Heckman & McFadden (2000): The Science of Individual Choices and Hidden Biases

Making Sense of Messy Real-World Data

Economists rarely get to run controlled experiments. They can’t randomly assign people to be educated or uneducated, employed or unemployed, to take the bus or drive a car. Instead, they observe the choices people actually make — and try to draw conclusions from this messy, self-selected data.

Two fundamental problems plague this enterprise. First, the people we observe aren’t random — they’ve selected themselves into the situations we’re studying, creating selection bias. Second, many of the most important economic decisions are discrete — buy or rent, work or stay home, take the train or drive — and standard statistical tools designed for continuous variables can’t handle them properly.

James Heckman solved the first problem. Daniel McFadden solved the second. Together, they built the toolkit of modern microeconometrics.

Heckman: The Selection Bias Problem

Imagine you want to know the effect of a job training program on wages. You compare wages of people who participated with those who didn’t. The participants earn more — so the program works, right?

Not necessarily. People who voluntarily enroll in training programs are probably more motivated, more ambitious, or more capable than those who don’t. They might have earned more even without the program. The difference in wages reflects both the program’s effect and the pre-existing differences between participants and non-participants. This is selection bias — and it’s everywhere:

• Studying the returns to college? College graduates differ from non-graduates in ways beyond education
• Measuring the wage gap between men and women? Women who work may differ systematically from women who don’t
• Evaluating a medical treatment? Patients who choose treatment may be sicker — or healthier — than those who don’t

Heckman’s breakthrough (1979) was the Heckman correction (or Heckit model) — a two-step procedure that accounts for selection bias:

Step 1 — Model the selection: Estimate the probability that each person is observed in the sample (e.g., the probability of working, of enrolling in training). This uses a probit model and produces the inverse Mills ratio — a measure of how selected each observation is

Step 2 — Correct the outcome equation: Include the inverse Mills ratio as an additional variable in the outcome regression (e.g., the wage equation). This variable captures the selection effect, and including it removes the bias from the other coefficients

The elegance: by explicitly modeling why people are in the sample, you can correct for the fact that they’re not random. The method is simple enough to implement in any statistical software, yet theoretically grounded in rigorous econometric theory.

Heckman’s Broader Contributions

Beyond the selection correction, Heckman transformed how economists think about causality:

• Treatment effects: Developed the framework for estimating the causal effect of a “treatment” (education, training, policy) when assignment isn’t random. Distinguished between the average treatment effect (ATE) and the treatment effect on the treated (ATT) — recognizing that the effect of a program on participants may differ from its effect on the general population
• Evaluation of social programs: Showed that naive comparisons of program participants and non-participants can be wildly misleading, and developed methods to extract causal estimates from observational data
• Early childhood interventions: Applied his methods to show that early childhood programs (like the Perry Preschool Program) have enormous long-term returns — influencing education policy worldwide

McFadden: The Theory of Discrete Choice

Most economic decisions aren’t about “how much” but “which one.” Take the bus or drive? Buy a house or rent? Choose this job or that one? These are discrete choices — selecting one option from a finite set of alternatives.

Before McFadden, economists had no rigorous way to model such choices. Standard regression assumes a continuous dependent variable. McFadden built the theoretical and econometric framework for discrete choice analysis.

The foundation: random utility theory. Each person assigns a utility to each alternative. They choose the option with the highest utility. But the researcher can’t observe all the factors that influence the choice — some are random from the researcher’s perspective. This randomness generates probabilities:

The conditional logit model (1974): McFadden derived the probability that a person chooses a particular alternative as a function of the characteristics of all available alternatives and the characteristics of the person. The model has an elegant mathematical form — the logit — that makes estimation tractable.

Key innovations:

• Characteristics of alternatives matter: Unlike previous models that only used characteristics of the chooser, McFadden’s model incorporates features of the options themselves — travel time, cost, comfort — allowing analysis of how changes in alternatives affect choices
• The independence of irrelevant alternatives (IIA): The basic logit model implies that the relative probability of choosing between two options doesn’t depend on what other options are available. McFadden recognized this limitation and developed more flexible models — nested logit, mixed logit — that relax it
• Welfare analysis: McFadden showed how to calculate the economic value of new alternatives or changes in existing ones — essential for cost-benefit analysis of public projects

The BART Application: Theory Meets Practice

McFadden’s most famous application was predicting ridership for San Francisco’s Bay Area Rapid Transit (BART) system — before it was built. Using his discrete choice model, he estimated how commuters would choose between driving, taking the bus, and taking the new train, based on travel times, costs, and individual characteristics.

His predictions proved remarkably accurate — and the methodology became the standard tool for:

• Transportation planning: Predicting demand for new highways, rail lines, and transit systems worldwide
• Environmental policy: Estimating how people respond to congestion pricing, fuel taxes, or electric vehicle subsidies
• Marketing: Understanding consumer choice among competing products
• Telecommunications: Predicting adoption of new technologies and services

Their 2000 Nobel Prize was awarded to Heckman “for his development of theory and methods for analyzing selective samples” and to McFadden “for his development of theory and methods for analyzing discrete choice.”

Explain It to a Child

Imagine you want to know if eating vegetables makes kids run faster. You compare kids who eat vegetables with kids who don’t. But wait — maybe the kids who eat vegetables also exercise more, sleep better, and have parents who care more about health. So you can’t tell if it’s the vegetables or everything else. Heckman figured out a way to untangle this — to separate the vegetable effect from all the other stuff. Now McFadden solved a different puzzle: when you’re at an ice cream shop choosing between chocolate, vanilla, and strawberry, what determines your pick? He built a math formula that predicts which flavor people choose based on the price, how much they like sweetness, and what other flavors are available. Cities used his formula to predict whether people would ride new trains — before the trains were even built.