of causality, because they are likely not affected by confounding. But even randomised trials can have biases when there are missing observations, as is often the case. Randomised trials are also not feasible in many cases, or if feasible, not representative for many risks, including environmental risks. Cohort studies can provide compelling evidence, but many cohorts do not adequately control for socioeconomic confounders and can suffer from many other issues related to the quality of exposure measurement or outcome ascertainment. To go beyond, the quantification of the number of studies of each type we have provided here will necessitate a deeper analysis of the potential limitations of all 2579 studies used across the risk-outcome pairs. In future work, we plan to evaluate the quality of each of these studies with a standardised approach and work toward an overall evidence summary. There is also a more fundamental philosophical question about the presentation of risk information. Should decision makers only pay attention to risk factor quantification for those risks supported by the strongest causal evidence such as randomised trials? Or do notions such as the precautionary principle suggest that we should pay attention to risk quantification even for risk-outcome pairs where the
