THE FALL OF OLS IN STRUCTURAL ESTIMATION

Abstract

A recurrent debate in the history of econometrics is the extent to which ordinary least squares (OLS) estimation can yield valid information about economic structure. This debate started well before the work of the Cowles Commission in the 1940's on simultaneity bias. For example, in 1912 Marshall was strongly condemning Henry Ludwell Moore, the first American econometrician, as a nightmare. The problem of interpreting econometric results has not diminished with time. This article discusses early developments in the analysis of identification and single equation estimation which do not seem to be widely known among modern economists, despite the insight they provide into econometric practice.' Economists such as Marshall became more tolerant of econometric work when multiple correlation was explained as a technique that could embody the ceteris paribus clause of pure theory. Nevertheless, the early experience of estimating demand and supply curves showed that econometricians faced unique difficulties. Statistical models in other fields used the logic of cause and effect to specify relationships with one dependent variable and several explanatory factors. It was apparent from the outset, however, that a scatter of simultaneous price and quantity observations could not directly reveal either demand or supply schedules without additional information. Substantial progress on what was later called the identification problem was achieved by the end of the 1920's. One result of this work was to suggest that OLS might be an inappropriate estimator for simultaneous systems. A form of instrumental variables estimation appeared in the literature as early as 1928 and an example of indirect least squares was given in 1929. The paper is organized as follows. The first section discusses the work of Sewall Wright in the 1920's as a kind of instrumental variables solution to the market equilibrium problem. The second section discusses an early discovery by Tinbergen of identification and estimation using exogenous variables in an indirect least squares setting. The third section discusses the work of Haavelmo and Koopmans in the emergence of structural estima-