Testing Control Theory and Differential Association: a Causal Modeling Approach*

TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION: A CAUSAL MODELING APPROACH*

Ross L. MATSUEDA Universityof California,Santa Barbara

A number of strong theoretical statements have been based on analyses of delinquency data from the Richmond Youth Project. Hirschi (1969) and Jensen (1972), in particular, found that Hirschi's control theory was empirically supported over Sutherland's theory of differential association. This paper reanalyzes these data and reassesses this negative evidence pertaining to differential association theory. It is shown that the ratio of learned behavior patterns favorable and unfavorable to violation of legal codes, the critical variable in Sutherland's theory, can be operationalized by explicitly modeling its measurement error structure. In turn, this allows the testing of specific hypotheses derived from the theory. The analysis based on this strategy finds differential association theory supported over control theory. Specifically, the unobservable construct representing the ratio of learned behavior patterns successfully mediates the effects on delinquency of the model's other variables.

A major contemporary controversy in the groups, the theory of differential association sociology of crime and delinquency concerns asserts that crime is rooted in normativecon- two dominant theories of criminal behavior: flict (Sutherland, 1947:19). In industrialized Sutherland'stheory of differentialassociation, societies, at least, definitions of legal codes and Hirschi's control theory. The most signifi- that favor law violation exist alongside defi- cant researchaddressing this issue is Hirschi's nitions unfavorableto law violation. (1969) landmark study. Hirschi developed, Sutherlandgave the name "differentialasso- operationalized,and empiricallyconfirmed his ciation" to the process by which persons expe- control theory, and presented evidence that rience these conflicting definitions about ap- seriously questioned the empiricalefficacy of propriatebehavior. Thus, definitionsfavorable differential association theory. Jensen (1972) and unfavorableto delinquent or criminalbe- reanalyzed these data from the Richmond havior are learned through interaction (com- Youth Project and, focusing on the relation- munication)in intimatepersonal groups.' This ships among parents, peers, and delinquency, differential learning includes the specific di- also found Sutherland'stheory unsupported. rection of motives, drives, rationalizations, The objective here is to develop a better strat- and attitudes-whether toward viewing legal egy for testing differentialassociation theory, codes as rules to be observed or broken. "A to investigate certain measurementproperties person becomes delinquentbecause of an ex- of the Richmond data, and to reassess this cess of definitionsfavorable to violationof law negative evidence pertainingto the theory. over definitions unfavorable to violation of law" (Sutherlandand Cressey, 1978:81). DIFFERENTIAL ASSOCIATIONVERSUS Both favorable and unfavorabledefinitions CONTROLTHEORY (behaviorpatterns) are weightedby frequency, duration,priority, and intensity. Thus, behav- Based on a conception of modern society as ior patterns presented with greaterfrequency, heterogeneous and segmented into conflicting presentedfor a longertime, presentedearlier in life, and presented from a more prestigious * Direct all correspondence to Ross L. Matsueda, source will have more weight in the process Department of Sociology, University of California, producing delinquent or nondelinquent be- Santa Barbara, CA 93106. havior An earlier version of this paper was presented at (differentialassociation). the Annual Meetings of the American Sociological In developing differential association, Association, Toronto, 1981. I thank William T. Sutherlandattempted to account for both the Bielby, Donald R. Cressey, and Thomas P. Wilson distributionof crime rates and for individual for helpful suggestions and invaluable comments on cases of criminalbehavior (Sutherland,[1942] earlier drafts. I also thank James N. Baron, Gary F. 1973:18-20; Cressey, 1960). Because true Jensen, Bruce C. Straits, and anonymous ASR re- crime rates are summarystatements about the viewers for their comments; Nancy C. Jurik for her encouragement and thoughtful advice; and Travis Hirschi, Joseph G. Weis, and Carol A. Zeiss for their I Our references to crime also pertain to juvenile assistance in obtaining the data. delinquency, and vice versa. American Sociological Review 1982, Vol. 47 (August:489-504) 489 490 AMERICAN SOCIOLOGICALREVIEW frequency of individualcriminal acts, they are business-calculates the risk of losing the in- determinedby the proportionsof persons re- vestment. Finally, involvement in con- ceiving an excess of criminalbehavior patterns ventional activity reduces delinquency by throughthe differentialassociation process. In limitingone's time to contemplateand commit other words, the extent to which a group or delinquentacts.2 society is organized in favor of crime, as In summary, according to differentialasso- against the extent to which it is organized ciation theory, definitions of the legal code against crime, determines its crime rate. mediate the effects of structural factors on Sutherlandgave the name "differentialsocial crime. According to control theory, in con- organization"to this process whereby certain trast, definitions of the legal codes, which re- structures translate normative conflict into flect the degree of belief in the moralorder, do various rates of crime. Moreover, he proposed not mediate attachment, commitment, or in- that structuralconditions such as class, age, volvement. This differenceprovides a basis for sex, ethnicity, and family status affect individ- empirically testing the comparative explana- ual criminality (and, thus, aggregate crime tory efficacy of the two theories (Hirschi, rates) only by affecting the probability of 1969:98-100; Jensen, 1972:564; Hepburn, learningbehavior patternsfavorable and unfa- 1976:450-51;Kornhauser, 1978:238). vorable to law violation (Sutherland, [1944] 1973:31).Thus, any effects that these factors HIRSCHI AND JENSEN'S TESTS OF have on either criminalityor crime rates are DIFFERENTIAL ASSOCIATION mediatedby the process of learningdefinitions favorable and unfavorableto delinquency. Hirschi's analysis of the self-report delin- In contrast, Hirschi's control theory insists quency data collected from the Richmond that definitionsof the legal code do not mediate Youth Projectprovides empiricalevidence that all such factors. Instead of asking why some purportedlycontradicts differential association persons engage in crime-as do most theories theory and supports his control theory. Spe- of deviance-control theory asks why most cifically, two of his findingsregarding parents, persons refrain from criminalbehavior. Delin- peers, and delinquency directly question the quency is taken for granted;conventional be- explanatorypower of differentialassociation.3 havior is problematic (Hirschi, 1969:10). Put First, Hirschifinds that the more intense the positively, control theory maintains that per- friendships-as measured by attachment- sons conform to legal codes because they are among boys with one or more delinquent bonded to society. Accordingly, when a per- friends, the less likely they will engage in de- son's bond to society is broken or weakened, linquency.4 Here, following Short (1957), he or she is free to engage in delinquency-but Hirschi operationalizesdifferential association is not requiredto do so. For Hirschi, then, the theory by using associations with delinquents motivationto commit delinquentand criminal ratherthan associations with behaviorpatterns behavior is constant across persons, and thus, favorable to delinquency.5But because delin- is not an explanatory variable (Hirschi, 1969:10-11, 24-25, 32). 2 Sutherland maintained that neutral (con- The bond to society is a strongcord consist- ventional)behavior affects criminalityin two princi- ing of four interwoven strands: attachment, pal ways. First, neutral behavior occupies one's belief, commitment,and involvement.Because time, which prevents one from associating with the fibers or elements of each strand resemble either criminal or anticriminalbehavior patterns those of the others, these strandsare positively (Sutherland,1947:7). Second, in a given situation,a intercorrelated.However, each affects delin- particularnoncriminal act can provide an alternative so the four are ana- to criminal behavior and thereby prevent that be- quency independently, havior from transpiring (Sutherland, [1944] lytically separable (Hirschi, 1969:27-30). At- 1973:35-36). This constitutes part of the objective tachment, perhaps the most importantstrand opportunityto engage in criminalbehavior. in the bond, contains a moral dimension that 3 We are here concernedwith only the two specific dissuades persons from delinquency. For findingspertaining to parents and peers. We do not Hirschi there are no delinquent subcultures. address other hypotheses, used by Hirschi Instead, there is variation in the extent to (1969:140)as indirecttests of differentialassociation, which people believe in society's norms, and because they are quite peripheral,if not irrelevant,to the less their belief the more likely they are to the theory. in to con- 4 In a partial replicationof Hirschi's study, Hin- engage delinquency. Commitment delang (1973) found that, contrary to predictions ventional activity dissuades persons from de- from control theory, attachmentto peers increased linquency because, when considering delin- the likelihood of delinquency. quent behavior, a person who has invested 5 Among the other studies operationalizingdif- time and energy in a conventional activity- ferential association using associations with delin- such as getting an education or buildingup a quent friends, rather than using behavior patterns TESTING CONTROLTHEORY AND DIFFERENTIALASSOCIATION 491 quent behavior patterns can be learned from process and situationally-induced-motives nondelinquents, and antidelinquentbehavior theories (Shortand Strodtbeck,1965; Briar and patternscan be learned from delinquents, this Piliavin, 1965)-or indirectlyby exposing a boy procedurecannot refute the theory, though it to delinquentbehavior patterns-as differential can provide support for it. Further, even if association specifies? friends are used as a crude measure of associ- As measures of the "availability of delin- ations with behavior patterns, attachment quent patterns," Jensen uses official delin- must, from the standpointof differentialasso- quency rates in schools, perceptionsof trouble ciation, refer to all friends, not just delinquent in the neighborhood,and numberof delinquent ones. friends (see also Short, 1957).Furthermore, he Specifically, by interpretingSutherland to uses self-reporteddelinquency as an outcome mean that "crime is more likely the more in- variable, but because the self-reportrates var- tense the associations with criminals,"Hirschi ied little across schools, he uses official delin- (1969:151)concludes that his finding about in- quency rates in schools to measure the avail- tensity contradicts differential association ability of delinquent definitions (Jensen, theory. Sutherland'stheory, however, pertains 1972:566).In addition, Jensen combines four to associations with behavior patterns, not to questionnaireitems to represent Sutherland's associations with criminals and noncriminals. "definitionsfavorable to violationof the law." As noted above, he specified "intensity" as a Jensen reports, with reference to the first modality that gives weight to whatever crimi- hypothesis, that when the differentialassocia- nal and anticriminalbehavior patternsare pre- tion variablesare held constant (by subdivision sented: an "intense" behavior pattern has in three-variabletables), the parental control greaterimpact on criminalbehavior than a be- variables (father's supervision and father's havior pattern that is not "intense." What is support) still depress delinquency. He con- needed is a specific measureof the ratio of the cludes that control theory is supported over two kinds of behavior patterns, explicitly differential association (Jensen, 1972:574). weightedfor intensity. Measuresof the number With reference to the second hypothesis, Jen- of delinquent and nondelinquent friends, if sen finds that the numberof delinquentfriends used at all, should be used to investigate the affects delinquencyindependently of the effect sources of learning delinquent and antidelin- of delinquent definitions. He concludes that quent behavior patterns. group-processand situational-motivestheories Second, Hirschi reports that when the are empiricallysuperior to differentialassocia- numberof delinquentfriends is held constant, tion theory (Jensen, 1972:573-74). certain indicators of attachment to parents, AlthoughJensen improvedon Hirschi's test attachmentto school, and commitmentto con- of differentialassociation theory, his study still ventional achievement all affect delinquency. leaves several questions unanswered. First, But, strictly speaking, this procedureformally does measurementerror in the indicatorsof the invalidates a "bad companions" theory of ratio of delinquent and antidelinquent defi- crime, not differentialassociation. This is be- nitions seriously attenuatetheir effects on de- cause it controls for the numberof delinquent linquency? Second, given that Sutherlandin- friendsrather than measuringdefinitions of the vented the differentialassociation principlein legal code. Again, a more direct operationali- part to account for certain variationsin crime zation of delinquent and antidelinquentdefi- rates-variations according to age, social nitions is needed. class, broken homes, and neighborhoods-can In reanalyzing these data but using more it in fact do so? And third, is the substantive extensive measures of "availability of delin- picture distorted because Jensen's three- quent patterns," Jensen (1972) goes beyond variable tables fail to capture more complex Hirschi to a more explicit test of differential relationshipsamong relevant variables?These association theory. He investigates two hy- issues are addressed in the following analysis. potheses. First, does parental control affect delinquency directly-as control theory SPECIFICATIONOF A CAUSAL predicts-or only indirectlyby influencingthe MODEL probability of learning delinquent The Richmonddata used by Hirschi and Jen- definitions-as differential association theory sen are reanalyzed here; thus the quality re- suggests? Second, do delinquent peers foster mains the same, as do the assumptions about delinquency directly-as suggested by group- the researchdesign. Perhapsthe most tenuous assumption pertains to the causal ordering of learned from all sources, Short (1957, 1958, 1960), the variables.6 Consistent with the theory of Voss (1964), Stanfield(1966), and Krohn (1974)find support for the theory, while Reiss and Rhodes 6 Anotherlimitation stems from nonresponsebias (1964) do not. in the originalsample. Hirschi(1969) noted that indi- 492 AMERICAN SOCIOLOGICAL REVIEW

Model Derived From DifferentialAssociation Theory Background Definitions Delinquency Variables n_ fn n D

Parental Peer Supervision - Relationships

Model Derived From Control Theory Background Definitions -) Delinquency Variables

Parenta l Peer Supervision -> Relationships (Attachment (Attachment to Parents) to Peers)

Model Derived From Multiple Factor Theories Definitions

BackgroundBackground ---'A~~~ Variables ~~~~~~~~~~~~Delinquency

Parental Peer Supervision - Relationships Figure 1. Causal Models of Delinquency differentialassociation, it is assumed that the unobservablevariable with multipleindicators, measure of definitions favorable and unfavor- is incorporatedinto the system of equations. able to delinquency is causally prior to delin- Third, specific hypotheses and the overall fit of quent behavior and causally subsequent to the model are evaluated using both tests of other variables in the model. This assumption point estimates and a more global goodness- cannot be tested with cross-sectional data; in- of-fit test. stead, a longitudinaldesign is required. (For evidence on this issue, see Cressey, [1953] 1973; Minor, 1981.) The Substantive Model The reanalysisattempts, in three steps, to go To simplify matters, the causal relations beyond Hirschi and Jensen's tests of differen- specified in the substantive model can be dis- tial association theory. First, certain causal re- cussed in terms of the five variables dia- lations impliedby the theory are translatedinto grammed in Figure 1. Primary focus will be a structuralequation model. Second, a mea- given to the differential association model, surement model, which treats definitions fa- using the alternativemodels derived from con- vorable and unfavorableto delinquency as an trol theory and multiplefactor theory as a point of contrast. viduals who failed to complete the questionnaire The four background variables, the age of were more likely to be delinquent. The critical ques- the respondent (AGE), his parents' socioeco- tion, however, is whether the relationships among nomic status (SES), whetheror not his home is relevant variables differ for those who failed to re- spond. Hirschi (1969:46) compared the relationships intact (BROKHOME),and his perceptions of between certain school-related variables and delin- trouble in his neighborhood (YOUNGTRO), quency and found no significant differences between are commonly cited as importantdeterminants respondents and nonrespondents. For a detailed de- of crime and delinquency. However, as scription of the Richmond data, see Hirschi (1969). Sutherland and Cressey (1978) show, dif- TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION 493 ferentialassociation theory explicitly stipulates ables, to affect peer relationships.Presumably, the causal mechanismthat makes these back- the kinds of friendshipsa boy makes, and the ground variables "work" to produce crime or closeness of those friendships,are affected by delinquency. (See Appendix A for brief de- the extent that his parents supervise him. Dif- scriptions of the variables.) ferential association theory predicts that be- For example, after summarizingthe empiri- cause delinquentbehavior, like all behavior, is cal literatureon age and crime, Sutherlandand learned primarilyin intimate groups, peer re- Cressey (1978:129-30) conclude that age has lationships have an important impact on importantdirect or indirect effects on crimi- learningdefinitions of the legal code. nality. The objective here is to determine The two aspects of peer relationshipscon- whether age affects delinquency directly, or sidered here are numberof delinquentfriends indirectlythrough its effects on parentalsuper- and attachmentto friends. Fromthe standpoint vision, peer relationships, and definitions fa- of differential association theory, delinquent vorable and unfavorableto delinquency. friends increase one's delinquency because Sutherland and Cressey (1978:220) further they are more likely thannondelinquent friends hypothesize that low socioeconomic status to transmit definitions favorable to delin- may affect delinquencyin two ways. First, be- quency. Jensen, however, finds that numberof cause poverty areas are often areas of high delinquentfriends affects delinquency regard- delinquency, a child from a low SES home has less of the number of definitions favorable to a greaterprobability of encounteringmany de- delinquency. This finding is reassessed. linquent behavior patterns. Second, being a Accordingto differentialassociation theory, member of a lower class may affect a child's behavior patterns learned in relationshipsthat denial or acceptance of conventional values. are intense, emotional, and prestigious have Similarly, they conclude that various home more significancefor subsequentconduct than conditions, including broken homes, promote behavior patterns without these charac- delinquency by increasing children's associ- teristics. Therefore, degree of attachment to ations with delinquent definitions, and de- friends should have an important effect on creasing their associations with antidelinquent learningdefinitions of legal codes. But the di- definitions (Sutherland and Cressey, rection of that effect depends on the content of 1978:219-24). the behavior patternslearned-whether favor- Perceptions of trouble in the neighborhood able or unfavorableto delinquency.Again, the (YOUNGTRO)is used as an indicatorof delin- importantpoint is that, from the standpointof quency area. Both Short (1957) and Jensen differentialassociation theory, attachmentto (1972) advocate using neighborhood delin- friends affects delinquency only insofar as it quency or troubleas perceived by respondents. affects the learningof definitionsfavorable and For Sutherlandand Cressey, what is important unfavorableto law violation. In contrast, con- is actual trouble in the neighborhood, which trol theory predicts that attachmentto peers indicates the number of delinquent behavior affects delinquency regardless of what is patterns in the area. But perhaps one's effec- learned (see Figure 1). tive neighborhood,as experienced throughas- In short, differentialassociation theory pre- sociation with persons and values of the area, dicts that the ratio of definitionsfavorable and is more significant for learning delinquency unfavorableto delinquencymediates the effect than is one's objective neighborhood.This is of each prior variable on delinquency. Thus, because some persons do not perceive trouble the effects of these variablesshould be zero, or in their neighborhoods because they are iso- at least trivial in size comparedto the effect of lated from it. definitions of the law. In contrast, multiple In the substantive model (Figure 1), each factor theories impute direct causal power to backgroundvariable is assumed to affect pa- each variable, while control theory predicts rental supervision. From the- perspective of that parental supervision (attachmentto par- differentialassociation theory, supervision re- ents) and peer relationships (attachment to duces delinquency by increasing exposure to peers) directly affect delinquent behavior. In antidelinquentdefinitions and decreasing ex- the followinganalysis, each of these alternative posure to delinquent definitions (Sutherland hypotheses is examined. and Cressey, 1978:222). In contrast, both The full structuralequation model, contain- Hirschi and Jensen treat parental supervision ing these relationships,is presentedin the path as an indicator of attachmentto parents, and diagramof Figure2. The substantivemodel is a find that, as control theory predicts, supervi- recursive system of five equations. Each dis- sion affects delinquent behavior directly (see turbance is assumed to be independentof all Figure 1). other disturbancesin the model and indepen- The substantive model also allows parental dent of the predictorvariables in its equation. supervision, as well as the backgroundvari- The disturbances contain numerous omitted 494 AMERICAN SOCIOLOGICALREVIEW

U1 U2 X4 U3 U5

AGE ~~~~~~~~~~~~DEFE

G / BROKHOME e1 /SE YOUNGTRO ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~e17 e2 BROOME \ SUCKERS\ \ \\ DELHURT\ \\ ~RSPECTPO Am TROUBLE e/ 3 YOUNGTROONTOATAHE , EVNBREAKOKLAW GETAHEAD

"/ 0 e 0a ;!fI t !

to l 5RPIKU BELIKFR RSPECTFR

PARIWITHPARWHERE FRPICKUP

e5 e6 e7 Figure 2. Path Diagramof the Full Causal Model Derived From DifferentialAssociation Theory effects which, we assume, constitute a random The Measurement Model shock.7 Also, the disturbancesare assumed to have a mean of zero and constant variance. In Perhaps the most damaging criticism of dif- that a addition, apart from BROKHOME,a dummy ferential association theory argues per- variable, all variables are treated as interval son's ratio of learnedbehavior patterns cannot (Cressey, 1952:52; scale measures, and all effects are assumed to be determined accurately Sutherland and Cressey, 1978:91). This is a be linear and additive.8 if this variable cannot be Hirschi and Jensen, the severe criticism, for Finally, following operationalized, the theory cannot be sub- measure of delinquency (DEL) is an index of acts committed in the jected to empirical verification. Cressey self-reporteddelinquent although differentialas- year prior to administrationof the question- (1960:57) argues that naire.9 sociation is not stated precisely enough to stimulate a rigorous empiricaltest, it remains an importantprinciple that focuses attentionon the fact that all behavior is oriented by norms, 7 For example, the disturbancepredicting delinquency contains the objective opportunityto engage and it also integratesand explains variationsin in delinquency,which includesalternative behaviors crime rates. Still, the theory would be more withinthe situation,interactions between person and valuableto social scientists if it could stimulate situation,and techniquesof delinquencynot learned explicit empirical investigations. throughdelinquent definitions. DeFleur and Quinney In developing the theory, Sutherland([1944] (1966), however, argue that underlyingdifferential 1973:36;1947:7) intended to express the ratio association is the postulate that knowingdelinquent of weighted definitionsfavorable and unfavor- definitionsimplies knowingthe techniquesfor com- able to delinquencyin terms of a precise math- mittingdelinquency. ematical formula, which he would then use to 8 We follow Hirschi (1969) in assuming interval predict criminalbehavior. It now seems clear scales. Although Jensen (1972) found some slight interaction effects, they are not substantively im- that the ratio cannot be determinedin the pre- portantfor our purposes. cise way that Sutherlandanticipated because 9 Although our primaryfocus is on the research the relevant behavior patterns cannot be di- frameworkof Hirschi and Jensen, we also investi- rectly observed, let alone weighted and sum- gated the measurement properties of the specific med to form a ratio. Nevertheless, when the items entering the index of delinquency. Incorpo- ratio of behavior patterns favorable and unfa- ratingthis measurementmodel-a single latent vari- vorable to delinquency is treated as a "latent" able with six indicatorsand three measurementerror or "unobservable"variable, it does have oper- correlations-into the system of equations yields similarsubstantive results. Most important,the hy- ational implications for relationships among pothesis tests again confirm differentialassociation variables than can be observed. More pre- theory. The validity coefficients of the indicatorsof cisely, observable items measuringdefinitions DEL are .46 for THEFT2, .45 for THEFT250,.37 for of the legal code can be specified as indicators THEFT50, .36 for CARTHEFT, .42 for VAN- of this underlying theoretical construct. Al- DALSM, and .46 for BATTERY. though each indicator is a fallible measure of TESTING CONTROLTHEORY AND DIFFERENTIAL ASSOCIATION 495 the underlying construct, the inaccuracy can component consists of two elements. One, the be taken into considerationby explicitly mod- eight indicators are each assumed to be ran- eling the indicators' measurement error domly generated from an infinite domain of structure. definitions items, and the resulting sampling However, this method of operationalizing error is picked up by the disturbance. Two, the "ratio" concept does not directly compute imperfectionsin the measuringinstrument (the a ratio. Sutherland ([1942] 1973:22) specified questionnaire)result in measurementerror (see "ratio"to indicate: Jdreskog, 1976). From a statistical standpoint, that some persons who have many intimate the two are treated the same way (Bielby and contacts with criminals refrain from crime Hauser, 1977:144). and that this is probably due to the coun- There are two reasons to expect some mea- teracting influence of associations with surement errors to be positively correlated. anti-criminalbehavior. Actual participation First, some of the items are substantively in criminal behavior is a resultant of two similar, so the components orthogonalto DEF kinds of associations, criminal and anti- tap some commonality.Second, the items were criminal,or the associations directedtoward all measuredon a single occasion, measuredon crime and the associations directed against identical scales, and (for the first four items) crime. measured consecutively on the questionnaire. This procedurefor operationalizingDEF re- Empirically,this implies that any operationali- quires that the indicators show content zation must capture associations with both validity-the items should tap the content of procriminaland anticriminalbehavior patterns. behaviorpatterns favorable and unfavorableto Because each definitionvaries by the weight it delinquency as Sutherlandintended. In addi- receives from frequency, duration, priority, tion, collectively, the sample of items should and intensity, each can, at least in principle,be represent all strata of the domain of content placed on a continuum. If each definition is (Bohrnstedt, 1970:92). The items were de- placed on a single continuum, ranging from signed to measure attitudestoward the law and highly antidelinquent to highly delinquent, Sykes and Matza's (1957) techniques of neu- each can be measured by the same scale. The tralization, both direct operationalizationsof theoretical construct, then, becomes a unidi- Sutherland's delinquent definitions (Hirschi, mensional variable measuring weighted defi- 1969:198-212; see also Austin, 1977).12 nitions favorable and unfavorable to delin- Also specified as an unobservable variable quency on a continuous scale. Although, with multiple indicators, the theoretical con- strictly speaking,the theoreticalconstruct here struct "parental supervision" (SUPER) is in- is not a ratio, it is a monotonic transformation dicated by PARWITH and PARWHERE.1 3 of a ratio, and, moreover, it measures both kindsof associations, which is what Sutherland 11 With domain samplingmodels the measuresare intended to accomplish.10 not literallysampled from an infinitedomain of mea- In short, after eliminating measurement sures. Rather, it is merely assumed that the sample error, the common varianceamong these indi- of items is a randomrealization of all possible mea- cators should adequatelycapture the variation sures (see Nunnally, 1967; and Bohrnstedt, 1970). among persons' definitions of the legal code. 12 Persons' responses along the continuous scale Consequently, the ratio of definitions of the should also reflect aspects of their ratios of defi- legal code can be operationalized, and the nitions. For the sample of items used here, this as- theory of differentialassociation can be sub- sumptionappears reasonable. For example, persons examination. answering"strongly agree" to the statement, "It is jected to rigorous empirical all right to get around the law if you can get away The measurementmodel is specified by the with it," should be those with high ratiosof weighted indicator-constructand indicator-errorpaths of definitionspertaining to this class of verbalizations. Figure2. As in a factor analytic configuration, Conversely, those answering "strongly disagree" each indicatoris specified as a linear combina- feel it is not "all rightto get aroundthe law if you can tion of the latent variable(DEF), plus a random get away with it"-which reflects antidelinquentat- measurement disturbance. The measurement titudes. Thus, these persons should have low ratios errors are assumed to be uncorrelated with of this class of weighted definitions. both structuralvariables and structuraldistur- 13 Parentalsupervision was originallyspecified as The stochastic measurement error a perfect linear combination of its two indicators, bances. and similar substantive results were found. How- ever, if these indicatorscontain measurementerror, 10 Jensen (1972), Hepburn (1976), Akers et al. that specification could bias the model in favor of (1979), and Johnson (1979) have tried to measure differentialassociation theory. Consequently,to in- definitionsof the legal code on a continuous unidi- crease the strengthof the test of differentialassocia- mensionalscale, but without modeling its measure- tion, the resultsof the present specification,allowing ment error structure. for errors of measurement,are reported. 496 AMERICAN SOCIOLOGICALREVIEW Similarly, ATTACHPE is indicated by BE- sampled. After using listwise deletion, missing LIKFR and RSPECTFR, and, (following values reduced the sample size to 1140.16 Hirschi) labeled "attachmentto peers."14 The model's parameterswere estimated by In addition, the perhapsunrealistic assump- using Joreskog and Sdrbom's (1978) LISREL tion that AGE, SES, BROKHOME, IV program.Assuming that the joint distribu- YOUNGTRO, FRPICKUP, and DEL are tion of the 17 variables is approximatelymul- perfectly measured variables is relaxed. Since tivariate normal, the programcomputes con- each of these constructs has but a single indi- sistent and asymptotically efficient maximum cator, the measurementparameters of a model likelihood estimates of parameters for iden- allowing for measurement error cannot be tified models. The parametersof both the five identified-a prerequisitefor estimation.Thus, equation substantive model and the seventeen these parameterscannot be estimatedfrom the equation measurementmodel were estimated data; they can only be fixed to more plausible jointly as a single system.'7 values. The validity coefficients of SES, LISREL IV also allows one to test an over- BROKHOME,YOUNGTRO, and DEL were identifiedmodel's ability to reproducethe ob- fixed to equal .80, and that of AGE to equal served variance-covariance matrix. Specifi- .95; the correspondingreliabilities were fixed cally, the likelihood-ratio-teststatistic tests the to .64 and .90. These values appearlow enough null hypothesis, that the model's overidentify- to reduce adequatelythe chance that tests are ing restrictionsare satisfied in the population, biased in favor of differential association against the alternative, that the moments are theory, and high enough to avoid obtainingim- actually unconstrained.In large samples, such plausible estimates of the remainingparame- as the one used here, this statistic is distributed ters of the model.'5 approximately x2, with degrees of freedom equal to the number of moments minus the number of parametersestimated. In addition, ESTIMATIONOF THE MODEL specific hypotheses, or overidentifyingrestric- tions, can be tested by nesting the hypoth- Following Hirschi and Jensen, the present esized model within a less restrictive model. study focuses on the 1588 nonblack males The difference in x2S provides a likelihood- ratio test of the restrictions, with degrees of freedom equal to the difference in degrees of 14 The metric of an unobservableis arbitraryand freedom between the two models. Often the must be normalizedfor identificationpurposes. For descriptive ratio of X2/df is used to assess the metric models, by constraining one of the factor relative fit of various models (cf. Wheaton et loadingsof each construct to equal unity, the metric al., 1977;Isaac et al., 1980).This procedureis of SUPER was arbitrarilyfixed to equal that of followed here. PARWITH,the metricof ATTACHPEwas fixed to The programalso provides a matrixof first- equal that of BELIKFR, and the metricof DEF was fixed to equal that of TROUBLE. Thus the slope order partial derivatives of the minimized parametersof these indicatorscan be identifiedonly function with respect to each fixed and free relative to one another(Bielby et al., 1977:1251).In parameter,and a residualmatrix of discrepan- addition,for standardizedmodels, the metricwas set cies between observed and implied moments. by fixing the variance of all unobservablesto unity These may be useful in respecifying a poor and freeing all factor loadings(see Jdreskog, 1979). fitting model (see Sdrbom, 1975; Jdreskog, 's We also performed a sensitivity analysis on 1979). these error variances (see Duncan, 1975:110).With Table 1 lists the goodness-of-fittests for each the exception of FRPICKUP, varying the validity model estimated. Model 1, derived from con- coefficients from .95 to .60 does not alter the substantive picture in any meaningfulsense. However, trol theory and multiple factor theory, allows with a validityfixed at .70, FRPICKUPhas a direct each variableantecedent to DEF to affect de- effect on delinquency statistically distinguishable linquency directly and also restricts all mea- from zero, but still dwarfedby the effect of DEF; at surementerrors to be uncorrelated.The x2 is .60, multicollinearityprevents stable estimation.We over four times the degrees of freedom, indi- believe that the reliability estimates we chose for cating a relatively poor fit (p = .000). Model 2 these indicators are the most plausible values. Moreover, the values for SES, FRPICKUP, and DEL are similar to previously reported estimates 16 Estimationof a model based on a pairwisepre- (see Bielby et al., 1977;Hindelang et al., 1981). For sent covariance matrix yielded similar substantive metricmodels, reliabilitieswere fixed to an arbitrary results. value by fixing error variances:o-2 = (1-Reliability) 17 By using the full information given by the o-2,where o-2 is the varianceof the observable. For model, this strategyprovides efficient parameteres- standardizedmodels, reliabilitiesare fixed by fixing timatesbut has the drawbackthat misspecificationin Pye = 1 - Reliability,where Pye is the squaredpath one portion of the model can spill over and bias coefficient of the measurementdisturbance. estimates of another portion. TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION 497

Table 1. Chi-square Goodness-of-Fit Tests for Models Estimated: Nonblack Males (N = 1,140)

Model X2 df p X2Idf (1) No correlated errors, 8 direct effects on DEL 368.84 89 .000 4.14 (2) 9 correlated errors within DEF, 8 direct effects on DEL 222.76 80 .000 2.78 (3) 9 correlated errors within DEF, 9 correlated errors between constructs, 8 direct effects on DEL 97.65 71 .020 1.38 (4) 9 correlated errors within DEF, 9 correlated errors between constructs, 1 direct effect on DEL 107.56 78 .016 1.38

allows certain measurementerrors in the indi- allowing for error correlations between indicators of DEF to correlate positively. Adding cators of SES, AGE, FRPICKUP, and PAR- these nine parametersinto the model substan- WHERE to certain indicatorsof ATTACHPE tially improves the fit: the difference in values and DEF were added (see Appendix B). The of 146.09 with nine degeees of freedom is resulting model fits the data reasonably well statisticallysignificant (p = .000) and the X2/df (X2 = 97.65, df = 71, p = .020). Althoughthere ratio is substantially reduced. Consequently, was no strong a priori theoretical reason for the hypothesis of purely randommeasurement including these parameters, they do appear error is rejected in favor of Model 2. (The pa- plausible, and, moreover, they work to at- rameter estimates of the measurementmodel tenuate importantrelations that were explicitly are discussed in Appendix B.) derived from the theoretical perspective. Model 2, however, still shows only a margi- Therefore, the results of Model 3, which are nal fit to the data (X2= 222.76, df = 80, p = reportedhere, are the more conservative find- .000). But with a large sample size and a large ings, and thus the stronger test of differential number of overidentifying restrictions, even association theory. small residuals can produce a significant x2 Table 2 presents the unstandardizedcoeffi- (Jdreskog, 1969, 1979). Thus, the lack of fit cient estimates for the substantiveequations in may simply representsampling variability. On their reduced, semireduced, and structural the other hand, it could indicatea misspecified forms; their standardizedcounterparts appear model, which, in turn, could produce biased in Table 3. The discussion will focus on those and inconsistent parameter estimates. For findingsmost significantfor our examinationof example, errorcorrelations between indicators differentialassociation theory. For equations of DEF and one or more backgroundvariables predicting the unobservable variable (DEF) could resultfrom social desirabilityeffects (see underlyingthe indicators of definitions of the Costner, 1969). To test for this, the first-order legal code, parameterestimates are presented derivatives were examined to locate those re- in lines 7 through 10. Of the backgroundvari- strictions that may be untenable. In so doing, ables, YOUNGTRO, followed by AGE, has care was taken to avoid contradictingthe logic the largesttotal impacton DEF (line 7 of Table of the strongtheoretical framework guiding the 3). As predicted by differential association initialmodel specification,and to avoid adding theory, boys who are older, who are from more parameters that were theoretically implausi- modest socioeconomic backgrounds,who are ble.'8 from broken homes, and who perceive more Proceeding incrementally, nine parameters trouble in their neighborhoods,are exposed to more definitionsfavorable than unfavorableto

18 delinquency. In general, a poor fit may result because the Together, the backgroundvariables explain number of factors is untenable, the hypothesized structureis untenable,or both (Jdreskog,1969). Be- 19 percent of the variance in DEF; adding cause of the strongtheoretical reason for specifying SUPER to the reduced form increases this to DEF as a unidimensionalconstruct, attemptsto im- 39 percent (line 8). For the most part, SUPER prove the model's fit by introducingadditional fac- affects DEF directly (comparelines 8 and 10). tors were not made. For this reason, too, the final Presumably, close parental supervision in- model relaxes constraints involving error correla- creases boys' exposure to antidelinquentbe- tions across constructs, rather than constraints on havior patternsin the home, and thus reduces direct effects from certain backgroundvariables to DEF. Nearly all of the effect of BROKHOME various indicators. From the standpointof restric- on DEF is mediatedby SUPER (comparelines tions on the covariance matrix, the two are nearly identical. If, however, the latter constraints are 7 and 8). Thus, boys from broken homes learn relaxed-allowing for a second form of differential slightlymore delinquentpatterns because their bias (Costner, 1969)-the substantive picture re- parents supervise them less. In addition, be- mains unaltered. Furthermore,principal axis factor cause they are less supervised, older boys and analysis of the seven indicatorsof definitionsof the boys from trouble-riddenneighborhoods are law supportsa unidimensionalsolution. exposed to slightly more delinquent than an- 498 AMERICAN SOCIOLOGICAL REVIEW

o o o > N ot _0- oo oo 00

CZ CZ:< en ON F o _ b U en o Io t V) o> c, b - en o _ _ en en en en VI) VI)

= VI crO- o en r- t- t- en vo I", ao = oo eq 'I 0 b O oo oo t- ? o V) en 0 oo oo oo o Eob en _ o en en t- W) X o VI)m b~~~~~~~~~~~~~~

S~~~~~~r 1- 0 r- - ON ---N

8~~~~~~~~~~~~~~~~~~~"0 r

S O < t o N m ? N r0~~~~~~~~~~~~~~e t- 00 00i S ~~~~~- V)e en entt -- . CZ Z.....NooN o oNo o o No>o?

D~~~~~~~~~ to C's z~~~~~~~~~~~~~~~aO7 ONtt>X -O m-N tUe

en > _nn888_~oooo__~

X~~~w' o o N .o ) eno eno0 eno- o,- o- o- aso ooo o o o4 I",o o C

E~~~~" " "I - en t' t' " env " "^ en ^- ON ^ 00 o0 o^ ONW) V^

rs o~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~' cd

cl ts _ CN- Pa 0 0 00 o _ N e cd TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION 499

Table 3. Standardized Parameter Estimates of the Substantive Model 3: Nonblack Males (N = 1,140) Predetermined Variables Dependent BROK- Variable AGE SES HOME YOUNGTRO SUPER FRPICKUP ATTACHPE DEF 1. SUPER -.128 -.018 -.126 -.275 2. FRPICKUP .243 -.043 -.032 .226 3. FRPICKUP .209 -.047 -.066 .152 -.268 4. ATTACHPE .105 .095 .065 -.190 5. ATTACHPE .137 .099 .096 -.121 .252 6. ATTACHPE .183 .089 .082 -.088 .193 -.220 7. DEF .152 -.029 .064 .393 8. DEF .090 -.036 .003 .260 -.481 9. DEF -.016 -.014 .037 .184 -.344 .508 10. DEF .035 .011 .059 .159 -.291 .446 -.279 11. DEL .105 .008 .052 .324 12. DEL .067 .003 .014 .244 - .293 13. DEL -.034 .026 .047 .170 -.163 .485 14. DEL -.017 .034 .054 .162 -.144 .464 -.095 15. DEL -.040 .027 .014 .054 .053 .162 .094 .678 tidelinquent definitions (compare lines 7 and These are importantfindings, for the empirical 8). test of differential association-the ability of Also consistent with differentialassociation DEF to mediate the antecedent variables' ef- theory, number of friends picked up by the fects on delinquency-requires something police has a large positive total impacton DEF nontrivialto be mediated. (line 9 of Table 3). In fact, FRPICKUPhas the Adding DEF to the equation predictingde- largestrelative direct effect (line 10of Table 3). linquency gives the structuralform (line 15). In addition, FRPICKUP mediates nontrivial The impact of DEF on delinquencyis negative amounts of AGE, YOUNGTRO, and SUPER and, as revealed by the standardizedcoeffi- in their effects on DEF. Therefore,being older, cients (line 15 of Table 3), comparativelylarge. being in a neighborhoodperceived to be more Thus, as differential association theory trouble-ridden, and being supervised less specifies, increasingthe numberof definitions causes boys to acquireslightly more delinquent favorable to violation of law relative to unfa- friendswhich, in turn, increases their exposure vorable definitions increases delinquent be- to delinquent definitions. Adding FRPICKUP havior. The model accounts for over half of the into the equation increases the R2 to over .60. variance in delinquency. The negative effect of ATTACHPEon DEF The hypothesis, derived from differentialas- (line 10) indicates that, on the average, in sociation theory, that DEF mediates the other closer friendships a higher proportion of an- variables' effects on delinquency, can be as- tidelinquentdefinitions is transmitted.As a re- sessed by comparinglines 14 and 15. Line 14 sult of being more attached to their friends, reveals that before adding DEF to the equa- boys who are from more trouble-free neigh- tion, AGE and SES already have trivial direct borhoods, who are more closely supervised, effects on delinquency. On the other hand, and who have fewer delinquentfriends tend to YOUNGTROhas a substantialand statistically learn more antidelinquentdefinitions relative significant direct effect. But when DEF is to delinquentdefinitions. With the addition of added to the equation, the effect becomes both ATTACHPE (line 10), the structural form trivial in size and statisticallyindistinguishable equation explains over 67 percent of the vari- from zero (line 15). Therefore, as differential ance in DEF. Thus, the model does well in association predicts, perceptions of neighbor- accounting for variabilityin the unobservable hood trouble increase delinquencyby increas- underlyingdefinitions of the legal code. ing the probabilityof learning delinquent be- Lines 11 through 15 present estimates of the havior patterns. While modest in size and not parameters predicting delinquency (DEL). quite significant, the direct effect of While the backgroundvariables explain 12 per- BROKHOMEon delinquencyis also mediated cent of the variance in delinquency (line 11), by DEF (compare lines 14 and 15). adding SUPER, FRPICKUP, and AT- As both Hirschi and Jensen found, and as TACHPEincreases this to 40 percent (line 14). control theory predicts, attachmentto parents, Without DEF, then, the model still accounts indicatedby SUPER, has a negative effect on for a substantial amount of variance in delin- delinquency unmediated by variables repre- quency. Also, the total effects of each variable senting peer relationships(line 14). However, except SES and BROKHOMEare nontrivial. contraryto control theory, but consistent with 500 AMERICAN SOCIOLOGICAL REVIEW differential association theory, this effect is measurementerror in certain variablesis con- entirely mediatedby DEF. In fact, not only is sidered. This is a significant issue, for had the remainingeffect (line 15) statistically non- errors in measurement not been considered, significant, but it is opposite in sign to that the estimates would have been biased and the predictedby control theory. Thus, delinquent substantiveconclusions misleading. behavior is reduced by parental supervision because boys are exposed to more antidelinquent definitionscompared to delinquentdefi- DISCUSSION nitions. According to the argument developed here, Also consistent with Hirschi and Jensen's contrary to popular criticisms of differential results, the semireduced form effect of association, the theory can and should be in- FRPICKUPon delinquency is large and posi- vestigated empirically.The oft-cited difficulty tive (line 14). But, unlike Jensen's finding, of operationalizingthe ratio of learned behav- DEF mediates all but a trivial and statistically ior patterns, so crucial to the theory, can be nonsignificantportion of this effect (line 15). conceptualizedas a problem of measurement Because delinquent friends affect boys' delin- error. Stated in this way, the problemcan be quent behavior only by increasingtheir expo- addressed by explicitly modeling the measure to delinquent definitions, differentialas- surementerror structureof items tapping the sociation theory is supported over group- underlying definitions construct, which then process and situationally-induced-motives allows the testing of specific hypotheses de- theories. rived from the theory. The total effect of attachmentto peers is to After following this strategy, and examining reduce the number of delinquentfriends (line a number of alternativemodel specifications, 14). Again, this is consistent with Hirschi's the conclusion is clear: differentialassociation findings, and with control theory in general. theory is supported over control theory. Be- But with the additionof DEF into the equation, cause of limitationsin the data, however, sev- this direct effect becomes statisticallyindistin- eral caveats are in order. First, as mentioned guishablefrom zero, and, from the standpoint earlier, the causal ordering of the variables of control theory, implausibly positive (line could be incorrect, since the data are cross- 15). Again, differentialassociation is supported sectional. If so, this could cause serious biases over control theory. in the parameterestimates. Second, both the In short, individual hypothesis tests of indicators and the outcome measure refer to point-intervalestimates indicate that DEF, the delinquent behavior in general, whereas latent variable underlying the indicators of Sutherland specified that specific forms of definitions of the legal code, successfully criminalor delinquentbehavior, such as theft mediates effects that other structuralvariables or assault, are determinedby specific ratios of have on delinquency. In addition, the more behaviorpatterns pertaining to those classes of global likelihood-ratiogoodness-of-fit test can acts (Sutherland, [1944] 1973:36).This study be applied to these hypotheses (see Table 1). has only had access to general indicators of By constraining each coefficient-tested definitions of the legal code; consequently, individually-to equal zero, Model 4 (derived these have been used to predicta generalindex from differential association) can be tested of delinquent behavior. Accordingly, it has against Model 3 (the less restricted alterna- been assumed that any differences between tive). The differencein x2 values of 9.48 with 7 these results using generalindicators and those degrees of freedom (compare lines 3 and 4 in obtaining from specific indicators (had they Table 1) is not statisticallysignificant (p > .20). been accessible) are random. Collectively, these parametersfail to improve Third, recent research has identifiedcertain the model's fit significantly;Model 3 is conse- problems in using self-reported measures of quently rejected in favor of the differentialas- relatively trivial delinquent offenses-as are sociation model. In addition, in Model 4, defi- used here-to make statementsabout all delin- nitions of the legal code (DEF) has a stan- quent behavior. For example, Elliot and Age- dardized effect on delinquency of .738, and ton (1980) note that the use of trivial items, explains 54 percent of the variation in delin- items that overlap, and items with truncated quent behavior. scales can mask significant relationships be- To sum up, tests of the specific hypotheses tween delinquency and certain demographic formulated here confirm the theory of dif- variables (see also Braithwaite, 1981). Hin- ferential association. This contrasts with the delang et al. (1979, 1981) conclude that while findings of Hirschi and Jensen who assumed such measures fail to capture relations that that all variables were measured perfectly. hold amongcertain demographicvariables and Using recently developed methodological official measures of delinquentbehavior, they techniques, this assumption is relaxed and remain,within the specific domainof behavior TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION 501

they are intended to tap, valid and reliable in- In sum, this study has located possible dicators of offending behavior. Consequently, sources of unreliability in the Richmond data generalizationof the findings reported here to and also has explicitly modeled that unreliabil- more serious forms of delinquencyis probably ity by relaxing a number of untenable restric- unwarrantedwithout furtherresearch. tions. This is important, for previous work that Fourth, this analysis is necessarily based on ignored these issues has produced misleading Hirschi and Jensen's assumptions about the results. Thus, if one chooses to reject the content validity of the indicators of DEF. Richmond data outright, this research locates Ideally, the indicators should be consistent possible grounds for doing so. If, on the other with Sutherland'sintended meaning of "defi- hand, one chooses to accept these data (which nitions favorable and unfavorableto violation are still perhaps the best data addressing the of the legal code." Further,they should mea- theoretical issue at hand and easily the most sure, at the very least, the most dominantdefi- widely cited) this research exploits these data nitions of the legal code communicated,trans- more fully than previous work, takes various mitted, and applied to delinquent behavior forms of unreliability into account, and thus within the population studied. The content of provides an improved test of control theory such definitions is likely to vary across sub- versus differential association. culturalgroups, where communicationis dis- tant and impersonal,and yet remain invariant APPENDIX A within subcultural groups and geographic areas, where communicationis relativelyopen, KEY TO VARIABLE LABELS intimate, and personal. AGE Age of respondent. 0 = 12 years or Although the measures used here appear younger, 1 = 13 years, 2 = 14 reasonably consistent with Sutherland's spe- years, 3 = 15 years, 4 = 16 years, 5 cification, it is not certain that they are appro- = 17 years, 6 = 18 years, 7 = 19 priate for this particularpopulation, because years, 8 = 20 years or older. the content of the items was determined de- SES Father's occupation measured on the DuncanScale; if there is no fa- ductively from a priori theoretical and con- ther living in the home, then moth- ceptual considerations, ratherthan also induc- er's occupationis used. For the few tively from a subset of the respondents them- cases in which father's occupation selves. had a missing value, and father's Finally, some of the evidence reportedhere education was reported, values of raises questions about the validity and reliabil- father's occupation were predicted ity of the indicatorsof DEF (see Appendix B). by regressing occupation on educa- The relatively low reliability coefficients are tion. similarto those reportedfor other social psy- BROKHOME A dummy variable coded as one if either the mother or father did not chological data (cf. Miller et al., 1979; and live with the respondent. Isaac et al., 1980),and in fact supportthe very FRPICKUP "Have any of your close friends reason for disentanglingunreliability from ac- ever been picked up by the police?" tual variation in the theoretical construct 0 = no or don't know, 1 = one (Kohn and Schooler, 1978). Nevertheless, the friend has, 2 = two friends have, low reliabilitiescould also signal that the indi- 3 = three friends have, 4 = four or cators have undesirable measurement prop- more friends have. erties. Furthermore,although it is unreason- PARWITH A composite asked regarding each able to expect these measurementerrors to be parent: "Do your parents know correlated who you are with when you are uncorrelated, the finding of many away from home?" 0 = never- errors among indicators within and between never, 0.5 = sometimes-never, theoretical constructs could suggest problems 1.0 = sometimes-sometimes, of validity. Ultimately, however, questions 1.5 = usually-sometimes 2.0 = about validity should be resolved on theoreti- usually-usually. cal rather than statistical grounds.19 PARWHERE Same as above but with the question: "Do your parents know where you are when you are away from 19 In the strategy used here, specifying the items home?" as fallible indicators of an unobservable variable BELIKFR "Would you like to be the kind of amounts to correcting for attenuation due to unrelia- person your best friends are?" 0 = bility, a procedure requiring reasonably valid and not at all, I = in a few ways, 2 = in reliable measures (Crouse et al., 1979:359-63). If in most ways. fact these indicators do not tap the theoretical do- RSPECTFR "Do you respect your best friend's main specified by Sutherland, any use of them for opinions about the important things examining differential association theory is unwar- in life?" 0 = not at all, 1 = a little, ranted. 2 = pretty much, 3 = completely. 502 AMERICAN SOCIOLOGICALREVIEW

DEL An index of delinquencycommitted YOUNGTRO [In my neighborhood] "Young during the last year containingthe people are always getting into following six items: trouble.' THEFT2 "Have you ever taken little things EVNBREAK "Policemen try to give all kids an (worth less than $2) that did not even break." belong to you?" DELHURT "Most things that people call 'de- THEFT250 "Have you ever taken things of linquency' don't really hurt any- some value (between $2 and $50) one." that did not belong to you?' OKLAW "It is all rightto get aroundthe law THEFT50 "Have you ever taken things of if you can get away with it." large value (worth over $50) that RSPECTPO "I have a lot of respect for the did not belong to you?" Richmondpolice." CARTHEFT "Have you ever taken a car for a GETAHEAD "To get ahead, you have to do some ride without the owner's permis- things which are not right." sion?" TROUBLE "I can't seem to stay out of trouble VANDALSM "Have you ever banged up no matter how hard I try." somethingthat did not belong to SUCKERS "Suckers deserve to be taken ad- you on purpose?" vantage of." BATTERY "Not counting fights you may have had with a brotheror sister, have you ever beaten up on any- APPENDIX B one or hurtanyone on purpose?" Analysis of the Measurement Model The following items are all measuredon the scale, Parameterestimates of the measurementmodel ap- "stronglyagree, agree, undecided,disagree, strongly pear in Table 4. All unrestrictedmetric slopes (listed disagree." in column3), except those of EVNBREAKand DEL-

Table 4. Measurement Parameter Estimates of Model 3: Nonblack Males (N= 1,140)

Variable (1) (2) (3) (4) (5) ______Observed Error Metric Validity Reliability Latent Observable Variance Variance Slope Coefficient Coefficient AGE AGE 3.012 .277f 1.000q .94Y .90() SES SES 5.663 2.037f 1.OW0 .80() .640f BROKHOME BROKHOME .204 .073f 1.0W .80() .640f YOUNGTRO YOUNGTRO .932 .336f 1.000f .800' .640f SUPER PARWITH .292 .125 1.00o0 .756 .571 (.014) PARWHERE .279 .132 .938 .725 .526 (.012) (.075) FRPICKUP FRPICKUP 2.235 .804f 1.00O .800 .640f ATTACHPE BELIKFR .374 .289 1.00O0 .477 .227 (.019) RSPECTFR .499 .319 1.456 .600 .361 (.033) (.244) DEF EVNBREAK 1.500 1.361 .904 .301 .091 (.060) (.128) DELHURT 1.123 .989 .904 .348 .121 (.044) (.113) OKLAW 1.079 .834 1.213 .476 .227 (.042) (.130) RSPECTPO 1.189 .893 1.327 .496 .246 (.044) (.136) GETAHEAD 1.222 .988 1.183 .436 .190 (.047) (.130) TROUBLE .967 .801 1.00o .415 .172 (.037) SUCKERS 1.219 1.010 1.132 .418 .175 (.048) (.120) DEL DEL 1.200 .432f 1.000' .800 .640f Fixed coefficient. Note: Standard errors appear in parentheses. The following measurement errors were found to be significantly correlated: p(ele8) = -.433 p(elel0) = .467 p(eje12) = .265 p(ele15) = -.467 p(elel6) = -.262 p(e2e 1) = -.186p(e2el3) = .137p(e,3e15)= .098p(e7e,0) = .166p(e10el3) = .308p(el0el4) = .077p(ellel6) = .089 p(ejje12) = .159 p(el2el3) = .054 p(el2el4) = .130 p(el2e16) = .104 p(el4e16) = .107 p(e,5e1,6)= .127. TESTING CONTROL THEORY AND DIFFERENTIAL ASSOCIATION 503

HURT, depart significantly (by at least one stan- Bielby, WilliamT. and Robert M. Hauser darderror) from the normalizedvalue of unity. For 1977 "Structuralequation models." Annual Re- the indicatorsof definitions of the law (DEF), this view of Sociology 3:137-61. indicatesthat, relative to TROUBLE, persons scor- Bielby, WilliamT., Robert M. Hauser and David L. ing highly prodelinquenton each of these other indi- Featherman cators tend to understatetheir prodelinquency,and 1977 "Response errors of black and nonblack vice versa. For the indicatorsof parentalsupervision males in models of the intergenerational (SUPER), which like those of DEF are measuredon transmission of socioeconomic status." identical scales, relative to PARWITH, persons AmericanJournal of Sociology 83:1242-88. scoring highly supervised on PARWHEREtend to Bohrnstedt,George W. understatetheir degree of parentalsupervision, and 1970 "Reliabilityand validity assessment in at- conversely. titude measurement."Pp. 80-99 in Gene F. As is commonly the case with measures of atti- Summers (ed.), Attitude Measurement. tudes, the indicators of attachment to peers (AT- Chicago: Rand McNally. TACHPE) and definitions of the legal code (DEF) Braithwaite,John containlarge portionsof measurementerror (column 1981 "The myth of social class and criminality 2). In general, larger validity (indicator-construct reconsidered.' American Sociological Re- correlations)and reliability(squared validities) coef- view 46:36-57. ficients than are found here are desirable. These Briar, Scott and Irving Piliavin findingsraise questions about the qualityof the indi- 1965 "Delinquency, situational inducements, cators, and the Richmond data in general. On the and commitment to conformity." Social other hand, given that random measurementerror Problems 13:34-45. was assumed to be generatedby both domain sam- Costner, Herbert L. pling processes and imperfectionsin the instrument, 1969 "Theory, deduction, and rules of corre- low reliabilitywas expected. spondence." American Journal of Sociol- The relative accuracy of the indicatorscan be as- ogy 75:245-63. sessed by evaluatingthe error variances (column 2) Cressey, Donald R. and the validity and reliabilitycoefficients (columns 1952 "Application and verification of the dif- 4 and 5). PARWITHand PARWHEREare both rea- ferential association theory." Journal of sonably reliable indicators of parental supervision. CriminalLaw, Criminologyand Police Sci- As a measure of attachmentto peers, RSPECTFR ence 43:43-52. appears more reliable than BELIKFR (columns 4 [19531 Other People's Money: A Study in the So- and 5). The indicatorsof definitionsof the law, all 1973 cial Psychology of Embezzlement. measuredon the same scale, have error variances Montclair,NJ: PattersonSmith. around1.0, showingsome stabilityacross indicators. 1960 "Epidemiology and individual conduct: a TROUBLE and OKLAW appear to be particularly case from criminology." Pacific Sociologi- accurate measures, while EVNBREAK does not. cal Review 3:47-58. The validity and reliability coefficients reveal that Crouse, James, Peter Mueser, ChristopherJencks these indicators have similar reliabilities, with and Charles S. Reichardt EVNBREAKand DELHURT being slightly less re- 1979 "Latent variable models of status attain- liable.The indicatorsused by Jensen (EVNBREAK, ment." Social Science Research 8:348-68. DELHURT, OKLAW, and GETAHEAD)are about DeFleur, Melvin L. and RichardQuinney average in reliability. 1966 "A reformulationof Sutherland'sdifferen- As expected, the largest error correlationamong tial associationtheory and a strategyfor em- indicators of DEF involve EVNBREAK and piricalverification." Journal of Researchin RSPECTPO,which both refer to attitudes toward Crime and Delinquency3:1-22. the police. For pairs of indicatorsin which at least Duncan, Otis Dudley one errorvariance is fixed, ratherthan estimated, the 1975 Introductionto StructuralEquation Mod- errorcovariance is an identifiableparameter, but not els. New York: Academic Press. the error correlation. Thus, the magnitude of the Elliot, Delbert S. and Suzanne S. Ageton error correlationdepends on the value of the error 1980 "Reconciling race and class differences in variance(s)fixed. Consequently, the error correla- self-reportedand official estimates of delin- tions involving AGE are anomalously large due to quency." American Sociological Review the small error variance fixed for AGE. 45:95-110. Hepburn,John R. 1976 "Testingalternative models of delinquency REFERENCES causation." Journal of Criminal Law and Akers, Ronald L., Marvin D. Krohn, Lonn Lanza- Criminology67:450-60. Kaduce and Marcia Radosevich Hindelang,Michael J. 1979 "Social learning and deviant behavior: a 1973 "Causes of delinquency: a partial replica- specific test of a generaltheory." American tion and extension." Social Problems Sociological Review 44:635-55. 20:471-87. Austin, Roy L. Hindelang,Michael J., Travis Hirschiand Joseph G. 1977 "Commitment, neutralization, and delin- Weis quency." Pp. 121-37 in Theodore N. Fer- 1979 "Correlates of delinquency." American dinand (ed.), Juvenile Delinquency: Little Sociological Review 44:995-1014. Brother Grows Up. Beverly Hills: Sage. 1981 MeasuringDelinquency.Beverly Hills: Sage. 504 AMERICAN SOCIOLOGICAL REVIEW

Hirschi, Travis Journal of Research in Crime and Delin- 1969 Causes of Delinquency. Berkeley: Univer- quency 18:295-318. sity of CaliforniaPress. Nunnally, Jum C. Isaac, Larry, ElizabethMutran and Sheldon Stryker 1967 Psychometric Theory. New York: 1980 "Politicalprotest orientationsamong black McGraw-Hill. and white adults." American Sociological Reiss, Albert J., Jr., and Lewis Rhodes Review 45:191-213. 1964 "An empirical test of differentialassocia- Jensen, Gary F. tion theory." Journalof Research in Crime 1972 "Parents, peers, and delinquent action: a and Delinquency 1:5-18. test of the differentialassociation perspec- Short, James F., Jr. tive." American Journal of Sociology 1957 "Differentialassociation and delinquency." 78:562-75. Social Problems4:233-39. Johnson, RichardE. 1958 "Differential association with delinquent 1979 Juvenile Delinquencyand its Origins.Cam- friends and delinquent behavior." Pacific bridge: CambridgeUniversity Press. Sociological Review 1:20-25. Jdreskog, Karl G. 1960 "Differential association as a hypothesis: 1969 "A general approach to confirmatory problems of empirical testing." Social maximum likelihood factor analysis." Problems8:14-25. Psychometrika34:183-202. Short, James F., Jr., and Fred L. Strodtbeck 1976 "Causal models in the social sciences: the 1965 Group Process and Gang Delinquency. need for methodologicalresearch." Univer- Chicago: University of Chicago Press. sity of Uppsala, Departmentof Statistics Sbrbom, Dag Research Report 76-8, Uppsala, Sweden. 1975 "Detection of correlated errors in lon- 1979 "Analyzingpsychological data by structural gitudinal data." British Journal of Mathe- analysis of covariance matrices." Pp. 45- matical and Statistical Psychology 100 in Karl G. Joreskog and Dag Sorbom 28:138-51. (eds.), Advances in Factor Analysis and Stanfield, Robert E. Structural Equation Models. Cambridge: 1966 "The interaction of family variables and Abt Books. gang variables in the aetiology of delin- J6reskog, Karl G. and Dag Sorbom quency." Social Problems 13:411-17. 1978 LISREL IV: Analysis of Linear Structural Sutherland,Edwin H. Relationshipsby the Method of Maximum [1942] "Development of the theory." Private Likelihood. Chicago: National Education 1973 paper published posthumously. Pp. 13-29 Resources, Inc. in Karl Schuessler (ed.), Edwin Sutherland Kohn, Melvin L. and Carmi Schooler on Analyzing Crime. Chicago: University 1978 "The reciprocal effects of the substantive of Chicago Press. complexity of work and intellectual flexi- [1944] "Critique of the theory." Private paper bility: a longitudinalassessment." Ameri- 1973 publishedposthumously. Pp. 30-41 in Karl can Journalof Sociology 84:24-52. Schuessler (ed.), Edwin Sutherland on Kornhauser,Ruth R. Analyzing Crime. Chicago: University of 1978 Social Sources of Delinquency. Chicago: Chicago Press. University of Chicago Press. 1947 Principlesof Criminology.Fourth Edition. Philadelphia:Lippincott. Krohn, Marvin D. Sutherland,Edwin H. and Donald R. Cressey 1974 "An investigationof the effect of parental 1978 Criminology. Tenth Edition. Philadelphia: and peer associations on marijuanause: an Lippincott. empirical test of differential association Sykes, Greshamand David Matza theory." Pp. 75-89 in Marc Reidel and Ter- 1957 "Techniquesof neutralization:a theory of rence P. Thornberry(eds.), Crimeand De- delinquency." American Sociological Re- linquency: Dimensions of Deviance. New view 43:643-56. York: Praeger. Voss, Harwin L. Miller,Joanne, CarmiSchooler, MelvinL. Kohn and 1964 "Differential association and delinquent Karen Miller behavior." Social Problems 12:78-85. 1979 "Woman and work: the psychological ef- Wheaton,Blair, Bengt Muthen,Duane F. Alwin and fects of occupational conditions." Ameri- Gene F. Summers can Journalof Sociology 85:66-94. 1977 "Assessing reliabilityand stability in panel Minor, W. William models." Pp. 84-136 in David Heise (ed.), 1981 "Techniques of neutralization: a recon- Sociological Methodology 1977. San Fran- ceptualizationand empiricalexamination." cisco: Jossey-Bass.