EFFECTIVE CONDITIONING (Q4029915)

Avoiding the usual ``pseudo-conditioning'' approach (one in which conditioning is introduced during the course of an argument, while the final form of the statistical inference is unconditional), the author studies here ``the logic of true conditioning'', in which the value of a statistic upon which one conditions is fixed at its observed value throughout the statistical analysis. Criteria of utility and relevance for judging a conditional variable are introduced, quantitative measures of them in terms of conditional and unconditional information being provided. These criteria are observed to be logically, mathematically and conceptually distinct. Both are shown to be needed in the evaluation of the effectiveness of conditioning, and it is also noted that a variable which is optimal under one of these criteria may be far from optimal under the other. The ideas are applied to the nuisance parameter case and the contingency table.