Social transmission in the wild can reduce predation pressure on novel prey signals

Study site

The experiment was conducted at Madingley Wood, Cambridgeshire, UK (0◦3.2´E, 52◦12.9´N) during summer 2018. Madingley Wood is an established field site with an ongoing long-term study of the blue tit and great tit populations. During the autumn and winter birds are caught from feeding stations using mist nets and they are fitted with British Trust of Ornithology (BTO) ID rings. Since 2012, blue tits and great tits have been fitted with RFID tags (BTO Special Methods permit to HMR), which enables collecting data remotely about their foraging behavior and social relationships. The study site has 90 nest boxes that are monitored annually during the breeding season. In 2018 chicks (n = 325) fledged successfully from 45 nest boxes (blue tits = 21, great tits = 24) and they were all ringed and fitted with RFID tags when they were approximately 10 days old. Because new juvenile flocks were arriving at our study site throughout the summer, we also conducted several mist-netting and ringing sessions in July and August to maintain a high proportion of blue tits and great tits ringed and RFID tagged for the experiments (on average 89%, see below). The study protocol was approved by the Animal Users Committee at the Department of Zoology, University of Cambridge.

Food items

We investigated birds’ foraging choices by offering them colored almond flakes at bird feeders that were distributed throughout the wood. Before beginning the experiments, we allowed the birds to become familiar with the food items by providing plain ‘control’ almonds (plain and not colored) in paired feeders (1.5 m apart) at three locations (approximately 170 m from each other). The feeders were surrounded with metal cages to exclude larger birds, and we placed plastic buckets under the feeders to collect any spilled almonds and minimize birds’ opportunities to forage from the ground. We introduced the feeders at the beginning of June when the nestlings had fledged and were beginning to forage independently, and continued to provide these plain almonds in between our learning experiments (Fig. 3c).

In the learning experiments, almond flakes were dyed with non-toxic food dye (Classikool Concentrated Droplet Food Colouring). We used three different color pairs: green (Leaf Green) and red (Bright Red), purple (Lavender Purple) and blue (Royal Blue), and orange (Satsuma Orange) and yellow (Dandelion Yellow). Almond flakes were dyed by soaking them for approximately 20 min in a solution of 900 ml of water and 30 ml of food dye and then left to air dry for 48 h. In the avoidance learning experiments, we made half of the almond flakes unpalatable by soaking them for one hour in 67% solution of chloroquine phosphate, following previously established methods from avoidance learning studies with birds in captivity^14,23,24,25. The food dye was added to the solution during the last 20 min.

Red and green are common colors used by aposematic, or cryptic prey, respectively⁹. Therefore, we investigated whether blue tits and great tits had initial color biases towards these colors before starting the main experiment. Because we did not want the birds in our study population to have any experience of the colors before the main experiment, this pilot study was conducted in Newbury, which is 130 km from our main study site. Birds were simultaneously presented with two feeders containing red and green almonds (both palatable) for 30 min and the number of almonds of each color taken by blue tits or great tits was recorded using binoculars. The position of the feeders was switched after 15 min to control for any preferences for feeder location, and the test was repeated on 9 different days. We did not find any evidence that birds had initial color preferences (t-test: t = 0, df = 15.69, p = 1). For the other two learning experiments, we chose color pairs that were available as a food dye and as different from red and green in the visible spectrum as possible to avoid generalization across experiments. These color pairs (blue/purple and yellow/orange) had similar contrast ratios as green and red, based on their RGB values (measured from photographs, see Supplementary Information). Although avian and human vision is different, the discriminability of colors is likely to be similar⁵¹, and rapid avoidance learning in each experiment shows that all colors were easily distinguishable. This was the main requirement for testing social information use, and subtle differences in color pair discriminability should only introduce noise to our data but not influence our conclusions.

Learning experiments with colored almonds

We conducted three avoidance learning experiments with different color pairs throughout the summer: red/green, blue/purple, and yellow/orange (unpalatable/palatable). In addition, we conducted a reversal-learning experiment with the blue/purple color pair by making both colors palatable after birds had acquired avoidance to blue almonds. Each experiment followed a similar protocol, in which birds were presented with colored almonds at the same three feeding stations where they were previously offered plain almonds. Each feeding station had two feeders, where one contained the palatable color and the other contained the unpalatable color (except in the reversal learning test when both colors were palatable). We switched the side of the feeders every day to make sure that birds learned to associate palatability with an almond color and not a feeder position. The feeders were filled at least once a day (or more often if necessary) to make sure that birds always had access to both colors. We continued each avoidance learning experiment until >90% of all recorded visits were to the feeder with palatable almonds, indicating that most birds in the population had learned to discriminate the colors. This took 7 days in the red/green experiment and 8 days in the other two color pairs (blue/purple and yellow/orange). The reversal learning experiment was finished after 9 days when 50% of the visits were to the previously unpalatable color (blue), indicating that most birds had reversed their learned avoidance towards it.

Recording visits to feeders

We monitored visits to all feeders using RFID antennas and data loggers (Francis Scientific Instruments, Ltd) that scanned birds’ unique RFID tag codes when they landed on a perch attached to the feeder. During the learning experiments, each day we also recorded videos from all three feeding stations (using Go Pro Hero Action Camera and Canon Legria HF R66 Camcorder). From the videos, we monitored the proportion of blue tits and great tits that did not have RFID tags and were therefore not recorded when visiting the feeders. We calculated the estimated RFID tag coverage for each day of the experiments by watching at least 100 visits to the feeders from the videos (divided equally among the three feeding stations) and recording whether blue tits and great tits had an RFID tag or not. We realized that the number of untagged individuals was very high (approximately 50% of all visiting birds) when we started the experiment with the first color pair (red/green; see Supplementary Fig. 3). We, therefore, stopped the experiment after two days and caught birds from the feeding stations with mist nets to fit RFID tags to new individuals. To maintain a high number of individuals RFID tagged for the other color pairs, we conducted a mist netting session a day before starting each experiment, as well as 4–5 days after it. We always switched the feeders back to containing plain almonds during mist-netting sessions to ensure that this would not interfere with the learning experiments. Apart from the first two days of the red/green experiment, the RFID tag coverage was on average 89% throughout the experiments (varying between 80 and 95%, Supplementary Fig. 3).

Birds were recorded every time that they visited the feeders, i.e., landed on the RFID antenna. However, it is possible that birds did not take the almond during every visit. To get an estimate of how often birds landed on the antenna without taking the almond, and whether this differed between palatable and unpalatable colors, we analyzed the visits to the feeders from the video recordings. We watched videos from the five first days of each experiment (i.e., different color pairs) and analyzed 60 visits to each color (divided approximately equally among the three feeding stations). We recorded whether the feeding event happened (birds ate the almond at the feeder or flew away with it) or whether birds left the feeder without sampling the almond. Because the number of visits to the unpalatable feeder was low during the last days of the avoidance learning experiments, we decided not to analyze avoidance learning videos after day five (but recorded visits from all days of the reversal learning experiment). We found that in avoidance learning experiments birds started to ‘reject’ unpalatable almonds after two days, i.e., they sometimes landed on the feeder but flew away without taking the almond (see Supplementary Fig. 4a). This change was not observed at palatable feeders where birds continued to consume almonds at a similar rate as at the beginning of the experiment (Supplementary Fig. 4a). In reversal learning, the proportion of visits that did not include a feeding event did not differ between purple and blue almonds: birds showed similar hesitation towards both colors at the beginning of the experiment, but this wariness decreased when the experiment progressed, with birds taking the almond during most of their visits (Supplementary Fig. 4b).

Statistical analyses and model validation

Foraging choices in learning experiments

We first analyzed how birds’ foraging choices changed during the learning experiments using generalized linear mixed-effects models with a binomial error distribution. The number of times an individual visited each feeder on each day of the experiment was used as a bounded response variable, and this was explained by species (blue tit/great tit), individuals’ age (juvenile/adult), and day of the experiment (continuous variable), as well as bird identity as a random effect. When analyzing avoidance learning, initial exploration of data suggested that results were similar across all three experiments, so we combined the experiments in the same model. To investigate whether learning curves differed between the species or age groups, the day of the experiment was included as a second-order polynomial term, and we started model selections with models that included a three-way interaction between species, age, and day². Best-fitting models were selected based on Akaike’s information criterion (see Supplementary Tables 1 and 2).

Social network

To investigate if birds used social information in their foraging choices, we first constructed a social network of the bird population based on their visits to feeders outside of the learning experiments, i.e., when birds were presented with plain almonds (in total 92 days, see Supplementary Information for the robustness of analysis to exclusion of network data before or after the experiment). We used only these data as individuals were likely to vary in their hesitation to visit novel colored almonds. We used a Gaussian mixture model to detect the clusters of visits (‘gathering events’) at the feeders⁵² and then calculated association strengths between individuals based on how often they were observed in the same group (gambit of the group approach). These associations (network edges) were calculated using the simple ratio index, SRI³⁵.

where x is the number of samples where individuals A and B co-occurred in the same group, y_A is the number of samples where only individual A was seen, y_B is the number of samples where only individual B was seen, and y_AB is the number of samples where both A and B were observed in the same sample but not together. Network associations, therefore, estimated the probability that two individuals were in the same group at a given time, with the values scaled between 0 (never observed in the same group) and 1 (always observed in the same group).

Social information use during avoidance learning: model description

If social avoidance learning was occurring, then the more birds observed negative responses of others feeding on the unpalatable feeder, the less likely they would be to choose the unpalatable feeder themselves. Thus, we expected the probability of bird j choosing the unpalatable option at time t to decrease with ({R}_{-,j}left(tright)) (the real number of negative feeding events observed by j prior to time t). Likewise, if appetitive social learning was occurring, then the more birds observed positive responses of others feeding on the palatable feeder, the more likely they would be to choose the palatable feeder themselves (rather than the unpalatable feeder). So, we also expected the probability of j choosing the unpalatable option at time t to decrease as ({R}_{+,j}left(tright),)(the real number of positive events observed by j prior to time t) increased.

However, we could not test for an effect of ({R}_{-,j}left(tright)) and ({R}_{+,j}left(tright)) directly, since birds often ate the almond away from the feeder, and therefore the real number of observed feeding events could not be measured. Instead, we aimed to test for a pattern following the social network that is consistent with these social learning processes. We reasoned that the probability that one individual i, observes a specific feeding event by another individual j, was proportional to the network connection between them, a_ij (probability they are in the same feeding group at a given time). Therefore, in each avoidance learning experiment (i.e., different color pair), we calculated the expected number of negative feeding events observed, prior to each choice (occurring at time t) as

where ({N}_{-,j}left(tright)) was the number of times j had visited unpalatable almonds prior to time t (i ≠ j), and summation is across all birds in the network, and likewise for the expected number of positive feeding events:

where ({N}_{+,j}left(tright)) was the number of times j had visited palatable almonds prior to time t (i ≠ j).

We analyzed whether the expected observations of positive and/or negative feeding events of others influenced the foraging choices in the avoidance learning experiments using generalized linear mixed-effects models with a binomial error distribution. We used each choice (i.e., visit a feeder) as a binary response variable (1 = unpalatable chosen, 0 = palatable chosen), with the probability that unpalatable feeder is chosen on feeding event E given by ({p}_{E}={p}_{-,{i}(E)}left({t}_{E}right)), where i(E) is the individual that fed during event E and ({t}_{E}) is the time at which event E occurred. We then modeled the probability of i choosing the unpalatable option at time t as:

where ({N}_{+,i}left(tright)) is the number of times a choosing individual had visited the palatable feeder (positive personal information), ({N}_{-,i}left(tright)) is the number of times a choosing individual had visited the unpalatable feeder (negative personal information), ({O}_{+,i}left(tright)) is the expected number of observed positive (positive social information) and ({O}_{-,i}left(tright)) observed negative feeding events (negative social information). Bird identity was included as a random effect, ({{rm{{B}}}}_{i}) (age and species were later added as variables, see below). Parameters ({beta }_{{rm{asoc}}+}) and ({beta }_{{rm{asoc}}-}) are the effects of asocial learning about the palatable and unpalatable foods, ({beta }_{{rm{soc}}+}) is the effect of social learning about the palatable food, and ({beta }_{{rm{soc}}-})is the effect of social avoidance learning about the unpalatable food. Estimation of these parameters, with associated Wald tests and confidence intervals, allowed us to make inferences about which effects were operating and the size of these effects. To aid model fitting we standardized all predictor variables and then back-transformed the effects to the original scale (see Supplementary Tables 3–5 for the model outputs). To assess the importance of asocial and social effects, we also ran separate models that excluded either asocial or social parameters and compared them to the initial model in Eq. (5) using Akaike’s information criterion (see Supplementary Table 6). However, in most cases, this reduced model fit significantly, and we, therefore, kept all parameters in the final models.

Our approach took ({O}_{-,{i}}left(tright)) as a measure of ({R}_{-,j}left(tright)), and ({O}_{+,{i}}left(tright)) as a measure of ({R}_{+,j}left(tright))-, which we termed the ‘expected’ number of observations of each type. Strictly speaking, ({O}_{-,{i}}left(tright)) and ({O}_{+,{i}}left(tright)) were upper limits on the expected number of observations, assuming that birds observed all feeding events in the groups in which they were present, whereas only an unknown proportion of such events (({p}_{o})) was observed. Therefore, the real expected number of negative/positive observations would be (Eleft({R}_{-,j}left(tright)right)={p}_{o}{O}_{-,{i}}left(tright)) and (Eleft({R}_{+,j}left(tright)right)={p}_{o}{O}_{+,{i}}left(tright)) respectively. Thus, the coefficient, ({beta }_{{rm{soc}}-}), for the effect of ({O}_{-,{i}}left(tright)) could be interpreted as ({beta }_{s{rm{oc}}-}={p}_{o}acute{{beta }_{{rm{soc}}-}}) where (acute{{beta }_{s{rm{oc}}-}}) is the effect per observation. Note that since ({p}_{o}le 1), and(,{beta }_{{rm{soc}}-}=acute{{beta }_{s{rm{oc}}-}}{p}_{o}), ({beta }_{s{rm{oc}}-}) is more likely to underestimate than overestimate the effect per observation of a negative feeding event. An analogous argument applies to the coefficient, ({beta }_{{rm{soc}}+}), for the effect of ({O}_{+,{i}}left(tright)).

Social information use during avoidance learning: extension to test for species effects

After fitting the initial model shown in Eq. (5), we further broke down the model to test whether individuals were more likely to learn socially by observing conspecifics than heterospecifics. This was done by splitting the expected number of observed positive and negative feeding events to observations of conspecifics (({O}_{+{rm{C}},i}left(tright)), ({O}_{-{rm{C}},i}left(tright))) and heterospecifics (({O}_{+{rm{H}},i}left(tright)), ({O}_{-{rm{H}},i}left(tright))), and including these in the model as separate explanatory variables thus:

with ({beta }_{{rm{soc}},-{rm{H}}}) and ({beta }_{{rm{soc}},-{rm{C}}}) giving the effect of a negative observation of a heterospecific and conspecific, respectively, whereas ({beta }_{{rm{soc}},+{rm{H}}}) and ({beta }_{{rm{soc}},+{rm{C}}}) give the effect of positive observation of a heterospecific and conspecific, respectively. In general –/+ subscripts refer to negative/positive feeding events and C/H subscripts to feeding events by conspecifics/heterospecifics. By re-parameterizing the model thus:

we were able to test for a difference between observations of negative feeds by conspecifics and heterospecifics (left({beta }_{{rm{soc}},{rm{C}}-}-{beta }_{{rm{soc}},{rm{H}}-}right)) and between observations of positive feeds by conspecifics and heterospecifics (left({beta }_{{rm{soc}},{rm{C}}+}-{beta }_{{rm{soc}},{rm{H}}+}right)).

For all experiments there was no evidence for a difference between ({beta }_{{rm{soc}},-{rm{H}}}) and ({beta }_{{rm{soc}},-{rm{C}}}) (yellow/orange: Z = 0.803, p = 0.42; red/green: Z = 0.065, p = 0.95; blue/purple: Z = 1.113, p = 0.27). However, there was some evidence of a difference between ({beta }_{{rm{soc}},+{rm{H}}}) and ({beta }_{{rm{soc}},+{rm{C}}}) in two of the three experiments (yellow/orange: Z = 1.359, p = 0.17; red/green: Z = 1.417, p = 0.16; blue/purple: Z = 0.729, p = 0.47). Consequently, we reduced the model down to:

for further analysis, i.e., with different effects for observations of conspecific/heterospecific positive feeds, but not of negative feeds. We did this for all color combinations (including blue/purple) to allow comparison across experiments (see Table 1). The R code used to run these models can be found in Supplementary data⁵³ in ‘GLMM models Orange Yellow final.r’.

Social information use during avoidance learning: simulations to test for a network effect

Next, we tested whether the social effects we detected followed the social network. When using a network-based diffusion analysis (NBDA⁴³), researchers can compare a network model with one in which the network has homogeneous connections among all individuals, but we found this to be unreliable for our model. Instead, we used a simulation approach to generate a null distribution for the null hypothesis of homogeneous social effects, taking the size of the social effects from the fitted models. We ran 1000 simulations (using the same procedure described above) for all social effects that were found to be significant in each avoidance learning model (each color pair; see Table 1). The total number of expected observations was kept equal, but we homogenized the observation effect across all birds by replacing the probability of bird i observing a feed by bird j, previously ({a}_{{ij}}), with ({sum }_{i}{a}_{{ij}}/n), where n is the number of birds in the experiment, (i.e., all birds had the same probability of observing each feeding event). The model was fitted to the simulated data each time to extract the Z value (Wald test statistic) of the social effect of interest. The distribution of these values was then used as a null distribution to test whether our observed social effect differed from the effects that did not follow the social network. To this end, we calculated the proportion of simulations that yielded a Z value as extreme or more extreme than that observed (judged by distance in either direction from the mean of the null distribution). The R code used to run these simulations can be found in Supplementary data⁵³ in ‘Simulations to test if network effects follow network Orange Yellow.r’.

Social information use during avoidance learning: extension to test for age effects

We then aimed to test whether each of the three social effects detected differed based on the age class of the observed individual (adult versus juveniles). We, therefore, split the negative expected observations ({O}_{-,i}left(tright)) into the expected observations of adults ({O}_{-{rm{A}},i}left(tright)) and juveniles ({O}_{-{rm{J}},i}left(tright)), each with its associated coefficient in the model ({beta }_{{rm{soc}},-{rm{A}}}) and ({beta }_{{rm{soc}},-{rm{J}}}). Likewise, we split positive observations of conspecifics as ({O}_{+{rm{CA}},i}left(tright)) and ({O}_{+{rm{CJ}},i}left(tright)) and positive observations of heterospecifics as ({O}_{+{rm{HA}},i}left(tright)) and ({O}_{+{rm{HJ}},i}left(tright)) to give the model:

As before, –/+ subscripts refer to negative/positive feeding events, C/H subscripts to feeding events by conspecifics/heterospecifics, and A/J subscripts to feeding events by adults/juveniles. We also fitted a re-parameterized version allowing us to test for a difference between expected observations of adults and observations of juveniles for each of the three social effects:

The R code used to run these models can be found in Supplementary data⁵³ in ‘GLMM models Orange Yellow final.r’. The main results of each model are presented in Table 2 and full model outputs in Supplementary Tables 3–5.

Social information use during reversal learning

To investigate social information use during reversal learning, we used the order of acquisition diffusion analysis (OADA), a variant of NBDA⁴³, which explores the order in which individuals acquire a behavioral trait⁴⁴. The rate of social transmission between two individuals is assumed to be linearly proportional to their network connection, and the spread of trait acquisition is therefore predicted to follow the network patterns if individuals are using social information. We used NBDA to investigate whether the order of individuals’ first visit to the previously unpalatable blue almonds (mimics) followed the network. We fitted several different models that included (i) only asocial learning, (ii) social transmission of information following a homogeneous network (equal associations among all individuals), or (iii) social transmission of information following our observed network. Models that included social transmission were further divided into models with equal or different transmission rates from adults and juveniles, and from conspecifics and heterospecifics, by constructing separate networks for each adult/juvenile and conspecific/heterospecific combination. To investigate whether asocial or social learning rates differed between blue tits and great tits, we included species as an individual-level variable. We then compared different social transmission models that assumed that species differed in both asocial and social learning rates, only in asocial or only in social learning rates, or that they did not differ in either (see Table 3). The best-supported model was selected using a model-averaging approach with Akaike’s information criterion corrected for small sample sizes. All analyses were conducted with the software R.3.6.1⁵⁴, using lme4⁵⁵, asnipe⁵⁶, and NBDA⁵⁷ packages.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Source: Ecology - nature.com