“Physical activity improves school achievement.” “Walk more briskly and you live longer.” “Diet influences the development of intestinal cancers through bacteria.” These are all recent headlines from the Finnish media, and each suggests some kind of causality, that is, they express a cause-and-effect relationship. For general readers, journalists and researchers alike, it is often difficult to find confirmation for the reliability of statements made in news stories about science. Could there be an alternative explanation for an observed connection, such as a coincidence or an unobserved intervening factor?

The basic problem of causal inference relates to the impossibility of measuring a particular response both when an action is performed and when it is not performed. A student cannot simultaneously do and not do a physical exercise. Despite this basic problem, causal inference is not impossible, but it requires substantiated assumptions and carefully planned research designs.

In modern causal inference, field-specific expertise is first refined into one or several causal models. After this the actual reasoning can be largely automatised. Data are needed for estimation and with them a given model can be proved wrong. However, no set of data – no matter how large – can alone provide a sufficient basis for causal inference.

In recent decades, significant advances have been made in the theory of causal inference, especially thanks to the Turing Award winner Judea Pearl and his research team. This work has produced tools researchers can use to determine what kind of causal inferences can be made under the given premises. Examples of these tools have also been developed at JYU’s Department of Mathematics and Statistics as part of the profiling area of “decision analytics utilising causal models and multiobjective optimisation (DEMO)”.

Finding causal relationships is the most central aim of science. A phenomenon becomes understood only once its causes and the effects related to it are known. The degree of a causal effect is usually a more important issue than the existence of causality as such. Causal inference is necessary in decision-making as well: decisions lead to actions, the consequences of which should be understood before making the decision. Better understanding of causal inference leads to better science and to better decisions.

Juha Karvanen, Professor of Statistics, Department of Mathematics and Statistics





Get latest articles from The University of Jyväskylä’s stakeholder magazine into your email. You can cancel your subscription at any time.