This work investigates how data availability can lead to different conclusions about the so-called fading-out effect of early childhood education (ECE). We explore a extensive dataset from 2008 and 2012 on elementary school children from a small municipality in Brazil. This data allows us to evaluate the effect of ECE attendance on language proficiency score using either cohort-based or grade-based data. We estimate and test if there is a difference in ECE effect on proficiency between 2008 and 2012. Our results indicate no fading-out of ECE effects while using cohort-based data, but a complete fade away of effects with grade-based data. We interpret those results as a sample selection issue caused mostly by ECE effect on grade progression.