Dealing with non-detections in biological surveys

Our big brains struggle to coherently combine probabilities. It’s something that’s been noted time and again in studies on how doctors interpret a medical diagnosis but it’s a basic problem whenever we are presented information with several points of uncertainty. And this is a big problem commonly encountered in environmental management whenever we have imperfect detection.

Techniques to estimate the chance of not observing a species if it is in fact present (false absence) have been around in ecology for some time. But often we’re interested in the probability a species is present in circumstances where it is not detected. This is especially the case in environmental impact assessments, where failure to detect a population of a threatened species (a false absence) can have dire consequences.

It’s easy to be duped into mistakenly thinking a simple reworking of conventional techniques will give you the answer. Psychologists even have a name for this mistake – the inverse probability fallacy.

The alternative to relying on our intuition is using Bayes Theorem. Brendan Wintle and colleagues have re-jigged Bayes Theorem for the specific setting of repeat survey visits with imperfect detection.

Applying their approach suggests that if the species you’re after is especially cryptic such that you’re unlikely to detect it in any one visit, but the habitat is ideal, you need to revisit the site many times. If the habitat is poor and the species is easy to detect, a lesser effort is required.

It’s a curious thing that epidemiologists and psychologists, who have known how to approach and solve these kinds of problems for decades, have had limited success in getting doctors and vets to think straight. It’d be no bad thing if we could begin to routinely use Bayes theorem or logic trees to avoid similar blunders in environmental management.

See Decision Point #57 for the complete story


Wintle BA, TV Walshe, KM Parris, & MA McCarthy (2012). Designing occupancy surveys and interpreting non-detection when observations are imperfect. Diversity and Distributions. 18: 417–424.

Leave a Reply