Today we will talk about causality, the tools we use to ask if and how a variable influences other variables. Causal questions are everywhere in neuroscience. How do neurons influence one another? How does a drug affect neurons? How does a stimulus affect behavior? We will talk about how we can answer questions of a causal kind. In tutorial 1 we will learn about the definition of causality and see how it naturally emerges from the consideration of perturbations to a system. In tutorial 2 we will learn how correlations often seem to be a proxy for causality but also how for larger systems this intuition is typically misleading. In tutorial 3 we will see the standard approach, where we correct for unmeasured variables with regression and we will see the biases introduced by this. Lastly, in tutorial 4 we will learn about instrumental variables, a strategy that sometimes allows us to learn about causality in an unbiased way. The intro teaches the basic concepts while the outro gives a rather broad overview about the need to do careful science. Overall, the day teaches the unavoidable problems that come with asking causal questions and teaches approaches that can sometimes make it possible to estimate.
Causal questions are important all across neuroscience. For example, model fitting (W1D3), machine learning (W1D4), and dimensionality reduction (W1D5), are often used to argue for or against causal models. For example, a regression may be used to argue that a brain region influences another brain region based on fMRI data. Today’s materials give us a better understanding of the problems that come with the approach. There are tight links between causality and Bayesian statistics (W3D1) where Bayesian techniques are used for the estimation of causality (see e.g. the work of Judea Pearl). Causality is often seen as the bedrock of science, today’s materials above all produce clarity about what it is.
Causality approaches are central across neuroscience. When we run experiments, we often randomly assign them to treatment groups vs control. Alternatively we stimulate animals at random points of time. These methods are all versions of randomized perturbations that we talk about in tutorial 1 and probably constitute a good part of all of neuroscience. We also use model fitting (see tutorials 2 and 3) frequently to drive arguments about how brains work. This is common for spike data, EEG data, imaging data etc. Lastly, we should be able to sometimes use instrumental variable techniques to estimate the effects of e.g. treatments with drugs. Today’s materials are simultaneously at the heart of the field and are frequently ignored.