How Can Electronic Medical Records Improve Patient Outcomes? Sergei Koulayev Keystone Strategy

Emilia Simeonova1 Tufts University and NBER

Introduction Innovations in information technology have found widespread application in economic activity. Until recently, the health industry had been slow to adopt computerized record-keeping that could aid the organization of health care delivery and the ease of interaction within the system. The recent government initiative to introduce electronic medical records (EMRs) in routine medical practice has been lauded as a major step forward. However, the mechanisms through which EMRs alter the health production process may differ across health care settings and patient populations. In a study of communication processes in hospitals Angst, Devaraj and D’Arcy (2012) find heterogeneous effects of EMR adoption on administrative and clinical outcomes, and suggest that the use of health IT in routine encounters might interfere with physician-patient interactions. The net impact of health IT implementation on patient satisfaction and the quality of care is difficult to predict. We are not aware of any theoretical research in health economics that models the mechanisms through which electronic health records could alter the process of health care delivery in an outpatient setting. A growing empirical literature studies the effect of health IT in hospitals on patient health outcomes (see e.g. DesRoches et al, 2010; McCullough and Parente, 2009; McCullough et al, 2010; Miller and Tucker, 2011; McCullough, Parente and Town, 2013). Different papers report findings that vary significantly depending on the setting and the patient population. In this paper we model the effects of electronic medical records on patient health behaviors by explicitly accounting for their impact on physician-patient interaction. We construct a conceptual framework that allows us to study patient compliance with prescribed medication as one of the factors through which the quality of physician-patient interaction affects health outcomes. Our theoretical model yields testable predictions indicating that introducing electronic health records into the process of care will have heterogeneous effects on patient health behavior that could lead to heterogeneous health outcomes. We show some empirical evidence that corroborates the implications of the theoretical framework. We use imperfect agency approach to model the doctor-patient interaction in the spirit of Aghion and Tirole (1997) and Balsa and McGuire (2001). The game proceeds in two stages.

1

Corresponding author. Email: [email protected]

1

First, the doctor applies a costly effort to learn the patient's condition, receiving a noisy signal about the true health status. Based on this imperfect information she assigns therapy. Second, the patient observes the doctor's effort and decides whether to comply with the prescribed treatment depending on her beliefs about expected benefits net of possible side effects. In this framework medication non-compliance is rationalized as patient's response to risks associated with treatment under imperfect agency. The key parameter is the quality of communication, which in its turn affects the productivity of the doctor's information gathering effort. Our focus on communication is motivated by sociological and medical studies which show that the ability to communicate with the doctor is one of the most important determinants of patient satisfaction with care (see e.g., Williams and Calnan, 1991; Vick and Scott, 1998). These attributes of the relationship have also been shown to affect health outcomes and patient compliance with medication therapy (Kaplan et al, 1989). Our model shows that if a doctor-patient match is characterized by poor communication, the EMR serves as an additional source of clinical information, reducing the doctor's uncertainty and improving patient adherence. However, if the quality of communication is already high, EMR may crowd out part of the doctor's effort, resulting in less customized care and suboptimal patient response. In other words, our model offers two ways in which EMRs may affect patients' decision making. First, by providing the doctor with clinical information, electronic systems reduce the possibility of medical error and thereby improve the patient's trust in the prescribed therapy. Second, access to EMRs may alter the doctor's behavior during the visit, which is another input into the patient's inference about the quality of the treatment.

2. Background This study bridges the literature on patient health behaviors and on the impacts of health IT adoption on the process of care. Patient non-compliance is a significant problem in health care and as such has generated a large volume of discussion in the medical and medical sociology literatures. Economists could contribute to understanding this phenomenon by conceptualizing economic models of medication compliance as an outcome of the doctor-patient relationship. Such models are particularly helpful in organizing our thinking about policies directed at improving adherence, as well as the effects of various innovations in health care - such as the wide-spread use of information technology. The traditional framework for economic analysis of physician-patient interactions is based on principal-agent theory in the context of severe asymmetry of information (initially pointed out by Arrow, 1963). In a later contribution to the contracts literature Aghion and Tirole (1997) present a model of formal versus informal authority that we use as a starting point for 2

modeling patient compliance. In their framework, the principal (the patient) holds the formal authority - she has the veto power over decisions made by her agent (the doctor). The agent has informal authority because he has an informational advantage over the principal and proposes a therapy at her discretion. In yet another influential paper Dewatripont and Tirole (2005) provide theoretical tools for modeling communication between the principal and the agent. In the health economics literature, Mooney and Ryan (1993) offer an informative discussion of the theory of imperfect agency, as it applies to modeling the incentives of the doctor to apply effort - in the context of informational asymmetry and financial incentives. The paper that is closest to ours is by Balsa and McGuire (2001) who model the doctor-patient encounter to study discrimination in health care. They show how sub-optimal care may result from poor quality of communication between the patient and the doctor. Although their model assumes perfect patient adherence to therapy, it brings together issues of communication, clinical uncertainty and doctor effort - all necessary components for analyzing patient's compliance. In subsequent work, Balsa and McGuire (2003) explore the role of stereotypes that doctors may hold about certain group of patients, and show how such stereotypes can be sustained in equilibrium. The key component of the conceptual framework is the prisoner's dilemma between the doctor and the patient: if the doctor believes that the patient would not follow her recommendation, there is no incentive for exerting effort; and vice versa, if the patient expects low effort from the doctor, the benefit of complying with the therapy is low, too. A drawback of this model is that the source of stereotypes is outside the model - in other words, it is not clear how the non-cooperative equilibrium is selected, instead of the cooperative one. In this paper, we show how the quality of communication can support such a set of mutual beliefs. This study also relates to the economics literature analyzing the effects of adoption of information technology in organizations. The idea that information sharing could result in productivity gains has been tested in several areas of economic research (e.g. Hubbard, 2003) and evidence on the effects of IT diffusion on economic activity abounds (see Brynjolffson and Yan (1996) and Brynjolffson and Hitt (2000) for surveys). However, there has been little empirical or theoretical analysis of the effects of computerization and information-sharing in a health care setting - in particular at the micro-level of physician-patient interaction. Among the exceptions is the work by Javitt, Rebitzer and Reisman (2008), who analyze the effects of introduction of a decision support system on resource utilization in a population of commercial HMO patients. They find that the new information technology resulted in 6% lower charges, mainly among the most costly patients. Athey and Stern (2002) study the adoption of superior emergency-response systems (911) centers and its effect on patient outcomes in Pennsylvania, and find significant reductions in mortality that can be attributed to the new technology. Miller and Tucker (2011) use infant health records to examine the effects of EMR adoption in inpatient care. They show significant reductions in infant mortality in hospitals that adopted EMRs. In a series of studies McCullough and co-authors explore the effects of health IT adoption on hospital productivity, patient safety, and inpatient health outcomes (Parente and McCullough, 2009; 3

McCullough et al, 2010; Lee, McCullough and Town, 2012; McCullough Parente and Town, 2013). They report modest positive effects of health IT adoption on hospital productivity and heterogeneous effects of the quality and outcomes of care depending on the medical setting. Thus, the existing empirical evidence on EMRs’ impact on inpatient care ranges from strongly positive in infant health care (Miller and Tucker, 2011) to neutral for the average adult inpatient (McCullough, Parente and Town, 2012). Inpatient and outpatient care settings vary in a number of aspects, and perhaps the most important difference in the context of computerized patient records is the intensity and importance of physician-patient communication. In the rest of this study we set out to conceptualize the potential channels through which EMRs can affect the interaction between providers and patients in regular outpatients care and their impact on patient health outcomes. 3. Doctor-patient interaction: a game-theoretic approach The traditional view on doctor-patient relationship is one-directional: the doctor prescribes the treatment, which the patient then follows. Instances of non-compliance are interpreted as deviations from rational behavior: patients do not understand the benefits from treatment or they forget to take the medication. Our goal is to construct a theoretical model of doctor-patient interaction as two-way process, thereby putting the patient’s contribution to therapy back into the set of determinants of treatment’s success. To do so, we build on the existing theoretical approaches to modeling the physician-patient interaction in the health economics literature. We consider the patient as an active participant in the health maintenance process. While such participation may take many forms, arguably the most important one is the decision to comply with the assigned therapy. Since the patient does not possess medical knowledge, the compliance decision is typically made in the face of uncertainty about possible effects of the treatment. Notably, this uncertainty has two sources: one is the patient's inability to evaluate the adequacy of the prescription; another is the patient's understanding that there is scope for medical error. A rational patient weighs the costs of compliance - possible side effects as well as costs of a doctor's mistake - against the benefit of improvement in health status. This view on compliance has already received support in the existing economics literature, notably in Balsa and McGuire (2001), as well Ellickson, Stern and Trajtenberg (1999); it has also received positive reviews from doctors, such as Escarce (2005). It is worth noting that this setup is mostly suitable to study outpatient interactions and non-acute health conditions. We begin with an observation that with EMRs doctors are supplied with an extra source of clinical information, which enables them to prescribe adequate treatment. Even if the patient herself does not have access to that information, she understands that the prescription is now made by a better informed doctor. That improves her trust and participation in the therapy.

4

Direct communication between the doctor and the patient during the visit is another important channel2 through which the doctor can learn about the patient's condition. It provides the kind of "soft" information that cannot be recorded in a computer: the patient's preferences for therapy, her emotional state or her perception of social support. According to sociological and medical studies, the ability to communicate with the doctor is one of the most important determinants of patient satisfaction with care (see e.g., Williams and Calnan, 1991; Vick and Scott, 1998). These attributes of the relationship have also been shown to affect health outcomes and patient compliance with medication therapy (Kaplan et al, 1989, Ashton et al, 2003, Cooper et al, 2003). The quality of communication varies among doctor-patient matches, which may contribute to the observed differences in health outcomes of otherwise similar patients. The connection between the quality of communication and health outcomes has been previously emphasized by Balsa and McGuire (2001, 2003) in the context of racial discrimination. age, education level, nationality, and language skills - such factors may affect the ability of the patient to explain her condition to the doctor, as well as the ability and willingness of the doctor to listen. For instance, Simeonova (2013) shows that life events such as the loss of a spouse can affect the process of outpatient care. Using registry data covering the universe of primary physician-patient encounters in Denmark, Koulayev, Simeonova and Skipper (2013) find that patient adherence with therapy varies over time even within the already established doctorpatient matches. As compared to reading the information off the computer screen, direct communication with the patient requires an investment of time and effort on the doctor’s part. Since effort is costly, doctor's incentives to acquire information through communication are sub-optimal, relative to the patient's best interest. A rational patient takes this into account when making a decision on whether to comply with the doctor’s recommendation. Therefore, the imperfect agency framework is suitable for describing the doctor-patient relationship (Mooney and Ryan (1993)). Doctor-patient matches characterized by higher quality of communication should exhibit better compliance for two reasons: first, the doctor is more informed ceteris paribus; second, as has been pointed out by Balsa and McGuire (2003), the doctor has more incentives to apply effort, further improving her information and the patient's trust. Having identified the main building blocks of the doctor-patient interaction – EMR, communication, and imperfect agency – we now combine them to study the interaction between health IT and communication, and the resulting effects on patients’ compliance.

2

While the quality of communication may have independent effect on patient's trust and compliance, we focus on the communication as an information channel.

5

3.1 A game-theoretic model of doctor-patient relationship The game between a patient and a doctor occurs in two periods. In the first period, the patient meets with the doctor and communication takes place, during which the doctor attempts to learn the patient’s true health condition, denoted by the parameter 𝜃. The doctor makes a costly effort 𝑒 and receives a noisy signal 𝜃̃ about the true value of 𝜃. Patients differ in their communication style, as represented by the parameter 𝜁. Language skills, education, prior experience with the health care system - all these factors affect parameter 𝜁, which can be interpreted as the ability of the patient to explain her condition to the doctor3. The effectiveness of the doctor’s effort depends on the quality of interaction with the patient. In a simplest example, a doctor can get much less relevant information from an immigrant worker who speaks poor English, than from a native speaker. If the patient’s condition is complicated enough, poor quality of communication affects the doctor’s ability to prescribe adequate treatment. For simplicity, we assume a linear communication technology: 𝑃(𝜃̃ = 𝜃) = 1 − 𝑝(𝑒, 𝜁) 𝑝(𝑒, 𝜁) = 1 − 𝑒𝜁 , 𝜁 ≥ 0 With probability 1 − 𝑝(𝑒, 𝜁) = 𝑒𝜁 the doctor learns the true value 𝜃 perfectly. With probability 𝑝(𝑒, 𝜁) = 1 − 𝑒𝜁, he makes an incorrect prescription. The probability of mistake decreases both with the doctor’s effort and with the patient’s communication type: higher type patients enjoy better informed doctors at any given level of doctor’s effort. Both the effort level and the parameter 𝜁 are commonly observed. An important assumption is that the patient understands the communication technology and takes it into account when choosing a compliance level. Therefore, due to their different communication abilities, an immigrant worker and a native speaker would make different inferences from the same observed doctor effort. A similar point is made in Balsa and McGuire (2001). In the second period, the patient decides whether or not to follow the recommended therapy, based on her perceptions of expected costs and benefits of the therapy. In our model, patient’s beliefs are based on two inputs: first, her prior beliefs, possibly based on past experience, as well as her communication type; second, the doctor’s behavior during the visit, as measured by the observable level of effort. Everyone can read such signs as the doctor’s attitude, attention to detail, as well as the total time spent on the visit. To the extent that the doctor’s effort

3

The communication parameter can also be interpreted as the quality of a doctor-patient match along a certain dimension (or several dimensions) between a physician and a patient that is pertinent to the curative process. For example, there is plenty of evidence that Hispanic doctors have proportionately more Hispanic patients (Stinson and Thurston, 2002) and that the racial profile of the patient population is a good predictor of the race of the physician (Komaromi et al, 1996). Patients express higher satisfaction if treated by a doctor of similar ethnicity or race, are likely to get more preventive care, and to maintain treatment for longer periods of time (Saha et al, 1999; Takeuchi et al, 1995; Cooper-Patrick et al, 1999).

6

is related to the gathering of information about her condition, the patient can use it to infer the potential benefits of the prescribed treatment. As a result, the doctor’s effort in our model has a dual purpose: it reveals clinical information about the patient to the doctor and at the same time determines the level of patient trust, and, eventually, of compliance. Normalizing the cost of non-compliance to zero, the relative benefit that the patient derives from following the therapy is: 𝑈𝑃 (𝜃̃, 𝜃) = 𝐵̃ − 𝐿(𝜃̃ − 𝜃) where 𝐵̃ is the perceived (net) benefit of the treatment and 𝐿(𝜃̃ − 𝜃) is the loss in treatment efficacy due to an imprecise diagnosis: 𝐿 = 𝐶 > 0 if 𝜃̃ ≠ 𝜃 and 𝐿 = 0 if 𝜃̃ = 𝜃. The parameter 𝐶 characterizes the scope for medical error, which depends on the severity and complexity of the patient’s condition. With these assumptions, the patient will adhere with the prescribed therapy if and only if: 𝑈𝑃 = 𝐵̃ − (1 − 𝑒𝜁)𝐶 > 0

(1)

During the visit, the doctor needs to form a belief about the likelihood of patient compliance: indeed, there is no need to apply a costly effort if the patient would not follow recommendations. Vice versa, if the patient observes low effort, she will follow up with low compliance. This relation between patient’s compliance and doctor’s effort resembles the prisoner’s dilemma, as emphasized previously by Balsa and McGuire (2003). In equation (1), the expected cost of compliance (1 − 𝑒𝜁)𝐶 is common knowledge to both parties. However, there is asymmetry of information regarding the patient’s perceptions of the benefits of the treatment, 𝐵̃ – this parameter is known to the patient (equation (1)) but not to the doctor. For tractability, we assume a uniform distribution of 𝐵̃ in the population: 𝐵̃ = 𝐵 + 𝜂

(2)

𝜂 ∼ 𝑈[−𝐵; 0] The doctor knows 𝐵 – the objective (actual) benefit of treatment, but not 𝜂 – an idiosyncratic shock to the patients’s perception of the treatment. This shock could be due to patient’s private information about her health condition, associated health-related behaviors or any prior (but erroneous) beliefs about the benefit of treatment. To avoid corner solutions, we assume a wide enough support, so that at its lowest realization, 𝜂 = −𝐵, the patient does not comply; at the highest realization, 𝜂 = 0, she complies with certainty, as follows from Assumption (1): Assumption 1. 𝐵 > 𝐶 This assumption ensures that non-compliance happens only because of low realizations of patient beliefs, and not because of the incompetence of the doctor (indeed, if it was known in 7

advance that the true expected benefit is negative, the doctor would not have prescribed the therapy). In our model, the doctor is a benevolent one in the sense that he internalizes the patient’s benefit from the treatment4. The optimal physician effort is determined from the following program: 𝑈𝐷 = 𝑃(𝑈𝑃 (𝑒) > 0)(𝐵 − (1 − 𝑒𝜁)𝐶) − 𝑐(𝑒) → max 𝑒

(3)

s.t. 𝑒 ≥ 0, (1 − 𝑒𝜁) ≥ 0 In reality, the doctor would also choose a therapy, based on her signal 𝜃̃ and the likelihood of compliance. For example, in order to induce higher compliance levels, a doctor may choose a drug with slower curative but softer symptomatic effects. In this model, we take the treatment function as given. In fact, in our setup there would be no incentives for a strategic choice of treatment, as the patient does not update her beliefs from that choice. To proceed, we need a technical assumption that guarantees an internal solution to the program (3). 1

Assumption 2. 𝜁 2 ≤ 2𝐶 Assuming a quadratic cost function, 𝑐(𝑒) = 𝑒 2 /2, Lemma 1 summarizes the solution to the optimal doctor effort: Lemma 1. With assumptions (1) and (2), the optimal physician effort is: 𝑒 ∗ = 2𝜁𝐶

𝐵−𝐶 𝐵 − 2𝜁 2 𝐶 2

(4)

which increases with 𝜁. If the cost of mistake is not too high, 𝐶 < 𝐵/2, then the doctor’s effort increases with 𝐶. For higher cost levels, 𝐵/2 < 𝐶 < 𝐵, there exists a threshold 𝜁̄1 , such that the 1 doctor’s effort increases with 𝐶 for patients with 𝜁 > 𝜁̄1 , and decreases for others. As 𝜁̄12 < 2𝐶, the latter set of (𝜁, 𝐶) combinations is non-empty. Proof. See Appendix. The optimal effort is increasing with the patient’s communication style, which follows from the assumption that the productivity of physician effort increases with 𝜁. The effects of the cost of mistake 𝐶 on the doctor’s effort are non-linear: among patients with mild conditions, 𝐶 < 𝐵/2, the doctor is motivated to apply more effort in order to avoid a mistake. However, once the 4

By postulating a benevolent doctor, we abstract from possible financial incentives that may alter the choice of therapy. This is an adequate assumption in the VA hospital system in the US and in most European health care systems, where pay is equalized. In US private sector, financial incentives could be important.

8

condition becomes more complex, the doctor’s effort responds positively to higher cost only if the patient is sufficiently communicative, 𝜁 > 𝜁̄1 . For patients with lower communication quality, the expectation of low future compliance discourages the doctor from applying more effort. This result is not obtained in a model with perfect compliance: when 𝑃(𝑈𝑃 (𝑒) > 0) = 1, the solution to (3) is 𝑒 ∗ = 𝜁𝐶 - which always increases with 𝐶.

3.2 The effect of electronic medical records When electronic medical records are introduced, they enlarge the information set of the doctor, at any given level of effort. That is, the doctor becomes more informed regardless of the signal 𝜃̃ he receives from the patient. Accordingly, he is able to adjust the treatment function in a way that minimizes the cost of a mistake, 𝐶. This is consistent with evidence from diabetes care – Cebul et al (2011) report that practices that have adopted electronic health records consistently outperform the comparison group along a number of dimensions of quality of care and outcomes. We do not explicitly model this change of treatment: instead, we assume that the introduction of EMR decreases 𝐶 for all patients. To study the effect of EMRs on compliance, we perform a comparative statics exercise of the expected cost of compliance, (1 − 𝑒 ∗ 𝜁)𝐶, at various levels of 𝜁. According to (1), the expected cost of compliance determines the compliance rate, 𝑃(𝑈𝑃 (𝑒) > 0). Previously, in Lemma (1) we have established the reaction of the optimal effort to changes in 𝐶 for various types of patients. We found that for many patients, lower 𝐶 means lower effort; to translate this into changes in compliance, we need to find if the effort drops faster than 𝐶 for some patients. Proposition 1. Suppose assumptions (1) and (2) hold. Then the introduction of EMR has 1 the following effects, depending on the communication style, 𝜁. There exists a threshold 𝜁̄2 , 2𝐶 > 𝜁̄2 > 𝜁̄1 , such that the expected cost of compliance, (1 − 𝑒 ∗ 𝜁)𝐶, decreases for patients with 𝜁 < 𝜁̄2 and increases for patients with 𝜁 > 𝜁̄2 . Among patients with 𝜁 < 𝜁̄2 , those with 𝜁 < 𝜁̄1 actually experience an increase in the doctor’s effort. Proof. See Appendix. The results of this Proposition are illustrated on Figure 1.

9

Figure 1. Gradient of compliance with respect to doctor's effort

We can interpret these results by keeping one of the parameters 𝜁, 𝐶 fixed and changing the other. For many patients with a given 𝜁, the introduction of EMR (lower 𝐶) decreases doctor’s effort. However, there is a category of patients - low communication type and high cost of mistake - for whom the doctor’s effort actually increases. For patients with high enough value of 𝜁, the doctor’s effort drops faster than 𝐶, leading to a higher expected cost of compliance and, consequently, to a deterioration in adherence. On the other hand, patients with a low value of 𝜁 exhibit higher compliance rates after the EMRs are introduced. To get some intuition for this result, suppose that physicians’ cost has changed by Δ𝐶 = 𝐶1 − 𝐶0 , inducing a change in effort Δ𝑒 = 𝑒1 − 𝑒0 . The corresponding change in the cost of compliance is: (1 − 𝑒1 𝜁)𝐶1 − (1 − 𝑒0 𝜁)𝐶0 = Δ𝐶(1 − 𝜁𝑒0 ) − 𝜁Δ𝑒𝐶1 This expression has two parts, whose relative weight depends on 𝜁. For patients with low 𝜁, the first part dominates: they care more about the actual decrease in the cost of mistake than about changes in doctor’s effort. This is the scenario where patients are primarily interested in “avoiding disaster”: they understand that doctor may be misinformed because of their low communication skills, and put more weight the on external sources of doctor’s information. For patients of the better communication skills type, the second part dominates: changes in doctor’s effort are more important that “avoiding disaster”, because the probability of misinformation is already low. We could say that these patients are seeking "customization of care" - they value the extra time the doctor spends with them. Some of that time is replaced by the time physicians spend with the patient’s EMR, hence we expect to see a decrease in the rates of compliance among these patients. It is important to emphasize here that the effect of EMRs on physician effort does not mean doctors become negligent. Also, our model is not informative about other roles of doctor’s 10

effort. Instead, we observe that the information received through EMRs crowds out the information gathering component of the doctor’s effort during the outpatient visit. For example, doctors who have access to EMRs may adopt a more active approach to treatment. This can manifest in (1) the physician spending less time listening to the patient; (2) the physician opting for an authoritative instead of a joint decision on the course of treatment; or (3) a more active approach to treatment management on behalf of the physician. These changes may decrease the quality of communication, as perceived by the patient. In a way, while it is important that doctors are well-informed, it is also important that patients know about it. The aggregate welfare effect of EMRs is difficult to evaluate, as it depends on the relative weight of patients to each side of the communication threshold, 𝜁̄2 . Yet we expect that for the majority of patients the quality of care should increase or remain at the pre-EMR level, owing not only to information effects, but also due to better coordination of care and in particular improved chronic disease management. In other words, we expect that changes in doctors’ effort, although important for a subset of patients, are not likely to outweigh the benefits from the technology. To summarize, by supplying doctors with information through independent channels, the electronic health record works to improve equality among patients, reducing the variation in quality of care that is due to communication problems.

3.3 Testable predictions Our model of doctor-patient interaction yields several testable implications of the effect of introduction of EMR on various types of patients. In this framework, patients differ by 𝜁 communication type and by 𝐶 - cost of mistake. Depending the combination of these parameters, the doctor may increase or decrease her effort in response to EMRs availability, and the patient may decrease or increase her compliance, as shown on Figure 1. An empirical test of these predictions requires proxies for 𝜁 and 𝐶, as well as a measure of doctor’s effort. Real-world counterparts of the parameter 𝐶 include indicators of the complexity and acuteness of the condition, severity of the disease, co-morbidities - factors that require an active interaction between the patient and the doctor. We should observe higher sensitivity of compliance to doctor’s effort among patients with higher values of these factors. The communication parameter 𝜁 can be broadly interpreted as a quality of match between a doctor and a patient, which includes all factors that facilitate interaction. Let 𝑖 - index of the patient, and 𝑗 - index of the doctor, and 𝑡 - time of the interaction. Then 𝜁𝑖𝑗𝑡 will have components: 𝜁𝑖𝑗𝑡 = 𝜁𝑖 + 𝜁𝑖𝑗 + 𝜁̃𝑖𝑗𝑡 , where 𝜁𝑖 - characteristics of the patient (such as language skills), 𝜁𝑖𝑗 - time-invariant factors that characterize the doctor-patient pair (such as common race), and 𝜁̃𝑖𝑗𝑡 - transitive factors pertaining to a particular interaction, unobserved to the 11

econometrician. The availability of proxies to these components of doctor-patient match value depends on the quality of the data. The effect of EMR on physicians’ treatment strategies One of the predictions of the model is that EMRs will reduce physician uncertainty about a patient’s 𝜃 and that may result in changes in treatment strategies employed by the doctor. A panel of individual outpatient health records or outpatient electronic pharmacy data that uniquely identifies physicians and patients can be used to test for changes in treatment strategies after EMR adoption. Such data are becoming more accessible with the spread of electronic claims processing systems and electronic pharmacy. One testable prediction of our models is that controlling for patient demographics and health status, physicians will attempt more medication groups in the treatment of the condition post EMR adoption. An increase in the number of medications prescribed conditional on the patient’s health state can be interpreted in two ways. First, we can think of it as stepping up the effort exerted by the physician in designing the optimal therapy. Second, a change in treatment strategy can be a response to a falling cost of non-compliance for the patient. As the model predicts, EMRs reduce the cost of non-compliance for all patients, but the reduction is relatively larger for patients with low levels of 𝜁 and high Cs. We would expect intensity to pick up for patients endowed with 𝜁 and 𝐶 illustrated in region C on figure 1. Patients’ response Empirically testing the model’s predictions about individual patients’ response to EMR adoption requires electronic pharmacy data or detailed surveys of physicians and their patients. These data can be used to construct measures of individual patient adherence with therapy and to identify differences in patients’ reactions to physician effort across demographic groups and health conditions. Throughout the rest of the text, the index 𝑖 indexes the patient, 𝑗 indexes the doctor, and 𝑡 is the year in which the interaction took place. The patient-specific parameter 𝜁 is not observed directly in the data. We think of the parameter 𝜁 as having two parts – a patientspecific part, and a patient-doctor specific part. The patient-specific part of 𝜁𝑖 is common across all physicians and affects each patient-physician interaction in a similar way. For example, if a patient has had negative previous experiences with the medical care system, he will be more distrustful of all health care providers and have a lower 𝜁𝑖 . We assume that the patient-specific 𝜁𝑖 does not change over time. The patient-provider specific part of 𝜁 -𝜁𝑖𝑗 , is particular to the doctorpatient pair. We think of 𝜁𝑖𝑗 as the match-specific portion of the parameter, which captures the level of trust between a patient and a physician inspired by factors exogenous to the physician’s efforts. For example, one such factor is common ethnicity or race. We assume that 𝜁𝑖𝑗 does not change with time. We can write an empirical model that exploits variation across and within patients as:

12

𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑗𝑡 = 𝜁𝑖 ∗ 𝐸𝑀𝑅𝑡 + 𝜁𝑗𝑡 ∗ 𝐸𝑀𝑅𝑡 + 𝐸𝑀𝑅𝑡 + 𝜁𝑡 + 𝜁𝑖𝑗 + 𝑋𝑖𝑡 𝛽 + 𝜀𝑖𝑗𝑡 where 𝐸𝑀𝑅𝑡 is an indicator equal to one if the electronic medical records system has been fully implemented in the outpatient clinic visited by the patient and X is a vector of patient characteristics. We omit the coefficients to avoid confusion with the parameter terms. The analysis can be conducted using doctor-patient-year-level variation in outcomes. The ideal outcome variable is the doctor-patient-year specific medication adherence rate or the sum of unpicked prescriptions written by doctor 𝑗 for patient 𝑖 in year 𝑡. Estimating this model with patient fixed effects will yield a coefficient on the interaction term EMR*patient type that will capture both the EMR-specific effects of 𝜁𝑖 and of all the 𝜁𝑖𝑗′ 𝑠 Including provider-patient match fixed effects would absorb both patient-specific (common across physicians) and patient-doctor-match specific (within physician-patient pair) unobservables that are not changing over time. Both 𝜁𝑖 and 𝜁𝑖𝑗 will be captured by the physicianpatient pair fixed effect. The coefficient on patient-provider specific interaction term will capture the match-specific within-pair effect of the introduction of the EMR. The two best proxies for 𝜁 available in most US-based health utilization data are the patient’s race and her annual income. We assume that black race is a proxy for the level of compatibility in the provider-patient pair as well as for the general familiarity and trust of the patient in the health care system. Black patients with cardiac conditions are less satisfied with the health care they receive and more likely to mistrust the system overall (LaVeist et al., 2000). Most general practitioners currently working in the US private and public health system are not African American. There may be differences in satisfaction with care and physician-patient cooperation based on racial matching. For example, Saha et al. (1999) find that minority patients who see minority physicians are more likely to rate physicians highly and to report receiving preventive care. Patients holding negative stereotypes about their physicians are less likely to be satisfied with the care they receive and less likely to adhere to physician therapy recommendations (Bogart et al., 2004). Thus, we expect that on average black patients would exhibit lower levels of 𝜁 higher levels of C. Studies have found that African American patients have lower average levels of compliance with medication (Simeonova, 2008; Simeonova, 2013). Income is a direct measure of socioeconomic status and a good proxy for educational attainment. Medical sociology studies have shown that patients of higher SES are more active in seeking and supplying information about their condition (Pendleton and Bochner, 1980; Boulton et al, 1986) and prefer to be more directly involved in decision-making. We expect that higher income is associated with higher levels of 𝜁. 4. Empirical evidence The most adequate way to empirically test the implications of the theoretical model is by using a panel of individual physician-patient interactions that contains information on patient 13

compliance with therapy and patient outcomes. Unfortunately we do not have access to such detailed data. Still, we find evidence on heterogeneous effects on EMRs consistent with the predictions of our model even in aggregated data that does not distinguish between individual patients or physician-patient pairs. We use outpatient data aggregated to medical center-by year-by race cells on mortality and compliance with medication. The cell means are calculated over outpatients who were diagnosed with chronic heart failure in the Veterans Health Administration between October 1998 and October 2004. The underlying individual sample was restricted to male patients who used outpatient care at least twice in the first year after a diagnosis of chronic heart failure. The month and year of full implementation of the EMR system are available for 104 medical centers and their satellite outpatient clinics. The distribution of the timing of implementation electronic medical records is recorded in Table 1. Table 1: Timing of EMR implementation in medical centers and outpatient clinics Year of full implementation N centers 1995 2 1996 1 1997 2 1998 16 1999 32 2000 15 2001 13 2002 14 2003 6 2004 2 Still in progress in 2004 4 The computerized patient records system was introduced between 1995 and 2004 in different medical centers and their satellite outpatient clinics. The electronic record contains information on all patient medical conditions, the outpatient visits and inpatient episodes, as well as the past and current medication therapy. It also records the identity of the providers whom the patient has encountered and their recommendations. Hence, if the patient met two different doctors in consecutive meetings, the second physician has a complete record of the medication therapy prescribed by the first physician, as well as all vitals, lab results, and previous adverse health events. The electronic medical records system is part of a much larger electronic patient data infrastructure (VISTA) which has been in use in the veterans’ health care system since the late 1970s (Brown at al, 2003). The new elements of VISTA, which were implemented around the period of interest, were the Bar Code Medication Administration (BCMA), used in inpatient services, and the Computerized Patient Record System (CPRS), recording patient information 14

across inpatient and outpatient encounters and pharmacy data. CPRS includes provider order entry and provider-entered electronic progress notes. It was released as a separate IT product initially in 1996, and its implementation was mandated nationally in 1999. Among other applications, CPRS contains patient-specific records of pharmacy order, lab reports, progress notes, vital signs, inpatient and outpatient encounters. In September 2002, providers entered over 90% of medication orders electronically (Brown et al, 2003). Other features of CPRS include a notification system that immediately alerts clinicians about clinically significant events such as abnormal test results, a strategy that helps prevent errors by requiring an active response for critical information. A patient posting system, displayed on every CPRS screen, alerts clinicians to issues related to the patient, including crisis notes, special warnings, adverse reactions, and advance directives. Table 2 shows means and standard deviations of the outcome and demographic variables by EMR implementation status. Table 2: Average medical center characteristics and patient outcomes by EMR adoption status Variable No EMR EMR Black 0.0896 0.0516 1-year survival 0.986 0.976 2-year survival 0.964 0.955 3-year survival 0.945 0.949 Mean compliance with medication 0.592 0.564 therapy Chronic Heart Failure The clinic-by year means we use in the empirical analysis are for patients who have received a diagnosis of chronic heart failure (CHF). There are several reasons why this condition is an interesting diagnosis to explore in relation to EMRs. First, heart disease is the leading cause of death in the elderly and is the most costly single condition in Medicare in recent years (33.2 billion dollars in 2007)5 . Second, heart disease is an Ambulatory Case Sensitive Condition, which makes it particularly susceptible to policy interventions in an outpatient setting. It has been shown that expensive hospitalizations and re-hospitalizations can be avoided with adequate preventive care and disease management. Finally, heart failure is rarely misdiagnosed, and there are clear guidelines for pharmacologic outpatient-based treatment. Chronic heart failure is a progressive health disorder with fatal outcomes. Mortality rates in the first year after diagnosis are about 10 per cent. However, if care is managed well, patients’

5

According to the AHA statistical abstract, 2007 (http://www.americanheart.org/downloadable/heart/1166711577754HS_StatsInsideText.pdf)

15

chances of living longer and their quality of life can be improved significantly. The recommended medical therapy is well publicized. Once the first year of treatment has passed successfully, chances of longer-term survival increasingly depend on the patients’ and doctors’ ability to adapt the treatment and lifestyles to counter the progression of the disease. Short-term (one-year) mortality is more likely to be influenced by the patient’s initial physical condition at diagnosis, while longer-term survival would be more sensitive to medical therapy and the ability of the patient and the doctor to effectively coordinate the management of the disease. Measuring patient adherence with therapy We use data on average clinic-level adherence with therapy across different racial groups before and after the introduction of EMRs. If there is little substitutability between doctor and patient effort, no therapy would work without the patient’s active participation. While health studies evaluate the effect of doctor inputs, they rarely account for the effect of patients’ response to physicians’ efforts. Leonard and Zivin (2005) provide one of the few models of health production that explicitly accounts for patient input. Patient response could be especially important for chronic conditions such as chronic heart failure that are managed on an outpatient basis, and that require an investment of daily effort by the patient. The aggregated data used here are medical center-by year by race means based on individual patient adherence. The individual data have been used in Simeonova (2008). Data on prescription refills were used to define a measure of patient adherence to therapy. The VHA pharmacy data contain a "days supply" variable attached to each prescription, as well as the time when the first dose was dispensed and the time of subsequent refills. Using the "days supply" variable one can determine whether the prescription was refilled on time. A refill is defined as “compliant” if it was picked up within 3 days of the expiration of the previous days’ supply. The adherence measure is defined as the number of prescriptions which were not re-filled on time over the total number of prescriptions in a patient-year cell. According to the most comprehensive study of adherence measures, the one defined here is ranked the best in the context of an integrated pharmacy system6 (Ostenberg and Blaschke, 2005). Medication adherence ratio = ((N prescriptions filled on time)/(Total N prescriptions)) Note that this measure is defined over prescriptions that were picked up by the patient, and does not include prescriptions written by the provider, but ignored or forgotten by the patient. In this study we use aggregated medical center-by race and year averages of this measure and weight all regressions by the number of patients in each cell.

6

The VA pharmacy only fills prescriptions that were ordered by a physician within the VA health care system. The pharmacy keeps electronic records for all transactions which is independent of the EMR system. Pharmacy records cover the entire medication history of the patient and are exhaustive within the VA health care system.

16

Empirical evidence - survival analysis We first present some empirical evidence on patients’ mean survival rates following implementation of EMR systems in outpatient medical centers. Table 3 presents the results from a series of regressions comparing one-, two-, and three-year survival rates in medical centers that implemented electronic medical records before and after the implementation and medical centers that did not implement the technology. All models include calendar year and medical center fixed effects. Survival means are available for 100 outpatient medical centers for an average of 5 calendar years each. In the regression analysis cell-level means are weighted by the number of patients in each cell. The empirical results suggest that electronic medical records do not significantly affect average patient survival rates. The coefficient estimates on the EMR implementation dummy in the first three columns in Table 3 are economically small and statistically insignificant. Based on these estimates, we can exclude EMR effects that exceed 0.1% reductions in the one-year average patient survival rate; effects on the two-year survival rate conditional on surviving one year of more than 0.3%; on the average three-year survival rate conditional on making it past the first two years that is greater than 0.6%. Interestingly, black race is associated with improved survival probabilities in the first two years after diagnosis. This is most likely explained by the younger age at first diagnosis for African American patients (see Simeonova, 2008). In the specifications reported in columns (4) through (5) we allow the effect of electronic medical records to vary by race. Health IT clearly has heterogeneous effects of different types of patients. African American patients are significantly more likely to survive the second and third years after initial diagnosis if they visit a medical center that has implemented EMRs. The associated increase in survival probabilities is sizeable, in the order of one standard deviation improvements in survival rates for black patients. If we consider the practical implications of our theoretical model, it is not surprising that the benefits of EMRs manifest in the second and third year after diagnosis. We hypothesized that electronic health records will affect the quality of the physician-patient interaction during the outpatient care visit. For patients suffering from chronic conditions, improved communication with their physician will take some time to translate into improved health outcomes.

17

Table 3: The effects of electronic medical record adoption on clinic-level survival probabilities; linear probability regressions; medical center-year cell averages by race weighted by the number of patients in each cell Survival post-diagnosis EMR Black race

(1) One-year

(2) Two-year

(3) Three-year

(4) One-year

(5) Two-year

(6) Three-year

-0.001 (0.001) 0.002** (0.001)

-0.003 (0.003) 0.008*** (0.002)

-0.002 (0.004) 0.002 (0.003)

Yes Yes 0.987*** (0.002) 325655 0.326

Yes Yes 0.963*** (0.004) 288481 0.551

Yes Yes 0.944*** (0.003) 243658 0.727

-0.001 (0.001) 0.002 (0.002) -0.001 (0.002) Yes Yes 0.987*** (0.002) 325655 0.327

-0.004 (0.003) 0.002 (0.005) 0.010* (0.006) Yes Yes 0.963*** (0.004) 288481 0.553

-0.003 (0.004) -0.009 (0.006) 0.017** (0.007) Yes Yes 0.945*** (0.003) 243658 0.729

EMR*black MC fixed effects Year fixed effects Constant Observations R-squared

Robust standard errors in parentheses clustered at the outpatient clinic level; * significant at 10%; ** significant at 5%; *** significant at 1%

Patients’ response We next investigate one of the mechanisms through which electronic health records can result in improved patient outcomes as suggested by our conceptual framework. Our theoretical model implies that patients with worse communication skills will benefit from EMRs through improved trust in their physician’s ability to assess their condition and prescribe the appropriate medication therapy. Thus, we expect that the group of patients who benefit from EMRs would also show improved compliance with physicians’ therapy recommendations post-EMR. In the testable implications section, we hypothesized that due to communication and cultural barriers to efficient physician-patient interaction, minority groups will benefit most from health IT. In the specifications reported in Table 4 we investigate the effects of EMR adoption on average compliance with recommended therapy. We report results from specifications that include only medical center fixed effects and year dummies. Again, the observations are medical center-year-race cell means. Observations are weighted by the number of individuals in each cell. In the first column we report results from the basic specification testing for a level effect of EMR adoption of average medication adherence levels in the medical center. The estimated coefficient is statistically imprecise, but also very small. In the second column we report the coefficients from a model including an interaction term between black race and EMR, which allows the effects of electronic health records to differ between cell means for patients of different racial backgrounds. Recall that our model predicts a positive effect of EMR adoption on the adherence of patients with low 𝜁 and that we hypothesized that on average, minority groups will comprise most of these patients. Thus, we 18

expect that the coefficient on the interaction term between EMR implementation and black race will be positive. Indeed, we find a sizeable positive effect of electronic records adoption on average compliance among African American patients. The effect is large enough to account for half of the gap in average adherence between black patients and the rest that is reported in column (1). Introducing electronic medical records into routine outpatient care practice works to reduce inequities in patients’ responses to physician effort. This is consistent with our model’s prediction that health IT has heterogeneous impacts on different patient types and that these heterogeneities might be lost in analysis focusing primarily on differences in outcomes evaluated at the mean. Table 4: The effect of electronic medical records on compliance with prescribed medication therapy; OLS regressions at the clinic-year-race cell weighted by the number of patients in each cell.

EMR Black race

(1) Mean compliance

(2) Mean compliance

-0.010 (0.010) -0.056*** (0.010)

-0.012 (0.010) -0.076*** (0.013) 0.027** (0.011) Yes Yes 0.590*** (0.008) 270447 0.739

EMR*black MC FE Year FE Constant Observations R-squared

Yes Yes 0.588*** (0.008) 270447 0.737

Robust standard errors in parentheses clustered at the outpatient clinic level; * significant at 10%; ** significant at 5%; *** significant at 1%

5. Conclusions This paper studies the effects of the introduction of electronic medical records in outpatient practice on patient health outcomes. We develop a model of doctor-patient interaction that reflects findings in the medical and sociology literature underscoring the importance of communication for successful medical therapy. In this framework, differential patient compliance with medication therapy is influenced by the physician-patient interaction and is the result of an optimization process by the patient. The conceptual framework laid out in the model has empirically testable implications. This study presents some empirical evidence on patient health outcomes and patient health behavior before and after the introduction of EMRs that is consistent with the predictions of the theoretical model. A clear direction for future research is to

19

use data on individual physician-patient interactions to conduct sharper and more detailed empirical analysis to test the implications of our model.

References [1] Aghion, P. and Tirole J. (1997), "Real and Formal Authority in Organizations", Journal of Political Economy [2] Angst, Corey, Sarv Devaraj and John D’Arcy (2012) “Dual Role of IT-Assisted

Communication in Patient Care: A Validated Structure-Process-Outcome Framework” Journal of Management Information Systems, Volume 29, Number 2 / Fall 2012 [3] Arrow, Kenneth (1963) “Uncertainty and the Welfare Economics of Medical Care” The American Economic Review, Volume 53, Issue 5 [4] Ashton CM, Haidet P, Paternitti DA, Collins TC, Gordon HS, O’Malley K, Petersen LA, Sharf BF, Suarez-Almazor ME, Wray NP, Street RL., Jr (2003), "Racial and Ethnic Disparities in the Use of Health Services: Bias, Preferences, or Poor Communication?", Journal of General Internal Medicine.2003;18(2):146–52. [5] Athey, Susan and Scott Stern (2003) “the Impact of Information Technology on Emergency Health Care Outcomes” Journal of Economics, 33 (3), Autumn 2002, pp. 399-432. [6] Balsa, Ana and Thomas G. McGuire (2001), "Statistical discrimination in health care", J Health Econ. Nov;20(6):881-907. [7] Balsa, Ana and Thomas G. McGuire (2003), "Prejudice, clinical uncertainty and stereotyping as sources of health disparities", Journal of Health Economics 22 (2003) 89–116 [8] Balsa, Ana, Thomas G McGuire, and Lisa S Meredith (2005), "Testing for Statistical Discrimination in Health Care", Health Serv Res. 2005 February; 40(1): 227–252. [9] Cebul, Randall, Thomas Love, Anil Jain and Christopher Hebert (2011) “Electronic Health Records and Quality of Diabetes Care” The New England Journal of Medicine, VOl 365, Issue 9, September 2011 [10] Cooper LA, Roter DL, Johnson RL, Ford DE, Steinwachs DM, Powe NR. (2003), "Patientcentered communication, ratings of care, and concordance of patient and physician race", Ann Intern Med. 2003 Dec 2;139(11):907-15. [11] DesRoches, Catherine, Eric Campbell, Christine Vogeli, Jie Zheng, Sowmya Rao, Alexandra Shields, Karen Donelan, Sara Rosenbaum, Stefanie Bristol and Ashish Jha (2010) “Electronic Health Records’ Limited Success Suggests More Targeted Uses” Health Affairs, April 2010, Vol 29 Issue 4 [12] Dewatripont, M and Jean Tirole (2005) "Modes of Communication", Journal of Political Economy, vol. 113, n. 6, December 2005, p. 1217-1238. 20

[13] Ellickson Paul, Scott Stern, Manuel Trajtenberg (1999), "Patient Welfare and Patient Compliance: An Empirical Framework for Measuring the Benefits from Pharmaceutical Innovation", NBER Working Paper No. 6890 [14] Escarce, J. (2005), "How Does Race Matter, Anyway?", Health Serv Res. 2005 February; 40(1): 1–8. [15] Goldman, Dana and James P. Smith. "Socioeconomic Differences In The Adoption Of New Medical Technologies," American Economic Review, 2005, v95, 234-237 [16] Hubbard, Thomas and George Baker “Make or Buy in Trucking: Asset Ownership, Job Design, and Information”, American Economic Review, June 2003 [17] Hubbard, Thomas and George Baker "Contractibility and Asset Ownership: On-Board Computers and Governance in U.S. Trucking." Quarterly Journal of Economics, November 2004 [18] Hubbard, Thomas and George Baker "Contractibility and Asset Ownership: On-Board Computers and Governance in U.S. Trucking" American Economic Review, September 2003 [19] Javitt, J. Rebitzer, J. and Reisman, L (2008), “Information Technology and Medical Missteps: Evidence from a Randomized Trial,” Journal of Health Economics, 27(23):585-602 (May 2008) [20] Kaplan, S.H., Greenfield, S., Ware, J.E., 1989. Impact of the doctor–patient relationship on the outcomes of chronic disease. In: Stewart, M., Roter, D. (Eds.), Communicating with medical patients. Sage, Newbury Park. [21] Koulayev, Sergei, Emilia Simeonova and Niles Skipper (2013) “Who Is in Control? The Determinants of Patient Adherence with Medication Therapy” Working Paper, March 2013 [22] LaVeist, Thomas, Kim Nickerson and Janice Bowie (2000) "Attitudes about Racism, Medical Mistrust, and Satisfaction with Care Among African American and White Cardiac Patients," Medical Care Research and Review, Vol 57 Supplement 1: 146-161 [23] Jinhyung Lee, Jeffery S. McCullough, and Robert J. Town (2012) “The Impact of Health Information Technology on Hospital Productivity” NBER Working paper 18025, April 2012 [24] Leonard, Kenneth and Joshua Graff-Zivin (2005) “Outcome Versus Service Based Payments in Health Care: Lessons from African Traditional Healers” Health Economics, 14: 575-593, 2005 [25] McCullough, Jeffery, Stephen Parente and Robert Town (2013) “Health Information Technology and Patient Outcomes: The Role of Organizational and Informational Complementarities” NBER Working Paper 18684, January 2013 [26] Miller, Amalia and Catherine Tucker (2011) “Can Health Care Information Technology Save Babies?” Journal of Political Economy, April 2011 [27] Mooney, Gavin and Mandy Ryan (1993) “Agency in Health Care: Getting Beyond First Principles” Journal of Health Economics, 12, 125-135 21

[28] Simeonova, Emilia (2008) “Doctors, Patients, and the Racial Mortality Gap” Columbia University Discussion Paper 0708-13, 2008 [29] Simeonova, Emilia (2013) “Marriage, Bereavement and Mortality”, Journal of Health Economics, 2013 [30] Scott, Anthony and Sandra Vick (1999) “Patients, Doctors and Contracts: An Application of Principal-Agent Theory to the Doctor-Patient Relationship” Scottish Journal of Political Economy, Vol. 46, No 2, May 1999 [31] Stern, Scott and Manuel Trajtenberg (1998), "Empirical Implications of Physician Authority in Pharmaceutical Decisionmaking", NBER wp 6851 [32] van Ryn M, Burke (2000), "The effect of patient race and socio-economic status on physicians’ perceptions of patients", J.Soc Sci Med. 2000 Mar;50(6):813-28. [33] Vick, Sandra and Anthony Scott (1998) “Agency in Health Care. Examining Patients’ Preferences for Attributed of the Doctor-Patient Relationship” Journal of Health Economics 17, 587-605 [34] Waitzkin, H. (1984). “Doctor–patient communication. Clinical implications of scientific and social research.” JAMA. 17, 2441–2446 [35] Waitzkin, H.(1985) “Information giving in medical care.” J. Health Social Behavior 26, 129–146.

22

Appendix: Proofs Proof of Lemma 1. We solve the unconstrained version of the optimization problem and then verify that in the optimum both constraints are satisfied. After differentiating, the first order condition is: 2𝜁𝐶𝐵 − 2𝜁𝐶 2 (1 − 𝑒𝜁) = 𝐵𝑒

(1)

Since 𝐵 > 𝐶, the restriction 𝑒 ∗ > 0 implies: 𝐵 − 2𝜁 2 𝐶 2 > 0, or 𝜁 2 < 𝐵/(2𝐶 2 ). Since 1/(2𝐶) < 𝐵/(2𝐶 2 ), this condition is satisfied by assumption 1. We also need to verify that (1 − 𝑒 ∗ 𝜁) ≥ 0. From the first-order condition, 𝐵−𝐶 𝐵 − 2𝜁 2 𝐶 2 𝐵 − 2𝜁 2 𝐶 2 − 2𝜁 2 𝐶𝐵 + 2𝜁 2 𝐶 2 = 𝐵 − 2𝜁 2 𝐶 2 𝐵(1 − 2𝜁 2 𝐶) = ≥0 𝐵 − 2𝜁 2 𝐶 2

(1 − 𝑒 ∗ 𝜁) = 1 − 2𝜁 2 𝐶

Since 𝐵 − 2𝜁 2 𝐶 2 , we need (1 − 2𝜁 2 𝐶), which is again implied by assumption 2. The effort is always monotonically increasing with 𝜁. The derivative of the optimal effort with respect to 𝐶 is: (𝐵 − 2𝐶)(𝐵 − 2𝜁 2 𝐶 2 ) − 𝐶(𝐵 − 𝐶)(−4𝜁 2 𝐶) 𝜕𝑒 ∗ = 2𝜁 (𝐵 − 2𝜁 2 𝐶 2 )2 𝜕𝐶 2 2 𝐵 + 2𝜁 𝐶 − 2𝐶 = 2𝜁𝐵 (𝐵 − 2𝜁 2 𝐶 2 )2 whose sign is determined by (𝐵 + 2𝜁 2 𝐶 2 − 2𝐶). Solving 𝐵 + 2𝜁 2 𝐶 2 − 2𝐶 = 0, we obtain the first 1 2𝐶−𝐵 1 threshold 𝜁̄12 = < . If 𝜁 > 𝜁̄1 we have 𝜕𝑒 ∗ /𝜕𝐶 > 0 - the effort increases with cost of mistake. 2𝐶

𝐶

2𝐶

Since 𝜁 > 0, this constraint matters only if 𝜁̄1 > 0, or 𝐵 > 𝐶 > 𝐵/2; for lower costs, 𝐶 < 𝐵/2, we have 𝜕𝑒 ∗ /𝜕𝐶 > 0 for all 𝜉 > 0. QED. Proof of Proposition 1. The expression for the expected cost comes from the first-order condition (1): 𝐵 𝑒 (1 − 𝑒𝜉)𝐶 = 𝐵 − 2𝜉 𝐶 from which we see that expected cost increases with introduction of EMR only if the effort falls faster than 𝐶, i.e., 𝜕(𝑒/𝐶)/𝜕𝐶 > 0. Differentiating 𝑒/𝐶, 𝜕(𝑒/𝐶) −(𝐵 − 2𝜁 2 𝐶 2 ) − (𝐵 − 𝐶)(−4𝜁 2 𝐶) = 2𝜁 (𝐵 − 2𝜁 2 𝐶 2 )2 𝜕𝐶 2 2 2 −𝐵 − 2𝜁 𝐶 + 4𝜁 𝐶𝐵 = 2𝜁 (𝐵 − 2𝜁 2 𝐶 2 )2 −𝐵 + 2𝜁 2 𝐶(2𝐵 − 𝐶) = 2𝜁 >0 (𝐵 − 2𝜁 2 𝐶 2 )2 23

1 𝐵 This inequality holds if 𝜉 2 > 2𝐶 2𝐵−𝐶, which gives the second threshold, 𝜁̃2 . We should have 𝜁̃2 > 𝜁̃1 ,

because 𝜕𝑒 ∗ /𝜕𝐶 > 0 is a necessary condition for 𝜕(𝑒/𝐶)/𝜕𝐶 > 0, but not a sufficient one. Indeed, 1 𝐵 2𝐶 2𝐵 − 𝐶 2𝐵𝐶 2𝐵𝐶 (𝐵 − 𝐶)2

1 2𝐶 − 𝐵 2𝐶 2𝐶 > (2𝐵 − 𝐶)(2𝐶 − 𝐵) > 5𝐵𝐶 − 2𝐶 2 − 2𝐵2 >0 >

1 We also need to verify that the threshold is binding 𝜁̃2 , i.e. 𝜁̃2 < 2𝐶, which is an upper limit on 𝜉, by

assumption 1 - which ensures an interior solution in this model. 1 𝐵 1 < 2𝐶 2𝐵 − 𝐶 2𝐶 𝐵 < 2𝐵 − 𝐶 0<𝐵−𝐶 QED.

24