Psychological Bulletin Copyright 1983 by the

1983, Vol. 94, No. 1,68-99 American Psychological Association, Inc.

Early Predictors of Male Delinquency: A Review

R. Loeber and T. Dishion
Oregon Social Learning Center, Eugene, Oregon

A systematic review is presented of prediction studies on delinquency. The main

aim is to identify etiological variables for delinquency that, in different studies
and across different populations, show good predictive validity. To achieve this
goal, a measure of predictive efficiency was chosen that could be applied to studies
from the United States and from abroad. The principal predictors of delinquency
were the parents' family management and techniques (supervision and discipline),
the child's conduct problems, parental criminality, and the child's poor academic
performance. Data are presented to show the earliest age of the child at which
these predictors have been measured. Results of the prediction data are used to
demonstrate utility functions in which false positive and false negative errors are
minimized. Recommendations are put forward to improve prediction studies in
criminality.

Claims about the early identification of Prediction in criminology serves two main
youths at risk for delinquency have some- purposes. More accurate prediction helps
times been extravagant. The Gluecks (Glueck parents, teachers, court officials, and thera-
& Glueck, 1950) claimed to be able to cor- pists to take adequate action when discov-
rectly identify 90% of all future delinquents. ering early warning signs of a juvenile delin-
The psychiatrist Glover, testifying before the quent career. Second, known predictors can
Royal Commission on Capital Punishment be used for the construction of theories of
(1949, cited in Hakeem, 1957-1958), ex- delinquency. Here some predictors are
pressed the belief that psychometric tests equated with causal and moderating factors
could do wonders. He stated that "if suffi- that may interact and form constellations of
cient trouble were taken, pathological cases factors likely to bring about delinquency.
liable to commit murder could be detected Theory building and predictive assessment
during early childhood" (p. 492). However, have a major common element: the selection
prediction research has not substantiated and weighting of variables in terms of their
these claims (Hakeem, 1957-1958). More- explanatory power for delinquency. Some
over, the research by Glueck and Glueck variables may only account for small differ-
(1950, 1959), with substantial claims of pre- ences between delinquents and nondelin-
dictability of delinquency, has been criticized quents, whereas other variables may account
on methodological grounds (Hirschi & Sel- for much of the variance between delinquent
vin, 1967;Prigmore, 1963;Reiss, 1951;Weis, and nondelinquent groups.
1974). On the more positive side, there is a Prediction in delinquency is possible for
large and rather scattered body of studies with two reasons: the first is that problem behav-
encouraging results for the identification of iors of children have a high degree of conti-
youths at risk for later delinquency. These nuity over time. Problematic conduct early
studies will be reviewed here. in life for certain groups of children tends to
continue rather than abate (Gersten, Lang-
The authors are indebted to L. R. Goldberg, D. J.
ner, Eisenberg, Simcha-Fagan, & McCarthy,
Farrington, J. C. Gersten, R. F. Sparks, and an anony- 1976; Ghodsian, Fogelman, Lambert, &Tib-
mous reviewer for their helpful comments. The paper benham, 1980; Loeber, 1982; Olweus, 1979;
was the result of frequent and fruitful discussions with Patterson, 1982; Robins, 1966; Werner &
the staff of the Oregon Social Learning Center, especially Smith, 1977; West & Farrington, 1973). The
G. R. Patterson, J. B. Reid, and P. Holleran.
Requests for reprints should be sent to R. Loeber,
challenge is to identify conduct problems that
Oregon Social Learning Center, 207 East 5th Avenue, precede delinquency and are ultimately pre-
Suite 202, Eugene, Oregon 97401. dictive of its occurrence. This will make it
68
 EARLY PREDICTORS 69

possible to identify on the basis of the con- and their implicit limitations for clinical or
duct problem those children at highest risk prevention activities. It was selected because
for delinquency. Another challenge is to es- of the care taken by the investigators to pro-
tablish the earliest age at which such conduct vide complete data that permitted additional
problems become predictive, so that preven- analyses. Robins and Hill (1966), in a well-
tive efforts can take place while the conduct designed investigation of theoretical impor-
problems are not yet firmly stabilized. One tance, studied 296 nonwhite youths and pos-
more challenge is to detect the interrelation- tulated that delinquency would be highest for
ships between different conduct problems youths with parents from a low employment
and to see whether this pattern is more pre- background. To begin with, they determined
dictive of delinquency than the individual the employment status of the child's guardian
components. using retrospective evidence from school rec-
The second reason prediction in delin- ords. The criterion of delinquency was de-
quency is feasible is that we know what sit- fined as a police or court record by the age
uations or social variables tend to enhance of 17. Ideally, the prediction of delinquency
children's eventual engagement in delinquent on the basis of employment status should
activities. Known circumstantial variables have a high degree of accuracy. Those chil-
are, for example, the social class of the par- dren with a guardian of low employment sta-
ents and the parents' child-rearing methods. tus should be more at risk for delinquency
The present article has three main goals: than those with a guardian of higher em-
(a) to establish a measure of predictive effi- ployment status. The results reported by Ro-
ciency that can be applied to the delinquency bins and Hill (1966) are shown in Figure 1:
prediction studies in our review, (b) to iden- 46 out of the 148 boys with guardians of low
tify variables from the literature that predict employment status became delinquents. This
delinquency in adolescence or in early adult- was compared with 30 out of 148 boys with
hood, and (c) to use the results of the pre- guardians of high employment status who
diction data to demonstrate the utility func- also became delinquent. Robins and Hill
tions of predictors and to minimize false pos- (1966) concluded that "about a third of the
itive or false negative errors. The article closes boys with . . . a lower status guardian even-
with methodological considerations and sug- tually became delinquent, compared with
gestions for improving prediction studies in one-fifth of the remainder" (p. 331). These
criminality. results provided support for the relationship
between economic stress and delinquency
Predictive Efficiency hypothesized by Robins and Hill (1966).
Researchers in criminology have used a Whereas such an analysis is extremely useful
variety of methods to assess the power or ef- for establishing theoretically useful relation-
ficiency of particular variables to predict la- ships, its potential application to issues of
tent delinquency (Gottfredson, 1970; Simon, prevention is limited (i.e., it would be useful
1971). However, predictive methods useable only for clinical situations in which one is
on the same population are not necessarily confronted with lower or higher class parents;
useable on different populations, because dif- the social class of the parent could then be
ferences in the base rates of delinquency be- used to predict the delinquency of the off-
tween groups affect most measures of pre- spring).
dictive efficiency. The following discussion Because it was not relevant to their theo-
outlines a method that is less sensitive to such retical hypotheses, the authors did not ex-
variations and that makes it possible to com- plicitly report on the total number of correct
pare the predictive efficiency of a wide variety predictions in the table, that is, those youths
of predictors over a wide range of studies. We who were assumed to be at risk (because of
will first explain how this method of predic- low employment background) who ulti-
tion has been developed and then review mately become delinquent, and those youths
studies by using this method. who were identified as not at risk (on the basis
The following example illustrates widely of higher employment background) who ul-
used means of assessing predictive efficiency timately did not become delinquent. Robins
70 R. LOEBER AND T. DISHION

and Hill's (1966) evaluation of predictive ef- (1981) has pointed out that for judicial de-
ficiency is not idiosyncratic, but very typical cisions about guilt, typically the percentage
for delinquency studies in general (e.g., of false negatives should be low; that is, the
Glueck & Glueck, 1950; West & Farrington, predictors should not miss youths who are
1973; Wadsworth, 1979). Moreover, the for- actual delinquents. For clinical decisions
mulation of predictive efficiency given above about treatment or prevention, the emphasis
does not discriminate between the types of is put on reducing false positives; that is, cap-
errors that are made in the process, nor do turing only those youths truly at risk for de-
conclusions about the prediction reflect linquency and minimizing the identification
chance and maximum limits in identifica- of those seemingly at risk but who do not
tion. become delinquent. Implicit here is the rel-
These points can best be illustrated by us- ative cost of certain types of errors for certain
ing prediction methods widely used in per- types of decisions. Sometimes costs and ben-
sonnel selection (Wiggins, 1973). First, the efits of decisions in predictions are weighted.
method should optimally identify youths The application of such utility values will be
who eventually become delinquent and iden- discussed in a later section. Until then, we
tify those who do not become delinquent. will treat prediction without discriminating
The first kind of correct identifications are between the relative utility of false positive
called valid positives, the second kind valid or false negative errors.
negatives (see Figure 1). Errors in identifi- Figure 1 summarizes the results obtained
cation are of two kinds: youths who are pre- by Robins and Hill (1966) in terms of valid
dicted to be at risk for delinquency but who versus false positives and negatives, which
do not become delinquent, false positives, have been calculated on the basis of the total
and youths who are not identified to be at number of subjects (N = 296). The percent-
risk for delinquency but later become delin- age of total correct predictions consists of the
quent, false negatives. Depending on one's sum of valid positives (15.5%) and valid neg-
priorities, the percentages of false positives atives (40%), which amounts to 55.4%. Thus,
and false negatives should be low. Monahan only about half of the youths were correctly

Delinquent Nondelinquent
Valid Po False Posith

Guardian's Low
46 102 148
Employment Status
(50%)

(15.5% I [34.5%)

False Negatives Valid Negatives

Guardian's High
Employment Status.
30 118 148

110.1%) (39.9%)

76 220 296

(25.7541 100%

Figure 1. The relationship between juvenile delinquency and guardian's employment status as shown by
the frequency of each. (Adapted from "Assessing the contribution of family structure, class and peer
groups to juvenile delinquency" by L. N. Robins and S. Y. Hill, Journal of Criminal Law, Criminology,
and Police Science, 1966, 57, 325-334. Copyright 1966 by Northwestern University School of Law.
Reprinted by permission.)
 EARLY PREDICTORS 71

identified on the basis of the guardian's em- .372, respectively. Translated into frequen-
ployment status. The delinquency outcome cies, the numbers expected by chance alone
of less than half of the subjects was not cor- amount to 38 valid positives (12.8%) and 110
rectly predicted. Most of these errors in pre- valid negatives (37.2%; see Figure 2). Taken
diction were false positives (34.5%), and fewer together, 148 (50%) of the subjects in these
were false negatives (10.1%). two cells could be predicted by a random se-
The simplistic logic of these computations lection of subjects on the basis of the mar-
is deceptive; the values in the 2 X 2 prediction ginal values in the prediction table.
tables are influenced by chance and by a The random correct value can be com-
maximum ceiling in prediction. We will dis- pared with the observed correct prediction
cuss each of these limiting factors. (see Figure 2). The observed valid positives
Chance occurrence of frequencies within and valid negatives together account for
a 2 X 2 table is a function of the marginal 55.4% of the correct predictions. The differ-
values of the table. These marginal values ence between the observed correct predic-
represent, on one side, the number of delin- tions and the random correct predictions re-
quents found in the population (called base veals the extent to which low employment
rate), and on the other side, the number of status of the guardian improved the identi-
individuals selected as delinquent by means fication of youths who would later become
of the prediction method (called selection ra- delinquent. In the given example, the percent
tio). In Figure 1, the base rate and selection improvement over chance was 5.4%. This
ratio, respectively, amounted to 25.7% and measure, as far as we know, has been used
50%. These proportions determine the chance in only one delinquency study (McCord,
occurrence of frequencies within the table 1980), despite its obvious advantages. How-
(Meehl & Rosen, 1955; Wiggins, 1973). Most ever, such a measure has different meanings
authors take chance occurrence into account from study to study. To control for this, an
by the calculation of chi square. In the ex- index was devised to represent the improve-
ample, Robins and Hill (1966) found that chi ment over chance as a function of the range
square equalled 3.98, which was significant of its possible predictive efficiency. The range
at the .05 level. However, this only means that is delimited by two values: the random cor-
one or more cells in the 2 X 2 table had rect and maximum correct percentages. As
frequencies that could not be expected by has been described in detail by Loeber and
chance alone. Although this knowledge is es- Dishion (Note 1), each identification table
sential, it fails to clarify which cell frequen- has a maximum value for the highest correct
cies deviate from what could be obtained by identifications possible within the table. This
chance alone. As Meehl and Rosen (1955) maximum ceiling is determined by the base
have stated, a significance test in this case rate and selection ratio. Figure 3 shows the
does not clarify "the number of correct de- best possible identification given the base rate
cisions for individuals within [delinquent and and selection ratio used by Robins and Hill
nondelinquent] groups" (p, 194). Two cells (1966). The number of subjects in the cell of
in prediction tables are most important for valid positives can only maximally be 76 be-
assessing predictive efficiency—the valid pos- cause of the base rate limitation of 76 out of
itives and the valid negatives, For that reason, 296. Once this maximum value is set, the
the degree that observed values in these cells frequencies in all other cells are fixed (see
deviate from random or chance values pro- Figure 3). The resulting maximum percent-
vides a more accurate assessment of predic- ages of valid positives and valid negatives are
tive efficiency than is possible by means of 25.7% and 50%, respectively. Taken together,
a chi-square measure. The calculation of ran- correct predictions in the study by Robins
dom correct prediction for valid positives and and Hill can never be higher than 75.7%.
valid negatives has been outlined by Wiggins Thus, the percent improvement over chance
(1973). In the example of Robins and Hill in a given study always falls between the ran-
(1966), the random correct values for valid dom correct value and the maximum correct
positives and valid negatives are 76/296 X value. As the difference between the latter two
148/296 = .128, and 220/296 X 148/296 = measures varies from study to study, it is ap-
72 R. LOEBER AND T. DISHION

Delinquent Nondelinquent

Valid Positives

Guardian's Low
38 148
Employment Status

112.8%)

Valid Neqafives
Guardian's High
Employment Status
110 148

137.2% I

76 220 296

Figure 2. Frequencies of valid positives and valid negatives expected by chance. (Adapted from "Assessing
the contribution of family structure, class and peer groups to juvenile delinquency" by L. N. Robins and
S. Y. Hill, Journal of'Criminal Law, Criminology, and Police Science, 1966, 57, 325-334. Copyright 1966
by Northwestern University School of Law. Reprinted by permission.)

propriate to express the improvement over culated as follows:

chance (IOC) as a function of the difference %IOC
between the random correct (RC) and max- RIOC = X 100
imum correct (MC) values in a given study. %MC - %RC
This value will be called relative improve- A major problem in the evaluation of pre-
ment over chance (RIOC), and can be cal- dictive efficiency is that it depends to a great
Delinquent Nondelinquent

Guardian's Low
76 72 148
Employment Status

(25.7%) (24.3%)

Guardian's High

Employment Status
148 148

(50%)

76 220 296
Figure 3. The distribution of cell frequencies maximizing correct predictions. (Adapted from "Assessing
the contribution of family structure, class and peer groups to juvenile delinquency" by L. N. Robins and
S. Y. Hill, Journal of Criminal Law, Criminology, and Police Science, 1966, 57, 325-334. Copyright 1966
by Northwestern University School of Law. Reprinted by permission.)
 EARLY PREDICTORS 73

extent on how well the selection ratio matches met this criterion because a number of stud-
the base rate in a given study (Loeber & Di- ies did not clarify the exact nature of pre-
shion, Note 1). A discrepancy between the dictor scoring. Excluded, for example, was
selection ratio and the base rate, which is very Powers and Witmer's (1951) report on the
common in delinquency studies, influences Cambridge-Somerville study, in which teacher
the magnitude of the maximum correct ratings were reported as a predictor, but these
value. The present use of the RIOC measure, ratings were, in fact, based on the researcher's
which is partly based on the maximum cor- interview of the teacher or on the researcher's
rect value, largely eliminates this problem. interpretation of written protocols of such
To test this, the IOC and RIOC indices were discussions with a teacher.
correlated with both the selection ratio and 5. Most of the studies referred to inde-
the base rate for each of the studies reviewed. pendent variables that predicted acceleration
It was thought that the best index of predic- or an increased probability of delinquency
tive efficiency would be the least correlated over time. We also included studies that pre-
with either of the selection parameters. As dicted the offset or decreased probability of
expected, the IOC correlated .54 and .38 and delinquency later in time.
the RIOC correlated. 13 and .22 with the base 6. The prediction and outcome variables
rates and selection ratios. We concluded that included in this review comprised not only
the RIOC measure is more independent of the incidence of police contacts youths had,
varying base rates and selection ratios and but also outcomes such as arrest rates, re-
therefore superior as an overall evaluative in- conviction rates, and high self-reported de-
dex of predictive efficiency. linquency. Wherever possible, we did not in-
clude traffic violations in delinquency out-
Inclusions and Exclusions in the come measures. Studies on parole violations
Present Review were excluded (see, e.g., Monahan, 1981;
An extensive search of the delinquency lit- Ohlin & Duncan, 1949).
erature resulted in a large number of studies 7. Both retrospective and prospective
with predictive data. A description of the studies were included.
studies included in this review follows: 8. Whenever data were presented by au-
1. The studies contained data that made thors in other than dichotomized fashion, we
it possible to reconstruct prediction tables as followed the cutting scores for predictor vari-
shown in Figure 1. Studies that contained ables mentioned by authors. If the authors
only percentage information and did not al- did not mention cutting scores, we set the
low the reconstruction of raw scores were cutting scores in such a way as to obtain the
excluded. most optimal improvement in prediction
2. The review focused on males, but stud- over chance.
ies on mixed male and female populations 9. Personality tests to identify youths at
were included. Studies solely on female pop- risk for delinquency were not included (see,
ulations were not included. e.g., Hathaway & Monachesi, 1953, 1963).
3. Only studies that contained predictors 10. Whenever recidivism studies were re-
that had taken place at least a year prior to viewed, the rate of recidivism was compared
the measurement of outcome of delinquency with the rate of one-time offenders rather
were included. Thus, whereas studies con- than with the rate of delinquency in general.
currently measuring independent and depen- It should be noted that the authors of the
dent variables are valuable for identification following studies do not all emphasize the
of potential predictors, such studies were not predictive utility of their studies. Many of
included in the present review. We included them have emphasized theoretical or empir-
only studies that predicted juvenile delin- ical aspects of their data. However, the fact
quency and delinquency in early adulthood; that independent variables in these studies
that is, we included only predictors and out- often preceded dependent variables in time
comes that occurred before age 22. allows us to examine the predictive efficiency
4. Only studies that used objective first- of a wide range of independent variables.
hand predictors were included. It was some- In the following review, Tables 1 through
times difficult to determine whether a study 10 supply information on the raw scores in
(Text continues on page 78)
Table 1
Outcome Statistics for Behavioral Predictors of Delinquency
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval R1OC

West & Farrington (1973)

Troublesoineness (teacher and 41 51 12.4 276 43 10.5 411 20.4 22.4 One or more adjudi- 8-10 b tol7 34.1
peer ratings)3 cated offenses
Farrington (1979)
Troublesomeness (teacher and 31 60 14.7 269 49 12.0 409 19.6 22.2 21 or more self- 8-10 to 14-16 21.1
peer ratings) reported acts
Craig & Click (1963) ?3
Problem behavior in Grades 35 70 23.3 187 9 3.0 301 14.6 34.9 "Serious and/or 6-9 to 16 66.3
1, 2, & 3 (teacher ratings) persistent" 5en
Mitchell & Rosa (1981) CD
Problem behavior (parent & 30 71 22.9 179 30 9.7 3 10 "deviant 19.3 32.6 One or more adjudi- 5-15 to 20+ 25.7 73
rn
teacher report) group" cated offenses >z
Kirkegaard-Serensen & Mednick o
(1977) H
Disciplinary problem (teacher 10 15 5.3 236 21 7.4 282 (select 11.0 8.9 Conviction 10-20 to 23-33 32.3
rating) sample) Eo
Robins (1966) 3
o
Antisocial referral 185 27 7.7 75 63 18.0 350 (racially 70.9 60.6 Arrest 14 to 18 48.2
mixed)
Stott & Wilson (1968) Marsh
(1969)
Delinquency Prediction scale 29 53 6.5 683 53 6.5 818 10.0 10.0 Conviction 18 to 21
(teacher rating)
Scarpitti (1964)
Potentially delinquent (teacher 27 43 24.9 99 4 2.3 173 17.9 40.5 Court contact 12 to 16 78.0
nomination)
Reckless & Dinitz (1972)
Likelihood of future 213 881 51.0 600 32 1.9 1726 (an experi- 14.2 63.4 Police contact 13 to 16 64.4
delinquency (teacher mental, con-
nomination) trol, & com-
parison
group)
Table 1 (continued)
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC

Simcha-Fagan(1979)
Mother-reported delinquency 32 54 13.5 581 62 11.7 729 12.9 11.8 Police contact 7 to 12 28.2
8 to 13
11 to 15
13 to 18
14 to 20
16 to 21
Roff, Sells, & Golden (1972)
Who is liked best (peer 20 76 5.5 1177 100 7.3 1373 8.75 7.0 Court contact 11-12 to 16 15.3
choice) w
Havighurst, Bowman, Liddle,
Matthews, & Pierce (1962)
Aggressiveness (teacher and 32 40 19.1 122 15 7.2 209 22.5 34.4 Police contact 12-13C to 18-19 51.4 js
peer ratings) w
O
West & Farrington (1973)
Aggressiveness (teacher rating) 23 59 15.1 264 45 11.5 391 17.4 21.0 One or more adjudi- 8-10 to 17 16.4
cated offenses 73
1/3

Kirkegaard-Sarensen & Mednick

(1977)
Violence & aggressiveness 7 20 7.1 231 24 8.5 282 (select 11.0 9.6 Conviction 10-20 to 23-33 17.0
(teacher rating) sample)

Easily angered (teacher rating) 13 50 17.7 201 18 6.4 11.0 22.3 25.2
Feldhusen, Thurston, & Benning
(1973)
Aggressiveness (teacher rating) 273 295 19.0 766 216 13.9 1550 (boys 31.6 36.7 Police record 9 to 17 30.3
& girls) 12 to 20
15 to 23
Mulligan, Douglas, Hammond,
& Tizard (1963)
Aggressiveness (teacher rating) 54 189 9.2 1700 120 5.8 2063 8.4 11.8 Conviction 13 to 15 21.5
Table 1 (continued)
-j
Os
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % ilegatives n % N PER PSR of delinquency interval RIOC

Mitchell & Rosa (1981)

Destructiveness (parent 12 14 4.4 244 51 15.9 321 "deviant 19.6 8.1 Court appearance 5-15 to 20+ 31.8
reported) group" for indictable
offense
Kirkegaard-Sorensen &
Mednick(1977)
Passivity (teacher rating) 13 50 17.7 201 18 6.4 282 (select 11.0 22.3 Conviction 10-20 to 23-33 25.2
sample)
Farrington (Note 4)
Truancy 16 23 5.6 304 68 16.5 411 20.4 9.5 One or more adjudi- 8-10 to 17 26.3 t-
O
cated offenses w
a
Robins & Hil! (1966)
Truancy 44 83 28.0 137 32 10.8 296 (nonwhite) 25.7 42.9 Police or court 6-12 to 17 26.2
record z
a
Mitchell &Rosa (1981) H
Wandering (p. 25) (parent 15 23 7.2 235 48 15.0 321 "deviant 19.6 11.9 Court appearance 5-15 to 20+ 24.7
reported) group" for indictable a
offense
Stealing (p. 24) (parent
reported)
Lying (p. 25) (parent
12
22 37
6 1.9

11.5
252

221
51

41
15.9
12.8 321
19.6
19.6
5.6
18.4
57.8
§
22.0
reported)
Farrington (Note 4)
Dishonest (peer rating) 33 55 13.4 272 51 12.4 411 20.4 21.4 One or more adjudi- 10 to 17 22.7
cated offenses
Daring (peer rating) 38 62 15.1 265 46 11.2 24.3 27.5
Daring (p. 99) (peer and 42 79 19.3 250 38 9.3 409 19.6 29.6 21 or more self- 8-10 to 14-16 32.6
parent rating) reported acts
Farrington (1979)
Involved in fights after 49 76 19.5 223 41 10.5 389 23.1 32.1 Conviction 18 to 21 49.3
drinking
Involved in antisocial groups 39 42 10.8 257 51 13.1 20.8 32.5
Unstable job 42 50 12.9 249 48 12.3 23.6 30.0
No money saved up 54 93 23.9 206 36 9.2 37.8 35.4
Drug user 45 77 19.8 222 45 11.6 31.4 27.0
Table 1 (continued)
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % rtegatives n % N PER PSR of delinquency interval RIOC

Prediction: Recidivism
Mitchell* Rosa (1981)
Problem behavior (p. 27, 17 13 21.7 20 10 16.7 60 45,0 50.0 Two or more court 5-15 to +20 26.0
parent and teacher appearances
reported)
Stealing (p. 29-30, parent and 11 3 5.3 28 15 26.3 57 45.6 24.6 60.5
teacher reported)
Lying (pp. 29-30, parent and 14 8 14.0 22 13 22.8 57 47.4 38.6 31.0
teacher reported)11
Osborn & West (1980) en
Unemployment over 5 weeks 14 9 20.9 13 7 16.3 43 48.8 53.5 Persisting vs. tempo- 1 8 to 23 28.1
in last year rary recidivism
Drug taking in past year 15 7 16.3 15 6 14.0 51.2 41.6
Self-reported aggression 15 8 18.6 14 6 14.0 53.5 38.5
en
Mulligan, Douglas, Hammond, O
& Tizard (1963)
Aggressiveness (teacher rating) 22 31 17.8 94 27 15.5 174 66.7 30.5 Reconviction 13 to 15 38.3 §
93
C/3
Robins (1966)
Antisocial referral 127 133 38.0 74 16 4.6 350 (racial 40.9 74.3 Three or more 14 to 18 56.2
mix) arrests
Buikhuisen & Hoekstra (1974)
Not moved after 188 59 13.1 82 122 27.1 451 (inmates) 68.7 54.8 One reconviction Juveniles to 5 23.7
imprisonment yrs. later
Knight & West (1975)
Offenses committed alone 16 3 3.9 27 31 40.2 77 61.0 24.7 Continued delin- 1 7 to 18-19 59.2
quent after age 17
vs. not cont.

Note. PER = percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance.
"For teacher ratings and peer ratings of troublesomeness separately as predictors of delinquency, see West and Farrington (1973) and Farrington and West (1971).
b
c
Average rating by different teachers at ages eight and ten.
Excludes 356 boys with lowest SES for whom peer ratings did not discriminate.
d
See Mitchell and Rosa (1981, p. 30) for combination of parent- and teacher-reported stealing or lying as predictor of delinquency.
78 R. LOEBER AND T. DISHION

the prediction table, the percentages of false was also predictive of delinquency, producing
positives and false negatives, and the per- a relative increment of 34.1% for official de-
centages of base rate and selection ratio. The linquency and 21.1% for high self-reported
tables also indicate the brief definitions of the delinquency. In these studies, the degree of
predictors, the criterion or outcome of delin- troublesomeness was assessed by teachers on
quency, the time interval between measure- two occasions, and in addition, peer ratings
ments of the predictor (expressed in terms of were incorporated in the assessment.
the youth's age) and the criterion (also ex- In some studies, good predictors are based
pressed in terms of the youth's age). Finally, on multiple assessments and/or assessment
as the principal evaluative index, the per- by different respondents (e.g., Craig & Click,
centage of relative improvement over chance 1963; Havighurst, Bowman, Liddle, Mat-
is provided. The first part of each table re- thews, & Pierce, 1962; West & Farrington,
ports studies with delinquency in general as 1973). Multiple assessment has the advantage
criterion, and the last part of each table re- of identifying a relatively high-risk group of
ports on prediction studies of recidivists youngsters with stable behavior, at least across
among populations of one-time offenders. the times of assessments, but most probably
The reader should keep in mind that samples also through ensuing years (see Loeber, 1982;
in some studies were specially selected and Olweus, 1979). The use of two or more re-
matched with control groups (Buikhuisen spondents, such as teachers and peers or
& Hoekstra, 1974; McCord, 1979; Robins, teachers and parents, has the additional ad-
1966; Robins & Lewis, 1966; Robins, West, vantage that the subject's behavior is ob-
&Herjanic, 1975;Tait&Hodges, 1972; Trev- served in more than one setting, which im-
vett, 1972; Voss, 1963) and consequently proves the generalizability of the findings. In
have much higher base rates than unselected addition, if these individuals are nonprofes-
populations of males. sionals or service providers who can easily be
We will first review predictors of delin- contacted, then the assessment is cost effi-
quency that can be measured early in a per- cient as compared with that done by mental
son's life. This category has been separated health workers. It is assumed that when the
into two subsets: (a) predictors that are ex- problem behavior occurs across situations, as
trapolative of the subject's own behavior, reported by different service providers, the
such as problematic or delinquent behavior stability of that behavior is higher than when
as witnessed by parents, peers, or teachers; it only occurs in only one setting (Loeber,
(b) predictors that are circumstantial—that 1982). Moreover, multiple assessments and/
is, representing the characteristics of the sub- or multiple respondents tend to reduce mea-
ject's family or social environment (Toby, surement error. In general, it is difficult to
1961). Following the section on early predic- assess which predictors are poor because of
tors is a brief review of predictors of delin- little association with a particular outcome
quency evident in late adolescence. or because of inadequate validity or reliabil-
ity of measurement. The present review, how-
Early Youth Behavior as a Predictor ever, by considering numerous studies, partly
of Later Delinquency overcomes this problem because the recur-
Table 1 lists a number of studies with ex- rence of particular good predictors across
trapolative predictors based on a subject's studies may emerge. One such predictor
behavior prior to the occurrence of delin- seems to be antisocial or problematic pre-
quency. Among the earliest predictors was delinquent behavior at the age of 12 to 14
the child's problem behavior over ages 6 to years (Mitchell & Rosa, 1981; Robins, 1966;
9 years (Craig & Click, 1963) that improved Scarpitti, 1964), which improved the predic-
predictability by about 66.3%. The predic- tion of delinquency by about 51% and the
tion was based on teachers' ratings of chil- prediction of recidivism by about 41%.
dren who presented problems in the first, sec- It is very likely that these problem cate-
ond, and third grades. Another set of studies gories include the child's aggressiveness.
(West & Farrington, 1973) showed that the Aggression at the ages of 9 to 15 years pro-
child's troublesomeness at the age of 8 to 10 duced improvements in prediction of 34% for
 EARLY PREDICTORS 79

delinquency in general and 38% for recidi- truancy before age 12 and wandering between
vism (Feldhusen, Thurston, & Benning, 1973; the ages of 5 and 15 (Mitchell & Rosa, 1981;
Mulligan, Douglas, Hammond, & Tizard, Robins, 1966; Farrington, Note 4).
1963; Havighurst et al., 1962). Unlike studies on older adolescents (Os-
Some of the specific problem behaviors in born & West, 1978, 1980), no studies on
Table 1 are correlated and thus may predict younger adolescents in our survey used com-
the same group of individuals who eventually posite predictors of more than one problem
become delinquent. Ideally, predictors should behavior such as, for example, high aggres-
not be highly correlated, so that the addition siveness and theft. Considering the high im-
of each new predictor adds to the group of provements in prediction such composite
individuals at risk rather than confirming measures have produced for the older age
those individuals who already have been group (see Table 9), we anticipate sizeable
identified. In the context of the present re- improvements in prediction through the use
view, children who are daring or disobedient of composite measures for the younger age
are often also aggressive (DeBlois & Stewart, group (see Loeber & Schmaling, Note 5, for
1980; Robins, 1966; Loeber & Schmaling, some concurrent evidence). This has an
Note 2). added advantage because of the relative in-
The second specific problem behavior par- stability of problem behaviors in childhood
ticularly predictive of recidivism is the sub- and adolescence (Rutter, 1978). A number
ject's stealing as reported by teachers and of longitudinal studies have shown that 30%
parents. Mitchell and Rosa (1981) measured to 43% of children who engage in maladap-
stealing in a sample of 5- to 15-year-olds, tive behavior at the ages of 4 to 11 years con-
which improved the prediction of delin- tinue to show such behavior 4 to 9 years later
quency by 57.8% and of recidivism by 60.5%. (Farrington, 1978; Ghodsian et al., 1980;
Unfortunately, the authors do not specify at Glavin, 1972; Janes, Hesselbrock, Myers, &
which of these ages stealing becomes not only Penniman, 1979; Werner & Smith, 1977).
prominent but also predictive. Moore, Thus, 57% to 70% of the children ultimately
Chamberlain, and Mukai (1979) followed up improved and did not show the problem be-
samples of stealing, aggressive, and normal havior years later. On the other hand, some
children. The first two groups had been re- who did not display problem behavior early
ferred for treatment between the ages of 4 in life revealed such behavior years later. Ex-
and 14 years because of problem behavior. pressed in percentages, 12% to 27% of those
At intake, the mothers reported on the nature who were initially free from the problem be-
and frequency of the child's problematic be- havior, as defined by the investigators, joined
haviors. At follow-up 2 to 9 years later, 84% the ranks of the chronic children over time
of the stealer sample had incurred a criminal (Farrington, 1978; Ghodsian et al., 1980;
record, compared with 24% of the children Janes et al., 1979). In absolute numbers, the
in the aggressive sample and 21% in the nor- percentages of newcomers given above refer
mative sample. Thus, the child's stealing as to as large or larger numbers as the chronic
reported by a parent in the family home was risk group (Loeber & Dishion, Note 1), The
highly indicative of later official delinquency. presence of chronic and newcomer groups
A number of studies have shown that has an effect on the most profitable predic-
youths engaging in stealing are also likely to tion strategy. The chronic group can be iden-
be involved in other, covert or more con- tified by means of, for example, aggressive-
cealing antisocial acts such as lying, wander- ness. However, we assume that those who
ing, or truancy (Miller, Court, Knox, & Bran- were not aggressive at the early assessment
don, 1974; Patterson, 1982; Reid & Hen- possibly showed other problem behaviors
driks, 1973; Loeber & Schmaling, Note 2; such as vandalism or lying that might even-
Reid, Hinojosa-Rivero, & Lorber, Note 3). tually lead to delinquency. For that reason,
In the present review, Table 1 shows that dis- we expect that composite measures of more
honesty by the age of 10 is moderately pre- than one precursor of delinquency will iden-
dictive of later delinquency (Farrington, Note tify an optimal proportion of children at risk
4; see also Mitchell & Rosa, 1981), as is for delinquency.
80 R. LOEBER AND T. DISHION

8 ^ p p rt; "I «n Investigators have a choice of including

"n o r^i o v~} ^o
2 various numbers of problem behaviors in the
prediction exercises. Mitchell and Rosa
~a O "^ (1981), for example, considered the total
S3
number of child problem behaviors reported
o °° m S °
c ^
1
"^
O
(M i
1 o\ OO
fy
^
O
by the mother between the ages of 5 and 15
—t
5
00
—
<-•
r-
O\ —.
" o o
CS *••
•r, years as a predictor of delinquency. A cutting
O ^- o +j +J —'
'3 score of five or more on the total deviation
'•§
S 1
1 «7 7 •»
J£? 1 ^f L/~l »5
measure produced an improvement of 20.8%
& l/^ tlJ OO —i ~-• nJ
for delinquency in general and 16.4% for re-
•o
cidivism (not shown in Table 9). The inclu-
G
o c S 8 sion of child problem behaviors in this in-
'C o* ° B S dex—such as anxieties—that are not predic-
If 'S'5 - tive of delinquency, or correlate negatively,
§1 Si C *-'
.^ g « •>•> b S
|l 1
o o •« +i
decreases the predictive power of such an
index.
:
'S5^o Delinquency is often said to be best pre-
Q
£, a; d! K! < c^ dicted by taking into account prior delin-
„ _ _^, a, ^,
quency (Monahan, 1981). Studies reviewed
a;
on ^t" ^O ON in CTs "^i" in Table 2 reinforce this idea, and produce
n,
o
G
increments averaging 40.0% (range 30.4% to
rt
oi ^o o ^o — — m 60.0%). Two points deserve attention. First,
•s the
ft.
t—
rq
os
•"<
r~- o\
m^
o
n (N
^* u
rate of recidivism computed over a pop-
1 ulation of youths is often very small (usually
•*-»

in r~^ O O in ^ s below 15%). Normally, the prediction of such

*
g low base rate events is not feasible even when
W ~
- 00 KOOO ON t-
o.
multiple gating or screening methods are
> ^ ^ ^ f\z ^ ^ used (Meehl & Rosen, 1955; Wiggins, 1973).
1

c
1

c
1 This will be reviewed briefly in the discussion.
Predicting recidivists among a population of
"S
^ 5 ^J ~
t ^ JQ
£ one-time offenders on the basis of prior de-
ii
linquency is more realistic (see Table 2). Sec-
si
;> 8>
r^i so
in
o r~^ r*^
>—<
i
n
—'
8 ond, for the studies shown in Table 2, prior
c delinquency
Cfi
^
as a predictor of continued de-
^
linquency was measured by age 15 to 17.
oo ON *—< o in r^ S Compared with aggressiveness or other non-
U5 fcP G
o.g ~ ^ ^ ^ r-,.2 delinquent problem behavior measured at an
•gS earlier age, prior delinquency appears to be-
^ S. oc ^ .nt- £; JQ •s
M come a good predictor at a slightly later age.
e
o The statement, however, should be qualified
in the sense that the age range is a function
sl o oo r- •— —• in II of the investigator's setting of the cutting
*! 00 score at that age. Koller and Gosden (1980),
in a retrospective study on a prison popula-
"2 >, 1 tion, found an average age of 14 years for the
8 g first officially recorded offense for inmate re-
0
S 00 3 vo ^"^ s cidivists, whereas first-time prisoners were,
c ~ « g 70^^ £
0
<u on the average, 21 years old when they were
1 g* •* D ^-••.9 "o " te u
>. S^-S 5 ' S ^ > c c™ ? 1
I
11
arrested for their first officially recorded of-
fenses.
3 3 »! The minimal age of measurement of pre-
•g'g S1 S c u v — .g1 « c S dictors may seem arbitrary in the studies dis-
•—'" '~'«!^8°'t^ cW«
</5

cussed above. For example, why not use

Sj*=^
= '-^
nO^« - ) p ^ c / 3o^ Q fQ2 O
1 criminal record at age 13, or perhaps even
 EARLY PREDICTORS 81

earlier? Perhaps with the exception of the in the earlier age groups. Bell and Pearl
study by Robins and Hill (1966), none of the (1982) speak of age-specific manifestations
studies systematically reviewed the predict- of risk variables. For example, property of-
ability of measures taken over different age fenses are often correlated with financial
groups. On the one hand, problem behaviors hardship caused by chronic unemployment,
including criminal activities need time to inability to hold jobs for extended periods of
become recognizable as a stable phenomenon time, or expensive drug habits. The occur-
(Epstein, 1980). On the other hand, the sta- rence of these variables before the age of 12,
bility of these phenomena at earlier ages than as an obvious example, is rare, and because
measured in the studies above may still prove of that they do not sufficiently qualify as pre-
sufficiently predictive of later delinquency. dictors in normal populations of young ad-
Clearly, only empirical research can dem- olescents. Farrington (1979) and Osborn and
onstrate how much earlier predictors can be West (1980) have shown that by the age of
measured without losing predictive power. 18 unstable employment and extended
Such studies may make it possible to identify unemployment are highly predictive of de-
youngsters who are at risk for delinquency linquency and recidivism (see Table 1). The
at an even earlier age than has been dem- same studies found that drug use at that age
onstrated up to now, especially if early pre- was also predictive of delinquency and recid-
cursors of delinquency are used. ivism, which is in line with studies on prison
populations, especially for opiate users (Gen-
Later Youth Behavior as a Predictor dreau et al., 1979; Koller & Gosden, 1980;
of Delinquency Pritchard, 1979).
Almost without exception, the early be- Empirical research is necessary to make
haviors of the youths that are predictive of better use of peer involvement in the predic-
delinquency remain predictive at a later age. tion of delinquency. Only the Farrington
This is not to say that these acts occur in (1979) study considered a measure of in-
exactly the same form or situation when volvement in antisocial groups at age 18. It
measured, for example, in early or late ad- will come as no surprise that involvement
olescence. What should be stressed here is the with antisocial or delinquent peers often oc-
relative continuity of predictive categories, curs at a much earlier age and even then can
of behavior over time. For example, at age still correlate highly with delinquency (Loe-
18, involvement in fights after drinking im- ber, Dishion, & Patterson, Note 7). Thus, in-
proved predictability by 49.3%, while at that volvement with antisocial or delinquent peers
age high self-reported aggression improved is expected to become a more important
the forecasting of recidivism by 38.5% (Far- early predictor for delinquency. It should be
rington, 1979; Osborn & West, 1980). In noted, though, that committing crimes alone
prison populations, high aggressiveness or discriminates between persisting offenders
serious institutional misconduct is also known and offenders who cease delinquent activity
to be predictive of recidivism (Cowden & in late adolescence (Knight & West, 1975).
Pacht, 1967; Cymbalisty, Schuck, & Dubeck, Osborn and West (1978, 1980) have drawn
1975; Department of Corrections of the State up two profiles of behaviors that are, together,
of Michigan, 1978, cited in Monahan, 1981; very predictive of recidivism. One profile,
Koller & Gosden, 1980). Thus aggressiveness shown in Table 9, consists of six behaviors;
from adolescence onward is a more or less the total score improves predictions by 68.3%
continuous predictor of delinquency. In the (Osborn & West, 1978). The other profile lists
same vein, unofficial theft predicts delin- 11 behaviors. A minimum cutoff score pro-
quency both at an early and at later ages. duces a relative increment of 30.6%. This
When property offenses are compared with study demonstrates a well-tried method in
offenses against people, the former seem to prediction in which not all, but a minimum
be more predictive of recidivism than the lat- number, of problem behaviors or adversities
ter (Gendreau, Madden, & Leipciger, 1979; are associated with an improvement in pre-
Koller & Gosden, 1980). diction (Rutter, 1978; Simon, 1971). What
Not all predictors of delinquency mea- is important is that the predictability of in-
sured in late adolescence necessarily occur dividual behaviors can be augmented by ag-
82 R. LOEBER AND T. DISHION

gregating these behaviors into a composite dictor for recidivism in large and less selected
profile. populations (Wolfgang et al., 1972).
Parents of a low socioeconomic status of-
Educational Achievement as a Predictor ten have larger families. Farrington and West
of Delinquency (1971) found that when a child had more
The school studies in the present review than three siblings before the age of 10, the
(Table 3) reinforce the image of delinquency- chance of delinquency for that child in-
prone children who are underachievers in an creased by 57.3% (see Table 5). Families are
educational sense. At the end of elementary not necessarily intact all the time: family
school, low achievement, low vocabulary, and breaks or prolonged separations from par-
poor verbal reasoning improved the predic- ents, as shown in Table 6, led to little im-
tion of delinquency by 27% (Farrington, provement in prediction (M = 16.7%; range =
1979; Rutter, Maugham, Mortimer, & Ous- 7.4% to 30.2%). What seems to matter more
ton, 1979; Wolfgang, Figlio, & Sellin, 1972). is overall family functioning. Craig and Glick
The best predictors during the high school (1963) used teacher ratings when the child
years were low grade point average and school was about 6 years old, which improved pre-
retardation by age 15, which improved pre- dictions by an estimable 80.2%. Unfortu-
dictability by 33.5% and 34.1%, respectively nately, it is not clear from this study which
(Polk, 1975; Robins & Hill, 1966). Again, the family characteristics were taken into ac-
same point can be made about educational count by the teachers. One of these charac-
achievement as has been made about prior teristics may be known criminal activity by
criminal activity. The minimum age at which family members. Table 7 shows that when
poor educational performance becomes rec- one or more parent or a sibling has had police
ognizable, stable, and predictive still needs contacts, the prediction of the youth's delin-
to be better established than has been possible quency or recidivism is improved by 50%
up until now. (range = 28.4% to 100%; Knight & West,
1975; Osborn & West, 1979; Robins, West,
& Herjanic, 1975). For these studies, the rel-
Early Circumstantial Predictors
ative's delinquency occurred before the child
of Delinquency was 8 to 19 years old. This does not mean
The circumstances in which children grow that the subject engaged in delinquent activ-
up differ vastly from one family to another. ity with the family member, for that is un-
Some children live in reasonably affluent common (Farrington, Gundry, & West,
families and have loving parents who are usu- 1975). According to those authors, even when
ally aware of what their children are doing the father's official delinquency occurred
and who are not reluctant to discipline when prior to the child's birth, this still can aug-
necessary, others do not. The following sec- ment the chance that the child will eventually
tion reviews a variety of early circumstantial become delinquent. As is shown in Table 7,
indicators of delinquency. the improvement in prediction is often small,
Table 4 shows the socioeconomic class of even when biological or adopted fathers are
the parent as a predictor of the child's later taken into account as a predictor of delin-
delinquency. The studies recording socioeco- quency. Only in the case of two antisocial
nomic class when the child was 4 to 12 years parents or grandparents when the child was
old improved predictions by a lesser fraction: 13 does the increment in prediction rise to
M = 19.8.% (range = 10.5% to 30.9%; Far- 49.8% (Robins & Lewis, 1966).
rington, 1979; Robins & Hill, 1966; Rutter It can be assumed that some parents, in-
et al, 1979; Wadsworth, 1979; Wolfgang et cluding those diagnosed as antisocial, are less
al, 1972). Only one study, by Knight and skilled in rearing children than others. Thus,
West (1975), showed a substantial improve- in some households, parents maintain few
ment in prediction of 49.3% by considering rules, do not exercise discipline when needed,
social class as a predictor to distinguish those or do not supervise youngsters (Patterson,
who continued their delinquent career after 1982). Table 8 shows only a few studies taking
age 18 versus those who did not. Otherwise, such parenting skills into account as single
socioeconomic class still proved a poor pre- predictors. The results are disappointing,
Table 3
Outcome Statistics for Educational Predictors of Delinquency
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC
Robins & Hill
(1966, p. 329)
School 22 72 24.3 184 18 6.1 296 nonwhites 13.5 31.8 Police or court Under 15 to 17 34.1
retardation record
before age 15
Farrington (1979,
p. 99)
Low vocabulary 40 87 21.3 242 40 9.8 409 14.7 31.0 2 1 or more self- 8-10 to 14-16 27.4
at 10 reported acts tfl
Poor school- 30 66 16.1 263 50 12.2 409 19.6 23.5 11 to 14-16 . 18.3 ' 7*
leaving results %
Polk (1975) &
Grade point 30 26 10.3 105 46 15.9 252 30.2 10.3 One or more 15 to 28-30 33.5 pi
CJ
average below contacts with o
C police-adult
criminality §
en
Rutter, Maugham,
Mortimer, &
Ouston (1979)
Verbal reasoning 266 553 48.9 266 46 4.1 1131 27.6 72.4 Cautioned or found 12 to 18 46.1
guilty
Wolfgang, Figlio, &
Seffin (1972)
Low achievement 486 368 30.5 196 157 13.0 1207 nonwhites 53.3 70.7 Police arrest 6-12 to 18 16.4
level in school 242 320 12.1 1632 448 17.0 2642 whites 26.0 21.3 22.9
Wadsworth (1979)
Attitude toward 68 262 14.6 1290 178 9.9 1798 13.7 18.3 Court appearance or 10 to 20 11.2
school work caution by police
(teacher rating)
Note. PBR = percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance. oo
U)
Table 4
Outcome Statistics for Socioeconomic Status of Parents in Predicting Delinquency
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC

Wadsworth (1979, p. 30)

Lower manual Socioeconomic 197 704 32.1 1,156 134 6.1 2,191 15.1 41.1 All reported offenses 30.9
status
Wolfgang, Figlio & Sellin
(1972, p. 54)
Low Socioeconomic status
Sample 1 763 1,377 19.6 3,649 1,254 17.8 7,043 white 28.6 30.4 Police record 4-5 to 18 10.5
Sample 2 1,293 1,151 39.7 293 165 5.7 2,902 nonwhite 50.2 84.2 28.5
Robins & Hill (1966)
Guardian's low occupational 46 102 34.5 118 30 10.1 296 nonwhite 25.7 50.0 Police or court record 6- 12 to 17 21.0
03
status (selected tu
sample)
Farrington (1979)
Low social class 24 55 13.4 274 56 13.7 409 19.6 19.3 21 or more self- 8-10 to 14-16 13.2 o
reported acts H
Rutter, Maughan Mortimer, &
Ouston (1979) en
310 24.9 568 192 15.4 ffi
Parental occupational group 175 1,245 29.5 39.0 Cautioned or found 12 to 18 14.4
guilty

Prediction: Recidivism
Knight & West (1975)
Low social class 23 6 7.4 27 25 30.9 81 59.3 35.8 Continuing delinquency 17 to 18-19 49.3
vs. temporary
delinquency
Wolfgang, Figlio, & Sellin
(1972, p. 67)
Low Socioeconomic status
Sample 1 859 430 29.5 73 94 6.5 1,456 nonwhite 65.4 88.5 Recidivists vs. one-time 4-5 to 18 14.0
offenders
Sample 2 395 372 18.4 738 514 25.5 2,019 whites 45.0 38.0 11.7

Note. PER = percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance.
 EARLY PREDICTORS 85

partly due to the modest reliability of the

8 f"i
r-'
O
O *
measures used by psychiatric social workers
* m u-i m
assessing the parents (West, 1969, p. 125).
More recent research on parents' child-rear-
g 2 ing skills has demonstrated that considerable
O ui ° -1 S improvements can be made in the measure-
"2 c 2 o o
(fi- O <N 00
ment of such skills and, possibly, in their ul-
timate predictive utility (Stouthamer-Loeber,
C
Patterson, & Loeber, Note 8). The impor-
o
tance of the parents' child-rearing practices
| |
0
§•
«1 is most evident from composite measures in
c "8 „ o° Table 9. In fact, some studies on parenting
j| •I o ^ eI skills produced increments ranging from 77%
'w o 1 <-*° (Trevvett, 1972) to 82% (Craig & Glick,
0 O ^ 1968). At the time of the assessment, children
were between 5 and 17 years old in the Trev-
X i- f> 00 vett study and 6 years old in the Craig and
& cK 00 d
Glick study, but this is not fully clear from
^_ the published reports (see also Glick, 1972).
XCQ M:
(N
o
•-' Q>
The composite measure in these studies was
OL, 8
c an abridged version of the Glueck scale, con-
"rt ^ "5
sisting of ratings of the discipline and super-
b — o\ I—
b vision of the boy by his mother, and a rating
<U "^ ~~**
^ 00 (N
B o of the family's cohesiveness. The resulting
C
C
S scores were weighted for the frequency of
•6 &
>i ^
1 ^ delinquency associated with each rating, and
S W
TT r-i
tN

O.
1 "
then totalled. Voss (1963) also used a three-
c c
I
•^
0
oo ^J- C
0
G
^
family
.i
item version of the Glueck scale, but replaced
cohesiveness with discipline of the boy
1
hi
o by the father, improving prediction by 48.3%,
•ol
In 3
G o — I a f but thereby limiting the assessment to two-
> Jf
"S fM 00 parent families,
c 8 The original Glueck and Glueck (1950,
"3, 1959) "prediction" study has not been in-
« \O r«-> —< o cluded in Table 9 because it was retrospective
m
00
S •* S in nature and relied heavily on the respon-
o oc
D-
pi
dent's recall of the youth's behavior and fam-
S CO
"
f^
o,
6 ily functioning of 7 years earlier. In the pres-
u, 1
ent review, replications of the original five-
(A
go item Glueck scale (Tait & Hodges, 1972) did
m fs not perform as well as the three-item version
'•3 '^ M •*• 12
,Ca 'rt
II usable for single-parent families. Two re-
ports, not included in Table 9, have appeared
that show very poor predictive performances
a of the original Glueck scale (Dootjes, 1972;
^ S
Wahlen, 1954, cited in Lundman & Scarpitti,
Less than two years of I
Farrington & West (1971)
Three or more siblings

& 1978).
Study predictor

§ McCord (1979) also used a composite

Wadsworth (1979)

score
I crement of family functioning leading to an in-
only child

|1
^ o n of 45.7%, but the computation of
the total score is not clear from the published
| 0 IX report. In summary, among all circumstan-
tu tial measures, composites of family function-
-1
ing produced not only the highest improve-
 00
OS
Table 6
Outcome Statistics for Separation from Parents and Family Conflict as Predictors of Delinquency
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC

Wadsworth (1980)
Death or separation before 60 220 10.0 1,640 271 12.4 2,191 15.1 12.8 All reported offenses 8 to 21 7.4
3 yr., 4 mo.

Robins & Hill (1966)

Father absent before age 1 5 26 129 43.6 128 13 4.4 296 nonwhite 9.8 52.4 Police or court record 6-15 to 17 30.2 pa

Harrington & West (1971) O

Separation from parent 19 71 17.3 289 32 7.8 411 12.4 21.9 Conviction 0-10 to 14 19.6 m
tn
through extended fO
hospitalization >
O
Farrington (1979) H
Separated up to age 10 27 63 15.4 266 53 13.0 409 19.6 22.0 21 or more self- 0-10 to 14-16 15.0 o
reported acts K

Gregory (1965) 1
Separation from parent by — — — — 5,600 24.0 13.2 Police and court 16 to 19 11.5
death or divorce records; at least one
minor offense
Farrington & West (1971)
Parental disharmony 18 71 19.0 261 23 6.2 373 11.0 23.9 Conviction 8-9 to 14 26.4
Craig & Click (1963)
Family good/fair/poor 37 21 7.0 236 7 2.3 301 14.6 19.3 "Serious and/or 6 to 16 80.2
(teacher rating) persistant"

Note. PER = percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance.
 EARLY PREDICTORS 87

ments in prediction, but as in the Craig and children, the present repertoire of known
Glick (1968) study, could probably be mea- predictors can probably be expanded. This
sured at a very early age. Moreover, com- would also be an important opportunity to
posite predictors composed of family crimi- improve the measurement quality of existing
nality, low income or socioeconomic status, and forthcoming predictors.
family size, separation from parents, and so
forth produced relative increments in pre- Summary of the Results of the
diction of 32% to 63% (May, 1981; West & Prediction Studies
Farrington, 1973; Wadsworth, 1979). The preceding results have been summa-
The importance of family functioning rized in Table 10 and have been ranked in
variables in the prediction of delinquency is terms of median percent relative improve-
also evident from studies showing the inci- ment over chance. Composite measures of
dence of delinquency within families. For parental family management techniques
example, in the study samples of Farrington tended to be most predictive of delinquency,
et al. (1975) and Wilson (1975), 47% to 62% followed by the child's problem behavior.
of all offenses committed by youngsters in Reports of the child's stealing, lying, or
the sample were committed by children from truancy come next, followed by criminality
11% to 16% of all families. In fact, certain or antisocial behavior of family members,
families were more at risk for delinquency and the child's poor educational achieve-
than were other families. It is likely that par- ment. The lowest ranking predictors are so-
ents' child-rearing practices set such families cioeconomic status and separation from par-
apart from less delinquent families (Wilson, ents. This ranking should be accepted with
1975). Although the early identification of caution, as the ranges of relative improve-
those families seems a promising avenue for ment over chance are considerable for some
delinquency prevention, such studies are still categories of predictors. Table 10 also shows
in their infancy. the ranking for the predictors of recidivism.
Even when the object is the prediction of The best predictors were reports of the child's
delinquency by youngsters rather than by stealing, lying, or truancy, followed by the
members of a family, the Glueck method re- child's own problem behavior or prior delin-
lies heavily on trained professionals to assess quency. In comparison, socioeconomic status
family functioning. Thus, such assessment, was the worst predictor. However, again the
in sharp contrast with the teacher, parent, or ranges in the percent of relative improvement
peer assessment of the youngster's problem over chance were large. The rankings of vari-
behavior, is relatively costly. Only in one ables predicting delinquency in general and
study did teachers assess family functioning predicting recidivism are largely in the same
with good predictive success (Craig & Glick, direction. The two rankings are only partly
1963). There is an obvious need to replicate comparable, however, due to the absence of
this study, as teachers are probably not usu- studies reporting on the predictive efficiency
ally as familiar with families as was the case of the following variables on the youth's re-
in this study. Bell and Pearl (1982) have cidivism: parental family management tech-
pointed out the necessity for using providers niques, the child's poor academic achieve-
of services to children more systematically ment, and the child's separation from par-
as assessors of children at risk for malad- ents.
justed behavior. They advocate a more sys- In summary, there is no doubt that certain
tematic assessment of high-risk children as extrapolative and circumstantial variables
part of a community-based rather than a greatly improve predictability. However, none
clinic-based identification program. Bell and of the studies reviewed here attempted to
Pearl (1982) rightly point out that such iden- combine extrapolative and circumstantial
tification of high-risk children requires ade- variables to further improve predictive effi-
quate intervention methods so that these chil- ciency. It is known from prediction studies
dren can be detoured from the anticipated in the field of mental health that such com-
undesirable outcome. In the process of a binations can be very fruitful. For example,
more widespread identification of high-risk Rutter (1978) used a family adversity index
 oo
Table 7 oo
Outcome Statistics for Delinquency in Other Family Members as a Predictor of Delinquency
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC
Robins, West, & Herjanic (1975)
One or both parents arrested 27 3 5 40.7 24 0 0 86 31.4 72.1 Court or police contact 10-11 to 18+ 100.0
(nonwhite before age 17
select (nontraffic)
sample)
One or both paternal 18 28 32.6 31 9 10.5 86 31.4 53.5 28.4
grandparents antisocial (nonwhite
select r3
sample)
5
m
Osborn & West (1979) w
One or more convictions for 52 50 13.0 214 67 17.5 383 31.1 26.6 One or more 8 to 24-25 29.1 en
7°
father convictions
One or more convictions for 46 45 29.4 49 13 8.5 153 38.6 59.5 45.7 Z
father or sibling o
Farrington (1979) H
Criminal parent 35 69 16.9 260 45 11.0 409 19.6 25.4 21 or more self- Under 10 to 24.4 a
C/l
reported acts 14-16
Delinquent sibling 17 29 7.1 300 63 15.4 409 19.6 11.2 21.6 5
Hutchings & Mednick (1975) ^
Criminal biological fathers 21 79 7.0 914 106 9.5 1,120 11.3 8.9 Criminal record Unknown to 10.8
(felony or misdemeanor) (nonadopted excluding minor 30-44
adults) offense
Criminal adoptive fathers 39 105 9.4 834 141 12.6 1,119 (adults 16.1 12.9 13.0
(felony or misdemeanor) adopted as
child)
Criminal biological fathers 80 273 28.1 534 84 8.7 971 (adults 16.9 36.4 19.5
of adoptees (felony or adopted as
misdemeanor) child)
Robins & Lewis (1966)
Two antisocial parents or 17 25.4 37 4 6.0 67 (includes 19.4 38.8 Arrest 13 to 18+
grandparents control
group)
Table 7 (continued)

False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PER PSR of delinquency interval RIOC

Kirkegaard-S0rensen &
Mednick (1977) £
Criminal father & 16 63 31.3 102 20 10.0 201 (sons & 17.9 39.3 Conviction Unknown to 8.3 JO
schizophrenic mother daughters of 23-33 %
schizophrenic 13
mothers) V
tn
0
Prediction: Recidivism
Knight & West (1975) po
Parent or sibling convicted 31 9 11.1 24 17 21.0 81 59.3 49.4 Delinquency continued 9 to 18-19 44.8 en
before S's age 10 after age 17 vs. not
cont.
Osborn& West (1979)
Father with two or more 56 34 18.4 61 34 18.4 185 48.6 48.6 Two or more 8 to 24-25 26.4
convictions convictions of son

Note. PBR = percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance.

oo
>o
90 R. LOEBER AND T. DISHION

83 (N
rn r-i
containing both a measure of the subject's
early behavior and measures derived from the
" (N
subject's family environment. The resulting
analysis showed that the risk of psychiatric
r-
|
| VO
disorders accelerated when four or more ad-
4 o versities were present. Similar combinations
« c o O
& S 1
of circumstantial and extrapolative predic-
oo tors have been used by Werner and Smith
a (1982) to predict which of those high-risk
IB- , youths would be resilient and not show later
•§1
(1 3 ^i
E-o
maladjusted behavior. In the study of delin-
quency, we need to allow such examples, if
Definition i
of cielinq

§
if 0 only to make predictors as powerful as pos-
i-. G. '> sible. Another important reason is that a bet-
(N U ter understanding of causal variables in de-
linquency is possible not only by identifying
a! <•<! q IN those variables predicting delinquency, but
^j- od
£ 25 M (N also by discovering the variables forecasting
absence of criminal behavior. These in a
a
SO VO Os
oc sense negative predictions are most useful to
^ si os r~^
ta better understand why certain youths do not
o engage in delinquent activities, even when
4J
fe; 1 tt (N O
their peers or relatives do so or even when
they reside in a neighborhood characterized
U5 Q
in C3 O 5 by antisocial lifestyles, unemployment, and
(N m d
<y
i)
;>
6^
1
u, ample opportunities for crime.
^ S1
^ rt
0.
B Figure 4 illustrates which early predictors
|
G
K £ iri r*i of delinquency produced the best results. It
1 shows the age of the child when various pre-
1 $ H dictors were last measured, graphed against
•S .^ -^
OS
r-
00 O
the measure of the percent relative improve-
"55 */"! OO

C)
o1
£| (N tN —<
82 ment over chance. If we can be assured of the
correctness of Craig and Click's (1963) data,
is (N M m o" family characteristics when the child was 6
Q
al *
(N t-~'
—i
o
(N 1 years old were both the earliest and among
1 31 the best predictors. From 9 years onward,
a
1 specific and generally problematic behaviors
c 0
P 9i
8 1 appear as predictors, with antisocial referrals,
S S aggressiveness, and predelinquency being the
•0 S : best predictive triad. At age 10, parent crim-
'C ll CTs
(N (N (N 1
i inality appears to be very predictive of later
child delinquency, at least as measured ret-
t 1 rospectively in Robins et al.'s (1975) study
1 •?
1
u on a select sample. At age 15, grade point
1 0 P
ff
« average becomes predictive, and a year later
West & Fanington (1973)
Poor parental child rea
Poor parental supervisi
(see West, 1969, pp.

high self-reported delinquency and officially

Outcome Statistics for

(see West, 1969, pp.

Study predictor

Authoritarian mother

I recorded delinquency come to the fore. These

"minimum" ages for predictors are, as men-
Farrington (1979)

tioned, not the result of systematic research.

178-180)

II
We anticipate that empirical studies will fur-
ther decrease the age at which children can
Table 8

ta be reliably identified as at risk for delin-

1 quency.
 EARLY PREDICTORS 91

The Evaluation of Delinquency Prediction icy 3 assigns a cost value to false positive er-
Through Utility Estimation rors (EU2).
The computational formula for the ex-
The concern with simple accuracy in crim- pected utility (EU) of a prediction strategy
inological prediction is the policy underlying recommended by Wiggins (1973) is
our use of the measures of relative improve- EU = Ul X p(VP) + C/2 X p(FP)
ment over chance. These indices do not con-
sider the errors concomitant with the valid + t/3 X + f/4 X j?(VN),
predictions within a study. It is often the case,
however, that investigators are not neutral to where j9(VP) is the probability of a valid pos-
the two types of prediction errors (i.e., false itive, p(FP) is the probability of a false pos-
positives and false negatives). One investi- itive, p(FN) is the probability of a false neg-
gator may be primarily concerned with min- ative, -J9(VN) is the probability of a valid neg-
imizing the false positive errors in a predic- ative, and t/j is the utility value of the
tion strategy to reduce the economical cost outcome.
of treating individuals not actually in need Wiggins (1973) further suggests that the
of treatment. Conversely, another investiga- cost of testing be estimated on the utility scale
tor may be concerned with reducing false and incorporated into this formula. We have
negative errors for the purpose of adequately simplified the formula by eliminating the es-
covering the population of potential criminal timated cost of testing.
offenders in a prevention program. A well- We will now reevaluate four of the predic-
established component of personnel selection tion studies (Craig & Glick, 1963; Polk, 1975;
and decision theory involves the numerical Robins, 1966; Scaipitti, 1964) discussed
estimation of the overall utility of selection above, using the three utility estimation pol-
or prediction procedures by attaching differ- icies (including the total percent correct and
ential weights to false positive and false neg- the relative improvement over chance as
ative errors (Cronbach & Gleser, 1965; Rorer, one). Table 1 1 provides the base rate, selec-
Hoffman, Hsieh, 1965; Wiggins, 1973). The tion ratio, total percent correct, relative im-
advantages of utility estimation in crimino- provement over chance, and the expected
logical prediction are (a) utility values as- utilities of the two alternative policies (EU[
signed to each prediction outcome provide and EU2).
a representation of the institutional or per- Note that, depending on the evaluation in-
sonal values underlying a prediction proce- dex one adopts, the rank-ordered preference
dure, (b) the overall utility of various predic- of the four studies radically varies. The Craig
tion strategies may be evaluated diversely ac- and Glick (1963) study is superior to the
cording to the idiosyncratic concerns of other three studies according to EUi because
various institutions or social service settings, a cost is assigned to false negative errors.
and (c) optimum cutting scores may be ra- However, on EU 2) the Craig and Glick study
tionally selected that maximize the overall is ranked third in terms of overall utility.
utility of a prediction procedure (Rorer et al., Although the four studies are relatively equal
1965). according to the total percent correct, their
The assignment of utility cost and gain rates of false positive and false negative errors
values in criminological prediction is clearly determine their practical utility in the two
a subjective process. One prediction strategy hypothetical applications. The reader is in-
may be said to be superior to another only vited to use these utility policies on further
relative to a specific utility estimation policy. studies in our review, or apply other utility
As examples three utility estimation policies functions adjusted to needs to those studies.
are presented in Figure 5. The first policy
consists of using our measure of relative im- Methodological Considerations and
provement over chance, which assumes the Recommendations in Prediction Studies
same utility values for false positive and false
negative errors. Policy 2 assigns a cost In this section we will consider prediction
value to false negative errors (EUi), and Pol- methodology in delinquency research and
Table 9
Outcome Statistics on Composite Measures as Delinquency Predictors
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PBR PSR of delinquency interval RIOC

Tait & Hodges (1972)

Glueck's Prediction Table 77 26 22.4 9 4 3.5 116 (selected 69.8 88.8 Known to juvenile 5 to 14 56.4
(5 factors) sample) court
Trevvett (1972)
Modified Glueck's Prediction 119 15 9.3 23 5 3.1 162 (same as 76.5 82.7 76.9
Table (3 factors) Tait &
Hodges)
Craig &Glick( 1968)
5
rfl
t»
Modified Glueck's Prediction 28 5 1.7 252 16 5.3 301 14.6 11.0 "Serious & 6 to 16 81.9 w
Table (3 factors) 73
persistent
delinquents"
Voss(1963) O
Modified Glueck's Prediction 13 24 10.8 176 10 4.5 223 (whites, 10.3 16.6 Police contact? 6 to 12 48.3
Table nonwhites, O
(1 parent: 2 factors) Puerto 53
(2 parents: 3 factors) Ricans) ffi
McCord(1979) 27 19 9.5 133 21 10.5 200 24.0 23.0 One or more 5-13 to 21 + 45.7 1
Home atmosphere: conviction for
1. Mother affection serious crime
2. Supervision
3. Parental conflict
4. Parental aggress.
5. Mother's self-confidence
6. Father's deviance
7. Father's absence
West & Farrington (1973)
Combination of 10 4 1.0 323 74 18.0 411 20.4 3.4 Conviction 8-10 to 17 63.0
1. parent criminality
2. low family income
3. poor parental behavior
Table 9 (continued)
False False
positives negatives
Valid Valid Definition criterion Prediction
Study predictor positives n % negatives n % N PBR PSR of delinquency interval RIOC

Wadsworth (1979)
1. Birth order 136 711 39.2 915 50 2.8 1,812 10.3 46.7 Court appearance Oto21 49.5
2. Family size or cautioned by
3. Family growth police
4. Parental divorce,
separation, or death
5. Lengthy hospital stay
6. Social group
tn
Prediction: Recidivism
Osborn & West (1980)
1. 9 wks. unemployed 19 9 20.5 13 3 6.8 44 50.0 63.6 Persisting vs. 18 to 23 68.3
2. Heavy smoking temporary M
3. Heavy drinking recidivism a
4. Sexual intercourse with
more than one partner
5. Drug use
6. Involved in fights
Osborn & West (1978)
Antisociality (score of 5 or 33 19 18.6 33 7 16.7 102 49.0 50.1 Reconviction 18 to 19-23 30.6
more on 11 self-reported
antisocial behaviors: see
West & Farrington, 1977,
P- 195)
May (1981)
Social Disadvantage Score 341 842 14.9 4082 389 6.9 5654 12.9 20.9 Police contact 11-16 to 13-18 32.4
(score of 4 or more out of
16 items)

Note. PBR percent base rate; PSR = percent selection ratio; RIOC = relative improvement over chance.
94 R. LOEBER AND T. DISHION

recommend ways in which prediction effi- officials for law-violating behavior. However,
ciency may be improved. criminological prediction research would
A formidable problem in criminology in- benefit from the use of multiple criteria for
volves the use of police contact or court ad- juvenile delinquency. Some suggested criteria
judication as prediction criteria. Hood and of delinquency are self-reported delinquency,
Sparks (1970, p. 12) have specified two major parent-reported delinquency, teacher-re-
problems with the use of criminal justice sta- ported school behavior problems, and peer
tistics as a criterion variable in crime cau- nominations of antisocial behavior.
sation studies: (a) individuals engaging in Another consideration to be made when
identical delinquent behavior who remain developing a prediction procedure or evalu-
undetected by law enforcement personnel are ating past work in criminological prediction
virtually ignored; and (b) factors explaining is the base rate (i.e., violation rate) of the
delinquent behavior are theoretically entan- principle criterion variable. As Meehl and
gled with factors responsible for official pro- Rosen (1955) have discussed, the overall ac-
cessing. Hood and Sparks suggest further that curacy of prediction decreases as the inci-
we must start, not with known delinquents, but with
dence of the criterion variable decreases. This
representative samples of the juvenile population drawn phenomenon is evident in the prediction
without regard to their known or probable delinquent studies reviewed here. For example, only 7
histories. Then, on the basis of interviews, question- of the 71 delinquency prediction studies with
naires, and tests we must differentiate these samples into criterion base rates less than or equal to 30%
delinquents and nondelinquents of various degrees and
kinds. (1970, p. 46) improve over the accuracy obtainable from
the base rate prediction alone. In contrast,
A majority of the prediction studies we have 17 of the 32 prediction studies with criterion
reviewed here rely on court records as the base rates greater than 30% exceeded the
only criterion of juvenile delinquency. This overall accuracy of the base rate prediction.
is quite reasonable given that juvenile delin- Meehl and Rosen (1955) argue that the utility
quency is usually defined as a legal phenom- of low baserate prediction is severely limited
enon, that is, juvenile delinquency refers to by the fact that the overall accuracy is likely
youths having contact with police and court to be less for predictions based on an expen-

Table 10
Rank Order of Predictors of Delinquency and Recidivism in Terms of Median Relative Improvement
Over Chance (Summarized from Tables 1-9)
R1OC

Predictor Mdn Range

For delinquency
1. Composite measures of parental family management techniques .50 .31-.82
2. Child problem behavior .32 .15-.78
3. Stealing, lying, or truancy .26 .22-.58
4. Criminality or antisocial behavior of family members .24 .08-1.0
5. Poor educational achievement .23 .11-.46
6. Single measures of parental family management techniques .23 .13-.23
7. Separation from parents .20 .07-.80
8. Socioeconomic status .18 .10-.31

For recidivism
1. Stealing, lying, or truancy .46 .31-.60
2. Child problem behavior .38 .26~.56
3. Criminality or antisocial behavior of family members .36 .26-.4S
4. Prior delinquency .36 .30-.60
5. Socioeconomic status .14 .12-.49

Note. RIOC = relative improvement over chance.

 EARLY PREDICTORS 95

sive test battery than for mere baserate pre- imately 50%. If an alternative criteria of de-
dictions. linquency is adopted—for example, self-re-
Obviously, one mode of attacking the low ported delinquency—cutting scores may be
baserate problem in criminological predic- established that provide an optimal base rate
tion is to restrict prediction efforts to criteria for prediction.
with base rates approaching the optimal 50% A recent development in the prediction of
level (Meehl & Rosen, 1955). Most recidi- delinquency involves the systematic appli-
vism studies based on samples of first-time cation of a multiple stage assessment battery
offenders report recidivism rates of approx- to a successively reduced risk sample. This

Age of Child when Predictor was (Las!) Measured

5 6 7 8 9 10 11 12 13 14 15 16 17 18
100 • Parent C r i m i n a l i t y ^

90

oFamily Functioning 0
80 o'Poor" Family 0 oPotential Delinquency 6

70-
<u
u
c •Problem Behavior 0
_°
u 60-
$ . . d
O •Aggressiveness
50-
C
01
oAntisocial Referral
E • Prior
40- Delinquency9
Q. •Troublesomeness
•Grade Point Average9
30- r, . i -High Self-
Ia •Daring report
a> Delinquency
20-
S?

10-

o-
•Certain age of child

oApproximate age of child

Craig & Click, 1963 Worn & West, 1978

b
Robins et al., 1975 ^olk, 1975

'Robins, 1966 "West & Farrington, 1973

d
Havighurst et al.,1962 'Farrington,1979
6
Scarpitti, 1964
Figure 4. The accuracy of various predictors of delinquency in relation to the age of the child at the time
of prediction.
96 R. LOEBER AND T. DISHION

Total
Percent Policy Policy
Corrfict 1 2

0, Valid Positives - U| +1 +1 +1
02 False Positives = U? 0 0 -1
Prediction Strategy
-03 False Negatives= U3 0 0
_04 Valid Negatives U, +1
Figure 5. The utility values assumed for the total percent correct index compared with two alternative
utility estimation policies.

procedure has been labeled multiple gating predictor and criterion, and, further, the in-
by Loeber, Dishion, and Patterson (Note 7) tercorrelations between the predictor vari-
and is adapted from the field of personnel ables. The best prediction battery would in-
selection (Cronbach & Gleser, 1965; Wiggins, clude predictors with relatively strong pre-
1973). In the multiple gating procedure each dictor-criterion correlations and low
assessment stage (i.e., predictor) is seen as a interpredictor correlations.
screening gate. The procedure being devel- Finally, we will make three general com-
oped by the above investigators involves ap- ments concerning method and geographic
plying the first gate to a full sample of both factors in the cross-validation of identifica-
risk and nonrisk subjects. On the basis of the tion studies. Regarding method factors, Far-
first gate, a given proportion of subjects is rington (1978) has compared various statis-
temporarily defined as a risk group and is tical prediction methods and found that the
further assessed at the second gate. Based on use of simple Burgess-type weights, unlike
the scores of the second gate, a smaller pro- discriminant analysis and multiple regres-
portion of subjects is temporarily denned as sion, are less susceptible to shrinkage in
a risk group and is retained for further as- cross-validation. Dawes and Corrigan (1974)
sessment on the third gate. The final gate also found that unit weightings perform sat-
provides for the final formulation of a group isfactorily in comparison to optimal weight-
predicted to be at high risk for future delin- ing strategies and are most robust when the
quency. The hypothesized result of each gate research is cross-validated. Duncan and Ohlin
in the multiple gating procedure is the suc- (1949) found that the predictive validity of
cessive increase in base rates and the decrease prediction strategies diminished as the vali-
in false positive errors, without a concomi- dation sample differed in sample parameters
tant increase in false negative errors. The in- from the construction sample. Reiss (1951)
clusion and order of predictor variables in concurs with these results and further dem-
the multiple-gating design involves two pri- onstrates that inefficient and psychometri-
mary considerations: (a) the lowest economic cally unstable predictors have limited pre-
cost of administering the predictor instru- dictive validity. These authors' results suggest
ments, and (b) the correlations between the that to ensure optimal predictive validity in

Table 11
Evaluation of Six Delinquency Prediction Studies Using Diverse Utility Estimation Policies
Base Selection Total %
Study rate ratio correct RIOC EU, EU2

Craig & Glick (1963) (Teacher ratings) 14.6 34.9 73.8 .66 .708 .505
Polk (1975) (School achievement) 30.2 10.3 71.4 .34 .532 .611
Robins (1966) (Antisocial referral) 70.9 60.6 74.3 .48 .563 .666
Scarpitti (1964) (Teacher rating of
potentially delinquent) 17.9 40.5 72.8 .78 .705 .480

Note, RIOC = relative improvement over chance; EUi = Expected Utility Policy 1 (cost value attached to false
positive errors); EU2 = Expected Utility Policy 2 (cost value assigned to false negative errors).
 EARLY PREDICTORS 97

cross-validation, one needs to include a small 8. Stouthamer-Loeber, M., Patterson, G. R., & Loeber,
number of stable and efficient predictors with R. Parental monitoring and antisocial child behavior.
Unpublished manuscript, 1981.
a high criterion-predictor correlation within
the prediction battery. It is also advantageous
to construct a prediction strategy from a sam- References
ple homogeneous with the target population. Bell, R. Q., & Pearl, D. Psychosocial change in risk
Second, there is a clear need to cross-val- groups: Implications for early identification. Journal
idate screening measures to identify youths of Prevention in Human Services, 1982, 7(4).
at risk for delinquency in different geograph- Buikhuisen, W., & Hoekstra, H. A. Factors related to
recidivism. British Journal of Criminology, 1974, 14,
ical areas to replicate promising criminologi- 63-69.
cal prediction results. As Rutter (1978) has Cowden, J. E., & Pacht, A. B. Predicting institutional
shown in the field of child psychopathology, and post release adjustment of delinquent boys. Jour-
diverse geographical areas can be character- nal of Consulting Psychology, 1961,31, 377-380.
ized by risk factors that only partly overlap. Craig, M. M., & Glick, S. J. Ten years experience with
the Glueck Social Prediction Table. Crime and Delin-
The present review indicates that many pre- quency, 1963, °, 249-261.
dictors from different locales operate in the Craig, M. M., & Glick, S. J. School behavior related to
same direction, although their predictive ef- later delinquency and nondelinquency. Criminologica,
ficiency may vary from site to site. Such gen- 1968, 5, 17-27.
erality makes it possible to develop screening Cronbach, L. J,, & Gleser, G. C. Psychological tests and
personnel decisions. Urbana: University of Illinois
instruments that can be widely used. Press, 1965.
These suggestions are provided in the con- Cymbalisty, B. Y, Schuck, S. Z., & Dubeck, J. A.
viction that improved methodology applied Achievement level, institutional adjustment, and re-
to criminological prediction will result in im- cidivism among juvenile delinquents. Journal of Com-
munity Psychology, 1975,,?, 289-294.
proved predictive efficiency. It is further rea- Dawes, R. M., & Corrigan, B. Linear models in decision
sonable to propose that through the system- making. Psychological Bulletin, 1974, 81, 95-106.
atic testing of objective predictors and a rig- DeBlois, C. S., & Stewart, M. Aggressiveness and anti-
orous adherence to sound empirical social behavior in children: Their relationships to
procedures, our understanding of causal pro- other dimensions of behavior. Research Communi-
cations in Psychology, Psychiatry, and Behavior, 1980,
cesses leading to criminal careers will be 5, 303-312.
greatly enhanced. Dootjes, I. Predicting juvenile delinquency. Australian
and New Zealand Journal of Criminology, 1972, 5,
Reference Notes 157-171.
1. Loeber, R., & Dishion, T. J. Strategies for identifying Duncan, O. D., & Ohlin, L. E. The efficiency of predic-
at-risk youths in the context of children's development. tion in criminology. American Journal of Sociology,
Unpublished manuscript, 1981. (Available from the 1949, 54, 441-452.
Oregon Social Learning Center, 207 East 5th, Suite Epstein, S. The stability of behavior. I. Implications for
202, Eugene, Oregon). psychological research. American Psychologist, 1980,
2. Loeber, R., & Schmaling, K. B. Empirical evidence 35, 780-806.
for overt and covert patterns of antisocial conduct prob- Farrington, D. P. The family background of aggressive
lems. Unpublished manuscript, 1982, youths. In L. A. Hersov, M. Berger, & D. Shaffer (Eds.),
3. Reid, J. B., Hinojosa-Rivero, G., & Lorber, R. A so- Aggression and antisocial behavior in childhood and
cial learning approach to the outpatient treatment of adolescence. Oxford: Pergamon Press, 1978.
children who steal. Unpublished manuscript, 1980. Farrington, D. P. Environmental stress, delinquent be-
(Available from the Oregon Social Learning Center, havior, and conviction. In I. G. Sarason & C. D. Spiel-
207 East 5th, Suite 202, Eugene, Oregon 97401.) berger (Eds.), Stress and anxiety (Vol. 6). Washington,
4. Harrington, D. P. Personalcommunication, 1981. D.C.: Hemisphere, 1979.
5. Loeber, R., & Schmaling, K. B. The utility of differ- Farrington, D. P., Gundry, G., & West, D. J. The familial
entiating between mixed and pure forms of antisocial transmission of criminality. Medicine, Science, and the
child behavior. Unpublished manuscript, 1982. Law, 1975, 15, 177-186.
6. Wolfgang, M. E. From boy to man—From delin- Farrington, D. P., & West, D. J. A comparison between
quency to crime. Paper presented at the National early delinquents and young aggressives. British Jour-
Symposium on the Serious Juvenile Offender, State nal of Criminology, 1971, 11, 341-358.
of Minnesota, Minneapolis, Minnesota, September Feldhusen, J. R, Thurston, J. R., & Benning, J. J. A
19-20, 1977. longitudinal study of delinquency and other aspects
7. Loeber, R., Dishion, T. J., & Patterson, G. R. Multiple of children's behavior. International Journal of Crim-
gating: A multi-stage assessment procedure for iden- inology and Penology, 1973, 1, 341-351.
tifying youths at risk for delinquency. Unpublished Gendreau, P., Madden, P., & Leipciger, M. Norms and
manuscript, 1981. recidivism for first incarcerates: Implications for pro-
98 R. LOEBER AND T. DISHION

gramming. Canadian Journal of Criminology, 1979, Australian and New Zealand Journal of Criminology,
1-26. 1980, 13, 117-123.
Gcrsten, J. C., Langner, T. S., Eisenberg, J. S., Simcha- Loeber, R. The stability of antisocial and delinquent
Fagan, D., & McCarthy, E. D. Stability and change in child behavior: A review. Child Development, 1982,
types of behavioral disturbances of children and ad- 53, 1431-1446.
olescents. Journal of Abnormal Child Psychology, Lundman, R. J., & Scarpitti, F. R. Delinquency preven-
1976,4, 111-127. tion: Recommendations for future projects. Crime and
Ghodsian, M., Fogelman, K., Lambert, L., & Tibben- Delinquency, 1978, 24, 207-220.
ham, A. Changes in behaviour ratings of a national Marsh, R. W. The validity of the Bristol Social Adjust-
sample of children. British Journal of Social and Clin- ment Guides in delinquency prediction. British Jour-
ical Psychology, 1980, 19, 247-256. nal of Educational Psychology, 1969, 39, 278-282.
Glavin, J. P. Persistence of behavior disorders in children. May, R. D. The Aberdeen delinquency study. In S. A.
Exceptional Children, 1972, 38, 367-376. Mednick, A. E. Baert, & B. P. Bachmann (Eds.), Pro-
Click, S. J. Identification of predelinquents among chil- spective longitudinal research. Oxford: Oxford Uni-
dren with school behavior problems as a basis for versity Press, 1981.
multiservice program. In S. Glueck & E. Glueck McCord, J. Some child-rearing antecedents of criminal
(Eds.), Identification of predelinquents. New York: In- behavior in adult men. Journal of Personality and So-
tercontinental Medical Books, 1972. cial Psychology, 1979,9, 1477-1486.
Glueck, S., & Glueck, E. Unraveling juvenile delinquency. McCord, J. Patterns of deviance. In S. B. Sells, R. Cran-
Cambridge, Mass.: Harvard University Press, 1950. dall, M, Roff, J. S. Strauss, & W, Pollin (Eds.), Human
Glueck, S., & Glueck, E. Predicting delinquency and functioning in longitudinal perspective. Baltimore,
crime. Cambridge, Mass.: Harvard University Press, Md.: Williams & Wilkins, 1980.
1959. Meehl, P. E., & Rosen, A. Antecedent probability and
Gottfredson, D. M. Assessment of predictive methods. the efficiency of psychometric signs, patterns, or cut-
In N. Johnston, L. Savitz, & M. E. Wolfgang (Eds.), ting scores. Psychological Bulletin, 1955,52, 194-216.
The sociology of punishment and aggression. New Miller, F. J. W., Court, S. D. M., Knox, E. G., & Brandon,
York: Wiley, 1970. S. The school years in Newcastle upon Tyne. London:
Gregory, I. Interospective data following childhood loss Oxford University Press, 1974.
of a parent: I. Delinquency and high school drop out. Mitchell, S., & Rosa, P. Boyhood behavior problems as
Archives of General Psychiatry, 1965, 13, 99-109. precursors of criminality: A fifteen-year follow-up
Hakeem, M. A critique of the psychiatric approach to study. Journal of Child Psychology and Psychiatry,
the prevention of juvenile delinquency. Social Prob- 1981, 22, 19-33.
lems, 1957-1958, 5, 194-205. Monahan, J. Predicting violent behavior: An assessment
Hathaway, S. R., & Monachesi, E: D. (Eds.), Analyzing of clinical techniques. Beverly Hills, Calif.: Sage Press,
and predicting juvenile delinquency with the MMPI. 1981.
Minneapolis: University of Minnesota Press, 1953. Moore, D. R., Chamberlain, P., & Mukai, L. A follow-
Hathaway, S. R., & Monachesi, E. D. Adolescent person- up comparison of stealing and aggression. Journal of
ality and behavior. Minneapolis: University of Min- Abnormal Child Psychology, 1979, 7, 345-355.
nesota Press, 1963. Mulligan, G., Douglas, J. W. B., Hammond, W. H., &
Havighurst, R. J., Bowman, P. H., Liddle, G. P., Mat- Tizard, J. Delinquency and symptoms of maladjust-
thews, C. V., & Pierce, J. V.. Growing up in River City. ment. Proceedings of the Royal Society of Medicine,
New York: Wiley, 1962. 1963, 56, 1083-1086.
Hirschi, T., & Selvin, H. C. Delinquency research: An Ohlin, L. E., & Duncan, O. D. The efficiency of predic-
appraisal of analytic methods. New York: Free Press, tion in criminology. American Journal of Sociology,
1967. 1949, 54, 441-452.
Hood, R., & Sparks, R. Key issues in criminology. Lon- Olweus, D. Stability of aggressive reaction patterns in
don: Weindenfeld & Nicholson, 1970. males: A review. Psychological Bulletin, 1979, 86,
Hutchings, B., & Mednick, S. A. Registered criminality 852-857.
in the adoptive and biological parents of registered Osborn, S. G., & West, D. J. The effectiveness of various
male criminal adoptees. In R. R. Fieve, D. Rosenthal, predictors of criminal careers. Journal of Adolescence,
& M. Brill (Eds.), Genetic research in psychiatry. Bal- 1978, /, 101-117.
timore, Md.: Johns Hopkins University Press, 1975. Osborn, S. G., & West, D. J. Conviction records of fa-
Janes, C. L., Hesselbrock, V. M., Myers, D. G., & Pen- thers and sons compared. British Journal of Crimi-
niman, J. H. Problem boys in young adulthood: nology, 1979, 19, 120-133.
Teacher ratings and 12-year follow-up. Journal of Osborn, S. G., & West, D. J. Do young delinquents really
Youth and Adolescence, 1979, 8, 453-472. reform? Journal of Adolescence, 1980, 3, 99-114.
Kirkegaard-Sorenson, L., & Mednick, S. A. A prospec- Patterson, G. R, A social learning approach, Vol. 3: Coer-
tive study of predictors of criminality. In S. A. Med- cive family process. Eugene, Oregon: Castalia Pub-
nick & K. O. Christiansen (Eds.), Biosocial bases of lishing Company, 1982.
criminal behavior. New York: Gardner Press, 1977. Polk, K. Schools and the delinquency experience. Crim-
Knight, B. J., & West, D. J. Temporary and continuing inal Justice and Behavior, 1975,2, 315-338.
delinquency. British Journal of Criminology, 1975,15, Powers, E., & Witmer, H. An experiment in the prevention
43-50. of delinquency: The Cambridge-Somervi/le youth study.
Roller, K. M., & Gosden, S. D. Recidivists: Their past New York: Columbia University Press, 1951.
and families compared with first time only prisoners. Prigmore, C. S. An analysis of rater reliability on the
 EARLY PREDICTORS 99

Glueck scale for the prediction of juvenile delin- survey-reported delinquent behavior. In R. G. Sim-
quency. Journal of Criminal Law, Criminology, and mons (Ed.), Research in community and mental
Police Science, 1963, 54, 30-41. health: An annual compilation of research. Greenwich,
Pritchard, D. A. Stable predictors of recidivism: A sum- Conn.: JAI Publishing, 1979.
mary. Criminology, 1979, 1, 15-21. Simon, G. H. Prediction methods in criminology. Lon-
Reckless, W. C., & Dinitz, S. The prevention of juvenile don: Her Majesty's Stationery Office, 1971.
delinquency. Columbus: Ohio State University Press, Stott, D. H., & Wilson, D. M. The prediction of early-
1972. adult criminality from school-age behaviour. Inter-
Reid, J. B., & Hendriks, A. F. C. J. A preliminary analysis national Journal of Social Psychiatry, 1968, 14, 5-8.
of the effectiveness of direct home intervention for Tail, C. D., & Hodges, E. F. Follow-up study of Glueck
treatment of predelinquent boys who steal. In L. A. Table applied to a school population of problem boys
Hamerlynck, L. C. Handy, & E. J. Mash (Eds.), Be- and girls between the ages of five and fourteen. In S.
havior therapy: Methodology, concepts, and practice. Glueck & E. Glueck (Eds.), Identification ofpredelin-
Champaign, III.: Research Press, 1973. quents. New York: Intercontinental Medical Books,
Reiss, A. J. The accuracy, efficiency, and validity of a 1972.
prediction instrument. American Journal of Sociology, Toby, J. Early identification and intensive treatment of
1951,5(5,552-561. predelinquents: A negative view. Social Work, 1961,
Robins, L. N. Deviant children grow up: A sociological 6, 3-13.
and psychiatric study of sociopathic personality. Bal- Trevvett, N. B. Identifying delinquency-prone children.
timore, Md.: Williams & Wilkins, 1966. In S. Glueck & E. Glueck (Eds.), Identification of pre-
Robins, L. N., & Hill, S. Y. Assessing the contribution delinquents. New York: Intercontinental Medical Book,
of family structure, class and peer groups to juvenile 1972.
delinquency. Journal of Criminal Law, Criminology, Voss, H. T. The predictive efficiency of the Glueck Social
and Police Science, 1966,57, 325-334. Prediction Table. Journal of Criminal Law and Crim-
Robins, L. N., & Laws, R. G. The role of the antisocial inology, 1963, 54, 421-430.
family in school completion and delinquency: A three- Wadsworth, M. E. J. Roots of delinquency, infancy, ad-
generation study. Sociological Quarterly, 1966, 7, 500- olescence, and crime. Oxford, England: Robertson,
514. 1979.
Robins, L. N., & Ratcliff, K. S. Risk factors in the con- Wadsworth, M. E. J. Early life events and later behavioral
tinuation of childhood antisocial behavior into adult- outcomes in a British longitudinal study. In S. B. Sells,
hood. International Journal of Mental Health, 1979, R. Crandall, M. Roff, J. S. Strauss, & W. Pollin (Eds.),
7,96-116. Human functioning in longitudinal perspective. Balti-
Robins, L. N., West, P. A., & Herjanic, B. L. Arrests and more, Md.: Williams & Wilkins, 1980.
delinquency in two generations: A study of black ur- Weis, K. The Glueck Social Prediction Table: An un-
ban families and their children. Journal of Child Psy- fulfilled promise. Journal of Criminal Law and Crim-
chology and Psychiatry, 1975,76, 125-140. inology, 1974, 65, 397-404.
Roff, M., Sells, S. B., & Golden, M. M. Social adjustment Werner, E. E., & Smith, R. S. Kami's children come of
and personality development in children. Minneapolis: age. Honolulu: University Press of Hawaii, 1977.
University of Minnesota Press, 1972. Werner, E. E., & Smith, R. S. Vulnerable, but inconvinc-
Rorer, L. G., Hoffman, P. J., & Hsieh, K-C. Utilities as ible: A longitudinal study of resilient children and
baserate multipliers in the determination of optimum youth. New York: McGraw-Hill, 1982.
cutting scores for the discrimination of groups of un- West, D. J. Present conduct and future delinquency. Lon-
equal size and variance. Journal of Applied Psychology, don: Heinemann, 1969.
1965, 50, 364-368. West, D. J., & Farrington, D. P. Who becomes delin-
Rutter, M. Family, area, and school influences in the quent? "London: Heinemann, 1973.
genesis of conduct disorders. In L. A, Hersov, M. Ber- Wiggins, J. S. Personality and prediction: Principles of
ger, & D. Shaffer (Eds.), Aggression and antisocial be- personality assessment. Reading, Mass.: Addison-Wes-
havior in childhood and adolescence. Oxford: Perga- ley, 1973.
mon Press, 1978. Wilson, H. Juvenile delinquency, parent criminality, and
Rutter, M., Maugham, G., Mortimer, P., & Ouston, J. social handicap. British Journal of Criminology, 1975,
15,000 hours—Secondary schools and their effects on 75, 241-250.
children. Cambridge, Mass.: Harvard University Press, Wolfgang, M. E., Figlio, R. M., & Sellin, T. Delinquency
1979. in a birth cohort. Chicago: University of Chicago Press,
Scarpitti, F. R. Can teachers predict delinquency? The 1972.
Elementary School Journal, 1964, 65, 130-136.
Simcha-Fagan, O. The prediction of delinquent behavior Received March 12, 1982
over time: Sex-specific patterns related to official and Revision received December 27, 1982 •