Cognition helps wildlife exploit novel resources and environments. Raccoons (Procyon lotor) have successfully adapted to human presence in part due to their cognitive abilities. However, close interactions between humans and wildlife can create conflicts. A better understanding of the raccoon’s behavioral flexibility and learning ability could improve the mitigation of those conflicts. Learning can be evaluated over multiple exposures to a cognitive task. Our objective is to evaluate wild raccoons learning in contexts varying in terms of exposure to humans (recreational and preservation zoning within protected areas) and task difficulty. We used two food extraction tasks to measure how problem-solving performance varied between trials based on success probability and time to solve the puzzles.

Field work

We conducted experiments in three protected areas located in southern Québec, Canada (all located around latitude 45.6° N and between longitude 72.6° and 75.2° W): Plaisance (28 km²), Îles-de-Boucherville (8 km²) and Yamaska (13 km²) national parks. These parks are considered to have a high density of raccoons, causing “severe” nuisance problems (Denis 2017; Dellarosa 2012). All these parks are relatively small, encompassed in mostly urban or agricultural territories, and border large bodies of water (river or lake). Two site categories are studied based on management zoning: intensive recreation and preservation zones. Recreation zones were defined by the presence of camping sites, vehicular circulation at low speed (<20 km/h), campfires, dumpsters, and a mosaic of ground cover (gravel or paved roads, parking lots, forests, open fields, buildings, playgrounds). Preservation zones were strictly accessible to the public by walking and biking trails, with extensive forest cover. We ran the experiments for three summers (earliest-latest dates: May 31‒Sept. 14) between 2019‒2021. Plaisance Park was not visited in 2021 because of its lower raccoon activity compared to the two other parks and its lower accessibility.

We used species-specific baits, but all wildlife could interact with the devices (we only recorded eight interactions by stripped skunks in addition to raccoons). The experiments are non-invasive; animals voluntarily approached the apparatus and left. This ensured that only motivated animals participated. Raccoons were trap-shy, and although tested, capture-marking did not prove efficient in identifying individuals. Raccoons were identified solely by LL through careful observations of the video footage, based on their size relative to the puzzles, body characteristics (fur, tail, limbs), marking when available, and scars and injuries, in a similar manner to Chow et al. (2021a) with Eastern grey squirrels (Sciurus carolinensis). Juveniles were excluded because they showed very little initiative and were impossible to tell apart from the videos, often interacting together with the devices, therefore creating confusion to track one individual at a time. It was also impossible to identify an individual as a juvenile one year and as a grown adult on consecutive ones. To increase our confidence in the identification process, we conducted an intra-rater reliability test (Cohen’s kappa) on a small subset of recordings from the site with the highest activity level (Îles-de-Boucherville, recreation zone). We obtained an 87% agreement (κ = 0.851) corresponding to an “almost perfect” agreement (Landis and Koch 1977).

We used two puzzle box tasks to test problem-solving abilities (Fig. 1). Raccoons trying to open food containers (e.g. plastic boxes, bags, bottles) is a common occurrence when they are exposed to humans, making this task contextually relevant (Knight 2022). The first puzzle we used (hereafter, the Box) is similar to the model used with carnivore species in other studies (Benson-Amram and Holekamp 2012; Benson-Amram et al. 2016). Using the same type of puzzle facilitates the comparison of our results to similar experiments (Krasheninnikova et al. 2020). The Box measured 30 cm on each side and was made from steel mesh. To solve this problem, a raccoon had to slide a latch and pull on a door. The second puzzle (hereafter, the Tube) consisted of two horizontal plastic tubes (50 cm long in total and 7 cm in diameter), one sliding over the other. It required the animal to slide and turn the outermost tube to align two holes (approximately 5 x 10 cm) and access the food in the inner tube. Both puzzles necessitated two consecutive actions that can be performed with the paws, mouth, or muzzle of the animal.

Video analysis

We considered the puzzle to be solved when a raccoon opened it enough to have direct access to the food with its paw, even if it did not immediately reach in and consume the reward. An attempt began when an animal approached within one body length of the box and ended when the animal moved more than a body length away from the puzzle or when it opened the puzzle. A trial included all attempts at opening a puzzle within a single night, so the trial ended when the last attempt ended. We recorded interactions with night vision cameras (Argus 2, Reolink, Hong Kong) set up 3‒4 m away (see Supplementary Material for an example with each puzzle). We quantified cognitive performance in problem-solving ability from the videos. We used two response variables to quantify problem-solving efficacy: 1) success (binomial) when the subject opens the puzzle (or not) to have direct access to the bait and 2) time to solve (continuous), which is the cumulative time from the first attempt until the puzzle opening within a trial. Two discrete terms representing behavioral traits were included in our models: persistence is the sum of all attempts within a trial, including the one when the puzzle is solved, and exploratory diversity is the number of unique behaviors directed at the puzzles. We calculated exploratory diversity in a similar manner to previous studies (Benson-Amram and Holekamp 2012; Benson-Amram et al. 2013, 2014; Johnson-Ulrich et al. 2018; Daniels et al. 2019). The exploratory diversity score accounts for behaviors without contact that allow the gathering of information (circling around the puzzle and sniffing), as well as behaviors with physical contact (Table 1). NA ("not available") means it was not possible to determine from the video recordings.

Data analysis

We performed a generalized linear mixed-effect model (GLMM) to examine how success changed over successive trials. We used a binomial distribution with the logit function and included the fixed covariates zone and puzzle type in interaction with trials. Individual and year were included as random effects. To assess the effect of trial number on solving time, we ran a GLMM using only successful trial data. We used a Gamma distribution with the log function. We explored interactions between trials and zones, as well as between trials and puzzle types. Year and ID were included as random terms, and the model was optimized using the Nelder-Mead optimizer from the R package “nloptr” (Nelder and Mead 1965; Johnson 2014).

To see if persistence (number of attempts) at a task changed over trials, we performed a GLMM with a Poisson distribution and log link function, controlling for repeated measures within an individual by including ID as a random term. We also calculated the success probability and mean number of attempts on unsuccessful trials (as a proxy of persistence) for each individual and tested for a correlation with a Kendall rank correlation coefficient. To see if exploratory diversity changed over trials, we performed a GLMM with a Poisson distribution and log link function, controlling for repeated measures within an individual by including ID as a random term.

To assess the effect of year, we used a log-likelihood ratio test, comparing a model with year as a random effect to one without, while keeping the fixed effect structure constant (lmtest package; Zeileis and Hothorn 2002). We also performed McNemar’s Test (with continuity correction) to determine if the proportions of success significantly differed when matching pairs of subjects (Fagerland et al. 2014) at their last trial of the year was different from the first of the next year. We used the program R to run all statistical analyses (4.2.3, R Core Team 2023).