Reward Contingency

What it is

The structure, measured on free-choice behaviour across the meta-analysis's 128 studies, runs as follows. Engagement-contingent rewards, 55 studies, d = -0.40, splitting by age into 43 child studies at -0.46 and 12 college-student studies at -0.21. Completion-contingent rewards, 20 studies, d = -0.36. Task-noncontingent rewards, 7 studies, d = -0.14, not significant.

That third row decides whether any of this is about jobs, and the authors say so themselves. Their own example is a study in which participants were paid two dollars simply for taking part, independent of what they did, and their stated reason for the category is to simulate the typical work condition in which people are paid a salary without having their pay tied in any direct way to the specific tasks they perform. Seven studies, no significant effect.

Expectancy is a second cut and it is just as decisive. Expected tangible rewards, 92 free-choice studies, d = -0.36. Unexpected tangible rewards, 9 studies, d = +0.01, homogeneous, no effect at all. That independently corroborates the null Mark Lepper of Stanford printed alongside his own famous 1973 result, in which children rewarded unexpectedly did not differ from controls.

The performance-contingent split is the part almost nobody prints, and it cuts both ways at once. Because those effects were not homogeneous, the 1999 analysis divided them four ways by type of control group and feedback valence. The row the authors themselves call the real-world one, where rewards are a direct function of performance so the best performers get the largest rewards and others get less or none, produced the largest undermining effect in the entire meta-analysis, d = -0.88, on six studies. The neighbouring rows are far smaller: d = -0.15 on 18 studies where everyone received the maximum reward, and d = -0.20 on 10 studies with a comparable-feedback control.

The opposing camp reads the same category the other way. Judy Cameron, Katharine Banko and W. David Pierce state in their 2001 abstract that when rewards are linked to level of performance, measures of intrinsic motivation increase or do not differ from a non-rewarded control group. That is a live disagreement about coding and aggregation rather than about arithmetic, and the fair way to present it is as two defensible procedures reaching two answers.

One disclosed limitation belongs on this page because the authors disclosed it. The test for differences among the four contingency categories on free-choice behaviour did not reach significance, with Qb(3) = 5.97 and p < .12, and they analysed the categories separately anyway on theoretical grounds and said so in print. The categories are theoretically motivated and were not shown to differ statistically from one another, and both halves of that sentence are owed to a reader.

In effect

Six studies carrying the real-world row

The research case worth sitting with is the performance-contingent breakdown, because it is where the two strongest claims in the dispute both live. The largest undermining effect in a meta-analysis of 128 studies is the row that most resembles a commission scheme, and it rests on six studies. A reader told only the headline learns too little; a reader told only that row learns something built on six studies. Both facts have to travel together, which is what a moderator table is for.

A salary, a bonus and a thank-you

In everyday terms the four categories map onto four different conversations at work. Being paid a salary is task-noncontingent, the category with no significant effect across seven studies. Being paid for each unit is engagement- or completion-contingent, the categories where the effect is largest. Being paid according to how well the work was judged is performance-contingent, the contested row. And being told the work was good is verbal reward, which moves motivation the other way at d = +0.33, significantly so for college students at +0.43 and not significantly for children at +0.11, a split the authors call particularly noteworthy.

The piece rate that raised output by forty-four per cent

The commercial counterweight is Edward Lazear's study of Safelite Glass Corporation, which moved from hourly pay to piece rates between 1994 and 1995 and whose published abstract reports productivity effects amounting to a forty-four per cent increase in output per worker, with profits rising too. Two cautions belong with that number every time. The earlier working paper gives the effect as about twenty to thirty-six per cent depending on what is held constant, so the published article and the published figure should be cited together and not mixed. And the paper's own theory predicts the firm attracted a more able workforce, with about half the worker-specific productivity gain paid out in higher wages, so selection and incentive both sit inside the headline. Installing windscreens is also not a task anybody claims is intrinsically fascinating, which is exactly why the result is compatible with the undermining literature rather than a refutation of it.

What it does not say

It does not say rewards in general damage motivation. Three of the four contingency categories behave differently from one another, one of them shows nothing, and verbal reward runs the other way.

It does not say the four categories are statistically distinct. The omnibus test across them did not reach significance and the authors report that themselves.

It does not say performance pay is settled. The two meta-analytic traditions disagree on that specific row, and the disagreement is about how to code control groups and whether to aggregate effect sizes within a study.

It does not transfer to salaried work on its own. The category built to simulate a salary is the one with no significant effect, across seven studies, and seven studies is a thin base for a confident claim in either direction.

Who developed it

The taxonomy is Edward Deci's, with Richard Koestner of McGill University and Richard Ryan of the University of Rochester, published in Psychological Bulletin in 1999. The order matters: the sequence "Deci, Ryan and Koestner" that circulates belongs to their 2001 response paper, not to the meta-analysis. The opposing reading of the performance-contingent row is Judy Cameron's, with Katharine Banko and W. David Pierce.

---