rewards & intrinsic motivation.pdf

8/10/2019 rewards & intrinsic motivation.pdf

1/44

1

The Behavior Analyst No. 1 (Spring)2001, 24, 144

Pervasive Negative Effects ofRewards on Intrinsic Motivation:

The Myth Continues

Judy Cameron, Katherine M. Banko,and W. David PierceUniversity of Alberta

A major concern in psychology and education is that rewards decrease intrinsic motivation toperform activities. Over the past 30 years, more than 100 experimental studies have been conductedon this topic. In 1994, Cameron and Pierce conducted a meta-analysis of this literature and con-cluded that negative effects of reward were limited and could be easily prevented in applied settings.A more recent meta-analysis of the literature by Deci, Koestner, and Ryan (1999) shows pervasivenegative effects of reward. The purpose of the present article is to resolve differences in previousmeta-analytic findings and to provide a meta-analysis of rewards and intrinsic motivation that per-mits tests of competing theoretical explanations. Our results suggest that in general, rewards are notharmful to motivation to perform a task. Rewards given for low-interest tasks enhance free-choiceintrinsic motivation. On high-interest tasks, verbal rewards produce positive effects on free-choicemotivation and self-reported task interest. Negative effects are found on high-interest tasks when

the rewards are tangible, expected (offered beforehand), and loosely tied to level of performance.When rewards are linked to level of performance, measures of intrinsic motivation increase or donot differ from a nonrewarded control group. Overall, the pattern of results indicates that rewardcontingencies do not have pervasive negative effects on intrinsic motivation. Theoretical and prac-tical implications of the findings are addressed.

Key words: meta-analysis, rewards, reinforcement, intrinsic motivation, intrinsic interest

Most parents, educators, and behav-ior analysts would agree that the idealstudent is one who performs academictasks at a high level, shows high inter-est and involvement in school activi-ties, is willing to take on challenging

assignments, and is a self-motivatedlearner. To instill interest and to height-en student performance, many practi-tioners implement reward and incen-tive systems in educational settings. Inrecent years, the wisdom of this prac-tice has been debated in literature re-views, textbooks, and the popular me-dia. Many writers and researchersclaim that giving students high grades,prizes, money, and even praise for en-gaging in an activity may be effectivein getting students to perform a task,

This work was supported by a research grantfrom the Social Sciences and Humanities Re-search Council of Canada.

Correspondence concerning this article shouldbe addressed to Judy Cameron, Department ofEducational Psychology, 6-102 EducationNorth, University of Alberta, Edmonton, Alber-ta, T6G 2G5 Canada.

but performance and interest are main-tained only as long as the rewards keepcoming. In other words, rewards aresaid to undermine intrinsic motivation.This premise is based on the view thatwhen individuals like what they are

doing, they experience feelings ofcompetence and self-determination.When students are given a reward forperformance, the claim is that they be-gin to do the activity for the externalreward rather than for intrinsic reasons.As a result, perceptions of competenceand self-determination are said to de-crease and motivation to perform theactivity declines.

Those who decry the use of rewardssupport their position by citing exper-imental studies on rewards and intrin-sic motivation conducted in social psy-

chology and education. Since the1970s, dozens of experiments havebeen designed to assess the impact ofrewards on intrinsic motivation. A cur-sory examination of the studies, how-ever, reveals a mixed set of findings.That is, in some studies, extrinsic re-


2/44

2 JUDY CAMERON et al.

wards produce negative effects onmeasures of intrinsic motivation. Otherstudies find positive effects of reward;still others show no effect. A numberof reviewers have noted the contradic-

tory nature of the findings and have at-tempted to identify the conditions un-der which extrinsic rewards producedecrements on measures of intrinsicmotivation (Bates, 1979; Bernstein,1990; Carton, 1996; Dickinson, 1989;Flora, 1990; Morgan, 1984).

In 1994, Cameron and Pierce pub-lished a meta-analysis of 96 experi-mental studies on the topic (with ad-ditional analyses by Eisenberger &Cameron, 1996). Based on their re-sults, they argued that negative effectsof reward were minimal and could be

easily prevented in applied settings.The research and recommendationsmade by Cameron and Pierce and byEisenberger and Cameron generatedconsiderable debate (Hennessey &Amabile, 1998; Kohn, 1996; Lepper,1998; Lepper, Keavney, & Drake,1996; Ryan & Deci, 1996; Sansone &Harackiewicz, 1998) and seeminglyspurred the publication of a new meta-analysis on the topic. Deci, Koestner,and Ryan (1999) presented a meta-analysis that claimed to support theview that rewards have pervasive neg-

ative effects on intrinsic motivation.Deci et al. (1999) identified 128 ex-

periments on rewards and intrinsic mo-tivation, including 20 unpublishedstudies from doctoral dissertations.They outlined a number of concernsthey had with the meta-analyses con-ducted by Cameron and Pierce (1994)and Eisenberger and Cameron (1996).Deci et al.s meta-analysis was de-signed to rectify these concerns, to testcognitive evaluation theory, and toprovide a more comprehensive reviewof the literature. Their findings sup-

ported cognitive evaluation theory and,in general, rewards were found to havea substantial negative effect on intrin-sic motivation. Deci et al. concludedthat although rewards can controlpeoples behaviorindeed, that is pre-sumably why they are so widely ad-

vocatedthe primary negative effectof rewards is that they tend to forestallself-regulation (p. 659).

The assertion that rewards decreaseintrinsic motivation has captured the

attention of cognitive researchers,practitioners, and the general publicbecause such a claim (a) seems to offeran empirical basis for psychologicaltheories that assume that self-determi-nation and freedom from control arefundamental human motives, (b) ap-pears to question basic behavioral con-ceptions of human nature, and (c) sug-gests that rewards used in schools, hos-pitals, the workplace, and so on aremore harmful than beneficial. A re-viewer of this manuscript suggestedthat the claim that rewards are harmful

may be attractive to some practitionersand educators because detecting andrewarding performance improvementsis hard work and the negative effectclaim relieves us of a difficult and de-manding task.

Clearly, Deci et al.s (1999) findingof general negative effects of rewardhas important theoretical and practicalimplications and calls for a carefulanalysis of contradictory empiricalclaims. In this article, we argue thatpervasive negative effects of rewardare not a necessary outcome of a meta-

analysis of this literature. We contendthat a careful examination of Deci etal.s meta-analysis reveals several con-ceptual and methodological shortcom-ings. The disparate conclusions of thetwo major meta-analyses on the effectsof reward on intrinsic motivation(Cameron & Pierce, 1994; Deci et al.)suggest the value of correcting theflaws in each and building on theirstrengths to draw more definitive con-clusions. In this article, we offer a re-analysis of the effects of rewards onintrinsic motivation. Our reanalysis is

informed by a consideration of Deci etal.s decisions and procedures. In ad-dition, the concerns raised by Deci etal. about our previous research are ad-dressed. The purpose of the present ar-ticle is to resolve differences in previ-ous meta-analytic findings and to pro-


3/44

3THE MYTH CONTINUES

vide a meta-analysis of rewards and in-trinsic motivation that permits tests ofcompeting theoretical explanations.

We begin with a general descriptionof the experiments conducted on re-

wards and intrinsic motivation. This isfollowed by a brief description of theprocedure and logic of meta-analysis.The meta-analyses by Cameron andPierce (1994), Eisenberger and Cam-eron (1996), and Deci et al. (1999) aredescribed, and criticisms of each arepresented. We then provide a detailedaccount of how our reanalysis is de-signed to resolve differences betweenDeci et al.s and our earlier reviews ofthis literature. Results of our newmeta-analysis are presented, and dif-ferences between our findings and pre-

vious reviews are explained. Finally,our discussion focuses on theoreticaland practical implications of the find-ings.

THIRTY YEARS OFRESEARCH ON REWARDS

AND INTRINSICMOTIVATION

The termintrinsic motivationis gen-erally understood in contrast toextrin-sic motivation. Intrinsically motivatedbehaviors are those in which there is

no apparent reward except with the ac-tivity itself (Deci, 1975). Extrinsic mo-tivation, on the other hand, is said tooccur when an activity is rewarded byincentives not inherent in the task. Al-though these terms have been criticizedand debated (e.g., Bandura, 1986;Dickinson, 1989; Flora, 1990), they areaccepted by many researchers. The dis-tinction between intrinsic and extrinsicmotivation led psychologists to specu-late about the relation between thesetwo sources. One view was that intrin-sic and extrinsic motivation combined

in an additive fashion to produce over-all motivation. For example, in worksettings, organizational psychologistsargued that optimal performancewould occur when jobs were interest-ing and challenging and employeeswere externally rewarded (e.g., with

money) for their work (Porter & Law-ler, 1968; Vroom, 1964). Other theo-rists challenged the additive assump-tion, suggesting instead that extrinsicrewards might interfere with intrinsic

motivation (DeCharms, 1968).The idea that extrinsic rewards coulddisrupt intrinsic motivation instigated aseries of experiments carried out in theearly 1970s (Deci, 1971; Lepper,Greene, & Nisbett, 1973). In the initialstudies, researchers tested the hypoth-esis that external rewards would un-dermine intrinsic motivation either bysubverting feelings of competence andself-determination or by deflecting thesource of motivation from internal toexternal causes. Intrinsic motivationwas inferred from changes in time

spent on an activity once rewards wereremoved, performance during the non-rewarded phase, or expressed task in-terest. When rewards were found tolower time on task, performance, or in-terest, the researchers claimed that re-wards undermined intrinsic motivation.Results from the early studies appearedto offer some support for the under-mining hypothesis. That is, when in-dividuals were promised a material re-ward, their performance, time on task,and interest decreased once the rewardwas no longer forthcoming. Because of

the implications for education, busi-ness, and the psychology of motiva-tion, the early findings led to a greatdeal of research on the topic.

Since the 1970s, over 100 experi-ments have been performed to inves-tigate alleged undermining effects ofrewards. The vast majority of the stud-ies on rewards and intrinsic motivationhave been conducted using a between-groups design. In a typical study, par-ticipants are presented with an inter-esting task (e.g., solving and assem-bling puzzles, drawing with magic

markers, playing word games). Partic-ipants are rewarded with money orgrades, candy, praise, good-player cer-tificates, and so forth for performingthe activity. Rewards are tangible (e.g.,money, candy, gold stars) or verbal(e.g., praise, approval, positive feed-


4/44


back). In addition, the rewards may beoffered beforehand (expected reward)or presented unexpectedly after the ac-tivity (unexpected reward). In some ex-periments, reward is offered simply for

doing an activity; in other studies therewards are given for completing a taskor for each puzzle or unit solved. In anumber of experiments, the rewardsare offered for meeting or exceeding aspecific standard. Participants in a con-trol condition engage in the activitywithout receiving a reward.

The reward intervention is usuallyconducted over a 10-min to 1-hr peri-od. Rewarded and nonrewarded groupsare then observed during a nonrewardperiod (typically, 2 min to 1 hr) inwhich participants are free to continue

performing the target task or to engagein some alternative activity. The timeparticipants spend on the target activityduring this nonreward phase, their per-formance on the task during the free-choice period, or self-reported task in-terest are used as measures of intrinsicmotivation. If rewarded participantsspend less free time on the activity,perform at a lower level, or expressless task interest than nonrewardedparticipants, reward is said to under-mine intrinsic motivation.

The findings from the studies on re-

wards and intrinsic motivation havebeen diverse (positive, negative, andno effects have been reported). None-theless, the results from these studiesare often cited as evidence that rewardsand positive reinforcement can backfire(e.g., Kohn, 1993). External rewardsare said to be controlling and to inter-fere with a basic human desire for self-determination.

Because the detrimental effects ofrewards have been interpreted as achallenge to behavioral conceptions ofhuman nature and to the benefits of be-

havioral technology for education andbusiness, a few behaviorally orientedresearchers have used single-subjectdesigns to assess the generality of thefindings. In this type of study, partici-pants serve as their own controls. Mea-sures such as time on task are taken

over a number of sessions in a baselinephase, reinforcement procedures areimplemented over several sessions, andfinally, reward is withdrawn and timeon task is assessed on repeated occa-

sions. An increase or decrease in in-trinsic motivation is measured by thedifference in time spent on the task be-tween baseline and postreinforcementphases. In the five studies employingthis type of design (Davidson & Buch-er, 1978; Feingold & Mahoney, 1975;Mawhinney, Dickinson, & Taylor,1989; Skaggs, Dickinson, & OConnor,1992; Vasta, Andrews, McLaughlin,Stirpe, & Comfort, 1978), participantsperformance during the postrewardphase either exceeded or remained atthe same level as performance in the

prereward sessions. In other words,when the rewards were shown to func-tion as reinforcement and multiple-tri-als procedures were used, there was noevidence of a decremental effect of re-ward.

Those who argue that rewards de-crease intrinsic motivation are criticalof the single-subject designs. For ex-ample, Deci et al. (1999) state that onecan conclude very little from the sin-gle-subject designs because there aretoo few participants and none of thestudies had control groups. Instead,

claims about negative effects of rewardare based on results from the group-design studies. As noted, however, thefindings from such studies have notbeen entirely clear-cut. Although mostresearchers have found that verbal re-wards do not decrease measures of in-trinsic motivation, the results with tan-gible rewards have been more contra-dictory. To understand such diverse ef-fects, Cameron and Pierce (1994),Eisenberger and Cameron (1996), andDeci et al. used the methodology ofmeta-analysis to assess the group-de-

sign experiments and to determinewhen and under what conditions re-wards have detrimental effects on taskperformance and interest. Despite theseeming objectivity of this technique,these meta-analytic reviews reachedmarkedly different conclusions. Cam-


5/44

5THE MYTH CONTINUES

eron and Pierce and Eisenberger andCameron reported minimal negative ef-fects of tangible reward, whereas Deciet al. found tangible rewards to be det-rimental under a wide range of condi-

tions.Although the usefulness of meta-analysis and statistical testing in gen-eral has been questioned by behavioralr es ear ch er s ( e. g. , see Ba ro n &Derenne, 2000; Derenne & Baron,1999), research summaries based onmeta-analyses have become valuedsources of information for both policymakers and researchers. Deci et al.s(1999) meta-analytic finding of generalnegative effects of reward has impor-tant implications. Thus, to understandwhy the meta-analyses by Cameron

and Pierce (1994) and Deci et al. re-sulted in different findings, it is impor-tant to be familiar with the techniqueand logic of meta-analysis. The meta-analytic procedures described beloware based on Hedges and Olkin (1985);these were the basic procedures usedby Cameron and Pierce and by Deci etal.

THE TECHNIQUE AND LOGICOF META-ANALYSIS

Meta-analysis is a technique for

combining the results of a large num-ber of studies on the same topic. It in-volves combining data from concep-tually related studies to reach general-izations based on statistical criteria.Quantitative analyses, similar to meta-analysis, have been conducted on sin-gle-subject designs (e.g., see Kollins,Newland, & Critchfield, 1997); how-ever, meta-analysis is typically usedwith between-groups designs in whicha treatment group (e.g., a rewardedgroup) is compared to a control group(nonrewarded group) on a common de-

pendent measure (intrinsic motivation).The goals of a meta-analysis are to es-tablish the relation between indepen-dent and dependent variables (in thiscase, the relation between rewards andintrinsic motivation) and to determinewhat factors moderate or alter the mag-

nitude of the relation (e.g., type of re-ward, reward contingency). Conduct-ing a meta-analysis entails specifyingthe criteria for including and excludingstudies, collecting all experiments that

meet the criteria, and coding the stud-ies.Once all relevant studies are identi-

fied, the statistical results of each studyare transformed into a measure calledan effect size. An effect size is foundby converting the findings from eachstudy into a standard deviation unit. Inthe rewards and intrinsic motivationliterature, an effect size indicates theextent to which the experimental group(rewarded group) and the control group(nonrewarded group) differ in themeans on measures of intrinsic moti-

vation (e.g., free time on task after re-wards are removed, task interest). In itssimplest form, the effect size (g) is thedifference between the means of the re-warded group and the nonrewardedcontrol group divided by the pooledstandard deviation of this difference. Ina meta-analysis, the effect size fromeach study, rather than the individualparticipants within a study, becomesthe unit of analysis. If the effect sizesfrom all the studies present a randompattern, they will hover around zero,indicating no evidence for an effect.

On the other hand, the effect sizes maycluster in a positive or negative direc-tion, indicating that something is goingon.

One problem in meta-analysis ariseswhen studies do not provide enoughinformation to calculate effect sizes.When means and standard deviationsare not available, effect sizes can becalculated from ttests, Fstatistics, and

p values (see Hedges & Becker, 1986).However, in some cases, there may stillbe insufficient information to obtain aneffect size. The meta-analyst can con-

tact the researchers and try to obtainthe missing data. When the data cannotbe procured, the study can be excludedfrom the analyses or assigned an effectsize of 0.00 (indicating no differencebetween experimental and controlgroups). It has been argued that includ-


6/44


ing zero effect sizes is a conservativestrategy; if a significant effect is de-tected in spite of the inclusion of zeros,the contention is that the results wouldnot be altered if missing data were

available (for a discussion of this issue,see Light & Pillemer, 1984). On theother hand, if ones bias is toward noeffect (i.e., we are satisfied if the treat-ment is not harmful), including zerosfavors this conclusion. One strategy fordealing with this issue is to conduct theanalyses with zeros included and ex-cluded.

After effect sizes (g) are calculatedfor each relevant study, an overallmean effect size (d) is obtained.First, g is converted to dby correctingfor bias (g is an overestimation of the

population effect size, particularly forsmall samples; see Hedges, 1981). Theoverall mean effect size is obtained byweighting each effect size by the recip-rocal of its variance and averaging theweightedd. This procedure gives moreweight to effect sizes that are more re-liably estimated. The calculation ofmean effect sizes provides a signifi-cance test (whether the value differssignificantly from 0.00) and a 95%confidence interval (CI) (when the CIcontains 0.00, the results suggest thatthere is no evidence of a statistically

significant effect).In a hierarchical meta-analysis, all

studies are included in an overall anal-ysis. The researcher then searches formoderator variables. The studies arebroken out by one key moderator, thenanother, and so on. The moderators thatthe researcher chooses to examine maybe based on theoretical considerationsor on differences between the studies(e.g., different procedures used in thestudies, different characteristics of thesamples used, year of publication,etc.).

Hedges and Olkin (1985) recom-mend using homogeneity tests to as-certain whether a moderator analysis isnecessary. Essentially, the procedure isto use a chi-square statistic, Q, with K 1 degrees of freedom, whereKis thenumber of effect sizes. The null hy-

pothesis is that the effect sizes are ho-mogeneous (i.e., effect sizes in a givenanalysis are viewed as values sampledfrom a single population; variation ineffect sizes among studies is merely

due to sampling variation). When Q isstatistically significant, the implicationis that moderator analyses should beconducted. The original set of studiesis then broken into subsets until thechi-square statistics within the sub-groups are nonsignificant. When theresearcher has exhausted potentialmoderators and homogeneity is stillnot obtained, outliers (studies with ex-treme effect-size values) are examinedindependently and the analysis is con-ducted with outliers removed.

DIFFERENCES BETWEENCAMERON AND PIERCES (1994)

AND DECI ET AL.S (1999)META-ANALYSES

Although Deci et al. (1999) andCameron and Pierce (1994) used thesame meta-analytic procedures to eval-uate the research on rewards and in-trinsic motivation, their results dif-fered. Cameron and Pierce conducted ahierarchical meta-analysis of the re-wards and intrinsic motivation litera-ture. Studies were included if they had

a rewarded group and a nonrewardedcontrol group and if they used one ofthe two main measures of intrinsic mo-tivation (free time on the task after thereward was removed or self-reportedtask interest). The effects of reward onthe two dependent measures (free timeand task interest) were assessed sepa-rately. When a study did not provideenough information to calculate an ef-fect size, it was not included in theanalyses.

Cameron and Pierce (1994) werefirst interested in whether rewards,

overall, produced negative effects onmeasures of intrinsic motivation. Theirfindings indicated no overall negativeeffects on either measure of intrinsicmotivation. However, the set of effectsizes was significantly heterogeneous;thus, the researchers conducted a num-


7/44

7THE MYTH CONTINUES

ber of moderator analyses to determinewhen and under what conditions re-wards produced negative effects. Re-wards were broken down by rewardtype (tangible and verbal). Tangible re-

wards were subdivided into expectedand unexpected, and expected tangiblerewards were further separated by thereward contingency. Cameron andPierce used a behavioral framework tocategorize rewards by reward contin-gency; in addition, they used the cate-gories suggested by Deci and Ryans(1985) cognitive evaluation theoryframework. Their results indicatednegative effects on the free-time mea-sure only when the rewards were tan-gible, expected, and not contingent onmeeting a performance standard. The

same findings were reported by Eisen-berger and Cameron (1996), who car-ried out some additional analyses of re-ward contingencies.

Deci et al. (1999) suggested thatCameron and Pierces (1994) and Ei-senberger and Camerons (1996) fail-ure to detect more pervasive negativeeffects was due to methodological in-adequacies. Specifically, they criticizedCameron and Pierce and Eisenbergerand Cameron for the following: (a) col-lapsing across tasks with high and lowinitial interest and omitting a modera-

tor analysis of initial task interest, (b)including a study that used an inappro-priate control group (Boal & Cum-mings, 1981), (c) omitting studies ordata as outliers rather than attemptingto isolate moderators, (d) omittingstudies that were published during theperiod covered by their meta-analysis,(e) omitting unpublished doctoral dis-sertations, and (f) misclassifying stud-ies into reward contingencies as de-fined by cognitive evaluation theory.

To rectify these issues in their recentmeta-analysis, Deci et al. (1999) ex-

cluded the study by Boal and Cum-mings (1981), included studies thatwere missed in the previous meta-anal-yses, and included unpublished doctor-al dissertations. In addition, in contrastto Cameron and Pierce (1994), forstudies with insufficient information to

calculate effect sizes, Deci et al. im-puted effect sizes of 0.00 and includedthese in each of their analyses.

In terms of initial task interest, Deciet al. (1999) noted that the field of

inquiry has always been defined interms of intrinsic motivation for inter-esting tasks and the undermining phe-nomenon has always been specified asapplying only to interesting tasks in-sofar as with boring tasks there is littleor no intrinsic motivation to under-mine (p. 633). Given that cognitiveevaluation theory has little to say aboutthe effects of rewards on low-interesttasks, Deci et al.s meta-analysis fo-cused on reward effects on high-inter-est tasks. Studies or conditions withinstudies were included only if the tasks

used were measured or defined to beinitially interesting; studies or condi-tions within studies were excluded ifthe tasks used were measured or de-fined as initially uninteresting.

Thus, Deci et al.s (1999) meta-anal-ysis began with the overall effects ofrewards on intrinsic motivation fortasks of initial high interest only. Deciet al. analyzed the effects of reward onmeasures of self-reported task interestand free-choice intrinsic motivation.Their free-choice measure includedtime spent on a task after rewards were

removed. When a time measure wasnot reported in a study, Deci et al. usedmeasures of task persistence during thefree-choice period (e.g., number of tri-als initiated in a labyrinth game, num-ber of balls played in a pinball game,number of successes on a task). Hence,Deci et al.s analysis of the free-choicemeasure was broader than the analysisby Cameron and Pierce (1994), whoused only studies that assessed timemeasures.

On tasks of high initial interest, Deciet al. (1999) found a significant nega-

tive effect of rewards on the free-choice measure and a non-significanteffect on the self-report measure. Bothmean effect sizes were heterogeneous.To obtain homogeneity at each level ofanalysis, Deci et al. tested a number ofmoderator variables. When homoge-


8/44


neity could not be obtained, Deci et al.followed the procedure used by Cam-eron and Pierce (1994) and identifiedand removed outliers. First, Deci et al.tested whether verbal versus tangible

rewards were a moderator. Verbal re-wards were found to increase free-choice intrinsic motivation for collegestudents (a nonsignificant effect wasfound for children) and to enhance taskinterest for both children and collegestudents. Tangible rewards producednegative effects on both the free-choiceand self-report measures. In accordwith Cameron and Pierce, tangible re-wards were broken down into expectedand unexpected rewards. Unexpectedrewards had no significant effects; ex-pected tangible rewards were found to

significantly undermine both self-re-ported task interest and free-choice in-trinsic motivation.

Using cognitive evaluation theory astheir framework, Deci et al. (1999) fur-ther subdivided expected tangible re-wards into task-noncontingent, engage-ment-contingent, completion-contin-gent, and performance-contingent re-wards. Task-noncontingent rewardswere those given without specificallyrequiring the person to engage in theactivity (p. 636); engagement-contin-gent rewards were those offered to par-

ticipants for engaging in a task withouta requirement to complete the task, doit well, or reach some standard. Com-pletion-contingent rewards were thoseoffered and given for completing atask, and performance-contingent re-wards were defined as those offereddependent upon the participants levelof performance (p. 636). Deci et al.found no significant negative effectsfor task-noncontingent rewards; en-gagement-contingent rewards producedsignificant negative effects on bothfree-choice intrinsic motivation and

self-reported task interest. Completion-contingent and performance-contingentrewards also resulted in significantnegative effects on the free-choice in-trinsic motivation measure.

In addition, Deci et al. (1999) pro-vided a breakdown of performance-

contingent rewards into studies ofmaximum and not-maximum re-ward. In studies of maximum reward,participants were offered rewards grad-ed in terms of meeting a criterion or

performance standard; all met the cri-terion and received the full amount ofreward. Six studies were identified byDeci et al. as involving not-maximumreward. In these studies, some partici-pants failed to attain the criterion andwere given less than the maximum re-ward. Deci et al. reported that relativeto a nonrewarded control condition,participants receiving less than themaximum reward showed a large de-cline in free-choice intrinsic motiva-tion. In fact, the value (d 0.88) wasthe largest mean effect size in their en-

tire analysis.As a supplemental analysis, Deci etal. (1999) analyzed studies with chil-dren in which the free-choice assess-ment of high-interest activities wasconducted immediately following theremoval of reward, within a week, andafter a week. Deci et al. found negativeeffects at each time of assessment andsuggested that the undermining effectis not a transitory phenomenon. An ad-ditional analysis of the effects of re-wards on low-interest tasks was con-ducted by Deci et al.; no statistically

significant effects were detected.All in all, Deci et al.s (1999) meta-

analysis produced numerous negativeeffects of the various reward contin-gencies. Given the discrepancies be-tween Deci et al.s and Cameron andPierces (1994) findings, it is importantto examine carefully the proceduresused by Deci et al. The first noteworthydifference between the two meta-anal-yses occurs at the level of all rewards.Cameron and Pierce were interested inassessing the overall effects of rewardsacross all types of tasks. Deci et al. did

not conduct this analysis; instead, theyargued that the more theoretically rel-evant question concerned the effects ofrewards on tasks of high initial interest.

We contend that an analysis of theoverall effect of reward is central to anunderstanding of this complex area of


9/44

9THE MYTH CONTINUES

research. On a practical level, manyeducators, parents, and administratorshave taken the position that rewardsand incentive systems are harmful. Theview is that rewards negatively affect

students intrinsic interest across alltypes of activities (e.g., reading, math,science, computer games, etc.); no dis-tinction is made between low and highinitial levels of task interest. Writerswho caution against the use of rewardsand reinforcement frequently use ex-amples to illustrate their point. Moreoften than not, activities such as read-ing, lawn mowing, and mathematicsare cited as activities that people willlose interest in if they are given re-wards for performing the activity. Mostof these activities are not ones that in-

dividuals begin doing with high levelsof initial interest. Importantly, policymakers who adopt the view that re-wards are harmful rarely distinguishbetween high- and low-interest tasks.Because of this, an analysis of theoverall effects of reward is warranted.It is our contention that a more com-plete hierarchical breakdown of the ef-fects of rewards on intrinsic motivationshould begin at the level of all rewardsover all types of tasks. Following this,a breakdown of reward effects on high-and low-interest tasks would be appro-

priate.A further difficulty with Deci et al.s

(1999) meta-analysis concerns theirsupplemental analysis of reward effectson low-interest tasks. Several studiesthat used low-interest tasks were ex-cluded from their primary meta-analy-sis of high-interest tasks (e.g., Freed-man & Phillips, 1985; Overskeid &Svartdal, 1996). The problem is thatthese studies were not brought backinto their supplementary analysis oflow-interest tasks.

Another concern is that for some

studies in their analysis of high-interesttasks, Deci et al. (1999) omitted con-ditions that were relevant to their anal-yses. For example, in an experiment byWilson (1978), one group was offered$0.50 to engage in the target activity,a second group was offered $2.50 and

a control group performed the taskwithout the offer of reward. In Deci etal.s analyses, only one of the rewardedgroups was included. For other studiesthat used more than one level of re-

ward magnitude (e.g., Earn, 1982;McLoyd, 1979; Newman & Layton,1984), Deci et al. included all rewardconditions. Their omission of certainconditions within studies does not ap-pear to be systematic (e.g., rewardmagnitude was not examined by Deciet al. as a moderator), yet there are anumber of different types of cases inwhich this occurs. In addition, as didCameron and Pierce (1994), Deci et al.also missed a few experiments that mettheir inclusion criteria and that werepublished during the period covered by

their meta-analysis. Also, several stud-ies using high-interest tasks that mea-sured self-reported task interest wereeither excluded or inadvertently omit-ted from Deci et al.s analyses. Manyof these studies found positive effectson the self-report measure of task in-terest; Deci et al.s omission of theseeffects helps to explain why they foundeither negative effects or no effects onthe task-interest measure. A list ofstudies not included in Deci et al.sanalyses, dependent measures thatwere precluded, and a description of

conditions omitted by Deci et al. arepresented in Appendix A. Any com-putational differences in sample sizesand effect sizes are also outlined inAppendix A.

A final issue concerns the classifi-cation of studies into various rewardcontingencies. Deci et al. (1999) sug-gested that Cameron and Pierce (1994)miscategorized many experiments. Us-ing cognitive evaluation theory toguide their classification of studies,Deci et al. established the categories oftask-noncontingent, engagement-con-

tingent, completion-contingent, andperformance-contingent rewards. Al-though this categorization system maybe informative for cognitive evaluationtheory, the problem is that the catego-ries are too broad. Studies that usedvery different procedures were pooled


10/44


into overall categories of engagement-contingent, completion-contingent, andperformance-contingent rewards. Forexample, under performance-contin-gent reward, Deci et al. pooled exper-

iments in which participants were of-fered a reward for doing well, for eachproblem or unit solved, for obtaining acertain score, or for exceeding a norm.Eisenberger, Pierce, and Cameron(1999) examined some of these differ-ent reward procedures and found verydifferent effects on measures of intrin-sic motivation.

Rather than argue about which stud-ies belong in which category, we sug-gest that a more nuanced approach isto return to the methods section of theoriginal studies and code the reward

procedures actually employed in theexperiment. To the extent that one canobtain consistency in coding, such aprocedural categorization of rewardcontingencies would allow an assess-ment of the effects of the actual con-tingencies rather than those presumedto be an effect by any theoretical per-spective. The literature on rewards andintrinsic motivation is fraught withcompeting theories (e.g., cognitiveevaluation theory, the overjustificationhypothesis, social cognitive theory,

general interest theory, the competingresponse hypothesis, behavioral theo-ry). The problem with organizing stud-ies according to a particular theoreticalstance is that each theory could beused to organize the literature and, us-ing categories appropriate to the theo-ry, each theory could gain support. Us-ing a theoretical approach to guide theclassification of the reward proceduresdoes not provide us with a definitiveanswer about the effects of reward con-tingencies on measures of intrinsic mo-tivation. Instead, we propose that a

procedural description of reward con-tingencies not only allows us to assesswhere we stand in terms of the effectsof the actual reward contingencies butalso provides us with a test of alter-native accounts of the effects of re-wards on intrinsic motivation.

A NEW META-ANALYSIS:RESOLVING DIFFERENCES

BETWEEN PREVIOUSMETA-ANALYSES

To address criticisms and resolvediscrepancies among Cameron andPierces (1994), Eisenberger and Cam-erons (1996), and Deci et al.s (1999)meta-analyses, we provide a reanalysisof the effects of rewards on intrinsicmotivation. Our goal is to eliminateshortcomings of the prior meta-analy-ses (including ours) while building ontheir strengths. Our current meta-anal-ysis focuses on how reward affectsmeasures of free-choice intrinsic mo-tivation and self-reported task interest.In accord with Deci et al., free-choice

measures included free time on taskwhen the rewards were removed and,when time measures were not avail-able, performance during the free-choice period. As did Deci et al., wecombined performance and time mea-sures to make up the free-choice intrin-sic motivation index (we found no sig-nificant differences in the analyseswhen only time measures were ana-lyzed). Our analysis begins with an as-sessment of the overall effects of re-ward.

We then examine the effects of dif-

ferent moderator variables. To dealwith Deci et al.s (1999) criticism, ourfirst breakdown is in terms of high andlow initial task interest. Subdividingthe studies by high and low initial taskinterest allows us to directly compareour findings with those of Deci et al.In doing so, we favor cognitive eval-uation theory. On the other hand, fail-ure to find pervasive negative effectseven with high-interest tasks favors theconclusion that reward contingenciesdo not destroy interest. In other words,the strongest way to test Deci et al.s

claims is to use their requirement thattasks used in the studies must be bro-ken out by high and low initial interest.

At each level of our analysis, a ho-mogeneity statistic (Q) was calculated todetermine whether the set of effect sizescould be considered homogeneous.


11/44

11THE MYTH CONTINUES

When Q was significant, we proceededwith further moderator analyses. Re-wards were broken down by type (verbalor tangible), by expectancy (expected,unexpected), and by reward contingen-

cy. In addition, when there was enoughdata, we examined differences betweenstudies in which participants receivedmaximum or less than maximum re-wards. These breakdowns generally re-sulted in homogenous samples. Howev-er, in a few cases, homogeneity couldstill not be obtained even after a thor-ough examination of potential modera-tors. In these cases, we conducted theanalysis by removing outliers (as didDeci et al., 1999). Outliers were exam-ined in an attempt to explain their ex-treme values. At each level of our anal-

ysis, we report mean effect sizes, signif-icance tests, and 95% CI. However, weshould point out that making conclu-sions based on heterogeneous samplesmay be misleading. In a hierarchicalbreakdown, interpretations should focuson the homogeneous effects at the bot-tom level of the analysis (see Hunter &Schmidt, 1990).

Sample of Studies

Studies included in the present anal-ysis incorporated the databases of

Cameron and Pierce (1994) and Deciet al. (1999). In addition, in a searchof PsycINFO, we located a few newstudies and a few studies that were in-advertently missed in previous meta-analyses. The criteria for including astudy in a sample were as follows: Arewarded group was compared to anonrewarded group, the rewards weredistinguished as verbal (praise, positivefeedback) or tangible (e.g., money,candy, good-player awards), and intrin-sic motivation was measured as freechoice (time spent on the task follow-

ing the removal of reward or perfor-mance on the task during the free-choice period) or by self-reported mea-sures of task interest (task liking,enjoyment, satisfaction, or task prefer-ence). Two studies included in Cam-eron and Pierces research were omit-

ted; in one study (Boggiano & Hertel,1983), the dependent measure was as-sessed before all participants workedon the task; in the other study (Boal &Cummings, 1981), all participants (in-

cluding the control group) receivedmonetary payments. These studieswere also not included in Deci et al.sanalyses.

In addition to including publishedwork and in keeping with Deci et al.(1999), we included unpublished doc-toral dissertations. The resulting sam-ple consisted of 145 independent stud-ies (21 of the experiments were fromunpublished doctoral dissertations). Ofthese, 115 studies included a free-choice measure of intrinsic motivation;100 included a self-report measure of

task interest.

Classification and Coding of Studies

To code initial levels of task interest,we used the procedures described byDeci et al. (1999). If a measure of ini-tial task interest was reported in the ar-ticle, the study was classified as a low-interest task when the average on thatmeasure was below the midpoint of thescale for the activity and as a high-in-terest task when the average was abovethe midpoint. Two studies not included

in any of Deci et al.s analyses (Freed-man & Phillips, 1985; Phillips &Freedman, 1985) provided initial task-interest measures and included both ahigh- and a low-interest task. We in-cluded these studies in our analysis oflow- as well as high-interest tasks (seecomments in Appendix A). Studieswithout initial interest measures wereclassified as high or low depending onhow the researcher defined the task oron whether the task had been describedas interesting or uninteresting in priorexperiments.

Studies were also classified accord-ing to reward type (tangible or verbal),reward expectancy (expected or unex-pected), and reward contingency. Toclassify studies by reward contingency,we went back to the original studies,read the precise procedures used for re-


12/44


TABLE 1

Description of expected tangible reward contingencies

Reward contingency Description

Task noncontingent Reward is offered for agreeing to participate, for coming to thestudy, or for waiting for the experimenter.

Offer of reward is unrelated to engaging in the task.Rewards offered for doing well Reward is offered for doing well on the task or for doing a

good job.No specification is given as to what it means to do a good job

or to do well.Rewards offered for doing a

taskReward is offered to engage in the experimental activity.No instructions are given about how well participants must

perform or whether they must complete the task.Rewards offered for finishing

or completing a taskReward is offered to finish an activity, to complete a task, or

to get to a ceratin point on the task.The reward is not related to quality of performance.

Rewards offered for each unitsolved

Reward is offered for each unit, puzzle, problem, etc., that issolved.

Rewards offered for surpassinga score

Reward is offered for surpassing a particular specified score(absolute standard).

In some cases, the better the score, the higher the reward.Rewards offered for exceeding

a normReward is offered to meet or exceed the performance of others

on the task (relative standard).

ward delivery, and wrote down whatwas said to participants and how thereward was delivered. We then orga-nized the studies into seven main cat-egories of reward contingency: rewardsdelivered regardless of task involve-ment (task noncontingent); rewardsgiven for doing a task; rewards for do-ing well; rewards for finishing or com-

pleting a task; rewards given for eachproblem, puzzle, or unit solved; re-wards for achieving or surpassing aspecific score; and rewards for meetingor exceeding others. Although all stud-ies were coded for reward contingency,it was at the level of expected (offered)tangible reward that it became neces-sary to analyze studies in the variousreward contingencies. Other analysesresulted in homogeneity, and furtherbreakdowns were not required. In Ta-ble 1, we provide definitions and de-scriptions for each of these contingen-

cies at the level of expected tangiblereward. A comparison of our rewardcontingencies and those of Deci et al.(1999) is presented in Appendix B. Wereturn to a discussion of these compar-isons in our results section.

In some studies, there was not

enough information to code the contin-gency (e.g., Chung, 1995; Hom, 1987).In addition, a few studies used a con-tingency that did not fit into any of theseven categories; for example, W. E.Smith (1975) offered rewards to partic-ipants for showing signs of learning.These studies were included in overallanalyses, but were omitted from the

analysis of reward contingencies. A listof the studies used in each analysis, adescription of reward type, reward ex-pectancy, and reward contingency, to-gether with effect sizes are presentedin Appendixes C through G.

To ensure reliability of coding, thesecond author was given the definitionsfor each contingency (Table 1) and asample of 32 studies to code (each ofthe studies involved expected tangiblerewards). Reliability calculated as per-centage agreement was 97% (31 of 32studies). One study (L. W. Goldstein,

1977) included a condition in whichparticipants were offered a reward totake pictures. The issue was whetherthis contingency involved reward sim-ply for doing the task or for finishingthe task. The third author was broughtin to code the study; he pointed out


13/44


that participants in the reward condi-tion were not required to complete orfinish the task to obtain the reward andthat Goldstein stated that the rewarddid not imply that the subject had done

well on the task, only that s/he had en-gaged in it (p. 30). Hence, the rewardcontingency was classified as a rewardoffered for doing the task.

Finally, we identified studies that in-volved maximum or less than maxi-mum reward. Such studies involved of-fering participants a reward for doingwell, for finishing a task, for eachproblem or unit solved, for surpassinga score, or for exceeding a norm. Stud-ies were considered to produce themaximum reward if participants in thereward condition met the performance

requirements and received the full re-ward. Less than maximum reward oc-curred when there was a time limitsuch that some participants were un-able to meet all the requirements in thetime allotted and were given less thanthe full reward. For example, Decis(1971) experiment involved less thanmaximum reward. Participants wereoffered $1.00 for each of four puzzlessolved within a 13-min time limit. Notall participants were able to solve thepuzzles within the time limit and didnot receive the full reward.

Calculation and Analysis ofEffect Sizes

After all studies were coded, we cal-culated effect sizes (g) for each com-parison of a rewarded group to a non-rewarded group on the free-choice andself-report measures of intrinsic moti-vation. Positive effect sizes indicatethat rewards produced an increase inmeasures of intrinsic motivation rela-tive to a control group, negative effectsizes denote a decrease, and an effect

size of 0.00 indicates no difference.When there was not enough informa-tion to calculate an effect size, we at-tempted to contact the researchers.From a list of 22 researchers, we wereable to locate E-mail addresses fornine. E-mail messages were sent re-

questing the missing data. Eight peoplereplied; six could not locate the data,and two provided us with data for stud-ies by Wicker, Brown, Wiehe, andShim (1990) and by Dollinger and

Thelen (1978). When we could not ob-tain missing data, we imputed an effectsize of 0.00. Each analysis was con-ducted with zeros included and exclud-ed. In accord with Deci et al. (1999),we report the analyses with the zerosincluded; however, when mean effectsizes were altered to any extent by theinclusion of zeros, we report the anal-ysis with and without zeros.

Eisenberger, Pierce, and Cameron(1999) pointed out that there were twopossible types of control comparisonsfor some of the studies labeled perfor-

mance contingent by Deci et al. (1999).In some studies, the control group wastold the performance objectives and wasgiven performance feedback (completecontrol); in others, the control group wasnot told a performance objective and nofeedback was given (partial control). Ei-senberger, Pierce, and Cameron exam-ined differences between these two typesof comparisons (reward vs. partial con-trol, reward vs. complete control). Onesmall difference was detected on thefree-choice measure. When rewardswere offered to exceed others, reward

versus a partial-control condition result-ed in a nonsignificant positive effect; themean effect for reward versus a com-plete control was significantly positive(no other comparisons resulted in differ-ences). Because this difference wassmall and both mean effects were in thesame direction, we included studies witheither type of control condition in thepresent analyses. If a study containedboth types of controls (e.g., Harackiew-icz, Manderlink, & Sansone, 1984), oneeffect size was calculated comparing thereward condition to both controls.

In accord with Deci et al. (1999) andwith our previous procedures, morethan one effect size was calculated forseveral studies in our analyses. For ex-ample, if a single study assessed freechoice and used two types of expectedtangible rewards (e.g., rewards offered


14/44


TABLE 2

Hierarchical analysis of the effects of rewards on measures ofintrinsic motivation

Analysis of the effects of reward K N d 95% CI

Free-choice intrinsic motivation

All rewarda

Low initial task interestHigh initial task interesta

Verbal rewardTangible rewarda

Unexpected rewardExpected reward (offered)a

11512

11425

1029

101

8,176429

7,8881,3746,942

3756,703

0.080.28*

0.09*0.31*

0.17*0.02

0.18*

0.12, 0.020.07, 0.47

0.14, 0.040.20, 0.41

0.22, 0.120.18, 0.220.23, 0.13

Self-reported task interest

All rewarda

Low initial task interestHigh initial task interesta

Verbal rewarda

Tangible rewarda

Unexpected reward

Expected reward (offered)a

10011982483

5

81

8,028503

7,5471,5846,354

299

6,138

0.12*0.120.12*0.32*0.08*0.03

0.08*

0.07, 0.160.06, 0.30

0.07, 0.170.22, 0.430.03, 0.13

0.20, 0.26

0.03, 0.13

Note. K number of studies; N total sample size; d mean weighted effect size; 95% CI 95% confidence interval.

a The sample of effect sizes was significantly heterogeneous.*p .05.

for doing the task and rewards offeredfor surpassing a certain score) plus acontrol group, two effect sizes werecalculated. Each individual effect sizewas entered into the relevant analysis(expected tangible rewards for doing atask, expected tangible rewards for sur-

passing a score). For the analyses ofexpected tangible reward, tangible re-ward, and all reward, one effect sizewas calculated (the two groups werecompared to the control group) and en-tered into the overall analyses. Thisstrategy satisfies the independence as-sumption of meta-analytic statistics(Hedges & Olkin, 1985) and givesequal weight to each study analyzed.Thus, subcategories (e.g., rewards of-fered for doing the task, for doing well,etc.) may contain more effect sizesthan the superordinate category (ex-

pected tangible reward). For example,for all reward on the free-choice mea-sure (over both high- and low-interesttasks), there were 126 effect sizes, butonly 115 of these are independent (sev-eral are within the same study).

After all effect sizes were calculated,

the present analyses were run on thecomputer program Meta (Schwarzer,1991) using the weighted integrationmethod described in our section onmeta-analytic procedures. The programconverts effect size, g, to d; meanweighted effect size (d) is obtained;

95% CI is constructed around themeans, and a homogeneity statistic,Q,is computed.

RESULTS OF OURMETA-ANALYSIS

In Table 2, we present the results forour meta-analysis up to the level of re-ward contingency. Table 2 presentsmean weighted effect size (d) and 95%CI for each analysis. Mean effects areconsidered statistically significant whenthe CI does not include zero. In the pres-

ent meta-analysis, positive effect sizesindicate that reward produces increasesin intrinsic motivation, negative effectsizes support the claim that rewardsundermine intrinsic motivation, andzero effects indicate no evidence for aneffect of reward. According to J. Cohen


15/44


(1988), an effect size of 0.20 is con-sidered small, 0.50 is moderate, andgreater than 0.80 is large.

All Rewards

First, the overall effects of rewardwere analyzed across all conditionsand across high- and low-interest tasks.On the free-choice measure, Table 2 in-dicates that there was no significant ef-fect (d 0.08, CI 0.12, 0.02).On the measure of self-reported taskinterest, a small significant positive ef-fect was detected (d 0.12, CI 0.07, 0.16). This analysis was not con-ducted by Deci et al. (1999); therefore,the findings cannot be compared. Theresults are, however, in accord withthose of Cameron and Pierce (1994).On both the free-choice and self-reportmeasures, however, the sets of studieswere significantly heterogeneous, sug-gesting the necessity of a moderatoranalysis. Thus, at the next level ofanalysis, we divided studies into thosewith low- and high-interest tasks.

The Effects of Rewards onLow-Interest Tasks

When reward effects were analyzedfor tasks with low initial interest, Table2 shows a statistically significant pos-

itive effect on the free-choice measure(d 0.28, CI 0.07, 0.47); therewas no significant effect on self-re-ported task interest (d 0.12, CI 0.06, 0.30). These findings indicatethat when a task is not initially inter-esting, rewards enhance free-choice in-trinsic motivation but not verbal ex-pressions of task interest.

Although the studies in this analysiswere considered homogeneous (i.e., Qwas not significant), we examinedwhether there were any differencesamong different types of rewards, ex-

pectancies, and contingencies. On thefree-choice measure, only one study in-cluded a condition that used a verbalreward (the effect was positive). Fortangible reward, one study included anunexpected reward condition (the ef-fect was positive). All of the 12 studies

with low-interest tasks included an ex-pected tangible reward condition; com-pared with a nonreward control, themean effect was significantly positive(d 0.26, CI 0.06, 0.45). Nine

studies involved offering the rewardfor doing the task; on the free-choicemeasure the effect remained significant(d 0.26, CI 0.03, 0.48). For self-reported task interest, no significant ef-fects were found under any of the con-ditions.

In Deci et al.s (1999) supplementalanalysis of low-interest tasks (p. 651),fewer studies were included and nosignificant effects were found on eitherthe free-choice or the self-report mea-sures of intrinsic motivation.

The Effects of Rewards onHigh-Interest Tasks

For high-interest tasks, the mean ef-fect size on free choice (Table 2)showed a small but significant negativeeffect (d 0.09, CI 0.14,0.04); the set of effect sizes, how-ever, was heterogeneous. The mean ef-fect size for self-reported task interestwas significant, small, but in a positivedirection (d 0.12, CI 0.07, 0.17);the sample of effect sizes was also het-erogeneous. Deci et al. (1999) also re-ported a significant negative effect onthe free-choice measure but a nonsig-nificant effect on the task-interest mea-sure. As noted, Deci et al. omitted ormissed several self-report effect sizes.

Verbal Rewards

Verbal rewards were found to sig-nificantly enhance both free-choice in-trinsic motivation (d 0.31, CI 0.20, 0.41) and self-reported task inter-est (d 0.32, CI 0.22, 0.43).These results were also obtained byDeci et al. (1999), who reported similar

small to moderate positive effects ofverbal rewards.

On the free-choice measure, the setof effect sizes was homogeneous, sug-gesting that no further breakdownswere necessary. In most studies of ver-bal reward, the rewards were unex-


16/44


pected and the mean effect was posi-tive; a positive effect was also foundwith the five studies that used expectedrewards. In addition, verbal rewardswere generally delivered simply for

doing a task and were not contingenton any specific level of performance(again, the effects were positive).When the effects of verbal reward onfree choice were examined with chil-dren versus adults (mainly college stu-dents), children showed a smaller pos-itive effect (K 10, N 320, d 0.22, CI 0.04, 0.39) than adults (K 15,N 844,d 0.36, CI 0.22,0.49). Deci et al. (1999) also reporteda larger effect for adults but a nonsig-nificant effect for children (our effectsize for children was statistically sig-

nificant because we included morestudies than Deci et al.).On the task-interest measure, the set

of effect sizes for verbal reward wassignificantly heterogeneous. We con-ducted moderator analyses of childrenversus adults and expected versus un-expected reward. Mean effect sizes foreach of these analyses remained sig-nificantly positive, but homogeneitywas still not obtained. In almost allstudies, the rewards were given for do-ing the task; hence, this reward contin-gency could not be a moderator.

To obtain homogeneity, three studieswere removed from the analysis (thesame outliers were removed by Deci etal., 1999). Inspection of the outliers in-dicated that two of the studies (Butler,1987; Vallerand, 1983) produced largepositive effects; these studies did notdiffer in obvious ways from other stud-ies in the sample except for their ten-dency to generate extreme values of ef-fect size. The third outlier (Kast &Connor, 1988) produced a negative ef-fect (0.46). Kast and Connor com-pared control participants to partici-

pants who were praised for their per-formance on the task as well as to an-other group who were also praised butwho were told that they should be do-ing well. The second verbal rewardcondition produced a negative effectand was different from verbal reward

used in other studies; Deci et al. termedthis controlling reward. When theoutliers were removed from the anal-ysis of verbal rewards on the task-in-terest measure, the set of studies was

homogeneous and the mean effect re-mained significantly positive (K 21,N 1,194, d 0.32, CI 0.21,0.44). In this data set, there were sixstudies that did not provide enough in-formation to obtain an estimate of ef-fect size (these studies were given aneffect size of 0.00). When these studieswere removed, the mean effect size fortask interest showed a slight increase(K 15, N 981, d 0.40, CI 0.27, 0.53).

Tangible Rewards

When the effects of tangible rewardson high-interest tasks were analyzed,Table 2 shows a small significant neg-ative effect on the free-choice measure(d 0.17, CI 0.22, 0.12) anda small significant positive effect onself-reported task interest (d 0.08,CI 0.03, 0.13). Both of these sam-ples of effect sizes were significantlyheterogeneous and required a furthermoderator analysis.

Reward exp ect anc y. Tangible re-wards were subdivided into unexpect-

ed (rewards delivered without a state-ment of the contingency) and expected(rewards delivered after a statement ofcontingency) categories. No significanteffects were detected for unexpectedtangible rewards (see Table 2), and thesamples were homogenous (Deci et al.,1999, reported similar findings). Ex-pected tangible rewards produced anegative effect on the free-choice mea-sure (d 0.18, CI 0.23,0.13) and a positive effect on theself-report measure (d 0.08, CI 0.03, 0.13), but both of these samples

were significantly heterogeneous.Reward contingency. For the next

level of analysis, expected tangible re-wards were subdivided into various re-ward contingencies. Results of ouranalysis on the free-choice measure arepresented in Figure 1. No significant


17/44


Figure 1. The effects of expected tangible reward contingencies on free-choice intrinsic motivationunder high levels of initial task interest. K number of studies, N total sample size, d meanweighted effect size; statistically reliable effect sizes are marked with an asterisk (*p .05, **p .01). Positive effect sizes indicate higher intrinsic motivation for rewarded versus control groups;negative effect sizes indicate lower intrinsic motivation for rewarded groups. Numbers in parenthe-ses represent 95% confidence intervals. All effect sizes are based on homogeneous samples.

effects were detected when the rewardswere task noncontingent, were offeredfor finishing or completing a task, or

were offered for attaining or surpassinga score. Figure 1 shows significantnegative effects when the rewards wereoffered for doing a task, for doing wellon a task, and for each unit solved. Asignificant positive effect was foundwhen the rewards were offered formeeting or exceeding the performancelevel of others.

When rewards were offered for do-ing a task, the effect was significantlynegative (K 57, N 2,910, d 0.35, CI 0.43, 0.27) but not ho-mogeneous. Although we searched for

moderators (salient vs. nonsalient re-ward, children vs. adults, and time ofreward delivery), analyses of these var-iables did not result in homogeneoussamples. As a result, outliers wereidentified and omitted. The mean effectwith outliers removed is presented in

Figure 1. Two of the outliers producedpositive effects; the only differencesbetween these two studies and the bulk

of studies were that the study by Tri-pathi and Agarwal (1988) was con-ducted in India and the study by Bren-nan and Glover (1980) was designed toassess the effects of rewards when therewards were shown to function as re-inforcement. Other outliers (Chung,1995; Danner & Lonkey, 1981; Fabes,Eisenberg, Fultz, & Miller, 1988; Mor-gan, 1983, Experiment 1; Okano, 1981,Experiment 2) had large negative ef-fects but there was no common factorthat could explain their extreme values.

Our findings for free choice indicate

that when reward contingency is de-fined in terms of experimental proce-dures, negative, neutral, and positiveeffects are obtained. Using cognitiveevaluation theory as their frameworkfor the categorization of reward contin-gencies, Deci et al. (1999) found neg-


18/44


Figure 2. A comparison of our findings with Deci et al.s (1999) effects of expected tangiblereward contingencies on free-choice intrinsic motivation for high-interest tasks. Deci et al.s cate-gories of completion-contingent and performance-contingent reward contained studies that involvedreward offered for each unit solved.

ative effects for all but task-noncontin-gent rewards. One way to understand

these differences is to compare Deci etal.s effects and definitions of contin-gencies with our effects and proceduraldefinitions. Figure 2 shows this com-parison and indicates that Deci et al.scompletion-contingent and perfor-mance-contingent rewards consisted of

a variety of reward procedures havingdifferent effects.

Our results for the task-interest dataare presented in Figure 3. The analysisshows no significant effect for task-noncontingent rewards, a small signif-icant negative effect for rewards of-fered for doing, and significant positiveeffects for each of the other contingen-


19/44


Figure 3. The effects of expected tangible reward contingencies on self-reports of task interestunder high levels of initial task interest. K number of studies, N total sample size, d meanweighted effect size; statistically reliable effect sizes are marked with an asterisk (*p .05, **p .01). Positive effect sizes indicate higher intrinsic motivation for rewarded versus control groups;negative effect sizes indicate lower intrinsic motivation for rewarded groups. Numbers in parenthe-ses represent 95% confidence intervals. All effect sizes are based on homogeneous samples.

cies. In the analysis of rewards offeredfor doing, 14 studies were given effectsizes of 0.00; when these studies are

removed from the analysis, the nega-tive effect increased from 0.13 to0.22 (K 24, N 1,201, d 0.22, CI 0.33, 0.10).

In the analysis of rewards offered foreach unit completed, when all studieswere included the effect was positive (K 22,N 1,161,d 0.19, CI 0.08,0.31) but significantly heterogeneous.Two studies (Kruglanski et al., 1975,Experiment 1; Wimperis & Farr, 1979)had a large positive effect size; whenthese studies were omitted, homogeneitywas attained (Figure 3 presents the data

for homogenous samples).In Figure 4, we compare Deci et al.s

(1999) findings and reward contingen-cies with ours. For completion-contin-gent and performance-contingent re-wards, Deci et al. found no significanteffects, whereas our findings show a

number of positive effects for studiesthat would be included in these cate-gories. As discussed previously, many

studies with self-report measures werenot included in Deci et al.s analyses(see details in Appendix A).

Maximum versus less than maximumreward.On the free-choice measure ofintrinsic motivation, there was onlyone reward contingency (rewards of-fered per unit solved) that allowed acomparison between maximum andless than maximum reward. For otherreward contingencies, most studies in-volved maximum reward; a compari-son with less than maximum rewardwould be unreliable. When rewards

were offered for each unit solved, thefindings showed nonsignificant effectsfor studies of maximum rewards (K 6,N 345,d 0.03, CI 0.25,0.18) and a significant negative effectfor studies of less than maximum re-ward (K 14,N 749,d 0.22,


20/44


Figure 4. A comparison of our findings with Deci et al.s (1999) effects of expected tangiblereward contingencies on self-reports of task interest for high-interest tasks. Deci et al.s categoriesof completion-contingent and performance-contingent reward contained studies that involved re-ward offered for each unit solved.

CI 0.37, 0.07). These two sets ofeffect sizes were homogeneous. Theseresults suggest that the negative effect

of pay per unit is associated with par-ticipants receiving less than maximumrewards.

No analyses were conducted on dif-ferences between maximum and lessthan maximum rewards on the self-re-port measure. Most of the contingen-

cies had too few studies that used lessthan maximum reward. For studies in-volving the offer of reward for each

problem solved, there were too few ex-periments of maximum reward (seeAppendix F).

DISCUSSION

A major issue in psychology and ed-ucation is that rewards and reinforce-


21/44


ment have a detrimental effect on in-trinsic motivation. The concern is thatif people receive reinforcement or re-wards for activities they already enjoy,they will be less motivated to engage

in those activities than they were priorto the introduction of reward once therewards are no longer forthcoming. Inother words, rewards and reinforce-ment are said to decrease intrinsic mo-tivation. Since the 1970s, over 100studies have been conducted to assessthe effects of reward on intrinsic mo-tivation. The vast majority of studieson the topic have employed between-groups statistical designs. Rewardedparticipants are compared to nonre-warded controls. Intrinsic motivation ismeasured by the difference between

groups on task interest and free choice(time on task and performance on taskonce the rewards are removed). Ameta-analysis of this experimental lit-erature by Cameron and Pierce (1994)and Eisenberger and Cameron (1996)found limited negative effects of re-wards, whereas a more recent analysisby Deci et al. (1999) showed pervasivenegative effects. The meta-analysispresented in this article was designedto correct flaws in the previous reviewsand to resolve differences.

A Summary of Our Findings

In terms of the overall effects of re-ward, our meta-analysis indicates noevidence for detrimental effects of re-ward on measures of intrinsic motiva-tion. This finding is important becausemany researchers and writers espousethe view that rewards, in general, re-duce motivation and performance. Inaddition, many students of psychologyand education are taught that rewardsare harmful and that reward proceduresshould be avoided in applied settings.

Our finding of no overall effect of re-ward, however, must be treated withcaution. In our meta-analysis, the over-all reward category lacked homogene-ity, indicating the appropriateness of amoderator analysis. In other words, theoverall reward category is too inclu-

sive; rewards have different effects un-der different moderating conditions.

Figure 5 shows the effects of differ-ent moderating conditions. The effectsof rewards on free-choice intrinsic mo-

tivation and self-reported task interestare presented only for homogeneoussubsets. When a result was heteroge-neous, we broke down the subset of ef-fect sizes by different moderator vari-ables until homogeneity was attained.A positive effect indicates that rewardsenhanced the measure of intrinsic mo-tivation relative to a control condition,a negative effect indicates a decreasefor the rewarded group, and a zero ef-fect indicates no significant effect.

The effects of all rewards are firstbroken into high- and low-interest

tasks. When the tasks used in the stud-ies are of low initial interest, rewardsincrease free-choice intrinsic motiva-tion and leave task interest unaffected.This finding indicates that rewards canbe used to enhance time and perfor-mance on tasks that initially hold littleenjoyment. As Bandura (1986) recog-nized, Most of the things people en-

joy doing for their own sake had littleor no interest for them originally. . . .But with appropriate learning experi-ences, almost any activity . . . can beimbued with consuming significance

(p. 241). Our results suggest that re-ward procedures are one way to culti-vate interest in an activity. In educa-tion, a major goal is to instill motiva-tion and enjoyment of academic activ-ities. Many academic activities are notof high initial interest to students. Animplication of our finding is that re-wards can be used to increase perfor-mance on low-interest academic activ-ities.

For high-interest tasks, verbal re-wards are found to increase free choiceand task interest. This finding repli-

cates the results of Cameron and Pierce(1994) and Deci et al. (1999). Most so-cial interaction in business, education,and clinical settings involves the use ofverbal praise and positive feedbackfrom managers, teachers, and thera-pists. When praise and other forms of


22/44


Figure 5. A summary of the meta-analysis comparing free-choice intrinsic motivation and self-reported task interest. 0 no reliable effect; statistically significant negative effect of reward; statistically significant positive effect of reward.

positive feedback are given and laterremoved, our findings indicate that in-terest and performance increase.

The effects of tangible reward onmeasures of intrinsic motivation differ

by reward expectancy. When rewardsare delivered unexpectedly (without adescription of the reward contingency),there is no evidence of a significant ef-fect on either free choice or task inter-est. This suggests that it is not tangiblerewards per se that undermine moti-vation and interest; instead it dependson instruction and the statement ofcontingency.

For high-interest tasks, when the re-wards are tangible and expected (of-fered beforehand), there are differenteffects depending on the description of

the reward contingency. When the of-fer of reward is unrelated to task be-havior (task noncontingent), there is noevidence for an effect of reward on ei-ther free choice or task interest. On theother hand, when people are offered atangible reward for doing a task or for

doing well at a task, they often chooseto do the activity less in a free-choiceperiod. The negative effect of rewardsoffered for doing a task is also detectedfor the task-interest measure. We did

not find a negative effect on task inter-est when the rewards are offered fordoing well. One possibility is that thetrue effect is negative but, at this point,there are too few studies to yield a re-liable estimate. In general, when thedescription of the reward contingencyimplies that rewards are loosely tied toperformance, the evidence suggeststhat people show a small reduction inperformance and interest.

Figure 5 shows that rewards offeredfor finishing or completing a task havea nonsignificant effect on the free-

choice measure but a positive effect ontask interest. Again, there were fewstudies in this category, and a firm con-clusion about the effects is premature.Stronger conclusions can be drawn forthe analysis of rewards offered for eachunit solved. When participants are of-


23/44


fered a reward for each problem, puz-zle, or unit solved, our findings indi-cate a negative effect on free choiceand a positive effect on task interest. Asupplementary analysis involving less

than maximum reward and maximumreward shows that the negative effecton free choice occurs when partici-pants obtain less than the full reward.In studies of less than maximum re-ward, participants are given a time lim-it to solve problems. Thus, the negativeeffect may be a result of time pressurerather than reward. Another way to un-derstand this result is to consider whatless than maximum reward signifies toparticipants. If people are told they canobtain a certain level of reward but aregiven less than that level, they have re-

ceived feedback information that indi-cates failure. In other words, this typeof situation may represent failure feed-back, not reward. When participantsare not under time pressure and areable to obtain the maximum reward,there is no significant effect on thefree-choice measure.

When rewards are offered for meet-ing or surpassing a score, Figure 5shows no significant effect on freechoice but a significant positive effecton task interest. Rewards offered forattaining a criterion are tightly linked

to level of performance. In this situa-tion, the rewards are tied to challengeand mastery of the activity, and peopleexpress interest in the task (see Ban-dura, 1986). When rewards are givenfor exceeding the performance level ofothers, the results show a significantincrease on free choice and task inter-est. One possible explanation for thepositive effects of this type of rewardcontingency is that rewards signifycompetence, self-efficacy, or ability atthe task, and people enjoy doing activ-ities that reflect their competence.

Overall, our analysis shows that tan-gible rewards can be used to produceboth negative and positive effects onmeasures of intrinsic motivation. Pos-itive effects are obtained when the re-wards are explicitly tied to perfor-mance standards and to success; neg-

ative effects are produced when re-wards signify failure or are loosely tiedto behavior.

Durability of Reward Effects

Deci et al. (1999) have claimed thatnegative effects of rewards are nottemporary. In a supplemental analysis,Deci et al. examined studies of chil-dren in which the free-choice assess-ment was conducted within a week fol-lowing the removal of reward and aftera week. Their analysis showed nega-tive effects on free choice for eachtime of assessment. Deci et al. con-cluded that their results indicate quiteclearly that the phenomenon of extrin-sic rewards undermining intrinsic mo-

tivation is not merely transitory (p.650). An examination of the studies in-cluded in Deci et al.s supplementaryanalysis indicates that most of the ef-fect sizes were based on rewards of-fered for doing the task or for doingwell. When the free-choice assessmentwas conducted within a week follow-ing the removal of reward, 10 of 12studies involved rewards offered fordoing well or for doing the task. Of the14 studies with assessments conductedmore than a week later, 13 were con-cerned with rewards offered for doing

well. Our interpretation of Deci et al.sfindings is that it is rewards offered fordoing (or doing well) that continue toproduce a negative effect on freechoice, not extrinsic rewards in gener-al. According to Bandura (1986), thiskind of reward procedure imparts littleindication of competence, in that therewards are allocated without regard toquality of performance and are thusloosely tied to behavior.

An unresolved issue is whether thereis a change in free-choice intrinsic mo-tivation over time. We examined seven

between-groups design studies of re-wards offered for doing the task thatassessed whether negative effects weremaintained over time (Chung, 1995;Loveland & Olley, 1979; Morgan,1983, Experiments 1 and 2; Ogilvie &Prior, 1982; Ross, 1975, Experiment 1;


24/44


Shiffman-Kauffman, 1990). Thesestudies included two measures of free-choice intrinsic motivation, one afterthe removal of reward and a second afew weeks later. Only two of the seven

studies showed a significant negativeeffect on the second measure (Morgan,1983, Experiments 1 and 2). These re-sults suggest that rewards offered fordoing a task have transitory effectswhen multiple measures of free-choicemotivation are used. This conclusion isstrengthened by examining the resultsfrom studies using repeated presenta-tions of reward followed by repeatedassessments of intrinsic motivation fol-lowing the removal of reward.

As previously indicated, some op-erant researchers tested the effects of

rewards on intrinsic motivation by ex-perimental designs in which the sameindividual was exposed to a baselineperiod, a reward intervention, and a re-turn to baseline (Davidson & Bucher,1978; Feingold & Mahoney, 1975; Ma-whinney et al., 1989; Skaggs et al.,1992; Vasta et al., 1978). Participantswere measured repeatedly during eachphase of the experiment, and rewardswere shown to increase measures ofperformance, indicating that the re-wards functioned as reinforcement.The results of these experiments

showed that participants spent as much(or more) time on the target activity inthe postreward phase as they did in theinitial baseline period. These findingsindicate that negative effects of rewarddo not persist when task performanceis rewarded on repeated occasions.

Magnitude and Impact ofReward Effects

It may be informative to considerhow serious the negative effects are onhigh-interest tasks when the rewards

are tangible, expected, and given fordoing a task or for doing well on atask. In all of the studies involvingthese contingencies, time spent on thetask during the free-choice period wasthe measure of free-choice intrinsicmotivation. Using the free-time mea-

sure, one could ask how much lesstime students would spend on high-in-terest tasks (e.g., art, music, reading,drama) if a teacher implemented a re-ward system for doing the task (or do-

ing well) and then removed it. Resultsfrom our meta-analysis indicate thatthe average effect size for a compari-son between students who receive thisreward procedure and nonrewarded in-dividuals on time on task is about0.30.

In the original experiments, freetime on task was typically measuredover an 8-min period. To convert theeffect size of 0.30 to real time, oneneeds to know the pooled standard de-viation of rewarded and nonrewardedgroups. Because many researchers re-

port only t or F statistics that cannotbe converted to the overall pooledstandard deviation, we are unable toprovide an estimate of this parameter.Instead, we will use a well-designedstudy by Pretty and Seligman (1984)that provides a pooled standard devia-tion. Pretty and Seligman conductedtwo experiments with large samplesand readily available statistical infor-mation. Both experiments compared acondition of tangible rewards offeredfor doing a high-interest task (Somapuzzles) with a nonrewarded control

group on 8 min of free time. Thepooled standard deviation was 2.6 min(Deci, 1971, also used Soma, the free-time measure was assessed over an 8-min period, and the pooled standarddeviation was 2.4 min).

Using 2.6 min as the estimate of er-ror, we are able to convert the negativeeffect size from the meta-analysis intoreal time. An effect size of 0.30would mean than in an 8-min period,the average individual who is offereda tangible reward for doing the task (ordoing well) will spend about 47 s less

on the task when the reward is with-drawn than the average nonrewardedindividual. Given this result, whatwould happen if a teacher implementedthis incentive procedure in a readingprogram (for children who already en-

joy reading) and then removed it? Ac-


25/44


cording to this estimate, students whoare offered gold stars for readingwould spend about 4 min less readingin a 40-min free-choice period thanstudents not given the incentive. If we

assume that students without rewardspend about 30 min reading in the 40-min free-choice period, then rewardedstudents would spend about 26 minreading (based on Deci et al.s, 1999,analysis of engagement-contingent re-ward, rewarded children would spendabout 25 min reading). A 4- to 5-minreduction in free-time reading could bebehaviorally important if cumulatedover many successive opportunities toread, but there are no studies that haveaddressed this issue.

A cautionary note is in order. Our

example of reading and reward de-pends on the use of a standard devia-tion from a single well-designed study.It also depends on the ability to extrap-olate from an 8-min experimental pe-riod to longer ones. It is possible thatthe negative effects, such as they are,are evident only for a short time at thebeginning of the free-choice period.That is, it may well be the case that ifan hour of free choice were given, re-sults might look very different. Thepoint is that this is a hypothetical ex-ample. Further evidence is required to

generalize the findings to experimentswith longer free-choice periods or toeveryday settings in which choice isdistributed over long periods of time.

Given the state of the literature, weconclude that the negative effect oftangible rewards offered for doing ahigh-interest task (or doing well) is sta-tistically significant, but the size of theeffect does not suggest a strong impact.Of course, our conclusion with regardto the magnitude of the negative effectsof reward contingencies applies equal-ly to positive effects. That is, although

the positive effects are statistically sig-nificant, they too are small.

A Comparison of Our Findings toThose of Deci et al. (1999)

Our pattern of findings for expectedtangible reward contingencies differs

from the results of Deci et al.s (1999)meta-analysis. Deci et al. present a pic-ture of pervasive negative effects. Thepicture depicted in our analysis is oneof circumscribed negative effects. As

noted, Deci et al. used reward contin-gencies that were theoretically rele-vant, but that were collapsed over dis-tinct reward procedures. For example,on free-choice intrinsic motivation,Deci et al. showed a negative effect forperformance-contingent rewards. Theperformance-contingent category in-cluded some studies of rewards offeredfor each unit solved, rewards offeredfor doing well, rewards offered for sur-passing a score, and rewards offeredfor exceeding others. By combiningthese distinct procedures, Deci et al.

obtained an overall negative effect forperformance-contingent reward. Weshow that these diverse reward proce-dures produce different effects on freechoice; hence, it is unwise to collapsethem into a single category of perfor-mance-contingent reward. Similarly,Deci et al. collapsed over reward cat-egories for the task-interest measure,and similar problems arise. In additionto collapsing over different reward cat-egories, Deci et al. omitted several pos-itive effects that, when included, re-sulted in positive findings for task in-

terest. Overall, our meta-analysis indi-cate s th at r ew ar ds d o n ot h av epervasive negative effects when minorimprovements to Deci et al.s catego-rization of reward contingencies aremade and all available studies are in-cluded.

Using cognitive evaluation theory toguide the classification of studies, Deciet al. (1999) obtained negative effectsof tangible reward contingencies. Weshowed that by classifying studies ac-cording to the actual contingency used,different effects were obtained. That

the results of a meta-analysis can bedrastically altered by assigning studiesto categories based on a particular the-oretical orientation points to some im-portant issues and limitations in this lit-erature. The difference between ourfindings and those of Deci et al. points

8/10/2019 rewards & intrinsic mo

rewards & intrinsic motivation.pdf

Documents