Contribution
The most extensive meta-analysis of the reward literature to its date, and the first to include unpublished dissertations (20 studies), to separate engagement- from completion-contingent rewards, to test age as a moderator, and to test whether undermining lasts. It is also an adversarial reanalysis: it sets out, study by study, where its coding departs from Cameron and Pierce’s (1994) and why.
Key insight
Whether a reward undermines interest depends on what the reward is likely to mean to the person: a controller of their behaviour or an affirmation of their competence (CET’s “functional significance”, p. 628). Tangible rewards given for doing a task are mostly controllers; positive feedback is mostly information. So the reward literature is not “do rewards work?” but a set of predicted interactions, and averaging across them hides the effect. The paper’s methodological claim follows: meta-analysis “is inherently conservative” in a theory-driven field, because published studies were designed to find the moderators that a meta-analysis then averages over (p. 631).
Assumptions
- Only interesting tasks are in scope. Tasks below the scale midpoint of rated interest, or defined as uninteresting, are excluded from the primary analyses (p. 635). For dull tasks “the critical theoretical issue is how to facilitate internalization” (p. 635). Cameron and Pierce collapsed interesting and dull tasks (p. 651), and their 2001 reply (LIT-562, read since in NOTE-587) reports reward effects on low-interest tasks as well, raising free choice there, so the scope is part of the dispute.
- Measures after the reward stops. Intrinsic motivation is free-choice persistence or self-reported interest measured once the contingency is over; measures taken during the reward phase mix motives (p. 635).
- Missing effect sizes imputed as zero: 8 of 150 free-choice and 26 of 121 self-report effects (p. 637).
- Hedges–Olkin weighting, homogeneity by Q, moderators before outliers; fewer than 4% of effects dropped as outliers (p. 634).
- Free choice is the more valid measure (pp. 655–656): unobtrusive, less open to demand, and its main bias would shrink, not inflate, undermining.
Key results
- All rewards: free choice d = −0.24 (k = 101); self-report 0.04, n.s. (k = 84) (p. 638). CET makes no prediction at this level.
- Verbal rewards: free choice 0.33 (k = 21); college 0.43, children 0.11 n.s. Self-report 0.31 after three outliers (pp. 638–639).
- Tangible rewards: free choice −0.34 (k = 92); self-report −0.07 (k = 70) (p. 639). Children worse than college students on both measures (pp. 646, 656).
- Unexpected tangible: 0.01 and 0.05, n.s. (p. 640).
- Expected tangible: −0.36 and −0.07 (p. 640).
- Task-noncontingent: −0.14 and 0.21, n.s. (p. 640).
- Engagement-contingent: −0.40 (children −0.43 after outliers, college −0.21) and −0.15 (p. 641).
- Completion-contingent: −0.44 after one outlier, and −0.17 after two (pp. 641–643).
- Performance-contingent: free choice −0.28, self-report −0.01 n.s. Split by control group and feedback: maximum reward vs no-feedback control −0.15; less-than-maximum vs no-feedback control −0.88; reward signalling good performance vs positive-feedback control −0.20; vs negative-feedback control −0.03 n.s. (pp. 643–644).
- Published vs dissertations: no difference for verbal or tangible rewards (p. 649).
- Persistence over time (children): −0.40, −0.49, −0.53 (p. 650).
- Interesting vs dull tasks (13 studies): free choice −0.68 vs 0.18 n.s.; self-report −0.37 vs 0.10 n.s. (p. 651). Adding the dull-task conditions back weakens composites slightly and removes significance only for completion-contingent self-reports (pp. 651–652).
- Informational vs controlling feedback (4 studies): controlling positive feedback lowers intrinsic motivation relative to informational, d = −0.78; informational vs neutral 0.66, controlling vs neutral −0.44 (p. 652).
- The behavioural account fails (pp. 654–655). Eisenberger and Cameron predict undermining only for performance-independent rewards, through helplessness. But engagement-contingent rewards, which the person controls by doing the task, undermine, while task-noncontingent rewards, the only plausible helplessness case, do not. Their own analysis shows the same. Two studies with luck-based rewards (Weiner 1980; Kruglanski et al. 1975) went the opposite way to their prediction. Their 1998 claim that performance-contingent rewards with specific standards enhance intrinsic motivation is not borne out (free choice −0.21, self-report −0.02, both n.s., pp. 655).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Expected, task-contingent tangible rewards undermine free-choice intrinsic motivation for interesting tasks | strong (k = 92, homogeneous subsets, robust to dull tasks and to dissertations) | pp. 640–644, 651–652 |
| C2 | Unexpected and task-noncontingent rewards do not | moderate (k = 9 and 7) | p. 640 |
| C3 | Positive feedback enhances intrinsic motivation, except children’s free choice | moderate (k = 21–24; age split k = 7 for children) | pp. 638–639 |
| C4 | Controlling administration turns positive feedback negative | moderate (4 studies, large effect) | p. 652 |
| C5 | Undermining persists at least two weeks in children | moderate (k = 12 per delay group; children only) | p. 650 |
| C6 | Effects are much weaker on self-reported interest than on free choice | strong as a description; the explanation (reward enjoyment inflates self-report) is argued | pp. 646, 655–656 |
| C7 | Cameron and Pierce’s discrepant result comes from collapsing dull with interesting tasks, wrong controls, misclassification and omitted studies | moderate: documented study by study; but by its own test dull tasks explain only “a relatively small amount” of the gap | pp. 632–634, 652, App. A–B |
| C8 | The pattern supports CET and contradicts the helplessness/learned-industriousness account | moderate (pattern fit; no mediation tested) | pp. 653–655 |
Concepts
- engagement-contingent reward: for working on the task, regardless of finishing or quality (p. 628).
- completion-contingent reward: for finishing the task (p. 628).
- performance-contingent reward: for doing well, meeting a standard or surpassing a criterion (p. 628).
- task-noncontingent reward: for something other than the task, such as turning up (p. 628).
- functional significance: what a reward means to the recipient, controlling or informational (p. 628).
- free-choice measure: time spent on (or trials of) the target activity in an unobserved free period after the reward phase (p. 635).
Connections
- The earlier meta-analyses (pp. 631–634). Rummel and Feinberg (1988), Wiersma (1992) and Tang and Hall (1995) found undermining; Cameron and Pierce (1994, LIT-556) and Eisenberger and Cameron (1996) did not, and called for “abandoning cognitive evaluation theory” (p. 632, quoting Cameron & Pierce p. 396). The paper treats the last two as “essentially the same meta-analysis” reported in pieces (p. 632).
- Cameron, Banko & Pierce (2001), LIT-562. The reply, read since in NOTE-587. Re-coding by procedure, it finds negative effects only for tangible, expected rewards “loosely tied to level of performance” or earned at less than the maximum, on high-interest tasks, and no pervasive harm. It agrees with this paper on rewards for merely doing the task.
- SDT summaries (LIT-559, LIT-558, LIT-560) cite this paper as having settled the matter. The 2006 paper says the earlier analysis was wrong by errors “numerous enough to generate the ‘null’ results” (LIT-560, p. 1568). This paper’s own claim is more measured: it explains the discrepancy by several factors and finds that one of them, the dull tasks, accounts for little of it (p. 652).
Bearing on the record
- Reward undermining. The narrowest claim this paper supports, that expected tangible rewards for merely engaging in an interesting task lower later free-choice engagement (d = −0.40), is also the one Cameron and Pierce found by this paper’s account (task-contingent −0.23, p. 649; performance-independent −0.29, p. 641), and the one the 2001 abstract keeps (“loosely tied to level of performance”).
- The methodological point (meta-analysis averages over the moderators a theory-driven literature was built to find, p. 631) bears on how the record reads any meta-analysis of a contested effect.
- No ML instruction. “Reward” here is a psychological manipulation, not a training signal. Nothing here belongs in the anthology.
Limitations
- CET’s categories are both the coding scheme and the hypothesis tested. The performance-contingent category mixes per-unit pay, rewards for “doing well” and rewards for beating a norm (Appendix A). If those behave differently, as the 2001 reply finds for rewards “linked to level of performance” (LIT-562: neutral for surpassing a score, positive for exceeding others), the category hides it.
- Most effects are small. Self-report effects of tangible rewards are around −0.07 to −0.17.
- The age finding is post hoc. “This set of findings has never been predicted before” (p. 656); the explanations offered are speculation.
- Imputed zeros are numerous for self-report (21%), which the paper itself takes as a sign that self-report is the weaker measure (p. 656).
- Mediation was not tested. “A similar analysis has not been done with respect to reward effects” (pp. 653–654) for perceived self-determination.
- Long-term, multi-session rewarding is nearly unstudied with controls: two such studies, nine rewarded participants in all (p. 650).
Open questions
- Do repeated rewards over weeks, with no-reward controls and adequate samples, undermine? The paper calls for such studies (p. 650).
- Does perceived locus of causality mediate tangible-reward effects?
- Why does positive feedback fail to raise children’s free choice?
Corrections
- none to a seeded skim (there was no seed)
- The brief’s citation is right. Crossref gives only the year; the issue is no. 6 of vol. 125.
- The brief says Deci, Koestner and Ryan “explicitly reanalysed Cameron & Pierce”. They did, and also Eisenberger and Cameron (1996), which they treat as the same analysis. Appendix A compares their treatment of each shared study with Cameron and Pierce (1994), not with the 1996 paper.