Analyses
We used an intention-to-treat (ITT) analysis to assess the impact of
treatment assignment on four educational outcomes among children over 6
years old at midline. namely: 1) overall school enrollment, 2) domestic
responsibilities as the primary reason for non-enrollment (among the
unenrolled), 3) average days of school per week, and 4) average hours of
school per day among older siblings. We relied on the following survey
items to assess time spent in school: “In the current/most recent
school year, in a typical week, how many days does (the child) attend
school?” and “In the current/most recent school year, in a typical
day, how many hours does (the child) attend school?”. We used a more
general survey item (“Is the child currently attending school this year
as a student?”) to measure overall school enrollment. We also examined
non-enrollment in school due to domestic obligations by recoding the
survey item “Why is (the child) not in school this year?”. This item
originally contained 15 possible responses (“Too old”,
“Ill/disabled”, “Lost interest”, etc.). We were primarily interested
in capturing children who were “Needed for housework” or required to
“Care for younger siblings”. Because domestic duties and childcare
responsibilities likely overlap, both in practice and likely in the
minds of survey respondents, we combined these two categories to form a
single indicator variable reflecting whether a child was unenrolled in
school during the survey year due to domestic (household or childcare)
responsibilities.
All models accounted for the baseline value of the outcome variable of
interest and geographic clustering/block group; models also incorporated
cluster-robust standard errors (at the hamlet level). We used ordinary
least squares (OLS) models for continuous outcomes. For dichotomous
outcomes, we produced estimates with both logistic and log-binomial
models; logistic models are a more common approach, but binomial models
(despite their potential convergence issues) allow us to easily estimate
risk ratios instead of odds ratios and are appropriate when the outcome
is common. We extended our basic models in two ways: first, we
stratified by sex to examine the potentially heterogeneous effect of the
childcare intervention on girls versus boys. Second, we conducted a
per-protocol analysis to assess the effect of the intervention on
households that fully complied with their treatment allocation. We
defined a “compliant” household as one where all eligible children (6
years old or younger at midline) abided by the household-level treatment
assignment. This is an imperfect definition, as a household could be
coded as non-compliant even if most (but not all) of its eligible
children acted in accordance with their treatment assignment. We expect
that more sophisticated markers of compliance will be available at a
later date, but this preliminary strategy allows us to better estimate
the impact of the intervention among children who were actually exposed.
We also stratified these models by sex. All analyses were conducted in
Stata 14 (StataCorp. 2015. Stata Statistical Software: Release
14 . College Station, TX: Stata
Results
Descriptive characteristics of the subsample are summarized in
Table 1 . The sample remained well-balanced following
restriction: treatment and control groups were quite similar across all
attributes at baseline. Use of balwadi (Seva Mendir) was reported for
16.4% of all children in treated households at midline, but restricting
to eligible children (under 6 years old) yields a more accurate
treatment uptake estimate of 28.5%. This was lower than anticipated,
likely due to the short window of time (and thus limited opportunity for
uptake) from baseline to midline; we expect that household-level uptake
of the intervention will increase substantially by the final wave of the
study. Use of balwadi among children residing in control
households was low (3.2%), as anticipated.
Educational outcomes, including the ability to both read and write and
the average time spent in school per day/week, improved slightly in both
treatment and control groups from baseline to midline; this is to be
expected, as many of these factors would be expected to increase with
age. The striking shifts in where children spent most of their day from
baseline to midline are likely attributable to seasonal shifts in school
attendance (responses to this question would vary based on when in the
calendar year they are surveyed) (Verify?).