Causal inference with data subject to covariate-dependent observation times: An application to a cohort of new users of antidepressants