Stable temporal relationships as a first step towards causal inference: an application to antibiotic resistance

Epidemiological studies often analyze data as static, essentially averaging observed associations across time. Overlooking time trends is especially problematic in settings subject to rapid changes. A prominent example for such a setting is antibiotic resistance, which has reached concerning levels, and poses a global healthcare challenge. Bacteria constantly evolve and hence antibiotic resistance is characterized by time-varying relationships with clinical and demographic covariates. In this paper, we speculate that covariates with a causal effect are expected to have stable relationships with resistance over calendar time. To this end, we applied time-varying coefficient models in a retrospective cohort analysis of a large clinical dataset from an Israeli hospital, and have shown their advantages in describing covariate-resistance relationships. We found both time-stable and time-varying covariate-resistance relationships. These results serve as initial evidence towards causal interpretation of these relationships, as one may expect time-stable rather than time-varying relationships to correspond with causal effects. We further conducted data-driven simulations, that have illustrated how results from time-varying coefficient models must be carefully interpreted with respect to causal claims. Potentially, identification of causal covariate-resistance relationships can lead to new medical interventions and healthcare policies, and improve the generalization of existing predictive models for antibiotic resistance.