CFE 2019: Start Registration
View Submission - CMStatistics
B1289
Title: Causation, bias and optimal experimental design Authors:  Henry Wynn - London School of Economics (United Kingdom) [presenting]
Abstract: One of the profound issues in causal modelling, particularly in socio-medical areas is the protection of causal models from bias from different sources, the most problematical being the inability to do controlled experiments, often labelled as the ``absence of counterfactuals''. Building on previous work, we study optimal experimental designs for models in which either there is a well-defined bias term or in which we guard against vaguer possible sources of bias. We adopt a cooperative game-theoretic approach in which a hypothetical player Alice has ``ownership'' of the main model, and another player ``Bob'' tries to eliminate the effect of bias. The standard randomized control methods falls into this category. Motivated by minimax justifications of randomization we are, in certain circumstances, able to establish the existence of Nash equilibrium. This then leads to new optimal design criteria. Complex causal models may be non-linear with additional complexity. Removing or guarding against bias can then be seen as providing a protective Markov ``blanket''.