Prediction vs explanation

The search for causal explanations can be likened to a search for a needle in a haystack.

The development of predictive models is a way of identifying what part of the haystack we should be looking in. But the best performing model (i.e which identifies the part of the haystack which should be looking into) does not by itself provide a causal explanation that we may be looking for. Associations are a necessary but insufficient basis for a good causal claim.

Additional steps need to be taken once a good performing predictive models is found. There needs to be a detailed within-case analysis to investigate how, if at all, the attributes in the model are causally connected in real life. To extend the haystack metaphor, this is like deciding to open up the hay bales in the area where the predictive model said we should be looking

Even if it is found that there is no underlying causal connection for a given predictive model this is not necessarily a bad finding. Predictive models can still be useful on their own. A predictive model that would enable donors to fund projects that were successful in achieving their objectives 75% of the time, versus a chance based choice of 50% of the time, would still be a very useful product. Many grant making bodies are not even able to quantify their performance in these terms.

And different kinds of models can be useful for different kinds of people…


Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: