Letter by Owen et al Regarding Article, “Dog Ownership and Survival: A Systematic Review and Meta-Analysis”

To the Editor:

Kramer et al¹ systematically reviewed the association between dog ownership and all-cause mortality. This is an important question, as many studies have reported significant cross-sectional associations between dog ownership and human physical activity or mental health outcomes, but more intervention evidence and pooled longitudinal data are needed to assess the public health policy relevance of our much-loved canine best friends. However, we are concerned with some methodological issues in the meta-analysis which may influence conclusions on the benefits of dog ownership.

Adjusted Versus Unadjusted Effects

Confounding obscures the real effect of an exposure on an outcome.² As such, it is important to adjust for potential confounding variables in cohort studies to more accurately estimate real effects. For example, the effect of dog ownership on all-cause mortality could be confounded by physical activity, sociodemographic differences, or overall health status. People who are physically active or have overall good health might be more likely to own a dog, and as these people already have a lower risk of all-cause mortality, physical activity and better health status could obscure the independent effects attributable to dog ownership. Therefore, where possible, adjusted estimates should be extracted and included in meta-analyses. To assess the role and extent of confounding in the effect, the unadjusted estimates should also be extracted and included in a sensitivity analysis.³

Fixed- Versus Random-Effects Models

Another important decision is whether to use fixed- or random-effects models to pool effects from individual studies. The random-effects model assumes that the underlying true effect size varies across studies, due to study participants being sampled from different populations, for example, from countries, with different degrees of physical activity or health status.⁴ Therefore, weights in the random-effects models are calculated using a combination of within- and between-study variance, resulting in relatively similar weights for studies of different sizes. In contrast, the fixed-effects model assumes that there is one true underlying effect size that all studies are trying to estimate—an assumption that is rarely met in practice.⁴ As such, weights are calculated using only the within-study variance and so larger studies are weighted very heavily. The problem with the fixed-effects model assumption is that it is rare that a researcher would focus only on specific studies included in the meta-analysis and not be interested in estimating the effect size across broader populations and to inform the research area as a whole.⁵ Therefore, the authors of this dog ownership-mortality meta-analysis¹ used a random-effects models, as this is routinely used in meta-analyses to estimate effects that can be generalized to the whole population and research area. But we suggest that including adjusted estimates in the pooled meta-analysis estimate would provide a more accurate representation of the evidence on dog ownership and mortality.

Footnotes