A second‐class science? A defence of observational epidemiology to make causal inferences

We read the thoughtful discussion of alcohol as a universal preventive agent by Fekjær 1with great interest. While we are largely in agreement with many of his concerns about the evidence for alcohol's beneficial effects, one particular argument stood out: his view that observational epidemiology reflects nothing more than second-class evidence. He is not alone in his condemnation of observational epidemiology 2. However, one could argue that it is naive to believe that findings from experimental studies (randomized or not) necessarily qualify as more robust evidence for the existence of a causal relationship than those from observational investigations. First, this assertion relies on the assumption that all experimental studies are conducted without sources of bias being present and on samples that are representative of the population of interest. Often this is not the case, with stringent inclusion and exclusion criteria restricting the sample of eligible individuals (e.g. those with existing illnesses, particular age groups, etc) 3, 4 and the need for participants to provide informed consent to being allocated randomly to conditions presenting major issues of both generalizability and potential bias. This reduces the external and internal validity of findings from experimental studies 3-5 (some of these issues are also relevant to animal studies 6). Fekjær 1 highlighted findings from observational studies that were not replicated in interventional studies—but there are also examples of results from experimental studies that have not fared well when translated into clinical practice 7, 8. Findings from observational and experimental studies often overlap 9, 10, and it would seem unwise to discredit all results from one particular study design as being inferior to another when they each have opposing strengths and weaknesses 11. We do, however, share several of the concerns raised regarding the use of observational data to make causal inferences. As others have noted, achieving causal estimates involves implementing appropriate methods to ensure that the measurement of an effect is as accurate as possible 12. This heuristic framework encompasses choosing variables which suitably reflect the exposure of interest as well as the outcome, defining appropriate covariates 13, 14 as well as ensuring that participants come from relevant populations, and making necessary adjustment for missing information to avoid forms of selection bias 12, 15, 16. Statistical methods are continuously being developed to allow, with greater confidence, for causal inferences to be made from observational data. For example, Mendelian randomization 17-21 studies using genotypes concerned with the metabolism of alcohol 22-24 have been used to examine the causal relationship between alcohol consumption and several health consequences including coronary heart disease 25, myocardial infarction 26, blood pressure 27, cancer 28, 29 and cognitive decline 30. A plethora of alternative causal inference methods, e.g. 31, 32 are available to those working with observational data, and it is encouraging to see them being adopted by the wider addiction field 33-35. Observational epidemiologists should first try to minimize the amount of error in the measurement of the association of interest. We should then be as transparent as possible—applying thorough criticism of our findings to evaluate the extent to which known (and unknown) bias could impact on the validity of our findings. The failure to do so is more a reflection of second-class scientists than observational epidemiology itself being a second-class science. None. Both S.B. and A.B. are currently funded by the European Research Council (309337, Principle Investigator: Britton, http://www.ucl.ac.uk/alcohol-lifecourse).