The cross-levels hypothesis is presented as an explanation for program evaluation failures. It states that the usefulness of evaluation data as feedback for decision making varies inversely with the number of organizational levels between the action the data describe, and the decisions they are intended to influence. To be useful for decision making, evaluation data must meet three hierarchical information needs: syntactic, semantic, and behavioral. Syntactic need refers to the choice of organizational levels. To avoid syntactic errors, evaluators should specify their level of reference--individuals, classes, districts, states, nations--and realize that aggregating data across levels may confuse relationships among variables. Evaluation data must be on the same level as decision maker concerns, to satisfy semantic needs. Formal evaluation reports, for example, are not relevant to teachers. The behavioral need explains why decision makers are less motivated by evaluation data removed from their level--politically speaking, this information is not perceived as important to their own concerns. The cross-levels hypothesis is strongly supported by these information needs, and offers an alternative to research design or statistical procedures as an explanation for program evaluation failures. (CP)