The lay abstract featured today (for Observational Causality Testing by Brian Knaeble, Braxton Osting and Placede Tshiaba) is from Stat with the full article now available to read here.
How to Cite
Knaeble, B., Osting, B. and Tshiaba, P. (2024), Observational Causality Testing. Stat, 13: e70017. https://doi.org/10.1002/sta4.70017
Is marijuana a gateway drug? Can we reduce risk of stroke by treating diabetes? Does smoking cause lung disease?
The methodology of this paper has been developed to address questions like the example questions above. We have developed a test for causation in observational data. We can observe an association and infer causation, but only when the association is sufficiently strong, and only when there is sufficient randomness within the processes that give rise to the association.
Here is the basic idea. An observer may be tricked by processes that produce spurious (non-causal) associations. However, that production is hindered by the presence of randomness within those processes. With enough randomness (entropy) in those processes the trick becomes impossible. Remarkably, full randomness is not required.
In a preliminary paper titled An Asymptotic Threshold of Sufficient Randomness for Causal Inference we showed how to compute a threshold of sufficient randomness from observational data. In this subsequent paper, Observational Causality Testing, we show how to quantify randomness in processes by transporting measures of discordance from studies of identical twins. We also show how to conduct observational causality tests with measured covariate data and finite population corrections. Example applications are described. The details provide insight into principles to guide covariate selection.
More Details