Researchers have confirmed that there is a general intelligence factor in dogs — some dogs are more equal than others:
Our results indicate that even within one breed of dog, where the sample was designed to have a relatively homogeneous background, there is variability in test scores. The phenotypic structure of cognitive abilities in dogs is similar to that found in people; a dog that is fast and accurate at one task has a propensity to be fast and accurate at another. It may seem obvious that once a detour task (finding the treat behind a barrier) has been solved in one form, the solution to the other forms will follow naturally, but dogs are not people. Experiments have shown that dogs’ problem-solving skills do not transfer readily from one problem to a different form of the same problem as ours do (Osthaus, Marlow, & Ducat, 2010). The g factor we report is consistent with the prediction made by the many experts in the ‘dog world’ (trainers, veterinarians, members of dog societies, and farmers) who were consulted in the early stages of this study. Those experts said that in their experience some dogs were more likely to catch-on, learn and solve problems more quickly than others. Our results show structural similarities between canine and human intelligence. Individual tests have some test-specific variance, tests are influenced by a group-level factor, and the group-level factor is influenced by a g factor. We tested models without the g factor, without the group-level factors and with uncorrelated group-level factors; models positing correlated group-level factors (the unstructured model and the hierarchical g model) fit the data. We emphasize the hierarchical g model because the poor fit of the no-g model rules out uncorrelated first-order factors; the hierarchical g allows us to examine how those correlations arise.
Although we cannot calculate empirically the impact of range-restriction (of intelligence) on our results we surmise that our sample of farm dogs is somewhat analogous to a human university student population because farm dogs at the low tail of the intelligence distribution are more likely to be given away as companion animals. Range restriction attenuates correlations (Alexander et al., 1984 and Wells and Fruchter, 1970) so we cautiously interpret the g factor we found as being a low estimate of commonality. A plot showing the possible impact on our results given various estimates of range restriction is given in the Supplementary Information together with the zero-order correlation matrix for all test scores.
Noise may arise from variation in appetite for treats. We assume that dogs vary in their appetitive motivation—and that differential interest in food treats may be confounded with test scores. Our finding that speed and accuracy are positively correlated suggests that this has not been a major concern, yet we expect that performance on a problem-solving test is affected by more than just ‘smarts’. Affective traits such as motivation, persistence, and so on likely influence performance on cognitive tasks, but if they contribute to covariance among tasks, it may be hard to distinguish these aspects from g; there is no a priori reason why g should not have an affective component. The crucial point is that our study investigates the covariance, the structure, among test scores. In humans where g has been most studied, g arises among mathematical and vocabulary tests even though students often have different preferences and motivation to do these kinds of tasks. If g tapped motivation heavily, we would expect to see covariance among measures of motivation across different kinds of test; in humans we do not see this (Loken, 2004).