In the current study, we examined associations between owner perceptions of the health of their dog and a range of owner-related, animal-related and healthcare variables. Predictor variables that were most strongly associated with our outcome variables were dog age and healthcare variables (including number of veterinary visits and receiving medication), whilst the contribution from other variables was more limited, and the effect of dog diet was negligible. We utilised data from a previous study, where the opinions of owners about the health of their dog was gathered by questionnaire [15]. The primary aim of that study was to examine associations between owner opinions of dog health status and the type of diet that owners predominantly fed, with the key finding being that health status was positively associated with feeding either a vegan or raw diet, compared with other diet types (including conventional and vegetarian). Given the focus of that study, other information gathered in the questionnaire was not assessed and no account was taken of possible confounding amongst variables. Therefore, to extend the findings of the previous work, additional owner, animal and healthcare variables were studied. Previous work has reported associations between such variables and owner decisions about feeding. For example, animal variables, such as age and neuter status, were associated with the owner feeding choice in the previous study of Knight et al. [15]. In the same study, maintenance of pet health was the most-common reason cited by owners for choosing a particular food [15] whilst, in other research, owner characteristics (such as geographic location) are also associated with food choice [65]. Finally, owner characteristics can also be associated with aspects of dog health; for example, both owner age and income are associated with the prevalence of obesity in dogs [66]. Considering these previous findings, our decision to examine variables beyond those of diet type was justified.

To ensure that owner perceptions of dog health were examined, we created two outcome variables from a question in the original survey question where owners were asked to rate their dog’s health over the previous 12 months [15]. Such responses are not objective measures of health; rather, they require owners to formulate a subjective holistic opinion, which will necessarily require them to take many factors into account (Fig 9a). There will be a contribution from the actual health of their dog (which in turn is associated with other factors), but this will be influenced by their knowledge and recollection, whereas general opinions on what constitutes ‘health’ will be affected by owner attitudes, beliefs and behaviours.

This study also differed from the original study because we excluded dogs <1y because numbers were small (n=26) and, arguably, insufficient to reveal insights into how the perceptions of owners with dogs in their growth phase might vary. We were also concerned about the small numbers (n=111) of respondents who were not make the primary decision about their dog’s diet. Some of these owners reported playing “some lesser role”, whilst others reported playing “no role” in diet decision-making [15]. Not only would this raise concerns about the accuracy of the diet information, but these owners might also be less involved in other aspects of the dog’s life, including health and veterinary matters, which might adversely affect the accuracy of other information provided (Fig 9a). There might also be differences in their opinions and attitudes towards health matters in general. Analyses performed using decision-maker status as an outcome variable supported these concerns, not least since the odds of a dog being recorded as having a health problem were less for owners who were primary decision-makers than for those who were not. One way to account for this would be to include decision-maker status as a covariate in all analyses but the marked imbalance in group size (2,211 vs. 111) could adversely affect the performance of procedures such as logistic regression. Therefore, we instead decided to limit further analyses to owners who were primary decision-makers only. Decisions about excluding observations from a dataset should be made with caution, properly justified and consider the possible impact on results. Based upon our flow diagram of associations within in the dataset (Fig 9b), we determined that it would be possible to eliminate all pathways associated with the decision-maker status variable, thereby simplifying the overall structure without affecting any other associations amongst variables. This was because removing all data from owners who are not primary decision-makers effectively ‘collapses’ the decision-maker status variable into a single category such that variance is no longer associated with it. Estimates of regression coefficients and odds ratios were broadly similar in logistic regression models using the all-owner dataset (S1 Table, S2 Table), compared with models using the primary decision-maker dataset (Tables 2 and 3). Any differences are likely to be the result of differences in the characteristics, attitudes and beliefs of the owners that were not primary decision-makers. Of course, one main disadvantage of focusing on data from owners who are primary decision-makers is that it limits the generalisability of findings to the wider pet-owning public, not least to other members of a dog-owning family. However, given the original concern about data reliability, further studies involving all family members would be required to explore such opinions properly.

The approach of the current study also differed by our decision to include healthcare variables (received medication, veterinary visits, switched to a therapeutic diet) as predictors. In the previous study [15], these variables were instead used as proxy measures for dog health and compared amongst dogs fed different diet types. Whilst both owner perceptions of health and healthcare variables will be associated with actual dog health, they are materially different. As already discussed, owner perceptions of health were derived from a question where owners subjectively rated their dog’s health [15], and these perceptions will be influenced by other factors including owner knowledge and recollections, as well as their beliefs and attitudes (Fig 9). Conversely, the healthcare variables were derived from survey questions that required factual recall rather than the formulation of a subjective, holistic opinion about health. For example, owners were asked whether their dog “had received medications in the last year” and how many times the dog had “visited a veterinarian or received a home visit” [15]. Given that such healthcare information will contribute to the knowledge an owner utilises when constructing opinions about the health of their dog, we reasoned that they were valid predictor variables in our statistical modelling. Further, excluding them from statistical models could be problematic as illustrated in Fig 9c; unlike removing observations from owners who were not primary decision-makers (as discussed above), where any variance associated with the variable is removed, simply excluding a variable from a statistical model will not eliminate any associated variance. The ‘hidden’ variance associated with the healthcare variables might then be incorrectly attributed to other variables leading to erroneous estimates of their regression coefficients. Any unassigned variance also adversely affects model performance, which likely explains why the models that only included owner-animal metadata performed poorly as discussed below.

Using simple logistic regression, we identified associations between owner perceptions of dog health and several owner (e.g., location, diet), animal (e.g., age, breed, neuter status) and healthcare (e.g., veterinary visits, received medication, switched to a therapeutic food) variables. Further, owner perceptions of dog health were associated with the dog diet variable in simple logistic regression (both raw diet and vegan diet associated with the any health problem variable; raw diet associated with the serious illness variable). These findings are not surprising, and would be expected because they were the result of univariable analyses (simple logistic regression) and, therefore, were broadly similar to the univariable analyses (e.g., odds ratios, one-way ANOVA) conducted in the previous study [15]. In the current study, and as described in the methods section, data pre-processing was necessary to ensure that our statistical analyses were valid. The approach taken was to remove or combine groups where numbers of dogs was small, to ensure better balance amongst categories within a predictor variable. The fact that the results obtained from our univariable analyses (simple logistic regression) were similar to those of Knight et al. [15] suggests that, despite this pre-processing, the final dataset remained representative of the dataset from which it was drawn.

We next used a combination of multiple logistic regression and machine learning predictive modelling so that multiple variables could be analysed concurrently. This enabled the creation of models that best predicted owner perceptions of dog health, and to determine the relative importance of variables contributing to the final models. We chose this combined methods approach to maximise the benefits of each whilst minimising their disadvantages. Multiple logistic regression is a well-known, widely available technique that is familiar to most people and has been extensively used in veterinary studies including those in dogs [67–70]. Procedures for supervised model selection are well established, such as those based on BIC [60,61], which can reduce the risk of model overfitting and can take relevant prior knowledge into account. For example, in our exploratory analyses, we tested interactions with a possible clinical (e.g., between the veterinary visit and received medication variables), biological (between the dog sex and neuter status variables) or epidemiological (between the location and setting variables) and psychological (between owner diet and dog diet variables). One disadvantage of our manual selection method is that it could introduce bias into the selection process. Other disadvantages include the fact that logistic regression makes certain statistical assumptions (e.g., independence of errors, the requirement for a linear relationship between the logit of the outcome and predictors, absence of multicollinearity and lack of strong influential outliers; [71]) which, if not met, could lead to poor prediction accuracy. Key advantages of machine learning methods are the fact that they make fewer prior assumptions about data distribution and, therefore, can better model complex, nonlinear relationships between the outcome and predictors even when there is significant noise [72]. However, there are disadvantages including requiring a large dataset, the time and resources needed for computation, challenges with interpretation of the models produced and susceptibility to errors [73]. In this respect, any machine learning procedure is susceptible to overfitting, although this depends upon context; models are often intended to be generalisable to a wide range of datasets beyond the dataset that they were trained on. Arguably, this would be less of an issue for the current study because many of the variables of significance were identified using both multiple logistic regression and machine learning predictive modelling, whilst their relative contribution to model prediction were broadly similar (see below). Therefore, despite the limitations of the different modelling techniques, our findings should be valid.

Using multiple logistic regression, the main variables associated with dog health in the best-fit models were dog age, veterinary visits, and whether the dog had been prescribed medication; in some models, whether the dog had been switched to a therapeutic food was also of significance, especially when owners described their dogs as “generally healthy, but with minor or infrequent problems” (e.g., binary logistic regression on the any health problem variable). Depending on the model used, these variables explained between 34% and 52% of variance within the dataset, with very little additional variance (2% to 6%, depending on the model) gained by the addition of any or all other predictor variables. Many of the same variables were also identified using XGBoost, and prediction accuracy (as measured by AUC) was similar, but other predictor variables were also identified including owner age and education. However, in contrast to the findings of the previous study [15], the diet of the dog had little to no effect in all models, except for those that only included owner-animal metadata, albeit that these models performed poorly. Therefore, we could not find convincing evidence that particular food types (conventional, vegan, vegetarian or raw) have any clinically meaningful positive or negative association with dog health status. As with the original study [15], these analyses were limited by the dataset used, including its cross-sectional design, meaning that causality of associations cannot be assumed, and its reliance on subjective owner perceptions of health status, which might not reflect the actual health of the dogs. Future studies should be considered, for example, cohort studies or randomised trials utilising objective measures of health such as information from electronic patient record as or formal diagnoses of disease by a veterinarian.

A key limitation of using questionnaires to ascertain the perceptions of owners about dog health is that such information might be affected by owner recollection leading to errors in classification of disease status, so-called response bias [74]. Responses might be biased for various reasons including acquiescence, courtesy, the order of questions or even the perceived social desirability of the responses [74]. Recall bias is a form of response bias whereby the recollections of the owner are affected by their past experiences [75]. Owners of dogs with health problems might search their memories more thoroughly for possible risk factors (such as diet choice, number of visits to the vet, use of medication etc, diet changes) than owners of dogs that do not have health problems, with the potential to exaggerate any associations identified.

As discussed above, it is also feasible that the attitudes and beliefs of owners might either consciously or unconsciously have influenced responses about the health of their dog, a point that the authors of the previous study emphasised [15]. For example, owners who believe a particular type of diet to be optimal for dog health, might be more likely to perceive their dog to be healthy, whether or not this was objectively true. Whilst such a bias would feasibly affect any diet type, a greater bias might be expected with diets perceived to be ‘unconventional’, again, as acknowledged in the previous study [15]. Although increasing in popularity [1], both vegan diets and raw meat diets are still uncommon choices for owners, and many veterinary professionals either do not recommend them or might even advise against their use. Current evidence suggests that both such diet types are often not formulated appropriately to be complete and balanced for essential nutrients [5,8,9–11]. Further, there are also concerns with feeding raw food given the potential for contamination either with bacteria of possible pathogenic potential or bacteria resistant to antimicrobials, both of which might pose a risk to the health of owners and other in-contact people [11–13]. Given such health concerns, owners who make an active choice to feed either a raw or vegan diet could be more defensive about this diet choice, compared with owners who feed conventional diets, and this attitude which might unconsciously have biased responses about the health of their dog, for example, minimising the significance of any illnesses. In the original study questionnaire, the authors did attempt to minimise the influence of canine diet variables on reported health status, by placing the health-related questions before most other variables [15]. However, it is unclear as to whether this would adequately eliminate pre-existing unconscious biases resulting from diet choice.

One variable that featured consistently in all models was dog age, whereby illness severity worsened as age increased, especially when in dogs that were ≥6 years. Perceptions of health in senior dogs are likely to be associated with quality of life, which is known to be negatively associated with age [76]. However, it might also be because many chronic diseases have associations with increasing age, including osteoarthritis [68], obesity [77], diabetes mellitus [70] and many forms of cancer [78]]. It is noteworthy that such chronic diseases are also negatively associated with quality of life [79–82]. An alternative explanation for the negative association between age and health in the current study is the fact that owners of older dogs perceive their health to be poorer simply because they expect older dogs to be less healthy [83].

Owner perceptions about dog health were also positively associated with them reporting that their dog had received medication in the last 12 months. Once again, the observational study design means that causality cannot be assumed. It is unclear whether such dogs are receiving these medications because they are genuinely less healthy (which is why they were prescribed), whether the knowledge that their dog is receiving medication means the owner perceives their dog to be less healthy or both (Fig 9). Finally, the association might be indirect, if both the outcome and predictor variable were associated with a third variable. In this regard, there was also a strong positive association between owner perception of health and the frequency of veterinary visits, whilst a positive correlation between frequency of veterinary visits and receiving medication was also evident, meaning that all three variables are likely to be inter-related. Further, in some models, another healthcare variable (switching to a therapeutic diet) potentially associated with frequency of veterinary visits was also included. Besides these, several additional associations were identified with XGBoost, but not with logistic regression, the most important of which were owner age and education. Owners that were older and those with a higher level of education perceived their dogs to be healthier than did other owners. Further studies would be required to explore all such associations further and determine their true significance to canine health.

Although the study was large, the population studied was not representative of the typical dog-owning public. In this respect, 33% and 13% of owners reportedly fed their dog raw and vegan diets, respectively, which is greater than expected; for example, in a recent UK survey, the proportion of owners feeding raw and vegan diets were 7% and <1%, respectively [1]. Further, the proportion of the UK human population reportedly consuming a vegan diet is estimated to be between 2 and 3% [2] whilst, in the current study, 23% of owners reported consuming a vegan diet. This raises a concern about how representative this study population as a whole, and is of particular significance given the strong clustering based on owner diet. Although the reasons for such an imbalance is not clear, it might well have resulted from the method of owner recruitment. In this respect, the study was widely advertised via social media but, given concerns that owners feeding dogs unconventional diets might be under-represented, relevant interest groups were actively targeted and invited to participate. This strategy was successful in ensuring that adequate numbers of dog owners feeding unconventional diets were included, ensuring that statistical comparisons were meaningful in both the previous [15] and current studies. However, the recruitment strategy might have inadvertently generated an unrepresentative study population which, in turn, may have influenced the results obtained, as discussed above.

In addition to the unrepresentative study population, multiple associations amongst groups were identified with the potential for confounding amongst variables. For example, feeding a raw diet was more commonly reported by owners from the UK, whilst feeding a vegan diet was more common in owners from Europe. Further, there were positive associations between owner income and education, and between dog age and neuter status. Including healthcare variables in multiple logistic regression analyses might also have hidden the associations seen with the owner-animal metadata, including diet type. In logistic regression analyses, we accounted for these associations by determining multicollinearity using VIF, refining models where this might be a problem. No significant issues of multi-collinearity amongst predictor variables were identified in any of the final models. Whilst this does not necessarily guarantee that variance from each factor in the models was correctly assigned, any errors are likely to be small. The use of XGBoost would further address these concerns given that these techniques do not make prior assumptions about data distribution and can better take account of complexity in interactions amongst variables [69]. Nonetheless, to ensure any owner-animal metadata effects were not unfairly penalised by the inclusion of the healthcare variables, we also examined models containing only the owner-animal metadata. In these models, dog age remained as the variable of most importance, but other variables of lesser importance were identified that had not been seen on the original models (including breed size category, owner age, education level and urban setting), whilst a limited effect of diet remained (i.e., the vegan diet variable was only the 12^th most important predictor variable). The significance of any associations between these variables owner perceptions of health are not known. Nevertheless, since predictive accuracy of all models (both multiple logistic regression and XGBoost) utilising just the owner-animal metadata was poor, these analyses should be considered exploratory and any conclusions drawn should be made with caution.

A final approach we took, to ensure that the we fully accounted for any possible effect of the dog diet vegan variable, was to add it back into the best-fit models for both the any health problem and serious illness outcome variables. In both cases, this additional predictor variable did not improve the performance of either model and, interestingly, models with equivalent performance could be created when the owner diet vegan was instead added.

The main limitations to the current study have already been discussed above, including the fact that owner-reported health was the main outcome measure and the fact that the study population was not representative of the general dog-owning population. Further, significant data pre-processing was required to ensure adequate group sizes for statistical analyses. The disadvantage of such an approach was that many smaller groups needed to be excluded, meaning that we might have missed some variables with a potential impact on owner-reported canine health. That said, in preliminary testing, we coded several of our variables in different ways (e.g., dog age as a continuous or ordinal 5-category variable; veterinary visits as a 4- or 5-category ordinal variable), and selected the approach that produced the best model fit. Therefore, whilst some genuine effects from small categories might still have been missed, the final best-fit models we selected were those that most closely fitted the data.

A third limitation was the fact that we did not contact the authors of the original study before conducting our data analyses, to clarify any uncertainties with the dataset and seek guidance on our planned approach. We chose not to do so because the original paper was well written, with clearly presented study methodology and results. Further, since independent replication is critical to the scientific method, it was arguably better not to have direct contact with the primary study authors when conducting statistical analyses.

A fourth limitation was that the outcome variables tested in the current study were derived from only one of the health metrics examined in the original study, namely owner-reported dog health [15]. Other health metrics reported by Knight et al. included asking owners “to report what they believed their veterinarian’s assessment [of their dog’s health] to be” [15] and owner reports of 22 possible health disorders, grouped into broad categories (e.g., allergy, cancer/tumours, hormonal, skin/coat) [15]. Our conclusions might have differed if one or more of these alternative health metrics had been analysed, but we decided against this because, in our opinion, all were inferior measures. Not only were these metrics also based on subjective owner reports, but owners were required to speculate on what their veterinarian’s opinion about the health of their dog would be. Further, there were no clear definitions of the health disorder categories, meaning that these would likely be broad, comprising many different diseases with differing aetiology and severity. Finally, the number of dogs included in many categories was small, being ≤50 in 12 of the 22 [15]. Whilst further work could be considered, where these additional metrics are analysed using the techniques of the current study, it is questionable as to whether anything further would be gained. Instead, it would be preferable to conduct a prospective study utilising objective, veterinary-derived measures of health, as discussed above.

A final limitation was the fact that there might be confounding between healthcare variables (e.g., veterinary visits and received medication) and dietary choices. For example, owners who are sceptical about conventional diets and instead choose to feed either vegan or raw food, might also be less willing either visiting their veterinarian or allowing their dog to receive medication. In support of this, negative associations were identified in the current study between feeding a raw diet and both veterinary visit frequency and the use of medications; further, in the previous study, dogs fed raw diets were less likely to be neutered than dogs fed conventional diets [15]. However, results were inconsistent for dogs fed vegan diets; whilst there was a negative association between feeding a vegan diet and the odds of a dog receiving medications, there was no association with frequency of veterinary visits. Further, in the previous study, dogs fed vegan diets were more likely to be neutered than those fed conventional diets [15]. Given these complexities, further research would be required to investigate associations between different diet choices and healthcare variables.