Effects of kindergarten retention for at-risk children’s mathematics development

Abstract

When a child does not seem to be ready for primary school, a popular practice is to grant the child more time by letting it repeat kindergarten. However, previous quasi-experimental research demonstrated negative, though diminishing, effects of kindergarten retention on academic learning during the first years of primary school. The present study extends the existing evidence by addressing children’s post-treatment school trajectories. Analysing data from a large-scale longitudinal study, we find that, on average, kindergarten repeaters would perform better for mathematics until five years later, were they promoted to first grade instead. However, if promoted instead, kindergarten repeaters would also have a higher likelihood to be retained in first grade and, under that condition, have a lower growth rate and score lower for mathematics five years later.

Keywords:

• kindergarten retention
• mathematics development
• propensity score matching
• longitudinal study
• educational effectiveness

Acknowledgements

This work was supported by the Research Foundation Flanders (FWO) [Grant number G.0444.10N].The authors wish to thank the schools, students and their parents for the effort put into cooperating, as well as the SiBO team of the Centre for Educational Effectiveness and Evaluation for collecting the data.

Notes

1. These numbers are based on a weighted frequency table using a matched sample that is representative for Flanders. More details are available upon request from the first author.

2. Probability of promoted at-risk children to be retained in G1, G2, G3, G4, G5 or G6 is 31.93, 6.53, 5.84, 3.58, 1.10 and 0.94%, respectively.

3. To make it more easily readable, school years 2002–2003 to 2008–2009 are referred to as Years 0–Years 6, respectively. In Year 0, all children were in kindergarten, in Year 1, kindergarten repeaters were repeating kindergarten and children in the control group were in Grade 1; in Year 6, regularly progressing children were in Grade 6 and repeaters were in Grade 5.

4. The variable d represents the effect size measure, estimated in terms of the respective standard deviations of the retained children’s outcome that year. Cohen (1988) labelled d values of 0.20, 0.50 and 0.80 as small, medium and large effect sizes, respectively. CI refers to the confidence interval of d.

5. Across the three comparisons considered, the ATE at Year 1 is consistently the most robust (IPs of 90.07, 89.33 and 49.13%, respectively), which makes sense since this is the time point at which we expect the students in our study to be the best characterised and matched. Comparison 1 and 2 exhibit similar levels of robustness, while comparison 3 appears to be the most vulnerable to bias. The ATEs at Year 6 again have similar IPs for comparison 1 and 2. These estimates are more vulnerable to bias, but at approximately 60%, a significant number of observations from the hypothetical population in which there is no treatment effect would need to be introduced to sway the inference. The corresponding IP for comparison 3 is 36.74%, which is still a large portion of our estimates that would need to be due to bias to invalidate the inference. Finally, the ATE for growth is the weakest of the results for comparisons 1 and 2. We find that although the ATE for growth of comparison 1 corresponds to the lowest IP for this comparison, it is still relatively high, i.e. 50.74%. On the other hand, the IP of the ATE for growth in comparison 2 obtains with the lowest IP of our estimated effects. This inference is therefore the most vulnerable to invalidation. In order to invalidate it, 17.47% of the estimate would need to be due to bias, or correspondingly one would have to replace 17.47% of the observations used to obtain this estimate, and assume the limiting case of no-measured effect for these replacement observations. Turning to the ATE for growth in comparison 3, we see that it has a very similar IP as the ATE at Year 6, i.e. 36.74%, and thus we believe this is also a convincing inference.

6. This concerns the matched, unweighted groups. In the estimation of the treatment effects, the control group subjects were weighted to be similar to kindergarten repeaters, which resulted in weighted standardised means of the covariates of both control groups to be similar to those of kindergarten repeaters.