Stokes and Baer 1977

1977, 10, 349-367

AN IMPLICIT TECHNOLOGY OF GENERALIZATION'TREVOR F. STOKES AND DONALD M. BAER

THE UNIVERSITY OF MANITOBA AND THE UNIVERSITY OF KANSAS

Traditionally, discrimination has been understood as an active process, and a technologyof its procedures has been developed and practiced extensively. Generalization, by con-trast, has been considered the natural result of failing to practice a discriminationtechnology adequately, and thus has remained a passive concept almost devoid of atechnology. But, generalization is equally deserving of an active conceptualization andtechnology. This review summarizes the structure of the generalization literature andits implicit embryonic technology, categorizing studies designed to assess or programgeneralization according to nine general headings: Train and Hope; Sequential Modifi-cation; Introduce to Natural Maintaining Contingencies; Train Sufficient Exemplars;Train Loosely; Use Indiscriminable Contingencies; Program Common Stimuli; MediateGeneralization; and Train "To Generalize".DESCRIPTORS: generalization, treatment-gain durability, followup measures, main-

tenance, postcheck methodology

Traditionally, many theorists have consideredgeneralization to be a passive phenomenon. Gen-eralization was not seen as an operant responsethat could be programmed, but as a descriptionof a "natural" outcome of any behavior-changeprocess. That is, a teaching operation repeatedover time and trials inevitably involves varyingsamples of stimuli, rather than the same setevery time; in the same way, it inevitably evokesand reinforces varying samples of behavior,rather than the same set every time. As a conse-quence, it is predictable that newly taught re-sponses would be controlled not only by thestimuli of the teaching program, but by otherssomewhat resembling those stimuli (Skinner,1953, p. 107ff.). Similarly, responses resemblingthose established directly, yet not themselves ac-tually touched by the teaching procedures, wouldappear as a result of the teaching (Keller and

'Preparation of this paper was supported in part byPHS Training Grant 00183, Program Project GrantHD 00870, and Research Grant MH 11739. Reprintsmay be obtained either from T. F. Stokes, Departmentof Psychology, University of Manitoba, Winnipeg,Manitoba, Canada R3T 2N2, or D. M. Baer, Depart-ment of Human Development, University of Kansas,Lawrence, Kansas 66045.

Schoenfeld, 1950, p. 168ff.). Thus, generaliza-tion was something that happened, not some-thing produced by procedures specific to it.

If generalization seemed absent or insignifi-cant, it was simply to be assumed that the teach-ing process had managed to maintain unusuallytight control of the stimuli and responses in-volved, allowing little sampling of their varie-ties. This assumption was strongly supported bythe well-known techniques of discrimination: bydifferential reinforcement (in general, by anydifferential teaching) of certain stimuli relativeto others, and/or certain responses relative toothers, generalization could be programmaticallyrestricted and diminished to a very small range.Thus, it was discrimination that was understoodas an active process, and a technology of its pro-cedures was developed and practiced extensively.But generalization was considered the naturalresult of failing to practice discrimination's tech-nology adequately, and thus remained a passiveconcept almost devoid of a technology. Never-theless, in educational practice, and in the devel-opment of theories aimed at serving both practiceand a better understanding of human function-ing, generalization is equally as important as dis-

349

NUMBER 2 (SUMMER) 1977JOURNAL OF APPLIED BEHAVIOR ANALYSIS

TREVOR F. STOKES and DONALD M. BAER

crimination, and equally deserving of an activeconceptualization.

Generalization has been and doubtless willremain a fundamental concern of applied behav-ior analysis. A therapeutic behavioral change,to be effective, often (not always) must occurover time, persons, and settings, and the effectsof the change sometimes should spread to avariety of related behaviors. Even though theliterature shows many instances of generaliza-tion, it is still frequently observed that when achange in behavior has been accomplishedthrough experimental contingencies, then thatchange is manifest where and when those contin-gencies operate, and is often seen in only transi-tory forms in other places and at other times.The frequent need for generalization of thera-

peutic behavior change is widely accepted, but itis not always realized that generalization doesnot automatically occur simply because a behav-ior change is accomplished. Thus, the need ac-tively to program generalization, rather thanpassively to expect it as an outcome of certaintraining procedures, is a point requiring bothemphasis and effective techniques (Baer, Wolf,and Risley, 1968). That such exhortations haveoften been made has not always ensured thatresearchers in the field have taken serious noteof and, therefore, proceeded to analyze ade-quately the generalization issues of vital concernto their programs. The emphasis, refinement, andelaboration of the principles and procedures thatare meant to explain and produce generalizationwhen it does not occur "naturally" is an impor-tant area of unfinished business for applied be-havior analysis.

The notion of generalization developed hereis an essentially pragmatic one; it does notclosely follow the traditional conceptualizations(Keller and Schoenfeld, 1950; Skinner, 1953).In many ways, this discussion will sidestep muchof the controversy concerning terminology. Gen-eralization will be considered to be the occur-rence of relevant behavior under different, non-training conditions (i.e., across subjects, settings,people, behaviors, and/or time) without the

scheduling of the same events in those conditionsas had been scheduled in the training conditions.Thus, generalization may be claimed when noextratraining manipulations are needed for extra-training changes; or may be claimed when someextra manipulations are necessary, but their costor extent is clearly less than that of the directintervention. Generalization will not be claimedwhen similar events are necessary for similar ef-fects across conditions.A technology of generalization programming

is almost a reality, despite the fact that until re-cently, it had hardly been recognized as a prob-lem in its own right. Within common teachingpractice, there is an informal germ of a technol-ogy for generalization. Furthermore, within thepractice of applied behavior analysis (especiallywithin the past 5 yr or so), there has appeareda budding area of "generalization-promotion"techniques. The purpose of this review is to sum-marize the structure of that generalization litera-ture and its implicit embryonic technology. Some270 applied behavior analysis studies relevant togeneralization in that discipline were reviewed.2A central core of that literature, consisting ofsome 120 studies, contributes directly to a tech-nology of generalization. In general, techniquesdesigned to assess or to program generalizationcan be loosely categorized according to ninegeneral headings:

1. Train and Hope2. Sequential Modification3. Introduce to Natural Maintaining Contin-

gencies4. Train Sufficient Exemplars5. Train Loosely6. Use Indiscriminable Contingencies7. Program Common Stimuli

2Ninety per cent of the literature reviewed wasfrom five journals: Behaviour Research and Therapy;Behavior Therapy; Journal of Applied Behavior Anal-ysis; Journal of Behavior Therapy and ExperimentalPsychiatry; and Journal of Experimental Child Psy-chology. Seventy-seven per cent of the literature re-viewed has been published since 1970.

350

AN IMPLICIT TECHNOLOGY OF GENERALIZATION

8. Mediate Generalization9. Train "To Generalize".

This review characterizes each category, anddescribes some examples of research that illus-trate the generalization analyses or program-ming involved in each category. Obviously, allthe relevant references cannot be discussed in thisreview.3 The nine categories listed above wereinduced from the literature; they are not a prioricategories. Consequently, studies do not alwaysfit neatly into these categories. It should also benoted that not all studies reviewed were thoroughexperimental analyses of generalization. Ofteninferences were necessary to categorize the re-search. However, the following discussion stillmay provide a useful organization and concep-tualization of generalization and its program-ming.

1. Train and HopeIn applied behavior analysis research, the most

frequent method of examining generalization, sofar, may be labelled Train and Hope. After abehavior change is effected through manipula-tion of some response consequences, any existentgeneralization across responses, settings, experi-menters, and time, is concurrently and/or sub-sequently documented or noted, but not activelypursued. It is usually hoped that some generali-zation may occur, which will be welcomed yetnot explicitly programmed. These hopeful probesfor stimulus and response generalization char-acterize almost half of the applied literature ongeneralization. The studies have considerableimportance, for they begin to document the ex-tent and limits of generalization of particularoperant intervention techniques. While not being

3Complete reference lists and detailed tables de-scribing subjects, procedures, and generalization ofall studies reviewed are deposited with the NationalAuxiliary Publications Service (NAPS). For copies,order NAPS Document #02873. Order from ASIS/NAPS Co., C/O Microfiche Publications, 305 East46th Street, New York, New York 10017. Remitwith order for each copy $3.00 for microfiche or$19.50 for photocopies. Make checks payable toMicrofiche Publications.

examples of the programming of generalization,they are a sound first step in any serious analysisof generalization. When generalization is desired,but is shown to be absent or deficient, program-ming procedures can then be instituted.

For example, useful generalization across set-tings was documented by Kifer, Lewis, Green,and Phillips (1974). In an experimental class-room setting, parent-child pairs were taught tonegotiate in conflict situations. During simulatedrole-playing, instructions, practice, and feedbackwere used to teach the negotiation behaviors offully stating one's position, identifying the issuesof conflict, and suggesting options to resolve theconflict. The data showed increased use of nego-tiation behaviors and the reaching of agreementsin actual parent-child conflict situations at home.An assessment of generalization across experi-

menters was described by Redd and Birnbrauer(1969), who demonstrated that control over thecooperative play of retarded children did notgeneralize from an adult who dispensed contin-gent edible reinforcement to five other adultswho had not participated in training.

Studies that are examples of Train and Hopeacross time are those in which there was a changefrom the intervention procedures, either to a lessintensive but procedurally different program, orto no program or no specifically defined pro-gram. Data or anecdotal observations were re-ported concerning the maintenance of the origi-nal behavior change over the specified timeintervening between the termination of theformal program and the postchecks. An exampleof a followup evaluation was the study by Azrin,Sneed, and Foxx (1973). An intensive trainingprogram involving reinforcement of correct toi-leting and positive practice procedures promptlydecreased bedwetting by 12 retarded persons.The reduced rate of accidents was maintainedduring a three-month followup assessment.

Perhaps there are many more studies in theTrain and Hope category than would have beenexpected (about 135, of which 65% are acrossTime). However, despite its obvious value, thisresearch is frequently characterized by a lack of

351


comprehensiveness and depth of the generaliza-tion analysis. Even though generalized behaviorchange was frequently reported, extensive, wide-ranging, and practical generalization was notoften noted or even sought. The continued de-velopment of behavior analysis almost surelywill demand more extensive collection of gener-alization data than is presently the fashion. Theextent and limits of applied behavioral interven-tions may be well documented and understood ifmeasurement is extended over longer periods oftime, over more than one circumscribed part ofthe day, with more than one related response,and with more than a restricted part of the socialand physical environment. It is as important forthe field to formalize the conditions of the non-occurrence of generalization as it is to documentthe conditions associated with the display of un-programmed generalization.

Most of the Train-and-Hope research describedsuccessful generalization-approximately 90%of Train-and-Hope studies. By definition, therewas no further need to program generalizationin those studies where generalization had beenexhibited within the Train-and-Hope para-digm-presuming, of course, that the generali-zation exhibited was considered sufficient to meetthe therapeutic goals of the various modificationprograms (not necessarily a valid presumptionin the Train-and-Hope research). This prepon-derance of positive data may simply reflect thetendency of some researchers not to report theirgeneralization data if measurement procedureswere instituted to probe for any generalized be-havior changes, but generalization was shown tobe absent. Some researchers may view nongen-eralization as reflecting a deficiency or ineffective-ness of their procedures to develop a desirablegeneralized performance. Behavior analysts,nevertheless, should encourage their fellow re-searchers to document and to analyze experimen-tally their apparent failures, rather than allowingthem to slide into oblivion. A detailed and sys-tematic understanding of generalization and itsprogramming could result. Alternatively, re-searchers might view their generalization base-

lines as being essentially independent of the mod-ified baseline; thus, to report nongeneralizationwould serve no useful purpose, for its nonoccur-rence would be expected. Again, any such docu-mentation contributes to our understanding ofthe extent and limits of generalization, as well asserving as an indication of the frequent necessityof generalization-programming techniques.

There is another reason for the predominanceof positive results in this section: if nongeneral-ization was clearly evident, and the modificationof this state was important, then a form of lim-ited programming was frequently instituted. Ex-amples of this research will be discussed in thenext category, "Sequential Modification".

2. Sequential ModificationThese studies exemplify a more systematic

approach to generalization than the Train-and-Hope research. Again, a particular behaviorchange is effected, and generalization is assessed.But then, if generalization is absent or deficient,procedures are initiated to accomplish the desiredchanges by systematic sequential modification inevery nongeneralized condition, i.e., across re-sponses, subjects, settings, or experimenters. Thepossibility of unprogrammed generalization typi-cally was not examined in these sequential modi-fication studies, because after the initial demon-stration of nongeneralization, all other baselineswere exhausted. That is, after changes had beenproduced directly in all baselines, generalizationto nonrecorded responses, subjects, settings, andexperimenters may have occurred, but could notbe examined.

For example, Meichenbaum, Bowers, and Ross(1968) reported an absence of generalization ofbehavior changes from an afternoon interventionperiod to the morning period in a classroom forinstitutionalized female adolescent offenders.Money dispensed contingent on on-task behav-iors effected desired behavior changes during theafternoon, but generalization to the morningperiod required that the same manipulations beapplied there as well (sequential modificationacross settings). Similarly, generalization across

352


settings of the disruptive and oppositional behav-ior of two children was investigated by Wahler(1969). He demonstrated control of these behav-iors in the home by using differential attentionand timeout operations. When generalization tothe children's school behavior was not evidenced,similar contingency operations were employed toaccomplish changes in that setting as well.The category of Sequential Modification char-

acterizes much of the actual practice of manybehavior analysts. Sequential modification ismerely a systematized experimental procedurethat formalizes and allows evaluation of thesetypical therapeutic endeavors. The tactic ofscheduling behavior-change programs in everycondition to which generalization is desired isfrequently employed. The rationale for theseprocedures is as follows. If a desired generaliza-tion is not likely to be exhibited after changing abehavior in a particular condition, or a numberof conditions, e.g., settings, then the researcher orpractitioner works to effect changes across con-ditions as a matter of course, rather than as anoutcome of the display or nondisplay of general-ization. Thus, a behavior analyst is likely to ad-vise the scheduling of consequences in everyrelevant condition in preference to the dispens-ing of consequences in only one or a few condi-tions, while hoping for generalization, but likelynot seeing it.

3. Introduce to Natural MaintainingContingencies

Perhaps the most dependable of all general-ization programming mechanisms is one thathardly deserves the name: the transfer of behav-ioral control from the teacher-experimenter tostable, natural contingencies that can be trustedto operate in the environment to which the sub-ject will return, or already occupies. To a con-siderable extent, this goal is accomplished bychoosing behaviors to teach that normally willmeet maintaining reinforcement after the teach-ing (Ayllon and Azrin, 1968).

Baer and Wolf (1970) reported a study byIngram that illustrated the mechanism of "trap-

ping", where a preschool child was taught anentry response that exposed the child to thenatural contingencies of peers in the preschoolenvironment. Preschool teachers modified thelow rate of skillful interaction of the child bypriming others to interact with the subject andreinforcing appropriate interactions. The datashowed that over time the teachers lost controlof the interaction behavior, which remainedhigh; it was assumed that the group's naturalconsequences for interaction had taken controlof the subject's behavior. Thus, to program gen-eralization, the child perhaps needed only to beintroduced adequately to the natural reinforcersinherent in active preschool play and interaction.Some early analyses of preschool children's be-havior have stressed that if the child can be sointroduced (through the operation of differentialattention from teachers) to a reinforcing pre-school natural environment, then the behaviorseventually do not need to be maintained by con-tinued contrived modification of the environ-ment. For example, Hall and Broden (1967)modified the manipulative play, climbing, andsocial interaction of three subjects through socialreinforcement operations. Behavior changes wereshown to be durable and successful followupdata at three months were described.

Buell, Stoddard, Harris, and Baer (1968) dem-onstrated the collateral development of appro-priate social behavior (e.g., touching, verbalizing,and playing with other children) accompanyingthe reinforcement of increased use of outdoorplay equipment by a 3-yr-old girl. This entry re-sponse to the natural reinforcement communitywas tactically sound because the child's motorbehavior was modified in a setting where theresulting behavior would tend automatically toincrease social contact with other children, andthis natural social environment could maintainthe child's new skills, but indeed may also beexpected to sharpen and refine them, and addentirely new ones as well.Most of the research concerning natural main-

taining contingencies has involved children, per-haps because such techniques seem particularly

353


suitable, especially to their social behavior. Re-search would profit by determining what naturalreinforcement communities exist for various be-haviors and subjects, and what economical meansmay be employed to ensure entry to these behav-ioral traps.

Unfortunately, in some instances there may beno natural reinforcement operating to developand maintain skills. For example, in the case ofretarded and institutionalized persons whose de-pendency has become a stable fact in the lives oftheir caretakers, some re-arrangement of thenatural environment may be necessary. A fewstudies have introduced subjects to semicontrivedor redesigned "natural" reinforcement communi-ties. A simple but meaningful example was pro-vided by Horner (1971), who taught a 5-yr-oldinstitutionalized retarded boy to walk on crutchesin an experimental setting. The child was thenprompted to generalize the new walking skillto other settings and activities to which he previ-ously had been taken in a wheelchair by solici-tous caretakers, by enlisting those caretakers torefrain from offering this help. Within 15 daysafter treatment was concluded, the child walkedon crutches to all those activities and settings,eventually extending his ambulation skills to anypart of his world. Stolz and Wolf (1969) traineda 16-yr-old, "blind" retarded male to discriminatevisual stimuli. Then, the environment was sostructured that assistance was not given in situa-tions where it had previously been given as amatter of course. When the boy was required touse visual cues to help himself in a cafeteria line,he soon emitted the necessary behaviors. How-ever, these studies did not establish the function-ality of their procedures in the maintenance ofbehavior changes.

Another significant example was provided bySeymour and Stokes (1976). In their study, insti-tutionalized delinquent girls were taught to so-licit reinforcement (cf. Graubard, Rosenberg,and Miller, 1971) from their natural community,the staff of their residential institution. In theircase, the staff had rarely displayed any systematicattempts at reinforcing desirable behavior shown

by the girls, perhaps on the presumption that thegirls were "bad" and not reinforcible in any case.However, the experimenters were able to teachthe girls that when their work was objectivelygood, and when staff persons were nearby, asimple skill of calling these adults' attention totheir good work would result in fairly consistentreinforcement. Thus, this was a case in whichexperimental reinforcement was used to developa response in the subjects that would tap andcultivate the available but dormant natural com-munity. In theory, this new skill should haveobviated the need for further experimental re-inforcement, for the praise evoked should havefunctioned to maintain both the girls' work andcueing, and the cueing, in turn, should have func-tioned to maintain staff praise. The Seymour andStokes' study could not be continued long enoughto establish whether this would happen, and soit remains a logically appealing but still unex-plored method of enhancing generalization:teaching the subject a means of recruiting a nat-ural community of reinforcement to maintainthat generalization. Perhaps an even greater ad-vantage of such procedures is a change in thelocus of control: the subjects can become moreprominent agents of their own behavior change,rather than being hapless pawns of more-or-lessrandom environmental contingencies.

Restructuring the environment thus becomes atarget of research aimed at extending the gener-alization of newly taught skills; even though, ata technical level, this operation may not be con-sidered generalization, but rather transfer ofcontrol from one reinforcement contingency toanother. In any event, it is a much neglectedtopic of experimental research, although widelyrecognized as a desirable, and even essentialcharacteristic of any rehabilitative effort.

Some natural contingencies are inevitably atwork contributing to the maintenance of inap-propriate behavior. For example, peer-groupcontrol of inappropriate behavior has often beensuspected and sometimes documented (Buehler,Patterson, and Furniss, 1966; Gelfand, Gelfand,and Dobson, 1967; Solomon and Wahler, 1973).

354


It would seem reasonable, then, that if the pat-tern of reinforcement of inappropriate behavioris modified, the observed outcome may errone-ously, but happily be attributed to generalization.For example, Bolstad and Johnson (1972) pre-sented data that showed that both experimentaland control subjects in the same classroom wereall affected (although not to the same extent) byexperimental manipulation of the reinforcementcontingency for the experimental subjects,whereas control subjects in a different classroomwere not so affected. The authors presented datathat may account for these differences. The con-trol subjects in the experimental classroom, whowere also disruptive students, had fewer disrup-tive interactions with the experimental subjectsduring the treatment phases than during base-line. This possible generalization effect may bedue to the disruption of the natural contingenciesoperating in that environment. That is, otherdisruptive students previously supported some ofthe disruptive behavior of the control subjects,but during treatment these experimental subjectsdid not support the disruptive behavior of theirpeers and, thus, a "generalized" decrease in dis-ruptive behavior by the control subjects resulted.

4. Train Sufficient ExemplarsIf the result of teaching one exemplar of a

generalizable lesson is merely the mastery of theexemplar taught, with no generalization beyondit, then the obvious route to generalization is toteach another exemplar of the same generaliza-tion lesson, and then another, and then another,and so on until the induction is formed (i.e., untilgeneralization occurs sufficiently to satisfy theproblem posed). Examples of such programmingtechniques will be described in this category oftraining sufficient exemplars, perhaps one of themost valuable areas of programming. Certainlyit is the generalization-programming area mostprominent and extensive in the present literature.

In the research discussed previously underthe categories of Train and Hope and Sequen-tial Modification, the typical analysis of gener-alization concerned the measurement of gener-

alization to only a few (and often only one)extraexperimental responses, subjects, settings,experimenters, or times. When the absence ofgeneralization was noted, sometimes it was ac-complished by further direct intervention inevery nongeneralized condition (i.e., SequentialModification). Having completed such modifica-tions, the possibility of more extensive general-ized effects (i.e., beyond the two or three modifiedbaselines) was not examined. In the training ofsufficient exemplars, generalization to untrainedstimulus conditions and to untrained responsesis programmed by the training of sufficient ex-emplars (rather than all) of these stimulus con-ditions or responses.A systematic demonstration of programmed

generalization and measurement of generalizedeffects beyond intervention conditions was re-ported by Stokes, Baer, and Jackson (1974). Theyestablished that training and maintenance of re-tarded childrens' greeting responses by one ex-perimenter was not usually sufficient for thegeneralization of the response across experi-menters. However, high levels of generalizationto over 20 members of the institution staff (andnewcomers as well) who had not participatedin the training of the response were recorded,after a second experimenter trained and main-tained the response in conjunction with the firstexperimenter. Thus, when generalization did notprevail after the training of one stimulus exem-plar, it was programmed by training a greaterdiversity of stimulus (trainer) conditions. Simi-larly, Garcia (1974) taught a conversationalspeech form to two retarded children, and, upondiscovering a lack of stable generalization acrossexperimenters after one training input, pro-grammed generalization across experimenters byhaving a second experimenter teach the sameresponses.A sufficient-stimulus-exemplars demonstration

of programmed generalization across settings hasbeen described by Allen (1973). Allen modifiedthe bizarre verbalizations of an 8-yr-old boy bydifferential attention procedures. Ignoring bi-zarre verbalizations and praise for appropriate

355


interaction reduced bizarre verbalizations duringevening camp activities. However, there was nogeneralization to three other camp settings. Afteradditional training in a second setting, somegeneralization to the unmanipulated settings wasnoted. This generalization was further enhancedby intervention in the third setting. Unfortu-nately, the experimental procedures did not allowsufficient time to document the full extent ofgeneralization after training in two settings, butgeneralization after training in two settings wasclearly evident. Griffiths and Craighead (1972)similarly programmed generalization across set-tings. A 30-yr-old retarded woman receivedpraise and tokens for correct articulation inspeech therapy. Generalization to a residentialcottage was not observed until the same proce-dures were instituted there. Following trainingin these two stimulus exemplars, generalizationto a third nontraining setting (a classroom) wasobserved.

Very little research concerned with generaliza-tion programming has dealt with the training ofsufficient stimulus exemplars. The infrequentresearch that has been published is characterizedlargely by programming across experimenters.This work has been promising, for after a modestnumber of training inputs, generalization appar-ently will occur with persons not involved intraining-unquestionably a valuable and inex-pensive outcome. However, the present implica-tion of these studies is limited because of therestricted nature of the type of subjects andresponses analyzed. Further work is also neededto give direction to the optimal conditionswhereby the most extensive generalization willbe achieved with a minimal training expendi-ture. Nevertheless, it is optimistic to note howfrequently a sufficient number of exemplars isa small number of exemplars. Frequently, it isno more than two. In particular, there may wellbe reason to suspect that the use of two trainerswill yield excellent results in terms of generaliza-tion. This possibility, obviously an economicalone, certainly merits systematic study of its po-tential and limits.

Although very little research has been re-ported, the analysis of generalization program-ming by training in a number of settings is avirtually untapped area of far-reaching value.However, consistent optimism should follow ex-amination of the studies showing generalizationafter training in only a few settings. Unfortu-nately, behavior analysts seem too often satisfiedwith the modification of a single, well-definedbehavior in one setting, e.g., a laboratory pre-school. Discriminated programs are often accept-able, and sometimes even desirable. When gener-alization is a valid concern, but researchers andpractitioners do not act as if this were so, thediscriminated behavior of researchers is mostprobably inhibitory to the development of aneffective generalization technology.

Over the past 10 yr, there has developed anextensive literature discussing the programmedgeneralization of responses through the trainingof sufficient response exemplars. A response classhas been operationally defined to describe thefact that some responses are organized such thatoperations applied to a subset of responses in theclass affect the other members of that class in thesame manner. For example, Baer, Peterson, andSherman (1967) reinforced various motor imita-tions by retarded children. They found that aslong as reinforcement followed some imitativeresponses, other imitations continued to be per-formed without training or reinforcement.A topographical analysis of generalized imita-

tion has been made by Garcia, Baer, and Fire-stone ( 1971). Four retarded children were trainedto imitate three different topographical types ofresponse: small motor, large motor, and shortvocal. These subjects were also probed for theirimitation of other unreinforced responses: shortmotor, long motor, short vocal, and long vocal.Generalized imitation was observed with eachsubject, but this generalization reflected the par-ticular dimensions of the topographical responsecurrently being trained or having previously re-ceived training. Thus, generalization may occurwithin well-defined classes and may not gener-alize to other classes unless some special training

356


(generalization programming) occurs within thatclass as well. These data depict one possible limi-tation of the generality of generalized imitation,as well as pointing to the need to train responseexemplars that will adequately reflect the di-versity of the generalization being programmed.

Children's grammatical development has beenanother prominent area of research dealing withgeneralized behavior. The concept of responseclass is again pivotal in these studies, whichconceptualize the rules of morphological gram-

mar as equivalent to response class phenomena.For example, Guess, Sailor, Rutherford, andBaer (1968) developed the generative correct

use of plurals by a retarded girl. After teachinga number of exemplars of the correct pluralresponse, the girl appropriately labelled new

objects in the singular or plural without furtherdirect training relevant to those objects. Pluralusage had become a generalized response class;the morphological rule had been established.Schumaker and Sherman (1970) rewarded threeretarded children for the correct production ofpast- and present-tense forms of verbs. As past-

and present-tense forms of verbs within an in-flectional class were modified, there occurred a

generalized usage of untrained verbs to similartense forms.

There has been considerable research to estab-lish the importance of the training of sufficientresponse exemplars. A survey of these (approxi-mately 60) studies shows that the number ofexemplars found to be "sufficient" for a desirablelevel and durability of generalization varieswidely, probably determined primarily by thenature of the task and the subject's prior skillsrelevant to it. Most of this research was con-

cerned with the development of motor and vocalimitations, and the beginning development ofgrammar and syntax. The development of ques-

tion-asking and instruction-following is also wellrepresented.

In conclusion, examination of the sufficientexemplar research points to a significant (andlong-familiar) generalization-programming pro-

cedure: a number of stimulus and/or response

exemplars should undergo training. That is, toprogram the generalized performance of certainresponses across various setting conditions orpersons, training should occur across a (suffi-cient) number of setting conditions and/or withvarious persons. In a similar manner, generaliza-tion across responses can be programmed reliablyby the training of a number of responses. Diver-sity of exemplars seems to be the rule to followin pursuit of the maximum generalization. Suffi-cient diversity to reflect the dimensions of thedesired generalization is a useful tactic. However,diversity may also be our greatest enemy: toomuch diversity of exemplars and not enough(sufficient) exemplars of similar responses maymake potential gains disproportional to the in-vestment of training effort. The optimal combi-nation of sufficient exemplars and sufficient di-versity to yield the most valuable generalizationis critically in need of analysis. Is the best pro-cedure to train many exemplars with little diver-sity at the outset, and then expand the diversityto include dimensions of the desired generaliza-tion? Or is it a more productive endeavor totrain fewer exemplars that represent a greaterdiversity, and persist in the training until gen-eralization emerges'?

5. Train Loosely

One relatively simple technique can be con-ceptualized as merely the negation of discrimi-nation technique. That is, teaching is conductedwith relatively little control over the stimulipresented and the correct responses allowed, soas to maximize sampling of relevant dimensionsfor transfer to other situations and other formsof the behavior. A formal example of this mostoften informal technique was provided bySchroeder and Baer (1972), who taught vocalimitation skills to retarded children in both oftwo ways, one emphasizing tight restriction ofthe vocal skills being learned at the moment(serial training of vocal imitations), and theother allowing much greater range of stimuliwithin the current problem (concurrent training

357


of imitations). The latter method was charac-terized repeatedly by greater generalization toas-yet-untaught vocal imitation problems, thusaffirming "loose" teaching techniques as a con-tributor to wider generalization.

It will be appreciated that the literature of thefield contains very few examples of this type.Researchers always have attempted to maintainthorough control and careful restriction andstandardization of their teaching procedures,primarily to allow easy subsequent interpretationof the nature of their (successful) teaching tech-niques. Yet the import of this technique is thatcareful management of teaching techniques to aprecisely repetitive handful of stimuli or formatsmay, in fact, correspondingly restrict generaliza-tion of the lessons being learned. The ultimateforce of this recommendation remains to be seen.What seems required is programmatic researchaimed at assessing the generalization character-istics of lessons taught under careful, restrictedconditions, relative to similar lessons taughtunder looser, more variable conditions.

6. Use Indiscriminable ContingenciesIntermittent schedules of reinforcement have

been shown repeatedly to be particularly resistantto extinction, relative to continuous schedules(Ferster and Skinner, 1957). Resistance to extinc-tion may be regarded as a form of generaliza-tion-generalization across time subsequent tolearning. The essential feature of intermittentschedules may be their unpredictability-theimpossibility of discriminating reinforcement oc-casions from nonreinforcement occasions untilafter the fact. Thus, if contingencies of reinforce-ment or punishment, or the setting events thatmark the presence or absence of those contingen-cies, are made indiscriminable, then generaliza-tion may well be observed.

In generalization, behavior occurs in settings inwhich it will not be reinforced, just as it does insettings in which it will be reinforced. Then, theanalogue to an intermittent schedule, extendedto settings, is a condition in which the subject

cannot discriminate in which settings a responsewill be reinforced or not reinforced. A potentialapproximation to such a condition was presentedin a study by Schwarz and Hawkins (1970).In that experiment, the behavior of a sixth-gradechild was videotaped during math and spellingclasses. Later, after each school day had ended,the child was shown the tape of the math classand awarded reinforcers according to how oftengood posture, absence of face-touching, and ap-propriate voice-loudness were evident on thattape. Although reinforcers were awarded onlyon the basis of behaviors displayed during themath class, desirable improvements were ob-served during the spelling class as well. In thatreinforcement was delayed, this technique musthave made it difficult for the child to discriminatein which class the behaviors were critical forearning reinforcement. In other words, the gen-eralized success of the study may well be at-tributable to the partly indiscriminable natureof the reinforcement contingency.

In general, it may be suspected that delayedreinforcement often will have the advantage ofmaking the times and places in which the con-tingency actually operates indiscriminable to thesubject. However, this advantage is an advantage,by hypothesis, primarily for the goal of general-ization. Otherwise, delayed reinforcement wouldoften be considered an inefficient technique, mostespecially so for the initial development of a newskill. Indeed, it may be exactly in the realm ofdisadvantaged persons such as retarded childrenthat the usual inefficiency of delayed reinforce-ment may seem the most severe handicap to itsuse. However, its potential for fostering general-ization suggests strongly that further research beinvested in this procedure (and any others thatmake reinforcement contingencies properly in-discriminable), to develop methods of applyingit perhaps only after the initial development of anew skill, in the interests of promoting gen-eralization.

Less than a dozen studies of generalizationinterpretable as cases of indiscriminable rein-forcement contingencies can be found in the

358


literature. Kazdin (1973), for example, showedthat teacher attention to one retarded child wasresponded to by another child as if it were rein-forcement for on-task behavior. Indeed, theonlooker reacted with increased on-task behavior,even when the teacher attended to the targetchild's off-task behavior. Possibly, prior experi-ence with reinforcement contingent on the peers'on-task behavior was sufficient to make all futurepraise (contingent or not) discriminative for on-task behavior. In other words, with sufficientprior experience, the onlooker may have stoppedobserving the contingency in which the rein-forcement operated and responded only to thereinforcing stimulus' presence, making the con-tingency functionally indiscriminable.

Generalization across subjects has similarlybeen reported by Broden, Bruce, Mitchell, Carter,and Hall (1970) in a classroom of culturallydisadvantaged children. When positive teacherattention was given for one child's attention toacademic work, the attending of a peer also in-creased. This generalization was also a probablefunction of the cueing properties of teacher rein-forcement. However, the generalization observedmay also have been due to the manipulation ofnatural social consequences received by the non-target child through peer attention, or may havebeen caused by a slight increase in the amountof teacher attention to the nontarget child. Theseeffects deserve further systematic evaluation be-cause of their relevance to the classroom prac-tices of many teachers who strive to instructeffectively but are unable to devote extensivetime to individual children.

Pendergrass (1972) showed that timeout couldbe employed to decrease the destructive behaviorof two retarded children. With one subject, de-creased rates were also observed with anotherresponse (self-biting) which was sometimeschained to the destructive behavior, but notitself subjected to contingent timeout. However,with the second subject, generalization to a sec-ond response (autistic jerking movement) wasnot observed. Analysis of the data revealed thatthe two behaviors occurred simultaneously more

frequently with the subject with whom general-ization was evidenced. Thus, with this subject,punishment of the generalization response oc-curred more frequently when destructive behav-ior was punished. Unfortunately, it was notdetermined how often the self-biting occurredat times not simultaneous with the destructivebehavior. Therefore, the schedule of punishmentfor self-biting was not established, i.e., whetherbiting occurred only when destructive behavioroccurred and, therefore, always met the timeoutcontingency. In this example (which was notintended to be a careful analysis of the indis-criminable reinforcement concept), not only wasthe reinforcement contingency somewhat diffi-cult to discriminate, but the two behaviors (de-structive and self-destructive responses) also maywell have been only somewhat differentiated bythe subject.

Thus, preventing the ready discrimination ofcontingencies is a generalization-programmingtechnique worthy of application and research.Perhaps a random or haphazard delivery of re-inforcement will (if luck or good judgementprevails) function to modify targetted behavioras well as behavior occurring in proximal timeor space. Even noncontingent reinforcement,delivered at the outset of an intervention pro-gram, may retard initial effects, but may workto later advantage in generalization outcomes.

Finally, Kazdin and Polster (1973) showedonce again the usefulness of intermittent sched-ules to delay subsequent extinction, relative tocontinuous schedules of reinforcement. Socialinteraction by two retardates was reinforced withtokens. After establishing social interaction, onesubject received continuous reinforcement andthe other, intermittent reinforcement. Duringextinction, only the subject who received inter-mittent reinforcement continued to interact so-cially with peers. However, these results maysimply reflect different extinction rates by twosubjects. The research was essentially a groupstudy where N 1. Adequate single-subject ex-perimental control was lacking. Therefore, repli-cation of these procedures would be desirable.

359


7. Program Common Stimuli

The passive approach to generalization de-scribed earlier need not be a completely imprac-tical one. If it is supposed that generalizationwill occur, if only there are sufficient stimuluscomponents occurring in common in both thetraining and generalization settings, then a rea-

sonably practical technique is to guarantee thatcommon and salient stimuli will be present inboth. One predictor of the salience of a stimulusto be chosen for this role is its already establishedfunction for other important behaviors of thesubject.

Children's peers may represent peculiarlysuitable candidates for a stimulus common to

both training and generalization settings. Anexample has been provided by Stokes and Baer(1976). In their study, two children exhibitingserious learning disabilities were recruited to

learn several word-recognition skills. One childwas taught these skills and concurrently shownhow to teach them to the other child, thus actingas a peer-tutor. It was found that both childrenreliably learned the skills, but that neither gen-

eralized them reliably or stably to somewhat dif-ferent settings in which the other child usuallywas absent. However, when the peer-tutor was

brought into those settings, then each child simi-larly showed greatly increased and stabilizedgeneralization, even though there were never

any consequences for generalization. Similardemonstrations have been provided by Johnstonand Johnston (1972) for the skill of speecharticulation. In that study, peers were rewardedfor correct monitoring of the subjects' articula-tion. Generalization of correct articulation oc-

curred only when the "monitoring" peer was

present. Unfortunately, it was not determinedclearly whether generalization was evidenced be-cause of the discriminative properties of thepeers' presence in both settings, or whether thepeers actively continued their monitoring in thegeneralization setting.

Rincover and Koegel (1975) have also incor-porated functional training stimuli into the gen-

eralization setting. Autistic children were re-warded for imitation and instruction-followingin a training setting. Four of their 10 subjectsthen did not exhibit generalization to a differentsetting. Therefore, to program for this general-ization, various aspects of the training procedures(e.g., hand movement by therapist) or physicaltraining environment (e.g., table and chairs)were systematically introduced to the generaliza-tion setting to control generalization. Makingthe experimental setting more closely resemblethe regular classroom (generalization setting)was the programming procedure employed byKoegel and Rincover (1974). They decreasedthe teacher-to-student ratio in the experimentalsetting from 1-to-i to 1-to-8. After these specialprogramming conditions were instituted, therewas increased performance on previously learnedand new behaviors learned in the classroom.Walker and Buckley (1972) programmed gener-alization of the effects of remedial training ofsocial and academic classroom behavior by estab-lishing common stimuli between the experimen-tal remedial classroom and the childrens' regularclassroom by using the same academic materialsin both classrooms.

The literature of this field shows only a hand-ful of studies deliberately making use of a com-mon stimulus in both training and generalizationsettings. Obviously, this is a technological dimen-sion urgently in need of thorough development.The use of peers as the common stimulus hasmuch to recommend it as a practical and naturaltechnique. To what extent peers need to partici-pate in the training setting has not yet beendetermined, although the absence of generaliza-tion sometimes shown when peers are presentin nontraining settings, suggests that peers notinvolved in a training setting will not likelyacquire sufficient discriminative function to con-trol generalized responding. The use of commonphysical stimuli is in even greater need of sys-tematic research. A common stimulus approachto generalization would encourage the incorpora-tion into training settings of (naturally occur-ring) physical stimuli that are frequently promi-

360


nent or functional in nontraining environments.If these stimuli are well chosen, and can be madefunctional and salient in the training procedures,then generalization may thereby be programmed.

8. Mediate Generalization

Mediated generalization is well known as atheoretical mechanism explaining generalizationof highly symbolic learnings (Cofer and Foley,1942). In essence, it requires establishing a re-sponse as part of the new learning that is likelyto be utilized in other problems as well, and willconstitute sufficient commonality between theoriginal learning and the new problem to resultin generalization. The most commonly usedmediator is language, apparently. However, thedeliberate application of language to accomplishgeneralization is rare in the literature reviewed,and correspondingly little is known about whataspects of a language response make for bestmediation.A sophisticated analysis of mediated general-

ization was conducted by Risley and Hart ( 1968),who taught preschool children to report at theend of play on their play-material choices. Men-tion of a given choice was reinforced with snacks,which produced increased mentioning of thatchoice, but no change in the children's actualuse of that play-material. When reinforcementwas restricted to true reports of play-materialchoices, however, the children then changed theirplay behavior (the next day) so that whenqueried about that play, they could truthfullyreport on their use of the specified play materialand earn reinforcement. Control over any choiceof play materials proved possible with this tech-nique, which placed teaching contingencies noton the play, but on a potential mediator (verbalreport) of that play behavior. That the reportswere only potential mediators was apparent inthe early stages of the study, when the childrenreadily reported (untruly) their use of play ma-terials with no corresponding actual behaviorwith those materials; at that stage, they earnedreinforcement even so. When the reinforcement

was restricted to true reports, the reports thenbecame mediators of play behavior. The lessongeneralized, such that after several sequentialexperiences with these procedures, the childrenthen used reports about play as mediators, evenwithout reinforcement being restricted to onlytrue reports. Israel and O'Leary (1973) usedessentially the same paradigm to compare theeffects of having children report first what theywould play with later, in contrast to having themreport after play what they had done (the Risleyand Hart method); they found that reinforcingpostreports (when they were true) produced moreactual behavior (the next day) than reinforcingthe actual behavior when it agreed with theearlier promise to perform it. This technique hasbeen extended subsequently to the case of socialskills, specifically sharing and praising betweenyoung children (Rogers-Warren and Baer, 1976).In that case, modelling was added, such that theyoung children would have a thorough chanceto learn the nature of the relatively complexresponses at issue.

Obviously, verbal mediation can easily fail,most especially in those situations in which theverbal mediators have little meaning (i.e., tightlyrestricted discriminative value) for the subjects.It is commonplace to find children agreeing to aquery (e.g., about whether they praised or shared)without any knowledge of what that must entailin actual behavior. In the case of retarded chil-dren, it might be particularly true that the abilityto use verbal responses as mediators would lagbehind that of normal children using the samelanguage responses. It may be reasonable tosuggest that in the development of language-training programs, systematic attention be givento the training of language skills sufficiently wellelaborated to function as mediators of nonverbalbehavior. Language is a response, of course; it isalso, equally obviously, a stimulus to the speakeras well as to the listener. Thus, it meets perfectlythe logic of a salient common stimulus, to becarried from any training setting to any general-ization setting that the child may ever enter.It also perfectly exemplifies the essence of the

361


active generalization approach recommendedearlier.The mediation of generalization is also exem-

plified in the behavior analysis research of self-control and self-management procedures. Thatis, self-control procedures such as self-recording,taught as part of an intervention program, mayfunction to promote generalization: such tech-niques are easy to transport and may be em-ployed readily to facilitate responding undergeneralization conditions. Some research that hasemployed any or all of the various tactics of self-assessment, self-recording, self-determination ofreinforcement, and/or self-administration of re-inforcement (Glynn, Thomas, and Shee, 1973),has also displayed maintenance and generaliza-tion of behavior change; however, the correla-tion is not perfect.

Broden, Hall, and Mitts (1971) reported thatafter an eighth-grade girl experienced self-recording of study behavior and teacher praisefor improved study, her study behavior main-tained at a high level for a recorded three weeks.Although the individual effects of the self-record-ing and praise were not determined, it is possiblethat the self-recording procedures contributedsignificantly to this generalization.

Drabman, Spitalnik, and O'Leary (1973)taught disruptive children to match their teach-er's evaluations of their appropriate classroombehavior. Tokens were dispensed for appropriateclassroom behavior and accurate matching. Dis-ruptive classroom behavior decreased and wasmaintained at low levels during a 12-day phasewhen tokens were not dispensed for self-record-ing accuracy. Generalized behavior improvementwas also evident during a 15-min no-tokenperiod within the experimental hour. Thesechanges were possibly a function of the closetemporal proximity of the token periods, whichfrequently immediately preceded or followedthe generalization period.

The role of self-control procedures in medi-ating generalization has often been proposed.Research would do well to examine the contri-bution of self-control tactics in generalization

and maintenance, especially when formal inter-vention manipulations have ceased to operate.The effects of accompanying procedures shouldbe experimentally separated from self-controleffects, and the role of each of the various self-control tactics (Glynn et al., 1973) should beindividually analyzed. The potential of self-mediated generalization is apparent, but its im-plications and practical utility still remain to beassessed.

9. Train "To Generalize"If generalization is considered as a response

itself, then a reinforcement contingency may beplaced on it, the same as with any other operant.Informally, teachers often do this when they urgea student who has been taught one example ofa general principle to "see" another exampleas "the same thing". (In principle, they are alsoattempting to make use of language as a medi-ator of generalization, relying on the supposedcharacteristics of words like "same" to accom-plish the generalization.) Common observationsuggests that the method often fails, and thatwhen it does succeed, little extrinsic reinforce-ment is offered as a consequence. A more formalexample of the technique was seen in a studyby Goetz and Baer (1973), in which three pre-school children were taught to generalize theresponse of making block forms (in blockbuild-ing play). Descriptive social reinforcement wasoffered only for every different form the childmade, i.e., contingent on every first appearanceof any blockbuilding form within a session, butnot for any subsequent appearances of that form.Thus, the child was rewarded for moving alongthe generalization gradient underlying block-form inventions, and never for staying at anyone point. In general, the technique succeeded, inthat the children steadily invented new blockforms while this contingency was in use. Thus,there exists the possibility of programming rein-forcement specifically, perhaps only, for move-ment along the generalization gradient desired.

In largely unspecified ways, perhaps two otherstudies exemplify this logic. Herbert and Baer

362


(1972), for example, taught two mothers ofdeviant children to give social reinforcementonly to their children's appropriate behaviors,but taught the mothers from the outset to judgeall behavior according to criteria they helpedto develop, rather than attack only a few speci-fied child responses. These mothers learned ageneralized skill because they applied correctsocial contingencies to categories that includedvirtually all appropriate child behavior likely tooccur. Behavior changes were maintained at 20and 24 weeks after completion of formal train-ing. Similarly, Parsonson, Baer, and Baer (1974)taught two teachers of retarded children to applygeneralized correct social contingencies to alllikely appropriate and inappropriate behaviorsof preschool retarded children. These effects werealso durable over several months. Apparentlygeneralized changes were produced in these stud-ies by Herbert and Baer and Parsonson et al.,but the extent and quality of that generalizationwas not quantified as such.

Very few studies of this type are found in theliterature of applied behavior analysis, probablybecause of the preference of behaviorists to con-sider generalization as an outcome of behavioralchange, rather than as a behavior itself. Ulti-mately, this behavioristic stance may well provedurable and consistent. Meanwhile, it is worthhypothesizing that "to generalize" may be treatedas if it were an operant response, and reinforcedas such, simply to see what useful results occur.

Consequently, one other technique deservesdiscussion: the systematic use of instructions tofacilitate generalization. Thus, if a behavior istaught and generalization is not displayed, theleast expensive of all techniques is to tell thesubject about the possibility of generalizationand then ask for it. If that generalization thenoccurs, it may well be referred to as "instructedgeneralization". If the effects of that instructionare themselves to become generalized (yieldinga "generalized generalizer"?), then reinforcementof the generalized behavior, on a suitable sched-ule, might well be prudent, at least at first. Per-haps it is simply a very elaborate version of this

technique that is being practiced when a clientis taught to relax in a somewhat anxiety-arousingsituation, and reinforced (socially) for doing so;and then is instructed to relax in a somewhatmore powerful anxiety-arousing situation, etc.That is, systematic desensitization to a heirarchyof stimuli may be analyzed as reinforcing not justrelaxation, but also generalization along an al-ready constructed generalization gradient (cf.Yates, 1970, p. 64ff.).

CONCLUSION

The structure of the generalization literatureand its implicit embryonic technology has beensummarized. The most frequent treatments ofgeneralization are also the least analytical-thosedescribed as Train and Hope and SequentialModification. Included in the category of Trainand Hope were those studies where the potentialfor generalization had been recognized, its pres-ence or absence noted, but no particular effortwas expended to accomplish generalization. Bycontrast, some limited programming was imple-mented in the Sequential Modification research.In these studies, given an absence of reliablegeneralization, procedures to effect changes wereinstituted directly in every nongeneralized condi-tion. Although contributing significantly to ourunderstanding of the generalization of behavior-change programs, these studies are not examplesof the programming of generalization.

Seven categories were discussed that directlyrelate to a technology of generalization. First,the potential role of Natural Maintaining Con-tingencies was discussed. According to this tactic,generalization may be programmed by suitabletrapping manipulations, where responses are in-troduced to natural reinforcement communitiesthat refine and maintain those skills withoutfurther therapeutic intervention. The Trainingof Sufficient Exemplars is numerically the mostextensive area of programming: generalizationto untrained stimulus conditions and to un-trained responses is programmed by the trainingof sufficient exemplars of those stimulus condi-

363


tions or responses. Train Loosely is a program-ming technique in which training is conductedwith relatively little control over the stimuli andresponses involved, and generalization is therebyenhanced. To invoke the tactic of Indiscrimi-nable Contingencies, the contingencies of re-inforcement or punishment, or the setting eventsmarking the presence or absence of those con-tingencies, are deliberately made less predictable,so that it becomes difficult to discriminate rein-forcement occasions from nonreinforcement oc-casions. Common Stimuli may be employed ingeneralization programming by incorporatinginto training settings social and physical stimulithat are salient in generalization settings, andthat can be made to assume functional or obviousroles in the training setting. Mediated General-ization requires establishing a response as partof new learning that is likely to be utilized inother problems as well, and thus result in gener-alization. The final technique, Train "To Gener-alize", involves reinforcing generalization itselfas if it were an explicit behavior. These program-ming techniques should be researched furtherand usefully applied in programs in which gen-eralization is relevant.

This list of generalized tactics conceals withinitself a much smaller list of specific tactics. Thesespecific tactics can be presented as a small pic-ture of the generalization technology in its pres-ent most pragmatic form, not only to offer a setof what-to-do possibilities, but also to emphasizehow very small the current technology is andhow much development it requires:

1. Look for a response that enters a naturalcommunity; in particular, teach subjects tocue their potential natural communities toreinforce their desirable behaviors.

2. Keep training more exemplars; in particu-lar, diversify them.

3. Loosen experimental control over the stim-uli and responses involved in training; inparticular, train different examples concur-rently, and vary instructions, SDs, socialreinforcers, and backup reinforcers.

4. Make unclear the limits of training contin-gencies; in particular, conceal, when pos-sible, the point at which those contingen-cies stop operating, possibly by delayedreinforcement.

5. Use stimuli that are likely to be found ingeneralization settings in training settingsas well; in particular, use peers as tutors.

6. Reinforce accurate self-reports of desirablebehavior; apply self-recording and self-reinforcement techniques whenever possi-ble.

7. When generalizations occur, reinforce atleast some of them at least sometimes, asif "to generalize" were an operant responseclass.

There are many examples of generalizationand nongeneralization of behavior changes. Thefact that apparently unprogrammed generaliza-tion has been demonstrated (particularly acrosstime) is valuable. It heralds a practicality de-sirable in any technology of behavior: that everyone of a subjects' responses, in every setting,with every experimenter, and at every conceiv-able time does not need to meet specific treat-ment consequences for that program to accom-plish and maintain important behavior changes.Alternatively, the fact that generalization is notalways observed and durability is not inevitablemeans that there is hope for behavior modifica-tion: behavior can always be modified andchanges are not necessarily irreversible. That is,once behavior has been modified, there is stillthe possibility of reconditioning if changes areundesirable or inappropriate, or if new inappro-priate behaviors develop. If both appropriateand inappropriate behavior changes were to per-sist and prove irreversible, it would presage thedemise of any technology of behavioral inter-vention. This occurrence of nongeneralizationalso underlines the need to develop a technologyof generalization, so that programming will bea fundamental component of any procedureswhen durability and generalization of behaviorchanges are desirable.

364

AN IMPLICIT TECHNOLOGY OF GENERALIZATION 365

A most important question is prompted by anexamination of the previous research: does gen-eralization ever occur without programming?In the above research, generalization was not al-ways evident. In fact, the highly discriminatedeffects of some operant programs were some-times documented. We have seen that the behav-ior analysis literature describes various programsthat have shown that generalization may be pro-moted or programmed by particular interventiontechniques. It seems reasonable to suggest, then,that many of the successful Train-and-Hopeexamples cited above may be undiagnosed in-stances of informal or inadvertent programmingtechniques, rather than an absence of program-ming techniques. It cannot be discounted, and isindeed possible, that these generalization exam-ples may simply depict successful programmedgeneralization, and neither the authors of thosepapers, nor the present authors have recognizedor hypothesized the programming technique.

Perhaps the most pragmatic orientation forbehavior analysts is to assume that generalizationdoes not occur except through some form of pro-gramming. Thus, the best course of action seemsto be that of systematic measurement and analy-sis of variables that may have been functional inany apparently unprogrammed generalization.These analyses should be included as part of allresearch where "unprogrammed" generalized be-havior changes are evidenced, for discriminatedbehavior changes may well be the rule if gen-eralization is not specifically programmed. Suchanalyses, if successful, will contribute to a tech-nology of generalization by further developingthe understanding of critical variables that func-tion to produce generalization, and would furtheremphasize the need always to be concerned notonly with generalization issues, but with the vari-ous techniques that accomplish generalization.

In other words, behavioral research and prac-tice should act as if there were no such animalas "free" generalization-as if generalizationnever occurs "naturally", but always requiresprogramming. Then, "programmed generaliza-tion" is essentially a redundant term, and snould

be descriptive only of the active regard of re-searchers and practitioners.

REFERENCESAllen, G. J. Case study: Implementation of behav-

ior modification techniques in summer camp set-tings. Behavior Therapy, 1973, 4, 570-575.

Ayllon, T. and Azrin, N. H. The token economy.New York: Appleton-Century-Crofts, 1968.

Azrin, N. H., Sneed, T. J., and Fox, R. M. Dry-bed:A rapid method of eliminating bedwetting (enu-resis) of the retarded. Behaviour Research andTherapy, 1973, 11, 427-434.

Baer, D. M., Peterson, R. F., and Sherman, J. A. Thedevelopment of imitation by reinforcing behav-ioral similarity to a model. journal of the Experi-mental Analysis of Behavior, 1967, 10, 405-416.

Baer, D. M. and Wolf, M. M. The entry into natu-ral communities of reinforcement. In R. Ulrich,T. Stachnik, and J. Mabry (Eds), Control of hu-man behavior: Volume II. Glenview, Illinois:Scott, Foresman, 1970. Pp. 319-324.

Baer, D. M., Wolf, M. M., and Risley, T. R. Somecurrent dimensions of applied behavior analysis.Journal of Applied Behavior Analysis, 1968, 1,91-97.

Bolstad, 0. D. and Johnson, S. M. Self-regulation inthe modification of disruptive behavior. Journalof Applied Behavior Analysis, 1972, 5, 443-454.

Broden, M., Hall, R. V., and Mitts, B. The effect ofself-recording on the classroom behavior of twoeighth-grade students. Journal of Applied Behav-ior Analysis, 1971, 4, 191-199.

Broden, M., Bruce, C., Mitchell, M. A., Carter, U.,and Hall, R. V. Effects of teacher attention onattending behavior of two boys at adjacent desks.Journal of Applied Behavior Analysis, 1970, 3,199-203.

Buehler, R. E., Patterson, G. R., and Furniss, J. M.The reinforcement of behavior in institutional set-tings. Behaviour Research and Therapy, 1966, 4,157-167.

Buell, J., Stoddard, P., Harris, F. R., and Baer, D. M.Collateral social development accompanying rein-forcement of outdoor play in a preschool child.Journal of Applied Behavior Analysis, 1968, 1,167-173.

Cofer, C. N. and Foley, J. P. Mediated generaliza-tion and the interpretation of verbal behavior: I.Prolegomena. Psychological Review, 1942, 49,513-540.

Drabman, R. S., Spitalnik, R., and O'Leary, K. D.Teaching self control to disruptive children.Journal of Abnormal Psychology, 1973, 82, 10-16.

Ferster, C. B. and Skinner, B. F. Schedules of rein-forcement. New York: Appleton-Century-Crofts,1957.

366 TREVOR F. STOKES and DONALD M. BAER

Garcia, E. The training and generalization of aconversational speech form in nonverbal retardates.Journal of Applied Behavior Analysis, 1974, 7,137-149.

Garcia, E., Baer, D. M., and Firestone, I. The devel-opment of generalized imitation within topo-graphically determined boundaries. Journal ofApplied Behavior Analysis, 1971, 4, 101-112.

Gelfand, D. M., Gelfand, S., and Dobson, W. R.Unprogrammed reinforcement of patients behav-ior in a mental hospital. Behaviour Research andTherapy, 1967, 5, 201-207.

Glynn, E. L., Thomas, J. D., and Shee, S. M. Behav-ioral self-control of on-task behavior in an ele-mentary classroom. Journal of Applied BehaviorAnalysis, 1973, 6, 105-113.

Goetz, E. M. and Baer, D. M. Social control of formdiversity and the emergence of new forms in chil-dren's blockbuilding. Journal of Applied Behav-ior Analysis, 1973, 6, 105-113.

Graubard, P. S., Rosenberg, H., and Miller, M. B.Student applications of behavior modification toteachers and environments or ecological ap-proaches to social deviancy. In E. A. Ramp andB. L. Hopkins (Eds), A new direction for educa-tion: behavior analysis. Lawrence, Kansas: Sup-port and Development Center for FollowThrough, 1971. Pp. 80-101.

Griffiths, H. and Craighead, W. E. Generalizationin operant speech therapy for misarticulation.Journal of Speech and Hearing Disorders, 1972,37, 485-494.

Guess, D., Sailor, W., Rutherford, G., and Baer, D. M.An experimental analysis of linguistic develop-ment: the productive use of the plural morpheme.Journal of Applied Behavior Analysis, 1968, 1,297-306.

Hall, R. V. and Broden, M. Behavior changes inbrain-injured children through social reinforce-ment. Journal of Experimental Child Psychology,1967, 5, 463-479.

Herbert, E. W. and Baer, D. M. Training parentsas behavior modifiers: self-recording of contin-gent attention. Journal of Applied Behavior Anal-ysis, 1972, 5, 139-149.

Horner, R. D. Establishing use of crutches by amentally retarded spina bifida child. Journal ofApplied Behavior Analysis, 1971, 4, 183-189.

Israel, A. C. and O'Leary, K. D. Developing corre-spondence between children's words and deeds.Child Development, 1973, 44, 575-581.

Johnston, J. M. and Johnston, G. T. Modification ofconsonant speech-sound articulation in youngchildren. Journal of Applied Behavior Analysis,1972, 5, 233-246.

Kazdin, A. E. The effect of vicarious reinforcementon attentive behavior in the classroom. Journal ofApplied Behavior Analysis, 1973, 6, 71-78.

Kazdin, A. E. and Polster, R. Intermittent token re-

inforcement and response maintenance in extinc-tion. Behavior Therapy, 1973, 4, 386-391.

Keller, F. S. and Schoenfeld, W. N. Principles ofpsychology. New York: Appleton-Century-Crofts,1950.

Kifer, R. E., Lewis, M. A., Green, D. R., and Phillips,E. L. Training predelinquent youths and theirparents to negotiate conflict situations. Journal ofApplied Behavior Analysis, 1974, 7, 357-364.

Koegel, R. L. and Rincover, A. Treatment of psy-chotic children in a classroom environment: I.Learning in a large group. Journal of AppliedBehavior Analysis, 1974, 7, 45-59.

Meichenbaum, D. H., Bowers, K. S., and Ross, R. R.Modification of classroom behavior of institu-tionalized female adolescent offenders. BehaviourResearch and Therapy, 1968, 6, 343-353.

Parsonson, B. S., Baer, A. M., and Baer, D. M. Theapplication of generalized correct social contin-gencies by institutional staff: an evaluation ofthe effectiveness and durability of a training pro-gram. Journal of Applied Behavior Analysis,1974, 7, 427-437.

Pendergrass, V. E. Timeout from positive reinforce-ment following persistent high-rate behavior inretardates. Journal of Applied Behavior Analysis,1972, 5, 85-91.

Redd, W. H. and Birnbrauer, J. S. Adults as dis-criminative stimuli for different reinforcementcontingencies with retarded children. Journal ofExperimental Child Psychology, 1969, 7, 440-447.

Rincover, A. and Koegel, R. L. Setting generalityand stimulus control in autistic children. Journalof Applied Behavior Analysis, 1975, 8, 235-246.

Risley, T. R. and Hart, B. M. Developing corre-spondence between the nonverbal and verbal be-havior of preschool children. Journal of AppliedBehavior Analysis, 1968, 1, 267-281.

Rogers-Warren, A. and Baer, D. M. Correspondencebetween saying and doing: teaching children toshare and praise. Journal of Applied BehaviorAnalysis, 1976, 9, 335-354.

Schroeder, G. L. and Baer, D. M. Effects of concur-rent and serial training on generalized vocal imi-tation in retarded children. Developmental Psy-chology, 1972, 6, 293-301.

Schumaker, J. and Sherman, J. A. Training genera-tive verb usage by imitation and reinforcementprocedures. Journal of Applied Behavior Analysis,1970, 3, 273-287.

Schwarz, M. L. and Hawkins, R. P. Application ofdelayed reinforcement procedures to the behaviorof an elementary school child. Journal of AppliedBehavior Analysis, 1970, 3, 85-96.

Seymour, F. W. and Stokes, T. F. Self-recording intraining girls to increase work and evoke staffpraise in an institution for offenders. Journal ofApplied Behavior Analysis, 1976, 9, 41-54.

AN IMPLICIT TECHNOLOGY OF GENERALIZATION 367

Skinner, B. F. Science and human behavior. NewYork: Macmillan, 1953.

Solomon, R. W. and Wahler, R. G. Peer reinforce-ment control of classroom problem behavior.Journal of Applied Behavior Analysis, 1973, 6,49-56.

Stokes, T. F. and Baer, D. M. Preschool peers asmutual generalization-facilitating agents. BehaviorTherapy, 1976, 7, 549-556.

Stokes, T. F., Baer, D. M., and Jackson, R. L. Pro-gramming the generalization of a greeting re-sponse in four retarded children. Journal of Ap-plied Behavior Analysis, 1974, 7, 599-610.

Stolz, S. B. and Wolf, M. M. Visually discriminated

behavior in a "blind" adolescent retardate. Journalof Applied Behavior Analysis, 1969, 2, 65-77.

Wahler, R. G. Setting generality: some specific andgeneral effects of child behavior therapy. Journalof Applied Behavior Analysis, 1969, 2, 239-246.

Walker, H. M. and Buckley, N. K. The use of posi-tive reinforcement in conditioning attending be-havior. Journal of Applied Behavior Analysis,1968, 1, 245-250.

Yates, A. J. Behavior therapy. New York: JohnWiley and Sons, 1970.

Received 22 December 1975.(Final acceptance 3 June 1976.)

Documents

Stokes and Baer 1977