There are many estimation procedures that correct for sample selection. The most famous is Heckman's two-step selectivity correction (in two equations) that assumes bivariate normality of the error terms in both equations.

Olsen (1980) ["A Least Squares Correction For Selectivity Bias", Ecta.] showed that it is sufficient to assume a distributional form for the errors of the selection equation $\nu_i$ and linearity of the conditional expectation of $u_i$ given $\nu_i$, where $u_i$ is the error term of the regression model of interest.

How can the OLS selectivity correction be derived? This should be simple to see, but I would appreciate if someone writes down the derivation step by step.

Olsen starts from the following specification and assumptions:

The regression model of interest is $$y_i=X_i\beta+u_i.$$ $y_i$ is observed if and only if $I_i = 1$ where

$$I_i = 1 \text{ iff } \nu_i<Z_i\gamma,$$ $$I_i = 0 \text{ iff } \nu_i\geq Z_i\gamma.$$

He assumes $X_i$ and $Z_i$ are row vectors which conform to the column vectors of unknown coefficients $\beta$ and $\gamma$, respectively. The $X$'s and $Z$'s are exogenous.

Further, he assumes that the expected value of the error of the intensive regression is zero $$E(u_i)=0,$$ $$E(u_iu_j)=\sigma^2_u \text{ for } i=j,$$ $$E(u_iu_j)=0 \text{ for } i\neq j,$$ $$E(\nu_i)=\mu_\nu,$$ $$E[(\nu_i-\mu_\nu)(\nu_i-\mu_\nu)]=\sigma^2_\nu \text{ for } i=j,$$ $$E[(\nu_i-\mu_\nu)(\nu_i-\mu_\nu)]=0 \text{ for } i\neq j,$$ $$Cov(u_i,\nu_i)=E(u_i\nu_i)-E(u_i)E(\nu_i)=\rho \sigma_\nu \sigma_u \text{ for } i=j,$$ $$Cov(u_i,\nu_i)=0 \text{ for } i\neq j,$$ $$E(u_i|\nu_i)=\rho(\nu_i-\mu_\nu)\sigma_u/\sigma_\nu.$$

Here I got stuck.

Therefore my question 1 is: How can it be shown that $E(u_i|\nu_i)=\rho(\nu_i-\mu_\nu)\sigma_u/\sigma_\nu$ (see also p. 1816 in Olsen, 1980)?

Olsen goes on by assuming the conditional expectation of $u_i$ given $\nu_i$ is linear in $\nu_i$, so we can use the decomposition $$u_i=\rho(\nu_i-\mu_\nu)\sigma_u/\sigma_\nu+\varepsilon_i,$$ where $\varepsilon_i$ and $\nu_i$ are uncorrelated.

This is $$y_i=X_i\beta+\rho(\nu_i-\mu_\nu)\sigma_u/\sigma_\nu+\varepsilon_i.$$

And it follows that $E(y_i|X_i,\nu_i<Z_i\gamma)=X_i\beta+\rho\sigma_uE(\nu_i|\nu_i<Z_i\gamma)/\sigma_\nu-\rho\sigma_u\mu_\nu/\sigma_\nu$.

Now, assuming that the $\nu_i$'s were uniformly distributed over $[0,1]$, he finds $E(y_i|X_i,\nu_i<Z_i\gamma)=X_i\beta+\rho\sigma_u\sqrt{3}(Z_i\gamma-1)$. But how?

Therefore my question 2 is: What are the steps from $E(y_i|X_i,\nu_i<Z_i\gamma)=X_i\beta+\rho\sigma_uE(\nu_i|\nu_i<Z_i\gamma)/\sigma_\nu-\rho\sigma_u\mu_\nu/\sigma_\nu$ to $E(y_i|X_i,\nu_i<Z_i\gamma)=X_i\beta+\rho\sigma_u\sqrt{3}(Z_i\gamma-1)$ (see also p. 1816/1817 Olsen, 1980)?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.