Out of a total of 4,446 dogs, 2,967 (67%) were at an ideal weight, defined as an owner-reported body condition score of 4-5 on a previously-validated 9-point system (BCS 4-5; (13)) and 1,480 (33%) were overweight or obese (BCS≥6). Of these dogs with higher BCS, 1,124 (25% of total) were overweight (BCS 6) and 356 (8% of total) were obese (BCS≥7).
Identification of Risk Factors Associated with Overweightness and Obesity
Significant Risk Factors Identified via Univariate Analysis:
To identify factors associated with increased BCS, we performed two univariate analyses comparing ideal weight dogs (N=2,966) to overweight/obese dogs (N=1,480) as well as to obese dogs only (N=356). Of the 45 variables selected as outlined in Methods, 22 (49%) were significantly associated with overweightness/obesity in a univariate analysis (p < 0.05, N=4,446) and 18 (40%) were significantly associated with obesity (p < 0.05, N=3,322). The relationships between these variables and BCS are summarized in Table 1 [See end of document]. The 18 variables positively associated with both overweightness/obesity and obesity were diet combinations containing dry food, increased treat quantity, lack of probiotic consumption, increased age, decreased exercise per week, neutering, increased pet appetite, increased food motivation, lower overall mood, decreased conspecific interaction, increased tail chasing, decreased prey drive, presence of other dogs in the household, rural home environment, conventional-only medicine type, household tobacco use, food intolerances, and rescue or other acquisition method. Four additional variables positively associated with BCS in the overweightness/obesity analysis were the use of dental chews, sharing or cooking food, increased overall nervousness, and decreased dental visit frequency.
Significant Risk Factors Identified via Stepwise Multivariable Analysis:
Since body weight is a complex trait that can be influenced by multiple variables, we performed multivariate analysis to understand how different variables are associated with BCS when they are taken together. We performed these analyses in the overweight/obese and obese groups separately. The results from the two stepwise logistic regression models comparing ideal weight dogs (N=2,725) to overweight/obese dogs (N=1,384) as well as obese dogs (N=327) are presented in Table 2 [See end of document]. Due to missing data in the selected variables, 337 dogs from the total dataset were dropped from this logistic regression. Log odds for significant variables are presented in Figure 1. The variables significantly associated with overweightness/obesity were age, exercise per week, food motivation level, overall mood, pet appetite, sharing food, neutering, treat quantity, probiotic supplements, home environment, diet, and dental treatment frequency (p < 0.05, stepwise logistic regression, N=4,109). The variables significantly associated with obesity alone were age, exercise per week, food motivation level, other dogs in the household, overall mood, pet appetite, neutering, tail chasing frequency, treat quantity, probiotic supplements, medicine type, diet, and dental treatment frequency (p < 0.05, stepwise logistic regression, N=3,052). The intersection of variables in the two models were age, exercise per week, food motivation level, overall mood, pet appetite, neutering, treat quantity, probiotic supplements, diet, and dental treatment frequency.
Given the nonlinear relationship between age and body condition reported by others (3,5,6,9), age was entered as both a linear and quadratic term in the logistic regression model. In the context of the model, both the linear and quadratic terms had odds ratios that were statistically significant (p<0.0001). The inclusion of the quadratic term significantly improved both the overweight/obese and obese only logistic regression models, as determined by nested model ANOVA (p<0.0001, likelihood ratio test).
Significant Risk Factors Selected via Elastic Net Analysis:
Collinearity poses a major issue for stepwise models, since only a subset of a group of collinear variables may be selected, such that the final ensemble of variables may be influenced by noise. Elastic Net is a method that combines the L1 and L2 penalties used in the Lasso and ridge methods, respectively. It addresses the collinearity issue, as it exhibits a grouping effect (14) wherein coefficients of correlated variables tend to be similar. Thus, through the Elastic Net algorithm we may see if any important variables are being masked by their correlations with others.
For both the overweight/obese and obese only models, the most significant ensemble of risk factors was selected as detailed in Methods. Factors that appeared in both the optimal model for overweightness/obesity and the optimal model for obesity alone were: pet appetite, treat quantity, exercise, probiotic supplementation, diet, mood, food motivation level, and age. The overweightness/obesity model also included neutering, home environment, and sharing food, while the obesity model also included medicine type. With the exception of medicine type, all of these variables are within the subset of those selected by stepwise logistic regression. Variables selected by the Elastic Net are presented in [Additional File 1] and [Additional File 2]. Comparisons between variables selected by Elastic Net and other methods are presented in Table 3 [See end of document].
Significant Risk Factors Selected via XGBoost Analysis:
Another issue with the stepwise model is that non-linear effects are not accounted for. We addressed this by using XGBoost, a tree-based machine learning algorithm (15). A tree-based model can establish decision points at multiple different values, and thus the final variable importance in the model encompasses non-linear relationships including interactions between variables. Through XGBoost we may identify which variables, if any, we should examine in terms of higher-order interactions or polynomial models. Factors that appeared in both the optimal model for overweightness/obesity and the optimal model for obesity were: age, pet appetite, exercise, treat quantity, food motivation level, diet, and probiotic consumption. The overweightness/obesity model also included neutering. Each one of these variables was also selected by stepwise logistic regression and Elastic Net. The variable importance plots from the XGBoost models are presented in [Additional File 3] and [Additional File 4]. Comparisons between variables selected by XGBoost and other methods are presented in Table 3 [see end of document].
We repeated the multivariate logistic regression analysis with only the subgroup of dogs that were reported to have no major health conditions, in order to remove the possibly confounding effects of disease and treatment variables. This subset consisted of dogs that did not have pancreatitis, diabetes, kidney issues, liver disease, heart issues, cancer, or gastrointestinal conditions (N=3,173, 71% of total dataset). Of these dogs, 2,118 (67%) were at an ideal weight (BCS 4-5) and 1,055 (33%) were either overweight/obese (BCS≥6) or obese (BCS≥7); these were further categorized into 792 (25% of total) that were overweight and 263 (8% of total) that were obese. These proportions were not statistically significantly different from the proportions in the overall sample (p < 0.05, χ2 test). Thirteen of the 15 variables selected from the total dataset remained significant despite decreased power to detect significant effects. These findings are available in Table 4 [See end of document]. Due to missingness in the selected variables, 192 dogs were dropped from the healthy subgroup logistic regression.
Individual Contributions of Selected Risk Factors
Among the 7 risk factors identified by all of the eight selection methods (univariate, stepwise, Elastic Net, and XGBoost each undertaken for both overweightness/obesity and obesity alone) as presented in Table 3 [see end of document] were diet, age, exercise, probiotic supplementation, and treat quantity. We further examined the individual contributions of these five risk factors to BCS.
We examined the relationship between the most common diet types and both overweightness/obesity (Figure 2) and obesity alone (Figure 3). Since a fresh food only diet was the largest category (besides “Other”, see Methods) we used this as our reference level in the logistic regression.
Relative to dogs on a fresh food only diet, dogs fed dry plus canned food (OR=1.85, p<0.0001), dry food only (OR=1.45, p<0.0001), and dry plus fresh food (OR=1.3, p=0.03) were more likely to be overweight/obese. Dry plus canned food (OR=2.6, p<0.0001) and dry food only (OR=2.1, p<0.0001) were also risk factors for being obese, but not dry plus fresh food (p>0.05). Dogs fed raw food only were less likely to be overweight/obese (OR=0.5, p=0.005), but there was no effect on obesity alone (p>0.05).
Given the strong relationship between body condition and age in both the univariate and multivariate models, we investigated this in further detail. We found that likelihood of overweightness peaked around age 8-10, decreasing with further aging (Figure 4, A). Likelihood of obesity alone showed a similar non-linear pattern. This relationship supports our use of higher-order polynomial age terms in the logistic regression models.
Since exercise was shown to be statistically significant in both the univariate and multivariate models (p<0.0001), we further considered this relationship (Figure 4, B). We found that incrementally increasing the amount of exercise per week decreased the likelihood of overweightness/obesity. The same pattern holds for obesity.
We also examined the relationship between probiotic supplementation and BCS (Figure 4, C). In total, 844 dogs were currently taking probiotics, and respondents reported using a broad range of commercial formulations. We found that probiotic supplementation was significant both with respect to overweight/obese status (OR=0.65, p<0.0001) and obese status (OR=0.46, p<0.0001) with dogs taking probiotics being more likely to be at an ideal weight.
Given that dogs receiving probiotics may be more likely to be at a lower weight due to having a medical condition, we repeated this analysis using only the healthy cohort as previously defined (N=3,173). We found a significant relationship with both overweight/obese status (OR=0.65, p<0.0001) and obese status (OR=0.52, p = 0.002) in this cohort as well.
Since supplement usage may indicate increased owner health consciousness, we also performed a similar test with the other supplements included in our analysis in order to determine if this significant effect was specific to probiotics. We examined prebiotics (N=183), multivitamins (N=425), CBD oil (N=540), fish oil (N=812), herbal supplements (N=207), and immune support (N=121). None of these supplements showed a significant relationship with either overweight/obese status or obese status in chi-square tests (p>0.05).
Finally, we examined the univariate association between treat intake quantity (by percentage of caloric needs being met by treats) and BCS (Figure 4, D). Both the univariate and multivariate analyses confirmed that, while giving over 10% of a dog’s diet in treats was associated with higher BCS, there was no significant difference between giving under 10% of a dog’s diet in treats and abstaining completely, suggesting that giving treats in moderation is not a risk factor for either overweightness or obesity.
Given that some respondents (N=593) gave data for multiple dogs, and that joint households may exhibited shared environmental factors not captured by existent survey questions, we determined if there was a significant effect of household on reported body condition. We entered household into the logistic regression model as a random effect, and an analysis of deviance showed that it was not significant (p>0.05). Furthermore, given findings from other studies that identified a relationship between sex and BCS (e.g. (6), or an significant interaction between sex and neutered status on BCS (10), the analysis was repeated in the subset of animals for which sex data were provided (N=3,922) but sex was not significant either as a main effect or in interaction with neutering (p>0.05). Given previously reported overweightness rates of 52% and 41% for females and males respectively (5), and α=0.05, we had β>.99 power to detect a main effect, and following previously published guidelines for calculating power for interactions (16), we calculated that we had β>.90 power to detect an interaction between sex and neutering. Therefore, we are relatively confident in reporting that no relationship between sex and BCS exists in this sample.