Discussion
Our analysis suggests that the availability of formal, affordable childcare may have an impact on patterns of school participation in older siblings, but not through an associated change in caregiving or other domestic responsibilities. The intervention was associated with a slight shift in school attendance in the overall sample (when measured in days of school per week). We did not detect an impact of exposure on the general probability of enrollment, the probability of being unenrolled specifically due to domestic responsibilities, or on the hours spent in school per day. The observed effect of childcare provision on days of school attendance per week was relatively small, but given the robustness of these findings it is reasonable to posit that this is a true effect, and one that may be amplified once data from all study waves are available for analysis.
To our knowledge, this is the first study to formally examine the impact of formal care on older siblings using data from a cluster-randomized study. This study design protects our analyses against several common sources of bias, but it is still necessary to acknowledge a number of limitations. Perhaps most importantly, our findings were based on two waves of data collection; this is a key limitation because there may not have been a sufficient amount of time between randomization to treatment and the midline survey to observe the full impact of exposure. Findings should therefore be viewed as preliminary until the final data are available. It will also be important to assess balwadi-level factors in future analyses, including date of first availability and number of days available/open, as these may be important mediating factors.
A central challenge in compiling this dataset was the amount of missing birthdate data (particularly month and day). Collecting precise survey data on date of birth is a widely acknowledged challenge in developing settings (21), and is a particularly important limitation in the present analysis as our inclusion criteria were dependent upon age. While the DHS describes a complex method to handling missingness (21-22), which essentially imputes values based on other “clues” in the data (ages of other siblings, etc.), we used a comparatively straightforward approach by randomly imputing missing dates from a uniform distribution. While this raises the possibility that certain children were inappropriately included or excluded from our subsample, we expect any misclassification to be non-differential given the randomized nature of the study design. Along similar lines, it is plausible that data completeness is positively correlated with higher socioeconomic position, which would in turn be positively correlated with school participation, but this too should be non-differential.
Finally, our analytical subsample because increasingly restricted over the course of our analyses, which certainly had an impact on the precision of our estimates and may have compromised the balance of the treatment and control groups. Although we demonstrate in Table 1 that these two groups remained exchangeable following basic restriction, further modifications of the subsample may have distorted this balance. Our analyses may have also been underpowered to detect a meaningful effect of the treatment on our outcomes of interest due to (sometimes sizable) reductions in sample size; this concern is particularly true for the sex-stratified per-protocol analyses.
This analysis offers a first glimpse at the causal impact of formal childcare on educational uptake in older siblings. However, it is important for future work to more comprehensively explore the nuances of this relationship. Once the full dataset is available, we plan to more rigorously test this hypothesis by including a longer period of follow-up time (which should have a two-fold effect of increasing compliance and amplifying any benefits of the intervention).