This project investigates whether the crucial distinction between causation and spurious correlations can be drawn in terms of (in)sensitivities to natural changes in the circumstances.
Most causal relations are insensitive: they do not break down under natural changes in the causal situation. For example, regardless of how smoking is caused, it correlates with lung cancer. By contrast, mere correlation is often sensitive: the correlation between cancer and yellowed teeth breaks down if the yellowness is not caused by smoking. The main thesis of this project is that different types of insensitivity can help to distinguish causation from types of spurious correlation that pose challenges for state-of-the-art accounts of causation.
It is often important to distinguish between causation and mere correlation. Among other things, this distinction is crucial to identify effective strategies for interacting with our environment. For example, both smoking and yellow teeth are correlated with lung cancer but only smoking is a cause of lung cancer. This means that one cannot effectively change one’s chances of having lung cancer by changing the yellow colour of one’s teeth, whereas one often can effectively affect one’s chances of getting lung cancer by changing one’s smoking behaviour.
This crucial difference poses a particular challenge for the scientific community. As noted already by Hume (1740), we cannot observe any special connection of causal determination or causal production between distinct phenomena such as smoking and lung cancer. We can only observe the correlations between their occurrences. Detecting causal relations on the basis of correlations is often the primary aim in many areas of scientific research, such as medicine and sociology. As argued by Pearl and Mackenzie (2018), this challenge is all the more pressing given the reliance on increasingly big data sets and automatized extraction algorithms. Characterizing a precise distinction between causation and correlation would thus benefit the general scientific community.
The difference between causation and mere correlation also poses a challenge for philosophical theories of causation. Any adequate theory of causation needs to be able to tell these cases apart. This challenge is particularly pressing for regularity theories causation, which aim to reduce causal relations to a subclass of correlations without relying on notions like production or determination. My project aims to articulate a systematic difference between causal and non-causal correlations. The hypothesis is that this distinction can be characterized in terms of the sensitivity of mere correlations. The project starts from the following conception of sensitivity:
Sensitivity: A correlation between two phenomena X and Y is sensitive if and only if it easily breaks down under natural changes in the causal situation.
For example, the observed correlation between yellow teeth and lung cancer in a population is sensitive, as it breaks down if we incite smokers to bleach their teeth. By comparison, the correlation between smoking and lung cancer is insensitive, because it does not break down as easily. In particular, this correlation will not break down if we incite smokers to bleach their teeth. The central hypothesis in this project is whether this notion of sensitivity can provide a principled distinction between causal relations and correlations between phenomena that are not related as cause and effect – ‘spurious’ correlations for short.
The purpose of this project is to elucidate the difference between causation and spurious correlation. The research will focus on three types of spurious correlation that are prevalent in the literature and aims to provide a unified explanation of why these types of correlation fall short of causation. Given that the approach developed in the project does not rely on notions such as causal production or causal determination, its results will also support those theories of causation that aim to reduce causal relations to a subclass of correlations without relying on such notions. In doing so, the resulting analysis of causation can help developing methodologies to detect causal relations on the basis of correlations.