Assessing pollinator’s coextinction risks

“Life is What Happens To You While You’re Busy Making Other Plans” is a famous quote by John Lennon. Life can also distract you from the little achievements that happen alongside it. In 2017, I entirely missed to dwell on one of my little achievements: the publication of my second PhD chapter. I would like to take a few minutes to explain what it was about.

Previously, I have talked a bit about coextinction. To remind you briefly – it’s the loss of species when their interaction partner species go extinct. The risk of losing one species with another used to be assessed based upon a species’ level of dependency, sometimes also called a species’ host-breadth (mostly used in parasite-host literature). We collect such information by observing interactions between species in the field, e.g., parasites on mammal hosts, insects visiting flowers.

But we’re just human, when we do field work we will inevitably miss some interactions between species or even the species themselves. The reasons for such observation bias are not always a person’s  mistake; some species may just be rare or may hide or the weather conditions can be unexpectedly cold (which especially affects insects). Obviously, sometimes it’s our fault: we may choose too a short time for our observations, we’re not careful enough when approaching the observation site, we can miss mobile species or plants flower / fruit at a different time to what we expected. There are plenty of reasons that lead to biased observations.

To overcome this problem of miss observing interactions or entire species, we developed a model*  to estimate how probable an interaction between two species in a plant-pollinator network is. In detail, we estimated

  • the number of interaction partner species of each species (serves as a measure of coextinction risk), and
  • how representative our network was in regards to the whole community.

We used our method on

  • simulated interactions with different scenarios to test how variation in sampling effort, interaction probability, and animal abundances, affected the results, and
  • we applied it to a real-world dataset of interactions between flowering plants and their insect visitors, which we collected in the Stirling Ranges in Western Australia.
Bluff Knoll good weather
Bluff Knoll the highest mountain in the Stirling Ranges. The top was my field site with the plant community of interest.

Our study showed that on one hand, the model performed well in predicting the number of interaction partners for scenarios of high animal abundances. On the other hand, for networks with many rare species our model showed high parameter uncertainty. Despite this drawback, the model estimates were generally better than just relying on observations.

The Stirling Range plant-insect community proved to have many interactions, but most species were rare species, thereby it closely resembled the simulation of high interaction rates with low abundances. While our results showed that we may have only detected 14-59% of the insect species, uncertainty in this estimate was high. This indicated that with our study design we were unable to find the majority of insect species in our study area. While we followed common sampling designs, in our particular case it was likely not appropriate, because the south-west of Western Australia is particularly biodiverse. With this in mind, if you plan an observational study in a biodiversity hotspot, you should probably amp up your study effort.


Plot of observed (purple star) and median estimated (green dot) number of insect species of the plant species in the Stirling Ranges. 2012 results on the left 2013 on the right. The line gives the 95% credible interval.


Our study confirms previous ones in that imperfect detection can strongly affect the inferences from observed interaction networks, especially network metrics which are often used to describe networks. This means that analysing observational data for dependency measures needs to deal with the the uncertainty that arises from imperfect detection – otherwise derived estimates may be highly misleading for assessing a species’ coextinction risks. Our study also shows how models can help estimating coextinction risk, but only when the underlying data is good, a problem that can be addressed by adapting the observation protocols to the specific study.

If you like to go a bit more into depth about the study and you’d like to nerd out on the model go here:

Plein, M., Morris, W. K., Moir, M. L., and P. A. Vesk 2017. Identifying species at coextinction risk when detection is imperfect: Model evaluation and case study. PLoS ONE 12(8): e0183351.


*for all the nerds out there: a hierarchical N-mixture model to fit to interaction data between animal species and individual plants

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s