A meta-analysis approach is the ideal method for understanding NPA animals’ group performance when the findings are discrepant across studies. The strength of a meta-analysis is that it allowed us to run statistical tests on the results. By taking heterogeneity across studies (variability in effect sizes) into account, meta-analyses provide a general and combined estimate of NPA’s performance in the opt-out task. We also included a moderator analysis to examine the variability attributed to some key factors (e.g., species) across studies.

Our aim was to (1) estimate the overall effect size for the chosen-forced advantage, opt-out rate, and composite score across different NPA species; and (2) understand how the factors “species”, “task order” (prospective vs. retrospective), and “task domain” (memory vs. perception) would influence these animals’ performances.

The following information was extracted from each of the included studies: author(s) and publication year; type of publication (journal or dissertation); species; test domain (perception or memory); test order (prospective or retrospective); sample size; chosen-forced advantage; opt-out rate; data status (collected or estimated).

To allow cross-species and cross-experiment comparisons, we evaluated all of the extant data using the metric composed of chosen-forced advantage and opt-out rate. For each individual animal, the chosen-forced advantage is calculated by obtaining the average “percentage correct on chosen trials” minus “percentage correct on forced trials” on all difficulty levels. We expected the chosen-forced advantage to be statistically higher than zero if NPAs were able to use the opt-out option when they feel uncertain. The opt-out rate is calculated as “percentage of tests declined at the highest difficulty level” minus “percentage of tests declined at the lowest difficulty level”. Note that the magnitude of the opt-out rate is tied to the distribution between difficult and easy conditions. While the degree of difficulty varies across studies, it generally falls within a range of levels (e.g., 1 to 4). These levels were systematically calibrated to align with the cognitive capacities of the animals being studied. Therefore, in a within-subjects context, the across-study differences in this regard do not impose biases in the effect size estimation. The composite score is the average of an individual animal’s chosen-forced advantage and the opt-out rate. We extracted the average rate for these three scores, as well as the standard error of the means among the individual animals within-study for the meta-analysis.

The Comprehensive Meta-Analysis software version 2 (Borenstein, 2005) was used to synthesize the data and perform the statistical analysis. Due to the heterogeneity in sampling methods, assessment instruments, and sample size across studies, a random-effects model was used to estimate the effect size for the three scores. Cochran’s Q and the I² statistic was used to assess the degree of heterogeneity across included studies. Subgroup analyses were used to examine the sources of heterogeneity and the key moderators that contributed to the heterogeneity. Publication bias was evaluated with the funnel plots and Egger’s test.

Moreover, we obtained the species’ absolute brain weight and number of neurons in the brain from the published source (Herculano-Houzel, 2011) and ran correlation analyses correlating the species’ chosen-forced advantage, opt-out rate, and composite score with their brain weight and number of neurons in the cortex.

To delve further into the factors that lead to the high heterogeneity, we conducted subgroup analysis for the three scores. We found that the three key factors (Order, Species, and Domain) did not systemically play a significant role in modulating the heterogeneity. The factor “Domain” (memory/perception tasks, Q=6.163, p=0.013, Table 2 top) and “Order” (prospective/retrospective in opting-out; Q=4.655, p=0.031, Table 2 middle) showed some effects on heterogeneity, but the factor “Species” did not impose any statistical significance ( Table 2 ). Taking the composite score as an example (bottom Table), the p-values for “Heterogeneity” are 0.109, 0.944, and 0.411 for the three factors respectively; none of them was statistically significant. This implies that the three factors did not affect (modulate) the heterogeneity of our meta-analysis significantly. Note that the p-values for the factor “Species” for heterogeneity are also not significant for chosen-forced advantage (top table) and opt-out rate (middle table) either. This led to the conclusion that the effects of species differences on heterogeneity were not statistically significant.

To examine if publication bias existed in the study, we used funnel plots for chosen-forced advantage, opt-out rate, and the composite score; a symmetric plot would indicate a lack of publication bias (Sterne et al., 2005). Egger’s tests were also performed to examine whether the assumption of a symmetrical distribution of effect sizes is viable; a p value greater than.05 would suggest a lack of sufficient evidence for publication bias (Egger et al., 1997). Egger’s tests revealed that there was no publication bias for chosen-forced advantage (t=1.246, p =0.244; Figure 4 top panel), opt-out rate (t=1.034, p =0.331; Figure 4 middle panel) and the composite score (t=1.859, p =0.096; Figure 4 bottom panel).

These results may be at odds with other researchers who claim that the prefrontal cortex or a well-developed neocortex is a biological prerequisite to such high-level cognition. Some scholars also hold the belief that high-level metacognition involves consciousness (Dehaene et al., 2017), an ability closely linked to human and primates’ prefrontal cortex. Considering that these NPAs are not thought to possess a well-developed PFC, our results seem to be contrary to these claims. However, recent research shows that despite the lack of layered neocortex, large areas of the avian forebrain are homologous to mammalian cortex (Jarvis et al., 2005; Pfenning et al., 2014; Güntürkün and Bugnyar, 2016) and play similar roles in higher cognitive functions (Kirsch et al., 2008). Similarly, complex brain functions can emerge from simple neural circuits such as honeybees’ mushroom body (Caron and Abbott, 2017) and miniature brains can complete a range of cognitive operations (Chittka and Niven, 2009; Chittka, 2017).

In light of these possibilities, we tested if meta-ability is correlated with the animal’s absolute brain weight and number of neurons in the brain (Herculano-Houzel, 2011), as these factors are known to be a predictor for some high-level cognitive abilities (Deaner et al., 2007; MacLean et al., 2014). We ran a series of correlation analyses correlating the species’ chosen-forced advantage, opt-out rate, and composite score with their brain weight and number of neurons in the cortex, respectively. We found that there is no significant correlation between species’ meta-scores, absolute brain weight, and neurons in the brain (all Cor < 0.103, all ps > 0.051). We then ran correlation tests between species’ meta-scores and their brain-to-body ratio, which also yielded nonsignificant results (all Cor < -0.183, all ps > 0.114). This suggests that brain volume per se is not sufficient in accounting for the species’ meta-ability. Rather, the existence of meta-ability could be due to other factors, such as interneuronal distance and axonal conduction velocity (Dicke and Roth, 2016), fundamental neural circuits that are similar across species (Chittka and Niven, 2009) or higher neuron packing densities in the forebrain (Olkowicz et al., 2016).

Our findings from the meta-analysis that four NPA species passed the uncertainty monitoring test at the group level carry theoretical implications for the field. In the modulator analysis, we found that the factor “Species” for heterogeneity is not significant for the chosen-forced advantage, the opt-out rate, nor for the composite score. This leads to the conclusion that the effects of species differences on heterogeneity are not statistically significant in this study. However, to avoid overgeneralization, we maintain that our conclusion should only be applicable to the few species included here.

In the meta-analysis, we found that test order played an important role in leading to heterogeneity across studies. This result is in line with previous findings that animals are better in performance for retrospective tests than for prospective tests (Morgan et al., 2014; Siedlecka et al., 2016). Interestingly, we did not find the factor species to be a determinant factor leading to the heterogeneity of animals’ performance, and this is in alignment with the emerging evidence that metacognitive-like abilities detected from animals from various distant families. One study on pharaoh’s ants found that individual ants spontaneously upregulate or downregulate pheromones depending on the reliability of their own memories (Czaczkes and Heinze, 2015). This demonstrates certain metamemory-like ability in ants, echoing other insect studies wherein honeybees were found to be able to monitor their uncertainty (Perry and Barron, 2013). It is possible that the evolution of metacognition is not a representation of a linear sequence of cortex-dependent evolution, but rather as representatives of different clades that diverged at different points during evolution.

We acknowledge the potential influence of low-level mechanisms on uncertainty response. While our study suggests that the four NPAs exhibited metacognitive-like abilities, we cannot entirely rule out alternative explanations involving lower-level mechanisms and/or associative learning or other mechanisms that do not require invoking introspection or access to private internal states (Hampton, 2009). Therefore we should be cautious in the interpretation of the significance of our findings in advancing our understanding of cognitive processes in NPAs.

Moreover, we note that the selection of these four species in our uncertainty monitoring meta-analysis was primarily driven by data availability. While we acknowledge this limitation, our choice aimed to represent diverse taxonomic groups (mammals, birds, insects) and comparable experimental paradigms. This allowed for preliminary insights into metacognitive-like abilities across different species. We recognize the need for further research beyond these four species to expand our understanding of metacognition in non-primate animals.

There are several caveats with the analyses and our interpretation. Firstly, this all-in-one meta-analysis could not take into account the variations in individual studies, such as factors related to rewards, training intensity, and each task’s temporal variations. Secondly, the mechanisms of how the two scores operationally complement each other remains unclear and therefore we remain tentative in how much information might be missed out by considering these two scores separately. The meta-like performance on these two paradigms remains debatable as the measures do not unambiguously reflect second-order (i.e., metacognitive) computations. To address this issue, a number of transfer tasks related to both paradigms have been developed in the field. In these transfer tasks, animals were tested on whether they could transfer the concept of opting out or information seeking to a novel task. The impetus for developing such tasks is that if animals are able to pass a transfer task following a primary task, this would indicate that their uncertainty monitoring ability and information-seeking behavior represent a task-independent cognitive state (Washburn et al., 2006). For example, researchers have tested if uncertainty monitoring ability is transferable to novel stimuli in pigeons (Sole et al., 2003; Nakamura et al., 2011), bantams (Sole et al., 2003; Nakamura et al., 2011), and bees (Perry and Barron, 2013). Five out of six pigeons and two out of three bantams in Nakamura’s study (Nakamura et al., 2011) generalized their uncertainty responses to novel stimuli at least once. Four of the ten bees could transfer the concept of opting out to a novel task (Perry and Barron, 2013). Other researchers have also tested if pigeons’ information-seeking behavior (Castro and Wasserman, 2013; Iwasaki et al., 2013) is transferable to new tasks. One out of four pigeons in Castro and Wasserman’s (2013) study successfully transferred the information-seeking behavior into two transfer tasks. One out of 2 pigeons in Iwasaki’s study (2013) transferred the information-seeking behavior from a hint-seeking serial learning task to a hint-seeking visual search task.

In these tasks, subjects demonstrated a generalized mechanism whereby a first-order (cognitive) representation is internally assessed through a second-order (metacognitive) process that directly evaluates its quality, although it remains unknown if such a mechanism contains introspection. Such performance has prompted theorists to land a “middle ground” between the low level and the high level to discuss animals’ meta-behavior. Metcalfe and Son (2012) layer three levels of metacognition. The middle level is referred to as “noetic metacognition”, which allows an organism to be aware of, and to cognitively operate on, objects and events, as well as relations among objects and events, in the absence of those objects and events. Smith et al. (2012) advocates that animals’ uncertainty responses are not appropriately considered associative responses in the traditional sense. They can be independent from stimuli, reinforcement, and so forth. Similarly, Beran (2012) proposes that a metacognitive process can represent a first-order state, without acknowledging that this state has a certain representational function. It is possible that animals, like humans, are capable of making uncertainty judgments based on internal cues without awareness of meaningful self-reflection. Uncertainty monitoring capacity could emerge phylogenetically as a gradual feeling of the indeterminacy response from its primary responses (Beran et al., 2009). The middle ground emphasizes the functional feature of metacognition rather than true self-reflection and has allowed us to appreciate NPA’s metacognition performance in a more realistic way. It opens up possibilities to interpret the existing NPA metacognition literature in relation to their evolutionary significance.

There has been a great amount of discussion regarding the theoretical implications of these approaches. Different levels of metacognition have been proposed by scholars to describe animals’ meta behaviors, such as object versus meta level (Shields et al., 2005), first-order versus second-order (Crystal and Foote, 2009), public mechanism versus private mechanism (Hampton, 2009), anoetic metacognition, noetic metacognition, and autonoetic metacognition (Metcalfe and Son, 2012), and low-level versus high level (Smith et al., 2012). Although NPAs can pass these two paradigms and certain transfer tasks, there are alternative explanations that can explain the results without implicating metacognition per se (Carruthers, 2008; Jozefowiez et al., 2009; Le Pelley, 2012). One way to help support the metacognitive argument |is to follow the emerging trends in human and non-human primate metacognition research to measure an individual’s metacognitive sensitivity in a more quantitative way. A widely used approach to measure human and non-human primates’ metacognition is to calculate the comparison between the type 1 sensitivity (d’) and the type 2 sensitivity (meta-d’). This approach can quantify meta-ability under the signal detection theory framework (Maniscalco and Lau, 2012) or by a recently developed hierarchical Bayesian estimation method (Fleming, 2017). To attain these analyses, confidence ratings are needed. This subjective confidence measurement is deemed practically very difficult in animals but there are now viable ways to attain these, such as temporal wagering tasks (Kepecs et al., 2008; Lak et al., 2014; Cai et al., 2022), in which animals express choice confidence by the amount of time they are willing to wait for reward. This has been demonstrated to be a successful paradigm in mice (Lak et al., 2014), rats (Stolyarova et al., 2019), and monkeys (Cai et al., 2022). It has become a hot topic to examine whether metacognition in the primate shows domain specificity, with supporting evidence for both domain-specific and domain-general metacognitive representations (Morales et al., 2018; Ye et al., 2018; Kwok et al., 2019; Zheng et al., 2021; Cai et al., 2022). This aspect will provide an important line for further investigation in the NPA species. Finally, in order to reduce publication bias, standardize research protocols, and ultimately expand the data volume in this niche field, researchers in the field could benefit from taking the ManyPrimates project model as well as other similar initiatives as an approach to establish an infrastructure for large-scale collaboration in non-primate animals cognition research (Altschul et al. 2019; ManyPrimates et al., 2019; Milham et al., 2022).