Research on Sentencing: The Search for Reform, Volume I (1983)

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

Determinants of Sentences ISSUES A diverse body of research exists on the determinants of sentences. This subject has been pursued from widely varying perspectives exploring the roles of normative premises and conceptions of justice, social struc- ture, organizations, conflict, and politics in influencing sentence out- comes. Underlying much of this research has been a fundamental con- cern with accounting for the diversity of sentence outcomes observed in courts. This has involved attempts to identify the variety of variables, and the interrelationships among those variables, that combine to in- fluence observed sentence outcomes. The increasing complexity of variables considered as factors influ- encing sentences has been accompanied by increasing methodological sophistication of the statistical analyses of sentencing. The earliest stud- ies often involved no more than simple bivariate contingency tables examining the relationship of a single variable to sentences (e.g., the number sentenced to prison for each race). More recent studies use assorted multivariate techniques, usually applied to linear models, that permit simultaneous statistical controls for the variety of factors thought to affect sentences. To date, the general state of knowledge about the factors influencing sentence outcomes remains largely fragmented, and there is no widely accepted theory on the determinants of sentences. Indeed, research on sentencing derives from a variety of different theoretical and disciplinary perspectives. 69

70 RESEARCH ON SENTENCING THE SEARCH FOR REFORM THE RANGE OF VARIABLES CONSIDERED AS DETERMINANTS OF SENTENCES Research on sentencing has considered both the role of case attributes at the time of sentencing and the role of various aspects of the sentence decision-making process. The principal variable measures that are used in sentencing research are listed below. I. Case Attributes 1. Offense Attributes a. Offense Seriousness: crime type~s) charged or convicted; num- ber of charges; statutory maximum sentence; injury or threat of injury to victim; weapon use; value of property stolen or damaged; number of accomplices; role of offender as principal or accessory in offense; victim vulnerability; victim provoca- tion; nature of offender/victim relationship; intent b. Quality of Evidence: number of witnesses; cooperation of wit- nesses; existence of tangible evidence; strength of defendant's alibi 2. Offender Attributes a. Prior Criminal Record: number of arrests, convictions, or in- carcerations; types of offenses; recency of prior events; liberty status at time of offense release on bail, probation, or parole at time of offense b. Demographic Attributes: age; race; sex c. Socioeconomic Status: occupational prestige; income; educa- tion d. Social Stability: employment history; marital status; living ar- rangements; history of drug or alcohol abuse 3. Case-Processing Variables Charge reductions or dismissals; pretrial release status—on bail or detained; attorney type—none, court-appointed, or privately retained; method of case disposition—guilty plea, bench or jury trial; time of guilty plea; presentence recommendations by pro- bation officer, prosecutor, and defense counsel II. Attributes of Decision-Making Process 1. Structural Variables ("Where") Community attitudes toward crime and punishment; publicity surrounding this case or other similar cases; selection process of

Determinants of Sentences 71 judges—elected or appointed; timing of next election of court officials; stability of courtroom workgroups; processing time; his- torical time period 2. Individual Decision-Maker Variables ("Who") Individual identifiers of key decision makers in each case; de- mographic attributes of key decision makers; general political/ ideological orientation of decision makers—conservative or lib- eral; decision maker's philosophy of sentencing relative impor- tance placed on retributive, rehabilitative, deterrent, or incapa- citative goals; decision maker's "special hang-ups" (e.g., being especially harsh on drug offenses or weapons offenses) 3. Procedural Variables ("How") Local legal practices in criminal cases; role of judge in plea bar- gaining; plea bargaining over charges and/or sentencing options; statutory (e.g., criminal code) or administrative regulations gov- erning sentencing; richness of variables maintained for each case; accuracy of those variables (data sources and validity checks); accessibility of data (e.g., manual or machine-readable files) Variables on case attributes include attributes characterizing the of- fender and the offense, particularly variables that function as indicators of criminal culpability and the potential rehabilitative/deterrent/inca- pacitative effect of imprisoning the offender. These variables include various factors in offense seriousness and characteristics of the offender, such as prior criminal record, employment, age, and sex. Also among the case attributes at the time of sentencing are the outcomes of earlier decisions in case processing, like charging and bail decisions, mode of case disposition, and attorney type. The variables characterizing the sentence decision-making process relate to where the decision is made, who makes the decision, and how the decision is made. The "where" variables refer to the social context in which the decision is made (e.g., jurisdiction or region) and are meant to reflect differences in community attitudes toward crime and punish- ment and differences in system attributes (e.g., case load, backlogs, elected or appointed judges). The "who" variables refer to decision- maker attributes, particularly attributes of judges and perhaps of pro- bation officers, prosecutors, and defense counsel if they have contrib- uted to the sentence outcome. These variables might include indicators of primary cultural reference groups, political orientation, and philos- ophy of sentencing for individual decision makers. The "how" variables refer to procedural differences, such as whether or not there is a formal

72 RESEARCH ON SENTENCING THE SEARCH FOR REFORM pretrial conference, whether that conference involves the judge, and whether the conference is limited to consideration of charges or also explicitly includes sentence options. DISCRIMINATION AND DISPARITY Exploration of the determinants of sentences is often framed in the context of important policy questions. Motivated by charges that sen- tencing is unfair, a major concern in sentencing research has been the extent of unwarranted variation in criminal sentences, particularly the validity of claims of widespread discrimination against black and poor defendants, and of large disparities in sentences. While widely used, the concepts of "discrimination" and "disparity" are rarely defined consis- tently. In this report they are distinguished in terms of the legitimacy of the criteria for determining sentences and the consistency with which those criteria are applied to similar cases. Discrimination exists when some case attribute that is objectionable (typically on moral or legal grounds) can be shown to be associated with sentence outcomes after all other relevant variables are adequately con- trolled.i Such an association is taken as presumptive evidence of the existence and extent of deliberate discrimination. Race is the clearest example of an illegitimate criterion; it is a "suspect classification" from a legal perspective and is widely viewed as inappropriate on moral grounds. The range of potentially illegitimate variables is viewed broadly in this report and may include case-processing variables, like bail status or type of attorney, in addition to the personal attributes that are conventionally cited as bases of discrimination (see list above). Disparity exists when "like cases" with respect to case attributes- regardless of their legitimacy—are sentenced differently. For example, this might occur when judges place different weights on the various case attributes or use different attributes in their sentencing decisions. Dis- parity refers to the influence in sentence outcomes of factors that char- acterize the decision-making process. The most commonly cited ex- amples of disparity are differences among judges within the same jurisdiction or in different jurisdictions. ~ As a policy matter, concern with discrimination has been primarily concerned with deliberate behavior that is discriminatory in intent. Research on discrimination, however, rests on outcomes and cannot distinguish purposive discriminatory behavior from behavior that is discriminatory in effect. As a result, research findings of discrimination refer to findings of discriminatory outcomes that may or may not result from discriminatory intent.

Determinants of Sentences 73 By these definitions discrimination and disparity are quite distinct behaviors (see Table 2-1~. If all decision makers behaved similarly, and used race or bail status as a factor in sentences, for example, it would be possible (though unlikely) to have discrimination without disparity. If all decison makers held shared values about legitimate case attributes, but placed different weights on them, the result would be disparity without discrimination. If some decision makers gave weight to race in their sentencing decisions and some did not (or gave race less weight), sentences would exhibit both disparity and discrimination. Evaluating the extent of discrimination or of unwarranted disparity requires important normative judgments about how much and what types of variation are unwarranted. Concern with discrimination focuses largely on the invidious role of certain personal attributes of the of- fender, particularly race and socioeconomic status, and the use of various case-processing variables. Concern for disparity, on the other hand, centers on the role of the organizational or structural context in which sentencing decisions are made and on the attributes of individual de- . . clslon ma hers. Discrimination A finding of discrimination first requires evaluation of the legitimacy of the potential factors associated with sentencing outcomes. This assess- ment is likely to be highly subjective, involving disagreement over the goals of sentencing and a balancing of those goals with whatever con- straints on sentencing may prevail in a particular society at a given time. Consider, for example, the ambiguous status of variables like age and employment. The use of such variables in sentencing is often explicitly justified by statute, as in special sentencing provisions for juvenile and young adult offenders and in revisions to the Federal Criminal Code recently proposed in the U.S. Senate (S. 1722, 1980~. Youthfulness can TABLE 2-1 Characterizing Sentence Outcomes in Terms of Disparity and Discrimination Legitimacy of Sentencing Criteria Application of Sentencing Criteria Consistent Inconsistent Legitimate No disparity and Disparity no discrimination Illegitimate Discrimination Disparity and discrimination

74 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM be considered a mitigating factor based on the presumed diminished culpability of young offenders. Use of unemployment can be justified on grounds that it is an indicator of greater risk of further crime for offenders placed under supervision in the community. But arguments can also be offered that these variables are not legitimate sentencing criteria. It might be argued, for example, that the intensity of offending is high among the young and that they thus pose a serious threat of continued offending. On grounds of deterrence or incapacitation, then, youthfulness would not be a legitimate basis for being sentenced leni- ently (Boland and Wilson, 1978; Kennedy, 1978; Wolfgang, 1978~. Like- wise it might be argued that employment status is highly associated with race; to the extent that race is an illegitimate variable for sentencing, employment should be similarly suspect. For these reasons employment was recently removed as a factor in the Maryland statewide sentencing guidelines (Sentencing Guidelines Project, 19814. Similarly, because of considerations of legitimacy, education no longer appears in the federal parole guidelines (Hoffman et al., 1978~. The legitimacy of a variable for sentencing may also vary with the type of sentencing decision. Because of differences in the probabilities of recidivism, it could be argued that employment status is legitimate for determining whether to incarcerate or not, but that employment status should be immaterial to the length of a prison term. In this case, use of employment status would be nondiscriminatory (i.e., legitimate) in the prison/no prison decision, but discriminatory (i.e., illegitimate) in the decision on length of incarceration. Discrimination can also exist when an otherwise legitimate variable is given an illegitimately large weight in the sentencing decision. For example, it might be widely accepted that pleading guilty warrants a discount in sentence; the amount of that discount, however, would likely be unacceptable if type of plea were used to determine whether or not the prosecutor seeks the death penalty. Here discrimination occurs when the impact of an otherwise legitimate variable exceeds (or falls short of) some acceptable margin. Disparity When considering the extent of unwarranted disparity, it is useful to distinguish four types of disparity. These different forms of disparity cannot be evaluated equivalently; they may or may not be justified, and some may even be desirable.

Determinants of Sentences 7s First, there may only be the appearance of disparity. This occurs when cases seem alike to an outside observer but differ materially in case attributes observed by the judge. For example, if the facts in two cases are identical but one defendant exhibits remorse and the other does not, they might receive different sentences. What appears to be disparity to a researcher working only from case records could be explained by the variables evident to the judge but not available in the records. Improved observations of independent variables like offender culpability, includ- ing such subtle considerations as remorse, may reduce the amount of this seeming disparity. Second, there may be planned disparity or disparity that is deliberately introduced as a matter of social policy, such as use of exemplary sen- tences (Morris, 19824. Consider, for example, several tax evaders who have been tried and convicted and who are thus all vulnerable to in- carceration. If it has previously been decided that it is sufficient to incarcerate only one of these offenders to achieve the desired general deterrent effect and thereby reduce the social costs associated with pun- ishment, singling out the one offender among many for such punishment would represent planned disparity. Planned disparity might also arise if "like" offenders are entitled only to an equal opportunity of receiving a particular sentence, which might be imposed through means of a lottery, for example. Under both these schemes, justice is served when all like offenders are vulnerable to some range of acceptable sentences by virtue of conviction. They are, however, not all sentenced equally harshly. Instead, particular sanctions are allocated with reference to other social ends, such as crime prevention through deterrence or in- capacitation and minimizing the social costs of punishment. A deliberate social policy of planned disparity would be warranted to the extent that the interests of justice can be responsibly limited to concern for an offender's vulnerability to a range of acceptable (i.e., not unjust) sen- tences. If, however, one's concept of justice requires equal treatment for like offenders, planned disparity in forms like exemplary sentences or equal opportunities to sanctions would be unwarranted. The third type of disparity involves interjurisdictional disparity such as that found between urban and rural courts in the same state. Such juris- dictional differences may reflect differences in community standards of offense seriousness or punitiveness, or it might reflect local organizational conditions like court overcrowding. Whether these jurisdictional differ- ences are warranted or not depends on the resolution of competing values, such as concern for evenhandedness or uniformity of standards versus the value of preserving local community control. In either case, however,

76 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM jurisdictional differences arising from application of discnm~natory (ille- gitimate) sentencing criteria would remain unwarranted. The last type of disparity relates to individual judges.2 This type of disparity can arise from fundamental philosophical differences regarding the goals of sentences, which may not be shared universally, or, even if they are, cannot be applied consistently. These differences may reflect differences in the experiences, training, and background of individual judges (or of court personnel making sentence recommendations to the judge) and would show themselves in use of different sentencing criteria or the application of different weights to the various criteria. The in- terjudge or intrajudge disparity that results may or may not be war- ranted. On one side, it could be argued that some variation in sentences is to be expected and even tolerated in order to accommodate reasonable differences of opinion in the application of legitimate sentencing stan- dards. As long as vulnerability to a particular judicial perspective does not vary systematically with defendant or case attributes (e.g., defen- dants charged with offenses involving gun use are no more likely to appear before judges favoring strong gun control than are any other defendants), the differences among judges in sentencing similar cases may be regarded as an acceptable or tolerable reflection of variation in the legitimate standards held within a community and so be warranted. (From this perspective, however, differences between jurisdictions or judges that arise from use of discriminatory (illegitimate) sentencing criteria by some judges or jurisdictions would remain unwarranted.) Alternatively, it might be argued that the application of different legal standards to identical defendants is inconsistent with the rule of law. Normally, the U.S. legal system operates through appellate review and legislative change to eliminate conflicting legal rules, particularly when individual liberty is at issue, and does not tolerate the degree of incon- sistency that may today characterize the sentencing behavior of different judges. If sentencing is to be similarly constrained by legal rules (as some proponents of reform urge), philosophical differences among judges would have to be significantly reduced or eliminated, perhaps through some compromise among judges or through the selection of a preferred sentencing rule by some democratically accountable body. Under this perspective, convergence of sentencing standards is preferable to con- 2 While judges are the decision makers typically identified in discussions of disparity, disparity in sentence outcomes can also arise from differences among prosecutors or other criminal justice decision makers.

Determinants of Sentences 77 tinned toleration of disparity. Some proponents of change also argue that significant variations among judges based on different philosophies are also unwarranted, because many operational consequences of that variation like "judge shopping" by both defense and prosecuting at- torneys, and perceptions of arbitrariness in sentences~ontribute to a sense of impropriety and injustice that undermines confidence in the legitimacy of the courts and the entire criminal justice system. ALTERNATE METHODOLOGICAL APPROACHES TO ANALYSES OF SENTENCING In this chapter we focus primarily on statistical studies of sentencing that have used quantitative data on case attributes and decision-process variables; in Volume II, Garber et al. and Klepper et al. discuss the possibility of developing more sophisticated formal models of the sen- tencing process as a basis for improved statistical analyses. However, much work on criminal sentencing has used quite different research methods. Among the most common have been observation of the behavior of criminal court participants and interviews with them. Some of this work has used the paradigm of anthropological study of a new culture; some has used concepts from organization theory as the basis for data gathering and analysis; and some of this work has been primarily descriptive. Another body of research uses experimental simulations in which subjects are asked to "sentence" experimental cases. A major concern in this experimental research is the process of attribution of factors, like offender culpability and victim provocation, by decision makers. While the processes involved in forming these judgments are not fully under- stood, several factors have been suggested as potentially relevant. These include the individual's ability to carry out the act, the effort expended, the degree of planning involved, the level of psychological functioning, and the type of motivation.3 Experimental manipulation is particularly well suited for exploring the impact of these subtle and often unmeasured factors. Our focus on one research approach is due to the large number of 3 Research examining elements of attribution in the context of sentencing includes: Harvey and Engle (1978), Hogarth (1971), Hood (1972), Joseph et al. (1976), Kapardis and Farrington (1982), Monahan and Hood (1976), Sebba (1980), Thomas (1979), Walster (1966), and Wheeler et al. (1981~. More general treatments of attribution theory are available in Heider (1958) and Weiner (1974~.

78 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM studies and the technical questions that they raise; it should not be taken to imply that this approach is the only one of value. Indeed, we believe that statistical analysis of quantitative data about sentencing or attempts to model the process should include consideration of the particular court- house cultures in which the behavior is embedded. Such consideration requires gathering information from participants themselves. In addi- tion, the careful controls possible in experimental research provide the opportunity for isolating the potentially subtle effects of variables, like defendant demeanor, that are difficult if not impossible to measure in aggregate statistical analyses. Studies of criminal courts have repeatedly demonstrated that juris- dictions vary substantially in terms of norms of appropriate sentencing policy (e.g., levels of harshness) as well as in standard operating pro- cedures (e.g., use of trial versus guilty plea and the implications of selection of one mode of disposition for ultimate sentence outcome). These norms are crucial to explanations of why different sentence out- comes occur but are typically unmeasured by generally available statis- tical data. In some jurisdictions, for example, bench trials are the equiv- alent of "slow pleas" and are appropriately coded as guilty pleas rather than trials; in others, they are quite real trials. Thus, a decision to treat bench trials as trials or as guilty pleas for purposes of statistical analysis cannot sensibly be made without knowledge of the operating norm within the particular jurisdiction. Furthermore, the potential differences in processing cases across jurisdictions, and sometimes even between courts within a jurisdiction, raise important questions about the appropriate- ness of cross-sectional analyses that assume a single homogeneous pro- cess in different settings. Observation and Interviews In our discussion of the use of variables measuring crime seriousness and prior record, we note that problems of measurement error present a difficult obstacle. Interviews with court personnel may be useful in identifying the key dimensions of case seriousness (degree of harm ac- tually done? risk of injury? offender culpability? victim provocation?) and the important aspects of prior record (arrests? convictions? jail or prison terms? recency versus severity of prior arrests or sentences?), as well as in alerting a researcher to differences among jurisdictions that may be obscured in multijurisdictional comparisons that use only one set of measures. Formal modeling of justice system operations can be considerably

Determinants of Sentences 79 improved by field work attempting to assess accurately the actual goals and behavior of participants. Do prosecutors attempt to maximize con- viction rates or sentence severity? Interviews are essential to develop sensible models. By the same token, models that use realistically dif- ferent utility functions for different types of attorneys (e.g., public de- fenders versus marginal private practitioners versus well-established criminal lawyers) could be developed on the basis of interviewing par- ticipants. Research based on observations or interviews faces real issues of the validity and reliability of often qualitative and subjective judgments made by investigators. Moreover, whether using quantitative or quali- tative techniques, research from a single jurisdiction must confront issues of generalizability. Experiments Experimental manipulation of a small number of variables permits iso- lating the independent contribution of variables that covary or interact with other independent variables in natural settings (e.g., age and crim- inal record). It also provides an opportunity to explore the impact of the full range of variation in variables whose effect in natural settings is difficult to measure because of their limited variation in those settings (e.g., sex or conviction type guilty plea or trial). Small effects of some variables that may be obscured by the much larger effects of other variables in aggregate statistical analyses can also be highlighted in ex- periments. This is particularly important in considerations of variables that, despite their small effect in aggregate data, are nevertheless im- portant for conceptual or policy reasons (e.g., racial discrimination). Experimental studies face challenges to the external validity of results arising from the artificial and often contrived character of the experi- mental situation. These studies, for example, often use inappropriate decision makers, drawing from jury pools or college students who are markedly different from and lack the experience of typical sentencers. Recognizing the problems of having inexperienced respondents assign sentences, the studies often ask respondents to assign levels of respon- sibility or blameworthiness, factors that no doubt affect sentences but are not the sole determinants. Furthermore, the use of often limited case information leaves considerable room for respondent interpretation and imputation of relevant but missing information, which jeopardizes the validity of experimental controls. Experimental research is also vul-

80 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM nerable to response biases when respondents, aware that they are the subjects of research, give the socially desirable or expected response.4 Statistical and Combined Approaches Research on sentencing based strictly on available or uniformly coded quantitative data from several jurisdictions is likely to miss the influence of subtle and typically unmeasured factors as well as to obscure impor- tant differences that may exist across jurisdictions. Furthermore, most attempts to characterize the process quantitatively have been limited to simple linear models in which sentences are posed as functions of simple weighted sums of the independent explanatory variables. More complex characterizations of the process, which are likely to reflect the reality of sentencing decisions more closely and yet still be tractable to analysis, are possible. These models might include, for example, interactions among explanatory variables and hierarchical decision structures, in which some variables are determining factors of sentences when they are present, while in their absence a different set of variables prevails. Standard statistical techniques are available for estimating both simple linear models and more complex models. An overall research strategy that combines interviews, observation, and the familiarity with courthouse cultures that such approaches afford; experiments with their potential for isolating otherwise subtle effects; and statistical analyses of aggregate quantitative data on case attributes and decision-process variables is likely to be most useful in developing knowledge about the determinants of sentences. FINDINGS Despite the growing diversity of factors considered and the increasing methodological sophistication of statistical analyses of sentencing, large portions two-thirds or more—of variance in sentence outcomes remain unexplained. For the portion that is explained, we have reviewed the findings relating to the role of offense seriousness, prior record, race, socioeconomic status, gender, and various case-processing variables. The validity of statistical inferences about the determinants of sentences 4 This type of response bias can be reduced by having the research focus on variables that are not highly charged (as race is) and for which there is no consensus on their use and weight. Use of experienced respondents (i.e., real judges) is also likely to reduce respondent susceptibility to social influence.

Determinants of Sentences 81 depends crucially on the methodological rigor with which the effects were estimated. Thus, the findings presented here are weighed in light of potentially serious methodological flaws in the research. METHODOLOGICAL CONCERNS One methodological concern affecting most research on the determi- nants of sentences is the treatment of the outcome variable sentence imposed. The sentences available to judges typically include choices among a number of qualitatively different options, including suspended sentences, supervised probation, fines, and incarceration, as well as choices on the magnitude of any particular sentence type. Two different approaches have been used to reconcile the different qualitative and quantitative dimensions of sentences. Some researchers focus on the variations in the magnitude of only one sentence type typically the length of prison terms for incarcerated offenders. Other studies collapse different sentence types into a single, arbitrary scale of sentence severity. Analyses that attempt to estimate the effect of variables on the mag- nitude of a single sentence type are vulnerable to a number of different kinds of error. To begin with, it is not obvious that the addition of one month to incarceration or probation terms (or one dollar to fines or restitution sentences) should always be treated in the same way. For short sentences (or small fines) one additional unit may represent an important increase in sentence severity, while for longer sentences (or higher fines) each additional unit may be less important. Simple linear models in which the independent variables enter additively cannot cap- ture such decreases in the marginal severity of the sentence units. Fo- cusing on only one sentence type by assigning values of zero to all other sentence outcomes in ordinary least-squares regression will result in biased estimates of the effects (Hausman and Wise, 1977; Tobin, 1958~. Trying to avoid these biases by restricting the analysis to only those cases of a single sentence type (e.g., only those cases considered for a prison sentence) could introduce selection bias effects. (The sources and nature of these selection biases are discussed in detail below in the context of findings on racial discrimination.) Statistical techniques are available to adequately address nonlinearities in sentence outcomes while still limiting the analysis to a single sentence type. Correcting for the potential biases arising from variables truncated at zero and selected samples, however, requires that the analysis be extended to include choices among sentence types. The alternative approach of using a single scale to represent several different sentence types inevitably raises serious questions of commen-

82 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM curability across the different sentence types that affect the accuracy of both the order and proportionality of the single scale. Most attempts to order the different sentence types into a single scale arbitrarily impose a ranking intended to reflect differences in severity with no empirical substantiation. One commonly used scale is that devised by the Federal Administrative Office of Courts (as reported in Hindelang et al., 1975~. Use of such an arbitrary scale raises serious problems in ordering the different sentence outcomes: for example, there may be disagreement on whether 3 or more years on probation is necessarily more onerous than 6 months of incarceration. Similarly, problems of proportionality arise from the use of arbitrary numerical scores like "two" for probation terms of 13 to 36 months, "seven" for prison terms of 13 to 24 months, and "fourteen" for prison terms of 49 to 60 months. Furthermore, it is not yet empirically established that, for example, prison terms of 54 months are twice as severe as prison terms of 18 months or that prison terms of: 18 months are 3.5 times as severe as probation for 24 months. Estimates of effects obtained from statistical analyses that use a single scale presumed to measure sentence severity as the outcome variable are also vulnerable to several kinds of statistical errors. First, the scale introduces errors in the sentence outcome variable, with an associated loss of precision in estimates of the effects of the determinants of sen- tences. The arbitrariness of the scale also makes it difficult to interpret the magnitude of the measured effects of explanatory variables on dif- ferent sentence types: the impact of a change in a determinant can be interpreted only as an increment in the arbitrary scale units and not in terms of additional years in prison or dollars of fine. Furthermore, since determinants can be expected to affect individual sentence types dif- ferently, the effects associated with the single arbitrary scale may not be relevant to any of the individual sentence types. In single-scale analy- ses, for example, the same model (i.e., the same factors and the same weights on those factors) is assumed to influence both the choice of the sentence type and the choice of the amount of that sentence. Such a model cannot capture a situation in which unemployment, for example, might affect the decision to imprison an offender but would have no effect on the length of the prison term. Furthermore, the choices among different levels of each sentence type (e.g., how long a prison term or how large a fine) are assumed to be determined by the same factors with the same weights on those factors. This would not accurately reflect a situation in which income, for example, does affect the choice of fine amount but has no bearing on the length of prison terms. These problems are pervasive in research on sentencing, affecting both the comparability of results across different studies and the strength

Determinants of Sentences 83 ot conclusions drawn from that research. A more desirable approach would be to partition the sentence outcome into two related outcomes involving (1) a choice among different sentence types and (2) a choice on the magnitude of the selected type. Statistical techniques are available for analyzing the choice of sentence (e.g., PROBIT, LOGIT) type; then, taking account of the bound at zero in the analysis of magnitude, these separate aspects of sentence outcome could and should be estimated simultaneously. This approach would not require the use of arbitrary scales across qualitatively different sentence types. It is also more flex- ible, allowing for differences in the determinants of different aspects of sentences. Findings from qualitative analyses could be very useful in suggesting which variables are more likely to be factors in the different aspects of sentence outcomes. Furthermore, if scales reflecting the rel- ative severity of sentence outcomes are desired, techniques are available for estimating scale values from existing data rather than arbitrarily imposing them (see Klepper et al., Volume II). THE PRIMARY DETERMINANTS OF SENTENCES Using a variety of indicators, offense seriousness and offender's prior record have emerged as the key determinants of sentences. The strength of this conclusion persists despite the potentially severe problems of bias arising from measurement error that characterize most of the empirical research. As indicated in the list above, many different factors may influence judgments of offense seriousness and prior record; few of these are usually included in individual studies of sentencing. As a result, the effects on sentence outcomes of the included indicators of offense se- riousness and prior record are particularly vulnerable to biases arising from the excluded elements. Offense Seriousness Typically, offense seriousness measures are limited to use of the legally defined offense types or the statutory maximum penalties for each of- fense type. Some elements of the offense are often unavailable to re- searchers using court records. These unavailable elements include ex- cessive harm to the victim, weapon use, the role of the victi~partially reflected in the nature of the offender/victim relationship and victim provocation and the offender's role as a principal or accessory. Even when the necessary data elements for the different indicators re- flecting offense seriousness are available, researchers do not know how

84 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM the separate elements combine to influence sentence outcomes. In the most commonly used approach, the various elements reflecting seriousness are assumed to enter the decision in a simple additive fashion in which all factors affect sentence outcomes linearly.S In these models the different elements in offense seriousness are considered simultaneously, and they always have the same incremental impact on sentence outcomes. These models do not adequately capture a hierarchical assessment of the elements of seriousness where the weight given some factors depends on the presence or absence of other factors. Some elements, for ex- ample, may be extremely rare, but when present they may be deter- mining factors in sentences. In a particularly heinous crime, the brutal treatment of victims may be the only element of seriousness considered in determining sentence outcome. In less vicious crimes, a wide variety of factors reflecting different aspects of offense seriousness may enter the sentencing decision. Offender's Prior Record The potential elements of "prior record" including items like the number, recency, and seriousness of prior arrests, prior convictions, and prior in- carcerations—are generally more visible to researchers than elements of offense seriousness. The record data, however, are often subject to errors and incompleteness, both in the data available to decision makers and to researchers. In terms of statistically analyzing the role of record variables in sentence outcomes, data elements that are available to decision makers, but not available to researchers, are especially troublesome. This is often the case for juvenile records, which may be available either formally or informally to decision makers, but are not available to researchers as part of the case record. Much research thus focuses on the role of officially available adult prior records in sentence outcomes. There is currently considerable debate over the extent to which juvenile records are actually used in sentencing adults and over the propriety of using those records. A recent study of the use of juvenile records in adult courts (Greenwood et al., 1980) found that, contrary to the widespread perception that juvenile records are protected against access, these records (in varying quality) are accessible and used to varying degrees in most U.S. 5 A special case of this approach combines the different elements of seriousness linearly to form a single seriousness score, and this score is then posed as a factor in determining sentence outcomes.

Determinants of Sentences 85 junsdictions. The explicit role of juvenile records In sentence outcomes In adult courts, however, remains largely unexplored. Also, as with offense senousness, it is not at all clear how the various elements of a record should be combined to reflect the relative impact of prior record on sentence outcomes.6 Among the issues of concern are commensurability across types of past offenses (e.g., how many misdemeanors are equivalent to one felony?; commensurability across disposition types (e.g., should more or less weight be given to prior incarcerations compared to nonincarcerative sentences?; the form of a decay factor to accommodate diminished importance of older records; and the role of juvenile records. Methodological Issues Inadequate measures of important elements of offense seriousness and prior record can bias estimates of the effects of these variables on sen- tence outcomes. In characterizing the nature of these biases, the dis- cussion here is simplified by treating offense seriousness and prior record as though they were single variables, each resulting from some linear combination of a variety of different elements. Under this character~- zation, when important elements contributing to the unidimensional measures of seriousness or prior record are not measured, there is mea- surement error in the main variable of interest, which results in mea- surement error biases in the estimated effects of these variables on sentence outcomes.7 The bias in the estimated effects of offense seriousness depends on the nature of the measurement error. For a linear model of the deter- minants of sentences, measurement error that is independent of the true level of seriousness yields underestimates of the effect of seriousness on sentence outcomes (i.e., the estimated effect is in the same direction as the true effect but smaller in magnitude).8 If, however, the error in 6 For prior record, as for offense seriousness, a special case involves combining the various elements of prior record, usually linearly, to form a single record score that is posed as a determinant of sentence outcomes. 7 In a more general formulation, the different elements of offense seriousness or prior record are treated as separate measures contributing to sentence outcomes. The biases resulting from failure to include measures of important elements are called specification errors. For a linear model of the determinants of sentences, the nature and direction of the biases arising from these specification errors are similar to those described in terms of measurement error biases in unidimensional variables. ~ This is a standard result that can be found in any text on econometrics or linear statistical estimation (e.g., Johnston, 1972; or Rao, 1973~.

86 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM seriousness due to unmeasured elements varies systematically with ob- served levels of seriousness, the effects of seriousness on sentence out- comes can be underestimated or overestimated: a positive association between observed offense seriousness and its measurement error results in underestimates; a negative association between observed seriousness and its measurement error results in overestimates.9 One source of a positive association between observed seriousness and its measurement error is a negative correlation between observed and unobserved dimensions of seriousness in which high values on ob- served dimensions of seriousness are associated with low values on unob- served dimensions. In this case the mismeasured value of seriousness fails to include elements that offset observed dimensions of seriousness, and observed seriousness increasingly overstates true seriousness. Victim provocation and the existence of a prior relationship between offender and victim, for example, are both elements that might serve to decrease the overall seriousness of an offense. Failure to measure either of these elements would result in underestimates of the effect of offense seriousness on sentence outcomes when more serious observed offense types (based, perhaps, on statutory classifications) are also more likely to involve victim provocation or victims previously known to the offender. Such relationships between offender and victim are likely to be more common in more serious violent offense types, which involve direct contact or confrontation between offender and victim, and less likely in theft offenses, where direct contact is less common. Both victim provocation and the involvement of victims previously known to the offender would then be unobserved factors that decrease true serious- ness below its observed value. In this event, observed seriousness would increasingly overstate true seriousness (i.e., observed seriousness is pos- itively related to its measurement error) and would result in underes- timates of the effect of true seriousness on sentence outcomes. Alternatively, observed seriousness and its measurement error might 9 For error in measurement that is positively related to the observed value of seriousness, increasingly larger values of observed seriousness involve increasingly larger errors added to the true value of seriousness. The resulting relationship between the observed values of seriousness and sentence outcomes has a flatter slope, thus diminishing or underesti- mating the true effect of seriousness on sentence. With a negative relationship between the error in measurement and the observed value of seriousness, increased values of observed seriousness involve increasingly larger errors subtracted from true seriousness. This results in a steeper slope for observed seriousness, thus exaggerating or overestimating the true effect of seriousness on sentence.

Determinants of Sentences 87 be negatively related and yield overestimates of the effect of seriousness if the observed and unobserved elements of seriousness are positively correlated. In this case the positive contribution to true seriousness of unobserved elements is excluded, and true seriousness is increasingly understated. This would occur, for example, if more serious observed offense types were also more likely to involve unobserved elements of seriousness such as injury to a victim, weapon use, or economic loss. Both positive and negative associations between offense seriousness and its measurement error are likely to exist. These systematic errors in measuring seriousness would contribute to both underestimates and overestimates of the true effect of seriousness on a sentence. Any in- dependent errors would result in underestimates. Studies of sentencing vary in the quality of the data used, the juris- dictions examined, and the dimensions of offense seriousness included in the analysis. These variations leave some studies more vulnerable to underestimates and others more vulnerable to overestimates of the effect of offense seriousness. Despite these biases, in both directions, offense seriousness is consistently found to have a strong effect on sentences. The consistency of this result under a variety of different biasing con- ditions increases confidence in the validity of the conclusion that offense seriousness is an important factor in sentence outcomes. Prior record is often measured in terms of its length typically the number of prior contacts with the criminal justice system without re- gard for the content of that record. There is some evidence to suggest that longer prior records are more likely to involve less serious offenses. Using a Sellin-Wolfgang type of scale for offense seriousness (Heller and McEwen, 1973; Sellin and Wolfgang, 1964) on the arrest records of Washington, D.C., arresters, Moitra (1981:46) found that the more prior arrests an arrestee had, the less serious those arrests were likely to be (Figure 2-1~. This might occur because of differential sanctioning by seriousness. To the extent that more serious arrests are more likely to be sanctioned and that sanctions inhibit further arrests through some combination of incapacitation, deterrence, or rehabilitation, offenders engaging in more serious prior offense types would have fewer prior arrests. Such a negative association between observed and unobserved dimensions of prior record would contribute to underestimates of the effect of prior record on sentence severity. Despite the likelihood of biases toward underestimating the effect, prior record is consistently found to have one of the strongest effects on sentence (Bernstein et al., 1977; Chiricos and Waldo, 1975; Lizotte, 1978; Lotz and Hewitt, 1977; Pope, 1975a,b).

88 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM 3.5 O 30 en en cn he o a: LO Number of Prior Arrests ~ p 6 p =8 me, / p = 1 0 '; 1 1 1 0 5 10 15 ARREST SEQUENCE FIGURE 2-1 Average seriousness of prior arrests for arresters with different prior-record lengths Washington, D.C., 1973. SOURCE: Moitra (1981:Figure 2-4~. DISCRIMINATION BY RACE There are two types of evidence often cited in support of the assertion that there is racial discrimination in sentencing. The first is the important social fact that blacks are in prisons in numbers disproportionate to their representation in the population. In 1979, blacks were 10.1 percent of the adult male population, but they comprised 48.0 percent of inmates of state prisons.~° The second set of evidence appears in studies there are now more than That attempt to find a statistical association between the race of defendants and the sentences they receive in crimina is The general population data for 1979 are from the U.S. Department of Commerce (1980~. The data on racial distribution in state prisons are from the 1979 Survey of Inmates of State Correctional Facilities, as reported by the U.S. Department of Justice (1982b).

Determinants of Sentences 89 courts. Some of these studies find an association that has been inter- preted as evidence of racial discrimination in sentencing. Prison Populations The overrepresentation of blacks in prison is clear evidence that some interaction of individual behavior patterns and societal response leads to the imposition of severe punishments on one group of people at rates out of proportion to their numbers in the population. However, it is not by itself evidence that this outcome is in substantial measure the result of racial discrimination at the sentencing stage in criminal courts. The disproportionate rate of imprisonment of blacks may be the product of a wide variety of behaviors and processes. One source of the dispro- portion may be differences in the types and amounts of illegal behavior among races. These behavioral differences may interact with patterns in the deployment of law enforcement resources and differing rates of apprehension, conviction, and imprisonment for various crime types to affect the racial composition of prisons. There might also be racial dis- crimination in the arrest process, the charging process, or the sentencing decision; or decisions by parole authorities may result in longer terms for black prisoners. Some or all of these processes may exist and could contribute to the disproportionate number of black prison inmates; only some might involve racial discrimination in sentencing. The evidence about differential offense rates among races is scanty, and we cannot say with confidence whether the proportion of blacks arrested is the same as the proportion actually involved in illegal activ- ities. It is possible to investigate, as has been done using victimization studies, the racial identities of offenders as reported by their victims. One set of studies (Hindelang, 1976, 1978; Hindelang et al., 1979) re- ports a fairly close correspondence between the proportion of robbers and assaulters who are reported by victims to be black and the proportion of persons arrested for robbery and aggravated assault who are black. However, on the basis of available evidence for crimes more generally, we can conclude little about the degree to which blacks are arrested in true proportion to their offense rates by crime. Focusing only on the postarrest phases of the criminal justice system, one approach to assessing the extent of discrimination would be to examine the correspondence between racial proportions at arrest and in prison. Examination of arrest statistics as shown in Table 2-2, for example, finds a similar differential by race, with blacks accounting for 35 percent of adult arrests for index offenses nationwide in 1979. For the crime types most likely to be found in prison, namely murder and

9o RESEARCH ON SENTENCING: THE SEARCH FOR REFORM TABLE 2-2 Distribution of Total U.S. Adult Arrests (Over 18) by Race and Crime Type in 1979 Total Adult Black Adult Percent Crime Type Arrests Arrests Black Murder 16,534 7,942 48.0 Rape 24,427 11,339 46.4 Robbery 89,463 48,578 54.3 Aggravated assault 216,222 80,847 37.4 Burglary 238,621 74,610 31.3 Larceny 651,745 208,874 32.0 Auto theft 72,753 23,613 32.5 Violenta 346,646 148,706 42.9 Propertyb 972,450 309,327 31.8 Total index offenses 1,319,096 458,033 34.7 a Includes murder, rape, robbery, and aggravated assault. b Includes burglary, larceny, and auto theft. SOURCE: Federal Bureau of Investigation (1980: Table 35~. robbery, the differential is even larger, with blacks accounting for 53 percent of adult arrests. An analysis by Blumstein (1982), exploring the consequences for prison populations of racially differential involve- ment in arrests, estimates that if there were no race-related differences in treatment by the criminal justice system after arrest, 42 percent of the prison population in 1979 would have been expected to be black, in comparison with the actual rate of 48 percent. These data are con- sistent with the assertion that blacks are overrepresented in prison pop- ulations primarily because of their overrepresentation in arrests for the more serious crime types, an argument counter to the assertion that overrepresentation results largely from discrimination at postarrest stages of the criminal justice system. One problem in generalizing from such a result is the difficulty in accurately characterizing racial discrimination through global statements about the criminal justice system in the United States as a whole. If and when it occurs in criminal justice institutions, discrimination on the basis of race is likely to vary across jurisdictions, regions, crime types, and individual participants, and further research at more disaggregated levels is required to isolate those differences. i: Similar results are found for arrests throughout the 197~1979 decade.

Determinants of Sentences 91 There are several possible ways that aggregate statistics can mask discrimination in the criminal justice system. Aggregate national data can conceal important differences among regions, states, or local juris- dictions. For example, rural jurisdictions (where white defendants pre- dominate) may impose more and longer prison sentences than urban jurisdictions (where blacks predominate). The relative leniency of sen- tencing in urban areas could mask possible racial discrimination against blacks in both types of jurisdictions. Thus one next stage of research is a disaggregated analysis that compares sentencing patterns within local and regional units within states. Using data that aggregate different crime types may conceal racial differences in sentencing for particular crime types. For the most serious crimes, such as murder and robbery, prison is the penalty in the great majority of cases, and prisons are predominantly filled with persons who have committed those crimes. In an aggregate analysis of prison pop- ulation, racial neutrality in sentencing for these most serious offenses may obscure important racial differences in sentencing for the less se- rious offenses, for which prison is a possible but not an ordinary out- come. These less serious offenses leave more room for discretion in sentencing decisions and thus greater opportunity for discrimination. Future research should focus on these less serious offenses. There can also be important differences in case processing at different points between arrest and prison, some of which may work to the ad- vantage and some to the disadvantage of black defendants. Prosecutors, for example, may devalue the seriousness of crimes against black victims and be more likely to dismiss these cases. Since blacks are predominantly victimized by other blacks (U.S. Department of Justice, 1981a), such a practice would work to the advantage of black defendants (although it would constitute an important form of racial discrimination). Even if judges then discriminate against black defendants in sentencing, com- mitting higher proportions of them to prison or imposing longer terms, the proportion of blacks in prison could equal the proportion at arrest. Alternatively, if prosecutors were more likely to pursue cases against black defendants, it would increase the proportion of blacks among defendants who are prosecuted and convicted. If judges then sentenced convicted blacks more leniently than convicted whites, that could also leave the proportion of convicted blacks in prison the same as at arrest. It is also possible that the disproportionate numbers of blacks who are arrested might result from police arresting blacks on weaker evidence than they require for whites. If prosecutors dismiss the weaker cases (which would be found predominantly among black arrestees) but blacks are subject to discrimination at sentencing, the total effect of discrim-

92 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM ination at arrest and at sentencing could still leave race-specific arrest and imprisonment rates in correspondence, thus masking both forms of discrimination. Future research on discrimination in sentencing should compare black-to-white ratios by type of crime at each of the inter- mediate stages of the criminal justice system between arrest and prison. Offender-based transaction statistics systems are particularly appropri- ate for such analysis. It is possible that black offenders are sentenced both more severely and more leniently than white offenders and are more vulnerable to diverse racial attitudes among judges. In other words, there may be greater variation in the sentencing of minority group offenders than in that for white offenders. As a result, the black prison population could be in the same proportion as found in the arrest population by offense type, but those in prison still could have been treated disproportionately more severely than comparable white offenders, even though this effect was offset in the aggregate by the more lenient treatment given to other black defendants who received nonincarcerative sentences. In enumerating these possibilities, we have suggested that race may be taken into account in ways that either advantage or disadvantage defendants who are black or members of other minority groups. We cannot yet say how much of the similarity in the proportion of blacks arrested and blacks imprisoned reflects racial neutrality and how much of it reflects the net result of offsetting effects. Aggregate data cannot reveal such differences. The variety of possibilities for offsetting rela- tionships that might be obscured by aggregate data underscores the need for careful, disaggregated research on racial effects for individual crime types at different stages of the criminal justice system and within indi- vidual jurisdictions. Our overall assessment of the available research suggests that factors other than racial discrimination in the sentencing process account for most of the disproportionate representation of black males in U.S. pris- ons, although discrimination in sentencing may play a more important role in some regions, jurisdictions, crime types, or the decisions of individual participants. We also note, however, that even a small amount of racial discrimi- nation is a matter that needs to be taken very seriously, both on general normative grounds and because small effects in the aggregate can imply unacceptable deprivations for large numbers of people. Thus even though the effect of race in sentencing may be small compared to that of other factors, such differences are important. Whatever explains the disproportion of blacks in our prisons, the existence of this disproportion remains a significant matter of concern.

Determinants of Sentences When over 3 percent of all black males in their twenties are in state prisons on any day in this country, with approximately another 1.5 percent in federal prisons and local jails (Blumstein, 1982), we face a social problem of serious proportions that cannot be ignored. The ex- istence of the disproportion has already raised serious questions about the legitimacy of criminal justice institutions. Therefore, correctly iden- tifying the sources of the disproportionality is crucial to the quest for effective solutions. 93 Studies of Sentencing The second type of evidence on racial discrimination derives from studies of the process of sentencing itself. The role of race in sentencing has been extensively studied with uneven quality and varied results- see Table 2-3. Some studies find statistical evidence of racial discrimination; others find none. While there is no evidence of a widespread systematic pattern of discrimination in sentencing, some pockets of discrimination are found for particular judges, particular crime types, and in particular settings. The studies, however, are vulnerable in varying degrees to a variety of statistical problems that temper the strength of these conclu- sions. Many early studies of sentencing including those on capital punish- ment found substantial racial discrimination, with blacks apparently being sentenced more harshly than whites (Table 2-2~. These studies were seriously flawed by statistical biases in the estimates of discrimi- nation arising from: failure to control for prior record, offense serious- ness, and other important variables that affect case disposition. Of the 36 studies using data on sentencing before 1969, only 12 studies have any controls for prior record and offense seriousness (see Table 2-3~. The remaining 24 studies fail to control for one or both of these variables. The absence of controls is especially characteristic of studies on the use of capital punishment. All but 1 of the 15 pre-1969 capital punishment studies fail to control for prior record of the offender, a potentially important factor in choosing between life in prison and the death sen- tence and also in commuting death sentences. They also fail to go beyond crude controls for offense type to even distinguish between homicide cases that are eligible for capital punishment and those that are not. To the extent that race is associated with offense seriousness or prior record, with blacks having more serious offenses or worse prior records, the race variable will pick up some of the effect of these omitted vari- ables, resulting in overestimates of the discrimination effect. It is doubt- ful, however, that the large magnitude of the effect found in these early

94 Cal Cal o ._ Cal o o - o sit o g ._ Ct .E ._ sit U. ._ Ct · Cal Ct o Cal ·= 3 A' ~ .o ,o ,- ~ Pa U) .- o . - A ~ sol V) ~ o Cal I ~ I, o m ~ ~ Ct .= .= ~ .E as Cal .~ . Ct =1:~s Cal .= Cal zo C) O z _1 - ~ __ _ ~ ~ ~ O _ 11 ~ _ - - ~ 11 c: ~ - o ct c~ - - - - ~ - ~ ~ - ~ ~ 11 Ca) Ch ') - ~ - l l ~ ~ o ~ ~ o ~ ~ ~ ~ - - ~ - ~ ~ ~ ~ - - - ~ ~ o ~ 11 c) oo ~ ~ - - ~ - ~a ~ ~ - 11 ~ - ~ - ct c~ - ~ v: O a' z ~ s: e ~n : o ~ v' o o o ~ ~ 11 ~ ~ 11 ~ ~ - o ~ - ~ ~ ~ ~ ~ ~ - ~ ll ~ ~ ~ g o c)

9s - ~ ~ ~ ct ~ ~ x ;^ - ~ - ~ ~ o so ~ 30 C, <t _ 't , ~ Ct ~ ~ V ~ U. ° O C—) _ · _ _) of ~ ~ ~ ~ to ~ ~ Ct Do — ~ ~ Ct ox . _ . _ Pa I- Ct o Cal c r~ 3 - . C c m ~ m i_ ~ ~ ~ I - . ~ ~ . _ ~ . _ _ , _ A_ m ~ O^ 3 ~ ::: . of_ ;>- o ~ ~ ~ ~ ^ c) -], ' o ~4 ~ - i o C: - ~ ^ ~ ~ a — _ ~ _ ~ 0 0 4 - ~ _ 0 o>^ ~ ~ ~ ~ Ct ~ ~ _ ') 00 _ ~ 3 ~ ^— _ _ c_> 50 ~ ~ ~ ~ oOo . c.>~cq~= o~ ,~, ~ ~ ~ "o ~ ~ ~ _ O ~ O _ ~ ~ _ _ ~ . _ Ct V~ Ir) Ct C~ a~ ~ 3 — c ~ ~ ~ ~ c-) o ~ ~ '~' ;;~` ~ ct: ~ ~ o : =- ~ o ON 't ctl C ~: ~ C ~ ,, ~ c ~ 3 E _ c4—~ c ~ -~ ~ -~ .3 r~ ~ O _ . ~ ~ 3= ~ h_ ~_ o ._ au _ ~ :s o ~_ oo ~ _ _ _ V~ ~ — _ _1 _ ._ - o .e m Ct o S ~ o v, J Ct o = C) C o C~ 0- ~0 - Ct 7 ~ ~ ~ . —00 __ ,~ O4 Ct V' - \ e~ ~ D ·~ , ._ C~ r_ O ^ ~ ~ 0N _ ~D Cr~ , =, Ce _ ~ _ _ ,,,, . _ 1.= ~ ~ a, ~ O V ~_ Ct · C) Ct ._ C oc ._ Ct Ct Ct ._ .m O V: V' Ct . _ ~ ._ 4_ ·S:4 .' 0 5 ~ ,~ ~ ~ O O D O1) .' ~ ~ ,~ ~S .= ~ .= ~ O O ._ ._ V' ~— U) ~ Ct ._ ,,- ~ ~ O (L) ·— O t~ .m . ~ , ~ . Ct ~ ~ t ; - _ ~ Ct Ct ~ V =: ~ _ a' o ~ ·— · ~ ~ := ~ 3 Ct O ~

96 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM studies would be completely eliminated by the introduction of appro- priate controls, and some portion of the estimated race effect may indeed reflect discrimination in sentencing for some crimes in some areas ex- tensively studied, particularly for capital punishment in the South in the 1940s, l950s, and 1960s. More recent studies using a richer set of controls have yielded varied results, with some finding evidence of racial discrimination while others do not. As indicated in Table 2-3, the introduction of controls for offense seriousness and prior record reduces the widespread finding of racial discrimination in sentencing, especially in studies using pre-1969 data. Nevertheless, discrimination continues to be found in specific contexts in more recent studies, particularly in rural courts, for selected crime types, when the victim is white, or for some judges in a jurisdiction. Even in these contexts, however, offense seriousness and prior record remain the dominant factors in sentence outcomes (Hagan and Bumiller, Volume II). Despite substantial improvements in research in addressing the prob- lem of omitted variables, recent studies are still subject to potential biases arising from measurement error and sample selection. These biases arise from the use of incomplete measures reflecting offense seriousness and prior record, which fail to adequately control for the role of un- measured elements of seriousness or record in distinguishing the sen- tences of whites and blacks. In addition to biasing the estimates of the effects of seriousness and record on sentence, failure to adequately measure important elements of seriousness or record can also contam- inate estimates of the effects of other correctly measured variables, like race. This occurs because only a portion of the true effect of seriousness, for example, is captured in the estimated effects of the included ele- ments. Some part of the true effect is "picked up" by other correctly measured variables that are associated with the excluded elements of seriousness. i2 Considering seriousness and record as single-score variables, each formed from a linear combination of contributing factors, the biases of interest arise from measurement errors in seriousness or record. When only one variable is measured with error, the direction of the bias in a i2 The contamination or "smearing" effect is discussed in more detail in the context of measurement error in a variable in Garber et al. (Volume IT) and Garber and Klepper (1980~. For further treatments of the case of a single variable measured with error, see Aigner (1974), Blomqvist (1972), Chow (1957), Levi (1973), McCallum (1972), and Wick- ens (1972~.

Determinants of Sentences 97 correctly measured variable depends on the bias in the incorrectly mea- sured variable and the nature of the association between these variables. As illustrated in Table 2 - , when a variable like race is measured cor- rectly and race is related to the mismeasured variable of seriousness, with blacks committing more serious offenses, there are opposite biases in seriousness and race. When the effect of seriousness is underesti- mated, the discrimination effect is overestimated, and vice versa. On the other hand, when whites commit more serious offenses the biases in race and seriousness are in the same direction. Similar arguments would apply to the incorrectly measured variable of prior record. A number of studies have found associations of race with offense seriousness and prior record. For offense seriousness, blacks have been found to be substantially overrepresented in more serious offenses, par- ticularly in violent crimes. This relationship was first noted in analyses of official data on arrests (Mulvihill et al., 1969; Wolfgang and Ferracuti, 1967; Wolfgang et al., 1972~. The role of race in offense seriousness is illustrated in Table 2-5. The ratios of black to white arrest rates are highest for robbery (with black adult rates 9.80 times white adult rates) and for serious violent crimes (with black adult rates 6.12 times white adult rates) and much lower for less serious, nonindex offenses (with black adult rates only 2.38 times white adult rates). The same difference is also found in analyses of self-reported crime. While self-report meas- ures of total criminal involvement find little difference by race, exam- ination of self-reports disaggregated by crime type indicate progressively greater involvement of blacks as offense seriousness increases, especially in cases of violent offenses (Hindelang et al., 1979~. Direct evidence of a relationship of race with offense seriousness is also reported in studies examining sentence outcomes (Arkin, 1980; Gibson, 1978b; Spohn et al., 1982~. Further indirect evidence of this relationship is found in Table 2-3: the role of race in influencing sentence severity is reduced when controls for seriousness and prior record are added to analyses, with 77 percent of the studies without controls and only 45 percent of those with controls finding discrimination in sen- tences. A similar reduction in effect within the same data set is reported in Burke and Turk (1975), Clarke and Koch (1976), and Spohn et al. (1982~. Evidence for a relationship between prior record and race has been reported in several studies. In accounting for the large differences in sentences of whites and blacks convicted in Philadephia, Green (1961, 1964) found that, controlling for current conviction charge, there were pronounced racial differences in prior criminal records of convicted offenders. The differences in sentences by race were consistent with

98 . . a' A_ Ct o so ¢ - . ~t ,,, 5 - ~ o C) Cal a.) Cal o ~ Ct Ct Cal :, so ~ 8 ~ c: ~ A_ ~ C) U. ~ Ct i_ . _~ me O ~ .-o 3 a' ~ V, Cal lo 1 ~ o - ~ o m ~ ;^ ~2 Cal S An in SO Ct U. hi: Ct - .0 US .= .£ . ~ lo: _` > V, o :: . - C~ in o o . ~ Cal s ~ 1 .o ~ cq Cq c~ ~ v, o . - 5) c~ o ct . c) ~ ~ ~ ~ o o · ~ - 'e .= .- .- c~ a~ .= o 'e m ~ Ct O O .- Ct .~ m . . CD o .Ct m O ~ CD ._ S~ . C~ o o . - o C~ .= :e o ~ o . ~ _ Ce ~ O LU o 7 ~ ~ m _ _~ ~ C) _ ~ ~ 3 ° ° C: _ ~ ~ ~n ~ ~o E~ o ~ ~ ~i ~ o :t ~ ~ ~ 0 =-c ~ Om 3~o E~ ~ c~ z f_ L~ Z c,, Z ~7 O ~: O ~: O ~ _ . ~ ~ 0 ,= =~ E ~v ~ tn · ~ ° ~ 3 ~ ~ ce ~ C ~ — . C 0 ~ ~ ~ ~ ._ ~ ~ ~ oc ~ .~° ,= ~ =, 3 0 ~ ._ ~ ~0 ct ~ 0C 0 00 — x ~ c' ._ y E ;^ ~ ~ .E ~ c'

99 O C c) ~ C~~ C~ ct ~ C As C ~ - O ~ A a . c O O D C - 0 a ° D ~ · ° ~ D · E c O w e s c e O. > c,) C ,~ .L) o ~ ';,) =~) ~ ~ 3 0 s, c c a ,, 9, ~ · ~ . c ~ ~ ~ ~ 2 At c' ~ =' E O c' i, _ ~ ~ ~ _ ~ _ ~ O ~ O ~ ~ ._ O - , ._ ~ ._ C) ~ O ~ O ~ ~ .> on u' 0 ~ so cat ~ cat ct cat

Crime Type 100 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM TABLE 2-5 Comparison of Black Arrest Rates With White Arrest Rates (Arrests per Population) by Age and Crime Type in 1970 for U.S. Cities (Black Arrest Rate/ White Arrest Rate) Juveniles Adults Serious violent 4.84 6.12 Murder 5.87 8.32 Rape 5.07 6.23 Aggravated assault 4.75 5.87 Robberya 9.07 9.80 Serious property 2.46 3.65 Burglary 2.56 4.10 Larceny 2.47 3.37 Auto theft 2.27 4.43 Nonindex 1.61 2.38 Forgery, fraud, embezzlement, stolen 2.35 3.14 property, arson Simple assault, weapons, vandalism 2.46 3.84 Narcotics .57 2.06 Prostitution, other sex offenses, gam- 1.05 4.43 bring, liquor law violations Other (excluding traffic and juvenile 1.62 2.11 offenses) NOTE: Arrest rates are derived from data on arrests by age, race, and crime type reported to the Federal Bureau of Investigation's uniform crime reporting program for 55 U. S. cities with populations of 250,000 or more in 1970 and from the 1970 census of populations by age and race in those cities. The arrest data for individual cities were provided by the Federal Bureau of Investigation. The ratios of black to white rates are based on the mean arrest rates for the 55 cities. a Robbery is usually treated as one of the serious violent crimes, but because it is different from other violent crimes, it is treated separately in this table. differences in prior record, with blacks generally having more serious prior records than whites. A similar difference in prior record was found in Gibson (1978b) and Spohn et al. (1982~. In Burke and Turk (1975) the relationship between race and prior record involves an interaction with age. Nonwhites under age 35 were more likely to have prior in- carcerations than whites in the same age group; the relationship was reversed for offenders 35 years old or over. Further indirect evidence of the relationship of prior record and race is again provided in Table

Determinants of Sentences 101 2-3: the role of race is reduced when controls for offense seriousness and prior record are included. The observed association of race with offense seriousness might arise from differential involvement in different offense types for different races or from differential treatment through the exercise of victim or police discretion in differentially reporting offenses or in the process of investigating, arresting, and charging defendants. While not conclusive, Hindelang and associates (Hindelang, 1976, 1978; Hindelang et al., 1979) present a variety of evidence from official arrest data, self-reports of crime, and victimization surveys supporting the differential involve- ment hypotheses. Similarly, in the case of prior record, the association may reflect real behavioral differences in the intensity of offending or may result from differential treatment, particularly for a first offense, which then increases the likelihood of accumulating a prior record. There is some evidence to support this latter hypothesis of differential treat- ment resulting in differential accumulation of prior record by race (Chir- icos et al., 1972; Tiffany et al., 1975~. However, this may result from more serious first offenses for blacks than for whites. Some have argued that racial discrimination in sentencing reflects a response to the combination of the offender's and victim's race. Under a presumption of racial discrimination, one might expect that offenses by blacks against white victims would be sentenced more harshly than similar offenses of whites against whites, whites against blacks, or blacks against blacks. This might occur because black victims are regarded as less important than white victims or because offenses across racial lines by blacks are viewed very seriously. When such factors have been ex- plicitly considered in analyses, the empirical results strongly support the expected differences in sentences for various race combinations of of- fenders and victims. i3 Ten of 14 studies including 7 on the use of capital punishment find that black offenders against white victims are sen- tenced more harshly than other race combinations (Bowers and Pierce, 1980; Florida Civil Liberties Union, 1964; Garfinkel, 1949; Howard, 1967; Johnson, 1941; LaFree, 1980; Partington, 1965; Southern Regional Council, 1969; Wolfgang and Reidel, 1973; Zimring et al., 1976~. As noted in Kleck (1981), these studies are also subject to biases resulting from unmeasured aspects of offense seriousness. Aside from the obvious race differences, Kleck (1981) notes that interracial offenses are also more likely to involve strangers, more likely to involve other i3 Because of insufficient cases, there are no studies that separately examine sentence outcomes for white offenders against black victims.

102 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM TABLE 2~ Disposition of Felony Arrests Jurisdiction Washington D.C. California New York City Disposition 1973a l979b l979c Felony arrests that result in felony conviction 13 percent 18 percent 12 percent Felony arrests that result in any conviction 29 percent 20 percents Not available Convictions sentenced to prison 32 percent 25 percents 31 percent a Forst et al. (1977~. b California Department of Justice (1980~. c Chambers (1981~. Superior court convictions only. felonies, and less likely to involve victim provocation. These character- istics of interracial offenses are all factors contributing to increased seriousness of the offense and presumably also to more severe sentences. Failure to measure and include these important dimensions of serious- ness would lead to biased estimates of the race effect. The 10 studies finding an effect for offender and victim race either fail to include or only partially control for these dimensions of offense seriousness. Four other studies that do control for factors associated with interracial of- fenses do not find any effect on sentence for offender and victim race (Farrell and Swigert, 1978b; Green, 1964; Judson et al., 1969; Myers, 19794. The suppression of the estimated discrimination effect when con- trols for these other elements of offense seriousness are included suggests that the biases in the offender/victim race effect are likely to be domi- nated by overestimates.~4 The estimated race effect may also be biased by sample selection. The processing of criminal cases through the various stages in the criminal justice system is like a sequence of filters, screening cases from the system according to various criteria related to case attributes. As indi- cated in Table 2-6, only 13 of every 100 felony arrests in Washington, D.C., in 1973 resulted in felony convictions, while another 16 resulted i4 Unfortunately, the general lack of data for interracial offenses involving whites against black victims does not permit evaluating whether the particular race of the victim in interracial crimes is important, independent of other considerations like greater involve- ment of strangers, of other felonies, and of victim provocation in interracial crimes.

Determinants of Sentences 103 in misdemeanor convictions. Of those convicted, 32 percent were sen- tenced to prison. These experiences in Washington, D.C., are typical of other U.S. jurisdictions. As a result, the cases ultimately available for sentencing are a selected sample, including only a fraction of the population of "similar" offenses originally committed. Sample selection of this sort poses problems to the generalizability of results. Offenders who are ultimately convicted or incarcerated are likely to differ in important ways from the original population of offenders. This threat to the generalizability of the results is generally well under- stood, and findings from studies using selected samples are usually prop- erly restricted to an appropriately limited population. It is less well understood, however, that sample selection can also pose serious threats to the validity of statistical results even within the selected sample. In the case of sentencing, internal selection biases can arise when unobserved and thus unmeasured factors are common to both the selection and sentence processes, thereby inducing (or altering) correlations in the selected samples between the unmeasured variables and other included variables that are also common to selection and sentencing. 15 Examples of the process giving rise to selection biases are presented in Table 2-7. In that table, we consider separately cases in which pros- ecutor aggressiveness and elements of offense seriousness are unmea- sured factors in both selection and sentencing. For prosecutor aggres- siveness, there would be no bias in the estimated effects if there were no sample selection; the sample selection process, however, induces bias in the selected sample. For the unmeasured element of offense seri- ousness, on the other hand, there is already bias in the estimated effects resulting from measurement error alone; this bias, however, is reversed by sample selection. As illustrated in the first column of Table 2-7, selection biases can arise even when there is no correlation between the unmeasured and measured variables in the original population. In this example, prose- cutor aggressiveness is assumed to be an unmeasured factor both in selection and in more severe sentence outcomes. Since cases are ran- domly assigned to prosecutors, there is no correlation between unmea- sured prosecutor aggressiveness and other measured case attributes. In is See Klepper et al. (Volume II) for a detailed discussion of the role of sample selection biases in research on discrimination in sentencing. For more general treatments of sample selection biases, see Berk and Ray (1982), Goldberger (1981), Heckman (1976, 1979), Olsen (1980), and Tobin (1958~.

104 Cal C) A_ Cat o Cal A_ Ct c 4 - 4— o U. Cal ._ ._ Cal Cal Cal ._ m o .~ as a) o :, Ct m ¢ o U. CD o . cat Ct o C) a' Cal . CD ~ a' · CD S it> Ct m' .. it O O A O C) Ct ~ ~ .= 0 ._ ~ ~ .= a' ~ I: ~ O 0 ~ ~ Cal ~ O ·- ~ Cal Cal .= O in ~ c<S — ~~ O m ~ ~ ~L) V, O U' ~ ~ o ~° ~ ~ E A 3 O V ~ X0m o ~ ·-o o ~ cn c~ C,7 o ~ 4_ · - ~ c,' o ~ p~ ¢ o u' a' ~; ~ ~ c .o- o ~ C~? U v) o ~ ~ cn u, o0 so~ oD ¢ v, ~ ~ ac) .= a, ~ v) c) · - o c c4~f ~ o ~ o ~ o r~ o ~ 'e ~ ~ ·O ~ (v =O c~ ~ ~ o ~ o ~ ~ o ~ ~ ~ o o ~ · ;;~` r ~ 'v - . ~ .= ~ - ~ ~ .o c st~o~ . t~~ I-~ cD ~o ~ o ~ ~ ~ c ·i~ ~ ;> cO "c ~ c ~L) ~L ~ E m c' ° ^ce 010 ;^ - ~ ~ ~ ~ ° V)~a ~ O ~ ~ ~ 4- ~ ~ ', ~ ~ c~ ~: au 3 't ~ ~_ =: V, o . _ ,— Ct ~ _ Ct ~ G) o ~ . _ ~ ~: ._ s~ ~ .° _ Ct — . _ ,— ~ ~4 C) — ,= ~ , - . _ ~ ._ _ ~ ._ o ~ o o d_ ~ ~ C) a' o o ~ o · _ ~ o oo ~ . _ _

105 , ~ . ~ .2, ~ 6, ~ ~ ~ ~ ~ .. it, ~ ~ . ~ ,= oh ~ ~ 2 ~ ~ ~ ~ He g ' ~ ~ ~ ~ l~ ~ ~ ~ ~ b 2 " ~ ~ ~ ~ ~ ~ 3 -a ~ ~ } ' ' ~ ' g e ~ g ' ~ ~ 0 ~ ~ ~ ' ~ ~ ' ~ .= ~ ~ C ~ ~ ~ To o ~ .= O ~ ~ ~ , ~ e v :, ~ ~ ~ i,

106 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM this event, if there were no selection, failure to include prosecutor ag- gressiveness would not bias the estimates of the included determinants of sentence outcomes. The selection process, however, operates so that those cases with more aggressive prosecutors are more likely to have charges brought, less likely to be dismissed, and more likely to result in convictions and be available for sentencing. In the presence of racial discrimination, with blacks also more likely to be selected, those whites who are selected are more likely to have aggressive prosecutors than are the selected blacks. Selection thus induces a correlation between race and prosecutor aggressiveness in the selected sample. When pros- ecutor aggressiveness is left unmeasured, some of its effect on more severe sentence outcomes will be picked up by the selected whites, thus diminishing, or underestimating, the effect of any discrimination against blacks in sentencing. Considering the second column in Table 2-7, there is already the potential for biased estimates of the discrimination effect arising from the correlation between the correctly measured race variable and the incorrectly measured offense seriousness variable in the original pop- ulation. In this case, however, the biases arising from measurement error are confounded by sample selection. Selection occurs if more serious offenses are more likely to be prosecuted, less likely to be dis- missed, and more likely to be sentenced severely. However, when of- fense seriousness is not measured completely, the differences in seri- ousness cannot be fully controlled. Despite the likely role of factors like weapon use and offender-victim relationship in assessments of serious- ness by criminal justice decision makers, these factors may not be mea- sured and included in research on sentencing. Selection biases associated with this measurement error will arise if, in addition to considering seriousness as a basis for selection and sen- tencing, there is also racial discrimination throughout criminal justice processing for example, with blacks more likely than whites to be charged, less likely to have their cases dismissed, and more likely to be sentenced severely regardless of offense seriousness. The whites who are selected, then, are likely to have committed more serious offenses than selected blacks. (Note that the selection process has reversed the original correlation found between race and seriousness.) However, because of errors in measuring seriousness, only differences in observed seriousness can be measured and included. Selected whites who are identical to selected blacks on observed seriousness are still likely to have committed more serious offenses on unobserved dimensions of seriousness. This correlation between correctly measured race and in- correctly measured seriousness in the selected sample results in biases

Determinants of Sentences 107 in the estimated effects of both race and seriousness on sentence out- comes. As indicated in Tables 2 - and 2-7, independent; measurement error in offense seriousness results in underestimates of the effect of serious- ness on sentence. When selection operates to select more serious of- fenses and when there is prior racial discrimination, with blacks being more likely to be selected, as in the example above, whites who are selected would be likely to have committed more serious offenses. In this case, some of the unmeasured effect of seriousness on sentence would be picked up by selected whites with their more serious offenses, thus diminishing, or underestimating, the effect of any discrimination effect against blacks in sentencing. Selection bias arising from measurement error in offense seriousness may also operate to exaggerate, or overstate, the actual level of dis- crimination against black offenders in sentencing. Consider, for exam- ple, the situation when more serious offenses are selected, but whites are now more likely to be selected. This might arise if there were dis- crimination against black victims in prosecution decisions in which vic- timization of blacks is treated less seriously by criminal justice decision makers, resulting in higher proportions of dismissals or charge reduc- tions. Since blacks are overwhelmingly victimized by blacks, black offenders would be less likely to be selected for further processing. Due to the greater likelihood in this situation that whites are selected re- gardless of seriousness, the offenses of blacks who are selected are likely to be more serious on both observed and unobserved dimensions. Once again independent measurement error in offense seriousness would lead to underestimates of the effect of seriousness on sentences. In the ab- sence of adequate controls for unobserved differences in seriousness, some of the contribution of more serious offenses by selected blacks to sentences would mistakenly be attributed to race, thus exaggerating, or overestimating, the effect of discrimination against blacks in sentencing. The exact nature of the errors in estimates of the effect of racial discrimination at sentencing, arising from any selection bias associated with measurement error in offense seriousness, depends critically on both the direction and magnitude of the contribution of seriousness and discrimination in prior selection processes. Thus, resolving the ambiguity i6 The 1979 National Victimization Survey (U.S. Department of Justice, 1981a) reports that, for personal crimes of violence, 84 percent of victimizations of blacks by single offenders involved black offenders (Table 43~; for black victimizations by multiple of- fenders, 72 percent involved all black offenders (Table 47~.

108 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM about the determinants of sentences requires empirical research to iden- tify more clearly the determinants of earlier selection in processing cases through the criminal justice system. There is some evidence suggesting the existence of racial differences in treatment at earlier processing stages. For example, some evidence suggests that differences in sentencing outcomes may arise through racial differences associated with attorney type and employment status of the defendant, which in turn affect ability to post bail (Clarke and Koch, 1976; Farrell and Swigert, 1978a; Lizotte, 1978; Spohn et al., 1982~. Each of these factors then affects the likelihood of conviction and the vulnerability to sentence. Race may also enter through its role in sen- tence recommendations by probation officers and prosecutors (Hagan, 1975, 1977; Hagan et al., 1979; Myers, 1979; Unnever et al., 1980~. While these results are suggestive, considerably more research is re- quired on the determinants of prior decisions affecting arrest, charges filed, dismissal, bail release, plea-bargain offers, sentence recommen- dations, and the like. In providing estimates of racial discrimination prior to sentencing, such results will also help to clarify the role of sample selection biases in estimates of discrimination at sentencing. Measurement Errors and Their Consequences Some measurement error is present in all statistical analyses of sen- tencing. The crucial question is how much of the estimated effect of correctly measured variables is real and how much is statistical bias. For independent errors in mismeasured variables, the bias in the estimate of an associated correctly measured variable, like race, will be larger relative to its true effect (Gerber et al., Volume II; Garber and Klepper, 1980~: 1. the greater the fraction of the variation in sentence outcomes at- tributable to incorrectly measured variables like offense seriousness and prior record; 2. the smaller the fraction of the variation in sentence outcomes at- tributable to the correctly measured variable, like race; 3. the greater the correlation between the correctly measured vari- ables and the incorrectly measured var~ables;~7 or i7 In the case of induced correlations in selected samples, the greater the fraction of variation in selection that is attributable to the correctly measured variable, race, the larger the correlation in the selected sample.

Determinants of Sentences 109 4. the greater the fraction of the independent variation in the incor- rectly measured variables (after controlling for other explanatory vari- ables) attributable to the measurement errors. The evidence suggests a primary role for the incorrectly measured variables of offense seriousness and prior record in influencing sentence outcomes and a nontrivial relationship between these mismeasured var- iables and race. Conditions 1 and 2 above suggest that the bias in the effect of a presumably correctly measured offender attribute like race will be larger when race actually plays a small role in determining sen- tences relative to the role of the incorrectly measured variables of offense seriousness and prior record. Under condition 3, the correlation between race and the incorrectly measured variables also contributes to a larger bias. Furthermore, to the extent that these incorrectly measured vari- ables are in fact primary determinants of sentences, conditions 1 and 4 suggest that the bias in the correctly measured variables is large when the primary determinants of sentence are measured with considerable error. Thus, the possibility of nontrivial correlations of race with other poorly measured but key variables like offense seriousness and prior record raises the threat of serious biases in the estimates of discrimi- nation effects. Further complications are introduced by the possibility that the cor- relations vary with the selection process and by crime type or jurisdic- tion. In this event, the statistical biases attributable to measurement error may be critical in some cases and trivial in others. The biases may even work in opposite directions in different studies. This suggests that measurement error bias, operating either directly or through sample selection, could substantially obscure the true incidence of discrimina- tion in sentencing. The biases in the estimates of the effects of racial discrimination in sentencing discussed above result principally from inadequate measures of key aspects of offense seriousness and prior record. One obvious remedy to this problem is to obtain improved measures of these variables in order to more fully and adequately reflect the richness of factors taken into consideration in sentencing decisions. To address the problem of selection biases more generally, analyses must be extended beyond sentencing to include examination of the selection processes as well. i~ Berk and Ray (1982) summarize a variety of available estimation procedures that correct for selection biases when there is no correlation between unmeasured and measured factors in the original population.

110 RESEARCH ON SENTENCING THE SEARCH FOR REFORM This broader approach to analyzing the determinants of sentences also has the potential of indirectly resolving the measurement error problems in key concepts like offense seriousness and prior record without re- quiring explicit measures of currently unavailable and difficult-to-mea- sure variables.~9 DISCRIMINATION BY SOCIOECONOMIC STATUS In addition to being disproportionately black, state prison inmates are disproportionately poor and unemployed and otherwise rank low on measures of socioeconomic status. In 1979, 41 percent of state prison inmates who had been admitted to prison after November 19772° had either no income (22.2 percent) or incomes of less than $3,000 (19.2 percent) in the 12 months prior to arrest. Of those with incomes, the median income was $6,66~much lower than the 1979 national median income for males of $10,972.2i The unemployment rate prior to incar- ceration for state prison inmates was 16.5 percent, compared to an average male unemployment rate adjusted for race and year of 7.8 percent for the decade of the 1970s.22 In a study of prison inmates in three southern states, Chiricos and Waldo (1975) report that inmates are overwhelmingly characterized by low scores on a status measure that combines income, occupation, and education factors. The evidence of discrimination on grounds of social or economic status is, however, equivocal. Like research on racial discrimination, this much smaller body of research is characterized by inconsistent results. Some studies find discrimination by status (Clarke and Koch, 1976; Farrell, 1971; Farrell and Swigert, 1978b; Judson et al., 1969; Lizotte, 1978; i9 With a system of equations that includes common latent (i.e., unobserved) variables in several equations, the effects of the unobserved latent variables can be estimated from common movements observed in multiple outcome variables; see Garber et al. (Volume II) for a fuller discussion of this result. 20 The inclusion only of inmates admitted after November 1977 is to avoid the inflation factor in reported incomes over time. 21 The data on income are from the 1979 Survey of Inmates of State Correctional Facilities, as reported by the U.S. Department of Justice (1982b). 22 Based on the data on prearrest employment available from the U.S. Department of Justice (1982b), 84.1 percent of inmates were in the labor force, resulting in an unem- ployment rate of 16.5 percent (13.9/84.1~. The comparable noninmate unemployment rate for males is calculated by first weighting the annual unemployment rates available from the U.S. Department of Labor (1980:62) by the racial distribution found in prison; the resulting annual rates during the 1970s are then weighted by the distribution of inmates by time served in 1979.

Determinants of Sentences 111 Thornberry, 1973), and others do not (Burke and Turk, 1975; Chiricos and Waldo, 1975; Nagel, 1969; Terry, 1967; Willick et al., 1975~. Other studies report a mediating role, with status variables affecting sentence outcome indirectly through their effect on initial charge (Hagan, 1975) or on the conviction charge (Swigert and Farrell, 1977~. The research on the effects of socioeconomic status is subject to the same methodological difficulties that apply to race. In some cases im- portant control variables are omitted entirely (Bedau, 1964, 1965~; in others, incomplete measurement of important dimensions of offense seriousness and prior record contributes to possible biases in the effect of status arising from the correlation of status variables with the incor- rectly measured variables. A number of studies using official arrest data have noted an associ- ation between socioeconomic status and offense seriousness, with mem- bers of lower status groups substantially overrepresented in arrests for more serious offenses (e.g., Braithwaite, 1981; Gordon, 1976; Reiss and Rhodes, 1961; Shaw and McKay, 1942~. When adequate controls for offense seriousness are taken into account, similar differences in of- fending are also found in studies using self-report data (Braithwaite, 1981; Elliott and Ageton, 1980; Hindelang et al., 1979~. Evidence of such a relationship is also reported in studies examining sentencing outcomes (e.g., Hagan, 1975~. A similar relationship is found between status and prior record, with offenders of lower status more likely to have prior convictions (Willick et al., 1975) or prior incarcerations (Burke and Turk, 1975~. When such correlations are combined with errors in measuring offense seriousness or prior record, or with a failure to include these variables in the analysis, the estimates of the effect of status on sentence outcomes are vulnerable to the same serious biases that plague results on racial discrimination. Additional problems arise from the uncertainty over how best to measure social or economic status. Socioeconomic status is a complex variable reflecting an individual's location in a social structure. Different positions are presumed to be associated with characteristic sets of beliefs, attitudes, and expected ways of behaving that not only influence the behavior of individuals in those positions, but also the expectations that others have about people of different status. Status thus links a set of attitudes or beliefs with behavior; the question is how best to charac- terize that link. For example, there is considerable uncertainty over the relative importance of different aspects of status, such as education, income, and occupation. It is also unclear whether status-linked behavior is principally influenced by experiences in formative years and thus by one's parents' status, or by one's own status, or by one's anticipated or

112 RESEARCH ON SENTENCING THE SEARCH FOR REFORM desired status. This ambiguity is reflected in research on sentencing in which socioeconomic status is variously measured in terms of father's occupation (Terry, 1967), own occupation (Hagan, 1975; Judson et al., 1969; Lizotte, 1978), occupational prestige (Burke and Turk, 1975; Far- rell and Swigert, 1978b; Swigert and Farrell, 1977~; income (Clarke and Koch, 1976; Nagel, 1969; Thornberry, 1973~; and a scale combining income, education, and occupation (Chiricos and Waldo, 1975; Willick et al., 1975~. The resulting likely measurement error in the status var- iable contributes to biases in the estimates of the effect of a defendant's status on sentence. Moreover, research on this subject is hampered by the relative lack of variation in socioeconomic status among defendants charged with similar offenses. Research in this area is thus best pursued by focusing on those crime types with the most variation or through experimental studies of sentencing. Even if the available estimates of the effect of status on sentence were unbiased, a finding of discrimination by status would depend on the legitimacy of specific measures of status as determinants of sentences, and at this time there is considerable debate about the legitimacy of some socioeconomic components in sentence decisions. For example, indicators like employment or education may be valuable as predictors of criminal recidivism and thus may be considered legitimate factors in determining sentences. For this reason, employment history and edu- cational attainment were for several years explicitly included in the U.S. Parole Commission's guidelines. Alternatively, the strong association of status variables with variables like race or wealth, which are more une- quivocally illegitimate, raises questions about the legitimacy of using any variables that embody race or wealth effects as factors in sentencing. For these reasons, the Minnesota sentencing guidelines explicitly ex- clude status variables from judicial consideration at sentencing. Reflect- ing similar concerns about legitimacy, educational attainment has also been removed from the federal parole guidelines. Thus, even if empirical questions regarding the influence of status on sentence were resolved, conclusions about the discriminatory nature of these variables would depend on resolution of the normative questions involved. DISCRIMINATION BY SEX While the disproportionality of blacks in prison is large compared to their representation in the general population, the disproportionality of men is enormous, with women accounting for 52 percent of the adult (over age 18) population but only 4 percent of state prison populations

Determinants of Sentences 113 in 1979 (U.S. Department of Justice, 1982b). As with blacks, however, the disproportionality found in prison populations is mirrored in arrests. Women accounted for 20.5 percent of adult arrests for index offenses in 1979, but they accounted for only 8.7 percent of adult arrests for the violent offenses of murder and robbery that are most often found in prison. Larceny accounted for a large proportion 79 percent adult index arrests of women in 1979 (although women accounted for only 32.7 percent of all adult arrests for larceny). Despite the apparently large differences in the criminal activity and imprisonment rates of men and women, sex differences in sentence outcomes have not generated a large volume of research. A recent review of this body of research found only about 20 studies since 1970 in which sex of the offender was a consideration (Nagel and Hagan, 1983~. This small body of research is noteworthy for its consideration of the impact of sex differences at various stages of case processing, from pretrial release to sentence. No one study, however, considers outcomes at all stages. Based on their review of the literature, Nagel and Hagan (1983) conclude that differences in outcome by sex do exist, particularly in the pretrial release decision on type of release and in the sentence decision, especially for less severe sentence outcomes. When these differences are found, they are to the advantage of women of- fenders. The strength of the conclusions drawn from the existing body of re- search, like those on race and socioeconomic status, must be moderated by the potential biases arising from errors in measuring seriousness and prior record and from possible selection effects resulting from the dif- ferential filtering of cases to the sentencing stage. For example, to the extent that women tend to commit less serious offenses and are also less likely to be selected for sentencing regardless of offense seriousness, those women who end up being sentenced would be likely to have committed more serious offenses. However, when there are independent measurement errors resulting from incomplete measures of seriousness, the unobserved dimensions of seriousness cannot be adequately con- trolled, and some of the effect of seriousness on sentence outcomes would be picked up by sentenced women with their more serious of- fenses. This would diminish understate the true difference in sen- tence outcomes between men and women. Whatever the actual effect of sex on sentence outcomes, the question of discrimination by sex depends on the legitimacy of sex differences as a determinant of sentences. This remains an unresolved legal question: sex has not been granted the status of a "suspect classification" (as has

114 RESEARCH ON SENTENCING THE SEARCH FOR REFORM race). The fact that any sex differences that may exist are to the ad- vantage of the otherwise presumed disadvantaged group also makes sex discrimination in sentencing a somewhat unique problem. To the extent that there is discrimination in sentence outcomes by sex (or by race or by socioeconomic status), a range of "solutions" is available for eliminating that discrimination. If the objective is to equal- ize sentences, one can shift the outcomes of the disadvantaged group to equal those of the advantaged group, or vice versa, or one can shift both groups to achieve some average of past sentencing practices. In California's Uniform Determinate Sentencing Law, the averaging ap- proach was used. However, since women represented such a small por- tion of all sentenced defendants, the effect has been to markedly increase the sentences of women, especially for violent offenses.23 CASE-PROCESSING VARIABLES Three case-processing variables have frequently been cited as potential factors in differential sentence outcomes: mode of disposition (guilty plea, bench trial, or jury trial); pretrial release status (free on bail or detained); and type of attorney (none, court-appointed, or privately retained). The evidence varies in quality and in the consistency of find- ings for each of these factors. Of the three factors, the evidence on the role of guilty pleas in less severe sentences is most convincing. Pretrial detention is commonly found to be associated with more severe sen- tences, but this result is particularly vulnerable to biased estimates and hence is best viewed cautiously. The evidence on the role of attorney type is mixed and does not support a general conclusion that attorney type is independently related to sentence outcomes. The strongest and most persistently found effect of case-processing variables is the role of guilty pleas in producing less severe sentences. It appears in some jurisdictions that defendants who exercise their right to trial receive harsher sentences than similarly situated defendants who plead guilty. Such a sentence differential is sometimes thought to be an essential element of the process by which large numbers of defendants are induced to plead guilty. Evidence for this phenomenon comes from interviews with court par- ticipants (Alschuler, 1968, 1976; Casper, 1972; Heumann, 1978; Mather, 1974; Newman, 1956; Vetri, 1964; Yale Law Journal, 1956) and statis- 23 This effect is discussed in greater detail in the analysis of the Determinate Sentencing Law in Chapter 4. impact of the California

Determinants of Sentences 115 tical analyses of case records in a wide variety of jurisdictions. Several statistical studies report substantial sentence differences by plea when other factors like record and charge are controlled (Brereton and Cas- per, 1982; Nardulli, 1978; Rhodes and Conly, 1981; Rich et al., 1981; Uhlman and Walker, 1980~. One study reports sentence differences by plea in selected courtrooms but no aggregate differences in three juris- dictions (Eisenstein and Jacob, 1977), while another reports sentence differences for some crime types but not others (Rhodes, 19784. The statistical evidence on what is called the guilty-plea discount is subject to possible biases arising from measurement error and sample selection. These potential biases are particularly troubling because they would result in overestimates of the effect of the discount. Several studies have found an association between offense seriousness and mode of disposition, with more serious cases more likely to go to trial (Eisenstein and Jacob, 1977; Hagan, 1975; Klepper et al., Volume II: Table 14. This might occur because of a prosecutor's decreased will- ingness to accept guilty pleas to reduced charges in serious cases and a corresponding decreased willingness by a defendant to plead guilty when the risk of severe sanction is high. To the extent that offense seriousness is poorly measured, independent measurement error would contribute to underestimates of the effect of seriousness and overestimates of the effect of trial on severe sentences. This measurement error bias will be large relative to the true effect of guilty pleas when: offense seriousness in sentence plays a large role; the role of disposition type in sentences is small; the error in measuring seriousness is large; or the correlation between seriousness and dispo- sition type is large. Thus, measurement error bias from an association between disposition type and offense seriousness could lead to estimates of an effect of disposition type when in fact there is none. However, the interview data from court participants suggest that this statistical bias is likely to be small relative to the true effect. To begin with, the views of participants are informed by direct knowledge of the relative influence of dimensions of seriousness that may be unobservable to the researcher. Moreover, as participants in the plea negotiation process, judges, prosecutors, and defense counsel are privy to the offers made to defendants who go to trial; they thus have firsthand knowledge of the size of the guilty-plea discount reflected in the actual differences found between offers made and sentences received after trial for the same case. Sample selection bias also may be present through differences in con- viction rates, and hence different likelihoods of sentence, for trial and guilty-plea cases. Offenders who plead guilty are certain to be convicted

116 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM and thus selected for sentencing, while some portion of trial cases do not result in convictions. If the strength of evidence also affects con- viction rates independently of mode of disposition, stronger cases are more likely to end in a conviction. Hence it is possible that cases with the strongest evidence and those with the weakest evidence are more likely to go to trial. For the strongest cases, the prosecutor might not be willing to bargain down and accept a guilty plea to reduced charges, and there would be little advantage to the defendant to plead guilty. For the weakest cases, the defendant would have reason to hope for acquittal or dismissal in court. Among those cases going to trial, the cases with the strongest evidence would be more likely to end in con- viction. On the average, then, cases that result in convictions through trial would be stronger than cases resolved by a guilty plea. Strength (or quality) of evidence may also contribute to more severe sentences, perhaps as an indicator of greater defendant culpability for the offense. In this event, controlling for other factors, the stronger evidence against offenders convicted in trials would lead to more severe sentences for those offenders than for offenders who plead guilty. However, to the extent that strength of evidence is poorly measured and thus poorly controlled in an analysis, any contribution of evidence to more severe sentences for those convicted in trials may be misinterpreted as an effect of disposition type. In this event the observed sentence differential be- tween pleas and trials might be explained in terms of differences in the strength of evidence. The magnitude of bias due to sample selection depends on the relative strength of the relationship between case quality and sentence severity: the smaller the role of case quality in sentence severity, the smaller the potential bias. While playing a major role in case dismissals and con- victions, case quality is likely to be at most a minor factor in sentences. Certainly there is little empirical evidence supporting a claim of any major effect on sentences. Overestimates of the guilty-plea discount from sample selection are thus not likely to be large. The preponderance of evidence suggests that mode of disposition probably does exercise an independent effect on sentence outcomes. It is a common finding that defendants held in pretrial detention receive substantially harsher sentences than those who are free awaiting trial (Clarke and Koch, 1976; Foote et al., 1954; Goldkamp, 1979; Greenwood et al., 1973; Landes, 1974; Lizotte, 1978; Morse and Beattie, 1932; Rankin, 1964; Spohn et al., 1982~. This finding persists after controlling for factors like offense seriousness and prior record. A variety of processes have been suggested as factors in the observed

Determinants of Sentences 117 relationship between pretrial detention and harsher sentences. One pos- sibility is that detained defendants are less able to assist in the prepa- ration of their cases, both for trial and for subsequent sentence hearings. Some defendants may also lose their jobs while detained; the loss of income may affect their ability to retain private counsel, and their un- employment may be held against them in sentencing decisions. The conditions of pretrial detention may also induce detained defendants to plead guilty early and settle for less favorable outcomes. Those defen- dants who are free awaiting trial, on the other hand, are in a better position to delay disposition of their cases, possibly resulting in better offers from the prosecutor and decay in the strength of the prosecution case as witnesses tire of court appearances and memories fade. Finally, more severe sentences may result from a labeling process in which de- tained defendants are presumed to be more serious or dangerous (other- wise they would not have been detained) and hence deserving of harsher penalties. It is also possible that the relationship between pretrial detention and harsher sentences is at least partially spurious, resulting from the role of common determinants of pretrial detention and sentence after con- viction. Bail amount and subsequent release on bail, for example, have been found to be associated with the key determinants of sentences— offense seriousness and prior record (Lances, 1974; Lizotte, 1978~. The more serious the offense and the worse the prior record, the more likely it is that the bail amount is set high and the defendant is detained. While most studies attempt to control for any spurious role of pretrial detention by including offense seriousness and prior record in their analyses, these variables are often poorly measured. Independent measurement error in either of these important variables will yield underestimates of the contribution of seriousness or prior record and overestimates of the contribution of pretrial detention to severe sentences. With systematic measurement errors, on the other hand, the biases might be in the opposite direction (see Table 2~. Sample selection biases may also distort the estimated effects of pre- trial detention. The selection stage presumed to be most affected by pretrial detention is conviction, with detained defendants being more likely to be convicted. Selection biases arise when some poorly measured variable, like offense seriousness or prior record, affects both selection (in this case through conviction) and sentence severity. In the event that detained defendants are more likely to be convicted, regardless of se- riousness or record, those defendants who are not detained but are convicted would be likely to have more serious offenses or worse rec-

118 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM ords. Such a relationship would contribute to more severe sentences for defendants who are not detained, leading to underestimates of the impact of detention on sentence outcomes. The association of pretrial detention with poorly measured variables like offense seriousness and prior record raises the possibility of biases in either direction in the estimated effect of pretrial detention on more severe sentence outcomes. While there appear to be both empirical evidence and theoretical reasons to support the view that pretrial de- tention has an independent influence on sentences, further research is needed to establish the existence and magnitude of such a relationship. Anecdotal evidence suggests that defendants represented by public defenders or appointed counsel receive harsher sentences than those represented by privately retained counsel (Alschuler, 1975; Blumberg, 1964; Casper, 1972~. This difference has been attributed to heavier work loads or less criminal experience for public or appointed attorneys, which contributes to less adequate defense and increased pressure to dispose of cases through plea negotiations. The spirit of cooperation and com- promise that characterizes courthouse regulars is another factor that might jeopardize defendants' positions. At the same time, many pri- vately retained counsel represent large numbers of nonaffluent clients and depend upon rapid turnover of cases to generate adequate incomes from small individual case fees. Thus, their case loads and practice styles may not be very different from those of public attorneys. Moreover, the expertise and courthouse familiarity of public defenders may work to the advantage of their clients. It should be noted that there are also likely to be important jurisdictional differences in the quality of public defense counsel. Statistical analyses of the effects of attorney type have generally failed to control adequately for other determinants of sentences and are thus vulnerable to biases arising from measurement error and sample selec- tion. Furthermore, the studies result in mixed conclusions, with some studies supporting the proposition of an advantage for the clients of privately retained counsel (Bin" and Rosenfeld, 1970; Katz et al., 1971; Spohn et al., 1982) and others contradicting it (Beattie, 1935; Eisenstein and Jacob, 1977; Oaks and Lehman, 1968; Rhodes and Conly, 1981; Smith, 1970; Taylor et al., 1972~. The evidence to date does not support the conclusion that attorney type is independently related to sentence. DISPARITY In studying the determinants of sentences, it is not sufficient to consider only factors relating to the offense, the offender, and case-processing

Determinants of Sentences 119 variables. Although some statistical studies have included as many as 30 explanatory variables relating to case attributes, two-thirds or more of variation in sentence outcomes remains unexplained. Many research- ers have looked to elements of the decision-making process, especially differences among judges, for the sources of that remaining variation. Attempts to measure variation in judicial sentencing are not a 1970s phenomenon. As early as 1895 researchers tried to document the extent of interjudge disparity or the differences in sentencing attributable only to the identity of the judge (Francis Galton, Nature, 1895, cited in Banks, 1964~. Early approaches were relatively straightforward; they generally compared the rates of particular sentences given by different judges. Everson (1919) found that the frequency of suspended sentences given for public intoxication by 42 magistrates in New York City varied from less than 1 percent to 83 percent. Gaudet et al. (1933) studied the sentences imposed by six New Jersey judges and showed that the rates of incarceration for their cases varied from 34 percent of all individuals sentenced by the most lenient judge to 58 percent of those sentenced by the most severe judge. In order to conclude from these studies that judge differences ac- counted for the differences in sentencing patterns, it is necessary to assume that the samples of cases sentenced by each judge were com- parable. Even if initial case assignment was random—a practice unlikely in most courts due to management considerations and simple careless- ness—comparability of samples at the time of sentencing would probably not result. Since the judge who initially receives a case may affect its disposition by trial or guilty plea, the mix of cases ultimately available for sentencing by a judge may be a function of the judge's reputation and behavior. In order to correct for differences in the cases sentenced by different judges, some researchers have used statistical controls. The crudest of these is the matching strategy that identifies subgroups of cases sharing similar characteristics (e.g., offense, prior record) and compares the sentencing patterns of different judges for each subgroup of cases (e.g., Green, 19614. The difficulty with this approach is that a researcher can never be certain that the subgroups identified for each judge consist of strictly comparable cases; it is always possible that the cases of two judges are different on some unmeasured variable or set of variables that is crucial for the sentencing outcome. More elaborate versions of the same type of approach use regression and related statistical techniques (e.g., PROBIT) to control for case differences across judges. Variables identifying or describing judges are then introduced in the model as independent variables in addition to case attributes, and the researcher then tests to see whether a judge

120 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM variable or set of variables can explain any additional variation in sen- tencing. Judge variables may be in the form of individual judge identity (e.g., Rhodes, 1977; Shane-Dubow et al., 1979) or attitudinal/person- ality groupings (e.g., Clarke and Koch, 1977, who classified Alaskan judges as "strict" or "lenient"; Hogarth, 1971, who measured Canadian magistrates for cognitive complexity as well as attitudes toward punish- ment). Most of these studies have shown a substantial impact of judge variables. A few have shown no judge effect (e.g., Rhodes, 1977~. One reason for the lack of judge effects in some studies of sentence outcomes is that such studies include case characteristics that may anticipate or reflect judicial reaction. Bail status, for example, was a predictor of sentence in Rhodes's study. Yet, as Rhodes mentions, the bail decision may reflect an earlier judicial decision on probable sentence. In this event the role of judge effects in both sentence outcomes and bail de- cisions must be investigated together. A more general problem with using statistical controls to create com- parable subgroups of cases is that, whenever the models fail to measure some variables adequately or omit them altogether, the ability of these models to assess the effects of judicial variables will be impaired. In general, the statistical controls cannot be assumed to have adequately controlled for case differences in evaluations of the separate impact of judicial identity. To avoid the problems of lack of comparability, a number of re- searchers have submitted identical cases to several judges, asking each judge to indicate a recommended sentence for the case. The "cases" have varied in detail from a list of eight case characteristics—offense, age, record, defendant's role in the offense, plea, injury to victim, weapon, dollar amount (Forst and Wellford, 1981) through presentence reports (Partridge and Eldridge, 1974) to excerpts from trial record, testimony, and a detailed description of the offender (Hood, 1972; Kapardis and Farrington, 1982~. In each study, the results have shown substantial differences in the sentencing recommendations of different judges. Forst and Wellford (1981) found that for 9 of their 16 scenarios, some judges recommended sentences of at least 20 years, while other judges rec- ommended against imprisonment; for 2 of the cases half of the judges recommended prison and half did not. The judges in this study were all federal court judges and came from different districts. The results are similar, however, in studies comparing judges in a single district. In a study of the federal Second Circuit (Partridge and Eldridge, 1974), judges in one district disagreed on whether to incarcerate in 13 of 20 cases; in another district they disagreed in 15 of 20 cases. While the sentencing experiments described here are able to have

Determinants of Sentences 121 multiple judges "sentence" identical cases, it is possible that the "sen- tences" in the experiment would not reflect sentences given when the decision had real consequences for a flesh-and-blood offender. While the effect of personal interaction between judge and offender is probably very limited (the defendant usually pleads guilty, and the judge learns about the defendant through the presentence report and from statements by opposing counsel), the absence of real consequences in experiments and the use of often limited case information that leaves considerable room for judicial interpretation or imputation of relevant but missing information are potentially more troublesome. One study that reduced these problems took advantage of a naturally occurring collegial sentencing structure the sentencing council (Dia- mond and Zeisel, 1975~. Federal judges in several courts meet regularly to discuss their sentencing decisions. Before each meeting every council member receives presentence reports on the offenders to be discussed at the meeting. Before the council convenes, each judge privately re- cords a favored sentence for each case. These recommendations are discussed at the council meeting and are expected to influence the de- cision of the sentencing judge, who retains full power to determine the actual sentence. Thus, unlike a decision in sentencing experiments, a sentencing council recommendation has real consequences for the of- fender through its potential influence on the sentencing judge. The information supplied to the council judges also closely approximates the information available to the sentencing judge. The results of this study indicate substantial disparity in sentence recommendations: in 30 per- cent of the cases, a random sample of three judges disagreed about whether to incarcerate the offender. The figure is almost identical for sentencing councils in Chicago and in New York. The sentencing council study generally controls for case attributes and defendant vulnerability. Hence, the only remaining problem is the extent to which the measure of disparity is influenced by interpersonal processes of the council itself, so that the recommended sentences do not com- pletely reflect the sentences of individual judges sitting alone. Judicial disparity may be somewhat understated in council cases if the prospect of formal review of individual judicial decisions in council deliberations leads judges to be more circumspect in their sentence recommendations. It is also possible that the prospect of a moderating effect of council deliberations may lead individual judges to initially recommend sen- tences that are more extreme than they would actually desire as a result. This situation would exaggerate or overstate the extent of judicial dis- parity.

122 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM In considering potential sources of systematic judicial variation, it is generally acknowledged that pursuing different goals in sentencing can often result in very different sentences in the same case. For example, general deterrence may suggest a prison sentence for the first-offender tax evader, while the goals of specific deterrence and rehabilitation would argue for a fine or probation. To the extent that different judges emphasize different goals, as found in Forst and Wellford (1981), for example, one would expect their sentences to differ. Aside from general judicial predilections, the particular goals of sen- tencing deemed appropriate in any case may be influenced by a variety of cues reflecting the degree of offender culpability (or responsibility for the offense) and the stability or enduring quality of offending be- havior for the defendant. The extent of blameworthiness of the offender affects judgments of the punishment deserved, and increases in blame- worthiness may well evoke sentences based on goals of retribution. To the extent that an offender is judged to be fully responsible for his or her actions and the offending is viewed as a stable attribute of the offender, the likelihood of incapacitative sentences increases. Sentences for the purposes of rehabilitation or deterrence are more likely when offending is perceived to be a temporary attribute of the offender. This perception increases the potential that a sentence can actually affect future offending behavior, both for the sanctioned offender and for others who witness the sanction. Various elements have been suggested as influencing attributions of offender culpability and stability. The level of responsibility for an of- fense varies with the offender's motivation and ability to commit the offense. Motivational factors like victim provocation (Harvey and Engle, 1978) and the extent of planning or forethought involved (Harvey and Engle, 1978; Joseph et al., 1976) have been found to affect attributions of culpability, as have ability factors like level of mental or psychological functioning (Monahan and Hood, 1976) and abuses of authority or po- sition (Diamond and Herhold, 1981; Thomas, 1979~. Another factor in culpability is the level of harm done (Hood, 1972; Kapardis and Far- rington, 1982; Walster, 1966; Wheeler et al., 1981~. There is little em- pirical work on cues affecting judgments of stability; some potentially important factors might include remorse, cooperation with authorities, and indicators of more general social stability, like family support and employment opportunities. Few of these variables effort, planning, level of psychological func- tioning, provocation, harm, and stability cues- have been directly mea- sured in studies of judicial sentencing. To the extent that they influence

Determinants of Sentences 123 judges differently in different cases, they might well account for inter- judge and intrajudge disparity. It is also possible that the origins of judicial disparity may have little to do with judges. Several studies have identified the importance of the recommendations by the prosecutor or probation officers in determining sentence outcomes (Carter and Wilkins, 1967; Hagan, 1975, 1977; Ha- gan et al., 1979; Myers, 1979; Unnever et al., 19804. Variations in sentences among judges and even for the same judge thus may arise from variations in the individual prosecutor or probation officer making sentence recommendations in different cases. The evidence for sentence disparity is extensive, but data on the sources of that disparity are scarce. One plausible direction for research is to examine the sentencing goals of different judges, how the goals are formed, and where they lead. If sentences are in part a product of the goals they are meant to achieve, the absence of consensus on ap- propriate sentencing goals may be a major factor contributing to inter- judge disparity. The extent to which disparity is unwarranted remains an important policy question whose resolution depends on the weight given to com- peting values. On the one hand, there is concern that sentences result from the evenhanded application of general sentencing principles. On the other hand, there is a recognition that there are often legitimate social, cultural, and philosophical differences over what those principles should be, as reflected, for example, in conflicting interpretations of the goals of sentencing. Resolution of this policy issue would benefit from continued efforts to clarify and articulate the principles that currently do and those that ought to underlie sentence decisions. Such work would help to illuminate the dimensions of the choices that must be made. CONCLUSION Evidence on the determinants of sentences is beginning to emerge from several research approaches. The available research provides some gen- eral information on which factors may be important and which may not. Estimates of the magnitude of these effects are considerably less precise. One limitation of existing research is inadequate controls for poten- tially important determinants of sentences arising from omitted or poorly measured variables. This limitation contributes to statistical biases of often unknown direction and magnitude in the estimated effects. Sentence decisions are also typically analyzed using simple linear models involving weighted sums of individual variables to characterize the re-

124 RESEARCH ON SENTENCING: THE SEARCH FOR REFORM lationship between determinants and sentence outcomes. These analyses often fail to address even simple forms of interactions among explanatory variables. Instead, all variables are considered simultaneously and al- ways enter the decision with the same impact. However, sentence de- cisions may be more complex and may require richer characterizations of the decision process. For example, it may be that sentencing decisions are a multistage process that first involves an attempt by the decision maker to allocate the case to one of a small number of case patterns, where each case pattern is subject to a different sentencing rule. One pattern of cases, for example, may be viewed as particularly well suited to rehabilitation, and the sentences imposed would be intended to en- hance rehabilitation opportunities. Another pattern of cases may elicit an incapacitative response, while still another pattern may be distin- guished for its potential general deterrent effects and be sentenced ac- cordingly. The sentencing rules characterizing sentence decisions within each case pattern may vary in terms of the variables included and the weights given these variables and may invoke interactions among variables and hierarchical treatments of the variables. In a hierarchical sentencing rule, the sentence decision follows a branching process in which the weight given some factors depends on the presence or absence of other factors. For example, in a particularly heinous crime, the viciousness of the crime alone may be sufficient to lead to incarceration. In less heinous crimes, a variety of factors, like the defendant's prior criminal record and general community ties, may enter the decision to imprison or not. There may also be some cases that do not fit any of the identified case patterns. Such cases may be sentenced on the basis of the partic- ularly unique features of the case and so be difficult to characterize by a general rule. This characterization of sentencing decisions is quite different from existing analyses in which the same simple linear model is applied uni- formly to all cases. The alternate formulation involves first a process of pattern recognition and then the application of potentially complex de- cision rules. Specifying the actual forms of alternate models of sentencing decisions to be tried will probably benefit from the insights derived from interviews of participants and extensive observations of the process. It is also important to remember that sentencing decisions are not made in isolation; they occur in the context of a variety of earlier de- cisions that potentially influence sentence outcomes. As a result, when attempting to sort out the determinants of sentences, one cannot focus only on the outcomes of the convicted cases that appear before a judge for sentencing. Sentencing decisions must be viewed more broadly to

Determinants of Sentences 125 reflect the impact of earlier decisions that result in convictions in some cases, thus making offenders vulnerable to sentencing. This larger sys- tem approach to the process will also help to address the methodological problems arising from selection, as well as an indirect basis for resolving the measurement problems in key concepts like seriousness, prior re- cord, and case quality.24 24 See Garber et al. (Volume II), Klepper et al. (Volume II), and Berk and Ray (1982) for a more detailed treatment of the ways in which explicit consideration of the broader case-processing system can help to alleviate the biases arising from measurement error and sample selection.