From a Century of Statistics to the Age of Causation

Judea Pearl, UCLA

Thursday, April 17 at 7:30pm in PAB 110

Some of the main users of statistical methods -- economists, social scientists and epidemiologists -- are discovering that, contrary to most textbooks, their fields rest not on statistical but on causal foundations. These foundations have been blurred or avoided through the years for the lack of a mathematical notation capable of distinguishing causal from equational relationships. Recent advances in graphical methods and nonparametric structural models provide formal and natural explication of these distinctions, and are changing the way causality is used in the knowledge-rich sciences. Simple methods are now available for solving the following problems:

  1. Deriving algebraic expressions for identifiable causal effects (both total and direct) in nonparametric structural models with latent variables.
  2. Selecting sufficient set of measurements (covariates or confounders) that permit unbiased estimation of causal effects in observational studies.
  3. Predicting (or bounding) treatment effectiveness from trials with imperfect compliance.
  4. Estimating (or bounding) counterfactual probabilities from statistical data (e.g., John, who was treated and died, would have had 90% chance of survival had he not been treated)

The talk will survey these developments and will outline future challenges.

References for Judea Pearl talks

J. Pearl "On the Foundations of Structural Equation Models." Technical Report R-244-s, UCLA, Computer Science Dpt November 1996.

Pearl, J., ``On the Testability of Causal Models with Latent and Instrumental Variables,'' (with appendix: Graphs, Structural Eqautions and Counterfactuals) In P. Besnard and S. Hanks (Eds.), Uncertainty in Artificial Intelligence 11\fR, Morgan Kaufmann, San Francisco, CA, 435--443, 1995.

Pearl, J., ``On the Identification of Nonparametric Structural Models,'' Technical Report (R-207), Revision IV, November 1995. To appear in Lecture Notes Series: Latent Variable Modelling with Application to Causality, Springer-Verlag.

J. Pearl, "Causal diagrams for empirical research" (with discussion), Biometrika 82(4), Dec. 1995, pp. 669-710.

Other relevant papers can be found here.