DRAFT DRAFT 1 Representing and Reasoning with Correlated Effects in Probabilistic Causal Models

J. Lemmer

Abstract:This paper introduces a new form of knowledge representation for probabilistic Uncertain Causal Models. This representation allows for correlations among effects to be more easily represented than in chain rule based models, such as standard Bayesian Networks (BN's). The importance of effect correlations has been recognized since the early days of computerized causal models, especially in the medical community. Currently, when effect correlations are important, techniques based on the Bahadur expansion are usually used rather than techniques based on Bayesian Networks. This paper adds a new knowledge structure to Bayesian Networks capable of representing effect correlation. Bayesian Networks now use Conditional Probability Tables (CPT‟s) to represent uncertain causal knowledge. This paper defines an additional representation we call Effect Probability Tables (EPT's). Unlike CPT‟s, EPT‟s can efficiently represent effect correlations. We analyze the relation between EPT‟s and CPT‟s and show them to be duals when effects are conditionally independent and causes are causally independent. We present an algorithm for using both representations in the same model without requiring these independence assumptions. We do claim that in some domains, the parameters for EPT‟s are more natural to obtain than the parameters for CPT's especially when subjective parameters are to be elicited experts. We claim that EPT's can result in better models especially when effect correlations are important. No claim is made that there are joint probability distributions, which cannot be modeled using just CPT‟s. But we do claim that the structure and parameterization of models built with both representations can be much more natural to experts. The paper also shows by example how modeling effect correlations can produce significantly different inference results than if these correlations were assumed away, as is common Bayesian Network practice today. Since the paper provides techniques for combining both CPT's and EPT's in the same model, synergies among causes, now conveniently modeled using CPT's, can be combined with effect correlations, conveniently modeled using EPT's, to provide more accurate, powerful and satisfactory models.