Abstract
Will lowering the marginal tax rate increase labor supply? In social sciences, experimentation is often impossible and data-driven approaches to answering such causal questions must use non-experimental data. Statistical methods to learn about cause and effect can exploit a known temporal ordering of the random variables as cause must precede effect. In some ways, causal inference from time series data appears simpler than causal inference using data with no temporal ordering. However, this perceived simplicity may be deceiving, and in this talk we discuss some common issues that arise when we try to reason about cause and effect from time series data along with some partial solutions.
Granger causality and impulse response functions are examples of standard time series methods that are often, at least between the lines, given a causal interpretion. Causal effects in time series are closely related to impulse response functions that are commonly used in, e.g., econometrics, and the interpretation of impulse response functions is similar to that of causal effects. However, this conflation implicitly assumes that the multivariate time series is fully observed, i.e., that there are no unobserved confounding processes. Without this assumption, we need additional information to identify causal effects from time series data. One possibility is a known causal graph representing cause-effect relations, and we outline a recent result in this context which allows us to corroborate or invalidate a causal interpretation of parameter estimates obtained using standard time series techniques. The widespread notion of Granger `causality' can also be given an actual causal interpretation, most easily under the restrictive assumption of no unobserved confounding processes. Without this assumption, Granger causality can still teach us about causal structure, and we describe how this can be done.
If the causal system develops faster than the sampling frequency, unobserved confounding may occur between the sampling points, and this will also complicate causal effect estimation and causal structure learning. A similar problem occurs when data is aggregated over time as is common for economic data. We give examples of this and some potential solutions.