Solving conservation crimes using robust estimators for Randomized Response Technique

How do conservation practitioners and researchers assess what factors promote illegal activities that violate species or land protection laws? Interviews are one of the most common and cost-effective methods.

A critical problem is that respondents have a strong incentive to lie under direct questioning: admitting to activities such as poaching tigers or illegally logging rosewood puts one at social and legal risk. Political scientists and sociologists faced similar obstacles studying sensitive behaviors such as support for insurgent groups. Researchers in these fields developed indirect questioning methods that ensure anonymity to interviewees, with should reduce biased responses.

Randomized response technique (RRT) is arguably the most prominent indirect question method. There are a variety of RRT designs, but all employ some type of a randomizer—be it a die roll, spin top, or card deck—that masks the respondent’s answer. When an interviewee answers “yes” or “no”, only they know if their answer is a true “yes” or a true “no”. However, because the randomization distribution is known (e.g. a dice has 6 sides), statistical models can return the true distribution of yes’s and no’s.

Collaborating with a sociology researcher, Dr. Maarten Cruyff, I developed an R package, zapstRR (ZoologicAl Package for Randomized Response Technique Analysis) that provides a set of user-friendly and robust statistical functions for RRT data. To date, there have not been any R packages offering multivariate (multiple RRT item) models; we provide two functions (RRsumscore and RRirt) that respectively 1) calculate “sum scores” (the ordinal sum of sensitive traits—e.g. hunting tigers, leopards, and dholes would equal a sum score of 3) and 2) jointly estimate regression parameters across multiple sensitive traits to increase statistical power.

Another major breakthrough offered by this package is the ability to estimate “evasive response bias” (also known as “one-sided lying”), or the proportion of individuals who consistently say “no”, even if they should say a forced or truthful “yes”. Evasive responses have long been suspected to impact conservation applications of RRT, but previously, there was no way to control for this issue. RRirt permits for such assessments.

The package vignette can be found at RPubs and the package itself can be easily downloaded into R using devtools. We provide two example datasets—one on illegal hunting in Southwest China and the other on illicit activities among university students—for users to explore the package functionality.

Written on October 17, 2017