Explanatory analyses of randomized studies.

This paper considers randomized interventions which do not completely determine an intended determinant of response, and which may also manipulate additional, possibly unobserved, variables influencing response. The example we use throughout this paper is counseling for a low-fat diet for breast cancer prevention, where the intervention is counseling and dietary fat is hypothesized to reduce breast cancer risk. We use additive linear models to derive conditions and assumptions for considering fat to be the sole explanation of an observed treatment effect. A modified experimental design which supports stronger conclusions about causality is proposed.