Motivating Knowledge Agents: Can Incentive Pay Overcome Social Distance? Purchased

Economists tend to believe in the power of incentives and prices to improve efficiency, whether the aim is to motivate workers or eliminate social ills such as discrimination.¹ Yet both theory and evidence suggest that there are circumstances in which there are grounds for caution: First, if there are multiple tasks or output is hard to measure, financial incentives may have undesirable consequences (Holmstrom and Milgrom, 1991; Gneezy et al., 2011). Second, in jobs with an aspect of social service, as in public goods provision, or if reputation matters, workers may not be ‘in it just for the money’. It has been argued that financial incentives may interfere with or even ‘crowd out’ such intrinsic motivation (Gneezy and Rustichini, 2000,b; Bénabou and Tirole, 2006). Third, theory suggests that aligning the identities of economic agents can increase efficiency (Francois, 2000; Besley and Ghatak, 2005), and there is evidence that ethnic fragmentation and ‘social distance’ can lead to worse economic outcomes (Easterly and Levine, 1997). Akerlof and Kranton (2005) argue that when group identity is salient, monetary incentives can be ‘both costly and ineffective’.

However, evidence on the effect of incentive pay on performance in pro-social tasks is still limited. Ashraf et al. (2014) find that both non-financial and financial rewards have stronger effects for socially motivated agents. Dal Bó et al. (2013) conclude that higher wages do not have adverse selection effects in terms of public service motivation. Rasul and Rogger (2016) suggest that the use of incentives can negatively affect aspects of performance in the Nigerian Civil Service.²

Moreover, very little is known about the interaction of incentive pay and social distance, although the literature on discrimination has suggested that competitive markets can remove the effects of social distance where these cause inefficiencies (Lang and Lehmann, 2012).

In this article, we develop a theoretical model and provide empirical evidence on the role of incentive pay in spreading information about a public service in a socially heterogeneous population. We study whether incentive pay is effective in settings where output is noisy and crowding out is a possibility, and whether incentive pay ameliorates or exacerbates the potentially detrimental effects of social distance.

A simple theoretical framework is developed which combines elements of a motivated agent framework (Besley and Ghatak, 2005) with the multi-tasking model (Holmstrom and Milgrom, 1991). The framework predicts that when there is a single task and the agent is intrinsically motivated, effort is always weakly increasing in the part of the agent's compensation that is dependent on success (the ‘bonus’). But when there are two tasks, which differ in terms of the agent's intrinsic motivation to succeed and in the marginal cost of effort, the effect of bonus pay will depend in part on the degree of substitutability in the cost of effort across the two tasks. If substitutability is low, increasing bonus pay will lead to an increase in the agent's effort with respect to both tasks. But if the two tasks are relatively substitutable in the cost function, an increase in bonus may cause effort in one task to decrease while effort in the other increases. This can be interpreted as incentive pay ‘crowding out’ intrinsic motivation for one of the tasks.

We then analyse data from a field experiment conducted across 151 villages in Karnataka, India, in the context of a government-subsidised health insurance scheme aimed at the rural poor. In a random sub-sample of the villages (the treatment groups), one local woman per village was recruited to spread information about the scheme. These ‘knowledge agents’ were randomly assigned to either a flat-pay or an incentive-pay contract. Under the latter contract, the agents' pay depended on how a random sample of eligible households in their village performed when surveyed and orally presented with a knowledge test about the scheme.

Our main empirical findings are as follows: first, hiring agents to spread information has a positive impact on the level of knowledge about the programme. The effect is driven by agents on incentive-pay contracts. Households in villages assigned an incentive-pay agent score on average 0.25 standard deviations higher on the knowledge test than those in the control group, and are also 8 percentage points more likely to enrol.

Second, social distance between agent and beneficiary has a negative impact on knowledge transmission. But putting agents on incentive-pay contracts appears to increase knowledge transmission by cancelling (at our level of bonus pay) the negative effect of social distance. In contrast, incentive pay has no impact on knowledge transmission or enrolment for socially proximate agent–beneficiary pairs. This result appears to be symmetric across social boundaries, in the sense that it holds whether the agent is from a high or low-status caste group. Our preferred interpretation is that, with respect to their ‘own’ group (socially proximate households), agents were already at a maximum effort level and hence, introducing bonus pay has no impact. However, non-incentivised agents choose a lower level of effort with respect to the ‘other’ group (socially distant households). With incentive pay, effort goes up to the same level as for the agent's own group. One might say that incentives appear to ‘price out prejudice’, although social distance barriers can operate through channels other than prejudice. We do not observe crowding out empirically, but cannot rule it out for unobserved parameter values.

Third, incentivised agents appear to achieve higher knowledge scores by reallocating time away from socially proximate households (their ‘own group’) towards socially distant households (their ‘cross-group’), without increasing aggregate time spent. The findings are consistent with a story in which non-incentivised agents spend more time than needed with their ‘own group’ because it is enjoyable rather than productive (‘idle chatter’). Incentivised agents channel some of this time toward productive use with households in the ‘cross-group’.

The article makes three main contributions. First, to the best of our knowledge, it presents the first randomised evaluation of incentive pay for agents tasked with providing information about a public service. An important aspect of service delivery is to make intended beneficiaries aware of their entitlements. Even if there were no supply-side problems – if the quality of schools and health centres were excellent and these facilities were widely available – the outcome would be disappointing if beneficiaries were unaware of the services or did not value them sufficiently (due to, say, a lack of information or present bias). While this is a recognised problem in rich countries,³ the issue has not received much attention in developing countries. There is, however, reason to believe that the problem is no less important there: a report on public services in India shows programme awareness to be low among target groups (World Bank, 2011).⁴ It is thus important to understand the role of incentives in raising awareness of social programmes, a context in which pro-social motivation is likely to feature.

Second, we contribute to the broader literature on financial incentives and performance by showing that incentives can matter, even in the context of a pro-social task, a soft objective and agents with possible intrinsic motivation. As Finan et al. (2017) point out, we do not know enough about the effect of financial incentives in these settings; in particular, when incentives may cause agents to prioritise dimensions that are easy to measure to the detriment of those that are less so, or when incentives could crowd out intrinsic motivation.

Third, the article extends our understanding of the interaction between incentive pay and social distance. While we are not the first to document the detrimental effects of social barriers, the question of whether incentive pay alleviates or exacerbates the negative consequences of social distance has not received much attention. This is particularly important in developing countries, many of which are highly stratified along socio-economic lines. The novelty of our findings is that what is ‘crowded out’ is an anti-social tendency to favour interactions with one's own group, whereas most previous studies have focused on financial incentives crowding out pro-social tendencies, such as picking up one's children on time from day care (Gneezy and Rustichini, 2000a).

Our article is related to that of Bandiera et al. (2009), who study the interplay of social connections and financial incentives in the context of worker productivity in a private firm in the UK. They find that when managers are paid fixed wages, they favour workers with whom they are socially connected; but when incentive pay is introduced, managers' efforts do not depend on social connections. But as Bandiera et al. (2011) point out, provision of incentives for pro-social tasks raise different issues compared to private tasks for several reasons, including the possibility of crowding out.

Another article looking at the effects of social distance in a for-profit setting is Fisman et al. (2017). They analyse data from an Indian bank, and find that the volume of credit is larger, and repayment rates higher, when the borrower and the loan officer are matched on social identity. While they are able to exploit quasi-random variation in social distance, they do not study the interaction of social distance and pay.

There is a growing literature on the importance of information campaigns in economic decision-making and, in particular, in determining demand for public services. Previous work has explored how information campaigns affect local participation and educational outcomes in India (Banerjee et al., 2010), how providing information on measured returns increases years of schooling (Jensen, 2010) and how creating awareness about HIV prevalence reduces incidence of risky sexual behaviour among Kenyan girls (Dupas, 2011).⁵

There is substantial evidence that ethnic heterogeneity is linked to poor economic outcomes, including sub-optimal provision of public goods and poor governance (Easterly and Levine, 1997; La Porta et al., 1999; Kimenyi, 2006). A possible explanation for this is that people prefer to interact with those who are similar to themselves, leading to fragmented markets, lower social mobility (Bertrand et al., 2000) and reduced gains from trade (Anderson, 2011). Several studies find evidence of strong own-group bias (Banerjee and Munshi, 2004; Kingdon and Rawal, 2010), with potentially adverse implications for the flow of information. In the context of awareness campaigns, if people prefer to liaise with their own kind, information constraints on the demand for public services may be more severe in socially heterogeneous settings. However, micro-level evidence on the role of social distance in the spreading of awareness about public services is rare.

This article is also related to the rich literature on the impact of monetary and non-monetary incentives on the performance of agents. This body of work encompasses studies in the ‘standard setting’ of firms in developed countries where output or productivity is measurable but worker effort is not (Lazear, 2000), as well as articles on incentives for teachers and health workers in developing countries as surveyed by Kremer and Holla (2008) and Glewwe et al. (2009). Muralidharan and Sundararaman (2011) and Duflo et al. (2012) study the effect of financial incentives for teachers on absenteeism and test scores, while BenYishay and Mobarak (2014) look at the role of incentives in leveraging peer learning to promote the adoption of agricultural technologies in Malawi.

There are also studies looking at the role of agents' intrinsic motivation and identification with either the task at hand or the intended beneficiaries in reducing the need for explicit incentives (Bénabou and Tirole, 2003; Akerlof and Kranton, 2005; Besley and Ghatak, 2005). In a laboratory setting, Gneezy and Rustichini (2000b) find non-monotonicities in the effect of incentive pay on effort. As Bandiera et al. (2011) point out, there is little field-experimental evidence in this area, although Ashraf et al. (2014) is a recent exception.

The rest of the article is organised as follows: In Section 1, a simple theoretical framework is presented with the aim of analysing the impact of incentive pay on agents' effort and its interaction with social-identity matching. Section 2 describes the context, experimental design and data. Section 3 presents the empirical evidence and Section 4 concludes.

1. Theoretical Framework

In this Section, we develop a simple model of motivated agents. It extends the study by Besley and Ghatak (2005) by incorporating features of the multi-tasking model (Holmstrom and Milgrom, 1991). The aim is to provide a theoretical framework that can generate predictions about the effects of incentive pay and how these might interact with the effects of social distance.

Suppose agents exert an unobservable effort in spreading awareness of a scheme to potential beneficiaries. The goal may be either the transmission of knowledge itself or to increase programme enrolment. The principal can be thought of as a planner (say, the relevant government agency) who values either awareness of or enrolment in the programme among the eligible population. A given agent can interact with an exogenously fixed number of target households.

1.1. A Single Task

First, assume there is a single task. This could correspond to a situation in which the potential beneficiaries of the public service are relatively homogeneous. Let e be the unobservable effort exerted by the agent. Let the outcome variable Y be binary, with the value 0 denoting ‘bad performance’ or ‘failure’, and the value 1 denoting ‘good performance’ or ‘success’. For example, a household doing well in a knowledge test (say, scoring above a certain threshold level), or enrolling in the programme, might be considered a success.

Agent effort stochastically improves the likelihood of a good outcome. To keep things simple, assume that the probability of success is p(e) = e, so that attention is restricted to values of e that lie between 0 and 1. Let us further assume that the lowest value e can take is e|$\in$| (0, 1), and the highest value e can take is |$\overline{e}$||$\in$| (e, 1). This means that there is some minimum effort that any agent supplies and that even with this minimum effort, there is some chance that the good outcome will happen. There is also a maximum level of effort, and even at that level, the good outcome is not guaranteed to occur. Therefore, as is standard in agency models, there is common support. That is, either outcome (0 or 1) is consistent with any level of effort in the feasible range. It is also assumed that both the principal and the agent are risk neutral.

If effort is contractible, the principal can simply stipulate e**. For the problem to be interesting, and for incentive pay to have an effect, assume that there is moral hazard in the choice of effort. Also, agents have zero wealth and there is limited liability: the agent's income in any state of the world must be above a certain minimum level, say, ω > 0. From the principal's point of view, this creates a tension between minimising costs and providing incentives. In the absence of a limited liability constraint (LLC), the principal could have achieved the first-best outcome by imposing a stiff penalty or fine for failure. With limited liability, the only way the principal can motivate the agent, beyond relying on her intrinsic motivation θ, is to pay her a bonus that is contingent on performance. When setting the bonus, the principal has to respect the LLC and the incentive compatibility constraint (ICC). There is also a participation constraint (PC) which requires the agent's expected pay-off to be at least as high as her outside option. To keep things simple, it is assumed that the outside option is relatively unattractive so that the PC does not bind – the analysis would be qualitatively unchanged if this assumption were relaxed.

Experimentally, we only observe outcomes for two given values of b, so the focus here will be on the ICC (1) rather than the optimal bonus. If there is no bonus pay and the agent is not sufficiently intrinsically motivated, we may get a lower corner solution, namely e = e. This will be the case if e ≥ θ/c. At the other extreme, if the agent is sufficiently motivated (namely, θ/c ≥ |$\overline{e}$|⁠), then even without any bonus pay the agent chooses the maximum level of effort, |$\overline{e}$|⁠. Otherwise, effort is increasing in bonus pay. The solution is illustrated in Figure 1. The slope of the interior-solution segment (1/c) is positive and so is its intercept (θ/c). However, depending on parameter values, the value of e for any given value of b could range from e to |$\overline{e}$|⁠. For example, the case of a relatively unmotivated agent is captured by the dashed vertical line marked by ce > θ. In this case, the vertical axis (at which b = 0) intersects the effort curve at a flat section where e = e. Similarly, a case where the agent is relatively highly motivated is captured by the dashed vertical line marked by θ > c|$\overline{e}$|⁠, and an intermediate level of motivation is captured by the line marked ce < θ < c|$\overline{e}$|⁠. In the former case, the agent is at the minimum effort level for b = 0 and initially the marginal effort with respect to bonus pay is zero. As bonus pay increases further, the marginal effort becomes positive, before returning to zero once the effort curve has hit the upper bound. If the vertical axis is at the right-most dashed vertical line, then the agent is already at the maximum effort level when b = 0 and effort will be unresponsive to incentive pay at any level. If the vertical axis is at the middle dashed line, effort level is at an interior value when b = 0 and the marginal effort with respect to bonus pay is positive.

1.2. Two Tasks

Assume now that the agent has two tasks, as in the multi-tasking model. The tasks may be thought of as the agent exerting effort to transfer knowledge to, or enrol, two different types of beneficiary households. However, unlike in the classic multi-tasking model, the outcomes associated with the two tasks are assumed to be equally measurable. Instead, the differences between the two tasks are in the agent's intrinsic pay-off from success and her cost of effort. Extending the notation from the previous section, let Y₁ and Y₂ be the binary outcomes for the two tasks and e₁ and e₂ the corresponding effort levels.

It is assumed that the principal is constrained to offer the agent the same conditional payments for the two tasks. That is, the payment in the case of success must be the same for tasks 1 and 2, as must the payment in the case of failure. This is justified if the principal is politically, socially or legally constrained to offer the same pay rates for all tasks. The assumption is also justified if the relevant characteristics of the households are not observable to the principal. For example, a knowledge agent may be biased in favour of some social or economic group or may have purely idiosyncratic biases, but if the principal does not observe the relevant dimension, the remuneration scheme cannot correct for it.

Note that if c₁ = c₂ = γ = c and θ₁ = θ₂ = θ, the set-up collapses to the single-task model. Abstracting from the special case c₁ = c₂ we can, without loss of generality, assume that c₁ < c₂ and refer to task 1 and 2 as the easier and the harder task respectively.

Several aspects of the solution are worth noting. First, effort in the easier task, e₁, is always weakly increasing in b.

Third, when both effort curves are internal, the slope of e₁ is always greater than the slope of e₂.

Fourth, the slopes of all internal curves are completely determined by γ, c₁ and c₂. The role of θ₁ and θ₂ is to shift the intercepts, and hence the lengths and meeting points, of the effort curves' constituent line segments.

The main types of solutions can be classified using the relative magnitudes of γ, c₁ and c₂. Above, it was assumed without loss of generality that c₁ < c₂, and the second-order condition requires c₁c₂ > γ². The substitutability parameter γ must therefore be either less than both c₁ and c₂, or equal to c₁ and less than c₂, or lie between c₁ and c₂.

Figures 2–4 illustrate representative cases⁷ where task 1 is intrinsically preferred (effort in task 1 is greater at b = 0), and moreover, e₁ is already at the highest possible level |$\overline{e}$| but e₂ has an interior solution. The latter corresponds to the condition |$(\theta_{2} - \gamma \overline{e})/c_{2} < \overline{e} < (c_{2}\theta_{1} - \gamma \theta_{2})/(c_{1}c_{2} - \gamma^{2})$|⁠. As in the single-task model, other solutions can be generated by drawing the vertical axis just to the left of the crossing point of the two effort curves, in which case task 2 would be intrinsically preferred. Also illustrated are the ‘kinks’ in e₂ that arise as e₁ meets the upper or lower bounds.

Solutions with γ < c₁ < c₂ (relatively low task substitutability) are illustrated in Figure 2. In the centre of the Figure, both effort curves are internal and positively sloped, while the slope of e₁ is greater than that of e₂.

Figure 3 illustrates the case γ = c₁ < c₂. Here, effort in task 2 is temporarily satiated while both effort curves are internal. Again, which task is intrinsically preferred depends on the position of the vertical axis.

Figure 4 illustrates the case c₁ < γ < c₂ (relatively high task substitutability). This is the only case that permits ‘crowding out’, that is, a phase in which effort in one task (task 2) decreases with increasing bonus pay. As illustrated, crowding out can only happen when both effort curves are internal. Again, the intrinsically preferred task is determined by the position of the vertical axis.

Mapping the theory to the experimental setting, each of the model's two tasks can be thought of as corresponding to a group of eligible households in the agent's village. In the empirical analysis we find that, in the absence of bonus pay, agents tend to exert a greater effort with respect to households who are similar to themselves in terms of social characteristics. The model's ‘intrinsically preferred task’ therefore corresponds to households who are socially proximate to the agent. These households will also be referred to as the agent's ‘own group’. Households who are socially distant from the agent (the ‘other group’) correspond, in the model, to the task that is not intrinsically preferred.

Which task is intrinsically preferred depends on θ_i and c_i, both of which are in principle unobservable. Therefore, while the agent's ‘own’ group will be mapped to the intrinsically preferred task, it is not always possible to deduce whether this is task 1 (the easier task) or 2 (the harder task).

Note that we have modelled the agent's effort but not her time use. Some of the results presented below suggest that these are not the same: agents appear to be able to hold effort constant while varying the time spent on a task. Our interpretation is that agents can control the intensity of effort (effort exerted per unit of time) – in particular, they may engage in enjoyable but unproductive ‘idle chatter’ with their friends.

2. Context, Experimental Design and Data

2.1. The Programme

The experiment was conducted in the context of India's National Health Insurance Scheme (Rashtriya Swasthya Bima Yojana – henceforth, RSBY). The scheme was launched by the central government in 2007 with the aim of improving the ‘access of BPL (Below the Poverty Line) families to quality medical care for treatment of diseases involving hospitalisation and surgery through an identified network of health care providers’ (Government of India, 2009). Each state followed its own timetable for implementation, and a few districts from each state were selected for the first stage. In Karnataka, five districts were selected (Bangalore Rural, Belgaum, Dakshina Kannada, Mysore and Shimoga), and household enrolment in these districts commenced in February–March 2010 (Rajasekhar et al., 2011).

The policy covered hospitalisation expenses for around 700 medical and surgical conditions, with an annual expenditure cap of 30,000 rupees (US$ 652) per eligible household.⁸ Each household could enrol up to five members. Pre-existing conditions were covered, as was maternity care, but outpatient treatment was excluded.

The policy was underwritten by insurance companies selected in state-wise tender processes. The insurer received an annual premium per enrolled household,⁹ paid by the central (75%) and state (25%) governments. The beneficiary household paid only a 30 rupee (US$ 0.65) annual registration fee.

Biometric information was collected from all members on the day of enrolment and stored, along with photographs, in a smart card issued to the household.¹⁰ Beneficiaries were entitled to cashless treatment at any participating (‘empanelled’) hospital across India. Both public and private hospitals could be empanelled. Hospitals were issued with card readers and software. The insurance companies reimbursed the hospitals for the cost of treating patients, according to fixed rates.

2.2. Experimental Design

One hundred and fifty-one villages were randomly selected from two of the first-phase RSBY districts in Karnataka: Shimoga and Bangalore Rural. In the first stage of randomisation, some villages in our sample (112 of 151) were randomly selected to be part of the treatment group, that is, receive an agent, while the remaining form the control group. In each treatment village, our field staff arranged a meeting with the local self-help groups (SHGs).¹¹ All contacted SHGs were female-only. In the meeting, SHG members were given a brief introduction to RSBY and told that we were looking to recruit a local agent to help spread awareness of the scheme in the village over a period of one year. They were told that the agent would be paid, but no further details about payment were given at that time. In each case, a single candidate was nominated by the group and recruited on the same day. The nominated agent was a member of the SHG, except in two cases where the selected agent was a non-member recommended by the SHG. In about a third of the cases, the president of the SHG became the agent. All agents were female.

Once the meeting was concluded and the agent selected, she was taken aside and given a more thorough introduction to the scheme, including details on eligibility criteria, enrolment, benefits and other relevant information. An agent background questionnaire was also fielded at this time.

The payment scheme was revealed to the agent only after recruitment. Each treatment village had been randomly allocated to a payment structure, which constituted the second stage of randomisation, but this information was kept secret. Even our field staff did not know about the contract type until after the agent had been selected. The day after recruitment, the agent was called and informed of her payment scheme. There were two payment schemes, defining the two treatment groups: flat-pay agents were told that they would be paid 400 rupees every three months. Incentive-pay agents were told that knowledge of RSBY would be tested in the eligible village population every three months. The agent's pay would depend on the results of these knowledge tests. There would be a fixed payment of 200 rupees every three months, but the variable component would depend entirely on the outcome of the knowledge tests in the village.¹²

The bonus payments were determined as follows: a random sample of households eligible for RSBY in each village was surveyed and orally presented with the knowledge test.¹³ A household was classified as having ‘passed’ the test if it answered at least four of eight questions correctly. The proportion of passing households in a village was multiplied by the number of eligible households in that village in order to estimate the total number of eligible village households that would have passed if everybody had taken the test. The bonus was calculated as a fixed amount per eligible household estimated to pass the test in a village, and set in such a manner that the average bonus payment across each of the two study districts would be 200 rupees per agent. The households taking the tests were not told how they scored, nor were they provided with the correct answers.

Thirty-eight villages/agents were assigned to the flat-pay treatment group, and 74 to the incentive-pay treatment group. Agents were told that there would be other agents in other villages, but not that there was variation in the payment scheme.

The purpose of not revealing the payment scheme until after recruiting the agent was to isolate the incentive effect of the payment structure from its potential selection effect. None of the agents pulled out after learning of the payment scheme. However, four agents dropped out 6–12 months after recruitment (after at least two rounds of payments). Three of these were in incentive-pay villages, while the fourth was in a flat-pay village.¹⁴ In each case, the reported reason was either childbirth or migration away from the village. The agents were replaced, but the villages in question are excluded from the analysis. Hence, in the analysis presented here, there are 37 villages with flat-pay agents and 71 villages with incentive-pay agents, for a total of 108 agents in 108 treatment villages. The number of control villages remains 39, so the total number of villages in our final sample is 147.

One question of interest is whether eligible households knew the type of payment scheme the agent in their village was on. We do not have data on this, but on balance we believe that most did not. When asked to nominate an agent, the self-help group was told the work would be remunerated but given no further details. We were careful to tell the agent about the type of contract in private, after selection, and away from the group. When we returned to make payments, these were also always made in private. The agents were of course free to tell others how they were paid, but people locally tend to be reticent in talking about money. Anecdotally, we know that at least some agents on incentive pay were careful not to reveal this information because, once spread, it might reduce their credibility as pro-social volunteers and thereby the villagers' willingness to listen. In any case, from a policy point of view, one would probably want to capture the total effect of the contract types inclusive of any additional effect on eligible households who learn how their agent is paid.¹⁵

The original plan was to set the variable part of the pay scale for incentive-pay agents in such a manner that average pay would equal 400 rupees in each of the two treatment groups. The aim of equalising average pay across the incentive-pay and flat-pay groups was to isolate the incentive effect of the contract structure (‘incentive effect’) from that of the expected payment amount (‘income effect’). The pay did in fact average 400 rupees for one district (Shimoga) in the first survey round and for both districts in the second, third and fourth rounds. But due to an administrative error, a majority of incentive-pay agents in Bangalore Rural were overpaid in the first round of payments. In spite of the error, the rank ordering of agents was preserved in the sense that better-performing agents were indeed paid more. Nevertheless, we also present results only for Shimoga district, where average pay in the knowledge group was equal to that of flat-pay agents (400 rupees) in all rounds.

2.3. Data

Following agent recruitment, four consecutive rounds of ‘mini-surveys’ were fielded.¹⁶ In each wave, randomly selected eligible households in each sample village were interviewed to establish the state of their knowledge about the scheme and determine their test scores, as well as measure their enrolment status. An important purpose of these surveys was to provide information on agent performance so as to be able to pay the incentive-pay agents. The households were drawn at random (with replacement) for the first, second and fourth survey rounds, so that there is a partial overlap between the households in these rounds. The first, second and fourth rounds of mini-surveys were based on face-to-face interviews. For the third survey, the sample from the second survey was re-used, but this time the households were contacted by telephone. Although not everyone could be reached by phone, the re-survey rate was significant. A sample of 2,360 households were interviewed in the first mini-survey wave, 1,931 in the second, 1,346 in the third and 2,093 in the fourth. In all, the mini-surveys cover 3,998 households, of which 1,068 were interviewed twice, 642 were interviewed three times and 460 were interviewed four times. As the tests were conducted in every sample village, there was no difference between incentive and flat-pay agents in the intensity of monitoring.

Using each household observation as an equally-weighted data point would give more weight to households that were observed more than once. Observation weights were introduced to take account of this, so that the total weight across observations equals 1 for all households. All regressions using data from more than one mini-survey are weighted least squares. In addition, standard errors are clustered at the village level (Bertrand et al., 2004). Since serial correlation is probably more severe within a household than across households within a village, clustering at the village level yields consistent, but not necessarily efficient, estimates.

After the completion of each mini-survey, the agents were revisited and paid. At the same time, the agents' knowledge of the scheme was refreshed and added to.

Descriptive statistics on agents are presented in Table 1. Recall that all agents are female. The average agent is around 35 years old, 88% are married, 58% of the agents' household heads have completed primary school and 82% of agent households have a ration card,¹⁷ and 38% are from a forward or dominant caste.¹⁸ In 29% of the cases, the recruited agent was the president of a self-help group. We also constructed a ‘female autonomy’ score for the agents.¹⁹

Table 2 presents summary statistics for the villages. The average village has a little over 200 households, of which about 50 are eligible for the scheme. About a quarter of sample villages are gram panchayat (village council) headquarters, and the distance from the village to the nearest town is about 13 kilometres. None of the village-level variables differ significantly across treatment groups.

Table 3 presents summary statistics for households. The average household has 4.8 members, 17% are from a forward/dominant caste. In 27% of households, the household head has completed primary school, 92% have a ration card. It is interesting to note that agents are more likely than the average eligible household to belong to the forward/dominant caste category. Agent households are also more highly educated than the average eligible household.

The main outcome variable is the household ‘knowledge score’. A knowledge test was fielded, in each of the four mini-surveys, to all interviewed households across the three experimental arms. Each test consisted of eight questions about particulars of the RSBY scheme, including eligibility, cost, cover, exclusions and how to obtain care. The exact questions used in the knowledge tests are provided in online Appendix B. Each answer was recorded and later coded as being correct or incorrect. The number of correct answers gives each interviewed household a score between 0 (least knowledgeable) and 8 (most knowledgeable).²⁰

The test questions asked in the four surveys were different, so although the raw scores can be compared across households within a survey, they are not necessarily directly comparable across surveys, even for a given household. The scores on each test were therefore standardised by subtracting the test-wise mean and dividing by the standard deviation.

3. Evidence

3.1. The Impact of Agents on Knowledge

The results of regression (2) are presented in Table 4, column (1). Households living in a treatment village score 0.18 standard deviations higher on the knowledge test compared to households in the control villages. Column (2) indicates that this effect is robust to the inclusion of fixed effects for taluk (the administrative unit below district) and time (survey wave).

In column (3), the treatment effect is estimated separately for flat-pay and incentive-pay agents, while still including taluk and time fixed effects. The estimated effect of flat-pay agents on test scores, while positive, is not statistically significant. This is consistent with the argument that, since these agents are paid a constant amount irrespective of outcome, they are not incentivised to exert any effort beyond the level determined by their intrinsic motivation. In contrast, households in villages assigned an incentive-pay agent score 0.25 standard deviations higher on the knowledge test than those in the control group. Hence, providing agents with financial incentives leads to an improvement in knowledge about the scheme among beneficiaries. Moreover, equality of these two coefficients is rejected. This suggests that, looking at the sample overall, the effect of knowledge-spreading agents is driven by the agents on incentive-pay contracts.

As mentioned, an administrative error caused incentive-pay agents in one district (Bangalore Rural) to be overpaid after the first survey. To allay concerns that our findings are driven by these higher rates of pay, Table 5 presents results using only data from Shimoga district, where no error was made. Overall, the qualitative findings are similar to those obtained in Table 4, if not stronger. Hence, it appears that the main findings are not driven by the larger payments made to agents in one district for one of the four rounds.

Effects by survey round and by agent characteristics are presented in online Appendix C.

3.2. The Impact of Agents on Enrolment

Next consider the impact of knowledge agents on programme enrolment. Although the agents were not incentivised to enrol households into the scheme, it is conceivable that increasing households' knowledge about the scheme might induce enrolment. Enrolment is also of primary policy relevance.

The results are presented in Table 6. Households living in a treatment village are on average 4.5 percentage points more likely to enrol in the scheme than households in the control villages, although this effect is not statistically significant (column (1)). Controlling for taluk and wave fixed effects does not change this result (column (2)). Once we disaggregate the treatment effect by agent contract type, we find that households living in a treatment village where the agent was on an incentive-pay contract are 7.5 percentage points more likely to be enrolled relative to control group, significant at 5%. No significant impact is found for those living in a treatment village where the agent was on a flat-pay contract. As with the knowledge score results reported above, we are able to reject the equality of the two coefficients at the 5% level. Hence, incentivising agents to disseminate programme knowledge also boosted enrolment.

3.3. Incentives and Social Distance

The results so far suggest that monetary incentives matter for how effective agents are at disseminating information about the scheme and increasing enrolment. But previous work suggests that social identity is also an important determinant of insurance take-up. Cole et al. (2013) find that demand for rainfall insurance is significantly affected by whether the picture on the associated leaflet (a farmer in front of either a Hindu temple or a mosque) matches the religion of the potential buyer.

This subsection asks whether matching agents with target households in terms of social characteristics has an effect on knowledge scores that is independent of the effect of incentive pay. Also, it investigates whether the effects of social distance and incentive pay are purely additive or whether they reinforce or weaken each other.

A simple metric of social distance was constructed as follows. First, we created four binary variables which capture basic social dimensions and for which we have data for both the agent and eligible households: forward/dominant caste status (0/1), whether the household head has completed primary school (0/1), ration-card status (0/1) and home ownership (0/1). In each of these four dimensions, the social distance between an agent and a household is defined as the absolute difference in the agent's and the household's characteristics. To take ration-card status as an example, ration-card distance is set to 0 if either both have a ration card or if neither does. Ration-card distance is 1 if any one of them has a ration card and the other does not.²¹

The composite social distance metric is the simple sum across the four individual distance measures, normalised to lie between zero and one by dividing by four.

Before turning to the main specification, consider the basic difference-in-differences calculation in Table 7. We create a binary variable, ‘socially proximate pair’, indicating whether or not the composite social distance metric is equal to or less than 0.5. The mean knowledge scores for socially proximate and socially distant household–agent pairs are tabulated by agent contract type. Reading the Table row by row, flat-pay agents are significantly more effective at transmitting knowledge to socially proximate households than to the socially distant households – average scores are higher by 0.14 standard deviations. For incentivised agents there is no significant difference in performance between close and distant pairs, and the difference in differences is significant. Alternatively, reading the Table column by column, incentive pay seems not to affect knowledge transmission in proximate pairs, but it does increase knowledge transmission in distant agent–household pairs, up to about the same level as for proximate pairs. Note that the socially distant households who are assigned a flat-pay agent score significantly lower than any of the other three groups, which is indicative of the disadvantage created by social distance in the absence of incentive pay. In summary, the difference-in-differences analysis suggests that incentive pay has the effect of neutralising social distance as an impediment to the transmission of knowledge – a finding that will be corroborated in what follows.

The coefficient β captures the effect²² of social distance on knowledge when the agent is not incentivised. The coefficient γ captures the effect of incentive pay for socially proximate (non-distant) agent–household pairs. Finally, δ captures the differential effect of incentive pay for socially distant agent–household pairs relative to socially proximate ones.

The results are presented in Table 8. Column (1) confirms that incentive-pay agents have a significant and positive impact on knowledge compared to flat-pay agents, even when controlling for agent and household caste, education, ration-card status and home-ownership as well as taluk and time-fixed effects.

Column (2) presents results for the composite social distance metric. The un-interacted treatment effect is not significant, while the coefficients on social distance and the interaction of incentive pay with social distance are both highly significant and roughly opposite in magnitude. We interpret this in three steps: first, it confirms that social distance has a negative impact on knowledge transmission. Second, putting agents on an incentive-pay contract has a positive effect on knowledge transmission, but only for socially distant agent–household pairs. And third, the effect of providing financial incentives (at our level of bonus pay) is more or less exactly the level required to cancel the negative effect due to social distance. In other words, the effect of incentive pay seems to be to cancel the negative effect of social distance, but no more. Loosely speaking, we may say that incentive pay appears to ‘price out prejudice’, although the effects of social distance are not necessarily a function of prejudice alone. For example, socially proximate pairs may meet more often socially, reducing the cost of effort required to transmit information. See online Appendix D for a discussion of the magnitude of the main effect relative to the incentive.

In Indian villages, caste groups sometimes live in distinct sub-villages called hamlets. This means that the social distance between a pair of households in the village may be positively correlated with the physical distance between them. To the extent that this is the case, it is possible that the results so far confound the effect on knowledge transmission of social distance with that of physical distance. After all, it seems natural that the cost of knowledge transmission increases with the physical distance between the agent and a household.

We control for the role of physical distance using two separate measures. The first is whether or not caste groups in a village tend to live apart. This information is based on enumerator recall and hence only available for 107 of the 147 villages. Based on this information, we construct a binary indicator which is equal to 1 if, in a given village, the settlements of the major caste groups are physically separated, and 0 otherwise. The indicator is equal to 1 for 26 of 107 villages. Returning to Table 8, this indicator and its interaction with the incentive-pay variable are included in the regression in column (3).

While the sample size drops, the results in column (3) confirm that physical separation does have a negative effect on knowledge transmission. Like for social distance, the coefficient on the interaction of incentive pay and physical distance is of opposite sign and roughly equal magnitude to the un-interacted physical distance term. While the interaction term is not statistically significant, this may indicate that incentive pay can overcome barriers to physical distance. But the results also show that the social distance indicator and its interaction with incentive-pay are still significant at the 10% and 5% levels, respectively, and do not drop much in magnitude. This indicates that social distance matters, even after controlling for physical distance.

The measure of physical distance used above is crude, and it is conceivable that the social distance measure still captures some physical distance effect. Hence, we use a second and more accurate measure of physical distance in the form of the actual straight-line distance in metres between the agent's and the household's dwellings, constructed from GPS coordinates collected in the field. Since these data were only collected during the fourth survey wave, the analysis can only be undertaken for a subset of the total sample. Using this second measure, we find results in column (4) that are qualitatively similar to those obtained with the crude measure in column (3). The coefficient on social distance is no longer statistically significant at conventional levels, possibly due to the fact that the sample size drops by almost a quarter. But the coefficient continues to be of opposite sign to that of its interaction term with incentive-pay. In addition, both the coefficient on social distance and that on the interaction term are not significantly different from those obtained in column (2) without the inclusion of physical distance (p-values for the F-tests are 0.38 and 0.74 respectively), which increases our confidence that these results are less likely to be driven by confounding factors. Moreover, this measure of physical distance does not appear to exert any independent influence on knowledge scores.

Columns (5)–(8) repeat the exercise in column (2) for each of the component social distance metrics. For distance in caste, ration-card status and home ownership, the story appears to align with the findings for the composite metric presented above, albeit not always with full significance. However, for education, there appears to be no significant disadvantage due to social distance. In other words, agent–household communication appears to be hampered by differences in caste, ration-card status and possibly home ownership, but not by differences in education. Correspondingly, in this specification, un-interacted incentive pay has a large and positive co-efficient. That is, agents do not appear to be at a disadvantage when communicating with households with a different educational background (having completed primary school or not) from themselves, and the introduction of incentives correspondingly boosts results for socially proximate and distant households alike.

It is of interest to examine whether the impact of social distance and its interaction with incentive-pay is symmetric across the caste hierarchy. In other words, is the impact of social distance between agent and beneficiary household more severe when a lower caste agent interacts with a higher caste household than vice versa? To test this, we compute differences in differences in mean effects by agent caste group. The results, presented in Table 9, suggest that the qualitative findings are symmetric: irrespective of the agent's own caste group, the coefficient representing the effect of introducing incentive pay is greater with respect to the cross-group than to the own group. This is reassuring, although in this basic analysis the differences in differences are not statistically significant.

Table 10 presents results for our other main outcome variable, enrolment. The results are qualitatively similar to those obtained for knowledge scores, and with somewhat greater significance for the composite measure of social distance (columns (2)–(4)). Once again the main effect of incentives appears to be to cancel the negative effects of social distance. This finding is robust to the inclusion of controls for either measure of physical distance. Taken together, the analysis in this subsection suggests that social distance can lower the efficacy of welfare programmes through reduced knowledge transmission and enrolment, and that the use of incentive pay can counteract this effect.