Discovering Causal Knowledge by Design

Causal knowledge is frequently pursued by researchers in many fields, such as medicine, economics, and social science, yet very little research in knowledge discovery focuses on discovering causal knowledge. Those researchers rely on a set of methods, called experimental and quasiexperimental designs, that exploit the ontological structure of the world to limit the set of possible statistical models that can produce observed correlations among variables. As a result, designs are powerful techniques for drawing conclusions about cause-and-effect relationships. However, designs are almost never used explicitly by knowledge discovery algorithms. In this work, we provide explicit evidence that designs have the potential to be highly useful as part of algorithms to discover causal knowledge. We first formalize the basic elements of experimental and quasiexperimental designs to characterize a design search space. We then quantify the range and diversity of designs that can be applied to examine the central questions associated with a large and complex domain (Wikipedia). Finally, we show that explicit consideration of designs can substantially improve the accuracy of causal inference and increase the statistical power of algorithms for learning the structure of graphical models.