22,500 coin flips
The Levitt experiment. What it found, and the three things it does not show.
$ eldr --cite levitt-2016
The study
Steven Levitt, University of Chicago. Roughly 22,500 usable coin tosses, from people who arrived at a website with a decision they could not make and agreed to let a coin break the tie. Heads meant act, tails meant wait. The site opened in January 2013 and collected for about a year, with participants surveyed at two months and at six.
Reported in The Review of Economic Studies 88(1); working paper NBER 22487.
What it found
- On the decisions participants themselves rated as important, such as quitting a job or ending a relationship, those told to act were about 11 percentage points more likely to have acted at two months. Across the full sample the gap was larger, roughly 25 points, because the less weighty questions moved more easily.
- At six months, wellbeing was significantly higher in the change arm **for the important questions, and not for the less important ones**. That restriction is the finding, not a footnote to it.
- Per question, quitting a job and ending a relationship carry the clearest six-month results.
What it does not show
1. It does not show the coin was right. It shows that breaking a stalemate beats holding one, on average, for people who had already self-selected as stuck. 2. It does not transfer to every decision. The sample is people who volunteered a decision to a website and accepted a coin flip. That is not a random sample of human decisions and nobody should pretend otherwise. 3. It does not license prediction. Nothing in the paper forecasts an outcome for an individual. Neither does this product.
Why it is the foundation here
The finding is narrow and it is enough: an externally generated position you did not author interrupts the loop your own reasoning is running inside. Second replaces the coin with fourteen answers, and then — unlike the coin — has to show its work.