In statistics, causal inference techniques are tailored to predict the effect of potential interventions, such as a treatment, a public health campaign or a policy, either using randomized experiments or observational data. Causal inferences are most easily drawn from randomized trials, but frameworks for inferring causation from observational data have also been established. Conducting randomized experiment is, however, not always possible in practice, either for financial, ethical or logistical reasons. This is often the case in epidemiology and in some other medical sciences, where many studies are carried out utilizing observational data. Decision makers thus sometimes need to rely on evidences arising from observational studies. However, estimating causal exposure effects using observational data requires important subject-matter knowledge.
In fact, one first needs to identify the confounding covariates that might distort the relationship between the exposure and the outcome and to select an appropriate modeling approach to control for the identified confounding covariates. Unless these tasks are adequately performed, the evidences produced may be erroneous, sometimes importantly so, and may lead decision-makers astray. However, these are very difficult tasks, especially in subject areas where current knowledge is scarce or limited. It would thus seem natural to supplement the available subject-matter knowledge with information from the data. However, most classical data-analytic approaches are poorly suited to draw information from the data in order to perform causal inferences. This research program aims at developing new data-analytical tools that will help applied health researchers making adequate causal inferences utilizing observation studies.