Randomized clinical trials are conducted for each new drug or medical device to demonstrate safety and efficacy for the European Medicines Agency (EMA) and the U.S. Food and Drug Administration (USFDA), for example, to approve it. Moore et al. [2020] estimated the median cost of clinical trials for 101 therapeutic drugs approved by USFDA in 2015–17 was US$48 million per approved drug. The aim of a clinical trial for a drug is to assess the effects of intervening to give someone the drug. The effects include both the beneficial and harmful effects, where the benefits have to outweigh the harms.

The main assumption behind a randomized clinical trial is that the random assignment of the drug means that there are no confounders. Missing data – when some patients drop out of the trial – makes the naive analysis of such data problematic. The conductors of the trial need to find out why the patients dropped out, or to consider the worst case (similar to the analysis of instrumental variables).

Randomized controlled trials are a standard mechanism in much of science, including the social sciences, however they may not be appropriate or possible in all situations. For example, a randomized trial to gauge the impact of an intervention in schools might be unethical because, although the study might provide information that can help future students, some of the students in the trial are not being provided with the best education available. It is also difficult to only vary a single condition in a study on students. The tools of causal analysis and making causal assumptions explicit should enable more cases where the effect of interventions can be inferred. Explicit assumptions are open to scrutiny and debate.

One of the promising ways to explain a prediction of an otherwise inscrutable method, such as a neural network, is a counterfactual explanation. Given a prediction, a minimal counterfactual explanation is a minimal change to the inputs that would result in a different conclusion. For example, if someone was denied a loan, it is reasonable to ask for the smallest changes that would have resulted in the loan being approved. There are generally many minimal changes that could have resulted in a different conclusion.