Syndicate content

Add new comment

Behavioral design: slap or tax yourself into productivity?

David McKenzie's picture

One of those stories going the rounds about a month ago concerns a blogger in San Francisco, who worried he was wasting too much time on Facebook and Reddit. As he writes on his blog, he used a software app which tracked what he was doing with his time and found almost 19 hours a week went to these activities. His solution: hire someone through Craig’s list who would work next to him, and would slap him in the face if he went off task. He claims his productivity then increased from a base of spending 35-40% on productive activities to spending 98% on productive activities.

                This example came to mind while hearing Sendhil Mullainathan present at the World Bank last week on what he terms “behavioral design”. Martin Kanz summarizes some aspects of his talk here, while Sendhil has an excellent new paper related to the talk here. Reading this paper and the above example led me to re-read a paper by Sendhil, Supreet Kaur and Michael Kremer that I had seen a very early version of and then forgotten about. There is a lot more in it than I had seen, so here’s some highlights:

The idea: Not only do our bosses want us to work harder, but self-control issues mean workers do not work as hard as they themselves would like.

How do they test this? With an experiment with data entry workers in India. They work with an office that employs 64 data entry workers at a time – with worker turnover, 111 individuals enter into their experiment. They are typically young (average 24), three quarters are male, with 13 years average education, and most had taken a computer course before. They are hired on a temporary basis for a one-year job. Workers can choose how many hours they work each day, and whether or not they work each day.

This is a setting where the only output the firm cares about is the number of correctly entered fields – which can be observed, and where production is completely individualistic, and workers are paid piece rates. This makes it great for the experiment, since they can observe individual worker output at the daily and even hourly frequency. [Side note: I am puzzled that they don’t seem to care about error rates – presumably it should cost the company more if someone produces 5000 accurate fields and 5000 incorrect fields as if they produce 5000 accurate fields and 0 errors].

The experiment randomizes two things:

·         Payday: workers were randomized as to whether Tuesday, Thursday or Saturday would be their payday. This would allow the researchers to see whether workers work harder closer to their payday, when the gap between effort and reward was less.

·         Contract terms: the standard contract is a linear piece rate contract, paying workers Rs 0.03 per field correctly entered. This was the control condition. Three treatments were also used:

-          Mandatory Target treatment: workers were assigned a target (either 3000, 4000, or 5000 correct fields in a day). If they achieved this target, they were paid the standard rate per field correct. However, if they didn’t achieve this target they would only get half the rate per accurate field. i.e. there was a penalty for not achieving enough output.

-          Morning and Evening Choice treatments: workers in this group got to choose whether or not they had a target, and choose what this target was. They would either choose this in the morning right when they arrived in at work, or otherwise the night before. Note that choosing a target never gives the worker any more pay for a given level of output, and gives less pay if the worker doesn’t meet the target – so workers should only choose a positive target if they need this incentive to work harder (and thus earn more since they get paid piece rates).

Randomization of contract terms was done daily, stratified so that each 12 days each worker had spent 3 days in each of the treatment conditions and also 3 days in the control condition. This gives the authors lots of observations (over 8,000 person-days), but is a little weird – the authors check to see whether performance on one day seems independent of that on other days and it does, but it is still an unusual setting for workers to have contract terms changing back and forth so often for so long.

Results: The authors find several results consistent with self-control limiting productivity and with workers being sophisticated enough to realize this problem and use commitment devices when they are available:

·         Workers earn 8% more on paydays than in the beginning of the weekly pay cycle – consistent with them working harder when the time gap between effort and reward is lower.

·         When given a choice, workers choose a positive target 35% of the time, and workers are about 6% more productive when they do so.

·         It is the workers who show the largest pay-day effects who are most likely to choose targets

Things of particular interest to other researchers:

There are a couple of things in the paper that are likely to be of broader interest to those of us working on designing other studies and interventions:

Measuring discount and risk: First, while behavior around paydays predicts strongly who will choose the commitment work contracts, they find much less effects from more traditional survey and elicited measures of self-control – they do the usual choices between money today and money in the future (with real money), and also ask questions like agreement with statements like “some days I don’t work as hard as I would like” and “at the end of the day I get tempted to leave work earlier than I would like”, but find these have very little predictive power for contract choice – they note this may mean that observing behaviors in related contexts may be more powerful than subjective questions or abstract behaviors for predicting behavior. Thisaccords with my general feeling that many of our standard survey-based measures for risk-aversion and discount rates may not really be measuring what we want them to.

Monetarizing Impact: Second, at the end, they also randomize wages to see how sensitive production was to wages. This has the nice feature of allowing them to convert the magnitudes of the other effects they find into price terms – so they find, for example, that the production increase on paydays is approximately the same as the effect of raising the wage by 18%, while the extra production attained under the targets is approximately equal to a 14% increase in the piece rate. The work Sendhil did with Dean Karlan and others in South Africa provides another example of doing this – they showed not only did showing a nice picture lead to more loan applications, but by randomizing interest rates offered, could show what this impact was equivalent to in price terms.

Now give yourself a nice big slap in the face and get back to work!