Salience, Risk, and Effort

2020-10-28

2020-11-02

How often respondents choose to do more tasks rather than fewer.

Participants on Amazon Mechanical Turk will make a series of binary effort choices, where they can either complete 1 or 6 "tasks". Each task consists of transcribing blurry greek letters. Each option comes with a lottery, where the lottery for 6 tasks weakly dominates the lottery for doing 1 task. Holding the expected benefit of completing extra tasks constant, I vary the riskiness of exerting effort so that the upside of doing so becomes more or less salient.

Each participant will make 12 binary choices, as described above. A random one of these choices will be implemented, to make all choices incentive compatible. I vary within individual both the objective expected benefit from doing more tasks (4 different incentive levels), and the "shape" of the incentives, which can be linear (earn $X extra no matter what), convex (earn a big bonus with a small probability) or concave (avoid missing out on a big bonus with a small probability).

The respondents' payoff is (implicitly: it is not described to them this way) a function of their effort, which in my model allows them to produce a unit of output, and whether they by chance produce an additional unit of output (regardless of their effort decision). A key parameter is how likely they are to produce this additional output, which I denote p. This parameter is important because when p is low (in my experiment, 0.10), convex incentives should have a positive effect on effort, and when p is intermediate (0.50) or high (0.90) concave incentives should. I therefore have three treatments corresponding to these three cases.

I use a timer to implement the lotteries. Respondents get to stop a timer that counts up from zero at a time of their choosing. The last two digits of the timer (which includes milliseconds) are the numbers used to determine their payoffs in the earnings. In piloting, I have found both that respondents are unable to control these last two digits (which appear uniformly distributed as we would expect) and do not appear to believe they can do so (e.g., they do not systematically pick risky lotteries that would be especially attractive if they could control the value of the timer better than they in fact can).

To randomly assign respondents to treatments, I use the randomizer function in Qualtrics.

Across individuals, I randomize whether p (the parameter described above) is 0.10, 0.50, or 0.90. Within individuals, I randomize the order in which the binary effort choices are presented to them.

Randomization occurs within individual. In particular, each participant sees all twelve combinations of expected incentive strength in a random order. However, to account for correlation across choices within individual, I will cluster my standard errors at the individual level where relevant.

Yes

I plan to include 400 individuals per treatment arm, or 1,200 overall.

Each respondent makes 12 effort decisions, so the effective number of observations is 14,400.

400 people per treatment arm.

The unit of the main outcome is binary: whether or not they choose to exert effort. Call the mean of this variable a_bar. The variance is therefore a_bar*(1-a_bar). Using simulations, I estimate that the planned sample size will be sufficient to detect differences between incentive shapes of 5 percentage points at p < 0.05 and with power 0.80. I use simulation because I will control in the regression for individual fixed effects and expected strength of incentives, and cluster at the individual level.