Online supplement for Health Information Exchange as a Multisided Platform: Adoption, Usage and Practice Involvement in Service Co- Production A. Multisided HIE Platforms The value created by a HIE to members on any side is not only a function their own characteristics, but also depends on the members on other sides of the platform. Increase in the membership at each side can create positive or negative direct network effects among the members in the same side as well as indirect network effects among the members at the other sides. The direct effects are represented by the same-side externalities and the indirect effects by the cross-externalities. These effects are illustrated and denoted as positive or negative in Figure 1. We discuss the effects originating from these sides as follows. A.1. Patients The availability of the records to other users of an exchange is controlled by the informed consent of the patients (Goldstein 2010), and such consents are usually given with different levels of availability constraints. As the level and volume of patient consents increase, more data becomes available on the system. Hence, the sharing will be greater and the value of the HIE to its other participating user types will increase, leading to better service that patients receive from healthcare providers (Adler-Milstein et al. 2011). Furthermore, an increase in the level and volume of patient consents would increase the quality of healthcare services and reduce the probability of redundant tests. Both will eventually result in lower costs of healthcare services which payers including private insurance companies or state and governmental payers such as Medicare and Medicaid will benefit. On the other hand, although healthcare payers and providers will enjoy positive effects of indirect externalities from increased 1
number of patients with consent, the data providers such as laboratories and radiology centers will lose a part of their potential revenue by the decrease in the number of potential patients as the customers of their services. This happens due to the reduction in the number of redundant tests and increase in better care which also reduces the need for extra lab, radiology tests and other surplus clinical work. Finally, same-side direct network effects on the side of the patients are in general not significant. A.2. Healthcare Providers When more practices and physicians join a HIE and access medical data, the probability of receiving better care increases for patients. With a high number of physicians with access to previous medical records, patients will undergo fewer tests, receive more rapid healthcare service which in many emergency cases, may be vital for them. The better healthcare service and increased performance of healthcare providers will significantly lower the healthcare cost which benefits the payers. However, an increase in the number of physicians with access to previous medical records could also negatively affect the potential market share of laboratory and radiology centers in the same way that increased levels of patient consent do. The most interesting externality with this side is the direct network externality among physicians. When they become a member of HIE, the tests that they order will become available on the system and other physicians will be able to use them. In other words, the increased number of physician members will result in a richer medical dataset. For a detailed study of the network effects among physicians refer to (Yaraghi, et al. 2013a, 2013b). In conclusion, we can expect positive crossexternalities from the side of Healthcare Providers to the sides of the Patients and Payers, a potential negative effect on the side of Medical Data Providers, and positive same-side effects due to the reasons cited above. 2
A.3. Medical Data Providers As membership of medical data providers in a HIE increase, the chance of creating digital health records on the HIE platform also increase. This would positively affect both patients and healthcare providers. Patients would have a larger portion of their medical history online and thus will receive the benefits of HIE in increased healthcare quality and reduced costs at higher levels. In a similar way, with more data providers on the HIE system, healthcare providers can access larger pools of medical data of their patients and thus would be able to provide better care at lower costs. As discussed earlier, this would again benefit insurance companies and other payers by reducing the chances of paying for redundant tests. More importantly, the availability of more thorough and comprehensive medical histories reduces the chances of occurrences of unusual medical complications caused by wrong diagnoses and prescriptions. This would eventually reduce the healthcare costs for payers. When more data providers join a HIE and contribute to its digital database of medical records, less patients would need surplus tests and lab work. This happens due to the availability of previous medical records that reduces the chances of re-ordering redundant tests. Further, more comprehensive medical histories help physicians to make better decisions and provide better care which in turn would reduce the possibility of extra tests which would otherwise be administered based on wrong diagnoses and practices. Thus, the membership of more medical data providers in a HIE could create negative direct network externality among other data providers. In summary, we can expect positive cross-externalities from the side of Medical Data Providers to the sides of the Patients, Payers and Healthcare Providers, and a potential negative same-side effect due to the reasons cited above. A.4. Payers A significant value offered by a HIE to the participants on the Payers side is the capability it affords them to better control the quality of healthcare services and manage the billing and claims processes better 3
and smoother. The increase in membership on the Payers side increases the likelihood of better quality control over the healthcare services provided by medical providers which would result in better care for patients. It also enhances the precision and speed of coverage payments to medical providers as well as major data providers. A HIE provides a unique and rich pool of data which payers can utilize to better analyze the cost-effectiveness of their coverage policies and investigate the effects of many different options in healthcare coverage. As the number of members in the Payers side increases, their collective business intelligence leads to more sophisticated data analysis towards better coverage policies and healthcare services recommendations. In conclusion, we can expect positive cross-externalities from the side of Payers to the sides of the Patients, Healthcare Providers and Medical Data Providers and positive same-side effects due to the reasons cited above. As the time in which we collected the data, the practices could not push their medical data into the HIE system. All of the practices downloaded the data which is provided by labs, radiology centers and hospitals. Despite the fact that no direct exchange happened between the practices, they could access the data which others had previously ordered and was created by data providers on HIE system. As the number of member practices increases, the number of accessible medical documents on HIE system also increases. This is due to the fact that member practices encourage their patients to provide consent and allow their medical records to be shared with HIE members and thus the potential value of HIE for members increases. 4
Figure 1.Network externality effects among HIE members B. The HEALTHeLINK Platform HEALTHeLINK is the regional HIE that facilitates the electronic sharing of medical data among healthcare providers in Western New York using web portal technologies. The underlying HIE project began in 2004, led by the Buffalo Academy of Medicine. The project actively engaged Western New York physician community, SUNY at Buffalo, county and state public health departments, HealtheNet and UNYPHIED (Upstate New York Professional Healthcare Information & Education Demonstration Project), and was supported by a grant from the Community Health Foundation of Western and Central New York. HEALTHeLINK was created from these efforts in 2008 and has since been supported by funding from the Heal NY program of New York State. HEALTHeLINK is a collaborative effort among community healthcare providers, large hospital systems, major laboratories and radiology centers, and regional insurance providers. 5
Since its establishment in 2008, HEALTHeLINK has developed into a major four-sided platform as shown in Figure 1. Till recently, the exchange has attracted more than 2054 healthcare provider members within 430 practices. Healthcare professionals join HEALTHeLINK at the practice level. The medical data of over 500 thousand patients are available through HEALTHeLINK. All of the major hospitals, labs and radiology centers in Western New York have joined HEALTHeLINK and routinely push the data to the HEALTHeLINK database. Three insurance companies (payers) have joined HEALTHeLINK so far. HEALTHeLINK provides access to three types of patient medical records: Lab Reports, Radiology Reports and Hospital Transcriptions. Federal and state incentives cover the membership fees for practices so joining HEALTHeLINK is free for practices. When a patient visits a laboratory, radiology center or gets hospitalized, his medical records will be uploaded on a central data center managed by HEALTHeLINK. If the patient grants consent, then the participating practices will be able to view and download the patient s records through two different channels. The first channel is a fully automated channel which sends the digital data directly from labs, radiology centers and hospitals to the practices interoperable EMR systems. This is denoted as the Full-Service Channel. Practices will only receive the records of their own patients through this channel. The access to the records of the new patients is provided through the second channel: a web portal in which they have to manually search and download patient records. This is denoted as the Self-Service Channel. C. Data Processing and Analysis Procedure C.1. Degree Centrality Degree is the simplest yet the most appealing measure of in social network analysis (Ahuja et al. 2003). It reflects the number of other nodes that are directly connected to a particular node (Freeman 1979). Literature is abundant with evidence on the influence of social networks on innovation diffusion (Valente 1996, 2010, p. 14). Centrality is a measure of the prestige 6
and criticality associated with the position of a node in the network (Borgatti and Everett 2006) and is shown to be significant in influencing the behaviors of others in adopting new technologies and innovations (Carrington et al. 2005; Slater et al. 2007). Different metrics have been designed in the literature and each of these reflect different concepts with dissimilar interpretations (Freeman 1979). C.2. Betweenness Centrality Betweenness is the extent to which a node lies on the shortest paths between pairs of other nodes in the network (Freeman 1979). Nodes with high betweenness can influence the transmission of information among others by strategically withholding or distorting it (Shaw 1954). Social networks literature emphasizes the role of nodes with high betweenness in the maintenance of communication and their potential as key enablers of innovation diffusion (Grewal et al. 2006; Tucker 2008). Freeman (1979) introduced the simplest measure of betweenness. Consider a network of nodes. Assume that each link has a unit weight. Thus, the total weight of any path between a pair of nodes and is the number of links along this path. A shortest path between the pair nodes and is defined as a path with the minimum total path weight over all such paths connecting the pair in the network. Note that there could be several such shortest paths between nodes and. Let be the number of shortest paths between the pair. Let be the number of such shortest paths that contain a given node. The Freeman's (1979) measure of betweenness for node with respect to nodes and is. Despite its simplicity that has led to widespread use, Freeman s measure has two fundamental shortcomings. First, it does not take into account the fact that the weights on the links could vary in practical social networks since they represent the strengths of relationships between nodes. To overcome this, Brandes (2001) developed a new measure of betweenness where the 7
weights on the links are allowed to vary. This metric is more appropriate in networks where the link weights are of considerable importance in interpreting node relationships. The Brandes metric is computed similar to the Freeman metric, except that the total path weight is computed by adding the weights along a path. The details of this algorithm along with other alternative measures are discussed by Brandes (2008). Second, by adopting a constant unit weight for all links, the Freeman metric fixedly interprets paths with higher total path weights between a pair of nodes as weaker ties between them. However, in many social network contexts such as ours, higher total path weights could imply stronger ties. This is a clear limitation of the Freeman metric that is also efficiently overcome with the Brandes metric. In the current context of the network of common practitioners, we define the weight of a link as the inverse of the number of practitioners that are common to the two associated nodes. Hence, the paths with the minimum total weight represent the strongest ties between the two nodes. In this research, we employ the Brandes metric, computed by the Gephi software system (Bastian et al. 2009). We then used Gephi software (Bastian et al. 2009) to create two networks of common patients and common members and calculate the respective metrics. We used Gephi since it applies Brandes (2001) algorithm to calculate betweenness which as discussed in section 3.1 in the main paper is a much faster algorithm and also considers the weights of the links in calculating the betweenness. Both measures of are normalized so that they can be incorporated in a model together and the estimates of their coefficients be meaningfully interpreted. C.3. Service Value We merged the first two data sets described in the main paper, to create a panel data set of practices HIE access behavior. In this new data set, for each practice, we identify the number of access times to each medical record in each month through each channel. Monthly access to each of 8
the three different types of medical records (services) is considered as a proxy for the value of that specific service to each practice at the time. C.4. Urban/Rural Location The Population data set contains the population of 42 cities in which different practices are located in. The cities with a population of less than 17,000 are considered rural while others are considered urban. C.5. Market Share Market share of each practice is calculated based on the population of the city in which it is located in and the population of healthcare providers (physicians, nurses, etc) that practice in the same city. The market share of practice with a size of and a total healthcare providers population of members, in a city that has a total population of, is calculated as C.6. Tenure Tenure with HIE for each practice is calculated as the number of months since the adoption date until August 2011. C.7. Nurse Ratio Nurse ratio, is the ratio of nurses, nurse practitioners and physician assistants to specialists and primary care physicians in each practice. C.8. Isomorphic Quotients In this study, we define two forms of isomorphism: Patient-centric and Practitioner-centric. In patient-centric isomorphism, practices that share a large number of patients with other practices are considered as large, compared to the rest. In practitioner-centric isomorphism, practices that share a large number of practitioners with other practices qualify as large. The large practices are the most influential change agents in the local healthcare market, and the smaller ones tend to emulate their HIE adoption and usage behaviors. In general, large hospital systems and many well- 9
established large-sized practices usually satisfy both these criteria. However, some practices could also uniquely qualify as large under only one of these criteria. To calculate the physician and patient isomorphic quotients, we first rank practices by the number of physicians and patients that they share with the whole community. The practices in top 5% in the rankings of shared physicians and patients are considered as major practices. The ratio of physicians and patients that each smaller practice shares with these major practices is considered as its physician and patient dependency on major practices. Let denote the set of all the practices and hospital systems in a community. Let denote the set of practices where the number of patients shared by each practice with the rest in the community is more than a threshold value. The practices in the set are considered large by the patient criterion. Similarly, let denote the set of practices where the number of practitioners shared by each practice with the rest in the community is more than a threshold value. The practices in the set are considered large by the practitioner criterion. Consider a practice Let and denote the total number of patients that practice i shares with the entire market and the corresponding set of large practices, respectively. Then, the ratio { is a measure of the influence the set has over along the patient dimension. We term this ratio as the Patient-centric Isomorphic Quotient. Similarly, we define for any practice the quantities and as the total number of practitioners that shares with the entire market and the corresponding set of large practices, respectively. Then, the ratio is a measure of the influence the set has over along the practitioner dimension. We term this ratio as the Practitioner-centric Isomorphic Quotient. Both these quotients range between 0 and 1. Practices with larger values of these quotients are more likely to be influenced by the respective large practices in their HIE adoption and usage behaviors. 10
Figure 2 graphically represents the operationalization of each of the variables and their use in the respective analyses. Table 1 summarizes these six data sets. Data set name Full-service Self-service Demand Affiliation Adoption Population Description Logs of access through full-service channel to medical records Logs of access through self-service channel to medical records Logs of medical documents ordered by all practices, regardless of their membership status Access permissions of each practitioner at multiple practices Name, location and adoption date of member practices Population of cities in which member practices are located Table 1: Description of 6 data sets used in the empirical analysis Population dataset Adoption dataset Selfservice dataset Full service dataset Demand data set Affiliation dataset Market share Location Tenure, Tenure 2 Service value Patients network Members network Nurse ratio Degree Patient-centric isomorphism Physician-centric isomorphism Betweenness Usage Analysis by Eq. (1) Coproduction involvement Adoption Analysis by Eq. (2) Figure 2: The flowchart of variable operationalization 11
1. LogTCiVHR 2. LogTCiHUB -0.1311 3. Logtimetoadopt -0.1020 0.27650 4. loglab 5. logradio 0.21144 0.19831 6. logtrans 7. Tenure 0.11211 8. Tenure 2 0.08093-0.0896 9. Rural 10. logmarketshare 0.03920 0.0010 0.02513 11. Nurse 0.0354 12. Degree 0.41045 13.Betweenness 0.06537 14. Practice efficiency -0.1788 15. Common patients -0.0669 with large practices 16.Common physicians with large practices -0.1535 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0.03351 0.0050 0.49755 0.32697 0.54217-0.2963-0.1873-0.0324 0.0084 0.04013 0.0008 0.14800 0.38992 0.08710 0.25674-0.2426 0.29582-0.1674 0.01932 0.1058 0.06787-0.4904-0.4438 0.0718 0.1232-0.0501 0.2819-0.0546 0.2407-0.2208-0.0979 0.0354-0.1629 0.0004-0.116 0.0126 0.0445 0.3397 0.60151 0.49113-0.0593 0.00400 0.7379-0.0791 0.09216 0.04266 0.0004 0.45119 0.05307 0.08776 0.30762-0.2532 0.62901-0.2377-0.1523-0.1307 0.00646 0.5884 0.14094 0.55859 0.09290 0.03204 0.0073 0.21924-0.2625-0.2956-0.2363-0.0385 0.0018 0.06069 0.21595 0.50684 0.14976-0.0385 0.0012-0.0900-0.2339 0.95799-0.0425 0.0005-0.0331 0.0056-0.1035-0.0370 0.0019-0.0506 0.1387 0.14589-0.0240 0.0442-0.0325 0.0082-0.0354 0.0030-0.1046-0.0337 0.0048-0.0528-0.0884 0.18772-0.0534-0.065 0.1586 0.0016 0.9713-0.092 0.0462-0.061 0.1845-0.035 0.4453-0.015 0.7362 0.0547 0.2402-0.0285 0.5406-0.1002 0.0311-0.0363 0.4353-0.1473 0.0015-0.0030 0.9481 0.04079 0.3817 Table 2: Correlation of the variables used in the empirical models 0.1100 0.0180 0.5191-0.041 0.3736-0.014 0.7517 0.0802 0.0848 0.1451 0.0018-0.148 0.0014-0.005 0.9107-0.207 0.099 0.0327-0.000 0.9999 0.1163 0.0124 0.00711 0.8789-0.0176 0.7046 0.03111 0.5048 12
D. Conceptual Model Betweenness H2-1 Physician isomorphism H3-2 In-degree H1-1 Time to adopt H1-3 H3-1 Rural location Patient isomorphism Figure 3: Conceptual model of HIE adoption Tenure Betweenness H2-2 H2-3 Usage In-degree H1-2 H1-4 Rural location Figure 4: Conceptual model of HIE usage 13
Betweenness H2-4 Efficiency H2-5 Tenure Figure 5: Conceptual model of practice efficiency (co-production) in using HIE References Adler-Milstein, J., Bates, D. W., and Jha, A. K. 2011. A survey of health information exchange organizations in the United States: implications for meaningful use. Annals of internal medicine 154(10) 666. Ahuja, M. K., Galletta, D. F., and Carley, K. M. 2003. Individual and performance in virtual R&D groups: An empirical study. Management Science 49(1) 21 38. Bastian, M., Heymann, S., and Jacomy, M. 2009. Gephi: An open source software for exploring and manipulating networks. Borgatti, S. P., and Everett, M. G. 2006. A graph-theoretic perspective on. Social networks 28(4) 466 484. Brandes, U. 2001. A faster algorithm for betweenness *. Journal of Mathematical Sociology 25(2) 163 177. Brandes, U. 2008. On variants of shortest-path betweenness and their generic computation. Social Networks 30(2) 136 145. Carrington, P. J., Scott, J., and Wasserman, S. 2005. Models and methods for innovation diffusion. Models and methods in social network analysis Freeman, L. C. 1979. Centrality in social networks conceptual clarification. Social networks 1(3) 215 239. Goldstein, M. M. 2010. Health information technology and the idea of informed consent. The Journal of Law, Medicine & Ethics 38(1) 27 35. Grewal, R., Lilien, G. L., and Mallapragada, G. 2006. Location, Location, Location: How Network Embeddedness Affects Project Success in Open Source Systems. Management Science 52(7) 1043 1056. Shaw, M. E. 1954. Group structure and the behavior of individuals in small groups. The Journal of Psychology 38(1) 139 149. Slater, M. D., Snyder, L. B., and Hayes, A. F. 2007. communication network analysis. The Sage sourcebook of advanced data analysis methods for communication research 243 273. Tucker, C. 2008. Identifying Formal and Informal Influence in Technology Adoption with Network Externalities. Management Science 54(12) 2024 2038. 14
Valente, T. W. 1996. Network models of the diffusion of innovations. Computational & Mathematical Organization Theory 2(2) 163 164. Valente, T. W. 2010. Social Networks and Health: Models, Methods, and Applications: Models, Methods, and Applications, Oxford University Press, USA. Yaraghi, N., Du, A. Y., Sharman, R., Gopal, R. D., and Ramesh, R. 2013. Network Effects in Health Information Exchange Growth. ACM Transactions on Management Information Systems (TMIS) 4(1) 1. Yaraghi, N., Du, A. Y., Sharman, R., Gopal, R. D., Ramesh, R., Singh, R., and Singh, G. 2014. Professional and geographical network effects on healthcare information exchange growth: does proximity really matter? Journal of the American Medical Informatics Association 21(4) 671 678 15