Earlier, I had blogged about the use of cameras to tackle teacher absenteeism in rural India’s schools. One commenter asked for more details, and I went digging for the original research paper and found that it is quite interesting and worth writing about.
This data comes from a paper by Esther Duflo, a professor at MIT. The actual teacher program incentive was run by Seva Mandir an NGO. Duflo and her colleagues analyzed the results and reached some interesting conclusions.
What exactly is the question that we are trying answer?
We ask three main questions: If teachers are given incentives to attend school, will they actually attend school more? If they attend school more, will they teach more? Finally, if teacher absenteeism is reduced, will children learn more?
Consider the first question. It is not obvious that increased incentives to teachers will result increased attendance:
[…] the incentives could fail to improve attendance for a variety of reasons. First, teachers may be unable to take advantage of the incentives if they must participate in village meetings, training sessions, election or census duty. These pressures may be particularly high on para-teachers, who are often among the few literate individuals in the village. Second, the incentive schemes may crowd out the teacher’s intrinsic motivation to attend school (Benabou and Tirole, 2006). Finally, some teachers, who previously believed that they were required to work every day, may decide to stop working once they have reached their target income for the month (Fehr and Gotte, 2002).
The last two sentences of that paragraph point towards some interesting research showing how well-meaning schemes can actually be counter productive. (I am lazy so I haven’t included the full references here. If you are interested, go to the original paper and find the references at the end of the paper.)
Anyway, even if we manage to get teachers into the classrooms, that might not mean anything:
Even if incentives increase teacher attendance, it is unclear whether child learning levels will actually increase. Teachers may multitask (Holmstrom and Milgrom, 1991), reducing their efforts along other dimensions. Such schemes may also demoralize teachers, resulting in less effort (Fehr and Schmidt, 2004), or may harm teachers’ intrinsic motivation to teach (Kreps, 1997). On the other hand, incentives can improve learning levels if the main cost of working is the opportunity cost of attending school and, once in school, the marginal cost of teaching is low. In this case, an incentive system that directly rewards presence would stand a good chance of increasing child learning. Thus, whether or not the incentives can improve school quality is ultimately an empirical question.
Again, things are not as simple as one would initially have though. Will all that background out of the way, we get on to the real research:
We study a teacher incentive program run by the NGO Seva Mandir. Seva Mandir runs single teacher NFEs in the rural villages of Rajasthan, India. As in many rural areas, teacher absenteeism is high, despite the threat of dismissal for repeated absence. In our baseline study (August 2003), the absence rate was 44 percent. Faced with such high absenteeism, Seva Mandir implemented an innovative monitoring and incentive program in September 2003. In 57 randomly selected program schools, Seva Mandir gave teachers a camera, along with instructions to have one of the students take a picture of the teacher and the other students at the start and close of each school day. The cameras had tamper-proof date and time functions, allowing for the collection of precise data on teacher attendance that could be used to calculate teachers’ salaries.
Yup, the attendance records were used to determine the salaries of the teachers. So there was a financial (dis)incentive scheme at work:
Each teacher was then paid according to a non-linear function of the number of valid school days for which they were actually present, where a “valid” day was defined as one for which the opening and closing photographs were separated by at least five hours and both photographs showed at least eight children. Specifically, they received Rs 500 if they attended fewer than 10 days in a given month, and Rs 50 for any additional day (up to a maximum of 25 or 26 days depending on the month). In the 56 comparison schools, teachers were paid a fixed rate for the month, and were told (as usual) that they could be dismissed for repeated, unexcused absences.
Well, that is the experiment. Now on to the results:
The program resulted in an immediate and long lasting improvement in teacher attendance rates in treatment schools, as measured through monthly unannounced visits in both treatment and comparison schools. Over the 30 months in which attendance was tracked, teachers at program schools had an absence rate of 21 percent, compared to 44 percent baseline and the 42 percent in the comparison schools. Absence rates stayed low after the end of the proper evaluation phase (the first fourteen months of the program), suggesting that teachers did not change their behavior simply for the evaluation.
Absenteeism halved! And stayed that way.
Now on to another interesting question. How much of the improvement was because of cameras (the monitoring) and how much of the effect was because of the financial incentive scheme? What if the financial incentives were different, or not present at all? That is where the MIT economists broke out their complex math formulae to figure this out:
To answer these questions, we exploit the non-linear nature of the incentive scheme to estimate a dynamic labor supply model using the daily attendance data in the treatment schools. The identification exploits the fact that the incentive for a teacher to attend school on a single day changes as a function of the number of days they attend school in the month, and the number of days left in the month. This is because they have to attend at least 10 days in a month to begin to receive the incentive (by working in the beginning of the month, the teacher builds up the option to work for Rs 50 per day at the end of the month). Indeed, regression discontinuity design estimates show that teachers work significantly more at the beginning of the month than at the end of the previous month, when they had not accumulated at least 10 days of work in that month.
We use this fact to estimate the teachers’ marginal utility of money. We allow serial correlation in the opportunity cost of attending school and heterogeneity in teachers’ outside option, and we use the method of simulated moments to estimate the parameters. Allowing for serial correlation and heterogeneity considerably complicates the estimation procedure, but we show that these features are very important in this application.
And according to their calculations:
We find that teachers are very responsive to the financial incentives: our preferred estimates suggest that the elasticity of labor supply with respect to the level of the financial bonus is 0.306. Furthermore, decreasing the number of days that workers must work until they are eligible for the incentive by a single day increases the expected number of days worked by about 1.29 percent. An unusual feature of this application is the ability to carry out convincing out-of-sample tests based on the randomized evaluation (as in Todd and Wolpin (2007)). When allowing for serial correlation and heterogeneity, we find that our model accurately predicts the difference in attendance in the treatment and the control group, as well as the number of days worked under a new incentive system initiated by Seva Mandir after the experiment.
I don’t think I really understand the meaning of the numbers in that paragraph, but basically, the financial incentives were very important. i.e. You can’t implement this scheme with just cameras alone. Simple threats “that they could be dismissed for repeated, unexcused absences” are not good enough. You need to be able to control their salaries.
But at the end of all this, the fruits of labor are sweet:
Student attendance when the school was open was similar in both groups, so student in treatment group received more days of instruction. A year into the program, test scores in the treatment schools were 0.17 standard deviations higher than in the comparison schools. Two and a half years into the program, children from the treatment schools were also 10 percentage points (or 62 percent) more likely to transfer to formal primary schools, which requires passing a competency test.
See full paper. More interesting than watching Twenty20 these days…