How Rolling Dice Helps Save Leopards


Researchers working in South Africa are using a rolled-dice trick to solicit honest answers from farmers suspected of illegally killing leopards and hyenas. The trick uses randomized response, a method developed in the 1960s to eliminate people’s bias toward giving evasive answers. The basic idea is to prompt honest answers to sensitive questions (on topics such as sexual and even criminal behavior) by asking people to give an answer based on a random event. Researchers believe the resulting sense of confidentiality teases out truthful answers.

The study is called “Identifying Indicators of Illegal Behavior: Carnivore Killing in Human-Managed Landscapes,” and has been  published by Britain’s Royal Society; here’s the abstract:

Managing natural resources often depends on influencing people’s behavior, however effectively targeting interventions to discourage environmentally harmful behaviors is challenging because those involved may be unwilling to identify themselves. Non-sensitive indicators of sensitive behaviors are therefore needed. Previous studies have investigated people’s attitudes, assuming attitudes reflect behavior. There has also been interest in using people’s estimates of the proportion of their peers involved in sensitive behaviors to identify those involved, since people tend to assume that others behave like themselves. However, there has been little attempt to test the potential of such indicators. We use the randomized response technique (RRT), designed for investigating sensitive behaviors, to estimate the proportion of farmers in north-eastern South Africa killing carnivores, and use a modified logistic regression model to explore relationships between our best estimates of true behavior (from RRT) and our proposed non-sensitive indicators (including farmers’ attitudes, and estimates of peer-behavior). Farmers’ attitudes towards carnivores, question sensitivity and estimates of peers’ behavior, predict the likelihood of farmers killing carnivores. Attitude and estimates of peer-behavior are useful indicators of involvement in illicit behaviors and may be used to identify groups of people to engage in interventions aimed at changing behavior.

And here’s a summary of the field test from the Discover Magazine blog:

First, a researcher gave a farmer a die, which the farmer could roll without the researcher seeing its outcome.

Then, the researcher asked questions like “In the last 12 months, did you kill any leopards?,” and before the farmer answered, he rolled the die. If he rolled a 1, he should say “no,” no matter what the correct answer was, and say “yes” if he rolled a 6. For all the other numbers, he should answer honestly. The fact the researcher never had any way of knowing whether the farmer was saying “yes” because it was the truth or because he rolled a 6 gave the subject a sense of safety.

But after recording all the answers provided by their subjects, the researchers found that farmers said they had killed leopards far more frequently than one in six times, which is what would have been expected had they only admitted to the crime when the die made them. Under the cover of the die, farmers were admitting to having killed the creatures. In fact, judging from the results, at least 19% of the 99 subjects had killed leopards in the last year. That’s quite a bit more than researchers were expecting.

Mike B

Wow, that's really clever. I hope this sort of thing becomes standard for all such surveys.


Ah: why I like Freakonomics :D


The drawback is that you statistically destroy 1/3 of the information.
Here's my alternative proposition as an information theory buff : Use the power of xor to hide the information, and magically get it back on the way out.

Give the die, and say it a game to see if the respondent is able to lie without showing it off.
If they roll 2 to 6, they must say the truth whatever it is.
If they roll 1, they must lie, and give the answer that is the opposite of the truth.

If the percentage of positive is 0, you will get 16,6% no in the response. If the percentage of negative is 100, you will get 16,6% yes in the response. And it's a linear progression between the two in between. You just have to rescale the answer from the 16,66 - 83,33 to a 0 - 100 range to get back the true percentage. Of course nothing is free. Your error margin is also scaled up, but only by one sixth.

The limit of the technic is the risk that some respondents won't play straight :
- some respondent may use that as an opportunity to tell an unspeakable sin, even if the dice said he should have lied
- some other may find a thrill in lying, even if the dice said he should have told the truth because they find the truth too boring