Generosity and Prosocial behavior in Health Care. Provision: Evidence from the Laboratory and Field

Transcription

1 Generosity and Prosocial behavior in Health Care Provision: Evidence from the Laboratory and Field J. Michelle Brock Andreas Lange Kenneth L. Leonard May 19, 2014 JEL Classification: I15, O19, C91, C93, J2 Keywords: Intrinsic Motivation, Health Care Quality, Altruism, Professionalism, Tanzania, Experimental Economics, Hawthorne Effect, Encouragement Effect, Study Effect. This work was funded by a Maryland Agricultural Extension Station seed grant, a contract from the Human Resources for Health Group of the World Bank, in part funded by the Government of Norway, and the Eunice Kennedy Shriver National Center for Child Health and Human Development grant R24-HD through the Maryland Population Research Center and a support from the Knowledge Product Human Resources for Health group at the World Bank. We are grateful for the support of the Center for Educational Health, Arusha (CEDHA), specifically Dr. Melkiory Masatu and Dr. Beatus Leon. We thank participants at Michigan State, University of Minnesota and ESA for their comments. European Bank for Reconstruction and Development, London, UK University of Hamburg, Department of Economics, Von Melle Park 5, Hamburg, Germany Corresponding author: 2200 Symons Hall, University of Maryland College Park, MD kleonard@arec.umd.edu

2 Abstract Do health workers (sometimes) have intrinsic motivation to help their patients? We examine the correlation between the generosity of clinicians as measured in a laboratory experiment and the quality of care in their normal practices under three different intrinsic incentive schemes. Specifically, we observe clinicians 1) in their normal work environment, 2) when a peer observes them and 3) 6 weeks after an encouragement visit from a peer. Clinicians who give at least half of their endowment to a stranger in the laboratory (generous) provide 8% better quality care. In addition, the average clinician provides about 3% better quality when observed by a peer and 8% higher quality care after the encouragement visit. Importantly, we find that generous clinicians react to peer scrutiny and encouragement in the same way as non-generous clinicians. Many clinicians are intrinsically motivated to provide higher quality care, however most clinicians respond to increased prosocial incentives in the form of scrutiny and encouragement from peers.

3 1 Introduction Health care workers are commonly described as being intrinsically motivated and the literature on health care is full of references to prosociality terms like professionalism, esteem, and caring. At the same time, all health systems invest significant resources in regulation and quality assurance, thereby declining to leave quality up to the caring instincts of providers. Furthermore, where regulation is weak, quality is also often low (Das and Hammer, 2007; Das et al., 2008; Rowe et al., 2005). In particular, significant attention is paid to the know-do gap the gap between what health workers know how to do and what they actually do for their patients (Leonard and Masatu, 2010b; Maestad and Torsvik, 2008; Maestad et al., 2010; WHO, 2005). Thus, in these settings, health workers could choose to do more for their patients and quality is low, in part, because health workers are not sufficiently motivated to provide adequate effort. Does this mean that intrinsic motivations the caring instincts are not present or strong enough and that one must look exclusively to extrinsic incentives to motivate quality? 1 Or should policy makers refocus their efforts on finding intrinsically motivated health workers who will provide adequate effort without extrinsic incentives? In this paper, we examine evidence on a particular type of intrinsic motivation. We specifically consider prosocial preferences and prosocial incentives as sources of motivation. Here, prosocial means that an individual has an outward orientation such that the welfare and/or opinion of others enter into his or her utility. Prosocial preferences can motivate an individual to take costly actions that benefit (or harm) others (i.e. altruism, positive or negative reciprocity). They can also motivate esteem seeking. Esteem seeking can be categorized as a type of prosocial behavior because the individual who is seeking an esteem payoff acts on another s behalf, or according to another s preferences/biases, to gain esteem. Unlike in the case of altruism, the purveyor of the esteem may not gain anything from the interaction. Prosocial incentives are the features of the decision making environment that increase the util- 1 There has been increased focus on extrinsic motivation using monetary incentives (see Basinga et al., 2011; Meessen et al., 2006, for example). 3

4 ity individuals can earn from prosocial behavior. The same way that the wage is the extrinsic incentive for some actions, prosocial incentives such as being observed or appreciated by others increase the return to prosocial actions. We argue that some health care providers are inherently altruistic, but that the setting in which a worker practices, defined by prosocial incentives, is potentially more influential on effort choice. Using data drawn from a laboratory experiment and the field, we show evidence that altruistic health workers, as defined by behavior in the lab environment, provide higher quality care (exert more effort) for their patients. In addition, we show that changing the workplace environment to provide greater prosocial incentives increases effort (and therefore quality) for all types of health workers. To test the relative importance of types and environments we examine the behavior of health workers who provide outpatient care (clinicians) in urban and peri-urban Arusha region of Tanzania. We look at four different settings, each with different implied prosocial returns to effort. First, we examine the performance of the clinicians in their normal workplace (baseline). Second, we measure their performance when there is a peer present to observe their activities (scrutiny). Third, we measure their effort after participation in a trial in which a Tanzanian MD reads an encouraging statement and asks them to improve their performance on five specific items (encouragement). Finally, we examine clinicians in an economic laboratory experiment, examining their willingness to sacrifice on behalf of strangers using the dictator game. The comparison of clinicians behaviors across these four settings allows us to evaluate the role of intrinsic and prosocial motivations. We use the laboratory experiment to distinguish clinicians who are generous to strangers in that setting and compare their performance in their normal workplace to clinicians who are not generous in the laboratory experiment. By comparing the quality of care (as measured by effort exerted to adhere to protocol items required by the patients symptoms) in the three different clinical environments (baseline, scrutiny and encouragement) we can evaluate the response of all clinicians to the prosocial incentives implied by peer scrutiny and encouragement. And finally, we can compare the differential response of generous and non-generous clinicians to the changes in prosocial incentives implied by scrutiny and encouragement. 4

5 We find that clinicians who are generous in the laboratory perform better at work. As such, prosocial behavior appears to be linked across different environments; measuring prosocial preferences in the lab allows the classification of particular types of health workers who treat their patients differently than other health workers. In addition, we find that the average health worker responds to changes in prosocial incentives in the workplace. The average clinician increased his or her effort significantly both when subjected to peer scrutiny and when encouraged to provide better care. In the latter case, the improvements are large and significant even eight weeks after clinicians received an encouragement visit. The performance increases under scrutiny or encouragement are similar for generous and non-generous clinicians alike. The results suggest that an underlying degree of prosocial attitudes drives behaviors in the laboratory and the field: willingness to sacrifice own gains for a stranger s benefit implies willingness to exert costly effort on behalf of one s patients. Such a view would lend support to the hypothesis that prosocially motivated types are an important determinant of quality care. However, even those apparently without this source of motivation can be incentivized by scrutiny and encouragement. Doctors who are nicer to their patients than other doctors are not automatically more professional than others and doctors who are not naturally nice to their patients can be encouraged to be more professional. In the following section, we outline the view of prosocial behavior and intrinsic motivation from the management and experimental economics literatures and present a descriptive model of behavior in the health care setting. Section 3 outlines the data and empirical methodology for examining the data. Section 4 shows the results and section 5 discusses the implications and provides our conclusions. 2 Intrinsic Motivation and Prosocial Behavior The term intrinsic motivation takes on different meanings in different literatures. In the psychology literature, it literally refers only to the enjoyment of doing the job as distinct from enjoying that 5

6 others may benefit. Grant (2008) offers the example of a professor who enjoys the performance of lecturing (intrinsic) as opposed to a professor who enjoys seeing students learn (prosocial). Most of the literature in the behavioral economics field has focused on prosocial behavior, perhaps because it is easier to measure in the laboratory. The health care literature, in contrast, uses the term intrinsic as an umbrella term for both strict intrinsic and prosocial motivation. Thus, we use the term prosocial when we are referring to specific forms of motivation (altruism, generosity and esteem seeking) and the term intrinsic to refer to the broad policy related interests of the health field. Where prosocial behavior is discussed in the health care literature, there is little debate about its importance and a number of studies have shown that improved prosocial motivation results in higher quality care (Delfgaauw, 2007; Kolstad, 2013; Prendergast, 2007; Serra et al., 2011). There is some debate on the source of prosocial motivation. Whereas, some recent research suggests that inherently altruistic types are more desirable in health care (Coulter et al., 2007; Smith et al., 2013), the majority of the research currently focuses on inculcating professional ideals, not selecting altruistic individuals. A common feature of this literature is the belief that medical schools can create unconditionally prosocial health workers (Medical School Objectives Working Group, 1999; Wear and Castellani, 2000). 2 In contrast, recent evidence from experimental and behavioral economics can be interpreted to suggest that most prosocial behavior is context dependent (Levitt and List, 2007). Benabou and Tirole (2003) suggest multiple sources of intrinsic benefit, concentrating on happiness (or warm glow (Andreoni, 1989, 1990)) derived from others perceptions, including concerns for social reputation and self-respect. In this formulation, the presence of opportunities to earn social reputation, for example, define the context and the corresponding realization of prosocial behavior. 2 The taught ideals are explicitly altruistic: I will follow that system of regimen which, according to my ability and judgment, I consider for the benefit of my patients (the Hippocratic Oath: Adams, 1849) and The health of those in my care will be my first consideration (Declaration of Geneva: World Medical Association, 1995). 6

7 Importantly, the other can be the recipient of generosity or a witness to the act of generosity. In addition, generosity may be conditional on the identity of the recipient or the individual observing the giving (Ellingsen and Johannesson, 2008): individuals may like being seen as generous in the eyes of specific people. Thus, people s motivation to be kind or generous to others can be thought of coming from internally realized benefits: individuals enjoy seeing the recipient receive something, they enjoy the fact that the recipient knows they gave them something or they enjoy being seen as having given to the recipient. When the other is a peer rather than a patient, we observe a different, but perhaps more important form of prosocial motivation, the desire to be seen as professional. Professionalism is common in settings where dedication to group goals and values must promote service to a greater good (Akerlof and Kranton, 2000, 2005; Cullen, 1978; Freidson, 2001). As with other forms of prosocial behavior, professionalism can be seen as a feature taught in medical school or as a feature of the work environment. For example Leonard and Masatu (2010a) describe a form of latent professionalism in which individuals follow professional norms only when they believe their fellow professionals can observe or evaluate their behavior. In this case, the environmental factor driving alignment with organizational goals is the opinion of the peer. Similarly, Kolstad (2013) demonstrates that access to information comparing own performance to the performance of peers leads to significant improvements in quality for many surgeons. In the health care context, therefore, we can think of two potential sources of intrinsic motivation that may lead to compliance with prosocial norms being correlated with effort in the workplace; patient- and peer-oriented motivation. If health workers care primarily about the welfare of their patients (or if they care to be noticed by their patients), they will seek to do the best thing for their patients or to provide the care that patients seek, respectively. On the other hand, peer-oriented prosocial behavior is a type of professionalism, in which health workers gain utility from following and being seen to follow the norms of their peer group. These two sources of intrinsic motivation define both a type of health worker who should always provide high quality care (patient-oriented health workers) as well as environmental conditions that would increase the intrinsic incentives to 7

8 provide high quality health care (the presence or scrutiny of peers). Note that whereas generous health workers who care about their patients always have patients present and thus always have the chance to feel warm glow, these esteem-seeking health workers may depending on their work environment have the opportunity to garner the esteem of their peers. 2.1 A descriptive model of effort with prosocial incentives We introduce a descriptive model of effort provision to help clarify our empirical investigation. Clinicians provide effort (a) for many reasons, some of which may be described as social preferences. In order to illustrate the different motivational factors, we distinguish monetary motivation (W for wealth) and two types of prosocial motivation: patient-based (M for moral), and esteemseeking (R for reflective). Effort choices will depend on stimuli s i (i {w, m, r}) which may impact the respective motivations: s w can be thought of as the wage, s m as the individual s level of social obligation to others, and s r as exposure to peers. Utility is assumed to be additively separable as in Levitt and List (2007) and Leonard and Masatu (2008): U(a, S) = U w (a, s w ) + U m (a, s m ) + U r (a, s r ) (1) The health worker will choose effort to maximize utility. We denote the optimal effort choice by a ( U a (a, S) = 0). It is standard to assume utility to be concave in the action a ( 2 U i a 2 positive impact on marginal returns from actions ( 2 U i a s i < 0) and that stimuli have a > 0) for all i {w, m, r}. Given these assumptions, it follows that increasing stimuli from any source increases effort ( a s i 0). = 2 U i a s i / 2 U a 2 > In our empirical setting, we can observe effort (a ) but we cannot observe the current levels of wealth, moral stimuli or reflective stimuli. However, we can study the changes in effort due to exogenously increased levels of reflective stimuli ( a / s r ) as a result of increased exposure to peer scrutiny or encouragement. In addition, we use behavior in the lab experiment to define a set 8

9 of clinicians who have strong incentives to provide effort due to altruistic reasons; i.e. have higher responses to moral stimulus than other clinicians. Assuming that this moral stimulus transfers across different environments, we expect that these altruistic doctors also face higher levels of moral stimulus in their normal workplace. This leads us to the following conjectures: Conjecture 1 Clinicians with higher response to moral stimulus as measured in a lab environment will provide higher levels of effort in the field. In addition to moral motivation, a clinician may also respond to the opportunity to be seen by other as conforming to professional standards and thereby respond to reflective stimuli. We examine the marginal impact of increased reflective stimuli by varying the level of scrutiny in our study: Conjecture 2 The average clinician will increase his effort when faced with increased reflective stimuli (exposure to peer scrutiny): a s r > 0. It may also seem natural to expect that the different stimuli interact. That is, when one form of stimuli is high all else equal the gain in effort from increasing another form of stimuli may be expected to be lower ( 2 a s i s i < 0). 3 In other words, when an individual has high incentives to provide effort (for example a high wage or high levels of social obligation) then increases in other stimuli (exposure to peers, for example) should not lead to large increases in effort. On the other hand, when a clinician faces low levels of stimuli overall, increases in any form of stimuli can lead to large increases in effort. We therefore expect: Conjecture 3 Clinicians who provide relatively low levels of effort, possibly because of a low response to moral stimuli, will exhibit greater responses to additional reflective stimuli. That is, a s r is negatively correlated with a across a sample of clinicians. 3 Without additional restrictions on the utility function, this result cannot be derived from our simple model. However, as we show below, our empirical measures of effort have a natural upper bound 100% adherence to protocol and therefore it is natural to think that increases in effort are increasingly costly at higher levels of effort, even with multiple sources of motivation. 9

10 Combining conjectures 1 and 3, we also investigate the following: Conjecture 4 Clinicians with high response to moral stimuli as measured in the lab will respond less to increases in reflective stimuli than clinicians with low response to moral stimuli. We investigate these conjectures using the data explained below. 3 Methodology We studied 103 clinicians who practice health care in Arusha region by collecting data on the quality of care in the course of their normal practices. Sixty-three of these clinicians also participated in a laboratory experiment and this analysis focuses on these workers. 3.1 The Laboratory Experiment The laboratory experiment took place in Arusha, Tanzania in July 2010, after all data had been collected from the field. The subject pool consisted of 71 clinicians 4 and 78 non-clinician subjects. The clinician subjects were recruited with a letter that we sent to all clinicians who had participated in the field study. We recruited non-clinicians subjects with printed advertisements distributed in major market areas in Arusha. While fliers were distributed to a variety of people, the group of non-clinician subjects was ultimately a convenience sample. All of the non-clinician subjects that arrived to participate each day were allowed into the experiment. Clinician subjects were given a per-diem of 35,000 Tsh in addition to what they earned in the experiment. Non-clinician subjects received a show-up fee of 5,000 Tsh. One US dollar is approximately equal to 1,300 Tanzanian 4 Some of the clinicians in the laboratory experiment did not participate in the field study. Also, the gap between data collection ending and the laboratory experiment varies considerably by clinician, since clinicians entered the study on a rolling basis. For some, a year may have passed between the interventions and the laboratory experiment. 10

11 shillings. 5 Clinician subjects gathered in a classroom and non-clinician subjects gathered on a lawn outside of the classroom, near enough that both groups could see each other but far enough that there was no communication or individual identification. This was done to preserve anonymity while ensuring that subjects understood the concept of being paired with another player. Subjects recorded decisions using paper and pen. We provided a hard copy of the experimental instructions to each participant and read them aloud before the experiment began. The instructions explained the basic guidelines of the experiment and how earnings were determined. Subjects were given the chance to ask clarifying questions after the instructions were read. The experiment was a standard dictator game in which the dictator decides how to allocate 100 tokens between him or herself and an anonymous partner. The dictator in each pair was always a clinician and the receiver was always someone drawn from the non-clinician pool. 6 The receiver had no choice but to accept what was given. Each token was worth 150 TSH, so that the clinician was choosing the allocation of 15,000 TSH (approx 12 USD). 3.2 The Field Sample and data collection We collected data on clinician performance for 103 clinicians and 4,512 patients in the semi urban area of Arusha, Tanzania. 7 The field data collection ran from November 2008 until August Clinicians entered the study at different times and the time between enrollment and the final data 5 The imbalance in the show-up fees was never highlighted to participants but could have been inferred. It does parallel the power and income imbalance in a typical clinical encounter. 6 We never used the terms doctor or patient in the experiment, but the clinicians knew they were in a group of clinicians. 7 We sampled 100 percent of the health care facilities in the area with outpatient departments, though some facilities were excluded based on convenience; they were either too difficult to reach for obtaining consent or had too small of a patient volume. 11

12 collection for each clinician was about six and half weeks on average. The sample includes public, private, and non-profit/charitable facilities. The term clinician refers to primary health workers who provide outpatient care. They fill the role of doctor, though the majority of them do not have full medical degrees. 8 On each day of data collection we interviewed all the patients seen in the 4-hour window during which we visited the facility. The interviews with patients followed the Retrospective Consultation Review (RCR) instrument. It is a slightly modified version of the instrument used by Leonard and Masatu (2006) and it measures adherence to protocol. Immediately after their consultation, patients are asked a series of questions about their consultation based on the symptoms that they reported. The questions allow us to reconstruct the activities of the clinician, specifically the extent to which they followed protocol. Even though the interview took place within minutes of the consultation, patient recall is not perfect. It is, however, highly correlated with what actually takes place (Leonard and Masatu, 2006). The questions used to establish protocol adherence are listed in Table 8 in the appendix. Given the existence of medically defined protocol, we can assume that effort to increase protocol adherence is a reasonable measure of quality Workplace Environment Interventions Every health worker was examined in his or her normal workplace (baseline) as well as under two interventions to the workplace environment designed to expose him or her to two different types of reflective stimuli: scrutiny and encouragement. The sequence of interventions followed a 8 The four cadres of clinicians include assistant clinical officer (ACO), clinical officer (CO), assistant medical officer (AMO), and medical officer (MO). Each of these titles requires a specific degree. The medical training required for each depends on the degrees an individual already has. Typically, with no other degrees and 4 years of secondary school, it requires 3 years of training to become a CO. ACO s have less training. AMO s have on average 3.5 years of medical schooling. MO s have the equivalent of a United States MD degree. None of the MOs in our sample participated in the laboratory experiments, so they are not featured in this paper. 12

13 standard order: first we measured protocol adherence under normal circumstances (the baseline); second we measured protocol adherence when there was another clinician in the room observing (scrutiny); third we measured protocol adherence immediately after this clinician left the room (post-scrutiny); fourth a doctor on the team visited with the doctor and read an encouragement script (encouragement visit); and fifth, we measured protocol adherence about 6 weeks after this visit (post-encouragement). Scrutiny involves an immediate reflective stimuli: there is a peer present in the room. Previous work has shown between a 5 and 10 percentage point increase in quality in such circumstances (Leonard and Masatu, 2006; Leonard et al., 2007). For the encouragement intervention, Dr. Beatus, a Tanzanian M.D. and lecturer at a health research institution, visited each clinician and read the following script (numbers were added for clarity and were not in the script): We appreciate your participation on this research study. The work that you do as a doctor is important. Quality health care makes a difference in the lives of many people. Dedicated, hard working doctors can help us all achieve a better life for ourselves and our families. One important guideline for providing quality care is the national protocol for specific presenting symptoms. While following this guideline is not the only way to provide quality, we have observed that better doctors follow these guidelines more carefully. Some of the protocol items that we have noticed to be particularly important (1) are telling the patient their diagnosis, (2) explaining the diagnosis in plain language, and (3) explaining whether the patient needs to return for further treatment. In addition, it is important to (4) determine if the patient has received treatment elsewhere or taken any medication before seeing you, and (5) to check the patient s temperature, and check their ears and/or throat when indicated by the symptom. For this research, we look at clinician adherence to these specific protocol items. We chose specific items because our previous work shows that the best clinicians frequently 13

14 perform these activities but most clinicians do not. Mentioning these items also allows us to compare the performance on these items to performance on items not mentioned. This intervention has multiple potential impacts. The most direct is that the script itself encourages clinicians to improve their quality of care, either because it inspires them or because it contains information they did not previously know. However it may also involve the understanding that one is participating in research or that one s actions are being observed or measured. By the time we measure the quality of care after the encouragement visit, a health worker is likely to have received up to four visits from the research team. 9 Thus, encouragement involves explicit expectations and frequent contact, but, unlike the scrutiny visit, it never involves the immediate presence of a peer. By mentioning five items during the encouragement visit, we can examine if the changes in effort are due to information: if clinicians are responding to new information we should observe increases in quality only for those items for which information was provided. 3.3 Research Design We use a within-subjects design and measure the changes in quality of care (from the baseline) as a result of our two interventions. The post-scrutiny visit was included to test whether clinicians return to their normal quality of care after the scrutiny treatment. The fact that clinicians do return to lower levels of effort after the scrutiny visit allows us to analyze the scrutiny and encouragement as two different types of interventions, not one cumulative intervention. In addition, by treating every clinician in the sample, we are required to use baseline performance as our control rather comparing performance to a random selection of clinicians who received no intervention. We deliberately chose this path for two reasons. First, there is no reason to expect a secular trend in quality that 9 In between the encouragement visit and the post study visits, clinicians were randomized into four treatments in which they received gifts, prizes and follow up visits at different times. These treatments are ignored in the current study and we examine only the long run impact of having been encouraged and studied. 14

15 could lead to significant increases in quality within 6 weeks 10, thus the assumption of no change in the absence of treatment is reasonable. Second, previous work shows that measuring the quality of care can lead to changes in quality without any other treatment, ensuring the contamination of any control group for which we would be able to measure quality. Because the timing is the same for all clinicians, we cannot definitively rule out the impact of the scrutiny visit on the effect we assign to encouragement, but we can rule out the impact of encouragement on the affect we assign to scrutiny. 3.4 Empirical Specification Our investigation concerns the overall level of effort (and therefore quality) provided by clinicians in the three settings (baseline, scrutiny and encouragement) and how these compare for generous and non-generous clinicians. As such, the dependent variable in our regressions is always the effort of the clinician and the independent variables include the generous non-generous classification from the laboratory experiment and the environment in which effort was provided. Measuring changes in effort such that we can reasonably infer that changes are associated with changes in actual quality requires some careful analysis. An outpatient consultation with a clinician involves a series of discrete interactions, most of which are required by protocol. The RCR instrument is designed to measure whether the clinician did the clinical tasks he or she is required to do by asking patients if the clinician did those items (as soon after the consultation as possible). These items can involve greeting the patient and offering him or her a chair, asking the patient how long they have been suffering from particular symptoms, asking about additional symptoms, examining the patient, and explaining the diagnosis properly. The list of discrete items required by protocol differs somewhat according to the presenting symptoms of the patient. We have compiled lists of items required by protocol for four categories of presenting symptoms (fever, cough, diarrhea and general) and two types of patients (older than or younger than five years). 10 Previous and subsequent research in this area has resulted in essentially identical measures of quality for the average clinician, ruling out a secular trend in quality. 15

16 Overall, there are 74 different items (listed in subsection A.2), but only a subset will apply to any given patient. During the RCR interview, patients are only asked about items that apply to their symptoms and age category. Thus, our dependent variable is x ijk, a dichotomous dependent variable indicating whether clinician j followed protocol for item k as required for patient i. This is modeled as a function of clinician fixed effects (Γ j ), item fixed effects (Γ k ), patient characteristics (Z i ). Z i includes four age categories, gender and the order in which patients were seen by each clinician. Each of our conjectures corresponds to an estimating equation. Equation 2 models the impact of being designated as generous in the laboratory experiment (Gen) on the quality of care provided in the baseline: x ijk = Φ(α 1 Gen j + Γ k + Z i z) + e ijk (2) Since this is the baseline, we do not include the variables indicating changes in workplace environment and cannot include clinician fixed effects. Equation 3 models the impact of our workplace interventions on the quality of care provided: x ijk = Φ(α 1 Scr i + α 2 Post Scr i + α 3 Enc i + α 4 Enc i Trk k + Γ k + Γ j + Z i z) + e ijk (3) Scr i and Enc i indicate whether the clinician was subject to one of the work environment interventions at the time they treated patient i. Enc i Trk j captures whether the item is one of the items mentioned in the encouragement visit. 11. We also include a variable (Post Scr) indicating patients seen in the time immediately after the scrutiny intervention to test if quality remains high, falls below normal, or returns to normal. By eliminating a dummy variable for the baseline, we can include clinician fixed effects in this regression. Equation 4 examines the differential reaction of clinicians to the workplace interventions according to their performance gap ( j ): the difference between what is required by protocol and 11 The direct effect of being a tracked item is included in item fixed effects. 16

17 the average proportion of items completed in the baseline: x ijk = Φ(α 1 Scr i + α 2 Scr i j + α 3 Enc i + α 4 Enc i j + Γ k + Γ j + Z i z) + e ijk (4) We can also include clinician fixed effects because the performance gap is interacted with the interventions, not entered directly. Equation 5 examines the impact of the workplace interventions interacted with whether or not the clinician is generous. Because generosity is interacted with the interventions, we can also include clinician fixed effects. x ijk = Φ(α 1 Scr i + α 2 Scr i Gen j + α 3 Enc i + α 4 Enc i Gen j +α 5 Enc i Trk k + α 6 Enc i Trk k Gen j + Γ k + Γ j + Z i z) + e ijk (5) With all four regressions, it is important to consider the sources of variation in effort that are not driven by our interventions. Not all items are equally important, not all clinicians are equally qualified to do each item, and the patients who are at one facility might be unobservably different from the patients at another facility. Thus, comparisons across doctors are difficult. We address these problems in five ways. First, wherever possible, we include clinician fixed effects (Γ j ), allowing us to compare each individual clinician to his or herself in different situations (baseline compared to peer scrutiny for example). This helps to deal with the case mix and qualifications problem because these potential sources of bias do not change during the short period of our study. Second, we include fixed effects for each specific item (Γ k ), essentially asking if a clinician is more or less likely than the average clinician to provide a given item. For example, the clinician who asks about the duration of a cough 80% of the time is providing below average quality, whereas the clinician who asks about the history of vaccinations in infants 80% of the time is providing above average quality. This helps to control for case mix by adjusting expectations for each type of patient; otherwise, a clinician who sees many infants will look worse than a clinician who sees 17

18 few infants because his average score may be lower. Third, because we observe a series of outcomes for each patient (corresponding to all of the required items) we can cluster the standard errors at the patient level or include a patient-level random effect. 12 This allows us to control for the fact that some patients may be quite different from others (they are more demanding or critically sick, for example), the distribution of these patients across clinicians may not be even and the probability of performing one item is likely to be correlated with the probability of doing another for the same patient. Fourth, in addition to examining the probability that a clinician would perform any individual required item, we examine the results looking at average adherence to protocol for each patient, reducing the number of observations to the total number of patients (not potential items). Finally, we include a variable in patient effects (Z i ) indicating the order of patients on the day of the visit. In addition to tracking the illnesses conditions of patients (which change over the day, but are controlled for directly), this helps to deal with changes in case mix (the most severe cases are usually seen earlier in the day). This is particularly important for the scrutiny and post-scrutiny visits because they take place on the same day as the baseline and are always after the baseline. If quality is falling normally over the course of the day and we did not take this into account we would underestimate the effort provided under scrutiny. We include four specifications for each of the equations above, corresponding to the columns in each of the tables. The first specification is a logit model of whether the doctor performed each required item with item-specific dummy variables. Since the standard errors are not corrected or adjusted, this specification always has smaller standard errors than the other specifications. The second specification is a logit model of whether the doctor performed each required item with item-specific dummy variables and patient random effects. The patient random effect captures the possibility that an unobservable patient characteristic might simultaneously increase (or decrease) the probability that a clinician did all of the required items. 13 The third specification is a linear 12 It is not possible to include patient level fixed effects, because each patient is only seen once in our data. 13 We expect this affect to be uncorrelated with observable characteristics of the patient (age and 18

19 regression of the discrete variable of whether the doctor performed each required item with itemspecific fixed effects and standard errors clustered at the patient level. The fourth specification is a linear regression of the proportion of required items performed for each patient ( x ij ). Because we examine average performance over all items, coefficients for tracked items (Trk k ) are dropped. The patient-level regression also controls for the major symptoms reported (fever, cough or diarrhea by infant or non-infant), controls that are already embedded in the item-specific dummy variables for the other three regressions. An additional concern with the measurement of quality is the fact that clinicians might realize we are on site collecting data and change their effort in reaction. In fact, during the scrutiny visit, this is precisely what we expect to happen: clinicians will react to our presence by increasing quality. Appendix A.1 investigates the evidence for this behavior by looking for patterns in the quality of care during each site visit that would indicate that clinicians had increased the quality of care in response to discovering our team. We find no evidence for any of these patterns, suggesting either that no one realized our team was present until after the data had already been collected, or that they were discovered, but there was no reaction. As we discuss later, the evidence suggests that clinicians realized the team had been present after we left, but this discovery does not allow them to cheat by temporarily increasing effort. 4 Results 4.1 Laboratory Experiment Table 1 presents a summary of giving in the dictator game. The average number of tokens given was just over one third, but the mode in the data was half; 36.8% of the participants in the laboratory experiment gave at least half of their tokens to the stranger. Most of these gave exactly half but a few gave more. The fact that the mode was half suggests a norm in which people simply divide their allocation gender) and therefore include these characteristics independently as dummy variables. 19

20 evenly between themselves and their partner. Thus, we create a dichotomous variable indicating the clinicians who gave at least half and call these clinicians generous (i.e. conforming to the generosity norm). 14 About a third of all health workers qualify as generous types in the laboratory, a higher percentage than usually found with the dictator game in other populations. This result may be driven by observable income differences between clinicians and recipients (who were recruited from the general population where the average daily wage is lower than it is among clinicians). Recall that the recipient sample was chosen specifically with the purpose of matching the context in which clinicians see patients. While the income differential between dictator and recipient makes it difficult to directly compare our lab experiment results with the literature, it strengthens the comparability of this behavior with field data because the same gap exists in the field. 4.2 Effort in the Field Table 2 shows the basic statistics for the 63 clinicians who were involved in both the laboratory experiment and the field study. Of these 63 clinicians observed in the baseline, 59 were observed under peer scrutiny and 51 under the encouragement intervention. Doctors dropped out of the study for various reasons but attrition was not correlated with quality. 15 The average clinician completed 74% of the required items in the baseline and the standard deviation of average doctor quality was 16 percentage points. The percentage of items completed during the scrutiny visit is the same as for the baseline, but recall that this number does not control for case mix and that, normally since 14 Results are robust to alternative definitions of generous, including giving exactly half and giving in a small window around the 50/50 allocation. The trends we observe do not come through, however, if we consider a continuous measure of generosity (i.e. where generosity is simply measured by the number of tokens given in the dictator game, rather than giving above some threshold). Those who give more than half are not higher quality clinicians than those who give exactly half, and those who give more than zero but less than 50 tokens are not higher quality clinicians than those who give zero. 15 Results do not change when we run regressions that exclude all attriting clinicians. 20

21 the scrutiny visit is later in the same day effort would otherwise have fallen Are generous clinicians different from other clinicians? The purpose of the laboratory experiment was to document any norm of other-regarding preferences among clinicians, and to categorize clinicians according to this norm for analysis with the field data. We use subjects giving behavior in a standard dictator game to categorize clinicians as responsive to moral stimuli or not. Table 3 (corresponding to Equation 2) examines the quality of care provided by clinicians in the baseline (and therefore does not included clinician fixed effects) with the key variable being the dichotomous classification of whether a clinician is generous in the lab experiment. The tables shows that by all four ways of examining quality, generous clinicians provide significantly higher quality than non-generous clinicians in the baseline. Lab behavior is therefore informative of the relative performance of doctors in the field. This confirms Conjecture 1 that clinicians with greater moral stimulus as measured in the lab will provide greater effort under normal circumstances. The impact is between 7 and 9 percentage points, about half a standard deviation of quality. Note that although the regression implies a causal relationship between behavior in the experiment and behavior in the field, we assume that there is an underlying characteristic that affects both of these behaviors: prosocial preferences (innate or learned) drive behavior in both settings. We cannot rule out other possible links that are not driven by prosocial behavior, nonetheless, this is an important external validity result for the dictator game because it shows a parallel between behavior in the lab and the field Reactions to reflective stimuli (scrutiny and encouragement) Table 4 (corresponding to Equation 3) examines Conjecture 2 for all of the clinicians who took part in the laboratory experiments (not differentiated by generosity). Unlike Table 3, each clinician is compared to himself or herself by the inclusion of clinician fixed effects. The average increase in quality due to scrutiny is between three and four percentage points, depending on the type of 21

22 regression. The reaction to encouragement is about 8 percentage points (as seen in column 4). Columns 1 through 3 show the reaction to encouragement for items that were not mentioned in the encouragement (about 5 to 6 percentage points) and, differentially, for those that were mentioned in the encouragement script (an additional 4 to 7 percentage points, for a total response of about 10 percentage points). The overall reaction to encouragement represents an increase in quality that is about half of a standard deviation of quality. Thus, Table 4 confirms our conjecture that the average clinician will respond to scrutiny and encouragement. Note that after the scrutiny from the research team, the clinician returns towards his or her baseline levels of effort. Effort is slightly higher than in the baseline, though this result is not significant across the regressions. This suggests that the response to scrutiny is short-lived (post scrutiny is not significantly greater than zero) and that there is no need to readjust effort to catch up after the scrutiny visit (post scrutiny is not less than zero). If clinicians believed our research project could have extrinsic ramifications on their practices, they would not have returned to low quality while we were still present at the facility (but not present in their consultation room). Further, it suggests that, by the time the encouragement occurs, clinicians have returned to baseline effort levels and the marginal impact of encouragement is measured from baseline, thus providing a good approximation of the absolute effect of encouragement Heterogeneous responses to scrutiny and encouragement Table 5 (corresponding to Equation 4) examines Conjecture 3, that clinicians who face low levels of motivation under normal circumstances exhibit greater changes in effort when faced with additional scrutiny, the conjecture inherent in our descriptive model of motivation. Table 5 includes a measure of the baseline performance for each clinician transformed into a performance gap: the difference between what is required by protocol and the average proportion of items completed in the baseline. (If a clinician follows protocol for all of his or her patients, the average score would be 1.00 and the gap would be 0.00.) By interacting the gap with the treatments, we examine the degree to which the gap explains the change in performance when a clinician is subject to peer 22

23 scrutiny or encouragement. A coefficient of one would suggest that the gap is fully closed, and a coefficient of 0 would suggest that the reaction to scrutiny is independent of the gap. Confirming Conjecture 3, the coefficients (across all four regressions) suggest that the performance gap is highly correlated with the increase in effort and the gap closes by about one quarter under scrutiny and almost one-half under encouragement. The coefficient on the performance gap is significantly different from both 0 and 1, implying that the reactions to scrutiny and encouragement differ across clinicians, with the better clinicians exhibiting a smaller reaction. The coefficients suggest that, after encouragement, a clinician with 75% adherence in the baseline (a 25% performance gap) will increase his or her effort by about 2 percentage points (0.25* ), whereas a clinician with only 50% adherence will increase effort by 9.5 percentage points Reactions to reflective stimuli for generous clinicians Table 6 (corresponding to Equation 5) examines the impact of the two interventions scrutiny and encouragement and the differential response of generous and non-generous clinicians. Testing whether the response to scrutiny or encouragement is similar for generous and non-generous clinicians addresses Conjecture 4. We regressed quality on interactions of whether a clinician is generous in the laboratory with the timing of the two interventions. As with Table 5, Table 6 regressions include fixed effects for each clinician and therefore examine changes in quality, not the level of quality. The increase in quality due to the impact of peer scrutiny and encouragement for the average clinician is almost exactly the same as we found in Table 4. Since generous clinicians provide higher levels of quality overall, Conjecture 4 suggests that they might respond less to the additional stimuli inherent in peer scrutiny and encouragement. When we examine the marginal coefficients for generous clinicians, the small and insignificant coefficients show that generous clinicians are 16 A clinician with almost perfect adherence actually decreases effort. This result is driven by the asymmetry of measurement error in quality at the high end it is difficult to overestimate quality when the baseline is 98% but easy to underestimate it. 23