This review focuses on the portion of the study that used violent crime (robbery or assault) as the criterion variable. Rhodes reported that blacks recidivated sooner that whites, men sooner than women, and young offenders sooner than old offenders. Offenders who had prior jail or prison commitments recidivated sooner than offenders with prior convictions but no incarcerations, and they in turn recidivated earlier than offenders with no prior convictions. Offenders who had a known heroin/opiate dependence16 recidivated sooner than those who did not. There was some tendency toward specialization: offenders who had been convicted of a violent crime or robbery prior to being released from prison were somewhat more likely to recidivate for robbery or assault than for some other crime.

Rhodes did not report the predictive accuracy of his regressions using violence as the criterion variable. However, the statistics presented permit determining his predictions' ability to distinguish between those who will and those who will not be arrested for violent crimes. Within a five-year follow-up period,

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

the average offender (a hypothetical offender who has the mean value for all predictor variables) has a .03 probability of being arrested for a violent crime. Consider an offender who appears to be unlikely to recidivate: Caucasian, female, age 45, no prior convictions other than her federal conviction for selling drugs, not addicted to opiates, and employed at the time of her federal conviction. Her probability of being arrested for a violent crime during the five years is almost zero. Consider an offender who appears likely to recidivate: black, male, age 20, two incarcerations prior to his current federal conviction for robbery, four prior convictions not resulting in incarceration, drug dependence, under supervision at the time of this federal conviction, and unemployed at the time of this federal conviction. His probability of being arrested for a violent crime within five years is about .50.

These statistics seem to indicate that violent recidivists can be distinguished from nonviolent recidivists (and from nonrecidivists in general) with at least modest accuracy. However, the illustration exaggerates the strength of this prediction instrument because the hypothetical offenders used in the illustration are extremes.

Rhodes also ran separate regressions in which an arrest for robbery and an arrest for assault were the dependent variables (unpublished). These clarified the above results by showing there was a slight tendency for offenders who had been incarcerated for assault to commit an assault in the future, and for offenders who had been incarcerated for robbery to commit a robbery in the future.

Weiner (no date) Weiner used a failure-time model to assess violent recidivism among 1,355 boys in Philadelphia who had at least one prior police contact for a crime of violence. Weiner's data comprised three overlapping samples of boys, all of whom were born in 1958, resided continuously in Philadelphia from their tenth to their eighteenth birthdays, and had police contacts between their tenth and eighteenth birthdays. The first sample (very violent) had 1,084 observations; the second (violent) had 1,323; and the third (participated in illegal acts) had 1,100. The total sample size was 1,355.

"Very violent" delinquents were those youths who participated in at least one UCR violent index offense: criminal homicide, forcible rape, robbery, or aggravated assault. "Violent" delinquents met the criterion for a very violent delinquent or had a UCR violent nonindex offense (these include simple assault and

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

sex offenses other than forcible rape). The third sample consisted of youths who had "participated in illegal acts in which at least one victim sustained an injury, ranging in seriousness from minor harm or treatment and discharge by a physician to hospitalization or death." Despite the different definitions, these samples were very similar, and Weiner's analysis was robust with respect to the three definitions of violence.

Data were drawn from rap sheets prepared by the Juvenile Aid Division of the Philadelphia police, police investigation reports, police arrest reports, family court records, and the census.

Weiner fit his regressions using three different failure-time distributions: Weibull, log-normal, and log-logistic. He reports that the distributional assumptions made little difference, but that the Weibull distribution provided a somewhat better fit than the other two.

Weiner used a stratified stepwise procedure to estimate his models. Twenty-one personal and delinquency history factors were grouped subjectively into four sets of decreasingly "acceptable" categories for making predictions. (Race, for example, was a member of the least acceptable category.) In the analysis, all the variables in the most acceptable category were considered for inclusion in the statistical model prior to stepping through the variables in the second group and so on.

This estimating procedure was applied separately to different violent "transitions." That is, Weiner initially examined the time from the first violent event to the second violent event, where the second event was considered censored when only one violent event had been observed prior to the subject's eighteenth birthday. Then he examined the second transition, that is, the time from the second violent event to the third, and so on. Results were reported separately for these different transitions.

Weiner found that few factors were useful in predicting recidivism. He reports that the instantaneous failure rate (the recidivism hazard) seemed to decrease over time for blacks, but not for whites. Given the unmeasured heterogeneity in these data, however, this decrease may be an artifact (see Heckman and Singer, 1985), as Weiner (no date:216) was aware.

Weiner (no date:207, 208) reports that blacks are more likely to recidivate than whites; however, few other patterns seemed to emerge:

That several risk factors consistently failed to achieve statistical and substantive significance also bears noting: Neither the type nor gravity of a youngster's present offense, nor a selected offense

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

aggravation component, the presence of a weapon, was significant. Nor did several indices of the types and severities of the justice system response to the youngster achieve significance, such as the presence of prior UCR index offenses for which the youngster was adjudicated delinquent, the total seriousness of those offenses, and the presence of prior commitments to secure and nonsecure facilities for UCR index offenses. The extent and gravity of a youth's serious delinquent history was similarly nonsignificant. Age also failed to exhibit a relationship to recidivistic timing, including the ages at the first and last prior delinquent incident.

Consistent, then, with other research, prediction of dangerous conduct, whether assaultive or threatening, remains elusive.

A major problem with this analysis is its handling of incarceration that occurs after the first violent crime in a transition. Weiner was unable to collect data about the length of time that an offender was confined during the follow-up period, so the variable "time until recidivism for a violent crime" was likely to be measured with considerable error. Weiner (no date: 318-325) was aware of this problem and attempted to control for its effect. Although he did not know the length of time that an offender had been confined, he did know whether the youth had been confined for some period of time. Consequently, Weiner entered a dichotomous variable into his regressions whenever he had evidence that incarceration had occurred sometime during the follow-up.

It is not clear what type of a statistical model can be developed to justify the introduction of the variable "occurrence of incarceration" as an independent variable. After all, as is true of violent recidivism, a future period of incarceration is an endogenous variable. Using this endogenous variable as an explanatory variable would seem to produce biased parameter estimates. Nevertheless, it is unlikely that even this problem could have hidden a strong correlation between violence and a subject's background. Based on Weiner's analysis, we can say with some confidence that within this set of data, the ability to predict the timing of repeated violence is at best low.

Rate of Violence Models

Some researchers have attempted to predict violence by using as a dependent variable the number of violent events per time at risk (i.e., time not incarcerated) or some variant (e.g., number of events weighted by some severity index). Such studies are reviewed in this section.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Dembo and Colleagues (1991) Dembo et al. (1991) report on violent behavior during a one-year follow-up of boys who had been detained in a regional detention center in Tampa, Florida. Initial interviews were conducted with 399 boys. Of these, 305 were reinterviewed, but only 201 had been at risk of recidivism prior to the reinterview (they were either in the community at the time of the reinterview or recently detained). These 201 youths were the focus of this study.

Dembo and colleagues constructed an index of violent recidivism, which was a composite score of the following behaviors:

  • Carried a hidden weapon other than a plain pocketknife

  • Attacked someone with the idea of seriously hurting or killing him

  • Had been involved in gang fights

  • Hit (or threatened to hit) a teacher or other adult at school

  • Hit (or threatened to hit) parents

  • Hit (or threatened to hit) other students

  • Had (or tried to have) sexual relations with someone against that person's will

  • Used force (strong-arm methods) to get money or things from other students

  • Used force (strong-arm methods) to get money or things from a teacher or other adult at school

  • Used force (strong-arm methods) to get money or things from other people (not students or teachers)

Crimes against persons, as so defined, occurred frequently. At the time of the initial interview, only 24 percent of boys and girls reported no violent acts; 38 percent reported 1-4 acts; and another 25 percent reported 5-29 acts. At follow-up, violence was somewhat less frequent: 46 percent reported no violent acts; 27 percent reported 1-4 violent acts, and 20 percent reported 5-25 violent acts. Although the particulars are not reported for the composite score, it is apparent that the least harmful violent acts constitute most of this measure.

Youths were at risk for varying lengths of time during the follow-up period. Consequently, the authors controlled for time at risk by dividing the reported acts of violence by time not incarcerated, and the rates were then annualized. Finally, a log transform was applied, with the rate zero transformed to -1.

The dependent variable was regressed against several independent variables: age, race, gender, occupational status of the household

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

head, recency and number of days in the previous month used alcohol, lifetime reported use of cocaine, an index for physical abuse, lifetime reported use of marijuana/hashish, an index of sexual victimization, urine test for cannabinoids, urine test for cocaine, emotional/psychological index, number of previous referrals for crimes against persons, placement in detention for crimes against persons, and self-reported crimes against persons. All variables were measured at the time of the initial interview and refer to the period just before that interview.

Only three variables were statistically significant. Violence decreased with age, increased with the abuse of alcohol, and increased with the number of self-reported crimes against persons prior to the follow-up period. Notably, violence was not predicted by the following: physical abuse, sexual abuse, drug test results, and official record of violence.

The regression predicted 23 percent of the variance in the dependent variable. Given that there were 15 independent variables and only 201 observations, significant shrinkage is likely.

Menzies, Webster, and Sepejak (1985) These researchers report on a study of recidivism among 211 patients interviewed in a pretrial forensic clinic. Two raters, who observed interviews at the forensic clinic, scored each of the subjects using the researches' Dangerous Behavior Rating Scheme (DBRS). Subjects were unaware of the raters.

Fifteen items of the DBRS were used to predict dangerousness. In addition, these fifteen items were factor analyzed, yielding four factors with eigenvalues greater than 1.0 that collectively accounted for more than 72 percent of the variance in the independent variables.

To construct a dependent variable, ''for each patient a profile was constructed cataloging all (not only violent) officially recorded transactions during the two year follow-up . … Subsequently, these outcome danger profiles were quantified by rating the behaviors of subjects on an 11-point scale . … For this purpose, nine independent judges were used to assign 'danger outcome scores' to each patient" (Menzies et al., 1985:59-61). The scores were averaged across the judges, and subjects with no official incidents were assigned a score of zero.

Menzies et al. (1985:61) report that "the aggregate index of all four factors produces the most accurate prediction of dangerous behavior . … This complex weighing of 15 items still only accounts for 12 percent of the variance in followup violence."

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Individual items with the highest Pearson product moment correlation were, in the following order, tolerance, capacity for empathy, capacity for change, and hostility. The authors do not define these terms in the text.

A serious problem is that the criterion variable was a composite of violent acts that were committed by people who, during the follow-up period, spent different times in prison, in hospitals, and on the street.17 With such a criterion variable, it is impossible to disentangle the propensity toward violent behavior from adjustments to different institutional settings. Compounding this problem is the fact that selection into a prison setting, a hospital setting, or the community is not randomly applied to the subjects.

Farrington (1989) There appear to be few studies of early predictors of violence (Loeber and Stouthamer-Loeber, 1987). One exception is a study by Farrington of 411 London males whose criminal careers were followed for 24 years, from age 8 until age 32. The sample consisted of all boys in primary schools within a mile radius of the research office. Multiple interviews were conducted: The boys were interviewed when they were ages 8, 10, 14, 18, 21, 25, and 32; the boys' parents were interviewed yearly from the time the boys were age 8 until they were age 14-15; teachers completed questionnaires about the boys when the boys were 10, 12, and 14. The boys' official criminal records were collected throughout the 24 years.

Response rates were high. By the time the subjects were 32, interviews or questionnaires with either the subject or a proxy (such as a spouse) were completed for 94 percent of the sample.

Farrington predicted violence at four distinct points in time. The definition of violence differed at each time. "Adolescent aggression at age 12-14" was a scale developed from teachers' responses characterizing the boy as "disobedient, difficult to discipline, unduly rough during playtime, quarrelsome and aggressive, overcompetitive with other children, and unduly resentful of criticism or punishment." Teachers identified 134 boys as aggressive. "Teenage violence at age 16-18" was a scale "derived from self-reports … of getting into fights, starting fights, carrying and using weapons, and fighting police officers." The scale identified 119 violent males. ''Adult violence at age 32" was defined as "admitting having gotten into a fight in the last five years." By this definition, 140 males were so identified. "Convictions for

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

violence between ages 10 and 32" was the last category of violence. Fifty males had been convicted.

The predictor and criterion variables were not used in their natural scale. Rather, according to Farrington (1989:84), "each predictor and criterion variable was dichotomized, as far as possible, into the 'worst' quarter versus the remaining 3/4 of the sample." The author claims that this procedure did not lose much information, while statistics became easier to interpret.

Farrington presents both bivariate results and the forward stepwise discriminate analysis discussed here. The results demonstrate that variables describing a child's formative years can be used to predict violence during that child's youth and adult years. Some predictors would be recognized by lay persons as signs that a child has not adjusted well to school—poor attendance, behavioral problems, low academic ability, low attainment, and early dropout. Other predictors include general tendencies toward delinquent activity (hostility toward police, delinquent activities, and delinquent friends); the child's family situation (low income, authoritarian parents, and detached parents); observations of "high daring," "high aggressiveness," or "nervousness and withdrawal'' by teachers and others; and the child's size and intelligence. In general, the particular variables that are pertinent vary at different times in the child's life.

Malamuth (1986) Malamuth reported on a self-report study of sexual aggression among 155 male volunteers (80 percent were college students.) The volunteers were asked to respond to a scale of sexual aggression developed by Koss and Oros (1982). Malamuth (1986:956) does not provide details of this scale in his study but states, "It assesses a continuum of sexual aggression including psychological pressure, physical coercion, attempted rape, and rape."

As explanatory variables, Malamuth uses the following: a measure of dominance as a sexual motive, a scale of hostility toward women, a scale that measured attitudes facilitating violence (Acceptance of Interpersonal Violence), a measure of antisocial characteristics, and sexual experience. A measure of sexual arousal in response to rape and to mutually consenting depictions was assessed for a subset of 95 volunteers by using penile tumescence as a gauge of arousal.

Using multiple linear regression to analyze sexual aggression among the 155 volunteers produced an r2 of .30. All variables except the antisocial characteristics variable were significant at

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

.06. Using all-subsets regression methods, Malamuth added three interaction terms to his model. The r2 increased to .453, and the variable dominance as sexual motive was no longer significant. Covariance tests indicated that sexual arousal in response to rape increased the model's explanatory power.

A related study by Ageton (1983), as reviewed by Malamuth (1988:475), found the following:

[Ageton] … conducted a longitudinal study to gauge the extent to which a variety of measures including those reflecting general delinquency, attitudes about rape, and sex-role stereotyping predicted sexual aggression. Subjects, 11 to 17 years old drawn from a representative national sample, were interviewed in several consecutive years during the late 1970s. Ageton found that the general delinquency factors, such as peer support for antisocial behavior, were highly predictive of sexual assault. Attitudes regarding rape also enabled some discrimination between sexual aggressors vs nonaggressors, but sex-role stereotyping did not. She suggested that the same set of factors explains sexual assault and other delinquent behaviors.

Furby, Weinrott, and Blackshaw (1989) We reviewed no studies predicting recidivism among sex offenders. The following conclusions are from a comprehensive review of 42 selected studies by Furby et al. (1989:22, 25, 27):

  • With such variability among study results, it is difficult to make any meaningful statement about the number of sex offenders who continue to commit sex offenses.

  • We can at least say with confidence that no evidence exists that treatment effectively reduces sex offense recidivism.

  • There is some evidence that recidivism rates may be different for different types of offenders. Those trends must be viewed as only tentative conclusions, which are based on patterns we identified across varied (and sometimes few) studies, each with its own flaws.

Garrison (1984) Among the studies that we examined, one by Garrison was unique in that it examined the events that precipitated violence. Garrison examined violent behavior among 100 males, aged 7-15, who were being treated in a psychiatric treatment facility. Over a two-year period, these children engaged in 1,038 incidents of observed, interpersonal aggression during the five hours per week that they were expected to participate in educational and social activities. Violence was defined as "intense

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

physical attacks on other persons." Participant observers also recorded the timing, location, antecedents, choice of victim (staff or patient), objects used, immediate consequences, and staff observations regarding the presence or absence of external provocation. Antecedents were "external provocation, no external provocation, and unable to report." Victim status was "staff, peer, and combination."

Logistic regression was used to analyze the data. From this limited variable set, the author reports statistical significance among the following variables: antecedent, age group, and victim status. It is unclear why victim status was included as an independent variable, because this would seem to be a variable of choice. Garrison (1984:233) reports that the model's fit was "rather low" and that "a prediction table which employed these 1038 incidents to assess the utility of the model showed overall prediction to be only marginally better than that expected by chance."

ACCURACY OF PREDICTION

Although we could apply formal statistics, such as RIOC (Farrington and Loeber, 1989) to summarize the predictive power of the regressions that were reported above, there is little reason to do so. It is apparent that violence can be "predicted," but that for even the most recidivistic offenders, violence is less likely to occur than to not occur.

Moreover, RIOC tells only part of the story. The important question is what level of accuracy can be achieved when predicting who will and who will not be violent. However, statistical analyses generally involve officially recorded violence such as arrests and hospitalizations. Typically, we might find that for a designated class of subjects, the probability of violence, as so measured, is .3-.5. Whether such predictions are accurate or inaccurate cannot be determined from calculating a statistic such as RIOC. Instead, when interpreting accuracy, we are required to make many "leaps of faith." How likely is it that an act of violence will lead to an official record of its occurrence? How likely is it that the probability of detection is uniform across a population of subjects? Because we are interested in actual violence, not violence as revealed through officially recorded incidents, answering these questions is inescapable when judging predictive accuracy. At this time, only subjective impressions and some imprecise statistics (Blumstein et al., 1985) provide guidance. Although RIOC

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

and similar measures give an aura of precision and science, such precision may be illusionary.

There are other problems with applying formal measures of predictive accuracy to the above studies. The criterion variable changes markedly across these studies. There are major differences in the definition of violence, in the periods during which violence might occur, in the settings in which violence might occur, in the precision with which violence is measured, in the numbers and types of explanatory variables available to the researcher, and in the availability and size of validation samples. It is not possible to make credible comparisons of statistical measures of association in the face of this diversity.

Still another problem with using measures of association to compare predictions is that any measure of association depends on the population to which it is applied. It would seem that the best way to compare the accuracy of prediction across instruments would be to use a common population (see Cohen and Zimmerman, 1990), but this raises other complications. Suppose that prediction instrument A was developed in a population X that had no variation in variable z. Then z would not appear as a predictor within population X and, of course, would not be used when predicting in population Y, despite the fact that z may in fact distinguish between those who are violent and those who are not.

Further, predictions of violent behavior are confounded with the contamination of public policy. Transporting predictions across settings where policy responses differ is not a good test of the predictors. Indeed, because public responses to violence are never absent, there may be no way to validate an instrument directly (Rhodes, 1985). As we argue above, it is important to distinguish the behavioral elements of violence from the public response elements. Most studies of violence do not allow these distinctions to be made.

SUMMARY OF OUR REVIEW OF THE PREDICTION OF VIOLENCE

  1. Violence can be predicted, meaning that within a given population we can assign different probabilities of violence to population members based on the characteristics of those members. Furthermore, predictors have some roots in theory and have been replicated across studies conducted in diverse settings, leading us to believe that the predictors are manifestations of real behavior rather than pure artifacts of methodological limitations or social

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

responses to violence. Nevertheless, there are significant concerns with the validity, reliability, and accuracy of predictions of violence.

  1. Some variables have consistently been shown to predict violent behavior. Other variables are strong in particular studies and often unmeasured in other studies. Variables that are commonly observed to predict violence, and variables that are predictive although infrequently used, provide guidance for conducting future prediction studies.

    Family background variables appear to be valuable when predicting adult violence based on childhood records; they are of limited use in predicting repeated violence by adults. Criminal record variables are pertinent when predicting violent behavior within five years for offenders released from prison, but are of little use when predicting violence within a few weeks of release from a mental health clinic.

    There is little doubt that criminal records (official or self-reported) are among the best predictors of violence, although there is some dispute about which aspects of criminal records have the greatest explanatory power. There is general agreement that violent offenders do not specialize in violence; nevertheless, past violence seems to be among the best predictors of future violence. There seems to be agreement that a propensity toward violence tends to be revealed prior to adulthood, but not all violent adults are known to be violent as children. There seems to be little doubt that drug involvement has some predictive power, although perhaps only for some forms of violence; the same may be true of alcohol consumption. Age, sex, and race are generally found to be predictive. Variables related to a child's development predict violence as a juvenile and as an adult. Variables such as family support can be useful when predicting violence. Although the nexus between violence and mental illness is nebulous, and it appears that health service professionals overpredict violence during clinical assessments, a history of hospitalizations has some predictive power as do some psychological scales that indicate a propensity toward violence. Violence during a period of hospitalization does not necessarily predict violence postrelease, however. The same is probably true of violence during periods of incarceration. Although the evidence from naturalistic settings is limited, biosocial variables seem to have strong predictive power.

  2. Predictions are fairly accurate when violence during the short term is being predicted for a population whose past behaviors and current attitudes are well understood (e.g., people who

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

are admitted to a mental health clinic for recent violent behavior). Predictions may be tolerably good over a longer term among a heterogeneous population whose past behaviors and current attitudes are less well understood, provided we focus on predicting for a subset of that population (e.g., young bank robbers who have lengthy criminal records and a history of drug abuse). Predictions may be remarkably good when violent behavior is defined broadly, measured over a long period of time, and predicted from variables that reach deep into the subject's past (e.g., violence during the adult years, with early child development and delinquency used as predictors).

  1. It is easier to predict repeat violence than to predict a subject's first violent act, because the best predictor of future violent behavior is past violent behavior. Given that most violent acts are not recorded in official records, we may never be able to develop predictions that do not suffer from a high proportion of false negatives.

RESEARCH AGENDA

One of our original goals for this review was to compare findings across studies and report the relative accuracy and strength of different types of factors for classification and prediction. However, the existing literature displayed so much diversity in regard to populations studied, techniques used in analysis, and measures used for validity and reliability that this comparison was not possible by using published data. Even if independent secondary data analysts could obtain the original source data for a large number of the studies, they would still be challenged to draw meaningful comparisons.

Nonetheless, it would be desirable to compare the relative strengths and reliability of different variables for classifying and predicting violence. We urge that a single study or a small number of studies be commissioned to collect data on a wide variety of types of independent variables discussed in this paper. Because biological variables appear very promising in regard to their accuracy in classification and prediction, we strongly suggest that the recommended comprehensive study or studies involve an interdisciplinary approach, including batteries of medical tests such as those used for classification on the physical disorder axis of DSM-III-R.

Secondarily, based on this review we would urge that situational variables be collected, quantified, and analyzed. Researchers

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

have repeatedly found that classification and prediction models are not stable when applied in environments other than those in which they were developed. However, by learning which environmental factors most effect classification and prediction and by controlling for them, it may be possible to apply the same instruments across different situations. Moreover, and perhaps more important, if the situational factors that increased or decreased the probability of violent recidivism for specific types of violent persons were known, it would be possible to provide research-based policy recommendations for placement or treatment of violent persons. For example, if violent persons with certain types of brain dysfunctions were found to be more likely to recidivate in situations with a high level of sensory stimulation, a recommended treatment would be placement in a rural rather than an urban environment.

Although it may not be possible to compare previously published models across study populations, we do feel that considerable clarification of past findings could be achieved by secondary analysis that goes back to the raw data and defines comparable variables. More specifically, the definitions of outcome variables need to be standardized along the following dimensions:

  • The length of time at risk (adjusting for social restraint)

  • Specification of the intended period of prediction (short or long term)

  • The extent of physical harm or potential harm associated with violent acts, if uninterrupted (e.g., distinguishing between an open-handed slap to an adult and bludgeoning with a blunt weapon)

  • Specification of a relationship between the measure of violence and the actual violent behavior partially captured by the measure (e.g., adjusting for probabilities of police intervention, arrest, conviction, or self-initiated hospitalization).

Just as studies would be more comparable were the independent and criterion variables measured similarly, they would also be more comparable were statistical tools more uniformly applied. Failure-time models, occurrence models, and rate models—as defined above—are all appropriate for predicting violence, but "what" is being predicted differs so much that estimates based on one technique are difficult to compare with estimates based on another. Throughout this paper, we have emphasized the need for modeling to disentangle violent behavior from social restraints on violent behavior. We feel especially strongly that researchers should

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

attempt to overcome censoring, selection bias, and variable length follow-up periods.

Finally, the field needs additional research on appropriate measures for statistically comparing the accuracy of predictive models. Although techniques described above have been developed for comparing the accuracy of models that divide individuals into groups (these may be classifications, or they may be predictions with cutoffs for "predicted violent" and "predicted nonviolent"), many researchers, including the authors, use these types of measures when the outcome variable being predicted is actually continuous. For example, use of the RIOC statistic requires turning a probabilistic prediction into a binary prediction by arbitrarily establishing a cutoff above which an individual is classified as violent. The literature is replete with discussions of the best or most appropriate way of establishing the cutoff, but obviously a great deal of information about the accuracy of the prediction model is being lost by summarizing the output set of probabilities into two categories.

Developing measures for comparing predictions that result in probabilistic statements is not easy for reasons explained above. Probabilistic predictions can be absolutely correct, but useless, if they are in a practical sense close to the base rate. Nonetheless, this appears to be a promising area for research.

POLICY IMPLICATIONS

Much of the research reported in this paper has been carried out in response to the need to inform policy. Practitioners need relatively quick, simple ways to decide whether a person is violent and, if so, whether treatment or special supervision is necessary. Ironically, there is a mismatch between what models can do and what we intend for them to do. Much research has been carried out to provide the types of answers we are least capable of accurately giving—unidimensional yes or no categorizations. Perhaps it is time to tell practitioners about our most accurate classification and prediction tools, and help them shape policy based on what can been done, rather than the other way around.

The types of predictions and classifications that can now be carried out most accurately are multidimensional, probabilistic, and short-term . No one can apply any known instrument to any given population and say with confidence whether or not a particular individual will commit a future violent act. However, following Gottfredson and Gottfredson (1988a), it is currently possible

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

to give practitioners instruments to assess the relative risks of a specific person in a specific population being a particular type of violent offender and committing future crimes, but practitioners as the public's representatives must define the stakes—and make decisions based on both the probabilistic risks and the stakes. A benefit-cost analysis is an inescapable aspect of judging the adequacy of violence prediction, but when the benefits and costs are so elusive, the analysis is outside the realm of science and is appropriately in the domain of politics and public policy.

Obviously practitioners should not be expected to look at models with numerous Greek letters and parameters and then apply them correctly, but by refining instruments that help assess violent persons probabilistically along several dimensions, behavioral and medical scientists can help shape practices that consider a variety of options for dealing with violent persons rather than a simple choice of two alternatives. For example, a person with a high probability of violence related to an organic disorder but low probability of violence related to alcoholic drinking, social stress, or other factors may be a good candidate for home placement with appropriate medication and frequent medical tests, whereas a person with a moderate probability of violence related to multiple factors (social, psychological, substance abuse, and post-traumatic stress) may be a better candidate for a treatment program inside a prison setting.

It is important not only to consider risks and stakes, but also to consciously develop a wider range of options for dealing with persons classified as or predicted to be violent. Research on classification and prediction can then be more easily focused on choosing among realistic options, which will help produce results and findings that are more immediately applicable by practitioners.

NOTES

1.  

However, this is, in some sense, also a subgroup.

2.  

The distinction between case management and prediction is illustrated by the work of Megargee et al. (1988). They specifically deny that their classifications of offenders based on Minnesota Multiphasic Personality Inventory test results can be useful for predicting future violence, while at the same time they demonstrate that a reduction in the occurrence of assaults within institutions can be achieved by segregating offenders whose classification is "predatory" from those classified as likely victims of assault.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

3.  

Megargee (1977) includes these three qualities in a list of desirable features of a classification system. Also included in the list are comprehensiveness, cost-effectiveness, and potential to reflect changes over time.

4.  

Dichotomous items are those having only two possible alternative responses.

5.  

When the separation between test and retest is as long as months or years, the comparison of test with retest is more pertinent to the temporal stability of the behavior being measured than to the reliability of the measurement instruments.

6.  

Tay-Sachs disease is a fatal genetic disorder that is not apparent by outward physical signs at birth. If both parents are carriers of the recessive gene, there is a 25 percent chance that a pregnancy will result in a Tay-Sachs baby. The blood level of the enzyme hexosaminidase A is depressed in Tay-Sachs carriers and zero in an affected person. By testing both parents' blood, it is possible to determine if both are carriers; if so, amniocentesis or other means of sampling fetal tissue permits determining with certainty whether the fetus will develop Tay-Sachs disease.

7.  

For an example of prosecuting attorneys' classifying defendants in the context of the local district attorney's office policy, see Chaiken and Chaiken (1987, 1990b).

8.  

There were an additional 15 reports of neglect and 16 reports of failure to thrive, but we do not consider these.

9.  

We would like to thank Joel Garner, then at the National Institute of Justice, for preparing this literature search of titles related to violence.

10.  

Alternatively, the prediction might be made conditional on a specific form of social constraint, such as intensive parole supervision or weekly reporting to a mental health clinic.

11.  

In predicting violence, one is not uninterested in social responses and whether they control or fail to control violent behavior, but it is important to distinguish the behavior itself from the effectiveness of social control. One cannot judge the effectiveness of variations in social control without independent measures of the rate at which violence would have occurred in the absence of social controls. Furthermore, development of the theory of violent behavior is retarded if empirical studies of violence are muddied with the effects of social controls of criminal behavior.

12.  

Goldstein (1989) advances three explanations for the relationship between drugs and violence. Drugs may have a pharmacological effect that precipitates violence; violence may be instrumental as part of property crimes committed to get funds for drug

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

   

purchases; or violence may be a systemic part of the drug distribution business.

13.  

From the context, we infer that the low glucose nadir indicates an abnormality in the physiological processes that keep blood sugar concentration in a normal steady state, for example, a defective system of glycogen conversion to glucose in the liver due to cirrhosis or other disease. In this case, cerebral function would be impaired in turn, because brain cells could not adequately store glucose or utilize other forms of nutrients for energy.

14.  

The signs of the coefficients for rape were consistent with the hypothesis that the probability of an arrest for violence decreased with age. The failure to find the difference to be statistically significant might be explained partly by the small sample size—277 for rape compared to 1,104 for robbery and 2,083 for assault.

15.  

The derivative of the logistic regression, when evaluated at the mean, indicates that subjects who suffered physical abuse as children were .08 more likely than others to commit violent acts, and that subjects who suffered neglect as children were .05 more likely than others to commit violent acts.

16.  

Other drug use was not recorded for this 1979 cohort.

17.  

The distribution of street time was reported as 20 or more months spent in the community (N = 110), 12 or more months in prison (N = 35), 12 or more months in a psychiatric hospital (N = 13), and a mixture of institutional conditions (N = 45).

REFERENCES

Ageton, G. 1983 Sexual Assault Among Adolescents. Lexington, Mass.: Heath.

Altschuler, D., and P. Brounstein 1989 Patterns of Drug Use, Drug Trafficking, and Other Delinquency Among Inner City Adolescents in Washington, D.C. Paper presented at the annual meeting of the American Society of Criminology, Reno, Nevada, November.

American Psychiatric Association 1987 Diagnostic and Statistical Manual of Mental Disorders, Third Edition-Revised. Washington, D.C.: American Psychiatric Association.

Ayoub, C., M.M. Jacewitz, R.G. Gold, and J.S. Milner 1983 Assessment of a program's effectiveness in selecting individuals ''at risk" for problems in parenting. Journal of Clinical Psychology 39:334-339.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Baum, M.S. 1981 Effectiveness of the Megarege typology in predicting violent behavior. Doctoral dissertation, University of California, Santa Barbara. (Dissertation Abstracts International 42/04H:1804. University Microfilms No. DEN81-20741.)

Belenko, S., K. Chin, and J. Fagan 1989 Typologies of Criminal Careers Among Crack Arrestees. Paper presented at the annual meeting of the American Society of Criminology, Reno, Nevada, November.

Black, T., and P. Spinks 1985 Predicting outcomes of mentally disordered and dangerous offenders. In D. Farrington and R. Tarling, eds., Prediction in Criminology. Albany: State University of New York Press.

Butcher, J.N., and L.S. Keller 1984 Objective personality assessment. Pp. 307-331 in G. Goldstein and M. Hersen, eds., Handbook of Psychological Assessment. New York: Pergamon Press.


Chaiken, J., and M. Chaiken 1982 Varieties of Criminal Behavior. Santa Monica, Calif.: The RAND Corporation.

1990a Drug use and predatory crime. Pp. 203-209 in J.Q. Wilson and M. Tonry, eds., Drugs and Crime. Vol. 13 of Crime and Justice: A Review of Research. Chicago: University of Chicago Press.

Chaiken, M. 1983 Crime Rates and Substance Abuse Among Types of Offenders. New York: The Interdisciplinary Research Center, Narcotic and Drug Research, Inc.

1990 Community or Individual Factors: What Matters More for Serious Criminal Behavior and Frequency of Arrest? Final report for grant 88-IJ-CX-0022 submitted to the National Institute of Justice.

Chaiken, M., and J. Chaiken 1987 Selecting "Career Criminals" for Priority Prosecution. NCJ 106310. Washington, D.C.: National Institute of Justice.

1990b Redefining the Career Criminal: Priority Prosecution of High-Rate Dangerous Offenders. NCJ 124136. Washington, D.C.: National Institute of Justice.

Chaiken, M.R., and B.D. Johnson 1988 Characteristics of Different Types of Drug-Involved Offenders . Washington, D.C.: National Institute of Justice.

Clements, C.B. 1986 Offender Needs Assessment. College Park, Md.: American Correctional Association.

Cocozza, J., and H. Steadman 1974 Some refinements in the measurement and prediction of dangerous behavior. American Journal of Psychiatry 131(9):1012-1014.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Cohen, J., and S.E. Zimmerman 1990 Improved Techniques for Assessing the Accuracy of Recidivism Prediction Scales. Pittsburgh, Pa.: School of Urban and Public Affairs, Carnegie-Mellon University.

Collins, J. 1989 Alcohol and interpersonal violence: Less than meets the eye. In N. Weiner and M. Wolfgang, eds., Pathways to Criminal Violence. Newbury Park, Calif.: Sage Publications.

Cook, P., and D. Nagin 1979 Does the Weapon Matter? Washington, D.C.: INSLAW.

Copas, J.B. 1983 Regression prediction and shrinkage. Journal of the Royal Statistical Society 129:311-354.

Copas, J., and R. Tarling 1986 Some methodological issues in making predictions. In J. Roth and C. Visher, eds., Criminal Careers and "Career" Criminals. Washington, D.C.: National Academy Press.

Corcoran, K., and J. Fischer 1987 Measures for Clinical Practice: A Sourcebook. New York: Free Press.

Cronbach, L.J. 1951 Coefficient alpha and the internal structure of tests. Psychometrika 16:297-334.

Dembo, R., L. Williams, A. Getreu, L. Genung, J. Schmeidler, E. Berry, E. Wish, and L. LaVoie 1991 A longitudinal study of the relationship among marijuana/hashish use, cocaine use and delinquency in a cohort of high risk youths. Journal of Drug Issues 21(2):271-312.


Elliott, D.S., D. Huiziuga, and S. Menard 1989 Multiple Problem Youth: Delinquency, Drugs, and Mental Health Problems. New York: Springer-Verlag.


Fagan, J., and J.G. Weis 1990 Drug Use and Delinquency Among Inner City Youth. New York: Springer-Verlag.

Fagan, J., E. Piper, and M. Moore 1986 Violent delinquents and urban youths. Criminology 24(3):439-470.

Farrington, D. 1989 Early predictors of adolescent aggression and adult violence. Violence and Victims 4(2):79-100.

Farrington, D., and R. Loeber 1989 Relative improvement over chance (RIOC) and phi as measures of predictive efficiency and strength of association in 2×2 tables. Journal of Quantitative Criminology 5(3):201-214.

Flinn, C. 1986 Dynamic models of criminal careers. Pp. 356-379 in J. Roth and

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

C. Visher, eds., Criminal Careers and "Career Criminals." Washington, D.C.: National Academy Press.

Fox, J.A., and J. Levin 1994 Overkill: Mass Murder and Serial Killing Exposed. New York: Plenum.

Furby, L., M. Weinrott, and L. Blackshaw 1989 Sex offender recidivism. Psychological Bulletin 105(1):3-30.

Garrison, W. 1984 Predicting violent behavior in psychiatrically hospitalized boys. Journal of Youth and Adolescence 13(3):225-238.

Gelles, R.J., and M.A. Straus 1988 Intimate Violence. New York: Simon and Schuster.

Gibbons, D.C. 1975 Offender typologies—Two decades later. British Journal of Criminology 15(2):140-156.

Glaser, D. 1964 The Effectiveness of a Prison and Parole System. New York: Bobbs-Merrill.

Goffman, E. 1959 The Presentation of Self in Everyday Society. Garden City, N.Y.: Doubleday.

1963 Stigma. Englewood Cliffs, N.J.: Prentice-Hall. Golden, C.J., R.F. Sawicki, and M.D. Franzen

1984 Test construction. Chapter 2 in G. Goldstein and M. Herson, eds., Handbook of Psychological Assessment. New York: Pergamon Press.

Goldstein, P. 1989 Drugs and violent crime. In N. Weiner and M. Wolfgang, eds., Pathways to Criminal Violence. Newbury Park, Calif.: Sage Publications.

Gottfredson, D., and S. Gottfredson 1988a Stakes and risks in the prediction of violent criminal behavior. Violence and Victims 3(4):247-262.

Gottfredson, S., and D. Gottfredson 1988b Violence prediction methods: Statistical and clinical strategies. Violence and Victims 3(4):303-324.


Hare, R.D. 1979 Biological and behavioral correlates of criminal psychopathy. In L. Beliveau et al., eds., Research on Diagnoses and Treatment. Montreal: International Centre for Comparative Criminology.

Hathaway, S.R., and J.C. McKinley 1940 A multiphasic personality schedule (Minnesota): I. Construction of the scale. Journal of Psychology 10:294-354

Heckman, J. 1979 Sample selection bias as a specification error. Econometrica 47:153-161.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Heckman, J., and B. Singer 1985 Longitudinal Analysis of Labor Market Data. Cambridge: Cambridge University Press.

Holden, R. 1985 Failure time models for thinned crime commission data. Sociological Methods and Research 14(1):3-30.

Holland, T., N. Holt, and G. Beckett 1982 Prediction of violent versus nonviolent recidivism from prior violent and nonviolent criminality. Journal of Abnormal Psychology 91(3):178-182.

Holmes, R.M. 1989 Profiling Violent Crimes: An Investigative Tool. Newbury Park, Calif.: Sage.

Howell, M., and K. Pugliesi 1988 Husbands who harm: Predicting spousal violence by men. Journal of Family Violence 3(1):15-27.

Irwin, J. 1970 The Felon. Englewood Cliffs, N.J.: Prentice-Hall.


Johnson, B.D., P. Goldstein, E. Preble, J. Schmeidler, D.S. Lipton, B. Spunt, and T. Miller 1985 Taking Care of Business: The Economics of Crime by Heroin Abusers . Lexington Mass.: Lexington Books.


Kagan, J. 1982 Psychological Research on the Human Infant: An Evaluative Study . New York: William T. Grant Foundation.

Kalbfleish, J., and R. Prentice 1980 The Statistical Analysis of Failure Time Data. New York: John Wiley & Sons.

Kennedy, T. 1986 Trends in inmate classification: A status report of two computerized psychometric approaches. Criminal Justice and Behavior 13(2):165-184.

King, G. 1988 Statistical models for political science event counts: Bias in conventional procedures and evidence for the Expontential Poisson Regression Model. American Journal of Political Science 32:838-863.

Klassen, D., and W. O'Connor 1988a A prospective study of predictors of violence in adult male mental health admissions. Law and Human Behavior 12(2):143-158.

1988b Crime, inpatient admissions, and violence among male mental patients. International Journal of Law and Psychiatry 11:305-312.

1988c Predicting violence in schizophrenic and non-schizophrenic patients: A prospective study. Journal of Community Psychology 16:217-227.

1989a Predictors of Violence: A Review. Paper presented at the Study

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Group Meeting on Risk of the MacArthur Foundation Program of Research on Mental Health and the Law. Palm Beach, Florida, January.

1989b Assessing the risk of violence in released mental patients: A cross-validation study. Psychological Assessment: A Journal of Consulting and Clinical Psychology 1(2):75-81.

n.d. Predicting Violence in Mental Patients: Cross Validation of an Actuarial Scale. Unpublished paper, Greater Kansas City Mental Health Foundation.

Koss, M., and C. Oros 1982 Sexual experiences survey: A research instrument investigating sexual aggression and victimization. Journal of Consulting and Clinical Psychology 50:455-457.

Kuder, G.G., and M.W. Richardson 1937 The theory of estimation of test reliability. Psychometrika 2:151-160.

Levin, J., and J.A. Fox 1991 Mass Murder: America's Growing Menace. New York: Berkeley Books.

Lewis, D.O., and D.A. Balla 1976 Delinquency and Psychopathology. New York: Grune & Stratton.

Lindesmith, A.R., and A.L. Strauss 1968 Social Psychology. New York: Holt, Rinehart & Winston. L

Loeber, R., and T. Dishion 1983 Early predictors of male delinquency: A review. Psychological Bulletin 94(1):68-99.

Loeber, R., and M. Stouthamer-Loeber 1987 Prediction. In H. Quay, ed., Handbook of Juvenile Delinquency . New York: John Wiley & Sons.


Maddala, G. 1983 Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press.

Malamuth, N. 1986 Predictors of naturalistic sexual aggression. Journal of Personality and Social Psychology 50(5):953-962.

1988 Predicting laboratory aggression against females and males targets: Implications for sexual aggression. Journal of Research in Personality 22:474-495.

1989 The attraction to sexual aggression scale: Part one. Journal of Sex Research 26:26-49.

Maltz, M. 1984 Recidivism. New York: Academic Press.

Mann, F.C., J. Friedman, and A.S. Friedman 1976 Characteristics of self-reported violent offenders vs. court identified violent offenders. International Journal of Criminology and Penology 4:69-87.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Marquis, K.H., and P. Ebener 1981 Quality of Prisoner Self-Reports: Arrest and Conviction Response Errors. Santa Monica, Calif.: The RAND Corporation.

McConaughy, S.H., T.M. Achenbach, and C.L. Gent 1988 Multiaxial empirically based assessment: Parent, teacher, observations, cognitive and personality correlates of child behavior profiles for 6-11-year old boys. Journal of Abnormal Child Psychology 11:485-509.

McCord, W., and J. McCord, with I.K. Zola 1959 Origins of Crime. New York: Columbia University Press.

Mead, G.H. 1934 Mind, Self and Society. Chicago: University of Chicago Press.

Megargee, E.I. 1977 A classification system for male youthful offenders based on the MMPI. Pp. 35-60 in R.A. Keil, ed., Mental Health for the Convicted Offender, Patient, and Prisoner. Raleigh: North Carolina Department of Corrections.

Megargee, E.I., and M.J. Bohn 1979 Classifying Criminal Behaviors. Newbury Park, Calif.: Sage Publications.

Megargee, E.I., M.J. Bohn, and J.L. Carbonell 1988 A Cross-Validation and Test of the Generality of the MMPI-Based Offender Classification System. Unpublished manuscript, Florida State University, Tallahassee.

Menzies, R., and C. Webster 1989 Mental disorder and violent crime. In N. Weiner and M. Wolfgang, eds. Pathways to Criminal Violence. Newbury Park, Calif.: Sage Publications.

Menzies, R., C. Webster, and D. Sepajak 1985 The dimensions of dangerousness. Law and Human Behavior 9(1):49-70.

Mercy, J.A., and P.W. O'Carroll 1988 New directions in violence prediction: The public health arena. Violence and Victims 3(4):285-302.

Milner, J. 1980 The Child Abuse Potential Inventory: Manual. Webster, N.C.: Psytec Corporation.

1989 Additional cross-validation of the Child Abuse Potential Inventory. Psychological Assessment: A Journal of Consulting and Clinical Psychology 1:219-223.

Milner, J.S., and R. Wimberley 1979 An inventory for the identification of child abusers. Journal of Clinical Psychology 35:95-100.

1980 Prediction and explanation of child abuse. Journal of Clinical Psychology 35:875-884.

Milner, J., R. Gold, and R. Wimberley 1986 Prediction and explanation of child abuse: Cross validation of

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

the child abuse potential inventory. Journal of Consulting Psychology 54(6):865-866.

Milner, J.S., K.R. Robertson, and D.L. Rogers 1988 Childhood History of Abuse and Adult Abuse Potential. Paper presented at the meeting of the Midwester Psychological Association, Chicago.

Milner, J., R. Gold, C. Ayoub, and M. Jacewitz 1984 Predictive validity of the child abuse potential inventory. Journal of Consulting and Clinical Psychology 52(5):879-884.

Monahan, J. 1981 Predicting Violent Behavior: An Assessment of Clinical Techniques . Newbury Park, Calif.: Sage Publications.

1984 Prediction of violent behavior—Toward a second generation of theory and policy. American Journal of Psychiatry 141(1):10-15.

Moss, C., M. Johnson, and R. Hosford 1984 An assessment of the Megargee typology in lifelong criminal violence. Criminal Justice and Behavior 11(2):225-234.

Nurco, D., T. Hanlon, T. Kinlock, and K. Duszynski 1988 Differential criminal patterns of narcotic addicts over an addiction career. Criminology 26(3):407-423.


Rhodes, W. 1985 The adequacy of statistically derived prediction instruments in the face of sample selectivity. Evaluation Review 9:369-382.

1986 A survival model with dependent competing events and right-hand censoring: Probation and parole as an illustration. Journal of Quantitative Criminology 2(2):113-137.

1989 The criminal career: Estimates of the duration and frequency of crime commission. Journal of Quantitative Criminology 5(1):3-32.

Rivera, B., and C. Spatz Widom 1991 Childhood victimization and violent offending. Violence and Victims 5:19-35.

Robertson, K.R., and J.S. Milner 1985 Convergent and discriminant validity of the Child Abuse Potential Inventory. Journal of Personality Assessment 49:86-88.

Robins, L.N., and E. Wish 1977 Childhood deviance as a developmental process: A study of 223 urban black men from birth to 18. Social Forces 56(2):448-471.


Schmidt, P., and A. Witte 1988 Predicting Recidivism Using Survival Models. New York: Springer-Verlag.

Shaffer, J., D. Nurco, and T. Kinlock 1984 A new classification of narcotic addicts based on type and extent of criminal activity. Comprehensive Psychiatry 25(3).

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

Simcha-Fagan, O., and J.E. Schwartz 1986 Neighborhood and delinquency: An assessment of contextual effects. Criminology 24(4):667-695.

Snyder, H.N. 1988 Court Careers of Juvenile Offenders. Pittsburgh, Pa.: National Center for Juvenile Justice.

Spitzer, R.L., and J.B.W. Williams, eds. 1987 Introduction. Pp. xvii-xxvii in Diagnostic and Statistical Manual of Mental Disorders, Third Edition—Revised. Washington, D.C.: American Psychiatric Association.

Steadman, H. 1987 How well can we predict violence for adults? A review of the literature and some commentary. Pp. 5-17 in F. Dutile and C. Foust, eds., The Prediction of Criminal Violence. Springfield, Ill.: Charles C. Thomas.

Steadman, H., and J. Morrissey 1982 Predicting violent behavior: A note on a cross-validation study. Social Forces 61(2).

Toch, H., and K. Adams 1989 The Disturbed Violent Offender. New Haven, Conn.: Yale University Press.

Tracy, P., M. Wolfgang, and R. Figlio 1990 Delinquency Careers in Two Birth Cohorts. New York: Plenum Press.


Virkkunen, M., J. DeJong, J. Bartko, F. Goodwin, and M. Linnoila 1989a Relationship of psychobiological variables to recidivism in violence offenders and impulsive fire setters: A follow-up study. Archives of General Psychiatry 46:600-603.

1989b Psychobiological concomitants of history of suicide attempts among violent offenders and impulsive fire setters. Archives of General Psychiatry 46:604-606.


Weiner, N 1989 Violent criminal careers and violent career criminals: An overview of the recent literature. Pp. 35-138 in Violent Crime, Violent Criminals . Newbury Park, Calif.: Sage Publications.

n.d. Violent Recidivism Among the 1958 Philadelphia Cohort Boys, Vol. I. Unpublished manuscript, Center for the Interdisciplinary Study of Criminal Violence, Sellin Center for Studies in Criminology and Criminal Law, The Wharton School, University of Pennsylvania, Philadelphia.

Weinstein, E. 1969 The development of interpersonal competence. Pp. 753-775 in D. Goslin, ed., Handbook of Socialization Theory and Research. Chicago: Rand McNally.

Widom, C.S. 1989a The intergenerational transmission of violence. In N. Weiner

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×

and M. Wolfgang, eds., Pathways to Criminal Violence. Newbury Park, Calif.: Sage Publications.

1989a The cycle of violence. Science 244:160-166.

Williams, T.M., and W. Kornblum 1985 Growing Up Poor. Lexington, Mass.: Lexington Books.

Wolfgang, M., R. Figlio, and T. Sellin 1972 Delinquency in a Birth Cohort. Chicago: University of Chicago Press.

Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 217
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 218
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 219
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 220
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 221
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 222
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 223
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 224
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 225
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 226
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 227
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 228
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 229
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 230
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 231
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 232
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 233
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 234
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 235
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 236
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 237
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 238
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 239
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 240
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 241
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 242
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 243
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 244
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 245
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 246
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 247
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 248
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 249
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 250
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 251
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 252
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 253
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 254
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 255
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 256
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 257
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 258
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 259
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 260
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 261
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 262
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 263
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 264
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 265
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 266
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 267
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 268
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 269
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 270
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 271
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 272
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 273
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 274
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 275
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 276
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 277
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 278
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 279
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 280
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 281
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 282
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 283
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 284
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 285
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 286
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 287
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 288
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 289
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 290
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 291
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 292
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 293
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 294
Suggested Citation:"Predicting Violent Behavior and Classifying Violent Offenders." National Research Council. 1994. Understanding and Preventing Violence, Volume 4: Consequences and Control. Washington, DC: The National Academies Press. doi: 10.17226/4422.
×
Page 295
Next: Incarceration and Violent Crime: 1965-1988 »
Understanding and Preventing Violence, Volume 4: Consequences and Control Get This Book
×
Buy Paperback | $65.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

This book analyzes the consequences of violence and strategies for controlling them. Included are reviews of public perceptions and reactions to violence; estimates of the costs; the commonalities and complementarities of criminal justice and public health responses; efforts to reduce violence through the prediction and classification of violent offenders; and the relationships between trends in violence and prison population during a period of greatly increased use of incarceration.

  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. ×

    Switch between the Original Pages, where you can read the report as it appeared in print, and Text Pages for the web version, where you can highlight and search the text.

    « Back Next »
  6. ×

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

    « Back Next »
  7. ×

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

    « Back Next »
  8. ×

    View our suggested citation for this chapter.

    « Back Next »
  9. ×

    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!