Introduction

The principal drivers of cognitive evolution have been debated for decades [1–8]. Researchers often fall into two broad camps, focusing primarily on either social or ecological factors. Briefly, the Social Intelligence Hypothesis (SIH) posits that cognitive evolution is principally driven by the informational challenges of navigating a dynamic social environment, such as the need to track, anticipate and respond to the behaviour of social partners, and monitor the relationships of others [6, 7, 9–11]. In contrast, the Ecological Intelligence Hypothesis (EIH) and its variants emphasise informational challenges posed by ecological variables, such as changeable food sources and climatic conditions [2, 8, 12–15]. A large body of research has investigated social and ecological correlates of brain size across a range of taxa, including primates, ungulates, carnivores and birds [3, 16–28]. However, results are often inconsistent and contradictory [29–32]. For instance, Dunbar (1992) [3], found that primate social group size positively correlated with a measure of brain size, which is commonly used as a proxy for cognitive ability. In contrast, DeCasien et al. (2017) [17] found that diet is an important driver of primate brain size but social group size is not. Wartel et al. (2019) [32], on the other hand, found that either diet or social group size could predict primate brain size, depending on where brain size data and predictor variables were sourced from. Although the majority of research interrogating the SIH and EIH has focused on primates, birds have emerged as a major model system in cognitive evolution over the last 20 years (eg. [21, 33–35]). Some species of bird show convergent cognitive performance to primates [33, 35], yet birds have divergent neuroanatomy [33] and differing constraints on brain size, such as those imposed by long-range migration [36]. Here we interrogate the potential pitfalls that arise in the comparative study of cognitive evolution in birds. Moreover, we highlight potential pitfalls of current methodologies that have not yet been investigated in any taxa.

The relationship between brain size and cognitive ability is largely unknown and highly contentious [29, 30, 37, 38]. Nevertheless, studies investigating comparative cognitive evolution very often use some measure of brain size as a proxy of cognitive ability [32]. Most comparative studies of brain size use one measurement of brain size per species, taken either from a single individual or averaged across multiple individuals (e.g. [16, 18, 25, 39–43]). How much estimates differ between datasets, and whether this may influence the conclusions drawn from comparative analyses of brain size, are, however, poorly understood [31]. Moreover, to control for the relationship between brain and body size, most studies of brain size control for body size (‘relative brain size’) [30]. However, although body size measures may have important implications for the conclusions of comparative analyses brain size [44], the degree to which body size estimates varies between published datasets, and how much this may influence results, has yet to be considered beyond primates [44].

The approach that researchers take toward model specification also has the potential to drastically influence results. While most studies utilise similar statistical techniques to test hypotheses of brain size evolution, approaches toward model specification differ. Indeed, some researchers opt to include covariates associated only with the hypothesis of interest (broadly, the SIH or EIH), and either omit (e.g. [43]) or include less detailed (e.g. [16, 21]) variables associated with the competing hypothesis. However, the combination of variables is known to have a substantial influence on the results of primate and carnivora brain size models [27, 32]. In addition, where covariates are sourced from can have a substantial impact on results. For instance, Wartel et al. (2019) [32] showed that changing the source from which covariates (e.g. diet) were collected substantially changed the results of a previous study [17]. Moreover, decisions on how to define variables can be a somewhat subjective decision made by authors, and this may also have a significant influence on results. For example, if some populations of a species are migratory, but most are resident (e.g. as in the jackdaw, Corvus monedula, where only northern and eastern European populations migrate [45]), should the species be classified as migrant, resident, or a different category altogether? How such classification decisions influence model results is, as yet, unquantified.

In this study, we collate data from multiple datasets of brain and body size, which span more than 1000 bird species, to quantify the degree to which estimates vary across datasets (Aim 1). Given that most comparative studies of brain size use only one estimate of brain and body size, variation in estimates has the potential to substantially change results. Second, we use detailed data from the well-studied Corvides infra-order to interrogate whether conclusions drawn from models testing alternative hypotheses for cognitive evolution differ depending on the combination, source and classification of variables included (Aim 2). Together, these investigations allow us to both identify novel pitfalls in the study of comparative cognition, and highlight parallel pitfalls to those previously identified in the field of primate comparative cognition [31, 32, 44].

Methods

Statistical modelling

All statistical analyses were undertaken in R v4.0.2 [70]. We used a phylogenetic generalized least squares (PGLS) modelling framework [71] in the package caper [72], which controls for non-independence of datapoints due to relatedness. We used this method because it is the most commonly used technique in the comparative brain size literature (e.g. [21, 36, 39, 42, 55, 60]). We constructed a consensus tree by downloading 1000 equally plausible phylogenetic trees for the species in our sample from www.BirdTree.org [21]. We used the Hackett rather than Ericson backbone because it is the most recently constructed; however, differences between backbones are small and they tend to produce consistent results [73]. Using TreeAnnotator in BEAST v1.10.4 [74], a maximum clade credibility consensus tree was built from these equally plausible trees. This tree was then used to control for phylogenetic non-independence in the following PGLS models. Lambda was estimated using Maximum Likelihood. Model diagnostics and variance inflation factor (VIF) were checked to ensure assumptions were met and variables were not unacceptably collinear, respectively.

Corvides analysis Here, we tested how variable combination, source (environmental variability: temperature variation or PPC) and classification (partial migrant/resident; cooperative breeder/non-cooperative breeder) changed conclusions. See Table 1 for a summary of model formulations. We used Sayol et al.’s (2016) [21] brain size data only. We chose to use this dataset because only one method was used to measure brain size, and all body mass data came from the same specimens that brain volume was measured from. It is therefore likely to be the most precise data currently available. Using this data, we built three models: an SIH model (brain size in response to body size, cooperative breeding and social foraging), an EIH model (brain size in response to body size, migration, environmental variability and diet breadth), and a ‘combined’ model with all covariates included. For the EIH model, we tested two sources of environmental variation: temperature variation [39] and PPCs [21]. We used temperature variation as the measure of environmental variation in the combined model because the limited number of species with PPC data available resulted in some social foraging categories having extremely limited sample sizes. We tested ‘initial’ and ‘reclassified’ variables for species movement and cooperative breeding across all models. Brain and body size measurements were log-transformed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Qualitative conclusions drawn from models testing different hypotheses (SIH/EIH/combined).

Significant predictors are shown in bold in column 3.

https://doi.org/10.1371/journal.pone.0270771.t001

Results

Aim 1—Quantifying variation in brain and body size between datasets

Altogether we collated brain size data for 1473 bird species. Of these, 1057 species had brain measurements in more than one dataset. Fig 1A visualises variation in log-transformed brain size estimates across datasets. All but one of the collated datasets had one brain size estimate per species; we therefore present García-Peña’s [47] sex-separated data as sex-averaged in Fig 1A. Only two of the five studies that did not measure sexes separately explicitly stated that brain size datapoints were sex-averaged [18, 21]. While brain size estimates tended to be the mean value of multiple specimens (on average, ~six specimens per species), ten or more brain size estimates were based on a single speciman in at least two studies [46, 48].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

https://doi.org/10.1371/journal.pone.0270771.g001

Across datasets, some species varied considerably in brain size estimate (Fig 1A). For instance, the minimum and maximum brain size estimates of the maroon-bellied parakeet (Pyrrhura frontalis) (with the largest overall difference in estimates) overlapped with the brain size estimates of 159 other species in our full sample of 1473 species (thus, the brain size of this species spanned 10.79% of the species in our sample, depending on the dataset used). Similarly, the eastern spinebill’s (Acanthorhynchus tenuirostris) minimum and maximum estimates (with the second largest difference) overlapped with 150 species (10.18% of our sample) and the great tinamou’s (Tinamus major) estimates (the third largest difference) overlapped with 141 species (9.57% of our sample). Variation was not driven by any one dataset in particular (Fig 1A).

Our collated body size dataset contained 2331 species from eleven datasets. 1599 of these species had body mass measurements in more than one dataset. There was a general paucity of details for body mass data. Only one of 11 studies explicitly reported the number of samples from which body mass estimate was calculated [51]. Only two of 11 explicitly stated that body mass estimates were sex-averaged [53, 54]. Ten of the eleven datasets recorded both brain size and body mass data; however, only three of these sometimes collected body mass from the same specimens as brain size [18, 46, 55] and only two always collected body mass from the same specimens as brain size [21, 51]. Three studies did not mention the primary sources of the body mass data used [48, 52, 53].

Fig 1B visualises variation in log-transformed body mass estimates across datasets. Again, some species varied considerably in body mass estimates. For instance, the minimum and maximum body mass estimates of the blue-breasted fairywren (Malurus pulcherrimus), with the largest overall difference in estimates, overlapped with the body mass estimates of 140 other species in our full sample of 2331 species (6.01%). Similarly the three-banded plover’s (Charadrius tricollaris) minimum and maximum estimates (with the second largest difference) overlapped with 759 species (32.56% of our sample), and the Carolina wren’s (Thryothorus ludovicianus) minimum and maximum estimates (with the third largest difference) overlapped with 429 species (18.36% of our sample). As in the brain size data, variation was not primarily driven by any one dataset (Fig 1B).

Aim 2—Influence of variable inclusion, classification and source on model results

The full dataset used for this analysis included 59 species, where all species had known social and ecological variables (excluding PPCs). The PPC subset contained 46 species. Conclusions on the principal drivers of brain size evolution in Corvides differed depending on modelling approach (see Table 1 for a general summary; see Table 2 and Fig 2 for model results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

https://doi.org/10.1371/journal.pone.0270771.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Phylogenetic generalised least squares model results, comparing different model formulations.

All significant pairwise contrasts for categorical variables are presented. Significant predictors are shown in bold.

https://doi.org/10.1371/journal.pone.0270771.t002

In the EIH model with temperature variation included, species movement was significantly associated with brain size. Specifically, resident species were found to have bigger brains than nomadic species. In the EIH model with PPCs (one measure of environmental variation) rather than temperature variation (a different measure of environmental variation), both PPC1 and PPC2 were significantly associated with brain size. However, species movement was not. Thus, depending on where we sourced our measure of environmental variation, we could have concluded that environmental variation drives brain size and species movement does not or, indeed, the exact opposite. Social variables also changed in significance depending on model specification. In the SIH model, species that forage in non-nested small groups were shown to have significantly larger brains than species that forage in pairs. In the combined model, with both social and ecological variables included, solitary foragers and species that forage in non-nested small groups were shown to have significantly larger brains than species that forage in pairs.

Changing facultative or suspected cooperative breeders from their initial categorisation of cooperative breeders to non-cooperative breeders did not qualitatively change SIH model results. However, changing partial migrants to residents did change EIH model results. For the EIH model with PPCs, PPC1 and PPC2 changed from significantly influencing brain size (PPC1: β = 0.01, SE = 0.01; 95% CI[0.001,0.03]; p = 0.03; PPC2: β = 0.03; SE = 0.02; 95% CI[0.002,0.06]; p = 0.04) to having no significant effect (PPC1: β = 0.01, SE = 0.01; 95% CI[0.0001,0.02]; p = 0.055; PPC2: β = 0.03, SE = 0.01; 95% CI[-0.002,0.06]; p = 0.07). In the combined model, where both changed variables were included, solitary foraging (relative to foraging in pairs) changed from having a significant (β = 0.23, SE = 0.10; 95%, CI[0.09,0.37]; p = 0.03) to no significant effect on brain size (β = 0.16, SE = 0.10; 95%, CI[0.03,0.35]; p = 0.11).

Note, however, that although results were unstable in regards to significance, and thus in qualitative conclusions, effect sizes and confidence intervals were not drastically different between models.

Discussion

In agreement with a growing body of literature [29–32], our analyses raise concerns that comparative analyses of brain size studies may not provide robust means of testing hypotheses about cognitive evolution. We show that there is considerable variation in bird brain and body size estimates across datasets (Fig 1), most of which is likely to be due to intraspecific variation. The most common methods of comparative brain size analysis do not take this variation into account, although it has the potential to substantially influence results. The combination, source and classification of social and ecological variables also changed results (Fig 2). Indeed, we could have come to several contradictory conclusions depending on our modelling approach. Our results chime with and add to concerns raised in the primate comparative cognition literature [31, 32] that current methods in the comparative study of cognitive evolution, using brain size as a proxy of cognitive ability, give unreliable results.

Comparative brain size studies that test hypotheses of cognitive evolution typically use brain and body sizes averaged from multiple specimens of a single species to obtain one brain and body size estimate per species. Despite all brain sizes in our sample being estimated either from endocranial volume, or brain mass converted to volume (which has a strong positive correlation with volume [50]), we found substantial variation in brain size estimates across datasets. In several extreme cases, the minimum and maximum brain size measures for one species overlapped with 10% of our sample. Similarly, we found considerable variation between body size estimates of the same species: in extreme cases, the minimum and maximum body size of a species overlapped with more than 30% of the species in our sample.

Variation in brain size was not driven by any dataset in particular, suggesting that it was not the result of a specific methodological approach, and that it is likely to be the result of natural intraspecific variation. Indeed, previous work has found that brain size shows substantial intraspecific variation across taxa, including birds [75]. In some species of bird, sex and age class have been shown to be particularly important predictors of brain size [76]. Nevertheless, key information such as sex and age of specimens, and variation around the single reported estimate, were not reported for most brain size datasets. Body mass estimates had even sparser associated information: in most cases, the number of specimens from which the estimate was derived was not given, and neither was sex (an important covariate of body mass in many birds [77]) or variation around the estimate. The fact that brain and body size estimates tend to be derived from few (or an unknown number of) specimens, often of unrecorded sex and age, raises concerns as to how well they represent species-average values. Our results suggest that variation in brain and body size estimates across datasets is sometimes substantial. While the source of variation between datasets is likely to be due to intraspecific variation in the majority of cases, variation may sometimes be due to errors in the datasets. For instance, the variation between blue-breasted fairywren estimates is likely to be due to a typo in one dataset, where the species’ body mass was recorded as 0.98 grams [18] but should have been an order of magnitude higher (other datasets record this species’ body mass as 9.8 grams). At best, variation between datasets–whether due to natural intraspecific variation or due to errors in data entry–may introduce noise into comparative brain size studies, and at worst may lead to spurious conclusions about the drivers of cognitive evolution. We were not able to quantify the influence of variation between datasets on comparative brain size model results here, due to non-independence of the datasets, but this would be a valuable focus for future work. Another important and related issue is that recent research suggests the relationship between brain and body size is often taxa-dependent [78, 79]. Thus, the popular method of including body size as a covariate in models, in order to control for the relationship between brain and body size across a diverse range of species, may be flawed even when brain and body size estimates are accurate.

As well as investigating variation in brain and body size estimates across datasets, we used data for the Corvides infraorder to interrogate how robust models are depending on variable combination, source and classification. Researchers typically have varying approaches to model building, and in agreement with Wartel et al.’s (2019) [32] analyses of primate brain size data, we found that modelling approach influenced results. For instance, depending on whether we chose to include social variables or social and ecological variables, we could have concluded that there is no difference in brain size between species that forage in pairs and species that forage solitarily or that species that forage solitarily have significantly larger brains than those that forage in pairs. In addition, we showed that the source of covariates has the potential to substantially change results. Using temperature variation from Fristoe et al. (2017) [39] as a proxy of environmental variation resulted in no support that environmental variation drives the evolution of bigger brains. Meanwhile, using more detailed measures of environmental variation from Sayol et al. (2016) [21] resulted in support. Note, however, that models with temperature variation rather than PPC had a larger sample size, which may influence these results. Nevertheless, these findings parallel those reported in the primate brain size literature [31], where using differing variable sources resulted in differing results even when sample sizes were matched.

In addition to the previous issues, we also show that variable classification can drastically influence results. Classifications of variables are sometimes subjective; for instance, species with both cooperatively and non-cooperatively breeding populations could be classified as either. We therefore changed categorical variables that could justifiably be re-classified, and tested how this influenced results. Re-classifying suspected/facultative cooperative breeders as non-cooperative breeders did not change SIH model results; however, re-classifying partial migrants (i.e., where at least one population of a species migrate) as residents substantially changed EIH model results. While two measures of environmental variation (PPC1 and PPC2) were significantly associated with bigger brains before re-classification, there was no significant effect following re-classification. Thus, depending on the classification of a different model covariate, we could have concluded that environmental variation is associated with bigger brains or that there is no association. It must be considered, however, that while P-values crossed the threshold of significance, estimates and confidence intervals for predictor variables did not substantially change. With larger sample sizes, models may be less volatile. Nevertheless, many studies of brain size evolution use sample sizes in the same range as ours (e.g. [22, 25, 36, 80, 81]). The concerns raised here are thus pertinent to a wide range of studies.

While our study focused on highlighting problems with current methods of using brain size to test hypotheses of cognitive evolution, there are also conceptual issues with such studies. A key and often overlooked point is that the relationship between relative brain size and cognitive ability is not clear [29, 30, 38]. Thus, even if studies show a strong correlation between specific variables and brain size, it is difficult to draw strong conclusions about how this translates to selective forces acting on cognition itself. In addition to this rather fundamental issue, we add our voices to a growing number in the field who argue that framing the SIH and EIH as dichotomous and competing hypotheses is not logically sound. The hypothesised underlying driver of cognitive evolution for both hypotheses is variation in environmental conditions (including social environment) in which individuals must gather and process information to mitigate uncertainty [82, 83]. Thus, according to theory, the foundational driver of cognitive evolution is the same between both the SIH and EIH. Moreover, social and ecological variables are not independent, i.e., social species solve ecological problems in a social context, and sociality itself may evolve in response to ecological variables [84, 85]. We therefore suggest not only that our methodological approach to studying comparative brain size evolution needs to change, but also the conceptual framework itself. Rather than splitting often-correlated variables into dichotomous and competing predictors, we could benefit from quantifying the environmental uncertainty animals face in specific contexts and examining how this may drive cognitive evolution.

When considering the accumulating literature on issues associated with current methods in the comparative study of cognitive evolution using brain size (here; [29–32]), we add our voices to a growing number in the field suggesting that we must (i) use caution when interpreting the results of studies using the same methods as presented here, and (ii) start to move away from these methods when interrogating hypotheses of cognitive evolution. It is important to note that the work done by previous researchers using the methods available at the time have helped us to form a deeper understanding of which variables may influence brain size evolution, and is key to guiding our path forwards in regards to what future research should endeavour to focus on. As more data become available and methods improve further, broad-scale comparative brain size studies will still offer novel and valuable insights into cognitive evolution [86]. However, we suggest that for a more nuanced understanding of the drivers of cognitive evolution we must also test how uncertainty in the social and ecological environment influences cognitive performance at the intra-specific level [84, 87–89], and between closely related species [90–94]. Conceptually, we also recommend a shift away from the treatment of the SIH and EIH as dichotomous and competing hypotheses. We propose that instead, we should work to understand whether and how uncertainty across a range of contexts drives cognitive evolution.

Supporting information

Acknowledgments

Thank you to Matilda-Jane Brindle for help with creating Fig 1.