Heckman correction

The Heckman correction (the two-stage method, Heckman's lambda or the Heckit method) is any of a number of related statistical methods developed by James Heckman at the University of Chicago in 1976 to 1979 which allow the researcher to correct for selection bias. Selection bias problems are endemic to applied econometric problems, which make Heckman’s original technique, and subsequent refinements by both himself and others, indispensable to applied econometricians. Heckman received the Economics Nobel Prize in 2000 for this achievement.

Method

Statistical analyses based on non-randomly selected samples can lead to erroneous conclusions and poor policy. The Heckman correction, a two-step statistical approach, offers a means of correcting for non-randomly selected samples.

Heckman discussed bias from using nonrandom selected samples to estimate behavioral relationships as a specification error. He suggests a two-stage estimation method to correct the bias. The correction is easy to implement and has a firm basis in statistical theory. Heckman’s correction involves a normality assumption, provides a test for sample selection bias and formula for bias corrected model.

Suppose that a researcher wants to estimate the determinants of wage offers, but has access to wage observations for only those who work. Since people who work are selected non-randomly from the population, estimating the determinants of wages from the subpopulation who work may introduce bias. The Heckman correction takes place in two stages.

In the first stage, the researcher formulates a model, based on economic theory, for the probability of working. The canonical specification for this relationship is a probit regression of the form

$\operatorname{Prob}( D = 1 | Z ) = \Phi(Z\gamma),\,$

where D indicates employment (D = 1 if the respondent is employed and D = 0 otherwise), Z is a vector of explanatory variables, $\gamma$ is a vector of unknown parameters, and Φ is the cumulative distribution function of the standard normal distribution. Estimation of the model yields results that can be used to predict this employment probability for each individual.

In the second stage, the researcher corrects for self-selection by incorporating a transformation of these predicted individual probabilities as an additional explanatory variable. The wage equation may be specified,

$w^* = X\beta + u\,$

where $w^*$ denotes an underlying wage offer, which is not observed if the respondent does not work. The conditional expectation of wages given the person works is then

$E [ w | X, D=1 ] = X\beta + E [ u | X, D=1 ].\,$

Under the assumption that the error terms are jointly normal, we have

$E [ w | X, D=1 ] = X\beta + \rho\sigma_u \lambda(Z\gamma),\,$

where ρ is the correlation between unobserved determinants of propensity to work $\varepsilon$ and unobserved determinants of wage offers u, σ _u is the standard deviation of $u$, and $\lambda$ is the inverse Mills ratio evaluated at $Z\gamma$. This equation demonstrates Heckman's insight that sample selection can be viewed as a form of omitted-variables bias, as conditional on both X and on $\lambda$ it is as if the sample is randomly selected. The wage equation can be estimated by replacing $\gamma$ with Probit estimates from the first stage, constructing the $\lambda$ term, and including it as an additional explanatory variable in linear regression estimation of the wage equation. Since $\sigma_u > 0$, the coefficient on $\lambda$ can only be zero if $\rho=0$, so testing the null that the coefficient on $\lambda$ is zero is equivalent to testing for sample selectivity.

Heckman's achievements have generated a large number of empirical applications in economics as well as in other social sciences. The original method has subsequently been generalized, by Heckman and by others.

Disadvantages

• The two-step estimator discussed above is a limited information maximum likelihood (LIML) estimator. In asymptotic theory and in finite samples as demonstrated by Monte Carlo simulations, the full information (FIML) estimator exhibits better statistical properties. However, the FIML estimator is more computationally difficult to implement.

• The covariance matrix generated by OLS estimation of the second stage is inconsistent. Correct standard errors and other statistics can be generated from an asymptotic approximation or by resampling, such as through a bootstrap.

• The canonical model assumes the errors are jointly normal. If that assumption fails, the estimator is generally inconsistent and can provide misleading inference in small samples. Semiparametric and other robust alternatives can be used in such cases.

• The model obtains formal identification from the normality assumption when the same covariates appear in the selection equation and the equation of interest, but identification will be tenuous unless there are many observations in the tails where there is substantial nonlinearity in the inverse Mills ratio. Generally, an exclusion restriction is required to generate credible estimates: there must be at least one variable which appears with a non-zero coefficient in the selection equation but does not appear in the equation of interest, essentially an instrument. If no such variable is available, it may be difficult to correct for sampling selectivity.

Implementations in statistics packages

• R: Heckman-type procedures are available as part of the sampleSelection package.
• Stata: the command heckman provides the Heckman selection model.

See also

• Propensity score matching

External links

• Encyclopædia Britannica article
• Article on Heckman-MacFadden Nobel Prize.
• Nobel prize official description.