Possible endogeneity of treatment variable

The treatment has been allocated across the tea plantations within the sampling frame based on managerial decisions about the order of roll-out of the program across plantations. Management report no systematic factors that influenced the decision about order, other than practical / administrative matters. However, if there are systematic unobserved differences between the treatment and control groups that are correlated with child development outcomes, this risks creating an endogeneity bias in model estimates. Baseline data can be used to check for similarity between treatment and control plantations. Since the midday meals intervention applies to children aged at least 6 months, we compare weight records of children up to 6 months of age across the treatment and control samples. Any differences cannot be attributed to the program, and instead could indicate relevant difference between the two groups.

The histograms of weight-for-age zscores (WAZ) in the 0–6 month age category across the treatment and control groups, together with the standard normal distribution curve, are presented in Fig. 1. According to this, the distribution of WAZ appear to be closer to the standard normal distribution within the control group than the treatment group. The two-sample Kolmogorov-Smirnov test for equality of distribution functions, rejects the null hypothesis of equality in distribution between the treatment and control groups (Combined K-S: 0.22, p < 0.000). These results suggest that children in the treatment group may be starting at a relative growth disadvantage compared to children in the control group, at the baseline.

Histogram of weight-for-age zscores in the 0–6 month age category in treatment and control groups

A number of methods are used to overcome potential endogeneity of the treatment variable, the most significant of which is to control for measures of past growth, using a lagged dependent variable (H_{i, t − 1} in eq. (1)) as a control variable in the models. We also use other measures of past growth (birthweight, first recorded weight and height) as control variables. Proxy variables are also included to control for important unobserved time-invariant factors, where applicable. Together these measures should account for any significant time-invariant factors that may drive systematic differences between the treatment and control samples with respect to the growth of children. In addition to this we also include proxy variables to control for possible time-varying factors which could also cause systematic differences between the treatment and control groups. Added together, these methods provide reasonable confidence that we are capturing a causal effect with the treatment variable, having overcome any possible significant bias in our identification strategy. We next give details of the main relevant unobserved time-invariant and time-varying factors whose impacts are accounted for through this analytical approach.