DARK SIDE OF INCENTIVES: EVIDENCE FROM A RANDOMIZED CONTROL TRIAL IN UGANDA* DAGMARA CELIK KATRENIAK§ Abstract Throughout our lives, we are routinely offered different incentives as a way to motivate us. Many researchers have studied the effects of incentives on people’s performance. There can also be important psychological outcomes in terms of stress and happiness. The current paper contributes to the literature by explicitly accounting for this performance-versus-well-being tradeoff introduced by incentives. I implement two types of social comparative feedback regimes, within and across-class group comparisons, and two types of incentive regimes, financial and reputation rewards. The results show that rewards can improve performance up to 0.28 standard deviations, but at a cost of higher stress and lower happiness, whereas comparative feedback alone (without rewards) increases performance only mildly, by 0.09 to 0.13 standard deviations, but without impact on student’s stress and happiness. More stressed students exert less effort, perform worse and attrite by 29 percent more compared to those who are stressed minimally. The results also help to identify gender-specific responses to incentives. While boys strongly react to rewards, girls do so only if they are also provided with feedback. Final contribution comes from a rich dataset of more than 5000 primary and secondary school students in Uganda, who were repeatedly tested and interviewed over one academic year.
Keywords: field experiment, Uganda, incentives, education, group competition JEL Classification: C90, C93, D04, I21, I29, O55 * This research was supported by GA UK grant No. 338911, and by GDN grant No.60. All errors are mine. ¶ CERGE-EI (a joint workplace of the Charles University in Prague and the Economic Institute of the Czech Academy of Sciences, Politickych veznu 7, P.O.Box 882, 111 21, Prague 1, Czech Republic. Email:
[email protected].
1
1. Introduction A trophy for the best student in a class, a certificate for the most improving learner of a
course or a bonus payment for the employee of the month, etc. We are routinely faced with
incentives of different types (symbolic, reputation or financial rewards) throughout our lives.
Rewards are often believed to motivate subjects and subsequently increase their performance, and
are therefore implemented in many different environments (Lazear, 2000, Fryer, 2010, etc). We are
also routinely compared to classmates/colleagues/competitors by receiving feedback about our performance. Feedback may motivate subjects to improve their performance (Andrabi et al, 2009; Azmat and Iriberri, 2010) even though the evidence on such positive effects is more scattered. 1 Feedback and incentives may also influence our well-being (Azmat and Iriberri, 2016) and change
in well-being may further influence people’s decision making and economic outcomes (e.g.,
Veenhoven, 1998, Juster et al., 2010, Helliwell et al, 2012; for more details see section Literature review).
This is a unique study implemented in the field that analyzes the effects of all types of
motivation schemes on performance and on the well-being (measured by perceived stress and happiness) of students evaluated in teams. The design offers a direct comparison of the effects of two feedback groups and two reward groups as well as their interactions (each feedback interacted
with each reward). Feedback differed across feedback-treatment groups with respect to its content.
Each student in the within-class feedback group (class randomly divided into groups of three to four students) received information about how he scored in Math and in English, how his group-mates scored and the position of the group within his class. Students in the across-class comparative
feedback group (comparison of entire class) received information about how they scored in Math 1
According to psychologists, positive feedback is believed to increase intrinsic motivation and foster longterm motivation, whereas negative feedback decreases intrinsic motivation (Burgers et al., 2015; Arnold, 1976; Deci, 1972).
2
and in English personally (i.e., they were not given information about their classmates) and the
position of their class compared to other classes. Students were not offered rewards until testing round 4 was finished. Students were orthogonally randomized into financial, reputation and noreward groups. Students in the financial reward group could win 2000 UGX per person (which was
approximately 0.80 US cents according to that day’s exchange rate). Students in the reputational reward group were promised that if they qualified for the reward, their names would be announced
in the local newspaper Bukedde (the most popular in the region) and they would receive a certificate. The general criterion I used was to reward 15% of the top performing and 15% of the
most improving students/groups/classes. The novelty of the experiment comes from the wide scope of outcome measures observed, rich design and its unique data set. The sample size consists
of more than 5000 primary and secondary school students from 146 classes located in Southern Uganda, who are repeatedly tested and interviewed during one full academic year. In total, five testing rounds were administered.
Looking at aggregated average treatment effects of feedback provision on students’ overall
performance (Mathematics and English pooled), students scored by 0.07 standard deviation more
compared to the control group students and the difference is insignificant on conventional levels (p=0.13 and 0.16). The results of feedback are, however, subject specific. While the students
improved significantly in Mathematics (0.08 to 0.13 standard deviations), they did not improve in English. Students exposed to rewards increased their performance on average by 0.1 standard
deviations if rewarded reputationally and 0.18 standard deviations if rewarded financially. The effects are similar when decomposed by subject. The effects are amplified if students face any of the
treatment combinations (the effect size is between 0.19 and 0.28 standard deviations). The results
on the outcomes other than learning - such as happiness and stress - put the benefit of reward
provision into the shade. The students who were offered only rewards (without any feedback) had their stress levels elevated and happiness levels decreased, whereas the well-being of students who 3
received only feedback remained unchanged. Moreover, most of the treatment combinations lead to
a decrease in students’ stress and increase or no effect on happiness. Thus, we can speak of an important trade-off: the introduction of rewards increases performance more than pure feedback,
but at the same time they lower students’ well-being. The effects persist when I control for multiple
comparison testing by adjusting the p-values using Simes step-up method (Simes, 1986).
In some experiments, boys and girls responded very differently to certain incentives. The
second contribution of this paper is to shed light on the underlying reasons behind these gender
differences. I attribute this difference to the existence of two types of competitions – intrinsic, or internally driven competition, developed by personal feelings based on comparison to others, and extrinsic competition induced by rewards. According to the results, if girls are given rewards but no
feedback, they will significantly underperform boys. If girls are, however, repeatedly informed
about their positions (no matter what type of feedback they receive), their performance will improve and will become comparable to boys. In other words, comparative feedback in a tournament environment plays a crucial role for girls motivating them to improve their
performance. Boys, in contrast, react only to rewards. The current design does not allow me to
distinguish whether gender differences are caused by the fact that students were evaluated in groups (group identity effect), or that they were repeatedly informed about their standing.
Nevertheless, since both within- and across-class feedback groups deliver similar effects, it seems
more likely that the effect is driven by social comparison rather than group identity. Such result would be in line with “reference group neglect” introduced by Camerer and Lovallo (1999) - students neglect information about others and focus solely on the feedback regarding her own performance.
The results of this experiment may be important especially for policy-makers in finding the
optimal strategy for improving performance and well-being of students in primary and secondary
schools. Despite numerous studies in the literature that are designed to improve students’ 4
performance and/or their attendance, concerns about students’ well-being have been discussed
minimally. Current well-being serves as an important prerequisite for future performance. Therefore, policy-makers should use a great amount of caution in designing educational rewards
and take into account the impacts on student well-being. Further research should aim to study the
long-term effects of changes in student well-being on performance to shed light into whether increase in stress today may increase or decrease benefits of tomorrow.
2. Literature Review
According to social comparison theory 2, informing a child about his/her performance without
comparing it to other children causes unstable evaluations of the child’s ability and can influence effort negatively (Festinger, 1954 3; the founder of the social comparison theory). On the contrary,
comparison enables a child to find his/her relative position within a particular group which can lead, via enhanced competitiveness, to an increase in effort and performance improvement. Feedback provision, as a way to inform subjects about their absolute or relative standing, has
been analyzed in different environments and has delivered opposing results. Andrabi, Das and IjazKhwaja (2009), for example, provided parents, teachers and headmasters with report cards
informing them how children are doing in a particular school. The intervention resulted in 0.1 standard deviation improvement in student test scores. Azmat and Iriberri (2010) informed high
school students about their relative standing and in this way helped to improve student grades by 5
per cent. Additionally, university students in the United Kingdom responded positively when they
improved their performance by 13% in response to feedback regarding their own absolute 2
Social comparison theory is about “our quest to know ourselves, about the search for self-relevant information and how people gain self-knowledge and discover reality about themselves” (Mettee and Smith 1977, p. 69–70). 3 Festinger in his original paper focused on the social comparison of abilities and opinions only. Since then, however, have many different dimensions of social comparison been studied (e.g., Buunk & Gibbons,1997, 2000; Suls & Wheeler, 2000). See e.g. Locke and Latham, 1990; Suls and Wheeler, 2000, for an overview of papers in psychology and management science. See Buunk and Gibbons (2007) for an overview of work in social comparison and the expansions of research on social comparison.
5
performance (Bandiera et al., 2015) 4. Not all studies, however, find positive responses to feedback provision. Azmat et al. (2015) do not find any effect of relative feedback on university student
performance (on the contrary, in a short period right after the feedback was provided they even find a slight downturn in student performance). More evidence on the negative effects of incentives
on performance can be found in experiments implemented at the workplace. Workers in two experiments lowered their performance after they received information about their rank position
Bandiera et al. (2011a, 2011b). Health workers also decreased their performance during a training program in Zambia when exposed to social comparison (Ashraf et al., 2014) 5.
The effect of feedback depends on who the subjects are compared with, how they are compared
and whether they are rewarded for their performance. Students face social comparison in their classrooms on a daily basis and it can strongly influence their self-esteem and their performance
(Dijskstra et al., 2008) as well as their well-being (Azmat and Iriberri, 2016). It is therefore important to understand with whom to optimally compare the students. If students are compared to the ones slightly better, their effort and performance tend to increase. Performance and effort
decrease if the comparison target is too far from a student’s ability (Ray, 2002). Students can be
compared individually or in groups. A group’s outcome depends on each member’s contribution and may foster mutual help (Slavin, 1984) in addition to positive peer effects (Hoxby, 2000; Sacerdote, 2001). Groups can be formed endogenously (e.g., by students themselves based on friendship) or exogenously (Blimpo, 2014) and they can be exposed to competition. In some
4
Other studies with positive effects of feedback provision: Tran and Zeckhauser (2012), Blanes-i-Vidal and Nossol (2010) or Fryer (2010) 5 There are also controlled lab environments studying the effects of feedback provision, e.g. Falk and Ichino (2006) and Mas and Moretti (2009) found that if one lets people observe the behavior of their peers, their performance would improve. Kuhnen and Tymula (2012) and Duffy and Kornienko (2010) find a positive effect to the provision of private feedback. Eriksson et al. (2009) on contrary find that rank feedback does not improve performance (even if pay schemes were used). Hannan et al. (2008) find a negative effect of feedback on relative performance under a tournament incentive scheme (if feedback is sufficiently precise).
6
studies, the effects of interventions are more pronounced if students are involved in tournaments (Eriksson et al., 2009; Bigoni et al., 2010; VanDijk et al., 2001) 6.
Subjects often improve their performance if they are rewarded financially. Bettinger
(2012), Angrist et al. (2002, 2006, 2009, 2010), Kremer (2004), Bandiera (2010), and others
studied the effects of the provision of cash payments, vouchers or merit scholarships to students
who successfully completed a pre-defined task. In such experiments knowing the relative position
is not crucial since success does not depend on the performance of other mates. In order to induce stronger competitive pressure, subjects need to be put into a tournament with a limited number of
winners. VanDijk et al. (2001) conclude, based on the results of their experiment in which they experimentally compared different payment schemes, that it is superior for a firm to introduce a tournament-based scheme over a piece-rate or team payment scheme. In the case of Blimpo (2014), groups involved in the tournament improved similarly compared to groups rewarded for reaching a
performance target. All treatments (with or without competition) resulted in positive improvement in student performance, increased by 0.27 to 0.34 standard deviations. The evidence is mixed. Fryer (2010) and Eisenkopf (2011) studied the impact of different financial rewards on student
performance and did not find any significant improvements (even though in the case of Fryer (2010) the effect might have not been detected because of lack of power, the author claims).
Even if the financial rewards result in performance improvements, they may not be
necessarily cost-effective (e.g., Bandiera et al., 2012 7). Alternative rewards 8 that would be possibly more cost-effective drew researchers’ attention. For example, Kosfeld and Neckerman
(2011) designed a field experiment where students in the treatment group were offered symbolic 6
See Hattie and Timperley (2007) for a review of the literature on the provision of feedback. Bandiera et al. (2012) find the financial rewards cost-ineffective since only a fraction of the students from the second quartile of initial ability distribution react positively to financial rewards. 8 See also theoretical models studying the effects of reputation and symbolic rewards on subjects’ performance in work of Weiss and Fershtman (1998), Ellingsen and Johannesson (2007), Besley and Ghatak (2008), Moldovanu et al (2007) or Auriol et al. (2008). 7
7
rewards (a congratulatory card) for the best performance while students in a control group were
not offered anything. Their results provide strong evidence that status and social recognition rewards have motivational power and lead to an increase in work performance on average by 12
percent. Subjects in the real-effort experiment conducted by Charness et al. (2010) increased their effort in response to the relative performance and expressed their “taste for status”. Jalava et al.
(2015) offered sixth grade students in Swedish primary schools different types of non-financial rewards (criterion-based grades from A to F, grade A if they scored among the top 3 students, a certificate if they scored among the top 3 students or they received a prize (in the form of a pen) if
they scored among the top 3 students). The effects were heterogeneous with respect to original ability (students from two middle quartiles respond the most to the incentives) and with respect to
gender (boys improved their performance in response to rank-based incentives only, girls also to symbolic rewards). Rank-based grading and symbolic rewards, however, crowded out intrinsic
motivations of students. Even non-monetary rewards have the power to motivate subject to
improve their performance. Naturally, the questions arose. What can we learn from direct comparison of monetary and non-monetary rewards? Would financial rewards prevail? Levitt et al. (2012) present the results of a set of field experiments in primary and secondary schools, in which
they provided students with financial as well as non-financial rewards, with and without delay and
incentives framed as gains and losses. Non-monetary rewards were as effective as monetary rewards (and therefore more cost-effective).
Feedback and incentives may also influence our psychological well-being (Azmat and
Iriberri, 2016). Change in well-being has been found to influence people’s decision making and
economic outcomes An increase in happiness 9 is associated with stronger health, sharper awareness, higher activity in addition to better social functioning (Veenhoven, 1998). Education is
one determinant of happiness (higher education is associated with higher well-being (Helliwell et
See Fordyce (1988) on a review of happiness measures, MacKerron (2012) for a review of the economics of happiness, Dolan et al. (2008) review well-being. 9
8
al., 2012; Dolan et al., 2008). Happiness is negatively related to stress. Subjects under stress make
suboptimal decisions, which, in the case of students, could lead to incorrect answers during examinations, or suboptimal choices in their careers (e.g., to be absent from school, to drop out of
school or to exert low levels of effort). Both stress and happiness influence subjects’ health (Juster et al., 2010; McEwen, 2008; Schneiderman et al., 2005). Stress influences learning and memory, and
it creates learning problems (Lubin et al., 2007; Wolf, 2009). In the extreme, stress hormones may even influence brain structure (Lupien et al., 2009). Therefore, the consequences of interventions on the well-being of students should not be underestimated.
This experiment differs from existing studies in its complexity of incentive schemes that
have been implemented and the broader scope of outcomes – except commonly used performance I study students’ confidence, stress, happiness and their academic aspirations. The predictions of the effects of my interventions based on the existing literature are controversial. Evaluation of students
in groups should push via enhanced cooperation within groups to group average improvements. If
the group is, however, big enough, free-riding behavior may prevail and result in heterogeneity within the group outcomes. Informing students about the position of their group could lead to
improvements in performances via enhanced competition or demotivate students with a negative
attitude toward competition. Alternatively, students could neglect information about their group members and focus solely on own performance (as found in Camerer and Lovallo, 2002). The effect
potentially depends on group gender and/or ability composition (as found in Apesteguia, Azmat,, and Iriberri, 2012) and group position in the group ability distribution. Students included in both financial and reputational reward treatments are expected to improve their scores, at least the ones in the second quartile of ability distribution.
9
3. Randomization and experimental design In this experiment, I study whether the provision of comparative feedback about own
performance and the performance of respective group members can influence students’
performance and their psychological well-being measured by stress and happiness. To evaluate the
effect of the intervention, I designed a Randomized Control Trial (RCT) experiment. At the
beginning of the academic year, the sample was stratified and randomized into two feedback-
treatment groups and one control group (as shown in Figure 1). Students in within-class feedback group were randomly divided into groups of three to four classmates and were evaluated as groups
within the class. In other words, group averages were taken into account when comparing the
students’ performance rankings. The students in the across-classes feedback group were evaluated
as a whole class (using class average) and were compared to other classes of the same grade in different schools.
Feedback differed across treatment groups with respect to its content. Each student in the
within-class feedback group received information about how he scored in Math and English, how his group-mates scored and the position of the group within his class. Furthermore, starting from
testing round 3, the student received information about how he (and his group-mates) improved or
worsened in between two preceding testing rounds. Students in across-class feedback group
received information about how they scored in Math and in English personally (i.e., they were not
given information about their classmates) and the position of their class compared to other classes.
The positions in both treatments were emphasized on a rank-order graph (see Appendix B4 and
B5). Students in the across-class feedback group received their first feedback with one-round delay
due to logistical issues. Students in the control group did not receive any information, they only answered exam questions. Students were not offered further rewards until testing round 4 was finished.
10
Figure 1: Stratification and randomization scheme
The relative standing of the group was based on the average group score from Mathematics and
English. Students were tested repeatedly during an academic year and received feedback three to four times depending on the feedback group (across-class/within-class feedback, respectively).
In order to study the effects of monetary and non-monetary rewards, I orthogonally re-
randomized the sample at the school level 10. The randomization (which happened three to four
weeks before the final testing round) divided the sample into 9 groups – one pure control group,
four pure treatment groups (i.e., one type of treatment only) and four interacted treatment groups
(two types of feedback interacted with two types of rewards). The scheme with all treatments and
the project timeline can be found in Appendix B2. Students were acknowledged about the exact rules of the competition during our personal visit and also via posters we left in each class.
10
The randomization was done at school level in order to avoid spillover effects and possible confusion.
11
The aim of such a cross-cutting design was to observe whether the introduction of rewards
could enhance student performance, especially if interacted with the within-class and across-class
feedback treatments. Furthermore, the design allows me to study also the effects of treatments on students’ well-being measured by happiness and stress. Students in financial treatment could win
2000 UGX per person (which is approximately 0.80 US cents according to that day’s exchange rate).
Students in the reputational reward scheme were promised that if they qualified for the reward, their names would be announced in the local newspaper Bukedde (the most popular in the region).
The qualification criteria differed based on original randomization into treatments (see Table 1) but the general rule was to reward 15% of the top performing students/groups/classes as well as 15% of the most improving students/groups/classes.
Table 1: Qualification criteria for winning the rewards Financial rewards (2000 UGX)
Within-class social comparison (Treatment 1) Across-class social comparison (Treatment 2)
Control group
15% of best performing and 15% of best improving groups (524 students) 15% of best performing and 15% of best improving classes (409 students) 15% of best performing and 15% of best improving students (498 students)
Reputational Rewards (Winners’ names published in a local newspaper) 15% of best performing and 15% of best improving groups (666 students) 15% of best performing and 15% of best improving classes (543 students) 15% of best performing and 15% of best improving students (585 students)
No rewards
Pure within-class social comparison group, no rewards (1205 students) Pure across-class comparison group, no rewards (1460 students) Pure Control Group, no rewards (1260 students)
Note: In order to avoid confusion, students were given exact information regarding the number of winning groups (if in T1), the number of winning classes (if T2) and the number of winning students (if originally in control group). I used percentages in order to guarantee a comparable number of winners across all treatment groups.
12
4. Timing, logistics and final sample The experiment took two years (August, 2011 – August, 2012) 11. The intervention
implementation and the core data collection took place from January 2012 until December 2012. Students were tested twice per term, i.e., approximately every one and half month. The agenda of
each visit was similar. After we entered the class, students in feedback-treatment groups received their feedback, control students started immediately answering the questionnaires and Math and English exam questions 12.
The final sample consists of 52 schools 13, 31 primary and 21 secondary schools out of which
19 are public, 23 are private and 10 are community schools. All schools were located in rural areas.
In total, 146 classes (P6 and P7 in primary schools, S1 up to S4 in secondary schools) summing up
to more than 5000 students were repeatedly tested. Apart from Math and English scores, I also collected information about student academic aspirations 14, immediate effort, strategic effort in a
form of preparation for the exam and immediate happiness revealed right before/after each exam. I
also repeatedly inquired about student expectations of their own score from Mathematics and 11 In 2011, I collected information about students’ basic demographic questions, questions regarding family background and family composition, parental status, education and job, wealth of the family and additional questions regarding the students’ interests, opinions, self-esteem and aspirations. Due to large attrition between 2011 and 2012 and the admission of new students throughout the 2012 academic year, the detailed information collected in 2011 is available for only circa 52% of students participating in the 2012 experiment. I also collected data about school (school type, school area, school fee structure and school equipment), headmasters and teachers (demographic information, years of experience, salary and their opinions). 12 The order was as follows: “Before Math questionnaire”, followed by Math examination that lasted 30 minutes; “After Math Before English questionnaire”, English exam in the subsequent 20 minutes and finally “After English questionnaire”. The core questions of the questionnaires were student expectations regarding how many points they thought they would obtain from Math and English examinations, how much effort they planned to put/they put into answering the questions and the level of their current happiness. All of these questions we asked before as well as after each exam. No before-Math and before-English questionnaires were collected during the baseline survey since students saw the examinations for the first time. 13 Initially there were 53 schools in my sample; one decided not to participate after I conducted the baseline survey. The school was initially randomized into the control group and its exclusion did not lead to significant differences in terms of baseline observables. 14 Students answered questions “What would you do if you were given an hour of extra time every day after school?” and had 15 binary scenarios to choose from. Out of 15 scenarios, five asked for a choice between educational activities (such as revise material taught at school) and work for money (such as selling vegetables on the market), five educational versus relaxing activities (such as talking to friends) and five work versus relaxing activities. I combined the answers into three indicators of the ordered preferences.
13
English in the testing in order to measure their confidence. To study students’ well-being, I
collected data on their happiness based on the Subjective Happiness Scale (Lyubomirsky and Lepper, 1997) and subjective stress based on the Perceived Stress Scale (Cohen, Kamarck and
Mermelstein, 1983). Happiness score is calculated as an average from four questions using the 7likert scale (with 1 being maximum and 7 minimum). Similarly, stress score is based on the answers
from four questions based on 5-likert scale when 1 means no stress and 5 maximum stress.. The questionnaires can be found in Appendix B2 and B3 15
5. Baseline summary statistics and randomization balance Data on student performance, demographics and student responses to questions suggests that
randomization divided the sample into groups that are similar in expectations (see Tables 2, and 3,
and Appendix A for the treatment-control group comparisons). Few significant differences can be
observed between across-class feedback and the control group, indicating that students in the
across-class feedback group were slightly more stressed, slightly less happy and exerted slightly more effort compared to the control group. If the covariates are correlated with student performance, such an imbalance could bias the estimation of the treatment effect of the intervention (Firpo et al., 2014). One can expect some imbalances between treatment and control 15
The headmasters of all schools agreed to participate in the experiment. The headmasters had an option to withdraw from participation throughout the entire duration of the experiment, nonetheless no school opted so. I asked the headmasters to communicate the content of the project to parents during their regular parental meetings. Besides headmaster consent, I also had full cooperation with a non-governmental organization UCDT that provided child sponsorship programs in Uganda. All schools in my sample cooperated with the UCDT – at least one of their students received sponsorship for his/her studies from the UCDT. In the letter of accordance in Appendix H, you can see their full support explicitly stated and that my project is in line with their goals. In order to minimize possible costs coming from our presence at school, the duration of the meetings was minimized to maximum 120 minutes. All meetings were organized with the headmasters one week in advance to find the most suitable and least harmful time in terms of delivered curriculum. Exam questions were designed based on the leaving examination questions. Administering exams in Mathematics and English was supposed to serve students as additional training for the leaving examinations they face during the final years of their studies in primary (grade 7) and lower secondary (grade 4) schools. There is no Institutional Review Board (IRB) for social sciences in the Czech Republic which could issue an IRB approval for my experiment. The experiment was designed in line with the conventions of IRB standards.
14
groups to occur purely by chance - as the number of balance tests rises, the probability to reject zero hypothesis of no difference between treatment and control group also increases. In my case, treatment and control groups differ significantly in less than 5% of all cases.
The average student scored 8.06 points out of 50 in the Mathematics exam and 14.07 points out of 50 in English. The real scores are in most of the cases below the student expectations. The
miscallibration of own performance is approximately 100 per cent. The average student put “a lot
of effort” into answering the exam questions (intensity 4 in the 5-likert scale) and he seems to be
“very happy” according to the immediate happiness scale (intensity 2 in the 7-likert scale when 1 is
the maximum). He finds the Mathematics exam of comparable difficulty and the English exam easier compared to the regular exams at school. The average student is overall quite happy (based on the Perceived Happiness Scale) and he has a low level of stress (Perceived Stress Scale). If the average student had a chance to have one hour of extra time every day, he would choose education over rest
in 4.3 cases out of 5; in 3.9 cases out of 5 he would choose education over work; and in 3.1 cases out
of 5 he would choose work over rest. The aspiration measures reveal the pro-educational preferences of students compared to work and rest.
15
Table 2: Randomization balance: comparison of mean characteristics of students in treatment and control groups, baseline tests and interviews Means Mean Differences Joint PWithin-class Across-class Control value Feedback Feedback (C) (T1 – C) (T2 – C) (T1) (T2)
PERFORMANCE (Baseline) Mathematics English
Sum Mathematics + English OTHER THAN PERFORMANCE Gender
8.063
8.838
8.655
-0.564§
23.088
-0.923
14.072
14.630
14.432
0.539
0.516
0.517
22.134
23.468
-0.359
0.197
0.198
0.395
0.183
0.699
0.426
0.022 -0.001 0.239 (0.015) (0.014) Age 17.058 17.049 16.999 0.059 0.049 0.737 (0.079) (0.078) Average class size 43.912 47.245 43.337 0.575 3.908 0.546 (3.208) (3.776) Expected number of points 4.331 4.536 4.552 -0.221 -0.015 0.299 from Mathematics (0.150) (0.145) Expected number of points 5.715 5.757 5.796 -0.081 -0.039 0.879 from English (0.161) (0.144) Perceived difficulty of Math 3.341 3.495 3.423 -0.082§ 0.072 0.030 exam (0.053) (0.052) Perceived difficulty of English 3.644 3.644 3.677 -0.033 -0.033 0.752 exam (0.052) (0.049) Immediate happiness after 3.287 3.226 3.132 0.155* 0.094 0.230 Math exam (0.092) (0.092) Immediate happiness after 2.909 2.869 2.782 0.127§ 0.087 0.303 English exam (0.085) (0.085) Effort put into Math exam 3.447 3.535 3.504 -0.057 0.021 0.298 (0.053) (0.052) Effort put into English exam 3.547 3.627 3.553 -0.006 0.074* 0.141 (0.046) (0.044) Subjective stress 1.504 1.588 1.439 0.065§ 0.149*** 0.001 (0.041) (0.036) Subjective happiness 2.869 2.913 2.806 0.064 0.107* 0.155 (0.058) (0.055) Education over work 3.538 3.496 3.477 0.060 0.019 0.526 (0.057) (0.059) Education over relax 3.834 3.756 3.778 0.056 -0.021 0.269 (0.049) (0.049) Work over relax 2.766 2.701 2.803 -0.037 -0.102 0.524 (0.094) (0.090) T1 stands for within-class social comparison group, T2 for across-class comparison group and C represents control group with no feedback provided. Robust clustered standard errors at class level are in parentheses, adjusted for stratification. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%.
16
Table 3: Randomization balance: comparison of mean characteristics of students, by treatment and control groups, baseline tests and interviews Means Mean Differences Joint PFinancial Reputation No Rewards value Reward (Fin) Reward (Rep) (No) (Fin – No) (Rep – No)
PERFORMANCE (Baseline) Mathematics
9.92
English
10.796
OTHER THAN PERFORMANCE Gender
0.553
Sum Mathematics + English
20.718
10.72
10.49
10.394
10.853
21.116
21.353
-0.507 (1.215) -0.096 (1.211) -0.603 (2.199)
0.292 (1.258) -0.497 (1.418) -0.206 (2.660)
0.788
0.937
0.955
0.039** 0.031* 0.089* (0.019) (0.017) -0.088 -0.267 0.798 Age 14.376 14.196 14.437 (0.350) (0.359) 0.388 10.087** 0.067 Average class size 45.434 55.137 45.987 (4.173) (4.332) -0.028 0.096 0.882 Expected number of points 4.839 4.964 4.917 (0.257) (0.236) from Mathematics 0.018 -0.002 0.994 Expected number of points 5.152 5.132 5.162 (0.255) (0.276) from English 0.043 0.121 0.499 Perceived difficulty of Math 3.283 3.361 3.256 (0.088) (0.093) exam 0.017 0.027 0.975 Perceived difficulty of English 3.407 3.417 3.398 (0.080) (0.085) exam -0.108 -0.091 0.793 Immediate happiness after 2.616 2.633 2.713 (0.159) (0.145) Math exam -0.053 -0.029 0.911 Immediate happiness after 2.504 2.529 2.548 (0.099) (0.112) English exam -0.053 -0.040 0.722 Effort put into Math exam 3.617 3.631 3.684 (0.083) (0.086) -0.027 -0.064 0.603 Effort put into English exam 3.526 3.489 3.554 (0.055) (0.064) -0.034 -0.084 0.912 Subjective stress 6.849 6.799 6.883 (0.230) (0.192) -0.324 0.234 0.226 Subjective happiness 10.376 10.933 10.671 (0.343) (0.278) -0.088 -0.176* 0.209 Education over work 3.822 3.735 3.910 (0.111) (0.098) -0.053 -0.020 0.748 Education over relax 4.234 4.266 4.296 (0.081) (0.077) -0.369*** 0.001*** Work over relax 2.767 3.250 3.141 0.115 (0.125) (0.102) Fin stands for financially rewarded group, Rep for reputationally rewarded group and No represents the control group with no rewards. Robust standard errors adjusted for clustering at class level are in parentheses. § significant at 15%, * at 10%; ** at 5%; *** at 1%. 0.545
0.514
17
6. The effects of incentives on students’ performance and their well-being The core question of the experiment is how different incentive schemes (social comparison,
financial and non-financial rewards) influence student performance and their well-being. I first
analyze the aggregated treatment effects (i.e., the overall treatment effect of the within-/acrossclass feedback in each testing round and the effects of financial and reputational rewards in the final testing round). Later, I disentangle the pure treatment effects (pure feedback and pure rewards) and the interacted effects (each type of feedback interacted with each type of rewards). I
discuss the role of group gender/ability composition and I study whether the type of feedback
students received matter for the improvement. Finally I look at the distributional analysis. 6.1. Average treatment effects on performance
In Table 4 I summarize the aggregated average treatment effects of feedback and rewards on
students’ overall performance (columns (1) to (4)) and on their subjective well-being (happiness
and stress in columns (5) and (6)). The effects are expressed in standard deviations. The provision of feedback (pooled or separately by feedback type) increases students’ overall performance by
0.07 standard deviations, which is considered as a small effect. In other words, the average student who received within-class or across-class feedback scored higher than 52.8 percent of students in
the control group. The type of feedback does not play a significant role. The aggregated results are similar in size compared to the results of students who only received feedback during the academic
year and were not included in any competition for rewards (for decomposed pure and interacted
treatment effects see Figure 4 and Appendices C1 and C2).The results are very similar to the results
of Jalava et al. (2015) who tested the effects of different grading schemes - 0.077 standard
deviations for criterion based grading, and 0.080 standard deviations for tournament grading. In
Pakistan, parents and teachers received report cards regarding the performance of their 18
children/students, which led to a 0.1 standard deviation increase in student performance (Andrabi et al., 2009).
Table 4: Aggregated average treatment effects of the provision of feedback and rewards on the overall performance and students’ subjective well-being AGGREGATED AVERAGE TREATMENT EFFECT ON:
Aggregated feedback treatment Within-class social comparison Across-class social comparison
STUDENTS’ OVERALL PERFORMANCE (1)
(2)
(3)
0.073* (0.041)
0.064 (0.039)
(4)
STRESS
HAPPINESS
(5)
(6)
§
§
0.061 -0.001 -0.111* (0.043) (0.090) (0.058) § Aggregated reward treatment 0.069 -0.104 -0.058 (0.082) (0.057) (0.046) § Financial Rewards 0.226** -0.108 0.133** 0.176*** (0.107) (0.052) (0.062) (0.070) § § Repurational Rewards 0.177 0.102** -0.112 (0.051) (0.119) (0.070) Yes Controlled for stratas Yes Yes Yes Yes Yes R-squared 0.714 0.715 0.645 0.058 0.634 0.078 N 5108 5108 5102 4105 5102 4056 Note: columns (1) – (4) show the average treatment effects (ATE) of differently aggregated treatments on students’ performance. Columns (5) and (6) represent the average treatment effects on students’ well-being (stress and happiness respectively). 0.074 (0.045) § 0.071 (0.047)
The aggregated effects of rewards are weaker (financial rewards) or comparable (reputational
rewards) compared to the existing results – 0.18 standard deviations in response to financial
rewards and 0.102 standard deviations to reputational rewards. The effect size depends on
whether students received feedback or not. Decomposition of the aggregated effects to pure and
interacted treatments shows that while rewarded students without feedback improved by 0.06-
0.13 standard deviations, students with feedback improved their performance by 0.12 to 0.28 standard deviations (for further details see Figure 3 and Appendices C1 and C2). In other words, the
average student rewarded with financial (reputational) rewards without any feedback scored more
than 55.2 percent (52.4 percent) of students in the control group. If the average student rewarded 19
reputationally received repeated (within/across-class) feedback during the academic year, she
scored more than 54.8 per cent/57.3 per cent of the control-group students. Similarly, the average
student who was rewarded financially and received within-/across-class feedback scored higher
than 57.9 per cent / 61.4 per cent of students in the control group. Jalava et al. (2015) find similar results. Students in their study who competed for a certificate which was given to first three students improved their performance by 0.083 standard deviations. Students who competed for (non-monetary) prizes improved by 0.125 standard deviations. Blimpo (2014) studied the effects of
financial rewards provided to students in Benin on individual or group basis. Students improved their performance by 0.27 to 0.34 standard deviations.
Table 5: Aggregated average treatment effects of the provision of feedback and rewards on the overall performance (Math and English pooled) – by subject AGGREGATED AVERAGE MATHEMATICS ENGLISH TREATMENT EFFECT ON STUDENTS’ (1) (2) (4) (5) PERFORMANCE – BY SUBJECT Aggregated feedback treatment Within-class social comparison (T1) Across-class social comparison (T2) Aggregated reward treatment Financial Rewards
0.094* (0.051)
0.128** (0.063)
0.099* (0.059) § 0.089 (0.056)
-0.009 (0.036)
0.126** (0.055)
-0.028 (0.039) 0.012 (0.040)
0.158** 0.142* (0.065) (0.078) 0.108** Repurational Rewards 0.115* (0.053) (0.064) Yes Yes Controlled for stratas Yes R-squared 0.645 0.688 0.645 N 5102 5093 5102 Note: columns (1) – (4) show the average treatment effects (ATE) of differently aggregated treatments on students’ performance. Columns (5) and (6) represent the average treatment effects on students’ well-being (stress and happiness respectively).
20
The effects on students’ performance differ by subject. While the effects of feedback provision
are driven solely by improvements in Mathematics (no improvements in English), the rewards lead to similar improvements in both subjects (see also Table 5). One explanation is that Math is more
elastic (Bettinger, 2012). It may be easier to detect the areas of Mathematics in which the student is
failing, while in English it may be hard to prepare for the test. It may also be a case of overall motivation. Students may have low motivation to study science, because science subjects are
usually perceived as more difficult 16 and students may not see their usefulness in real life; but once they are incentivized (students see real rewards instead of abstract future benefits), they improve.
Figure 2: The evolution of the aggregated average treatment effects of within-/across-class feedback on students’ overall performance (Math and English pooled)
16
Judging also by a consistently lower number of applicants for Science subjects as opposed to Arts subjects in the National examinations held by the Ugandan National Board Examination Committee.
21
Current data show that students in the control group, whose performance is mimicking student
evolution in absence of the treatments, have stagnated in Mathematics during the whole academic year (their absolute performance decreased by 0.33 per cent) but their absolute score in English increased by 50.25 per cent. Based on such progress, it may be easier to improve in Mathematics compared to English. Alternatively, the pattern may be the result of an order effect (the Math
examination always preceded English examination so students lost motivation to perform better). A
significant improvement in Mathematics, but not in English can be also found in other studies, e.g. Bettinger (2012) or Reardon, Cheadle and Robinson (2009). The evolution of the treatment effect can be found in Figure 2.
6.2. Average treatment effects on students’ well-being Both types of feedback leave students’ stress level unchanged but the within-class feedback
slightly decreases their happiness (by 0.111 standard deviations). The decomposition of the aggregated treatment effects show however that pure feedback does not influence students’ well-
being if provided without any rewards (the summary of aggregated treatment effects can be found
in Table 4, disentangled pure and interacted effects in Figure 3). The stress is induced once the feedback group students competed for rewards. Rewards, on the contrary, significantly increase the
stress level of students and decrease students’ happiness 17. Students who were not informed about
their performance over the year but were included in a competition for money reported their stress
level by 0.466 standard deviations higher compared to the control-group students. In other words, these students reported higher stress level 67.9 per cent of the control-group students who
received no feedback and no rewards. Provision of feedback significantly lowers the stress level induced by financial rewards. Students who received within-class feedback and competed for 17
In aggregated terms, financial rewards increased stress by 0.226 standard deviations and reputational rewards by 0.177 standard deviations; students’ happiness decreased by 0.1 standard deviations.
22
Figure 3:Dis-aggregated average treatment effects of incentives on students’ performance and their subjective well-being
23
financial rewards increased their stress level by “only” 0.222 standard deviations and students
inacross-class feedback group did report higher stress level compared to the control group students. Financial rewards decrease students’ happiness level but having repeated feedback does
not significantly change the results. The effect of reputational rewards seems to work in an opposite
way compared to financial rewards – pure reputational rewards do not influence stress level but if students received repeated feedback introduction of the rewards increases their stress level by
0.196 (within-class feedback) to 0.237 standard deviations (across-class feedback). In terms of
percentages, within-class (across-class) feedback students reported higher stress level in response to reputational rewards compared to 57.7 per cent (59.4 per cent) control group students but the
differences are only significant on 15%. Further details including alternative specifications can be found in Appendix C.
A policy maximizer who would want to minimize the effects of interventions on students’ well-
being should therefore consider a class-level competition for financial rewards with regular feedback regarding student’s own performance and the performance of her class – students’
performance increases with no significant effect on stress or happiness. 6.3. Group composition
If we let students to choose whether they want to compete in teams or as individuals, the
studies have shown that average-ability subjects have higher tendency to choose team-work
compared to high-/poor-ability subjects (Amann and Gall, 2006, Breton et al., 1998, 2003). In this study, I am interested in behavior of students exogenously grouped with students from the whole
ability-spectrum. Students were grouped into groups of three students and received feedback about their own as well as group performance during the whole academic year. In some cases the number
of students in the class was not divisible by three. In that case there were one or two groups of four 24
students (in total 18 groups out of 717 groups). In the following analysis I take only three-people
groups into account. Three types of ability groups (poor, mixed and good performers groups) and
four types of gender groups (pure boys, two boys and one girl, one boy and two girls, or pure girls) were formed. The analysis helps us to understand how well-informed groups who differ in terms of
the ability/gender composition perform in response to financial and reputational rewards. The
applicability of the results goes beyond educational framework. Teams are increasingly used indecision making processes in organizations compared to individuals (Hamilton et al., 2003,
Woolley et al., 2010). Companies spend large amount of money on the incentivization of their employees and with the aim to maximize efficiency or team performance they often carefully select
high-ability performers to work on a project/represent the firm, etc. In such environment the
results of this research offer comparison of responses of different ability groups with or without further incentivization.
6.3.1. Ability composition First, I compare the performance of mixed-ability and high-ability groups to poor-ability
groups and observe differences in responses to provision of financial and reputational rewards.
Students’ ability is measured in terms of students’ initial performance in Math and in English. All
poor-ability group students scored below the median, all high-ability students scored above the
median and the ability of mixed-ability group students varies across the whole performance distribution.
Students in the mixed ability group who did not compete for rewards do not outperform
poor-ability groups of students in Mathematics, but they do outperform them in English by 0.133
standard deviations. It means that in English the average student from mixed ability group scored
more than 55.2% of students from poor ability groups. However, once incentivized, mixed-ability
group students improved significantly more both in Math (0.21-0.22 standard deviations) and in 25
English (0.38 standard deviations and 0.224 standard deviations in response to the financial and
reputational rewards respectively). The type of the reward matters for the incentivization of mixedability group students in English. Students competing for money improved by 0.164 standard deviation more compared to students competing for reputational rewards (the difference is
significant at 5% confidence level). Students in the high-ability group who did not compete for
rewards outperformed significantly poor-ability group students by 0.451 standard deviations in Mathematics and 0.387 standard deviations in English. It means that students randomized into
high-ability group outperformed 65%-67.5% poor-ability students in English and Math
respectively. There is no statistically significant value added of rewarding high-ability group
students with neither pecuniary nor non-pecuniary rewards (see Figure 4 and Appendix E1).
Figure 4: Comparison of performance of mixed- and high-ability ability groups to poor-ability groups and their responses to the provision of rewards
Note: groups were randomly divided into poor-ability, mixed-ability, and high-ability groups based on the baseline performance. Groups of four students are excluded from the analysis (18 out of 717 groups). All groups received within-class feedback during the whole academic year. The bars show the average treatment effects of different incentive schemes of mixed- and high-ability groups in comparison to poor-ability groups. Stars indicate significance of the difference in means. § significant at 15%, * at 10%; ** at 5%; *** at 1%
26
Except two cases, students do not differ across different ability groups in perceived stress
and happiness. The average student from the mixed-ability group incurred higher stress compared
to 62.2% poor-ability group students when offered financial rewards and the average high-ability
group student incurred higher stress compared to 62.6% poor-ability students when offered
reputational rewards. Ability-composition of the groups does not seem to determine the level of effort exerted to perform the task.
6.3.2. Gender composition The performance of teams may also be influenced by gender composition of the group.
Apesteguia, Azmat and Iriberri (2012) studied the effects of gender compositions on economic
performance of undergraduate and MBA students involved in a business game competition. All
men and mixed-gender groups outperformed groups of pure women. The group composition of two
men and one woman seem to be optimal as they perform the best. In this study the group also consists of three students and due to random assignment four different gender-compositions were formed (pure girls, majority of girls, majority of boys and pure boys). The findings are similar.
Mixed groups outperform pure-gender groups, but I do not find strong support for the groups of two boys and one girl to be dominant.
In the Figure 5 and Table in the appendix E2 I compare the performance of mixed groups
and groups of boys compared to groups of girls. The results suggest that in the absence of rewards,
mixed gender groups outperformed uniform gender groups by 0.16-0.18 standard deviations. Pure boy groups performed comparably with pure girl groups. Once the rewards are offered, pure female
group is outperformed by all other types of groups by 0.19 up to 0.50 standard deviations
depending on the group gender composition and type of the rewards. Group composition does not
seem to play significant role in terms of students’ perceived stress in reaction to different treatments. Pure female groups seem to be on average happier compared to other groups. 27
Figure 5: Comparison of performance of mixed-gender and pure-boy groups to groups consisting of pure girls and their responses to the provision of rewards
Note: within-class feedback groups were randomly divided into three-boy-groups, groups of two boys and one girl (2B1G), groups of one boy and two girls (1B2G) or pure three-girl-groups. Groups of four students are excluded from the analysis (18 out of 717 groups). The bars show the average treatment effects of different incentive schemes of 3-boy-groups and mixed-groups in comparison to 3-girl-groups. Stars indicate significance of the difference in means. § significant at 15%, * at 10%; ** at 5%; *** at 1%
6.4. Distributional Analysis It is also important to learn whether the treatment effect differs at different point of the
performance distribution. To learn to what extent high performers differ in their reactions to the
treatment effects from low performers, I estimate quantile regressions. Figure 6 shows the average treatment effect of the incentives on performance of students by their rank in the performance
distribution. The Figure consists of four graphs which differ in terms of the combinations of the treatments. Graphs in the first column compare the average treatment effects of the pure feedback with feedback-reward interacted treatments. Graphs in the second column compare average
treatment effects of pure-reward with feedback-reward-interacted treatments. The main 28
observation is that the bottom performers respond stronger compared to the top performers to the
provision of pure financial rewards, pure across-class feedback or their interaction. In all other cases the bottom performesrs respond comparably to the top performers.
Figure 6: Distribution of the average treatment effects of different incentives on overall performance (Math and English pooled), by deciles
6.5. Positiveness and negativeness of the feedback The nature of the feedback students receive may influence students’ well-being. Azmat and
Iriberri (2016) show that positive feedback increases students’ happiness. I find similar results but
mainly for students who received feedback but no rewards. Besides feedback regarding own performance and the absolute and relative performance of the group, students also received additional information stressing whether they improved or worsened in two subsequent rounds
29
(starting from testing round 3). Students are considered to receive mostly positive (negative) feedback if in at least two out of three cases they improved (worsened).
Students in the within-class (across-class) feedback group who received mostly positive
feedback reported to be happier than 55.2 per cent (58.9 per cent) of the students in the within-
class (across-class) feedback group who received mostly negative feedback. While the introduction
of the reputational rewards eliminated the happiness surplus in both feedback groups, across-class
feedback group rewarded financially sustained their happiness surplus (see also Table 6). Table 6: The impact of the content of feedback on students’ stress and happiness Dependent variable: Stress and happiness Mostly positive feedback:
Stress
(1)
Happiness
(2)
Within-class feedback aggregated Within-class feedback, no rewards Within-class feedback with monetary rewards Within-class feedback with reputational rewards Number of observations
-0.012 (0.154)
Across-class feedback aggregated Across-class feedback, no rewards Across-class feedback with monetary rewards Across-class feedback with reputational rewards Number of observations
-0.007 (0.056)
(4)
0.103* (0.056)
1453
-0.117 (0.080) 0.060 (0.115) 0.149 (0.123) 1453
1451
-0.082 (0.160) 0.037 (0.209) 0.282 (0.199) 1451
30
(3)
1454
0.132** (0.063) 0.014 (0.092) 0.120 (0.102) 1454
1416
0.259*** (0.094) 0.221* (0.110) 0.013 (0.146) 1416
0.218** (0.088)
7. Gender differences and disentangling the channels of the average treatment effects Girls performed differently to boys in various studies. Angrist and Lavy (2009) studied the
effects of cash incentives on matriculation rates among Israeli students. Girls, contrary to boys,
substantially increased their performance. A higher effect among girls was also found in the
analysis of voucher provision within the PACES program in Colombia (Angrist et al., 2002).
Stronger responsiveness to incentives among girls can be also found in studies of tuition provision by Dynarski (2008), early childhood interventions by Anderson (2008), housing vouchers by Kling et al. (2007) or public sector programs by Lalonde (1995) and others 18.
The results of this experiment show that girls react positively to feedback provision (0.12 –
0.14 standard deviations) even if they are not offered rewards. Once included in a competitive
environment, girls improve by 0.2 to 0.28 standard deviations (see Tables 7 and 8, or Appendix D).
Therefore, girls can perform the same way as boys if they receive feedback about their performance, the performance of their group and the group’s relative standing. In the absence of
feedback, girls do not improve at all. Boys improved if they were offered rewards (with or without feedback) by 0.18 to 0.28 standard deviations but do not react to pure feedback provision. I attribute the gender difference in reaction to different treatments to the existence of two types of
competition: intrinsic, or internally driven, competition developed by personal feelings based on comparison to others, and extrinsic competition coming from offered rewards. These results are of
special help to policy makers whose aim is to influence the performance of both girls and boys.
Figures 7 and 8 compare the treatment effects across the whole performance distribution by gender. While girls from bottom performance distribution seem to be the most responsive to
incentives, the most responsive boys are from the middle of the performance distribution. The
results from quantile regressions can be found in Appendices F5 and F6. There are no gender
18
For a review of gender differences in risk preferences, other-regarding preferences and competitive preferences, see Croson and Gneezy (2009)
31
differences of different incentive schemes on students’ psychological well-being – similar result can be found in Azmat and Iriberri (2016).
Figure 7: Distribution of the average treatment effects of different incentives on overall performance of girls (Math and English pooled), by deciles
FIGURE 8: Distribution of the average treatment effects of different incentives on overall performance of boys (Math and English pooled), by deciles
32
Table 7: OLS estimates of the average treatment effects of different motivation schemes on students’ performance and their subjective well-being – by gender (Pure within-class Pure within-class Within-class feedback Within-class feedback feedback and interactions) feedback rewarded financially rewarded reputationally Mathematics (st.dev) English (st.dev.) Stress Happiness Confidence (Math) Confidence (English) Aspirations Education over work Education over rest Work over rest
(Pure across-class feedback and interactions) Mathematics English Stress Happiness Confidence (Math) Confidence (English) Aspirations Education over work Education over rest Work over rest
Girls
Boys
Girls
Boys
Girls
Boys
0.121§ (0.081) -0.141** (0.059) 0.072 (0.124) 0.023 (0.087) -7.385*** (0.929) -5.023*** (0.994)
0.076 (0.107) -0.116§ (0.072) 0.043 (0.119) 0.213* (0.123) -4.13*** (0.954) -2.79*** (0.909)
0.229* (0.118) 0.016 (0.092) 0.258** (0.116) 0.304*** (0.101) -6.104*** (1.214) -5.528*** (1.375)
0.228* (0.137) 0.199* (0.116) 0.143 (0.189) 0.282** (0.112) -4.07*** (1.249) -4.604*** (1.363)
0.201** (0.102) 0.069 (0.088) 0.178 (0.159) 0.073 (0.115) -5.324*** (1.144) -5.722*** (1.115)
0.204§ (0.129) 0.092 (0.094) 0.313* (0.179) 0.297*** (0.111) -6.604*** (1.069) -5.129*** (1.193)
-0.035 (0.079) 0.017 (0.047) 0.038 (0.069)
0.098 (0.082) 0.219*** (0.068) -0.009 (0.113)
0.163** (0.081) 0.109** (0.044) -0.043 (0.091)
0.146* (0.086) 0.098 (0.074) -0.267** (0.110)
0.052 (0.094) 0.061 (0.061) -0.027 (0.093)
0.042 (0.101) 0.046 (0.099) -0.057 (0.117)
Girls
Boys
Girls
Boys
Girls
Boys
0.135* (0.077) -0.076 (0.066) -0.099 (0.119) 0.020 (0.089) -8.148*** (0.841) -6.013*** (0.980)
0.009 (0.088) -0.019 (0.072) 0.016 (0.119) 0.124 (0.098) -4.74*** (1.083) -4.49*** (1.058)
0.275* (0.159) 0.108 (0.101) -0.016 (0.146) -0.022 (0.109) -6.948*** (1.170) -6.528*** (1.154)
0.284§ (0.173) 0.249** (0.112) -0.022 (0.155) 0.193§ (0.130) -4.597*** (1.538) -4..047*** (1.363)
0.189** (0.091) 0.041 (0.083) 0.229* (0.121) 0.153 (0.143) -6.957*** (1.406) -6.411*** (1.580)
0.175* (0.103) 0.042 (0.103) 0.286§ (0.174) 0.241* (0.122) -6.125*** (1.675) -5.327*** (1.579)
0.101 (0.072) 0.023 (0.044) 0.038 (0.069)
0.174* (0.089) 0.140** (0.067) -0.103 (0.100)
0.099 (0.093) -0.049 (0.069) -0.043 (0.091)
0.219** (0.105) -0.091 (0.096) -0.011 (0.120)
0.101 (0.091) -0.006 (0.066) -0.027 (0.093)
-0.026 (0.136) 0.109 (0.087) -0.069 (0.112)
Pure across-class feedback
33
Across-class feedback rewarded financially
Across-class feedback rewarded reputationally
Table 8: OLS estimates of the average treatment effects of different motivation schemes on students’ performance and their subjective well-being – by gender (Pure financial rewards Pure financial Within-class feedback Across-class feedback and interactions) rewards rewarded financially rewarded financially Mathematics (st.dev) English (st.dev.) Stress Happiness Confidence (Math) Confidence (English) Aspirations Education over work Education over rest Work over rest
(Pure reputational rewards and interactions) Mathematics English Stress Happiness Confidence (Math) Confidence (English) Aspirations Education over work Education over rest Work over rest
Girls
Boys
Girls
Boys
Girls
Boys
0.018 (0.102) -0.038 (0.097) 0.431** (0.198) 0.015 (0.014) 1.869* (1.074) 2.239** (1.108)
0.207* (0.123) 0.139 (0.112) 0.482*** (0.162) 0.322** (0.132) -1.322 (1.429) -0.387 (1.099)
0.229* (0.118) 0.016 (0.092) 0.258** (0.116) 0.304*** (0.101) -6.104*** (1.214) -5.528*** (1.375)
0.228* (0.137) 0.199* (0.116) 0.143 (0.189) 0.282** (0.112) -4.07*** (1.249) -4.604*** (1.363)
0.275* (0.159) 0.108 (0.101) -0.016 (0.146) -0.022 (0.109) -6.948*** (1.170) -6.528*** (1.154)
0.284§ (0.173) 0.249** (0.112) -0.022 (0.155) 0.193§ (0.130) -4.597*** (1.538) -4..047*** (1.363)
0.163** (0.081) 0.109** (0.044) -0.043 (0.091)
0.146* (0.086) 0.098 (0.074) -0.267** (0.110)
0.099 (0.093) -0.049 (0.069) -0.043 (0.091)
0.219** (0.105) -0.091 (0.096) -0.011 (0.120)
0.046 (0.098) 0.009 (0.078) -0.017 (0.092)
0.006 (0.111) 0.016 (0.083) 0.137 (0.112)
Pure reputation rewards
Girls
Boys
0.059 (0.147) -0.039 (0.087) -0.008 (0.203) -0.005 (0.116) 1.905* (0.972) 0.989 (1.096)
0.218 (0.154) 0.079 (0.106) 0.158 (0.195) 0.144 (0.103) -0.399 (1.224) -1.301 (1.008)
0.021 (0.096) -0.017 (0.061) 0.164* (0.088)
Within-class feedback rewarded reputationally Girls
Boys
Across-class feedback rewarded reputationally Girls
Boys
0.201** (0.102) 0.069 (0.088) 0.178 (0.158) 0.073 (0.115) -5.324*** (1.144) -5.722*** (1.115)
0.204§ (0.129) 0.092 (0.094) 0.313* (0.179) 0.297*** (0.111) -6.604*** (1.069) -5.129*** (1.193)
0.189** (0.091) 0.041 (0.083) 0.229* (0.121) 0.153 (0.143) -6.957*** (1.406) -6.411*** (1.580)
0.175* (0.103) 0.042 (0.103) 0.286§ (0.174) 0.241* (0.122) -6.125*** (1.675) -5.327*** (1.579)
0.052 (0.094) 0.061 (0.061) -0.027 (0.093)
0.042 (0.101) 0.046 (0.099) -0.057 (0.117)
0.101 (0.091) -0.006 (0.066) -0.027 (0.093)
-0.026 (0.136) 0.109 (0.087) -0.069 (0.112)
0.165§ (0.101) 0.109 (0.078) -0.004 (0.131)
34
8. Robustness Checks 8.1. Multiple comparisons The probability that the coefficients are significant purely by chance increases with the number
of hypothesis tested. Multiple-test procedures take p-values from multiple comparisons testing and
uncorrected critical p-values interpreted either as FWER or FDR and result in adjusted critical pvalues. “If the input uncorrected critical p-value α ∈ (0,1) is an FWER, then we can be 100(1 − α)%
confident that all the null hypotheses in the discovery set are false. If the input uncorrected critical p-value α = β ∗ γ is an FDR, then we can be 100(1 − β)% confident that over 100(1 − γ)% of the
null hypotheses in the discovery set are false” (Newson, 2010, p.569).
In order to address these concerns about multiple inference I control for the familywise error
rate (FWER) using one-step methods (Bonferroni, in Dunn 1961; and Sidak, 1967 corrections), and
step-down methods - Holm (1979) and Holland-Copenhaver (1987) corrections) and for the false
discovery rate (FDR, using step-up methods - Simes (1986), Hochberge (1988), and Yakutieli-
Benjamini (2001) procedure). The detailed description of the procedures can be found in Newson
(2010). The corrected p-values are summarized in Table 9.
Table 9: Adjusted p-values for aggregated treatment effects and disaggregated treatment effects
Familywis e error rates
One-step method Step-down method
False Discov ery rate
Type of correction (corresponding to uncorrected alpha = 0.1)
Step-up method
Bonferroni correction Sidak correction Holm correction Holland correction Hochberg correction Simes correction Yekutieli correction
Correlation assumed
Arbitrary Nonnegative Arbitrary Nonnegative Independence Nonnegative Arbitrary
35
Aggregated treatment effect 0.0077 0.0081 0.0200 0.0210 0.0170 0.0620 0.0190
Disaggregated treatment effects 0.0053 0.0055 0.0083 0.0087 0.0077 0.0680 0.0150
Disadvantage of FWER procedure is that it can result in low power for testing single hypotheses
in large experiments with high number of multiple comparisons. In such cases FDR procedure is preferred since it controls for the proportion of Type I errors to true positives and therefore results
in higher power. In the case of this experiment, one-step and step-down methods rules out any of
the initially presented average treatment effects of interventions on students’ performance. FWER procedures seem to be too conservative due to the number of multiple comparisons I test. Among
FDR procedures I rule out the Hochberg corrections because I do not meet its restriction of
independence (I compare all groups with treatments to the same control group). The significance of
the average treatment effects of different incentive schemes on students’ performance and has been
confirmed when I used Simes correction and with some exception when I used Yekutieli correction (the effect of pure financial rewards and financial rewards interacted with within-class feedback
turned to be insignificant). Similar conclusion can be done regarding the average treatment effects on subjective well-being with one exception – the negative impact of pure financial rewards is
significant using all types of FWER and FDR corrections. Summary of corrected p-values for all disaggregated treatment effects can be found in the Appendices F7, F8 and F9. 8.2. Attrition
High drop-out and absence rates are common features of students in developing countries and
it is not an exception in my data. There are several reasons. Some students did not have the money
to pay the school fees and decided to change schools to avoid repaying their debt, others changed
their school because of family reasons (the family moved to a different area, they were sent to live with other family members, etc.), some completely dropped out of school, some just registered as
new students and some of the students passed away. Due to the constraints of the experiment, all
participation data are based on our visits only (it means that no random visits were organized).
36
The main concern in most project evaluations is whether the attrition of subjects is random or
whether there is a systematic difference between the attrition from the treatment group compared
to the control group caused by the intervention itself. Only uninformed students, who did not
receive feedback during the academic year and who were chosen to participate in a tournament rewarded with reputation rewards did not significantly change their attrition. All other treatment
groups lowered their absences compared to the control group ranging from 6.5 to 17 per cent.
Lower attrition means higher attendance.
Who are the attrited students? Random versus non-random attrition The treatments influenced the probability of always being present during our visits and the
probability to attrite. So in absolute numbers there are less students who drop out from treated classes compared to the control classes and more cases when students from the treatment group
attended all five testing rounds compared to students from the control group. Besides the differences in the number of attrited students, students who dropped from the within-class feedback group are worse in terms of their initial performance compared to students from the
across-class feedback group or the control group. That might re-introduce a bias if the treated
students who are present during the final testing round are systematically different compared to
the control-group students. As shown in Table 10, this is not the case in this project. The distribution of students who stayed in either of the treatment groups (based on their initial
performance) is not statistically different from the distribution of the initial abilities of students
from the treatment group. In such a case the OLS estimate should provide unbiased estimates of the
treatment effects. Nevertheless, I used inverse probability weights and imputation methods to check the stability of the results (for further details see the next section).
37
Table 10: Ksmirnov test on equality of distributions of students who attrited and students who stayed (p-values presented) Baseline differences Students who Students who Alwayspresent attrited stayed students (T1 – C) (T2 – C) (T1 – C) (T2 – C) (T1 – C) (T2 – C) (T1 – C) (T2 – C) Mathematics
0.123
0.274
0.000
0.158
0.752
0.192
0.677
0.958
English 0.952 0.168 0.003 0.546 0.230 0.282 0.211 0.840 Note: T1 stands for within-class social comparison group, T2 for across-class comparison group and C represents control group with no feedback provided. P-values are presented.
The effect of treatments on attrition Estimates of treatment effects can be biased if the attrition from control versus treatment
groups systematically differs and the difference is caused by the presence of the treatment.
Students in treatment groups attrite less often in absolute values and are more often present in all five testing rounds compared to their control-group counterparts. In order to see whether and to
what extent social comparison and reward treatments influence the probability of dropping out. I
run a probit model on attrition and full attendance on all treatment dummies controlling for strata variables (Table 11).
The attrition rate comprises of students who missed our last testing round but attended the
baseline testing at the beginning of the project. Non-rewarded students exposed to both within and
across-class social comparison feedback have from 6.5 to 6.9 per cent lower probability to miss the final testing round. Among rewarded students who did not receive any feedback only students
rewarded financially lowered their attrition by 7.9 per cent. Reputation rewards without provided feedback do not affect attrition rate. All treatment interactions lower the attrition rate (from 9.3 to 17.2 per cent).
As previously discussed, despite the different attrition across treatment and control groups,
students who remained at schools in the last testing round are on average the same in terms of 38
initial characteristics and therefore the OLS estimates should not be biased. In the following section I run alternative specifications to compare OLS estimates with estimates that correct for possible attrition bias.
Table 11: Probabilities of students’ dropouts, by gender
Overall treatment effects on:
Attrition
Overall
Girls
Boys
Girls Overall
Boys Girls
Boys
Within-class feedback, no -0.066* -0.064* -0.058 0.105** 0.091* 0.108** (0.039) (0.037) (0.049) (0.049) (0.055) rewards (T1_solo) (0.049) -0.071** -0.046 0.124*** 0.110** 0.137*** -0.097** Across-class feedback, no (0.032) (0.035) (0.042) (0.049) (0.046) (0.038) rewards (T2_solo) -0.130*** -0.128*** 0.127** 0.168** 0.093* -0.124** Financial Rewards, no (0.038) (0.033) (0.056) (0.067) (0.057) (0.055) feedback (Fin_solo) -0.056 -0.021 0.030 0.047 0.021 -0.100** Reputational Rewards, no (0.046) (0.052) (0.077) (0.087) (0.074) (0.050) feedback (Rep_solo) -0.158*** -0.127*** -0.196*** 0.233*** 0.228*** 0.247*** Within-class feedback with (0.033) (0.033) (0.062) (0.071) (0.064) (0.037) financial rewards (T1_fin) -0.147*** -0.128*** -0.157*** 0.263*** 0.257*** 0.252*** Across-class feedback with (0.032) (0.031) (0.060) (0.068) (0.063) (0.041) financial rewards (T2_fin) -0.157*** -0.146*** -0.171*** 0.208*** 0.209*** 0.217*** Within-class feedback with (0.038) (0.036) (0.067) (0.073) (0.064) (0.043) reputation rewards (T1_rep) -0.212*** -0.192*** -0.226*** 0.099* 0.079 0.126** Across-class feedback with (0.026) (0.026) (0.051) (0.057) (0.056) (0.031) reputation rewards (T2_rep) Controlled for stratas Yes Yes Yes Yes Yes Yes N 7050 3818 3139 7050 4672 3884 Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
8.3. Stability of the results In order to adjust the results for non-random attrition, I proceeded with imputation methods
and inverse probability-weighted regressions (Imbens, 2004; Woolridge, 2007; Kwak (2010), Hirano et al., 2000, etc.). Inverse probability weighting (IPW) can adjust for confounding factors
and selection bias. As the title suggests, IPW assigns a weight to every student which equals to the student’s inverse probability to be absent/to attrite and adjust for that in the estimation of the 39
treatment effects. An imputation method is used to fill the missing observations of students who
were absent or dropped out in the last testing round based on a predefined rule.
Tables 12 and 13 and Appendices F2, F3 and F4 provide the comparison of ordinary least
squares estimations (column 1) of the treatment effects to the weighted least squares using inverse
probability weights (column 2), separately for Math and English. Correcting for the probability of
dropping out, treatment effects are similar or slightly higher in absolute terms but not significantly different. The results of the imputation methods (columns 3 and 4) bring similar conclusions. I use
two different measures to impute missing observations – median ration and the class percentile ranks (inspired by Krueger, 1999). All of the measures take the advantage of repeated school visits
and follow the same logic – if the observation from the last school visit is missing, I look at the last
score available and adjust for the differences in test difficulty. The same procedure is done to impute Math and English scores separately. The median ratio measure imputes the last available observation and the class percentile ranks take into consideration the rank of the student in the last
available distribution and impute the score corresponding to the student of the same rank in the
final visit distribution. The imputation method artificially fills missing observations and the results serve only as bounds. Both imputation measures deliver similar or stronger results compared to
ordinary least squares. Ordinary least squares results are also comparable to the weighted regression estimates.
40
Table 12: Average treatment effects of different motivation schemes - alternative specifications Dependent variable: Math score PURE TREATMENTS Within-class feedback, no rewards (T1_solo) Across-class feedback, no rewards (T2_solo) Financial Rewards, no feedback (Fin_solo) Repurational Rewards, no feedback (Rep_solo)
TREATMENT INTERACTIONS Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputation reward (T1_rep) Across-class feedback, reputation reward (T2_rep)
OLS
IPW
Imputation (median ratio)
Imputation (class percentiles)
0.100 (0.085) 0.082 (0.073) 0.106 (0.101) 0.138 (0.141)
0.046 (0.092) 0.067 (0.079) § 0.151 (0.102) 0.188 (0.149)
0.133* (0.079) 0.129* (0.068) 0.169* (0.096) 0.206* (0.124)
0.123 (0.085) 0.087 (0.078) 0.143 (0.106) 0.177 (0.128)
0.231* (0.118) 0.277** (0.139) 0.209** (0.103) 0.188** (0.080)
0.338** (0.135) 0.456*** (0.132) 0.212* (0.108) 0.208** (0.087)
0.281** (0.129) 0.331** (0.128) 0.266** (0.073) 0.186** (0.073)
0.273** (0.124) 0.305** (0.139) 0.258** (0.112) 0.250*** (0.090)
Controlled for stratas Yes Yes Yes Yes N of observation 5102 5102 Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
41
Table 13: Average treatment effects of different motivation schemes - alternative specifications Dependent variable: English score PURE TREATMENTS Within-class feedback, no rewards (T1_solo) Across-class feedback, no rewards (T2_solo) Financial Rewards, no feedback (Fin_solo) Repurational Rewards, no feedback (Rep_solo)
TREATMENT INTERACTIONS Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputation reward (T1_rep) Across-class feedback, reputation reward (T2_rep)
OLS
IPW
Imputation (median ratio)
Imputation (class percentiles)
-0.128** (0.056) -0.049 (0.059) 0.045 (0.088) 0.016 (0.082)
-0.133* (0.070) -0.079 (0.072) 0.032 (0.085) 0.004 (0.084)
-0.133** (0.060) -0.052 (0.063) -0.006 (0.096) -0.089 (0.123)
-0.135*** (0.045) -0.046 (0.048) 0.041 (0.069) 0.036 (0.059)
0.103 (0.094) 0.173 (0.094) 0.087 (0.080) 0.047 (0.080)
0.145* (0.086) 0.258** (0.102) 0.041 (0.078) 0.071 (0.077)
0.096 (0.108) 0.113 (0.099) 0.069 (0.082) 0.024 (0.082)
0.072 (0.080) 0.137* (0.075) 0.069 (0.058) 0.059 (0.064)
Controlled for stratas Yes Yes Yes Yes N of observations 5093 5093 6736 7107 Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
42
9. Conclusion Various interventions have been conducted with the aim of lowering absenteeism and
increasing student performance. Authors usually focus on the main outcomes of their interventions,
such as subjects’ performance, absence or drop-out rates, leaving outcomes other than learning
aside. Evidence from psychology indicates that current well-being, measured in terms of stress and
happiness, serves as an important prerequisite of future performance. For instance, stressed students are absent and drop out from school more often compared to non-stressed students;
stress makes students exert less effort and perform worse. This paper contributes to the current
literature by studying the effects of different types of incentives on student performance and their
well-being (measured by happiness and stress). I bring new evidence on performance-versus-well-
being tradeoff by implementing two types of social comparative feedback regimes (within- and
across-class group feedback), two types of incentive regimes - financial and reputation rewards, and their interactions.
The results of this study show that providing students with pure feedback without further
incentivization deliver subject-specific results. While in Mathematics students improved by 0.08 to
0.13 standard deviations, there was no improvement in Egnlish. The results are driven purely by improvements in girls’ performance. Pure rewards (without feedback) on contrary lead to an
improvement of 0.1 to 0.18 standard deviations in both subjects. The results are driven mainly by boys. Interacted incentive scheme of feedback combined with rewards leads to an improvement between 0.20 to 0.29 standard deviations if rewarded financially and 0.12 to 0.18 standard
deviations if rewarded reputationally. There is, however, a trade-off between improvement in
performance and students’ well-being in response to different incentive schemes. Feedback and reputational rewards improve students’ performance mildly, but does influence neither their
happiness nor stress. Competing for financial rewards result in moderate to strong improvements 43
in performance but students’ stress significantly increases and their happiness decreases. Students
competing for monetary rewards reported significantly lower stress levels compared to those who
competed for money without any feedback. Stressed students exerted less effort, performed worse
on average and attrited by 29 percent more compared to relaxed students.
Furthermore, this paper sheds light on gender differences in responsiveness to different
incentive provisions. According to the results, girls did not improve if they received no feedback but they competed for rewards of any type and they significantly underperformed boys. If the girls
were repeatedly given feedback (and the type of feedback does not matter), they performed
comparably to boys. Moreover, girls also respond positively to pure feedback provision (without rewards). Comparative feedback plays a crucial role for girls in inducing their performance in a
tournament environment. Boys react only with respect to the provision of rewards. Provision of
feedback does not play any role in their performance improvements. There are no genderdifferences in the effects of incentives on girls’ and boys’ well-being.
The results of this experiment might be of important help especially for policy makers trying to
find the optimal incentive scheme. Policy makers must exercise a great amount of caution in designing educational rewards and consider the impact on student well-being. Further research
should be conducted with the aim to study the long-term effects of changes in student well-being on
performance.
44
References Andrabi, T., Das J. and Ijaz-Khwaja, A. (2009): Report Cards: The Impact of Providing School and Child Test-scores on Educational Markets, BREAD Working Paper No. 226
Angrist, J., Bettinger, E., and Kremer, M. (2006): Long-term educational consequences of secondary school vouchers: Evidence from administrative records in Colombia, The American Economic Review, 847-862. Angrist, J., and Lavy, V. (2009): The effects of high stakes high school achievement awards: Evidence from a randomized trial. The American Economic Review, 1384-1414.
Apesteguia, J., Azmat, G., and Iriberri, N. (2012): The Impact of Gender Composition on Team Performance and Decision-Making: Evidence from the Field, Management Science, Vol. 58(1) January 2012, pp. 78–93.
Arnold, H. J. (1976): Effects of performance feedback and extrinsic reward upon high intrinsic motivation, Organizational Behavior and Human Performance, 17(2), 275-288.
Ashraf, N., Bandiera, O., and Lee, S. S. (2014): Awards unbundled: Evidence from a natural field experiment, Journal of Economic Behavior and Organization,100, 44-63.
Auriol, E., and Renault, R. (2008): Status and incentives, The RAND Journal of Economics, 39(1), 305-326.
Azmat, G., Bagues, M., Cabrales, A., and Iriberri, N. (2015): What you know can’t hurt you (for long): A field experiment on relative performance feedback, Working paper, Aalto University.
Azmat, G. and Iriberri, N. (2010): The importance of relative performance feedback information: Evidence from a natural experiment using high school students. Journal of Public Economics, 94(7), 435-452
Azmat, G. and Iriberri, N. (2016): The Provision of Relative Performance Feedback: An Analysis of Performance and Satisfaction, Journal of Economics and Management Strategy, Vol. 25(1), pp.77110. Bandiera, O., Barankay, I., and Rasul, I. (2010): Social incentives in the workplace, The Review of Economic Studies, 77(2), 417-458.
Bandiera, O., Larcinese, V., and Rasul, I. (2015): Blissful ignorance? A natural experiment on the effect of feedback on students' performance, Labour Economics, 34, 13-25
Barankay, I. (2011). Rankings and social tournaments: Evidence from a crowd-sourcing experiment. In Wharton School of Business, University of Pennsylvania Working Paper.
Benabou, R., and Tirole, J. (2003): Intrinsic and extrinsic motivation, The Review of Economic Studies, 70(3), 489-520. 45
Benjamini, Y., and Y. Hochberg (1995): Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society, Series B (Methodological) 57: 289–300.
Benjamini, Y., and D. Yekutieli (2001): The control of the false discovery rate in multiple testing under dependency, Annals of Statistics 29: 1165–1188. Besley, T., and Ghatak, M. (2008): Status incentives, The American Economic Review, 206-211.
Bettinger, E. P. (2012): Paying to learn: The effect of financial incentives on elementary school test scores. Review of Economics and Statistics, 94(3), 686-698.
Blanes i Vidal, J., and Nossol, M. (2011): Tournaments without prizes: Evidence from personnel records, Management Science, 57(10), 1721-1736.
Blimpo, M. P. (2014): Team incentives for education in developing countries: A randomized field experiment in Benin, American Economic Journal: Applied Economics, 6(4), 90-109.
Dunn, O. J. (1961): Multiple Comparisons Among Means, Journal of the American Statistical Association 56 (293): 52–64
Buunk, B. P., Gibbons, F. X., and Reis-Bergan, M. (1997): Social comparison in health and illness: A historical overview. Health, coping and well-being: Perspectives from social comparison theory, 1-23. Buunk, B. P., and Gibbons, F. X. (2000): Toward an enlightenment in social comparison theory, In Handbook of Social Comparison (pp. 487-499), Springer US.
Burgers, C., Eden, A., van Engelenburg, M. D., and Buningh, S. (2015): How feedback boosts motivation and play in a brain-training game, Computers in Human Behavior, 48, 94-103.
Charness, G., Masclet, D., and Villeval, M. C. (2010): Competitive preferences and status as an incentive: Experimental evidence, Groupe d’Analyse et de Théorie Économique Working Paper, (1016). Cohen, S., Kamarck, T., and Mermelstein, R. (1983): A global measure of perceived stress, Journal of health and social behavior, 385-396. Croson, R., and Gneezy, U. (2009): Gender differences in preferences, Journal of Economic literature, 448-474.
Deci, E. L. (1971): Effects of externally mediated rewards on intrinsic motivation, Journal of personality and Social Psychology, 18(1), 105.
Deci, E. L., Koestner, R., and Ryan, R. M. (1999): A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation, Psychological bulletin, 125(6), 627. 46
Dijkstra, P., Kuyper, H., van der Werf, G., Buunk, A.P. and van der Zee, Y. (2008): Social comparison in the classroom: a review, Review of educational research, Vol. 78, No. 4, p.828-879
Dolan, P., Metcalfe, R., and Powdthavee, N. (2008): Electing happiness: Does happiness affect voting and do elections affect happiness, Discussion Papers in Economics, (30).
Duffy, J., and Kornienko, T. (2010): Does competition affect giving?, Journal of Economic Behavior and Organization, 74(1), 82-103.
Dynarski, S. (2008): Building the stock of college-educated labor, Journal of human resources, 43(3), 576-610. Eisenkopf, G. (2011): Paying for better test scores, Education Economics,19(4), 329-339.
Ellingsen, T., and Johannesson, M. (2007): Paying respect, The Journal of Economic Perspectives, 135-150.
Eriksson, T., Poulsen, A., and Villeval, M. C. (2009): Feedback and incentives: Experimental evidence, Labour Economics, 16(6), 679-688. Falk, A. and Ichino, A. (2006): Clean Evidence on Peer Pressure, Journal of Labor Economics, Vol. 24, Issue 1 Festinger, L. (1954): A theory of social comparison processes, Human relations, 7(2), 117-140.
Fordyce, M. W. (1988): A review of research on the happiness measures: A sixty second index of happiness and mental health, Social Indicators Research, 20(4), 355-381.
Fryer Jr, R. G. (2010): Financial incentives and student achievement: Evidence from randomized trials (No. w15898), National Bureau of Economic Research.
Hannan, R. L., Krishnan, R., and Newman, A. H. (2008): The effects of disseminating relative performance feedback in tournament and individual performance compensation plans, The Accounting Review, 83(4), 893-913. 57.
Helliwell, J. F., and Wang, S. (2012): The state of world happiness, World happiness report, 10-
Hastings, J. S., Neilson, C. A., and Zimmerman, S. D. (2012): The effect of school choice on intrinsic motivation and academic outcomes (No. w18324), National Bureau of Economic Research. Hattie, J., and Timperley, H. (2007): The power of feedback. Review of educational research, 77(1), 81-112. Hirano, K., Imbens, G. and Ridder, G. (2003): Efficient Estimation of Average Treatment Effects Using the Estimated Propensity Score, Econometrica, Vol. 71(4), 1161–1189
Hochberg, Y. (1988): A sharper Bonferroni procedure for multiple tests of significance. Biometrika 75: 800–802. 47
Holland, B. S., and Copenhaver. M.D. (1987): An improved sequentially rejective Bonferroni test procedure. Biometrics 43: 417–423
Holm, S. (1979): A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics 6: 65–70.
Hoxby, C. (2000): Peer effects in the classroom: Learning from gender and race variation (No. w7867), National Bureau of Economic Research.
Imbens, G.W. (2004): Nonparametric estimation of average treatment effects under exogeneity: A review, The Review of Economics and Statistics: Vol 86, No.1, pp 4-29.
Jalava, N., Joensen, J.S. and Pellas, E. (2015): Grades and rank: Impacts of non-financial incentiveson test performance, Journal of Economic Behavior and Organization 115 (2015) 161– 196 Juster, R. P., McEwen, B. S., and Lupien, S. J. (2010): Allostatic load biomarkers of chronic stress and impact on health and cognition, Neuroscience and Biobehavioral Reviews, 35(1), 2-16.
Kling, J. R., Liebman, J. B., and Katz, L. F. (2007): Experimental analysis of neighborhood effects, Econometrica, 75(1), 83-119. Kremer, M., Miguel, E. and Thornton, R. (2002): Incentives to Learn, NBER Working Papers 10971, National Bureau of Economic Research, Inc.
Krueger, A. B. (1999): Experimental estimates of education production functions, The Quarterly Journal of Economics, 114 (2): 497-532
Kosfeld, M. and Neckermann, S. (2011): Getting More Work for Nothing? Symbolic Awards and Worker Performance, American Economic Journal: Microeconomics, Vol. 3, Issue 3
Kremer, M., Miguel, E., and Thornton, R. (2004): Incentives to learn (No. w10971), National Bureau of Economic Research. Kuhnen, C. M., and Tymula, A. (2012): Feedback, self-esteem, and performance in organizations, Management Science, 58(1), 94-113.
Kwak, D. (2010): Inverse probability weighted estimation for the effect of kindergarten enrollment age and peer quality on student academic achievement for grades K-12, working paper
LaLonde, R. J. (1995): The promise of public sector-sponsored training programs, The Journal of Economic Perspectives, 149-168. Lavy, V. (2009): Performance Pay and Teachers’ Effort, Productivity and Grading Ethics, American Economic Review 99, 5
Lazear, E.P. (2000): Performance pay and productivity. American Economic Review, Vol. 90, No. 5, pp.11341-61. 48
Levitt, S. D., List, J. A., Neckermann, S., and Sadoff, S. (2012): The behavioralist goes to school: Leveraging behavioral economics to improve educational performance (No. w18165), National Bureau of Economic Research. Locke, E. A., and Latham, G. P. (1990): A theory of goal setting and task performance, PrenticeHall, Inc.
Lupien, S. J., McEwen, B. S., Gunnar, M. R., and Heim, C. (2009): Effects of stress throughout the lifespan on the brain, behaviour and cognition, Nature Reviews Neuroscience, 10(6), 434-445.
Lyubomirsky, S., and Lepper, H. (1999): A measure of subjective happiness: Preliminary reliability and construct validation, Social Indicators Research, 46, 137-155. The original publication is available at www.springerlink.com. McEwen, B. S. (2008): Central effects of stress hormones in health and disease: Understanding the protective and damaging effects of stress and stress mediators, European journal of pharmacology, 583(2), 174-185.
MacKerron, G. (2012): Happiness economics from 35 000 feet. Journal of Economic Surveys, 26(4), 705-735.
Markham, S. E., Scott, K., and McKEE, G. A. I. L. (2002): Recognizing good attendance: a longitudinal, quasi‐experimental field study. Personnel Psychology, 55(3), 639-660. Mas, A. and Moretti, E. (2009): Peers at Work, American Economic Review, Vol. 99, Issue 1 31
Mettee, D. R., and Smith, G. (1977): Social comparison and interpersonal attraction: The case for dissimilarity. Social comparison processes: Theoretical and empirical perspectives, 69, 101. Moldovanu, B., Sela, A., and Shi, X. (2007): Contests for status, Journal of Political Economy, 115(2), 338-363.
Newson, R.B. (2010): Frequentist q-value for multiple-test procedures, Stata Journal, Vol. 10, No. 4. Ray, D. (2002): Aspirations, poverty and economic change, Understanding Poverty, 2006, p. 409-443(35)
Reardon, S. F., Cheadle, J. E., and Robinson, J. P. (2009): The effect of Catholic schooling on math and reading development in kindergarten through fifth grade, Journal of Research on Educational Effectiveness, 2(1), 45-87.
Ryan, R. M., and Deci, E. L. (2000): Intrinsic and extrinsic motivations: Classic definitions and new directions, Contemporary educational psychology, 25(1), 54-67.
Sacerdote, B. (2011): Peer effects in education: How might they work, how big are they and how much do we know thus far?, Handbook of the Economics of Education, 3, 249-277. 49
Šidák, Z. (1967): Rectangular confidence regions for the means of multivariate normal distributions, Journal of the American Statistical Association 62: 626–633.
Simes, R. J. (1986): An improved Bonferroni procedure for multiple tests of significance. Biometrika 73: 751–754. Slavin, R. (1984). Meta-analysis in education: How has it been used? Educational Researcher. 13(8), 6-15, 24-27
Schneiderman, N., Ironson, G., and Siegel, S. D. (2005): Stress and health: psychological, behavioral, and biological determinants, Annual Review of Clinical Psychology, 1, 607.
Suls, J., and Wheeler, L. (2000): A selective history of classic and neo-social comparison theory, In Handbook of social comparison (pp. 3-19). Springer US. Bigoni, M., Fort, M., Nardotto, M., and Reggiani, T. (2011): Teams or tournaments? A field experiment on cooperation and competition among university students. Tran, A., and Zeckhauser, R. (2012): Rank as an inherent incentive: Evidence from a field experiment, Journal of Public Economics, 96(9), 645-650.
Van Dijk, F., Sonnemans, J., and Van Winden, F. (2001): Incentive systems in a real effort experiment, European Economic Review, 45(2), 187-214. Veenhoven, R. (1988): The utility of happiness, Social indicators research,20(4), 333-354.
Weiss, Y., and Fershtman, C. (1998): Social status and economic performance:: A survey,European Economic Review, 42(3), 801-820. Wolf, T. M. (1994): Stress, coping and health: enhancing well‐being during medical school, Medical Education, 28(1), 8-17. Wooldridge, J. (2007): Inverse Probability Weighted M-Estimation for General Missing Data Problems, Journal of Econometrics 141:1281-1301
50
Appendix APPENDIX A: Summary statistics and randomization balance Appendix A1: Balance between control and treatment groups
Variable
School Level: The number of primary schools The number of secondary schools School Type: Public Schools Private Schools Community Schools By Population By PLE/UCE results By testing results
Control
Within-class feedback
Across-class feedback
10 7
11 7
10 8
8 7 2 2345 (48 groups) 3.175 21.140
5 9 4 2415 (51 groups) 3.039 21.363
Note: min(PLE/UCE)= 1.7397, max(PLE/UCE)= 4.2857, mean(PLE/UCE)=3.1040 Note: min(TR)=8.3125, max(TR)=39.7765, mean(TR)=21.3192, where TR=Testing Results
51
6 8 4 2371 (51 groups) 3.102 21.648
Appendix A2: Comparison of mean characteristics of students in control and treatment groups Withinclass feedback (T1) Mathematics English
Sum Mathematics + English
Means Acrossclass feedback (T2)
Mean Differences
Control (C)
(T1 – C)
(T2 – C)
A. STUDENTS PERFORMANCE – ROUND 1 – BASELINE SURVEY 11.015 11.198 11.092 -0.077 0.106 (0.99) (0.96) 11.551 11.927 11.477 0.074 0.450 (1.53) (1.72) 22.566 23.125 22.569 -0.003 0.556 (2.30) (2.43) B.
B.1 After Math questionnaire Q1: Expected number of points [min 1, max 10] Q2: Subjective effort level [min 1, max 5] Q3: Perceived difficulty [min 1, max 5] Q4: Subjective level of happiness [min 1, max 7] B.2 After English questionnaire Q1: Expected number of points [min 1, max 10] Q2: Subjective effort level [min 1, max 5] Q3: Perceived difficulty [min 1, max 5] Q4: Subjective level of happiness [min 1, max 7] B.3 Aspiration questionnaire Aspirations Education over Relax [min 1, max 5] Education over Work [min 1, max 5] Work over Relax [min 1, max 5] Perceived happiness scale [min 4, max 28] Perceived stress [min 0, max 16]
QUESTIONNAIRES
4.331
4.537
4.551
3.341
3.494
3.423
3.447 3.319
3.525
3.504
3.253
3.184
5.715
5.757
5.796
3.644
3.644
3.677
3.547 2.950
3.627
3.553
2.904
2.856
3.833
3.756
3.778
2.766
2.701
2.803
3.538 11.479 6.018
3.496
3.477
11.653
11.223
6.352
5.756
52
-0.221 (0.150) -0.057 (0.053) -0.082 (0.053) 0.135 (0.092) -0.081 (0.161) -0.006 (0.046) -0.033 (0.052) 0.094 (0.084) 0.056 (0.049) 0.060 (0.057) -0.037 (0.094) 0.256 (0.231) 0.262 (0.164)
-0.151 (0.145) 0.021 (0.052) 0.072 (0.052) 0.069 (0.094)
Joint Pvalue
0.183 0.699 0.423
0.299
0.298 0.030 0.343
-0.039 (0.144) 0.074* (0.044) -0.033 (0.049) 0.048 (0.086)
0.879
-0.021 (0.049) 0.019 (0.059) -0.102 (0.090) 0.429** (0.222) 0.595*** (0.142)
0.269
0.141 0.752 0.534
0.526 0.524 0.155
0.000
Appendix A2: Comparison of mean characteristics of students in control and treatment groups (Continued)
Withinclass feedback (T1)
Means Acrossclass feedback (T2)
Mean Differences
Control (C)
(T1 – C)
(T2 – C)
Joint Pvalue
C. OTHER (continued) C.1 Attrition rates All schools
0.359
0.346
0.454
-0.095*** (0.034) -0.059* (0.030)
-0.108*** (0.033) -0.069** (0.029)
C.2 Alwayscomers All schools
0.202
0.186
0.082
17.058
17.048
16.999
0.121*** (0.033) 0.097*** (0.033) 0.059 (0.079)
0.104*** (0.104) 0.077** (0.031) 0.049 (0.078)
Restricted sample#
Restricted sample# C.3 Age
C.6 Gender All schools
Restricted sample#
C.4 Class size All schools
0.358
0.207
0.534 0.548
0.348
0.417
0.188
0.110
0.512
0.508
0.524
0.533
0.025* (0.015) 0.015 (0.015)
0.004 (0.015) -0.009 (0.015)
0.002
0.041 0.000 0.008 0.737 0.192 0.277
-7.741* -3.581 0.146 (4.045) (4.672) Restricted sample# 52.15 56.56 55.14 -2.985 1.428 0.489 (3.988) (4.651) Attrition rate is defined as the rate of students missing in the last testing round conditional on student participation in the baseline testing. T1 stands for within-class comparison, T2 for across-class comparison and C for control group. There was one school which experienced strong transformation (exogenous to the intervention), which resulted in change of their headmaster and a large dropout of students. Restricted sample (#) excludes that school from the analysis. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1% 52.26
56.42
60.00
53
Appendix A3: Comparison of mean characteristics of students in control and treatment groups – by treatment decomposition
SOLO TREATMENT EFFECTS Within-class competition with no rewards (T1_solo) Across-class competition with no rewards (T2_solo) Financial rewards with no feedback (Fin_solo) Reputation rewards with no feedback (Rep_solo) INTERACTION EFFECTS Within-class competition with Financial rewards (T1_FIN) Across-class competition with Financial rewards (T2_FIN) Within-class competition with reputation rewards (T1_REP) Across-class competition with reputation rewards (T2_REP) Pure control
Mathematics
Sum Mean 19.551
21.575 20.528 24.288 24.111 23.326 22.734 25.454 22.697
Difference from pure control -3.136* (1.792) -1.284 (1.814) -1.891 (2.127) 2.071 (1.898)
Mean 7.126
8.068 7.719 9.366
1.728 (2.339) 1.215 (1.700) 0.453 (1.685) 3.304** (1.491)
8.485
0
8.583
9.002 8.834 9.974
English
Difference from pure control -1.439* (0.771) -0.559 (0.706) -0.751 (1.035) 0.976 (0.815)
Mean 12.425
13.507 12.809 14.922
0.029 (0.934) 0.651 (0.719) 0.418 (0.719) 1.606** (0.767)
15.625
0
14.115
14.324 13.899 15.479
Difference from pure control -1.698§ (1.096) -0.725 (1.178) -1.139 (1.163) 1.095 (1.191)
1.698 (1.458) 0.563 (1.117) 0.035 (1.052) 1.698** (0.808) 0
Joint p-value 0.028 0.069 0.039 Rows represent treatment groups (either pure treatments or treatment interactions). Columns (1), (3) and (5) represent average scores from Math, English and their sum. Columns (2), (4), (6) represent differences between particular treatment and pure control group (group without any feedback and any reward). Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%.
54
APPENDIX B: Randomization and logistics in the field
In order to increase the balance between control and treatment groups, the sample was stratified
along three dimensions – school location (the sample was divided into four areas differing in the
level of remoteness), average school performance in national examination (above average or below
average) and student level (grade 6 and 7 of primary education and grades 1 up to 4 of secondary education) 19, 20. Within each strata, I randomized the sample into treatment and control groups. The
randomization was done in two stages (as shown in Figure 1). First, after the stratification of the
sample by school performance and area, I randomized the whole sample of 53 schools into
treatment and control group in a ratio 2:1. The randomization was done at the school level and resulted in 36 treatment schools and 17 control schools. School-level randomization in the first
stage was chosen in order to minimize control group contamination due to information spillovers. In the second stage, I divided classes of the treatment schools randomly into within-class feedback (T1) and across-class feedback group (T2) in a ratio 1:1 (class-level randomization). In this scenario, no student in a control-group school received any of the treatments and students in the
treatment-group schools might have received either within- or across-class feedback depending on the type of intervention their class was randomized into. Overall, 1/3 of the sample is the control group, 1/3 is treatment group 1 and 1/3 is treatment group 2. Exposure to the treatment is the only difference in the outcomes between the control and treatment groups. 19
Every year students of P7 in primary schools and S4 in secondary schools take the national leaving examinations that are compulsory in order to complete their study and to proceed to a higher level. Using the data on PLE and UCE, I was able to divide schools into better and worse performing schools. 20 Uganda introduced Universal Primary Education (UPE) for all in 1997, allowing up to four students to go to school for free. Later it was extended to all children. Primary education is a seven-year program and for successful completion students need to pass the national Primary Leaving Exam (PLE) at the end of grade 7. Without passing PLE they cannot be admitted to a secondary school. Secondary school consists of two levels - “O-level”, which is four year program from S1 up to S4 completed by passing Ugandan Certificate of Education (UCE); and “A-level”, which is a two year extension to the Olevel and is completed by passing the Ugandan Advanced Certificate of Education (UACE). In 2007 Uganda introduced Universal Secondary Education (USE) as the first African country. The school year consists of 3 trimesters and lasts from January until December. Students are supposed to be examined by midterm and final. Students, however, do not necessarily have access to their evaluations and have limited information about their improvements.
55
All schools in the sample were connected to local non-governmental organization called
Uganda Czech Development Trust (UCDT). UCDT is a local affiliation of the non-governmental
organization Archdiocese Caritas Prague, Czech Republic, which has been running a sponsorship program “Adopce na dalku” in Uganda since 1993. According to UCDT representatives, students
were located into primary and secondary schools based on their own choice, therefore supported students should not differ from not supported students in terms of their school choice.
During the academic year students in the feedback groups received feedback. The feedback
was provided to students in the form of a report card, which was stuck into a small progress report book that each child in the treatment group received from us. My team members explained the
content of the report card repeatedly until students understood the message fully. The books were
stored at schools and headmasters promised to let children check their books at any point. The
books contained all necessary information to keep a child’s attention and motivation active. After
the experiment, students could keep their books. After students learned their feedback, they were
asked to start to work on questionnaires and to solve the Math and English exam 21. Students in the control group immediately started to answer the questionnaires. In order to ensure transparency, I used my own constructed tests. In order to announce the competition, I organized additional
meetings with students to explain the conditions in detail. Moreover, I left fliers in their classrooms
so that their absent classmates could also learn about the competition. Students were reminded of the conditions of the competition before they sat for the final exams. It took me and my four local
enumerators on average 3 to 4 weeks to evaluate the examinations. Knowing the winners, we visited schools again to disseminate the rewards.
21 The order was as follows: “Before Math questionnaire”, followed by Math examination that lasted 30 minutes; “After Math Before English questionnaire”, English exam in the subsequent 20 minutes and finally “After English questionnaire”. The core questions of the questionnaires were student expectations regarding how many points they thought they would obtain from the Math and English examinations, how much effort they planned to put/they put into answering the questions and the level of their current happiness. All of these questions we asked before as well as after each exam. No before-Math and before-English questionnaires were collected during the baseline survey since students saw the examinations for the first time.
56
Appendix B2a: Project’s timeline Reward scheme introduced 2012 Testing 3
Testing 4
Testing 5
Testing 1
Testing 2
Baseline testing from Math and English and questionnaires; No treatment
Withinclass feedback group (T1) received first treatment;
Within-class feedback group (T1) received treatment including improvement status
Within-class feedback group (T1) received treatment including improvement status
Within-class feedback group (T1) received treatment including improvement status
Acrossclass feedback group (T2) no treatment
Across-class feedback group (T2) received first treatment
Across-class feedback group (T2) received treatment including improvement status
Across-class feedback group (T2) received treatment including improvement status Chosen students competed to win prizes
2013 Follow-up Session No treatment provided, students examined from Math and English; BREAK
BREAK
2011 Baseline Survey Students, teachers and headmasters interviewed
Rewards disseminated
Note: T1 (treatment 1) stands for within-class social comparison treatment; T2 (treatment 2) represents acrossclass social comparison group; Qualification criteria differed based on initial randomization (T1,T2,C).
Appendix B2a: Orthogonal randomization of the sample into reward treatments
57
Appendix B2: Short version of the Perceived Stress Scale
Appendix B3: Short version of the Perceived Happiness Scale
58
Appendix B4: Score cards for students in within-class comparison group
Appendix B4: Score cards for students in across-class comparison group
59
Appendix C: A average treatment effects Appendix C1: OLS estimates of the effects of different incentive schemes on students’ performance and well-being (Pure within-class feedback Pure within-class Within-class feedback Within-class feedback and its interactions) feedback rewarded financially rewarded reputationally Math and English pooled (in st.dev.) Mathematics (in st.dev) English (in st.dev.) Stress Happiness Confidence (Math) Confidence (English) Education over work Education over rest Work over rest
(Pure across-class feedback and its interactions) Math and English pooled (in st.dev.) Mathematics (in st.dev) English (in st.dev.) Stress Happiness Confidence (Math) Confidence (English) Education over work Education over rest Work over rest
0.017 (0.014) 0.100 (0.085) -0.128** (0.056) 0.059 (0.113) -0.131 (0.095) -7.081*** (0.775) -5.559*** (0.809) 0.023 (0.054) 0.103*** (0.039) 0.026 (0.075)
Pure across-class feedback 0.044 (0.059) 0.082 (0.073) -0.049 (0.059) -0.059 (0.107) -0.083 (0.082) -6.920*** (0.782) -6.267*** (0.895) 0.129*** (0.050) 0.073** (0.037) -0.033 (0.068)
0.201** (0.094) 0.231* (0.118) 0.103 (0.094) 0.222* (0.129) -0.271*** (0.094) -6.169*** (1.215) -5.190*** (1.406) 0.154*** (0.059) 0.101** (0.042) -0.147* (0.086)
Across-class feedback rewarded financially
60
0.282** (0.113) 0.277** (0.139) 0.173* (0.094) -0.054 (0.128) -0.063 (0.097) -5.607*** (0.897) -4.618*** (1.038) 0.154** (0.064) -0.067 (0.059) 0.045 (0.079)
0.187** (0.073) 0.209** (0.103) 0.087 (0.080) 0.197 (0.162) -0.196* (0.103) -6.468*** (0.914) -6.681*** (1.096) 0.063 (0.078) 0.059 (0.059) -0.045 (0.088)
Across-class feedback rewarded reputationally 0.122* (0.073) 0.188** (0.080) 0.047 (0.080) § 0.198 (0.124) -0.237* (0.122) -6.098*** (1.008) -5.782*** (1.186) 0.035 (0.089) 0.042 (0.044) 0.031 (0.082)
Appendix C1: OLS estimates of the effects of different incentive schemes on students’ performance and well-being (continued) (Pure financial rewards and Pure financial Within-class feedback Across-class feedback interactions) rewards rewarded financially rewarded financially Math and English pooled (in st.dev.) Mathematics (in st.dev) English (in st.dev.) Stress Happiness Confidence (Math) Confidence (English) Education over work Education over rest Work over rest
(Pure reputational rewards and interactions)
Math and English pooled (in st.dev.) Mathematics (in st.dev)
0.129* (0.068) 0.106 (0.101) 0.045 (0.088) 0.466*** (0.162) -0.166§ (0.105) 0.794 (0.892) 1.662§ (1.021) 0.031 (0.077) 0.014 (0.068) 0.045 (0.083)
Pure reputational rewards
0.201** (0.094) 0.231* (0.118) 0.103 (0.094) 0.222* (0.129) -0.271*** (0.094) -6.169*** (1.215) -5.190*** (1.406) 0.154*** (0.059) 0.101** (0.042) -0.147* (0.086)
Within-class feedback rewarded reputationally
0.282** (0.113) 0.277** (0.139) 0.173* (0.094) -0.054 (0.128) -0.063 (0.097) -5.607*** (0.897) -4.618*** (1.038) 0.154** (0.064) -0.067 (0.059) 0.045 (0.079)
Across-class feedback rewarded reputationally
0.062 0.187** 0.122* (0.106) (0.073) (0.073) 0.209** 0.188** 0.138 (0.103) (0.080) (0.141) 0.087 0.047 English (in st.dev.) 0.016 (0.080) (0.080) (0.082) § 0.197 0.198 Stress 0.074 (0.162) (0.189) (0.123) -0.196* Happiness -0.105 -0.237* (0.103) (0.097) (0.122) -6.468*** Confidence (Math) 1.498* -6.098*** (0.914) (0.782) (1.008) -6.681*** Confidence (English) 0.987 -5.782*** (1.096) (0.967) (1.186) 0.063 Education over work 0.083 0.035 (0.078) (0.076) (0.089) 0.059 Education over rest 0.037 0.042 (0.059) (0.047) (0.044) Work over rest 0.081 -0.045 0.031 (0.087) (0.088) (0.082) Note: Robust standard errors adjusted for clustering at class level are in parentheses. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
61
Appendix C2: Average treatment effects of different types of interventions on performance and students’ well-being Overall performance Mathematics English Stress (Math+English) Dependent variable: (All areas) (Without (All areas) (Without (All areas) (Without (All areas) area 1) area 1) area 1) Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Repurational Rewards, no feedback (Rep_solo) Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputat.reward (T1_rep) Across-class feedback, reputat.reward (T2_rep)
0.017 (0.060) 0.044 (0.059) 0.129* (0.068) 0.062 (0.106) 0.201** (0.094) 0.292** (0.113) 0.187** (0.073) 0.122* (0.073)
-0.031 (0.109) 0.014 (0.097) § 0.134 (0.087) 0.017 (0.116) 0.190* (0.101) 0.257** (0.117) 0.180** (0.087) 0.104 (0.087)
0.100 (0.085) 0.082 (0.073) 0.106 (0.101) 0.138 (0.141) 0.231* (0.119) 0.277** (0.139) 0.209** (0.103) 0.188** (0.080)
-0.036 (0.135) -0.004 (0.126) 0.069 (0.129) 0.045 (0.158) 0.182 (0.134) § 0.215 (0.141) 0.165 (0.123) 0.136 (0.108)
-0.128** (0.056) -0.049 (0.059) 0.045 (0.088) 0.016 (0.082) 0.103 (0.094) 0.173* (0.094) 0.087 (0.080) 0.047 (0.080)
-0.141 (0.115) -0.051 (0.109) 0.064 (0.111) -0.003 (0.105) 0.109 (0.109) § 0.168 (0.112) 0.093 (0.099) 0.045 (0.101)
0.059 (0.113) -0.059 (0.466) 0.466*** (0.163) 0.074 (0.189) 0.222* (0.129) -0.054 (0.128) 0.197 (0.162) § 0.198 (0.124)
Happiness (All areas) -0.131 (0.095) -0.083 (0.082) -0.166§ (0.105) -0.105 (0.097) -0.271*** (0.094) -0.063 (0.097) -0.196* (0.103) -0.237* (0.122)
Controlled for stratas Yes Yes Yes Yes Yes Yes Yes Yes N 5108 5102 5093 4105 3516 3512 3503 4105 Note: The treatment effects are calculated with respect to the control group. I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at class level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%.
62
Appendix C3: Aggregated average treatment effects, by subject
Dependent variable: Math score
(1)
Feedback provided (T)
0.102** (0.051)
Within-class feedback (T1) Across-class feedback (T2) Rewards provided (Rew) Financial Rewards (Finrew) Repurational Rewards (Reprew)
(2)
MATHEMATICS (3) (4)
(5)
(6)
0.094* (0.051) 0.112* (0.059) 0.093* (0.055)
0.099* (0.059) 0.089§ (0.056) 0.128** (0.063)
0.138** (0.066) 0.151* (0.082) 0.127* (0.066)
(7)
(8)
-0.001 (0.037)
0.125** (0.056) 0.142* (0.078) 0.115* (0.064)
ENGLISH (9) (10)
(11) -0.009 (0.036)
-0.015 (0.042) 0.014 (0.042)
0.126** (0.055) 0.153** (0.066) 0.103* (0.054)
(12)
-0.028 (0.039) 0.012 (0.040) 0.158** (0.065) 0.108** (0.053)
Controlled for stratas Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes N 5102 5102 5102 5102 5102 5102 5093 5093 5093 5093 5093 5093 Note: Rows represent aggregated treatment groups. The treatment effects are calculated with respect to the control group. Students in group “T” are those who received any type of feedback (either within-class feedback (T1) or across-class feedback (T2)). Students in the group “Rew” received any type of reward (either financial reward (Finrew) or reputational reward (Reprew)). Columns (1) – (6) represent the treatment effects in Mathematics, columns (7) – (12) in English. In all specifications I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%. Significant differences between estimates within particular specifications: there is a statistical difference between provision of feedback (T) and rewards (Rew) in the English performance (p=0.025, column 11); between within-class feedback (T1) and financial or reputational rewards (Finrew or Reprew) in the English performance (p=0.009 and p=0.029 respectively, column 11), across-class feedback (T2) and financial or rewards (Finrew or Reprew) in the English performance (p=0.038 and p=0.113 respectively, column 12).
63
Appendix C4: Different specifications used in estimation of the aggregated average treatment effects of the provision of feedback or rewards on performance in Mathematics MATHEMATICS Dependent variable: Math score
Within-class social comparison (Treatment 1) Across-class social comparison (Treatment 2) Financial Rewards Repurational Rewards Controlled for stratas Interactions N
Pure FB (round 4)
Pure FB (round 4)
0.024 (0.062) 0.005 (0.058)
0.037 (0.048) 0.043 (0.043)
No No 5245
Yes No 5245
Pure FB (round 5)
Pure FB (round 5)
0.084 (0.081) 0.024 (0.084)
0.112* (0.059) 0.093* (0.055)
No No 5102
Yes No 5102
Pure Rewards (round 5)
0.231** (0.092) 0.185** (0.079) No No 5102
Pure Rewards (round 5)
Mix FB and Rewards (round 5)
0.151* (0.082) 0.127* (0.066) Yes No 5102
0.086 (0.079) 0.046 (0.081) 0.233** (0.093) 0.184** (0.078) No No 5102
Mix FB and Rewards (round 5)
0.099* (0.059) 0.089§ (0.056) 0.142* (0.078) 0.115* (0.064) Yes No 5102
Note: Robust standard errors adjusted for clustering at class level are in parentheses. The treatment effects are calculated with respect to the control group. Columns (2), (4), (6) and (8) control for stratum fixed effects (areas (by distance from the capital city, Kampala), school performance at national examination and grade level; P6,P7, S1 up to S4). N stands for the number of observations. First two columns analyze the effect between testing round 4 and the baseline testing in round 1. The remaining estimates are based on the differences between round 5 (the final testing round) and the baseline round 1. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
64
Appendix C5: Different specifications used in estimation of the aggregated average treatment effects of the provision of feedback or rewards on performance in English ENGLISH
Dependent variable: English score
Pure FB (round 4)
Pure FB (round 4)
Pure FB (round 5)
Pure FB (round 5)
-0.102§ (0.067) -0.039 (0.071)
-0.015 (0.042) 0.014 (0.042)
Pure Rewards (round 5)
Pure Rewards (round 5)
Mix FB and Rewards (round 5)
Mix FB and Rewards (round 5)
OVERALL TREATMENT EFFECTS -0.099* -0.028 (0.058) (0.039) -0.007 0.012 (0.064) (0.040) 0.340*** 0.336*** 0.153** 0.158** (0.052) (0.055) (0.066) (0.053) 0.254*** 0.250** 0.103* 0.108** Repurational Rewards (0.067) (0.066) (0.054) (0.053) No No No Yes No Yes Yes Controlled for stratas Yes Interactions No No No No No No No No N 5246 5246 5093 5093 5093 5093 5093 5093 Note: Robust standard errors adjusted for clustering at class level are in parentheses. The treatment effects are calculated with respect to the control group. Columns (2), (4), (6) and (8) control for stratum fixed effects (areas (by distance from the capital city, Kampala), school performance at national examination and grade level; P6,P7, S1 up to S4). N stands for the number of observations. First two columns analyze the effect between testing round 4 and the baseline testing in round 1. The remaining estimates are based on the differences between round 5 (the final testing round) and the baseline round 1. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1% Within-class social comparison (Treatment 1) Across-class social comparison (Treatment 2) Financial Rewards
-0.040 (0.074) 0.027 (0.073)
0.023 (0.043) 0.062§ (0.042)
65
Appendix C6: Different specifications used in estimation of the dis-aggregated average treatment effects of different incentive schemes on performance in Mathematics and English Mathematics English Dependent variable: (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Repurational Rewards, no feedback (Rep_solo) Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputat.reward (T1_rep) Across-class feedback, reputat.reward (T2_rep)
0.100 (0.085) 0.082 (0.073) 0.106 (0.101) 0.138 (0.141) 0.231* (0.118) 0.277** (0.139) 0.209** (0.103) 0.188** (0.080)
0.104 (0.085) 0.081 (0.073) 0.097 (0.099) 0.115 (0.142) 0.215* (0.115) 0.281** (0.136) 0.196** (0.099) 0.138* (0.079) 0.002 (0.002)
Average class size Gender
0.101 (0.085) 0.077 (0.074) 0.101 (0.101) 0.137 (0.142) 0.229* (0.119) 0.279** (0.141) 0.203** (0.103) 0.183** (0.081)
Baseline score
Controlled for stratas N
0.095 (0.087) 0.070 (0.075) 0.085 (0.106) 0.135 (0.145) 0.215 (0.120) 0.251 (0.144) 0.212 (0.106) 0.179 (0.086)
-0.009 (0.063) 0.725*** (0.017) Yes 5102
0.723*** (0.017) Yes 5102
-0.128** (0.056) -0.049 (0.059) 0.045 (0.088) 0.016 (0.082) 0.103 (0.094) 0.173* (0.094) 0.087 (0.080) 0.047 (0.080)
-0.068*** (0.022)
Public Food
0.100 (0.084) 0.081 (0.075) 0.104 (0.105) 0.135 (0.145) 0.227* (0.127) 0.272* (0.146) 0.205* (0.111) 0.185** (0.087)
0.719*** (0.017) Yes 5065
0.725*** (0.017) Yes 5102
66
0.089*** (0.029) 0.723*** (0.017) Yes 4906
-0.127** (0.056) -0.049 (0.059) 0.043 (0.089) 0.009 (0.084) 0.099 (0.095) 0.174* (0.093) 0.083 (0.081) 0.034 (0.085) 0.001 (0.001)
-0.134** (0.056) -0.052 (0.059) 0.042 (0.088) 0.016 (0.081) 0.097 (0.093) 0.166* (0.093) 0.081 (0.079) 0.043 (0.079)
-0.129** (0.055) -0.045 (0.061) 0.053 (0.087) 0.027 (0.082) 0.124 (0.097) 0.193** (0.097) 0.104 (0.086) 0.059 (0.081)
0.052*** (0.020) 0.041 (0.049)
0.737*** (0.016) Yes 5093
0.737*** (0.016) Yes 5093
-0.134** (0.055) -0.043 (0.060) 0.063 (0.089) 0.035 (0.083) 0.108 (0.091) 0.189** (0.090) 0.097 (0.082) 0.053 (0.081)
0.737*** (0.016) Yes 5056
0.736*** (0.016) Yes 5093
0.058*** (0.021) 0.732*** (0.016) Yes 4896
Appendix D: Gender differences Appendix D1: Gender differences in the average treatment effects of aggregated interventions on the overall performance (Math and English pooled)
Dependent variable: Math score
(1)
Feedback provided (T)
0.104** (0.043)
(6)
0.097** (0.041) 0.098** (0.058) 0.109** (0.060)
Within-class feedback (T1) Across-class feedback (T2) Rewards provided (Rew) Financial Rewards (Finrew) Reputational Rewards (Reprew) Controlled for stratas N
OVERALL PERFORMANCE BY GIRLS (2) (3) (4) (5)
0.067 (0.065)
Yes 2862
Yes 2862
Yes 2862
0.089** (0.044) 0.106** (0.050)
0.114§ (0.076) 0.083 (0.059) Yes 2862
0.042 (0.061)
Yes 2862
(7)
OVERALL PERFORMANCE BY BOYS (8) (9) (10) (11)
0.026 (0.048)
0.134** (0.063) 0.105§ (0.072) 0.070 (0.057) Yes 2862
Yes 2209
Yes 2209
0.013 (0.047)
0.038 (0.056) 0.016 (0.053)
Yes 2209
0.133** (0.063) 0.259*** (0.066) 0.129** (0.059) Yes 2209
Yes 2209
(12)
0.019 (0.052) 0.021 (0.051) 0.258*** (0.065) 0.128** (0.059) Yes 2209
Note: Rows represent aggregated treatment groups. Students in group “T” are those who received any type of feedback (either within-class feedback (T1) or across-class feedback (T2)). Students in the group “Rew” received any type of reward (either financial reward (Finrew) or reputational reward (Reprew)). Columns (1) – (6) represent the treatment effects in Mathematics, columns (7) – (12) in English. In all specifications I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%.
Significant differences between estimates within particular specifications: there is a statistical difference between within-class feedback (T1) and financial or reputational rewards (Finrew or Reprew) in the English performance (p=0.111 and p=0.043 respectively, column 12).
67
Appendix D2: Girls’ aggregated average treatment effects on performance in Mathematics and English
Dependent variable: Math score
(1)
Feedback provided (T)
0.159*** (0.053)
Within-class feedback (T1) Across-class feedback (T2) Rewards provided (Rew)
(2)
MATHEMATICS (3) (4)
(5)
(6)
0.154*** (0.053) 0.157*** (0.058) 0.163*** (0.060) 0.093 (0.076)
0.149** (0.058) 0.159*** (0.061) 0.073 (0.071)
(7)
(8)
0.001 (0.041)
0.093§ (0.058)
ENGLISH (9) (10)
-0.016 (0.044) 0.019 (0.046)
(11) -0.007 (0.039)
0.094* (0.057)
(12)
-0.027 (0.042) 0.014 (0.045)
Financial Rewards 0.103 0.088 0.089 0.094 (Finrew) (0.095) (0.088) (0.069) (0.068) Reputational Rewards 0.085 0.062 0.095§ 0.099* (Reprew) (0.075) (0.071) (0.056) (0.058) Controlled for stratas Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes N 2858 2858 2858 2858 2858 2858 2854 2854 2854 2854 2854 2854 Note: Rows represent aggregated treatment groups. Students in group “T” are those who received any type of feedback (either within-class feedback (T1) or across-class feedback (T2)). Students in the group “Rew” received any type of reward (either financial reward (Finrew) or reputational reward (Reprew)). Columns (1) – (6) represent the treatment effects in Mathematics, columns (7) – (12) in English. In all specifications I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%. Significant differences between estimates within particular specifications: there is a statistical difference between within-class feedback (T1) and financial or reputational rewards (Finrew or Reprew) in the English performance (p=0.111 and p=0.043 respectively, column 12).
68
Appendix D3: Boys’ aggregated average treatment effects on performance in Mathematics and English MATHEMATICS Dependent variable: (1) (2) (3) (4) (5) (6) (7) (8) Math score Feedback provided (T)
Within-class feedback (T1) Across-class feedback (T2) Rewards provided (Rew)
0.027 (0.061)
0.192*** (0.022)
0.055 (0.073) -0.001 (0.064)
0.019 (0.060)
0.191*** (0.069)
0.038 (0.071) 0.003 (0.065)
-0.011 (0.047)
0.157** (0.068)
ENGLISH (9) (10)
-0.022 (0.054) 0.001 (0.053)
(11) -0.017 (0.046)
0.158** (0.068)
(12)
-0.038 (0.051) 0.005 (0.051)
Financial Rewards 0.213** 0.207** 0.226*** 0.234*** (Finrew) (0.089) (0.089) (0.078) (0.078) § Reputational Rewards 0.175** 0.170** 0.106 0.111* (Reprew) (0.074) (0.073) (0.067) (0.066) Controlled for stratas Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes N 2207 2207 2207 2207 2207 2207 2202 2202 2202 2202 2202 2202 Note: Rows represent aggregated treatment groups. Students in group “T” are those who received any type of feedback (either within-class feedback (T1) or across-class feedback (T2)). Students in the group “Rew” received any type of reward (either financial reward (Finrew) or reputational reward (Reprew)). Columns (1) – (6) represent the treatment effects in Mathematics, columns (7) – (12) in English. In all specifications I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%. Significant differences between estimates within particular specifications: in Mathematics, there is a statistical difference between provision of feedback (T) and rewards (Rew) (p=0.045, column 5), between within-class feedback (T1) and financial or reputational rewards (Finrew or Reprew) in the English performance (p=0.112 and p=0.027 respectively, column 12), and between across-class feedback (T2) and reputational rewards (p=0.092). In English, there are statistical differences between financial and reputational rewards in the specification 10 (p=0.046), between provision of feedback (T) and rewards (Rew) in the specification 11 (p=0.023), and in the specification 12 there are differences between financial and reputational reward estimates (p=0.044), between within-class feedback (T1) and financial or reputational reward estimates (p=0.002 and p=0.068 respectively), and between across-class feedback (T2) and financial reward estimates (p=0.009).
69
Appendix D4: Gender differences in aggregated average treatment effects on stress and happiness
Dependent variable: Math score
Feedback provided (T)
Within-class feedback (T1) Across-class feedback (T2) Rewards provided (Rew)
OVERALL (1) -0.133 (0.204)
0.516* (0.269)
(2)
0.003 (0.233) -0.275 (0.213)
PERCEIVED STRESS GIRLS (3) -0.089 (0.234)
0.508* (0.284)
(4)
0.105 (0.259) -0.295 (0.239)
BOYS (5)
OVERALL (6)
-0.183 (0.216)
0.533* (0.295)
(7)
(8)
-0.326* (0.195) -0.119 (0.255) -0.248 (0.239)
-0.437* (0.228)
PERCEIVED HAPPINESS GIRLS (9)
(10)
-0.313§ (0.214) -0.428* (0.224) -0.219 (0.219)
-0.349 (0.255)
-0.399* (0.235) -0.221 (0.253)
BOYS (11) -0.348 (0.251)
-0.544* (0.304)
(12)
-0.483§ (0.297) -0.213 (0.268)
Financial Rewards 0.549** 0.599** 0.477 -0.398 -0.331 -0.504 (Finrew) (0.277) (0.282) (0.334) (0.258) (0.277) (0.375) Reputational Rewards 0.408 0.307 0.547* -0.408 -0.314 -0.493§ (0.318) (0.327) (Reprew) (0.304) (0.331) (0.267) (0.298) Controlled for stratas Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Note: Rows represent aggregated treatment groups. Students in group “T” are those who received any type of feedback (either within-class feedback (T1) or across-class feedback (T2)). Students in the group “Rew” received any type of reward (either financial reward (Finrew) or reputational reward (Reprew)). Columns (1) – (6) represent the treatment effects in Mathematics, columns (7) – (12) in English. In all specifications I controlled for stratas (i.e., students’ performance at the national examinations, area and the level of studies). N stands for the number of observations. Robust standard errors adjusted for clustering at school level are in parentheses. § significant at 15%, * significant at 10%; ** significant at 5%; *** significant at 1%. Significant differences between estimates within particular specifications: there is a statistical difference between within-class feedback (T1) and financial or reputational rewards (Finrew or Reprew) in the English performance (p=0.111 and p=0.043 respectively, column 12).
70
Appendix D5: Gender differences in aggregated average treatment effects – by gender and by subject
Note: The graph illustrates the improvements in Mathematics and English (in standard deviations) of girls and boys in different treatment groups with respect to the control group. The bars correspond to robust standard errors. T1 stands for the within-class comparative feedback, T2 stands for the across-class comparative feedback, Finrew stands for financial-reward group and Reprew stands for the reputationalreward group.
71
Appendix D6: Gender differences in dis-aggregated average treatment effects on students’ overconfidence, by subject
Note: Students answered short questionnaires before and after every exam. Overconfidence in this case is measured by the difference between the expected number of points from the particular subject and the real score the student achieved. The calibration is measured in the number of points students overestimated their performance. Maximum number of points in each of the subjects was 50. G stands for the estimated effects on girls, B for boys. The bars correspond to robust standard errors.
72
Appendix D6: The evolution of the average treatment effects over time – by gender and by subject
Note: The graph illustrates the development in improvements in overall performance, in Mathematics and English separately (in standard deviations) of girls and boys in different feedback treatment groups with respect to the control group. The bars correspond to robust standard errors. T1 stands for the within-class comparative feedback, T2 stands for the across-class comparative feedback and C for the control group.
73
Appendix E: Group composition Appendix E1: Comparison of performance, exerted effort and subjective well-being of mixed- and high-ability groups to poor-ability groups and their responses to the provision of rewards ABILITY COMPOSITION OF GROUPS
Mathematics
English
Perceived Stress
Subjective Happiness
Effort Math
Effort English
Mixed ability: pure feedback Well performers: pure feedback
0.055 (0.059) 0.451*** (0.129)
0.133** (0.063) 0.387*** (0.078)
0.275** (0.127) 0.561*** (0.206)
0.521*** (0.086) 0.449*** (0.099)
0.096 (0.094) 0.078 (0.176)
-0.069 (0.094) 0.082 (0.096)
0.004 (0.096) § 0.126 (0.081)
Mixed ability: feedback and monetary reward Well performers: feedback and monetary reward
-0.082 (0.078) 0.001 (0.161)
Mixed ability: feedback and reputational reward Well performers: feedback and reputational reward
0.262*** (0.092) 0.322** (0.149)
0.357*** (0.078) 0.393*** (0.107)
0.181 (0.143) 0.327** (0.148)
Financial rewards
0.271*** (0.079) 0.253** (0.123) 0.672*** (0.029)
0.289* (0.169) 0.280* (0.141) 0.703*** (0.029)
0.219 (0.166) 0.129 (0.171)
0.252** (0.125) 0.127 (0.148)
0.284** (0.132) 0.071 (0.150)
Yes 1323
Yes 1392
Reputational rewards Initial value
Stratification variables N
Yes 1426
Yes 1425
0.313** (0.124) 0.051 (0.254)
0.208 (0.156) -0.028 (0.162) 0.082*** (0.026) Yes 1327
74
0.121 (0.155) 0.225 (0.182)
0.094 (0.172) -0.193 (0.172) 0.229*** (0.035)
0.109 (0.112) 0.053 (0.109)
0.050 (0.130) 0.058 (0.121) 0.254*** (0.032)
0.102 (0.128) 0.042 (0.128)
0.089 (0.128) -0.036 (0.166) 0.219*** (0.039) Yes 1365
Appendix E2: Comparison of performance, exerted effort and subjective well-being of mixed-gender and pure-boy groups to groups consisting of pure girls and their responses to the provision of rewards GENDER COMPOSITION OF GROUPS Three boys: pure feedback Two boys + One girl: pure feedback One boy + Two girls: pure feedback Three boys: feedback and monetary reward Two boys + 1 girl: feedback and monetary reward One boy + 2 girls: feedback and monetary reward Three boys: feedback and reputational reward Two boys + 1 girl: feedback and reputational reward One boy + 2 girls: feedback and reputational reward Financial rewards Reputational rewards Initial value Stratification variables N
Math 0.065 (0.071) 0.184*** (0.055) 0.164*** (0.044)
Perceived Stress
Subjective Happiness
Effort Math
Effort English
-0.009 (0.065) -0.026 (0.068) 0.002 (0.044)
-0.056 (0.165) -0.104 (0.111) -0.022 (0.104)
-0.288** (0.136) -0.160** (0.079) -0.204** (0.099)
0.067 (0.148) -0.121 (0.133) -0.192* (0.106)
-0.143 (0.139) -0.171§ (0.107) -0.258** (0.108)
0.197§ (0.129) 0.153 (0.107) 0.203** (0.106)
0.046 (0.272) 0.263 (0.205) 0.174 (0.197)
-0.258 (0.207) -0.302* (0.179) -0.182 (0.172)
0.033 (0.167) 0.071 (0.139) -0.046 (0.156)
-0.256 (0.173) -0.137 (0.134) -0.135 (0.146)
English
0.263 (0.301) 0.465*** (0.143) 0.359** (0.154)
0.264 (0.181) 0.310** (0.139) 0.303*** (0.106)
0.411*** (0.154) 0.403*** (0.125) 0.707*** (0.029) Yes 1624
0.309** (0.137) 0.382*** (0.119) 0.734*** (0.028) Yes 1623
0.504*** (0.149) 0.388*** (0.113) 0.206§ (0.131)
75
0.295 (0.366) 0.142 (0.236) 0.302§ (0.197)
-0.029 (0.205) -0.114 (0.215) 0.081*** (0.026) Yes 1624
-0.117 (0.234) -0.303* (0.172) -0.345** (0.123)
-0.096 (0.187) -0.019 (0.148) 0.077 (0.139)
-0.196 (0.157) -0.182 (0.178) 0.233*** (0.034) Yes 1313
0.195 (0.167) -0.071 (0.138) 0.258*** (0.032) Yes 1392
-0.259 (0.216) -0.007 (0.128) -0.034 (0.139)
0.029 (0.189) -0.167 (0.146) 0.224*** (0.039) Yes 1365
Appendix F: Other
Appendix F1: Aggregated average treatment effects of different incentive schemes on performance in Mathematics and English – by students being at official age in his class or older
Note: The graph illustrates the differences in improvements in Mathematics and English (in standard deviations) between students of official age and unofficial age (students older than usual age at the particular level) in different treatment groups with respect to the control group (e.g., if the official age of getting to the primary school is 6, then the official age at the P6 level is 12 (±1)). The bars correspond to robust standard errors. T1 stands for the within-class comparative feedback, T2 stands for the across-class comparative feedback, Finrew stands for financial-reward group and Reprew stands for the reputational-reward group.
76
Appendix F2: Comparison of the estimates of the average treatment effects of different motivation schemes on students’ performance in Mathematics and English using various specifications OLS
IPW
Imputation (median ratio)
Imputation (class percentiles)
Median Regression
MATHEMATICS Within-class feedback (T1) Across-class feedback (T2)
Financial Rewards (Finrew) Reputational Rewards (Reprew) Baseline Mathematic score Controlled for stratas ENGLISH
Within-class feedback (T1) Across-class feedback (T2)
Financial Rewards (Finrew) Reputational Rewards (Reprew) Baseline English score Controlled for stratas
0.099* (0.059) 0.089§ (0.056) 0.142* (0.078) 0.115* (0.064) 0.725*** (0.017) Yes
0.076 (0.065) 0.107* (0.066) 0.327*** (0.100) 0.152** (0.073) 0.731*** (0.021) Yes
0.124* (0.063) 0.116** (0.054) 0.198** (0.081) 0.164** (0.073) 0.757** (0.049) Yes
0.112* (0.055) 0.096* (0.053) 0.169** (0.079) 0.157** (0.067) 0.668*** (0.019) Yes
0.096** (0.052) 0.069 (0.051) 0.175** (0.083) 0.124** (0.062) 0.750*** (0.019) Yes
-0.028 (0.039) 0.012 (0.040) 0.158** (0.053) 0.108** (0.053) 0.739*** (0.016) Yes
-0.029 (0.048) 0.028 (0.048) 0.290*** (0.083) 0.155** (0.068) 0.696*** (0.024) Yes
0.019 (0.052) 0.056 (0.051) 0.159* (0.075) 0.103 (0.073) 0.737*** (0.026) Yes
-0.011 (0.042) 0.025 (0.043) 0.211*** (0.066) 0.158*** (0.056) 0.691*** (0.016) Yes
-0.018 (0.042) 0.017 (0.042) 0.176** (0.075) 0.095§ (0.058) 0.758*** (0.016) Yes
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. IPW stands for the inverse probability weight regression adjusting for students’ probability to dropout. The imputations based on median ration imputed the last available observation in Mathematics or English adjusted for the difference in the test difficulties using median ratio. In the imputations based on class percentiles I first seek for the percentile rank of the student in his last observed score within his class and then assign the student grade from the testing round 5 of a student from the same percentile rank. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
77
Appendix F3: Comparison of the estimates of the disaggregated average treatment effects of different motivation schemes on students’ performance in Mathematics using various specifications OLS
IPW
Imputation (median correction)
Imputation (class percentiles)
Median Regression
PURE TREATMENT EFFECTS Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Reputational Rewards, no feedback (Rep_solo)
0.100 (0.085) 0.082 (0.073) 0.106 (0.101) 0.138 (0.141)
0.046 (0.092) 0.067 (0.079) 0.151§ (0.102) 0.188 (0.149)
0.133* (0.079) 0.129* (0.068) 0.169* (0.096) 0.206* (0.124)
0.070 (0.082) 0.036 (0.627) 0.070 (0.097) 0.092 (0.115)
0.121* (0.072) 0.070 (0.069) 0.162* (0.093) 0.159 (0.132)
Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputational reward (T1_rep) Across-class feedback, reputational reward (T2_rep)
0.231* (0.118) 0.277** (0.139) 0.209** (0.103) 0.188** (0.080)
0.338** (0.135) 0.456*** (0.132) 0.212* (0.108) 0.208** (0.087)
0.281** (0.129) 0.331** (0.128) 0.266** (0.112) 0.186** (0.073)
0.202* (0.116) 0.209§ (0.130) 0.171* (0.099) 0.164** (0.076)
0.267** (0.113) 0.296* (0.160) 0.214** (0.100) 0.193*** (0.075)
0.725*** (0.017) Yes
0.732*** (0.021) Yes
0.755*** (0.048) Yes
0.658*** (0.019) Yes
0.747*** (0.019) Yes
TREATMENT INTERACTIONS
Baseline Math score
Controlled for stratas
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. IPW stands for the inverse probability weight regression adjusting for students’ probability to dropout. The imputations based on median ration imputed the last available observation in Mathematics or English adjusted for the difference in the test difficulties using median ratio. In the imputations based on class percentiles I first seek for the percentile rank of the student in his last observed score within his class and then assign the student grade from the testing round 5 of a student from the same percentile rank. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
78
Appendix F4: Comparison of the estimates of the disaggregated average treatment effects of different motivation schemes on students’ performance in English using various specifications OLS
IPW
Imputation (median correction)
Imputation (class percentiles)
Median Regression
PURE TREATMENT EFFECTS Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Reputational Rewards, no feedback (Rep_solo)
-0.128** (0.056) -0.049 (0.059) 0.045 (0.088) 0.016 (0.082)
-0.133* (0.070) -0.079 (0.072) 0.032 (0.085) 0.004 (0.084)
-0.151*** (0.043) -0.062 (0.046) 0.036 (0.069) 0.026 (0.059)
Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputational reward (T1_rep) Across-class feedback, reputational reward (T2_rep)
0.103 (0.094) 0.173* (0.094) 0.087 (0.080) 0.047 (0.080)
0.145* (0.086) 0.248** (0.102) 0.041 (0.078) 0.071 (0.077)
0.065 (0.079) 0.128* (0.074) 0.062 (0.057) 0.052 (0.065)
TREATMENT INTERACTIONS
Baseline Math score
Controlled for stratas
0.737*** (0.016) Yes
0.697*** (0.023) Yes
0.702*** (0.014) Yes
-0.207*** (0.062) -0.139** (0.065) -0.047 (0.093) -0.099 (0.086)
0.043 (0.101) 0.062 (0.104) -0.009 (0.081) -0.042 (0.087)
0.682*** (0.017) Yes
-0.149*** (0.053) -0.074 (0.055) 0.009 (0.084) -0.025 (0.078)
0.129 (0.107) 0.175** (0.084) 0.096 (0.077) -0.039 (0.079)
0.759*** (0.017) Yes
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. IPW stands for the inverse probability weight regression adjusting for students’ probability to dropout. The imputations based on median ration imputed the last available observation in Mathematics or English adjusted for the difference in the test difficulties using median ratio. In the imputations based on class percentiles I first seek for the percentile rank of the student in his last observed score within his class and then assign the student grade from the testing round 5 of a student from the same percentile rank. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
79
Appendix F5: Quantile regressions – aggregated treatment effects on performance in Mathematics and English OLS
Quantile Regression (q=0.25)
Quantile Regression (q=0.5)
0.069§ (0.047) 0.069§ (0.048) 0.145* (0.080) 0.078 (0.063) 0.702*** (0.024)
0.096** (0.052) 0.069 (0.051) 0.175** (0.083) 0.124** (0.062) 0.750*** (0.019)
Quantile Regression (q=0.75)
MATHEMATICS Within-class feedback (T1) Across-class feedback (T2)
Financial Rewards (Finrew)
Reputational Rewards (Reprew) Baseline Mathematic score Controlled for stratas ENGLISH
Within-class feedback (T1) Across-class feedback (T2)
Financial Rewards (Finrew)
Reputational Rewards (Reprew) Baseline English score Controlled for stratas
0.099* (0.059) 0.089§ (0.056) 0.142* (0.078) 0.115* (0.064) 0.725*** (0.017)
0.101§ (0064) 0.061 (0.063) 0.127 (0.092) 0.125* (0.075) 0.770*** (0.027)
Yes
Yes
Yes
Yes
-0.028 (0.039) 0.012 (0.040) 0.158** (0.053) 0.108** (0.053) 0.739*** (0.016) Yes
-0.061§ (0.039) 0.014 (0.044) 0.181** (0.071) 0.099* (0.058) 0.746*** (0.018) Yes
-0.018 (0.042) 0.017 (0.042) 0.176** (0.075) 0.095§ (0.058) 0.758*** (0.016) Yes
-0.021 (0.049) -0.021 (0.044) 0.160** (0.065) 0.105* (0.058) 0.764*** (0.020) Yes
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. In that case imputation would not work properly. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
80
Appendix F6a: Quantile regressions – dis-aggregated treatment effects on performance in Mathematics OLS Quantile Quantile Quantile Regression Regression Regression (q=0.25) (q=0.5) (q=0.75) PURE TREATMENT EFFECTS Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Reputational Rewards, no feedback (Rep_solo)
0.100 (0.085) 0.082 (0.073) 0.106 (0.101) 0.138 (0.141)
0.097§ (0.059) 0.084 (0.059) 0.131§ (0.089) 0.181* (0.102)
0.121* (0.072) 0.070 (0.069) 0.162* (0.093) 0.159 (0.132)
0.087 (0.092) 0.050 (0.093) 0.012 (0.123) 0.137 (0.156)
Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputational reward (T1_rep) Across-class feedback, reputational reward (T2_rep)
0.231* (0.118) 0.277** (0.139) 0.209** (0.103) 0.188** (0.080)
0.240** (0.123) 0.253* (0.137) 0.148§ (0.092) 0.163* (0.096)
0.267** (0.113) 0.296* (0.160) 0.214** (0.100) 0.193*** (0.075)
0.725*** (0.017) Yes
0.700*** (0.024) Yes
0.747*** (0.019) Yes
0.239§ (0.147) 0.257 (0.192) 0.237* (0.128) 0.141 (0.099)
TREATMENT INTERACTIONS
Baseline English score
Controlled for stratas
0.764*** (0.028) Yes
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. In that case imputation would not work properly. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
81
Appendix F6b: Quantile regressions – dis-aggregated treatment effects on performance in English OLS
Quantile Regression (q=0.25)
Quantile Regression (q=0.5)
Quantile Regression (q=0.75)
PURE TREATMENT EFFECTS Within-class feedback, no feedback (T1_solo) Across-class feedback, no feedback (T2_solo) Financial Rewards, no feedback (Fin_solo) Reputational Rewards, no feedback (Rep_solo)
-0.128** (0.056) -0.049 (0.059) 0.045 (0.088) 0.016 (0.082)
-0.137*** (0.052) -0.025 (0.055) 0.078 (0.100) 0.009 0.084
-0.149*** (0.053) -0.074 (0.055) 0.009 (0.084) -0.025 (0.078)
-0.127** (0.064) -0.074 (0.064) -0.029 (0.099) -0.002 (0.098)
Within-class feedback, monetary reward (T1_fin) Across-class feedback, monetary reward (T2_fin) Within-class feedback, reputational reward (T1_rep) Across-class feedback, reputational reward (T2_rep)
0.103 (0.094) 0.173* (0.094) 0.087 (0.080) 0.047 (0.080)
0.049 (0.091) 0.209** (0.097) 0.093 (0.084) -0.024 (0.089)
0.129 (0.107) 0.175** (0.084) 0.096 (0.077) -0.039 (0.079)
0.169* (0.089) 0.099 (0.078) 0.060 (0.087) 0.040 (0.077)
TREATMENT INTERACTIONS
Baseline English score
Controlled for stratas
0.737*** (0.016) Yes
0.749*** (0.017) Yes
0.759*** (0.017) Yes
0.759*** (0.018) Yes
Note: Robust standard errors adjusted for clustering at class level are in parentheses. Controlled for stratum fixed effects (four areas by distance from the capital city, Kampala, school performance at national examination and grade level (P6,P7, S1 up to S4). N stands for the number of observations. One school is eliminated from imputation method due to high turnover of students caused by frequent change of headmasters. In that case imputation would not work properly. § significant at 15%; * significant at 10%; ** significant at 5%; *** significant at 1%
82
Appendix F7: Comparison of uncorrected and corrected p-values using different multiple-comparison procedures: for aggregated and disaggregated average treatment effects on students’ overall performance
P-VALUE CORRECTIONS Aggregated ATE Within-class feedback Across-class feedback Disaggregated ATE
Within-class feedback, no rewards Across-class feedback, no rewards Financial Rewards, no feedback Reputational Rewards, no feedback Within-class feedback with financial rewards Across-class feedback with financial rewards Within-class feedback with reputation rewards Across-class feedback with reputation rewards
Uncorrect ed
Bonferroni
Sidak
Holm
Holland
Hochberg
Simes
Yekutieli
0.102 0.130
1.000 1.000
0.754 0.836
0.409 0.409
0.350 0.350
0.389 0.389
0.133 0.153
0.422 0.487
0.774
1.000
1.000
0.961
0.776
1.000
1.000
0.685
0.413
0.347
0.413
0.086
0.555
1.000
0.776
0.776
0.059
1.000
1.000
0.776 0.293
0.659
0.306
0.058
§
0.026
0.091
§
0.025
0.087
§
0.047
0.464
0.035 0.014 0.012 0.097
0.667 0.257 0.221
1.000
1.000 1.000 0.493
1.000 1.000 0.316
§
0.956 0.961 0.275
§
0.228
0.135
§
0.127
§
0.135
0.856
0.581
0.457
0.581
0.199
0.128
83
0.121
0.128
0.587
0.131
1.000 1.000 0.205
Appendix F8: Comparison of uncorrected and corrected p-values using different multiple-comparison procedures: for aggregated and disaggregated average treatment effects on students’ stress P-VALUE CORRECTIONS Aggregated ATE Within-class feedback Across-class feedback Disaggregated ATE
Within-class feedback, no rewards Across-class feedback, no rewards Financial Rewards, no feedback Reputational Rewards, no feedback Within-class feedback with financial rewards Across-class feedback with financial rewards Within-class feedback with reputation rewards Across-class feedback with reputation rewards
Uncorrect ed
Bonferroni
Sidak
Holm
Holland
Hochberg
Simes
Yekutieli
0.827 0.199
1.000 1.000
1.000 0.944
1.000 1.000
0.970 0.703
0.838 0.838
0.838 0.287
1.000 0.913
0.600
1.000
1.000
1.000
0.993
0.938
0.712
1.000
0.582
1.000
1.000
1.000
0.993
0.938
0.712
1.000
0.005
0.089
0.085
0.075
0.072
0.075
0.022
0.079
0.696
1.000
1.000
1.000
0.993
0.938
0.735
1.000
0.088
1.000
0.826
0.792
0.563
0.792
0.152
0.539
0.673
1.000
1.000
1.000
0.993
0.938
0.735
1.000
0.226
1.000
0.992
1.000
0.833
0.938
0.330
1.000
0.111
1.000
0.893
0.888
0.610
0.888
0.176
0.623
84
Appendix F9: Comparison of uncorrected and corrected p-values using different multiple-comparison procedures: for aggregated and disaggregated average treatment effects on students’ subjective happiness P-VALUE CORRECTIONS Aggregated ATE Within-class feedback Across-class feedback Disaggregated ATE
Within-class feedback, no rewards Across-class feedback, no rewards Financial Rewards, no feedback Reputational Rewards, no feedback Within-class feedback with financial rewards Across-class feedback with financial rewards Within-class feedback with reputation rewards Across-class feedback with reputation rewards
Uncorrect ed
Bonferroni
Sidak
Holm
Holland
Hochberg
Simes
Yekutieli
0.036 0.341
0.468 1.000
0.379 0.996
0.304 1.000
0.266 0.876
0.288 0.984
0.078 0.493
0.248 1.000
0.169
1.000
0.970
1.000
0.773
0.979
0.268
0.949
0.115§
1.000
0.902
1.000
0.667
0.979
0.199
0.705
0.316 0.281
1.000 1.000
0.999
1.000
0.998
1.000
0.874 0.874
0.979 0.979
0.399 0.381
1.000 1.000
0.005
0.089
0.086
0.071
0.069
0.071
0.018
0.064
0.517
1.000
0.999
1.000
0.946
0.979
0.614
1.000
0.054
1.000
0.654
0.706
0.516
0.706
0.140
0.497
0.059
1.000
0.684
0.708
85
0.518
0.708
0.140
0.497
