Causal Structure and Hierarchies of Models

Abstract

Economics places a high premium on completeness of explanation. Typical general-equilibrium accounts of economic phenomena are preferred to partial equilibrium accounts on the ground that important interactions are necessarily omitted in the latter. A similar preference for microfoundational explanations over macroeconomic explanations of aggregate phenomena is grounded in similar reasoning. Probabilistic accounts of causation frequently presume that greater detail is superior to less. Simpson’s paradox, for example, assumes that failure to account for distinctions within populations results in false conclusions. Strategies of causal refinement – e.g., distinguishing between direct and indirect causes – are similar. However, there are countervailing practices in economics. Representative-agent models aim to capture economic motivation but not to reduce the level of aggregation. Structural vector-autoregression models and dynamic stochastic general-equilibrium models with small numbers of variables are often practically preferred to ones with large numbers. The distinction between endogenous variables determined within a causal system and exogenous variables determined independently of the causal system suggests a partitioning of the world into distinct subsystems. This paper will explore this tension. I advocate a structural account of causation grounded in Herbert Simon’s Causal Order and Identifiability (1953), which defines cause with reference to complete systems. But any workable causal epistemology must deal with incomplete systems and piecemeal evidence. The main formal focus of the paper is to better understand the constraints that a structural account of causation places on the freedom to model complex or lower-order systems as simpler or higher-order systems. The main epistemological focus is to understand how and to what degree piecemeal evidence can be incorporated into a structural account.