In junior-high and high-school statistics, we learn to collect data and compute averages or correlations, but economic data are tricky because many variables influence one another at the same time. Haavelmo introduced probability models that explicitly include unobservable random factors (error terms) and even described the sampling process mathematically. He then showed how to treat a set of equations that are determined simultaneously—the so-called simultaneous-equation model—within this probabilistic framework. As a result, economists can test causal relationships even for interdependent pairs such as demand and supply. His framework underlies modern studies of financial crises, policy evaluation, and many other real-world economic issues.

In his 1944 monograph, “The Probability Approach in Econometrics,” Haavelmo reconstructed econometrics as an experimental science grounded in probability theory. He first recast structural equations derived from economic theory as stochastic models, defining error terms and parameters with explicit population interpretations. He then proved that in simultaneous-equation systems, endogenous variables are jointly determined, causing ordinary least squares to lose unbiasedness and consistency; this led him to formalize the concept of econometric “identification.” His probabilistic approach paved the way for estimation techniques such as Limited-Information Maximum Likelihood and Two-Stage Least Squares, both of which became workhorses of empirical economics. Moreover, by clarifying the sampling process and model assumptions, he opened theoretical paths to tackle sample-selection bias, natural-experiment designs, and many other applied issues.

Haavelmo defined an econometric model as a joint conditional distribution on a probability space (Ω,𝔽,P), specifying structural parameters β and disturbances u as IID under the null, thus delineating the parameter space Θ. He showed explicitly that simultaneity bias renders OLS inconsistent, E[β̂∣X]≠β, and introduced the rank condition Rank(Π)=k to assess identifiability. He further demonstrated that the likelihood-based estimator for simultaneous systems coincides with the mode of the posterior, granting a Bayesian interpretation, and derived asymptotic variances via the information matrix I(β). His work links directly to the Cauchy–Rees-Fisher problem and provides the mathematical justification for instrumental-variables estimation, laying the theoretical backbone for GMM and SVAR methodologies. Contemporary agendas in structural estimation—particularly the retrieval of policy-invariant parameters—remain fundamentally rooted in Haavelmo’s perspective.