Bias reduction by imputation for linear panel data models with nonrandom missing

When no variables are observed for endogenous non-respondents of panel data, bias correction is available only for a limited class of instrumental variable estimators, which require strong conditions for consistency and often suffer from substantial efficiency loss. In this paper we examine a convenient alternative method of imputing the missing explanatory variables and then using standard bias-correction procedures for sample selection. Various bias-corrected estimators are derived and their performances are compared by Monte Carlo experiments. Results verify efficiency loss by the instrumental variable estimators and suggest that the imputation method is practically useful if it is applied to first-difference regression.