Borrowing Constraints and Demand for Remedial Education: Evidence from Tanzania

Abstract

We use a cash transfer to relax households’ borrowing constraints, then elicit their willingness to pay for a remedial education programme offering tutoring and life skills training. Lottery losers were willing to pay 3,300 Tanzanian Shillings for the programme, which is 7% of per capita monthly expenditures. For those identified at baseline as able to borrow, willingness to pay increases by 3% upon winning a lottery prize of 3,200 Tanzanian Shillings. For those unable to borrow, willingness to pay increases by 27% upon winning the lottery. We conclude that borrowing constraints limit access to educational programmes, and may increase inequality of educational attainment.

A large body of recent research and policy debates have highlighted low levels of social mobility around the world (see, e.g., Chetty et al., 2014). Researchers and policymakers alike suspect a lack of equal access to education as a potentially important source of low social mobility. Currently, educational attainment strongly depends on the socio-economic status of the parents in many countries (Filmer and Pritchett, 1999) including many African countries (Azomahou and Yitbarek, 2016; Alesina et al., 2021). This holds also true for the setting of this study: Tanzania. The attendance rate in primary education is 76%, and falls to only 23% in secondary education (Ministry of Health, Community Development, Gender, Elderly and Children Dar es Salaam & Ministry of Health Zanzibar & National Bureau of Statistics Dar es Salaam & Office of Chief Government Statistician Zanzibar & ICF Rockville, Maryland USA, 2016). But the secondary school attendance rate is 41% among children from the top wealth quintile while it is only 6% among those in the lowest quintile. Given empirical estimates of returns to secondary education in Tanzania of around 15% per year (Montenegro and Patrinos, 2014) these figures suggest immense lost potential.

What keeps children from poorer backgrounds from achieving higher levels of education? One leading hypothesis is that poor households are borrowing constrained: despite high returns to education, they cannot raise the funds needed to pay for school or programme fees, or other complementary inputs. A competing hypothesis is that other correlates of poverty, such as the early childhood parental environment, affect children’s cognitive and non-cognitive abilities, and eventual schooling outcomes (Heckman and Carneiro, 2002). Evidence from developed countries suggests borrowing constraints may be of secondary importance for progression to higher education and better labour market outcomes (Heckman and Mosso, 2014). But they may be of primary importance in developing countries, where borrowing constraints are thought to limit the take-up of productive investment opportunities (e.g., Banerjee and Newman, 1993; De Mel et al., 2008; Beaman et al., 2023).

We report results from a randomised control trial designed to study how borrowing constraints affect families’ investment in a remedial education programme in Tanzania.¹ To that end we collaborated with a non-governmental organisation (NGO) that runs free study clubs for girls aged 12–14, corresponding to cohorts who should be attending the final two years of primary education, as part of its efforts to improve girls’ education outcomes. The NGO was interested in implementing a participation fee to support long-run programme sustainability, but was concerned about how this might affect access. We worked with them to elicit households’ willingness to pay (WTP) for the programme, a measure of each household’s demand for education. The experiment and surveys were designed to uncover how borrowing constraints depress demand for education.

The experiment took place in 69 villages that did not previously have a study club from our partner NGO. We conducted a baseline survey of a representative sample of eligible girls and their household head, which included two measures of borrowing constraints: a household-level dummy for ‘inability to borrow for an important expenditure’, and an index of four binary borrowing constraint measures. We then invited the girl and a responsible adult to a village meeting about girls’ education. We gave them a lottery ticket for a prize of 3,200 Tanzanian Shillings (TSh) (purchasing power parity (PPP) US$4.33), described as a ‘thank you’ for taking the survey, to be drawn during the meeting. This lottery is our experimental treatment of interest, the lottery prize acting as an unconditional cash transfer.

The village meetings started with the draw of the lottery to award prizes to 50% of eligible attendees. Subsequently, the programme officers explained the study clubs in detail. Last, we elicited WTP to join the club through a ‘multiple price list’ variant of the Becker–DeGroot–Marschak mechanism (Becker et al., 1964). Girls filled out the instrument along with the adult who had joined them, usually a household head, so we interpret the response as the household’s WTP.

The experiment and data allow us to (i) measure the household’s WTP for their daughter to participate in the study club, (ii) study the correlation between WTP and borrowing constraints, (iii) identify the causal effect of the cash transfer on WTP, and (iv) examine heterogeneity of the effect of the cash transfer with respect to borrowing constraints. We additionally examine long-run effects on programme participation.

What should we expect to see if borrowing constraints are depressing demand? Suppose household |$ i$| values the programme at |$ V_{i}$| units of consumption. After accounting for pressing spending needs their investable funds are |$ A_{i}(B_{i})$|⁠, where |$ B_{i}$| is credit available to them. It is natural to assume that |$ A^{\prime }\ge 0$|⁠: the more easily they can borrow, the less pressing will be their unmet spending needs and the more they can afford to invest in a new opportunity. As a result, the maximum amount they are willing to pay for the programme is |$ \textit{WTP}=\min \lbrace A_{i}(B_{i}),V_{i}\rbrace $|⁠.

Our cash transfer treatment can be thought of as increasing investable funds to |$ A_{i}(B_{i}) + T$|⁠, and thus WTP to |$ \min \lbrace A_{i}(B_{i})+T,V_{i}\rbrace $|⁠. So the treatment should have no effect on unconstrained households’ (with |$ A_{i}\ge V_{i}$|⁠) WTP, while those with |$ A_{i}\lt V_{i}$| will increase their WTP (the effect is intermediate for those lying close to the constraint). Thus, we expect the cash transfer will increase WTP for the programme, and this effect will be concentrated among the borrowing constrained.

We find five main results. First, on average households have substantial WTP for the remedial education programme. Non-lottery-winning households are willing to pay around 3,300 TSh (PPP US$4.47) on average, corresponding to 1.4% of total monthly household expenditure, or 7% of monthly per capita expenditure, and is approximately equal to the programme fee (3,000 TSh). There is notable heterogeneity: 9% of households were willing to pay 10,000 TSh, while 16% were not willing to pay anything.

Second, there is a negative association between WTP and two proxies for borrowing constraints: among lottery losers, those unable to borrow have approximately 500 TSh lower WTP compared to those who are able to borrow.

Third, winning the 3,200 TSh lottery prize increases WTP by around 300 TSh, or 9%, on average.

Fourth, the effect of the lottery treatment interacts strongly with borrowing constraints measured at baseline. The average effect of the lottery is almost entirely driven by borrowing constrained households. Those who report being able to borrow increase WTP by only 3% (roughly 120 TSh) when they win, while those who cannot borrow increase WTP by 27% (roughly 850 TSh). Put differently, while WTP is substantially lower for those unable to borrow amongst lottery losers, this association disappears for winners. The last finding suggests that the lottery was effective at relaxing borrowing constraints.

This interaction effect is robust to controlling for interactions with a host of observable characteristics that might be correlated with borrowing constraints and WTP, such as distance to school, girl’s cognitive skills, preferences, and household expenditures or poverty. This suggests that the heterogeneity we find is not proxying other girl- or household-level characteristics. We also discuss alternative explanations—including income, experimenter demand, ‘house money’, or anchoring effects—and argue that they are qualitatively and quantitatively implausible.

Fifth, due to implementation difficulties, the programme was launched with a delay and some changes, including elimination of fees (see Section 1.4). However, borrowing constraints remain strongly predictive of long-run programme take-up.

The collection of evidence suggests that borrowing constraints are an important impediment to demand for educational investments.

The paper is related to the literature on the role of borrowing constraints in suppressing profitable investments in general and education investments in particular. While the association between family income and schooling outcomes has been documented in a variety of contexts (e.g., Das et al., 2022), evidence on the role of borrowing constraints is mixed and is largely from developed countries (Heckman and Carneiro, 2002; Cameron and Taber, 2004; Dahl and Lochner, 2012; Caucutt and Lochner, 2020). In developing countries, a number of studies highlight the importance of prices and borrowing constraints for the take-up of health products (Kremer and Miguel, 2007; Hoffmann, 2009; Hoffmann et al., 2009; Ashraf et al., 2010; Cohen and Dupas, 2010; Dupas, 2014; Tarozzi et al., 2014; Fischer et al., 2019; Berry et al., 2020), insurance (Casaburi and Willis, 2018), and fuel-efficient stoves (Berkouwer and Dean, 2022). In the context of education there are a number of studies quantifying demand for education exploiting (downward) changes in school fees through vouchers (e.g., Angrist et al., 2006), scholarships (e.g., Kremer et al., 2009; Duflo et al., 2021) or fee abolition (e.g., Deininger, 2003; Riphahn, 2012; Bold et al., 2014).

Most notable amongst those studies is the work of Berry and Mukherjee (2019) as they also study out-of-school tutoring centres. They implement a two-part pricing design to induce random variation in the offer and final price. They find that higher offer prices select participants with higher attendance and higher final prices induce participants to drop out. But they do not find evidence that study centres increased test scores amongst the group of participants induced to sign up by a lower price. Their work takes the demand for educational investments as given. Our work complements their work in studying a potentially important source of low demand for education investments: borrowing constraints. In this sense, another closely related study is Dillon (2020), showing that a change in the timing of school fees forced farmers to sell crops early, forgoing profitable opportunities to store them. In a world without borrowing constraints, the timing of education expenditures should be irrelevant.

We also relate to the literature on cash transfers for education (see recent reviews by Baird et al., 2014; Bastagli et al., 2016). This literature mostly focuses on conditional cash transfers, which are linked to education take-up (e.g., Evans et al., 2023). Studies tend to focus on enrolment and participation in education, and typically find positive effects. Benhassine et al. (2015) study a cash transfer that is ‘labelled’ as being for education, and find it has similar effects on school participation to conditional transfers. In contrast, the lottery in our experiment is deliberately not labelled as being for education, but as compensation for participation in the survey, so is more closely related to unconditional cash transfer programs. As an example, Haushofer and Shapiro (2016) find in Kenya that large cash transfers increased monthly educational spending, but the effect is small relative to the size of the transfer ($1 PPP versus transfers of $404 or $1,525 PPP). We find larger impacts relative to the size of the transfer. This may be because our experiment provided liquidity shortly before the opportunity to make an educational investment. This is by design, to mimic the provision of liquidity through credit, and distinguishes our treatment from a typical unconditional cash transfer. Most importantly, our findings demonstrate strong heterogeneity: the lottery payment increases educational investments almost exclusively for households that our baseline survey identifies as unable to borrow, and for them it increases educational investments strongly.

1. Background and Methodology

1.1. The Program

We study demand for, and take-up of, a remedial education programme implemented by an NGO in Tanzania.² The central aim of the programme is to improve learning outcomes of girls at risk of dropping out of school, or who have recently dropped out, and to increase enrolment rates. As part of this effort, the NGO established study clubs, designed to provide subject-based tutoring in mathematics and English. According to the programme design, tuition for in-school girls who are in their final two years of primary education was scheduled to take place in the afternoon hours, three times a week, for three hours. The tutoring for out-of-school girls was to take place in the mornings, five days a week, for three hours. In addition, the NGO would register the out-of-school girls under the Institute of Adult Education, enabling them to complete their Form 1 and 2 courses. The tutoring was facilitated by trained teachers who were paid an honorarium for their work. The tutoring followed the primary education curriculum and is intended to prepare pupils for the Tanzanian secondary school entrance exams.³

The clubs were established inside villages to make them easily accessible. In the afternoon hours, the clubs were then used as safe spaces for both in-school and out-of-school girls to come together, interact, forge bonds and support each other in their studies. In addition to subject-based tutoring, the clubs provided life skills training through peer mentoring.

1.2. Sample Selection

Here we summarise how we arrive at our analysis sample; see Online Appendix C.1 for details.

The NGO implemented the remedial education programme at 20 branches in the regions of Dar es Salaam, Mwanza, Shinyanga, Tabora and Singida. In September 2013 we randomly selected eight study branches. Within those branches the NGO’s field staff identified 105 villages close to potential treatment schools.⁴

Villages were grouped in 42 ‘clusters’, with villages close to the same school assigned to the same cluster. Of those, we randomly selected 27 clusters containing 69 villages as study locations. The remaining villages were to be control villages for the purpose of programme evaluation, and not relevant to this paper.

We randomised study villages, stratified by branch, to receive a free club (36 villages) or a club with a one-time joining fee of 3,000 TSh (33 villages).⁵

We conducted a short census of girls aged 11 to 18 in the villages in November 2013.

The census served a sampling frame for the baseline survey. Girls in the census were screened for programme eligibility.⁶ The census sample consisted of 5,968 girls, of whom 5,048 were eligible for programme participation.

The baseline survey was conducted in December 2013. The main respondent was the selected girl, followed by a short module addressed to the household head. We aimed to sample only one girl per household, except where the number of available girls was small. In addition, due to challenges finding participants we allowed for some convenience sampling (53 girls). Our full baseline sample contains 1,717 girls.

At the end of the baseline survey, girls received a lottery ticket for a prize of 3,200 TSh if they came to an information meeting about the new education programme, and they were told that half of eligible (i.e., baseline-surveyed) attendees would win. This lottery was framed as a thank you for taking the survey. The lottery is the treatment of interest in this paper.

1.3. WTP Elicitation

The information meetings were organised in June 2014. Online Appendix C.4 contains the meeting script. All baseline girls were invited to attend, as well as any other girls living in the village. They were to be accompanied by a household member, ideally the household head. The meeting was described as an information session about the new education programme. Of the 1,717 girls in the baseline, 880 attended a meeting, plus 252 non-baseline girls (who are not included in the analysis).

First, we conducted the lottery for baseline participants (whether or not they remembered their ticket). Prizes were to be awarded through a public draw at which 50% (rounding up) of tickets would win. Winners were told that they were free to do whatever they wanted with the money.

Afterwards the programme officers described the study clubs in detail. Information about the programme was provided in exactly the same way to all participants. It was emphasised that to join the club, girls needed to sign up on the day of the meeting, and that any fee charged for participation was due at the first club meeting, scheduled to take place around one week later.

Last, we elicited WTP to join the club. Participants were told that joining the club might be free or might require a one-time fee. The price had already been decided and it was written inside an envelope that was shown to the audience, but they were not told about the price distribution. Before the envelope was opened, participating girls along with the accompanying household head needed to declare their maximum WTP. They were provided with a sheet of paper listing eleven prices uniformly spaced from 0 to 10,000 TSh.⁷ They were told to tick ‘Yes’ next to each price they would be willing and able to pay to participate in the club, and to tick ‘No’ for prices that they were not willing/able to pay (this could include that they would join the club only if it was free). If the price in the envelope was equal to or below their WTP, they would be required to join the club and pay the fee at the first meeting of the study club. Those whose WTP was below the price would pay nothing and receive nothing. For expected utility maximisers, bidding up to one’s true maximum WTP is a weakly dominant strategy.

Our elicitation mechanism is a ‘multiple price list’ variant of the Becker–DeGroot–Marschak mechanism (Becker et al., 1964; Andersen et al., 2006). This implementation helps the participants by breaking down the mechanism into simple take-it-or-leave-it questions. They were reminded that they could not influence the price, and our procedure—where the price was already determined, but not revealed—made this very clear.⁸

We began with a few comprehension questions, then a practice exercise, selling bars of soap using the multiple price list procedure. If their WTP was higher than the price in the envelope (400 TSh in all villages), they were required to buy the soap.⁹ Participants were instructed not to decline one price, then accept a higher one, on the WTP sheets (i.e., ‘multiple switching’).

After the soap exercise, we elicited WTP to join the study club. After everyone reported their WTP, the answer sheets were collected and the price inside the envelope was revealed (either 0 or 3,000 TSh). Everyone willing to pay at least as much as the price in the envelope was asked to sign a ‘contract’ promising to join the club and to pay the price at the first club meeting in around one week’s time.

We have WTP data for 825 of the 880 baseline girls that attended. We infer that the 55 for whom we do not have data chose not to participate in the elicitation. This could be because they were unwilling to participate even at zero price. Our results are robust to assigning zero WTP to these girls.

1.4. Implementation Challenges

Lottery winners were not recorded by the enumerators in four villages. This leaves us with 65 villages and 805 girls for whom we have WTP and lottery data. The lottery was not perfectly implemented in every village, but we find no evidence that this contaminated the randomisation, see Online Appendix C.2 for details.

After completion of the WTP meetings, the NGO experienced unanticipated difficulties with the programme launch, which was delayed by several months. When it was launched, an altered version of the programme was rolled out, with tutoring delivered in schools rather than in clubs (which instead focused on life skills, providing a social space, etc). Presumably due to the delay, there were significant difficulties in collecting fees, so de facto the programme became free. This is not an issue for our WTP analysis as the WTP elicitation was incentive compatible as the delays were unanticipated.

A follow-up survey was conducted two years after the baseline (18 months after WTP elicitation). We use two variables from this survey to capture programme participation: (i) a binary measure of whether the girl ever attended the club, and (ii) her frequency of attendance (days per week).

1.5. Borrowing Constraints Survey Measures

Our main treatment variable of interest is the lottery treatment, which we interpret as alleviating the effects of borrowing constraints among treated households. To explore how this treatment interacts with borrowing constraints, we construct survey-based measures of borrowing constraints.

In the survey, both girls and household heads were asked separately: ‘If you needed to borrow money for an important expenditure (e.g., health or school related expenditure), how easy would it be for you to borrow the money?’ Options were ‘easy’, ‘not easy, but possible’, and ‘not possible’. If the respondent said ‘don’t know’ we code them as missing. This gives us two dummy variables for the girl and two for the household, defined as not possible, and not possible OR not easy. Our first survey measure of borrowing constraints is the dummy for borrowing ‘not possible’ according to the household head. Our second measure is a standardised (to mean zero, standard deviation one) index of the four dummy variables, which we refer to as the borrowing constraints index.¹⁰ In the terminology in the introduction, a higher constraints measure is interpreted as a lower value of investable funds, |$ A_{i}$|⁠.

1.6. Balance Checks

In the Online Appendix, we perform a sequence of balance checks capturing each stage of the selection process outlined above. Importantly, girls who attended the WTP meeting were remarkably similar to the general population of baseline girls on a wide range of covariates. Standardised differences in covariates between winners and losers are small and quantitatively unlikely to drive any of our main results.

2. Results

2.1. Estimation

To identify the effects of winning the lottery on the demand for the programme, we estimate an ordinary least squares (OLS) regression of the form:

$$\begin{eqnarray} \text{{Y}}_{ihv} = \beta \cdot \text{Lottery}_{i}+ \sum _{j=1}^{65}\gamma _{j} {1\!\!1}(v=j) +\varepsilon _{ihv}, \end{eqnarray}$$

(1)

where |$\text{{Y}}_{ihv}$| is an outcome for girl i from household h in village v. |$\text{Lottery}_{i}$| is a dummy variable equal to 1 if the girl won the lottery, and |$\gamma _{j}$| are village fixed effects for the 65 villages in which we have lottery data. The parameter of interest is |$\beta$|⁠, the average effect of winning the lottery.

To examine how the lottery treatment interacts with borrowing constraints, we estimate:

$$\begin{eqnarray} \text{{Y}}_{ihv} = \beta \cdot \text{Lottery}_{i} + \lambda \cdot \text{Constraint}_{i} + \delta \cdot \text{Lottery}_{i} \cdot \text{Constraint}_{i} + \sum _{j=1}^{65}\gamma _{j} {1\!\!1}(v=j) +\varepsilon _{ihv}, \end{eqnarray}$$

(2)

where |$\text{Constraint}_{i}$| is a measure of borrowing constraints. In this specification, |$\beta$| identifies the treatment effect when |$ \text{Constraint}_{i}$| is zero. For our binary measure these are households who can borrow (either easily or with some difficulty). In our case these are households at the index mean. |$\lambda$| identifies the relationship between borrowing constraints and |$ Y$|⁠, for those who lost the lottery. |$\delta$| identifies the interaction effect between constraints and lottery win.

Although the lottery treatment was assigned by randomisation at the girl level, we have some households with multiple participating girls.¹¹ Intra-household decisions about different girls may be interrelated, and borrowing constraints are partially defined at the household level.¹² In the spirit of clustering at the level of assignment (Abadie et al., 2022), we cluster standard errors at the household level. We report estimates clustered at the village level in Online Appendix Table B6. We also report randomisation inference |$p$|-values for the randomised treatment effect and its interaction with borrowing constraints (Imbens and Rubin, 2015; Young, 2019).

2.2. Demand for Education and Borrowing Constraints

In our full sample, all households would sign up if the programme were offered for free. But fees significantly affect demand: 16% of households were not willing to pay more than zero. Roughly 50% were willing to pay the true (not-yet-announced) programme fee of 3,000 TSh. Less than 20% were willing to pay more than 5,000 TSh.

Figure 1 displays the demand curve for the programme splitting the sample according to whether the household head reported at baseline that they cannot ‘borrow money for an important expenditure’. In both subsamples lottery winners have higher WTP, indicated by a first-order shift to the right of the demand curve. While both subsamples show some response to the lottery, the shift is more pronounced for the borrowing constrained subsample. In particular this group shows a large increase in the share of households willing to pay high prices (5,000 TSh or more).¹³

Fig. 1.

Demand Curves and Borrowing Constraints.

Notes: These figures present demand curves for the remedial education programme, separately for four subgroups: in panel (a) we present results for the subsample of households whose household head responded that they would not be able to ‘borrow money for an important expenditure’, in panel (b) we present results for the subsample of households whose household head stated that this would be possible; in both graphs we present results separately for households who won and those who did not win the lottery. The full sample corresponds to the sample used in columns 2 and 3 of Table 1, Panel A.

Table 1 presents the regression equivalents. Panel A, column 1 displays estimates of (1) in the full sample for which we have WTP and lottery data. Lottery losers were willing to pay 3,335 TSh on average, around 7% of monthly per capita expenditures. Winning the lottery increases WTP by 311 TSh, or 9.3% (⁠|$p$|-value = 0.038) on average.

Table 1.

Demand for Education and Borrowing Constraints.

	WTP			Participation	Attendance
	(TSh)			(0/1)	(per week)
	(1)	(2)	(3)	(4)	(5)
Panel A: borrowing constraints dummy
Lottery win	311	362	119	0.008	0.085
	(152)	(175)	(210)	(0.065)	(0.207)
	[0.042]	[0.043]	[0.605]	[0.953]	[0.698]
Cannot borrow			−522	−0.160	−0.282
			(282)	(0.065)	(0.265)
			[0.065]	[0.015]	[0.289]
Cannot borrow |$\times$| lottery win			734	0.026	0.014
			(379)	(0.103)	(0.376)
			[0.047]	[0.669]	[0.924]
Village FE	Yes	Yes	Yes	Yes	Yes
Mean outcome (C)	3,335	3,414	3,414	0.183	0.425
Observations	805	642	642	269	272
R|$^2$|	0.474	0.493	0.497	0.429	0.316
Panel B: borrowing constraints index
Lottery win	311	297	312	0.016	0.050
	(152)	(160)	(159)	(0.041)	(0.125)
	[0.042]	[0.067]	[0.051]	[0.695]	[0.729]
Borrowing const. index			−387	−0.047	−0.237
			(126)	(0.035)	(0.128)
			[0.002]	[0.174]	[0.065]
Borrowing const. index |$\times$| lottery win			432	−0.010	0.034
			(168)	(0.045)	(0.146)
			[0.009]	[0.721]	[0.886]
Village FE	Yes	Yes	Yes	Yes	Yes
Mean outcome (C)	3,335	3,378	3,378	0.177	0.422
Observations	805	736	736	316	319
R|$^2$|	0.474	0.482	0.490	0.360	0.274

Notes: The table reports ordinary least squares estimates based on equations (1) and (2). The dependent variable in columns 1–3 is the household’s WTP (in TSh) for the remedial education programme, in column 4 it is a dummy for whether the girl ever attended the programme’s club, and in column 5 the number of times per week she visited. Column 2 shows results from the specification of column 1, but in the sample of column 3. Lottery win indicates whether the individual has been randomly assigned to receive a lottery payout. Cannot borrow is a dummy variable indicating if the household head reported that it would not be possible for them to borrow money for an important expenditure. Borrowing constraints index is an index combining four dummy variables indicating if the respondents (girl or the household head) states that it would not be possible or it would be anything, but easy to borrow money for an important expenditure. We calculate the index by first normalising each indicator by subtracting the sample mean and dividing by its standard deviation; then taking the average of the four normalised indicators, and normalising again. If only some of these dummies are available we impute the missing ones at the sample mean. All regressions include village fixed effects. Standard errors are clustered at the household level and given in parentheses. In square brackets |$p$|-values of the null hypothesis of no effect are provided. For the main effect of Lottery win and interactions with Lottery win these are calculated as randomisation inference |$p$|-values, for all other coefficients they are calculated analytically based on the reported clustered standard errors. The randomisation |$p$|-values are the percentile of the coefficient estimated under the true assignment in the distribution of coefficients estimated under 10,000 alternative assignments. Mean WTP among all lottery losers, and the number of observations, are reported at the bottom of the table. Mean WTP among lottery losers who ‘can borrow’ is 3,633 TSh. Mean participation among lottery losers who ‘can borrow’ is 26.2%.

Column 2 reports similar estimates for the subsample for whom we observe the dummy indicating the household cannot borrow (missing if they responded ‘don’t know’).

Column 3 estimates heterogenous effects of the lottery in (2). For households who can borrow, average WTP among lottery losers is 3,633 TSh, increasing by only 119 TSh or 3% when they win the lottery. Households who cannot borrow have initial WTP of |$ 3,\!633-522=3,\!111$| TSh, and are substantially more responsive to the lottery, increasing WTP by |$ 119+734=853$| TSh, or 27%, when they win. Thus the difference in WTP between constrained and unconstrained households is smaller among winners than losers.

Panel B uses the index measure of borrowing constraints, and finds very similar results to Panel A.¹⁴ A one standard deviation increase in borrowing constraints is associated with 387 TSh lower WTP among lottery losers, while the effect of winning the lottery increases by 432 TSh. The fact that the two coefficients approximately cancel one another implies once again that WTP of lottery winners is relatively insensitive to whether they are borrowing constrained.

Columns 4 and 5 of both panels report effects on long-run club participation and attendance, respectively.¹⁵ Notably, borrowing constraints are negatively associated with both, implying that just as borrowing constraints make club fees difficult to finance, they presumably make costly investments in remedial education (in terms of time and other resources) more difficult in general. Some 26% of lottery losers who ‘can borrow’ attended the programme, falling by 16 percentage points for those who cannot.¹⁶ Directionally, the coefficients on our lottery treatment and its interaction with borrowing constraints are almost always the same as in the main analysis (positive main effects in all specifications and positive interactions in three out of four). These coefficients are relatively small and never statistically significant. That is not surprising since the lottery treatment was designed and only expected to act as a short-run relaxation of borrowing constraints.

2.3. Robustness

We think there are five possible alternative interpretations of our results. First, our borrowing constraints measures might capture factors other than borrowing constraints. Second, our treatment effect might be an income effect rather than a constraint relaxation. Third, we address possible behavioural confounds. Fourth, we discuss the impact of the information session on the estimated effects. And, fifth, we discuss the role of households anticipating the option to default on the contract.

First, our borrowing constraints measures may be capturing some other underlying differences in girls’ or households’ characteristics. To address this concern, we assess robustness of our estimates to controls. In particular, we estimate (2), controlling for baseline covariates and their interaction with |$\text{Lottery}_{i}$|⁠. We include a wide range of covariates capturing education access and attainment (access to tutoring, cognitive skills, distance to school, perceived returns to secondary education), gender attitudes that might affect girls’ schooling, preferences (risk aversion and patience), health and household structure. A particular concern is that our constraint measures might simply reflect poverty, so we include measures of per capita expenditures and poverty.

Table 2 presents the results based on the index measure of borrowing constraints (Online Appendix Table B7 uses the binary measure). Each row reports results of a separate regression, one for each baseline covariate.

Table 2.

Demand for Education and Borrowing Constraints Index: Robustness to Controls.

Variable	Lottery	Constraint	Lottery |$\times$| Constraint	Covariate	Lottery |$\times$| Covariate	N
Tutoring	297	−399	421	−68	42	726
	(161)	(126)	(168)	(127)	(167)
	[0.076]	[0.002]	[0.011]	[0.592]	[0.876]
Cognitive skills	316	−389	428	125	−164	736
	(159)	(125)	(168)	(110)	(162)
	[0.048]	[0.002]	[0.010]	[0.254]	[0.343]
Distance to school	312	−387	435	63	−8	736
	(159)	(126)	(169)	(96)	(156)
	[0.050]	[0.002]	[0.008]	[0.514]	[0.814]
Returns second. ed.	310	−445	492	281	−350	664
	(174)	(128)	(177)	(129)	(175)
	[0.067]	[0.001]	[0.005]	[0.030]	[0.040]
Gender attitude	284	−380	447	73	80	716
	(162)	(127)	(169)	(114)	(161)
	[0.083]	[0.003]	[0.005]	[0.518]	[0.479]
Risk aversion	307	−383	463	−205	235	703
	(165)	(130)	(172)	(132)	(168)
	[0.070]	[0.003]	[0.009]	[0.122]	[0.104]
Patience	281	−386	409	162	−164	684
	(171)	(133)	(181)	(135)	(174)
	[0.101]	[0.004]	[0.020]	[0.231]	[0.506]
Illness	310	−384	417	31	72	727
	(162)	(128)	(171)	(126)	(163)
	[0.062]	[0.003]	[0.012]	[0.807]	[0.507]
HH kids (no.)	316	−399	455	−109	80	726
	(162)	(126)	(168)	(127)	(164)
	[0.052]	[0.002]	[0.007]	[0.390]	[0.645]
HH kids (f-share)	388	−418	512	68	−34	689
	(167)	(129)	(172)	(125)	(168)
	[0.023]	[0.001]	[0.004]	[0.588]	[0.841]
Per capita expenditure (TSh)	352	−412	439	−192	277	665
	(175)	(133)	(178)	(175)	(196)
	[0.043]	[0.002]	[0.013]	[0.275]	[0.302]
Poverty (⁠|$\lt $|2 US$/day)	352	−418	442	106	−114	665
	(175)	(135)	(178)	(135)	(176)
	[0.044]	[0.002]	[0.013]	[0.435]	[0.541]

Notes: The table reports ordinary least squares estimates based on specification (2). Lottery indicates whether the individual has been randomly assigned to receive a lottery payout. Constraint is an index increasing in borrowing constraints (see Table 1's notes for further details). Tutoring is a dummy variable indicating if the girl attended any tutoring or study group during the past year. Cognitive skills is a normalised index combining the girl’s score in a math exam (EGMA), a reading exam (EGRA) and a Raven’s test. Distance to school is the shortest time (in minutes) it takes to reach school. Gender attitude is based on the girl’s responses to questions capturing various gender roles in the family (e.g., ‘Who should earn money for the family?’). It is the fraction of questions (out of 7) to which the girl responded with gender-neutral roles. Risk aversion is the girl’s response to the question ‘On a scale from 0 (not at all willing to take risks) to 10 (very willing to take risks), which number do you give yourself?’ inverted. Patience is the girl’s response to the question ‘On a scale from 0 (very patient) to 10 (very impatient), which number do you give yourself?’ inverted. Illness is a dummy variable indicating if the girl reported having had any serious illness in the last year. HH kids (no.) is the number of household members younger than 20. HH kids (f share) is the percentage of females among household members younger than 20. Per capita expenditure is the monthly household consumption (in Tanzanian Shillings) as reported by the household head, divided by the number of people living in the household. Poverty (<2 US$/day) is a dummy variable indicating if the per capita daily expenditure is less than US$2 PPP. See Table B10 in the Online Appendix for further details on the covariates. The covariate variables have been standardised. All regressions include village fixed effects. Standard errors are clustered at the household level and given in parentheses. In square brackets |$p$|-values of the null hypothesis of no effect are provided. For the main effect of Lottery and interactions with Lottery these are calculated as randomisation inference |$p$|-values, for all other coefficients they are calculated analytically based on the reported clustered standard errors. The randomisation |$p$|-values are the percentile of the coefficient estimated under the true assignment in the distribution of coefficients estimated under 10,000 alternative assignments.

The coefficient estimates for |$\beta$|⁠, |$\lambda$| and |$\delta$| are highly robust in magnitude and precision. Moreover, the additional covariates and interactions mostly have small, non-significant coefficients. We conclude that our results are unlikely to be driven by omitted variable bias.¹⁷

Additionally, Online Appendix Table B6 shows that the results are robust to alternative fixed effects (branch, enumerator) and clustering (at the village level).

Second, an income effect interpretation of our findings says that winning the lottery increased household wealth, that education is a normal good, and so WTP for the programme increased accordingly. In the terminology from the introduction, |$ V_{i}$| might be increasing in |$ A_{i}$|⁠. We do not have a wealth-preserving borrowing treatment to fully rule this out, but we can assess its quantitative plausibility. On average, winning the lottery increased WTP by 9%. The 3,200 TSh prize is around 1.4% of total household monthly expenditures. Assuming the whole amount is spent within a month, this gives an implied ‘elasticity’ of 6. Among the constrained group WTP increases 27%, an elasticity of 19. While we do not have a clear benchmark with which to compare this (it is not a traditional income elasticity because the shock is not an income shock and the expenditure is a one-time expense), income effects of this magnitude seem unlikely. It is also comforting that per capita expenditure is not associated with higher WTP; see Table 2.

Third, certain behavioural mechanisms could increase the WTP of lottery winners. For example, an experimenter demand interpretation posits that winners believed they were expected to pay more, and did so out of reciprocity or perceived social pressure. (Other behavioural channels could include ‘house money’, ‘mental accounting’, or ‘anchoring’ effects.) To mitigate these concerns, we separated the lottery from the programme by framing it as a ‘thank you’ for participating in the already-completed baseline survey, and by telling participants they were ‘free to do whatever you like with this money’. We also spent time between the lottery draw and the WTP elicitation, explaining the study clubs.¹⁸ A particular version of this concern is that participants perceived a connection between the credit constraints question and the lottery treatment. Because the question was just one of many in a large baseline survey, and because of the temporal separation, we think this is unlikely.

To the extent that personality traits correlate with such behavioural responses, the results in Table 2 are helpful, as both risk aversion and patience are not strongly predictive of WTP. More importantly, these mechanisms are unlikely to explain the interaction that we observe between our treatment and the borrowing constraints measures. It would have to be that more constrained households are more sensitive to these behavioural mechanisms. However, the inclusion of the personality traits and their interaction with the lottery win never changes the magnitude of our main interaction effect in a meaningful way.¹⁹

Fourth, the information session surely impacted WTP. The same information was provided to all study participants, but baseline knowledge, and hence belief updating, might have been heterogenous and correlated with borrowing constraints. Indeed, the results in Table 2 indicate that households with more experience of tutoring services, or more optimistic beliefs about returns to higher education, had higher WTP. Both variables might proxy for differences in information about remedial education at baseline. The fact that their inclusion does not affect our results suggests that differences in information are not driving the findings.

Fifth, while we described the WTP choice as a firm commitment, bound by a ‘contract’, participants might not have viewed it as such. This could cause them to overbid, expecting to have the option to renege later. However, as with the other threats we do not believe this would generate our interaction effect of interest. If anything it could go the other way: limited liability protects households from being forced to pay a bill they cannot afford, so the cash transfer could actually decrease constrained households’ propensity to overbid. As described in Section 1.4, due to unanticipated programme issues the fees were ultimately not implemented, so we do not know if participants would have honoured their commitments. Maffioli et al. (2023) conduct a careful review of the WTP literature, and report a wide range of non-payment rates. Default appears to be more frequent for instalment plans; ours was intended to be a single up-front payment.

3. Conclusion

Despite improvements in access to primary education, learning achievements remain low in many developing countries, particularly for children from lower socio-economic backgrounds. Remedial education programs have emerged as one possible way to ameliorate inequalities in educational attainment.

We study the role of borrowing constraints in determining families’ willingness to pay for a remedial education programme in Tanzania. Through a lottery, we distribute cash prizes that exogenously relax some households’ constraints. Households are willing to pay 7% of average monthly per capita expenditure for their daughters to participate in the programme. Winning 3,200 TSh in a lottery increases willingness to pay by approximately 9%. This effect is almost entirely driven by those households our survey identifies as borrowing constrained, whose WTP is depressed absent the lottery, and who increase their willingness to pay by 27% when they win the lottery. It is robust to controlling for a host of observable correlates.

We conclude that borrowing constraints play a significant role in shaping demand for educational programs: households with the ability to borrow value those opportunities; borrowing constrained households also value them, and in fact value them similarly to unconstrained households, but are not in a position to explore those educational investment opportunities. To the extent that borrowing constraints are correlated with socio-economic status, these results suggest that they are likely to propagate inequality across generations.

Additional Supporting Information may be found in the online version of this article:

Online Appendix

Replication Package

Notes

The data and codes for this paper are available on the Journal repository. They were checked for their ability to reproduce the results presented in the paper. The replication package for this paper is available at the following address: https://doi.org/10.5281/zenodo.10887743.

We thank DFID for funding under the Girls’ Education Challenge (GEC) programme and BRAC for supporting and enabling this research. Sara Banfi, Mattia Chiapello, Francesca Larosa, Andrea Giglio, and in particular Camilla Fabbri and Lisette Swart provided outstanding research assistance. We thank Tessa Bold, Francesco Loiacono, Mireia Raluy i Callado, and Abhijeet Singh, and seminar audiences at EBRD, University of Milan-Bicocca, Trinity College Dublin, the IZA and CESifo education conferences, NEUDC, and the IIES Speed Brown Bag and Development Tea, for valuable comments. The experiment received ethical clearance from the Bocconi University Ethics committee and is registered at the AEA RCT registry, https://doi.org/10.1257/rct.5695-1.0. All remaining errors are our own.

Footnotes

1

Remedial education programmes targeting vulnerable children have emerged as a possible way to ameliorate inequalities in educational attainment; and are being implemented by many governments and NGOs around the world (e.g., Banerjee et al., 2007 ; 2016; Muralidharan et al., 2019). However, they need to be funded, and attempts to charge fees might jeopardise this objective in settings where ability or willingness to pay is low. Thus, even effective interventions may be discontinued in the absence of continued donor funding. For example, the highly effective MindSpark centres in India were forced to close down as insufficiently many families would or could pay the subsidised price (Muralidharan et al., 2019).

2

The NGO we collaborated with is BRAC Tanzania.

3

The programme is similar to remedial education programs aiming to provide a pathway for out-of-school youth to return to formal education. In the East African context, these programs are often provided by NGOs, labelled as ‘second-chance’, ‘bridging’ or ‘re-entry’ programs (Ngware et al., 2018). As an alternative, wealthier families can hire private tutors to assist youth prepare for the secondary school entrance exams.

4

For simplicity, we always refer to the communities in which clubs are situated as ‘villages’, though in peri-urban areas a better descriptor would be ‘neighbourhood’. Village boundaries were defined by NGO staff.

5

Note that this one-time fee is similar to an annual membership fee, since a given cohort of girls are expected to participate for one year and then take the secondary school entrance exams (i.e., ‘graduate’ from the club). This type of fee structure is common in clubs in our context. For example, BRAC’s Empowerment and Livelihood for Adolescents (ELA) programme evaluated by Bandiera et al. (2020) in Uganda also has a similar fee, although in that case the fee is voluntary and often waived. Non-formal education programs also often charge such one-off fees as a ‘community contribution’.

6

Eligibility required the girl either (i) had dropped out of school within the last two years (satisfied by 15% of girls), (ii) was at risk of dropping out (a grade of less than 50% in mathematics, science, or English in the last exam) (33%), (iii) belongs to a poor household, based on a poverty scorecard for Tanzania, developed by Grameen foundation (61%), (iv) has lost one or both parents (29%), or (v) displays signs of physical or mental disability (3%).

7

See Online Appendix Figure C2 for the answer sheet, and Figure C3 for its English translation.

8

Burchardi et al. (2021) test for optimal bidding under four WTP elicitation mechanisms, including one very similar to ours, with rural participants in Uganda. They find high rates of optimal bidding, averaging 86%, in all four.

9

We did not collect soap WTP data so cannot examine behaviour in this practice round. Lottery winners may have put some of their winnings into the soap purchase. If so, this would tend to attenuate treatment effects on WTP for the programme.

10

If some of the index components are missing we impute them with sample means, if all are missing we code the index as missing.

11

Our analysis sample of 805 girls contains 779 households: 755 with one girl, 22 with two, and 2 with three.

12

One measure, the household ‘cannot borrow’ dummy is fixed within household, while the index which depends also on the girl module can in principle vary within household.

13

The simple theory presented in the introduction predicts that WTP can increase by at most the size of the cash transfer, |$ T$|⁠. The large shift in the demand curve at high values of WTP in Figure 1(a) indicates that for some people WTP may have increased by more than 3,200 TSh. This could reflect sampling variation, or a non-linear response not captured by our model.

14

Because the index is standardised, the coefficient on lottery win has a different interpretation in Panels A and B. In Panel A it is the effect for participants who ‘can borrow’ while in Panel B it is for those at the mean of the index.

15

The sample size is smaller than in the other columns, which we attribute to programme and survey delays, and difficulty matching respondents across surveys. Reassuringly, Online Appendix Table B9 shows that our main results replicate when we restrict to the sample with follow-up data.

16

The overall drop in demand relative to when we elicited WTP is most likely due to changes in life circumstances, and lower WTP due to delays and changes to the programme. For sake of illustration, if we assumed everyone experienced a drop in WTP of around 5,000 Tsh we would predict take-up very similar to what we observe, see Online Appendix Figure A1.

17

We also probe the robustness of our results to the collection of channels tested in Table 2 by (i) including the first principal component of those variables and its interaction with the lottery variable and (ii) implementing the partialling-out LASSO estimator of Chernozhukov et al. (2018). The results are presented in Online Appendix Table B8. Both exercises leave our main effects essentially unchanged, qualitatively and quantitatively.

18

These are consistent with recommended practices in the experimental literature (Zizzo, 2010; De Quidt et al., 2019). De Quidt et al. (2018) and Mummolo and Peterson (2018) subsequently developed new techniques to measure demand effects, finding they are modest in behavioural experiments conducted online. While the setting is very different, this gives further cause for optimism.

19

As suggested by a referee we also investigated a couple of other personality-like measures: confidence, and self-perceived communication and persuasive abilities; our findings are robust to controlling for these as well (results available on request).