Mcfadden, Daniel L.

views updated

Mcfadden, Daniel L. 1937-

Daniel L. McFadden, a physics undergraduate who switched to behavioral science (economics) in graduate school, won the Nobel Prize in economics (Bank of Sweden Prize in Honor of Alfred Nobel) in 2000. McFadden’s best-known economic contribution and the one that led to his winning the Nobel Prize was the theoretical development and econometric evaluation of discrete choices involving multiple outcomes.

Discrete choice outcomes had been a vexing problem in economic analysis. Typically, economic decisions have been framed as trade-offs between goods in which quantities of goods could be traded over relatively continuous ranges. Discrete choices, by contrast, are “lumpy”: One choice precludes the possibility of choosing another option. For example, the same consumer cannot choose simultaneously to drive and to take the train to work on the same morning.

Modeling discrete outcomes requires a theoretical framework for discrete choice and empirical innovations that allow researchers to investigate the factors that influence the outcomes of choices. McFadden provided the theoretical linkage by adopting the basic precepts of a random utility model in which the agent is assumed to maximize utility but the researcher cannot observe the full set of factors that lead to the choice; the agent also cannot do this because of cognitive and information limitations.

The limited decision-making capacity of the agent, along with the limited information researchers have about the factors that lead to a decision, results in a set of determinants that are unobservable in the modeling process. This insight implies that the errors associated with decision-making are especially important in estimating discrete choice models. The idea of unobserved preference heterogeneity serves as the theoretical foundation of McFadden’s well-known multinomial logit model.

In a multinomial logit model McFadden assumes that there is a latent (unobserved) variable that represents the indirect utility (satisfaction) associated with each discrete choice. One set of independent variables measures the attributes of the choice. A second set of variables measures the individual attributes affecting tastes. A fundamental assumption of McFadden’s discrete choice models is that current economic conditions affect the feasibility of a choice through the budget constraint but not the preferences of the individual. Thus, prestige goods, or “snob” goods, are not modeled adequately using McFadden’s theoretical foundation.

Although the theoretical foundations for the multinomial logit model are complex, the empirical estimation of that model has become increasingly simple. Most of the empirical methods developed by McFadden, such as multinomial logit, conditional logit, and other nested models, have been standardized enough that they are easy to estimate. However, the interpretation of the results is less straightforward. Estimates are interpreted relative to the “base” category (often the most common choice), and every other outcome has its own set of parameter estimates. One consequence of this is that when there are many choice outcomes, interpretation of the results is difficult. In particular, neither the sign nor the magnitude of the parameter estimates indicates the direction of the influence of the independent variable on the outcome. Typically, marginal effects must be calculated for each variable of interest.

A second difficulty with the multinomial logit model is that the observed attributes of the choices must provide a systematic mapping onto the agent’s preferences. If arbitrary attributes are used to categorize choices, the multinomial logit model will provide incorrect estimates of the effects of independent variables. This difficulty is known as the independence of irrelevant alternatives (IIA) problem. The issue of IIA points to the importance of understanding, classifying, and defining the choice set.

SEE ALSO Choice in Economics; Economics, Nobel Prize in; Information, Economics of; Logistic Regression; Variables, Latent