Remedying Education: Evidence from Two Randomized Experiments in India

Publication Details

Quarterly Journal of Economics, August 2007, v. 122, iss. 3, pp. 1235-64. Available From:

Link to Source
Abhijit Banerjee, Shawn Cole, Esther Duflo and Leigh Linden
Institutional affiliations
None specified
Grant-holding institution
None specified
South Asia
Distance Education/ Education Technology, Educational Inputs, System Reform & Capacity Building
Distance Education/ Education Technology, Educational Inputs, System Reform & Capacity Building
Equity Focus
None specified
Evaluation design
Instrumental Variables (IV), Randomised Control Trials (RCT)
Journal Article


Education policies in developing countries have focused on increasing school enrolment but have paid less attention to quality. This study evaluates whether supplementary school inputs improve school outcomes in Gujarat, India.
In the first intervention, the weakest children worked with young women (balsakhi) for about 2 hours per day. In Vadodara, in the first year (2001–2002), schools were stratified on the basis of language, baseline score and gender before 49 randomly chosen schools received balsakhi in grade 3 (group A) and 49 received them in grade 4 (group B). In the second year (2002–2003), group A schools were assigned balsakhi in grade 4 and group B schools in grade 3. In addition, 24 new schools were randomly assigned to group A or B.
In Mumbai, in the first year (2001–2002), after stratification by baseline score and language of instruction, half of 77 schools were randomly chosen to receive a balsakhi in grade 3 (group C) and the remaining half in grade 2 (group D). In the second year (2002–2003), group C and D schools received balsakhi in grades 4 and 3, respectively.
In the second program in Vadodara, children in grade 4 played computer games solving math problems 2 hours per week. Schools were stratified by treatment status for balsakhi, gender, language of instruction and average baseline math scores. A total of 55 schools received the computer-assisted learning programme in the first year of implementation (2002–2003) and 56 schools served as the comparison group. The groups’ treatment status was switched in the second year (2003–2004).
Change in students’ test scores was regressed on treatment status and pretest score to measure the impact of the intervention. To determine if the impact of the remedial education was direct (that is, due to working with balsakhi) or indirect (that is, due to a reduction in class size), an instrumental variable (IV) regression was run in which predicted probability of being assigned to balsakhi was used as an instrument for actual assignment.

Main findings

The balsakhi programme improved average math and language test scores in the first and the second years of the intervention by 0.14 and 0.28 standard deviations, respectively. Being exposed to the program 2 years in a row caused gains of 0.25 and 0.41 standard deviations in Vadodara and Mumbai, respectively, an effect that is larger than being exposed for 1 year alone. Not surprisingly, the computer-assisted learning programme had an impact only on math scores (0.39 standard deviations) and not on language scores (–0.03 standard deviations).
Disaggregating test score gains by baseline ability shows that students in the bottom third of baseline test score distribution gained about twice as much as students in the top third (0.47 standard deviations vs. 0.23 standard deviations) due to the balsakhi programme. The computer-assisted learning programme increased math scores by 0.42 and 0.27 standard deviations for students in the bottom and top terciles, respectively.
A year after the end of the balsakhi programme, the gain in test scores had attenuated to 0.04 standard deviations. However, the effect for students in the bottom tercile hovered around 0.10 standard deviations for both math and language scores. This effect was higher than that for students in the middle and top terciles. Similarly, a year after the end of the programme, the impact of computer-assisted learning on math scores fell to 0.09 standard deviations.
The authors disentangle the direct and indirect effects of the balsakhi program using predicted probability of assignment to the balsakhi as an instrument for actual assignment. The effect of the programme on the children who worked with balsakhi was large (0.60 standard deviations), compared with the results for students who stayed in regular class (0.06 standard deviations).

Scroll to Top