Intrinsic Incentives, Exercises of Economics

Download Intrinsic Incentives and more Exercises Economics in PDF only on Docsity!

Intrinsic Incentives: A Field Experiment on Leveraging Intrinsic

Motivation in Public Service Delivery

Scott S. Lee∗

JOB MARKET PAPER

January 5, 2018

(Please click here for latest version.)

Abstract Although extrinsic and intrinsic motivation likely jointly explain the effort of many agents engaged in public service delivery, canonical models of incentives in firms focus on the former. In the context of a rural health worker program in India, I develop and test a novel mobile phone app designed to increase agents’ intrinsic returns to effort. At one year of follow-up, the self-tracking app leads to a 24% increase in performance as measured by the main job task (home visits). Moreover, the app is most effective when it leverages pre-existing intrinsic motivation: it produces a 41% increase in performance in the top tercile of intrinsically motivated workers, but no improvement in the bottom tercile. This treatment effect persists over time for the most intrinsically motivated workers, whereas early improvements decay among the least motivated workers. Supplementary evidence suggests that the treatment effect on performance is mediated primarily by making effort more intrinsically rewarding, and not by other mechanisms such as the provision of implicit extrinsic incentives. Despite these effects on worker performance, I find no effect on health outcomes. ∗Harvard Medical School and Brigham and Women’s Hospital, [email protected]. I thank the Chief Medical Office of Kaushambi District, Dr. Nandini Sharma/Maulana Azad Medical College, and Dimagi, Inc. for hosting this research; Brian DeRenzi, Andrew Ellner, and Neal Lesh for ongoing collaboration as co-Investigators; and Sapana Gandhi, Sangya Kaphle, Sugandha Nagpal, Robert Racadio, and Jeremy Wacksman for excellent research assistance. I am grateful to Nava Ashraf, Oriana Bandiera, Iqbal Dhaliwal, Paul Farmer, Rema Hanna, Michael Kremer, Matthew Rabin, Andrew Weiss, and seminar participants at the Northeastern Universities Development Consortium, Harvard Business School, Harvard Department of Economics, and Harvard Medical School for helpful comments. Generous financial support has been provided by the Massachusetts General Hospital Consortium for Affordable Medical Technologies, Child Relief International, and the Harvard Business School Doctoral Program.

1 Introduction

Public services such as governance, education, health care, and national defense rely on agents for their provision. In developing countries, the effort of these agents is often a binding constraint, prevailing over other factors such as the agents’ ability and market demand (Das & Hammer, 2007, 2014; Leonard et al., 2013; Maestad et al., 2010). The standard agency model offers both an explanation and a solution to this problem: agents dislike effort, and they can be persuaded with incentives to exert it. In line with this view, recent field experiments have shown that both monetary and non-monetary incentives can improve the performance of agents engaged in public service delivery (Ashraf et al., 2014; Basinga et al., 2011; Duflo et al., 2012; Miller et al., 2012; Muralidharan & Sundararaman, 2011).^1 What this perspective neglects, however, is the converse scenario: agents who exert effort despite having little extrinsic incentive to do so. In settings in which forty percent of health workers are absent from their posts on any given day (Chaudhury et al., 2006), what explains the presence of the remaining sixty percent? While other extrinsic factors (e.g., monitoring, social pressure, status-seeking) likely contribute, the persistence of effort in the face of weak incentives—and the decision to select into pro-social jobs in the first place—suggests a role for intrinsic motivation.^2 That agents can be intrinsically motivated is not new in economics (Benabou & Tirole, 2003; Fehr & Schmidt, 1999). But this motivation is typically taken as given—a fixed trait that or- ganizations may wish to select for but cannot influence (Besley & Ghatak, 2005). If it can be influenced, it is only for the worse, as proposed by theories of motivational crowd-out (Deci et al., 1999; Benabou & Tirole, 2006).^3 In contrast, the business literature has long posited that managers can leverage intrinsic motivation by manipulating job attributes such as autonomy (Deci & Ryan, 1985), purpose (Weick, 1995), and organizational culture (Schein, 1985).^4 But these attributes are (^1) For evidence challenging the effectiveness of financial incentives in public service delivery, see Banerjee et al. (2008), Glewwe et al. (2010), and, in the US context, Fryer (2013). For theoretical contributions on why financial incentives in public service delivery may fail, see Benabou & Tirole (2003, 2006). 2 For empirical evidence that supports the hypothesis that those with pro-social preferences select into pro-social jobs, see Kolstad & Lindkvist (2013) and Lagarde & Blaauw (2013). 3 A small but notable exception is theoretical and empirical work on the role of delegation and empowerment in enhancing intrinsic motivation. See, e.g., Aghion & Tirole (1997); Rasul & Rogger (2015). 4 For more recent evidence from laboratory experiments, see Grant (2007) and Ariely et al. (2008).

generally conceptualized as static elements of job design, as opposed to incentives that interact with heterogeneous preferences. To provide a richer view into the relationship between intrinsic incentives and intrinsic moti- vation, this paper explores the potential role of an intrinsic incentive technology in enhancing the effort of agents engaged in public service delivery. I define an intrinsic incentive as any variable in the principal’s choice set that modifies the agent’s marginal intrinsic utility of effort, analogous to how an extrinsic incentive modifies the marginal extrinsic utility of effort. In the setting of a rural health worker program in India, I develop a novel mobile phone technology—a “self-tracking” app—designed to act as an intrinsic incentive by delivering information that makes effort more intrinsically rewarding.^5 The app comprises a set of graphs that a health worker can access on her phone to view her performance with respect to the job’s primary task: visiting and provid- ing support and counseling to pregnant women in their homes. As a counterfactual, I develop an analogous app—a “generic encouragement” app—designed to be lower-powered in that it provides generic messages of encouragement that are independent of the agent’s effort.^6 I test these incentives head-to-head by randomly assigning 145 health workers to receive one of the two apps on their work phone, and then tracking both app usage and performance on a daily basis for one year. The experiment yields four main findings. First, both intrinsic incentive technologies are demanded. Across the two treatments, despite receiving minimal encouragement and no directive to do so, the average health worker accesses the software application approximately once every three days. Second, turning to effects on effort, compared to the generic encouragement app, the self-tracking app leads to a 23.8% increase in performance as measured by home visits. Third, the self-tracking app is most effective when it leverages pre-existing intrinsic motivation; it produces a 41.4% increase in performance in the top tercile of intrinsically motivated workers, but no improvement in the bottom tercile, indicating that, in this setting, intrinsic incentives and intrinsic (^5) The technology is intended to leverage intrinsic motivation both in the sense used in the psychological literature (motivation to do a task for its own sake—effort as its own reward) and in the sense of prosocial preferences (motivation to do a task due to its positive social externalities—impact as its own reward), both of which are distinguishable from extrinsic motivation in that no benefit external to the task and its output is needed to justify effort. Throughout this paper, unless otherwise stated, I use “intrinsic” motivation to encompass both intrinsic and prosocial preferences. 6 The conceptual distinction between the two lies in the slope with which they are expected to modify agents’ marginal intrinsic utility of effort, analogous to how a piece rate of x dollars per piece is higher-powered than one that pays x 2 dollars per piece.

motivation are complements. Moreover, those with high intrinsic motivation exhibit a durable response to the self-tracking app, whereas those who are least intrinsically motivated respond initially but eventually perform no better than their counterparts in the control group. Fourth, despite these gains in client visits, no aggregate impact on the health of the pregnant clients and their children is observed, which, while not desirable from a welfare standpoint, supports the notion that the increased effort is elicited primarily for intrinsic reasons. To support this, supplementary evidence suggests that the treatment effects of the self-tracking app are not mediated by implicit extrinsic incentives or effects on the production function. Rather, they appear to increase effort by making effort more intrinsically rewarding. This paper contributes to two nascent literatures. First, recent empirical work has evaluated information incentives as a tool for motivating prosocial behavior. In particular, various forms of relative performance feedback have been found to be effective in improving home energy conser- vation (Allcott, 2011), student learning (Tran & Zeckhauser, 2012; Azmat & Iriberri, 2010), and physician quality (Kolstad, 2013). Information incentives, however, can function as extrinsic or intrinsic rewards, and when the targeted task (such as energy conservation or school performance) confers financial benefits, an information incentive that increases effort in the task is likely to be operating at least in part via an extrinsic channel.^7 In contrast, the self-tracking app studied in this experiment is designed to increase only intrinsic returns to effort. In this regard, this intrinsic in- formation incentive conceptually shares more in common with other intrinsic rewards, such as task meaning and organizational mission, than with other information incentives such as performance feedback. The second literature examines the interaction between incentives and psychological traits such as intelligence and personality in prosocial behavior, in the context, for example, of civil servants (Dal Bo et al., 2013), health agents (Ashraf et al., 2014, 2015), and taxpayers (Dwenger et al., 2014). In particular, Callen et al. (2015) find that personality traits predict job performance among health (^7) Kolstad (2013) finds that cardiac surgeons respond to physician report cards in ways that cannot be explained solely by profit maximization. He defines this reduced-form residual as an “intrinsic incentive” effect, but is silent about what in the physicians’ utility function drives this effect. As such, the results are consistent with physicians behind motivated not only intrinsically, but also by non-financial extrinsic preferences such as those for prestige, recognition, and career promotion.

officials in Pakistan, and that the experimental response to a novel monitoring technology varies with these personality traits. I extend this approach by (a) testing an intrinsic rather than extrinsic incentive, (b) testing a specific, theoretically guided interaction—that between intrinsic incentives and intrinsic motivation—and (c) elucidating psychological mechanisms. The paper is organized as follows. Section 2 presents a simple principal-agent framework in which the principal can offer extrinsic and intrinsic incentives, and agents have extrinsic and intrinsic preferences. Section 3 describes the policy and program context. Sections 4 and 5 present the experimental design and results, respectively. Section 6 concludes.

2 Framework

In this section, I develop a simple framework for analyzing the potential role of intrinsic incentives in public service delivery. I take extrinsic incentives as given and instead focus on the principal’s ability to stimulate effort by increasing intrinsic—i.e., direct hedonic—returns to effort. The framework assumes that an agent may find an effort task tasteful or distasteful for intrinsic reasons. Modifying or augmenting some aspect of the task may make it more tasteful, thereby increasing its marginal intrinsic returns and, by extension, equilibrium effort. The possibility that the principal may be able to control this modification paves the way for intrinsic incentives. I show that, like an extrinsic incentive, an intrinsic incentive can be high- or low-powered, and the incentive interacts with the agent’s intrinsic motivation to determine the marginal intrinsic utility of effort. This interaction generates testable implications that I then take to the experimental data.

2.1 Principal’s choice

Consider a single-principal, single-agent framework. The principal (e.g., a government, nonprofit organization, hospital) is invested in the production of a social good Y (e.g., population health). The agent can contribute to the production of Y with effort ei ≥ 0. Assume that the production function Y ( ei ) is monotonically increasing in ei.

To motivate the agent to exert effort, the principal may use extrinsic and intrinsic incentives. The former provides a reward contingent on output, whereas the latter enhances the hedonic returns to effort itself. Denote by α ≥ 0 and β ≥ 0 the extrinsic incentive contract, where α is a fixed reward and β is paid linearly for each unit of Y produced.^8 Because α does not affect equilibrium effort, for notational convenience, assume that α = 0. Denote by ψ (·) an intrinsic incentive function. ψ (·) will enter into the agent’s utility function as described in Section 2.2 below. Here, I comment on its conceptual underpinnings. The ψ (·) function is left purposefully vague. Definitionally, ψ (·) is a function that alters the marginal intrinsic utility of effort. It is analogous to β , but whereas β is non-negative, the value of ψ (·) is unbounded—i.e., it can increase or decrease marginal intrinsic returns. Theories of motivational crowd-out (Deci et al., 1999), for example, imply ∂ψ∂β < 0 —i.e., an increase in extrinsic returns to effort reduces its intrinsic returns.^9 Potentially any job attribute could affect marginal intrinsic utility of effort: the nature of the effort task, the technology of production, the degree of monitoring vs. autonomy, organizational norms and culture, and so forth. For the current purposes, let the principal’s choice variable in the ψ (·) function be the psychological salience and observability to the agent of the effort task and its social impact. The principal can alter the agent’s information environment to achieve this effect. For example, she may provide a technology by which the agent is better able to self-observe effort and output and, in so doing, experience greater marginal utility (or disutility) of effort. The principal chooses intrinsic incentive regime j ∈ { h, l }, corresponding to “high-powered” and “low-powered,” respectively. The high-powered (low-powered) regime enables the agent to access information that makes the effort task and its social impact highly (minimally) salient. Assume that j maps onto ψ ( j ) such that ψh = 1 and ψl = 0. Assume also that providing j carries zero marginal cost, in terms of both the direct supply cost and the shadow cost of reputation.^10 Thus, (^8) The conventional interpretation for β would be a variable wage, but the conceptual intuition can be extended to other extrinsic goods such as social status and job security. 9 I abstract away from this relationship in this framework because it is not empirically relevant in this experiment; as I describe in Section 4, there is no variation in 10 β in the sample and hence no way to identify ψ ( β ). Benabou & Tirole (2003) describe this shadow cost as follows: “A teacher or a manager who makes very complimentary comments to every pupil or employee may lose her credibility....[W]hen disclosing soft information to several agents the principal must realize that they will see through her ulterior motivation, and believe her only if she builds a reputation for not exaggerating claims.” Why there is no shadow cost of j in the experimental intervention

the only cost incurred by the principal is βY , but, as is made explicit below, Y is affected by ψ ( j ) via the latter term’s effect on optimal effort. Let the principal’s utility be V ( Y − βY ), where, given risk neutrality, V ′^ > 0 and V ′′^ = 0. The principal chooses { e, β, j } to maximize her expected utility, subject to the agents’ individual rationality and incentive compatibility constraints.

2.2 Agent’s choice

I now turn to the agent’s choice of e. A risk-neutral agent i with preferences ( θEi , θIi ) chooses effort ei to maximize his expected utility:

Uij = θEi [ α + βY ( ei )] + θIi ( γi + ψj ei ) − e (^2) i 2 (1)

Equation 1 has three terms: an extrinsic payoff term, an intrinsic payoff term, and an effort cost function. The two payoff functions are weighted by extrinsic ( θEi ) and intrinsic ( θIi ) preference parameters, respectively. Assume θEi and θIi are independent and that θEi is distributed over [0 , 1]. In contrast, assume θIi is distributed over [− 1 , 1]. Similar to the discussion of ψ (·) above, whereas preferences for extrinsic rewards can logically only be weakly positive, intrinsic payoff can be positive or negative depending on the tastes of the agent toward the effort task. The convex cost function, e 22^ i , represents an intrinsic cost of effort encompassing quantities such as time, leisure, and caloric expenditure. The first term—the extrinsic payoff function—is simply the principal’s payment contract con- sisting of a fixed wage α and a piece rate βY weighted by the agent’s preference for extrinsic rewards, θiE. For simplicity, assume α = 0. As with the payment contract, the intrinsic payoff term has two components—one that is effort-independent and another that is effort-dependent. The effort-independent parameter, γi , can be thought of as an endowment of intrinsic (dis)utility—e.g., the warm glow experienced from having a prosocial job, or the effort-independent disutility of hav- ing a job that runs counter to one’s tastes. Since this endowment affects equilibrium utility but is explained in more detail in Section 4.1, but in brief, because j only contains objective information, it does not rely on scarcity for its value, whereas recognition, praise, or positive feedback does.

not effort choice, for simplicity, assume γi = 0. The second component of the intrinsic payoff term, ψj ei , captures the practical intuition that, in some circumstances (e.g., volunteering), even when there is no extrinsic payoff ( α = β = 0), agents are still willing to supply e ∗ i > 0 , implying that, over the interval [0 , e ∗ i ), marginal utility of effort is positive. In the absence of an intrinsic incentive (i.e., ψ = 0), the agent derives no intrinsic returns to effort, even if he has intrinsic taste θEi > 0 for the effort task. In other words, only inasmuch as the effort task is observable to the agent can he derive hedonic benefit from it. With the high-powered intrinsic incentive ( j = h = 1), marginal intrinsic benefit is positive if θIi > 0 , negative if θIi < 0 , and 0 if θIi = 0. The agent maximizes Equation 1 with respect to ei , such that:

e ∗ ij = θiE β + θIi ψj (2) This simple first-order condition illustrates the complementarity between intrinsic incentives and intrinsic motivation that this experiment tests. It implies that intrinsic incentives will lead to higher effort for intrinsically motivated agents, but may have no effect or even reduce effort for those who are intrinsically unmotivated. As such, the aggregate effect of an intrinsic incentive depends on the distribution of intrinsic motivation in the agent population, which implies that the principal can increase her utility both by selecting for agents whose intrinsic preferences align with the effort task and by providing such agents with strong intrinsic incentives.

3 Context

3.1 Program context

In 2005, the Government of India launched the Accredited Social Health Activist (ASHA) program, a nationwide effort to improve health services at the community level, especially in rural regions.^11 ASHAs are female community health workers who are selected by local village councils to provide for the health needs of the villages in which they reside.^12 Each ASHA is assigned a discrete (^11) As of 2014, 828,000 ASHAs had been recruited across IndiaGovernment of India (2015). (^12) Job qualifications include: female gender; married, widowed, or divorced status (due to patrilocality); grade 8 education or higher; age between 25 to 45 years; and preferably, literacy.

catchment area, which typically corresponds to a village or group of villages with a population of roughly 1,000 people. The ASHA’s primary job tasks revolve around supporting and counseling women during preg- nancy, childbirth, and the postpartum period.^13 The typical sequence of serving a pregnant client proceeds as follows. When the ASHA learns of a new pregnancy in her village, she visits the woman and offers to support the woman. If the woman accepts, the ASHA registers the client using a phone-based record-keeping tool.^14 Over the course of the pregnancy, ASHAs are expected to visit the client at home at least monthly. During these visits, ASHAs carry out a variety of tasks: counseling on nutrition, physical activity, and other day-to-day aspects of pregnancy; counseling on identifying pregnancy-related danger signs requiring urgent medical attention; encouraging the client to obtain facility-based antenatal care; providing iron and folic acid supplements; working with the client to develop a birth plan, which includes calculating the estimated delivery date, ad- vising the client on local health facilities for delivery, identifying means of transport, and engaging family support; and updating the client’s maternal health card. The ASHA records and submits these follow-up visits using her phone-based tool. At the time of labor, the ASHA is expected to accompany the client to a health center or hospital and remain with her throughout labor and delivery; the ASHA’s payment, as discussed below, is contingent upon this presence. After the mother is discharged, the ASHA visits the mother and the child several times over the ensuing six weeks to monitor their health and counsel on newborn care, breastfeeding, family planning, and immunizations. At six weeks postpartum, the ASHA “discharges” the client from care. The ASHA job is typically not a full-time position, and ASHAs do not receive salaries.^15 Instead, they are paid piece rates for discrete activities such as facilitating institutional delivery, assisting (^13) In 2006, an expert group convened by The Lancet identified one “overwhelming priority strategy” for reducing maternal deaths: “promoting delivery in primary-level institutions (health centers), backed up by access to referral- level facilities,” as opposed to home delivery (Campbell & Graham, 2006). 14 The mobile phone tool is not a feature of the national ASHA program but, rather, the program site in which this experiment takes place. See Section 3.2. 15 In the study population in the pre-experimental period, the average ASHA earned USD 372 in total annual ASHA payments. For context, auxiliary nurse-midwives (ANMs), the supervisory cadre directly above ASHAs and the actual providers of facility-based antenatal care services, earn approximately USD 2,000 per year. Anganwadi workers (AWWs)—child health and nutrition workers who are positioned laterally to ASHAs and have similar job qualifications but work full-time—earn approximately USD 880 per year.

with polio immunization campaigns, and mobilizing men and women to undergo sterilization.^16 ASHAs’ chief source of income arises from a federally sponsored conditional cash transfer scheme designed to encourage institutional delivery, called Janani Suraksha Yojana (JSY). In this scheme, pregnant women are paid INR 1,400 (USD 23 at 2015 exchange rates) for delivering in an accredited public or private health facility. In addition, an ASHA who accompanies the woman to the hospital for delivery is paid INR 600 (USD 10). ASHAs are not paid for visits to the client during the antenatal period. Thus, in the absence of intrinsic motivation or other non-pecuniary preferences, antenatal visits are rational only inasmuch as they increase the probability of institutional delivery.

3.2 Program site

With a population of 1.6 million, Kaushambi District is one of 19 (out of 70 total) districts in Uttar Pradesh designated by the state government as “high-focus” in view of its poor development indicators. Its maternal mortality ratio of 442 deaths per 100,000 live births is nearly twice the national average and 30% higher than the state average (Government of India, 2011a); the dis- trict’s neonatal mortality rate is twice the national average; and the district has the second-highest scheduled caste population share in the state (Government of India, 2011a).^17 This experiment takes place in Mooratganj, one of eight sub-districts in Kaushambi, with a population of 193, (Government of India, 2011b). At the time of the experiment launch in 2014, Mooratganj had 145 ASHAs, all of whom had been recruited and trained in 2006-2007 when the ASHA program was rolled out in the district. In 2012, a nongovernmental organization (NGO) established a partnership with the Kaushambi district health office to equip the ASHAs with mobile phones to facilitate their work. The phones contain a software application called CommCare, through which the ASHAs register clients and doc- ument home visits as described above.^18 These records, which are synchronized with a cloud-based server, provide the data on which the self-tracking app functions. By the time of the experiment (^16) For example, if an infant registered by the ASHA completes his or her complete course of routine immunizations, the ASHA receives a payment of INR 150 (USD 2.50). 17 Scheduled castes are castes designated in the Constitution of India as historically disadvantaged. Nationwide, 17% of the Indian population belongs to a scheduled caste; in Kaushambi, the proportion is 36%. 18 The application is developed by a US-based company called Dimagi, Inc.

launch, the ASHAs had been using CommCare for 15 months. As all of the ASHAs in the exper- iment were trained to use CommCare, this experiment is not designed to evaluate the underlying work technology. Table 1, Panel A shows descriptive statistics for the performance of the 72 ASHAs in the control group over the pre- and post-intervention periods. Despite the expectation that ASHAs should visit all clients on a monthly basis, the average control ASHA visits 46% of her 13.1 pregnant client s in the average month. Panel B shows descriptive statistics for a limited set of client characteristics and health outcomes as reported by ASHAs. Of note, the 77% institutional delivery rate is substantially higher than the 22% rate for rural Uttar Pradesh reported in the most recent (2005-2006) wave of the National Family Health Survey (International Institute for Population Sciences & Macro International, 2007); though possibly overstated due to selection (both in registration of pregnant women and in reporting of outcomes), this statistic is consistent with a broad increase in institu- tional delivery that has been observed across India since the introduction of the JSY conditional cash transfer scheme in 2005 (Lim et al., 2010).

4 Experimental design

4.1 Experimental interventions

I create a novel mobile phone-based “self-tracking app” designed to enhance the intrinsic utility that ASHAs derive from providing care to pregnant women.^19 In this paper, I mean “intrinsic” both in the sense used in the psychological literature (motivation to do a task for its own sake—effort as its own reward) and in the sense of prosocial preferences (motivation to do a task for its positive social externalities). Both types of motivation stand in contrast to extrinsic motivation, which may be rooted in individualistic (income, job security, career advancement) or social (status, reputation, recognition) concerns, but which, in either case, regards effort as palatable only for the extrinsic rewards it earns. (^19) Credit for the technological development of the app is due to Brian DeRenzi.

A key challenge in assessing the mechanism of an intrinsic incentive is that it may also operate via these extrinsic channels. This is a partly a matter of design and partly an empirical question; I address both in the analysis of effects and mechanisms in Section 5. Another challenge is that an incentive may function as a production technology; i.e., it may increase output not only by increasing effort but also by increasing the productivity of effort, Y ( e ). For example, an app that increases the frequency with which an ASHA interacts with her work phone may make her more adept at using the phone, which in turn may increase her productivity.^20 To mitigate this confound, I create an analogous app that mimics the user interface of the self-tracking app, but replaces its content with information that is, a priori , expected to be lower-powered as an intrinsic incentive. Because the two apps’ interfaces are identical, treatment effects on take-up and performance can be attributed to the information content of the apps as opposed to putative motivational or learning effects of the technology by which the information is delivered. There remains the possibility that the information content could alter the production function; I address this in Section 5.

4.1.1 High-powered intrinsic incentive: Self-tracking

The self-tracking app enables ASHAs to access data visualizations of their work performance.^21 The data are compiled from the ASHAs’ own submissions via their phone-based reporting tool. With one exception (the relative performance graph described below), all of the information contained in the visualizations is generated by the ASHA herself, which highlights the notion of “self-tracking,” as opposed to performance feedback, in which the agent is provided with information that is not observable to her. The app has four screens: a menu, a relative performance page, a calendar page, and a historical trends page. (^20) Such an app would need not deliver performance-related information—e.g., a mobile phone game that confers learning benefits (e.g., how to use the phone) that could increase job-specific ability (e.g., how to fill out forms on the phone). 21 The app is similar in concept to consumer-oriented mobile apps and wearable technologies that enable users to track and visualize data related to personal activities—e.g., Fitbit, Runkeeper, Apple Watch.

Figure 1: Self-tracking app: menu page (Hindi sample and English mock-up).

Menu (Figure 1). The menu page serves as the entry point for accessing the tool. It is accessed by pressing a pre-specified button on the phone. The menu displays a date interval spanning the first day of the current month to the current day. The calendar month is the primary performance interval for ASHAs; the official ASHA program guideline is that ASHAs should visit all of their pregnant clients at least once per month. Below the date interval are three rows that link to the three other pages, which the ASHA can select using the phone’s navigation buttons.

Figure 2: Self-tracking app: relative performance page.

Relative performance (Figure 2). This page features an ordered bar graph of the number of unique pregnant clients visited in the current month, for the ASHA and 15 other ASHAs. The peer ASHAs are chosen randomly from among the 73 ASHAs who belong to the self-tracking treatment condition. The ASHAs are not told the identities of the peers. Anonymity is important for both ethical and theoretical reasons—ethically, to avoid harmful repercussions that may result from publicizing individual ASHAs’ performance, and theoretically, to unlink the informational effect of social comparison from external peer effects, such as these very repercussions. Each ASHA’s peer

group is randomly chosen, such that all peer groups are asymptotically identical but individually unique.^22 In order to avoid serial correlation effects, all peer groups are redrawn at the beginning of each month. The graph updates in response to changes in the ego ASHA’s performance, as well as the performance of peer ASHAs. The relative performance graph is designed to provide an anonymous social benchmark that makes the information signal about performance more useful.^23 Conceptually, the relationship between social comparison and intrinsic and extrinsic motivation is nuanced. On one hand, pub- licly identifiable social comparison would be expected to interact with externally-oriented social preferences such as tastes for status and recognition. Private social comparison, however, does not engage these external preferences; status, for example, cannot be conferred on an anonymous entity.^24 On the other hand, private social comparison can interact with internally-oriented social preferences, such as competitiveness and taste for winning, or preferences that may not involve so- cial interaction but may be mediated by social information, such as social norms regarding private behavior. If these preferences are in relation to an abstract reference group and do not implicate social interaction, they more closely align with standard definitions of intrinsic rather than extrinsic motivation.^25 (^22) In consequence, each peer group has, in expectation, the same average performance, but any given peer group may be higher- or lower-performing than average (and this may fluctuate on a day-to-day basis). 23 In a different study population (such as students taking an exam), analogous information could be conveyed by disclosing only the group mean, which is also anonymous. In the ASHA population, piloting exercises revealed that the concept of “average” was difficult to convey to many ASHAs, and thus the visual display format was adopted. 24 In the absence of peers, motivation around managing one’s self-image could affect effort—e.g., an ASHA who is motivated to do ASHA work because she self-identifies as a prosocial type. However, whereas social image motivation (in which an agent has preferences over how others attribute her behavior) implicates social preferences, self-image motivation can exist in the absence of social preferences, and thus I classify this as a type of intrinsic, not extrinsic, motivation, consistent with Benabou & Tirole (2006); Bénabou & Tirole (2011). 25 To illustrate, consider a competitive athlete. While preferences for status and recognition may play a role in driving the athlete’s effort, there may also be a private component—a desire to excel—that is intrinsic. The athlete might set a performance goal that is absolute (running a mile in less than five minutes), relative to her own performance (setting a personal record), or relative to others’ performance (setting a world record). In each case, good performance can be an end in itself, not an instrument for attaining other rewards. Relative social benchmarks, a type of social norm, have been shown to exert substantial influence on behavior even when the behavior is private and does not entail social interaction, such as the case of home energy consumption (Allcott, 2011; Schultz et al., 2007).

Figure 3: Self-tracking app: calendar page.

Calendar (Figure 3). The remaining two pages are more straightforward. The calendar page displays a calendar in the standard seven-day format, which is the local convention.^26 On each day, if any visits to pregnant clients occur, the number of visits is indicated in a circled number. A counting rule restricts the number of times an ASHA can get credit for visiting a given client in a given day to one, but there is no restriction on the number of times an ASHA can report visiting a client in a given month.

Figure 4: Self-tracking app: historical page.

Historical (Figure 4). This page displays a line graph plotting the total number of clients visited each month for the current month (to date) and the five preceding months, in a rolling manner. Taken together, the self-tracking app provides the ASHA with the ability to access personalized information about her performance. The data elements conveyed include: total caseload, number of clients visited in the current month, number of clients visited each day in the current month, number (^26) Sundays and national holidays are shaded red; the current day is shaded black.

of visits conducted in the current month, number of clients visited monthly over the previous six months, and visit performance relative to an anonymous peer group. All of these elements are meant to make the effort task and its prosocial impact more psychologically salient to the ASHA, thereby, in theory, amplifying the marginal intrinsic utility of effort.

4.1.2 Low-powered intrinsic incentive: Generic encouragement

One counterfactual to the self-tracking app could be to continue the status quo in which ASHAs are not given access to any additional app on their phones. However, this would introduce a mechanism confound; any treatment effect could be due to the interactive nature of the app, rather than its information content.The intent of the experiment is to focus on the incentive effect of information itself. To this end, I develop a counterfactual intervention that preserves the technological interface of the self-tracking app, but provides generic information that is putatively less effective at making the effort task and its prosocial impact salient. The low-powered app also has four pages; sample pages are shown in Figure 5. The content includes generic encouragement messages. A different set of pages is generated on a daily basis, out of an inventory of 52 sets. The menu page points to three further pages:“Responsibilities of an ASHA,” “Advice for Healthy Mothers & Babies,” and “Inspiring Quotes.” Each page includes a statement accompanied by a picture that illustrates the statement. Content for the first two sections is drawn from ASHA training materials. The “Inspiring Quotes” section contains quotes drawn from Hindi-language websites, with many attributed to well-known South Asian cultural figures such as Gandhi and Mother Theresa.

4.1.3 Audio service

In the study sample, 28% of the ASHAs are illiterate.^27 To ensure that the interventions would be useful to these ASHAs, the research team developed an automated audio version of each interven- tion. Analogous to the act of accessing the app on the phone, an ASHA calls a designated phone (^27) Each ASHA was asked to read a Hindi sentence which stated, “The woman went to the market to buy vegetables.” Twenty-eight percent were able to read no words or only a few words; the remainder were able to read all of thewords and are classified as literate.

number from her work phone. An automated recording then reads aloud the information contained in the ASHA’s visual app. There is no limit on how often the ASHA can utilize the audio service. While there are differences in the user experience of the audio and the visual software interface (e.g., navigation is possible only in the latter), the underlying information content is the same. Thus, the theoretical framework is unaffected, and in the empirical analysis, unless otherwise stated, I combine usage of the visual and audio systems to yield a composite measure of take-up.

4.2 Randomization, implementation, and data sources

The 145 ASHAs in Mooratganj were randomly assigned to one of the two intrinsic incentive treat- ments: self-tracking and generic encouragement. Randomization was conducted in May 2014, one month before the launch of the experiment, and was stratified by six variables: Hindi literacy, total client visits conducted over the prior 4 and 12 months, respectively, and scores on three psy- chometric scales for extrinsic, intrinsic, and prosocial motivation, respectively.28,29^ In June 2014, the self-tracking and generic encouragement apps were installed on ASHAs’ phones, and all 145 ASHAs were trained in their use. All ASHAs were told that two different phone-based tools were being piloted, and that the eventual plan was to make both available if so desired by ASHAs. No complaints were raised to the research team or the implementing NGO during the training or at any point thereafter regarding the randomization. Once the experiment launched, care was taken to preserve the intent of the intervention: to alter the intrinsic utility of effort without altering real or perceived extrinsic returns to effort.To avoid a potential monitoring effect, no efforts were made to affect demand for the apps through, e.g., routine follow-up visits, marketing, or interactions during other program activities such as trainings.^30 (^28) Literacy was directly measured during the baseline survey by asking the ASHAs to read a Hindi sentence. The two visit measures were included to capture both short- and medium-term baseline work performance. The psychometric scales are described in detail in Section A. 29 I used the “T-min-max” method with 1,000 draws to carry out the randomization. For a discussion of this randomization method, see Bruhn & Mckenzie (2009). 30 The one partial exception was an automated SMS system created by the research team, in which a text message is sent to each ASHA every Monday stating either (for self-tracking app users), “Your visits information is available. Please press the shortcut button to access your information,” or (for encouragement app users), “Your advice and encouragement is available. Please press the shortcut button to access your encouragement.” ASHAs were told that these messages were sent automatically to all ASHAs. SMS use is low overall in this setting, and we find no significant Monday fixed effect for app usage.

Instead, the research team interacted with ASHAs only when an ASHA called a research assistant to request assistance with the self-tracking or encouragement app. Such assistance variously involved re-installing the software in the case of accidental deletion, re-saving the phone number used to access the audio service, and addressing questions regarding interpreting the information in the app. For the first eight months of the experiment, the research team had no contact with ASHAs other than through these troubleshooting visits. At nine months of follow-up, a midline survey was administered, in which 142 out of 145 ASHAs participated. In the course of each ASHA’s interview, the research team rechecked all phone settings related to the use of the visual and audio services, fixed settings as necessary, and documented any steps taken. In addition to the baseline and midline survey data, this experiment relies on performance data reported by the ASHAs through their phone-based record-keeping tool, and app usage data measured directly. Client visits and app sessions are timestamped with start and end times. Client visits are tied to individual clients, allowing differentiation of initial registration visits and follow-up visits. These and other data sources are described in further detail in Appendix Section A.

5 Analysis

5.1 Randomization balance

Tables 2 and 3 report tests of equality for variables measured at baseline for ASHAs and households, respectively. Table 2 reports baseline data across four ASHA-related domains: job performance, job-specific ability, psychological traits, and demographic characteristics. None of the variables is independently significant, and together they are not jointly significant. Similarly, Table 2 shows baseline data from a household survey that was conducted by the partner NGO in May and June 2014, prior to the launch of the experiment. Only the difference in household size is marginally significant, with households in the self-tracking treatment being 7% larger on average, though they statistically have the same number of children under five. Importantly, when comparing data sources, baseline rates of ASHA visits and institutional delivery in the household survey

corroborate rates self-reported by ASHAs (cf. 1; e.g., 77% vs. 70% institutional delivery rate in data collected from ASHA reports and the household survey, respectively). Some of this difference is likely attributable to selection—that is, ASHAs submit data only for registered clients, whereas the household survey samples all pregnancies. In addition, I run t -tests for differences in means between the two treatment conditions for all 263 numeric variables in the baseline ASHA survey. Of these, 4.9% of the differences are significant at the 10% level; 1.9% at the 5% level; and 0.4% at the 1% level. Repeating this exercise for daily home visits for all 455 days in the pre-experimental period, t -tests reveal that 6.9%, 2.5%, and 0.5% of daily visit counts are significantly different at the 10%, 5%, and 1% levels, respectively. Taken together, these results indicate failure to reject the null hypothesis that the two treatment groups are identical on observables. In the analysis that follows, to the extent possible, baseline differences are controlled for, either with explicit covariates or with fixed-effect estimators.

5.2 Empirical strategy and average treatment effects

To identify causal effects, I exploit the fact that performance is observed at the daily (or, in the case of earnings, monthly) level. This allows for panel analysis with ASHA fixed effects that control for all time-invariant ASHA characteristics. In addition, the availability of pre-experimental data allows for differencing out time-variant, ASHA-specific trends. I estimate the general equation:

yit = β 0 + β 1 At + β 2 At ∗ Ti + αi + Zmγm + uit (3)

where yit is an outcome of interest for ASHA i at time t ; At is an indicator for whether time t is after the launch of the experiment; Ti is an indicator for treatment assignment that takes value 0 in the generic encouragement condition and 1 in the self-tracking condition; αi is a vector of ASHA fixed effects; Zm is a vector of month-year fixed effects; and uit is the error term. I assume that errors are serially correlated and thus present standard errors clustered at the ASHA level throughout.

Randomization ensures, in expectation, that Ti ⊥ uit , and that trends in Yit between the two treatment groups, in the absence of treatment, would have been parallel. Furthermore, spillovers between treatments are unlikely given that ASHAs live in different villages and cover defined, non-overlapping catchment areas and thus are unlikely to experience spillovers via either market demand or through direct interactions. Finally, there is minimal attrition during the follow-up period of the experiment. Of the 145 ASHAs, all participate in app training, and at twelve months of follow-up, only 3 ASHAs—one in the generic encouragement condition and two in the self- tracking condition—are no longer working (all three due to out-migration). Under these identifying assumptions, β 2 measures the average causal effect of the self-tracking app relative to the generic encouragement app.

5.2.1 Average treatment effects on take-up and client visits

Table 4 presents average treatment effects on take-up of the interventions, on client visits, and on earnings. Columns 1-3 show results for take-up of the main visual software, the complementary audio service, and total use of the two modalities combined. All estimates are in units of sessions per day. Three findings are noteworthy. First, despite an experimental protocol that provides little explicit encouragement to use the interventions, demand for both apps is high. Over the course of the one-year experimental period, the average ASHA in the self-tracking (generic encouragement) treatment uses her app once every 3.89 (3.22) days. To contextualize this, client visits occur every 3.78 and 4.67 days in the self-tracking and generic encouragement conditions, respectively, during the experimental period. Thus, in both conditions, app usage is higher than visit frequency, suggesting that use of the phone for visits is not exclusively driving use of the apps.^31 Second, the fact that take-up of the two apps is similar (with the point estimate favoring the generic encouragement app) suggests that any treatment effects favoring the self-tracking app cannot be explained purely by differences in how often each app is used (e.g., a learning-by-doing effect that makes frequent phone users more efficient at filling out forms). (^31) That take-up of the the visual software is higher for the generic encouragement app than for the self-tracking app is likely because the encouragement app provides new content daily, whereas the self-tracking app is informationally static in the absence of visits, which occur less than daily for all ASHAs.

The remaining columns of Table 4 examine two measures of performance: visits per day and earnings per month. Partitioning visits into the initial registration visit and follow-up visits to a given client, the experiment reveals no effect of the self-tracking app on the former (Column 4), but a 33.3% increase in reported follow-up visits in the self-tracking treatment (Column 5). Driven by this effect, self-tracking ASHAs report 23.8% more total visits than their counterparts (Column 6). To investigate the time pattern of this effect, Column 7 divides the 12-month experimental period into two six-month halves. The treatment effect is concentrated in the first half of the experimental period. Figure 6 illustrates this finding graphically by plotting the treatment effect on total visits for each month of the experiment; the effect is largest in the first four months of the experiment, though all monthly point estimates remain positive for the duration of the experiment. Finally, reinforcing these results as well as the intuition that follow-up visits reflect intrinsically more so than extrinsically motivated effort, the self-tracking app has no impact on earnings (Column 8). Thus, assuming rational expectations and accurate measurement of earnings, the increase in follow-up visits in the self-tracking app treatment cannot be explained by financial preferences alone.

5.3 Heterogeneous treatment effects

The theoretical framework in Section 2 predicts complementarity between intrinsic incentives and intrinsic motivation and that, in the presence of intrinsic aversion, an intrinsic incentive may dampen effort. In this section, I assess both predictions. Table 5 reports the estimates of

ln ( yit ) = β 0 + β 1 At + β 2 AtTi + β 3 Atln ( θki ) + β 4 AtTiln ( θki ) + αi + Zmγm + uit (4)

where ln ( yit ) is log total visits reported by ASHA i on day t ; At is the post-intervention indicator; Ti is the treatment indicator; ln ( θki ) is the log motivation of ASHA i with respect to psychometric dimension k ∈ {extrinsic, intrinsic, prosocial, social desirability, competitive}; αi is a vector of ASHA fixed effects; Zm is a vector of month-year dummies; and uit is the error term. The coefficient

of interest is β 4 , which is the marginal elasticity of output with respect to psychometric trait θk in the self-tracking treatment, relative to that in the generic encouragement treatment. In other words, it is the difference in elasticity between two curves—those between total visits and θk^ in the self-tracking treatment and in the generic encouragement treatment, respectively—and it measures the self-tracking app’s degree of complementarity with a given motivational trait relative to that of the generic encouragement app. In Table 5, each row is a fixed-effects regression of Equation 4, where only the row variable changes. Column 1 (which reports estimates of β 3 in the above equation) shows that, in the control condition, intrinsic and prosocial motivation have negative elasticities of effort as measured by total visits. That is, a 1% increase in intrinsic (prosocial) motivation is associated with a 0.48% (0.44%) decrease in total visits per day. This itself is not remarkable, as motivation is not exogenous and could be correlated with other traits (e.g., ability) that affect performance. More important is the finding in Column 2 that the intrinsic and prosocial motivation elasticities of effort are significantly higher in the self-tracking condition than in the generic encouragement condition. In other words, the self-tracking condition is more effective at eliciting performance the more intrinsically/prosocially motivated an ASHA is; it leverages intrinsic/prosocial motivation. This relationship does not hold for extrinsic motivation, and as two additional placebo tests, it does not hold for social desirability or competitive motivation.^32 The two placebo traits are plausible confounds: the self-tracking app could in theory leverage social desirability or competitiveness, but this is not observed in the data. Table 6 further characterizes these results by partitioning the sample into terciles of psychome- tric traits and estimating Equation 3 for each tercile. Specifications 1, 3, and 5 use a pooled causal estimator for the entire post-intervention period, whereas Specifications 2, 4, and 6 estimate each tercile-specific treatment effect in the first half of the experiment, as well as the change in treatment effect in the second half. The estimates in the odd-numbered columns for intrinsic and prosocial motivation show that the self-tracking app has no effect on total visits in the least motivated tercile of ASHAs with respect to each trait. In contrast, the self-tracking app has positive effects in the (^32) Both of these psychometric scales were administered at baseline along with the other psychometric traits. The scale items are listed in the Appendix.

middle and top terciles of each trait. The largest treatment effect, which is observed for the top tercile of intrinsic motivation, is 41.4% of the mean in the generic encouragement condition. While the point estimates of the treatment effects for each tercile of extrinsic motivation have a positive slope, it is not significant, as was shown in Table 5. The even-numbered specifications in Table 6 illustrate how the treatment effects heterogeneously evolve over time. Across all three motivational traits, the least motivated tercile exhibits a positive but imprecisely estimated treatment effect in the first half of the experiment, and this effect decays to zero in the second half. A similar decay is observed in the middle tercile of extrinsic motiva- tion, whereas those who are most extrinsically motivated do not respond to the self-tracking app either at the outset or as the experiment proceeds. In contrast, in the middle and top terciles of intrinsic and prosocial motivation, the treatment effect is positive and precisely estimated in the first half of the experiment and persists through the second half, showing no decay. That is, the most intrinsically/prosocially motivated ASHAs respond to the self-tracking app from the outset, and their response persists, whereas less intrinsically/prosocially motivated ASHAs respond at the outset but only transiently. Taken together, this analysis of average and distributional effects suggests three main findings: the self-tracking app treatment leads to an average increase in client visits; consistent with theory, this effect interacts positively with intrinsic motivation; and this complementarity is driven by a treatment effect that is both larger and more sustained among more intrinsically motivated workers. As to the prediction that a high-powered intrinsic incentive may dampen effort for those who are intrinsically unmotivated, the results do demonstrate a decay in client visits over time among the least intrinsically/prosocially motivated ASHAs, suggesting the possibility that, had the experiment continued, their performance may have dipped below that of their counterparts in the control condition.

5.4 Compensating mechanisms

In the remainder of this sub-section, I examine the validity of the experiment’s main findings and the impact of the experiment on health outcomes.

Table 7 examines whether the effect of the self-tracking app on reported visits is compensated by negative effects on other measures of effort. One such mechanism could be socially inefficient allocation of effort both across space—e.g., visiting easy-to-reach clients many times, to the exclu- sion of other clients—and across time—e.g., visiting a given client multiple times in some months but none in others. Column 1 reports estimates of a month-level panel regression in which the dependent variable is the share of pregnant clients visited by an ASHA in a given month. During the experimental period, self-tracking ASHAs visit an 8.31-percentage point greater share of their clients each month, a 20.8% improvement over the control mean. That this effect is approximately equal to the treatment effect on total visits indicates that increased visits do not occur at the expense of coverage. Figure 7b graphically illustrates this increase in client coverage. Columns 2-5 test for fabrication of visits. How long an ASHA spends completing a visit form is tracked by the phone directly, from the time she opens the form to when she completes it. Shorter visit duration may reflect higher productivity—e.g., greater proficiency at typing on the phone. Nevertheless, it also raises concern for fabrication. Column 4 shows that self-tracking ASHAs spend 13.5% less time filling out forms; this effect is similar when considering new client visits and follow-up visits separately (Columns 2-3). Taken together, these results suggest two complementary interpretations. First, although average visit duration is 13.5% shorter, because self-tracking CHAs report 26.8% more total visits, they spend more aggregate time “inside” forms. While this does not rule out fabrication, it makes pure fabrication less plausible, since that would be expected to be a time-saving strategy manifesting as less aggregate time completing (fake) forms. Second, similar effects on duration are observed for both initial and follow-up visits. This too casts doubt on the extent of fabrication, as the nature of the two types of visits is that it is much easier to fabricate a follow-up visit form (which consists of checking off a list of counseling topics, which is difficult to verify) than an initial registration form (which requires typing the name of a client, her husband, her phone number, and so forth, all of which can be verified).

5.5 Health impacts

Finally, Table 7, Columns 5-12 assesses for treatment effects on four health practices and outcomes at the client level, as reported by ASHAs. These include practices such as attending antenatal care visits and receiving tetanus vaccinations, as well as pregnancy outcomes such as institutional delivery and maternal death. Columns 5-12 estimate

yijt = β 0 + β 1 Tj + β 2 Et + β 3 Tj ∗ Et + Ziθ + αj + uijt (5)

where yijt is the outcome of client i of ASHA j during period t ; Tj is ASHA j ’s treatment assignment, where Tj = 1 for the self-tracking treatment; Et denotes whether the client’s pregnancy was exposed to the experimental period; Zi is a vector of client-level controls; and αj is a vector of ASHA dummies. The estimation sample is restricted to clients who have completed their pregnancy. Et defines as exposed those clients who were registered by an ASHA prior to the launch of the experiment (so as to ensure no endogenous selection) but whose pregnancy concluded after the launch ( N = 1 , 779 ). Those who began and ended their pregnancy before the launch of the experiment are classified as non-exposed ( N = 4 , 820 ). The difference-in-differences estimator of interest is β 3 , which, under the parallel trends assumption, identifies the effect of the self-tracking app on client outcomes. Standard errors are clustered at the ASHA level. In brief, the results are equivocal. Regarding antenatal practices, the self-tracking condition leads to an increase in reported ANC visits but has no effect on ASHAs’ reports of how many tetanus vaccines the client received and whether the client has developed a birth plan. The average effect on these practices is not significantly different from zero (Column 8). Regarding pregnancy outcomes, these too are reported by ASHAs, but using a separate form at the conclusion of the pregnancy. Whether the ASHA submits a pregnancy outcome form is itself an important outcome, as it is likely to be (negatively) correlated with whether the client has been lost to follow-up. On this margin, we observe no significant effect (Column 9). Surprisingly, Column 10 shows that, conditional on delivery, the probability of institutional delivery is 4.8 percentage points lower in the self-tracking group than in the generic encouragement group, in which 78.8%