Sensitivity analysis of job-training effects on reemployment for Korean women

Abstract

The main difficulty in treatment effect analysis with matching is accounting for unobserved differences (i.e., selection problem) between the treatment and control groups, because matching assumes no such differences. The traditional way to tackle the difficulty has been “control function” approaches with selection correction terms. This paper examines relatively new approaches: sensitivity analyses—sensitivity to unobservables—in Rosenbaum (Biometrika 74:13–26, 1987), Gastwirth et al. (Biometrika 85:907–920, 1998) and Lee (J Appl Econ 19:323–337, 2004). These sensitivity analyses are applied to the data used in Lee and Lee (J Appl Econ 20:549–562, 2005) to see how the assumption of no unobserved difference in matching affects the findings in Lee and Lee, to compare how the different sensitivity analyses perform, and to relate the “sensitivity parameters” in the different sensitivity analyses to one another. We find (i) the conclusions in Lee and Lee are weakened in the sense that only the “strong” ones survive, (ii) the sensitivity analysis in Rosenbaum (Biometrika 74:13–26, 1987) is too conservative (and inferior to Gastwirth et al.’s), and (iii) Gastwirth et al.’s and Lee’s approaches agree on some findings to be insensitive, but the two approaches also disagree on some other findings.