Impact of Imperfect Disease Detection on the Identification of Risk Factors in Veterinary Epidemiology

Risk factors are key epidemiological concepts that are used to explain disease distributions. Identifying disease risk factors is generally done by comparing the characteristics of diseased and non-diseased populations. However, imperfect disease detectability generates disease observations that do not necessarily represent accurately the true disease situation. When not taken into account, this imperfect detectability is likely to impact the identification of risk factors to an unknown extent. In this study, we conducted an extensive simulation exercise to demonstrate the impact of imperfect sensitivity on the outcomes of logistic models which are the most popular models to identify risk factors. We used a probabilistic framework to simulate both the disease distribution in herds and imperfect detectability of the infected animals in these herds. These simulations show that, under the logistic model, true herd-level risk factors are generally correctly identified but their associated odds ratio are heavily underestimated as soon as the sensitivity of the detection is less than one. If the detectability of infected animals is not only imperfect but also heterogeneous between herds, the variables associated with the detection heterogeneity are likely to be incorrectly identified as risk factors. This probability of type I error increases with increasing heterogeneity of the detectability, and with decreasing sensitivity. Finally, the simulations highlighted that, when count data is available (e.g. number of infected animals in herds), they should not be reduced to a presence/absence dataset at the herd level (e.g. presence or not of at least one infected animal) but rather modelled directly using zero-inflated count models which are shown to be much less sensitive to imperfect detectability issues. In light of the simulations, we revisited the analysis of the French bovine abortion surveillance data, which has already been shown to be characterized by imperfect and heterogeneous abortion detectability. As expected, we found a substantial bias in quantitative outputs of the logistic model in comparison with results from the zero-inflated Poisson model. We conclude by strongly recommending that efforts should be made to account for imperfect disease detectability when modelling disease risk, or at least to discuss its potential impact on model outcomes.