AUSTIN, TEXAS - This series has covered a great deal thus far, going well beyond the common blurb ‘correlation does not mean causation.’ It has explored multiple types of correlation, such as the spurious relationships and confounding relationships covered in the second part of the series, titled Causation: The most misunderstood concept in AI. The series has also stressed the importance of understanding the differences between predictive analytics and prescriptive analytics, as well as the role that causation plays in the latter paradigm.

The previous article concluded with three considerations for transforming a predictive model into a prescriptive one. This one will explain each of these in greater depth. First, one should always assume causation based on well-documented scientific studies. Second, make sure to design experiments and randomized clinical trials that can actually establish causation. Third, think about causal inference methods that may include Bayesian networks. These methods will also be discussed in future installments.

Assume causation based on well-documented scientific studies

Prescriptive models use a set of predictor variables to predict an outcome. Some of these are set and cannot be manipulated - including age, gender, and family history. These variables may be merely correlated or causative. For prescriptive models, they do not matter because one has no control over them.

For variables that can be controlled, one must first be sure that they are in fact causal in nature. If there is scientific evidence to conclude these manipulated predictors - i.e., independent variables, factors, and features - cause the prediction (target, dependent variable), then a clinician deploying the model may cautiously move forward.