How important are omitted variables, censored scores and self-selection in analysing high-school academic achievement?

Using a rich longitudinal data set from birth, we explore three estimation issues related to academic performance analysis. Our paper primarily examines the effect of omitting childhood and teenage characteristics (childhood ability, parental resources at different times and peer effects), which are traditionally unavailable in data sets. Additionally, we explore the potential endogeneity of pre-exam school-leaving choices (self-selection) to academic performance; and we demonstrate the effect of accounting for censored academic performance measures. We find that omitting background characteristics results in overestimation of coefficients on other characteristics (the effect of current income is overestimated by 0.21 standard deviations of the average academic performance and the effect of ethnicity by 1.38 standard deviations). This then affects the policy implications drawn: for the group who did not take the exam, the predicted performance goes from a fail to a C (or pass). We also find that accounting for censored academic performance measures affects the estimation results, but allowing for selection correction does not.