“Correlation does not imply causation” is a basic motto of science. Every scientist knows that observing a correlation between two things doesn’t necessarily mean that one of them causes the other.
But according to a provocative new paper, many researchers in psychology are drawing the wrong lessons from this motto. The paper is called The Taboo Against Explicit Causal Inference in Nonexperimental Psychology and it comes from Michael P. Grosz et al.
The article makes a lot of points, but to me the main insight of the piece was this: many studies in psychology are implicitly about causality, without openly saying as much.
Consider, for example, this highly cited 2011 study which showed that children with better self-control have better health and social outcomes years later as adults.
This 2011 paper never claimed to have shown causality. It was, after all, an observational, correlational design, and correlation is not causation. But Grosz et al. say that the study only makes sense in the context of an implicit belief that self-control does (or probably does) causally influence outcomes.
The title of the 2011 paper suggests that it was a study about predicting the outcomes. Prediction can be an important goal, but Grosz et al. point out that if the study had really been about prediction, it would make sense to consider a whole range of possible predictors. A purely predictive study wouldn’t focus on a single factor. The paper also probably wouldn’t be so highly cited, if readers really thought it said nothing about causality.
Grosz et al. analyse three other influential “observational” psychology papers and in all cases, they find evidence of unstated causal claims and assumptions, swept under a correlational rug.
As they put it, “Similar to when sex or drugs are made taboo, making explicit causal inference taboo does not stop people from doing it; they just do it in a less transparent, regulated, sophisticated, and informed way.”
The authors go on to argue that there’s actually nothing wrong with talking about causality in the context of observational research – but the causal assumptions and claims need to be made explicit, so that they can be critically evaluated.
To be clear, the authors are not saying that correlation implies causation. They argue that it is sometimes possible to draw inferences about causation from correlational evidence, if we have enough evidence to rule out non-causal alternative explanations. This kind of inference is “very difficult. However, this is not a good reason to render explicit causal inference taboo.”
