UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Robust performance benchmarking : an application of multivariate and data envelopment analysis at the… Tang, Kevin Berenato 2002

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-ubc_2002-0268.pdf [ 4.22MB ]
JSON: 831-1.0090302.json
JSON-LD: 831-1.0090302-ld.json
RDF/XML (Pretty): 831-1.0090302-rdf.xml
RDF/JSON: 831-1.0090302-rdf.json
Turtle: 831-1.0090302-turtle.txt
N-Triples: 831-1.0090302-rdf-ntriples.txt
Original Record: 831-1.0090302-source.json
Full Text

Full Text

ROBUST PERFORMANCE BENCHMARKING: AN APPLICATION OF MULTIVARIATE AND DATA ENVELOPMENT ANALYSIS AT THE WORKERS' COMPENSATION BOARD by K E V I N B E R E N A T O T A N G Bachelor of Commerce, University of Alberta, 2000 A THESIS SUBMITTED I N PARTIAL F U L F I L L M E N T OF T H E REQUIREMENTS FOR T H E D E G R E E OF MASTER OF SCIENCE (BUSINESS ADMINISTRATION) T H E F A C U L T Y OF G R A D U A T E STUDIES (Faculty of Commerce & Business A<dmiriistration) We accept this thesis as conforming to the required standard T H E UNIVERSITY OF BRITISH COLUMBIA December 2001 © Kevin Berenato Tang, 2001 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of QwAWJytfl * ^ ^ V I N I ^ O ^r5T>ib<M The University of British Columbia Vancouver, Canada Date DE-6 (2/88) ABSTRACT Compensation Services manages return to work, wage loss, pension and health-care benefits to injured workers, and those suffering from occupational diseases. This thesis presents the methodology and results of a comprehensive study conducted to determine the relationships between existing performance measures, quantify the factors influencing these performance measures, and deterrriine relative efficiencies across case management operations. We use principal components analysis, cluster analysis, and multiple regression to derive the relationships between performance outcomes and influencing factors. We then use data envelopment analysis, incorporating these multiple inputs and outputs, to assess overall relative efficiencies and set performance targets. The analysis has brought about an increased understanding of service delivery location performance and performance measurement. Results may be used to provide managerial decision support, communicate best practices, and serve as a basis for further efficiency or quality initiatives. The factors accounted for in the multivariate analysis can explain between 20% - 50% of the variability in key performance outcomes across case management desks. The overall efficiency analysis revealed strong performers both within case management offices and across regions. Four case management offices consistently contain strong-performing case management desks across several methods of evaluating efficiency. Transferring best practices has the potential to significantly increase relative efficiency improvements for case management desks across the province. TABLE OF CONTENTS ABSTRACT ii LIST OF TABLES iv LIST OF FIGURES v A C K N O W L E D G E M E N T S vi 1. INTRODUCTION 1 1.1 Background: Case Management at the Workers' Compensation Board 1 1.2 Problem Description 5 2. LITERATURE REVIEW 6 2.1 Benchmarking Compensation Systems 6 2.2 Analysis of Workers' Compensation Board Performance Indicators 8 2.3 Data Envelopment Analysis Approaches 9 3. M E T H O D O L O G Y 10 3.1 Project Approach 10 3.2 Brainstorming Sessions 11 33 Data Collection 14 3.4 Identifying Key Performance Measures 17 3.5 Benchmarking Key Performance Measures 18 3.6 Overall Efficiency Analysis 21 3.6.1 Data Envelopment Analysis - Introduction 21 3.6.2 The CCR Model 23 3.6.3 The B C C Model 29 3.6.4 Slacks-Based Efficiency Measurement 30 4. RESULTS 32 4.1 Client Service Manager Feedback 32 4.2 Principal Component Analysis Results 36 4.3 Multivariate Analysis Results 41 4.3.1 Cluster Analysis 41 4.3.2 Multiple Regression 44 4.4 Data Envelopment Analysis Results 52 5. CONCLUSIONS A N D SUMMARY OF R E C O M M E N D A T I O N S 61 6. FUTURE R E S E A R C H A N D APPLICATIONS 63 6.1 Modeling Extensions 63 6.2 Ongoing Efficiency Initiatives 64 REFERENCES 67 APPENDICES : 69 Appendix I: Data Elements 69 Appendix II: SAS Multivariate Procedures and Output 70 Appendix III: Detailed DEA Formulations 76 LIST O F T A B L E S Table 1: Data Assembly Methodology 16 Table 2: Sample of Claim Records 18 Table 3: Sample of Categorical Variables Summarized by Desk 19 Table 4: Performance Indicators and CSM Feedback 33 Table 5: Case Management Outcomes and Interrelated Measurement Variables 40 Table 6: Injury Cluster Groups 41 Table 7: Industry Cluster Groups 43 Table 8: Body Part Cluster Groups 43 Table 9: Simple Regression of Factors on Claims Closed 46 Table 10: Simple Regression of Factors on Avg. Time in Case Mgmt 47 Table 11: Simple Regression of Factors on Proportion Reopened 47 Table 12: Factors Influencing Claims Closed 48 Table 13: Factors Influencing Time in Case Management 49 Table 14: Factors Influencing Reopenings 50 Table 15: Comparison of D E A and Regression Rankings 55 Table 16: Ordinal Rank Comparison of D E A Formulations 58 Table 17: Efficient Units Comparison of D E A Models 58 Table 18: Summary of Recommendations 62 iv LIST OF FIGURES Figure 1: WCB Organizational Structure 2 Figure 2: Compensation Services Business Units 3 Figure 3: Service Delivery Locations 4 Figure 4: Project Components 10 Figure 5: CSM Survey Sample - Ranking Performance Indicators 12 Figure 6: CSM Survey Sample - Ranking Influencing Factors 13 Figure 7: Sample D E A Optimization Input using Microsoft Excel 22 Figure 8: CCR Example: 2 Outputs, 1 Input 23 Figure 9: Comparison of Efficiency Frontiers 30 Figure 10: Wage Loss Duration Incurred by Case Management 35 Figure 11: Magnitude of Principal Component Variables 37 Figure 12: Scree Plot of Principal Components 37 Figure 13: Distribution of Time Between Initial Closure and Reopening 40 Figure 14: Dendogram of Clustered Injury Variables 42 Figure 15: Distribution of Qaims Closed 44 Figure 16: Distribution of Average Time in Case Management 45 Figure 17: Distribution of Reopenings 45 Figure 18: Regression-Based Performance Targets for Claims Closed 50 Figure 19: Regression-Based Performance Targets for Avg. Time in Case Mgmt 51 Figure 20: Regression-Based Performance Targets for Proportion Reopened 51 Figure 21: Distribution of D E A Efficiency Scores 54 Figure 22: Distribution of Efficiency Scores by SDL 55 Figure 23: Comparison of Regression and D E A Based Performance Targets 57 Figure 24: Reference Set Frequency 59 Figure 25: Cycle of Performance Measurement & Control 65 V ACKNOWLEDGEMENTS I would like to thank the Centre for Operations Excellence (COE) for providing opportunities for applied research projects and financial assistance. C O E management and staff have been very supportive in assisting with project deliverables and conference presentations. I cannot overstate my gratitude to my project team: Jason Goto and Johnny Yeung. Jason has provided invaluable mentorship and encouragement, and I deeply appreciate his tutelage and support. Johnny's contributions were essential in facilitating successful data analysis and project management. Thanks to the Workers' Compensation Board of British Columbia's Compensation Division for sponsoring this project. Special thanks go to Executive Director Steve Baraett and Divisional Controller Brian Erickson for spearheading the project sponsor team, and the numerous client service managers and data resources who provided essential advice, feedback and data support. I would also like to acknowledge Dr. David Glenn for his time, input, and support as my thesis supervisor, and Dr. Chunyan Y u for agreeing to serve as external examiner. From the University of Alberta, special thanks go to Dr. Erhan Erkut for introducing me to management science and encouraging the pursuit of graduate studies, and Dr. Tarja Joro for introducing me to data envelopment analysis and providing feedback during the initial stages of the project. I wish to thank my family and friends for their encouragement and support during my graduate studies at the University of British Columbia. And finally, to my classmates in the COE-sponsored Master of Science program for the friendship and energy expended in extracurricular activities, encouragement and comic relief when things started to get crazy. Your support during academic and project challenges was very much appreciated. I wish you all the best in future endeavours. VI 1. INTRODUCTION 1.1 Background: Case Management at the Workers' Compensation Board The Workers' Compensation Board (WCB) is an administrative agency that operates under the authority of the Workers Compensation Act of British Columbia. The Workers Compensation Act is provincial legislation that protects injured workers, their dependents, and their employers from financial hardship in the event of a workplace injury or disease. The WCB promotes occupational health and safety practices, provides rehabilitation and return-to-work plans for injured workers, administers indemnity payments, and protects public interest of work environments. The WCB is funded completely by employers in industries covered by the Workers' Compensation Act. In return for employers funciing the system, employees cannot sue an employer or another worker for a workplace injury or disease. The WCB is organized into five main divisions: Prevention, Rehabilitation and Compensation Services, Finance and Information Services, Legal Services, and Human Resources. Rehabilitation and Compensation Services is the division of the WCB that manages return-to-work, adjudicates wage loss claims, and administers wage loss payments, pension, rehabilitation, and health care benefits to injured workers, those suffering from occupational diseases, and the dependents of deceased workers. 1 Figure 1: W C B Organizational Structure Panel of Adminstrators Chief Executive Officer 1 I 1 Prevention Rehabilitation & Finance & Legal Human Compensation Information Services Resources Services Systems Rehabilitation & Compensation Health Care Services Services Figure 1 displays the main organizational structure of the WCB, highlighting the division primarily associated with this study. In addition to Compensation Services, we also utilized resources in the Finance & Information Systems, Legal Services, and Human Resources for data requirements and confidentiality agreements. When a worker is injured or suffers from an occupational disease, the worker, employer or doctor reports the injury to the WCB and starts a workers' compensation claim. The injured worker must have been working when injured and the injury must have been caused by something to do with his or her job in order to be covered by the WCB. For a disease, the disease contracted must be caused by the work or the work environment itself in order to be covered by the WCB. The WCB covers both physical and psychological injuries. Approximately 180,000 claims are received every year. 2 Figure 2: Compensation Services Business Units Claim Types 1 • • I • Opportunity: Avoidable Delay Time Call Centre Entitlement Case Management Costs: Resources in each Business Unit . _ __ , Figure 2 describes the flow of claim processing through Compensation Services. These units used to be integrated, but are now in separate business units, although some business units are still collocated in the same physical location. Claimants move between the business units based on their claim types and the severity of the injury. Claim types can be classified as straightforward, complex, and activity-related soft tissue disorder (ASTD) claims. ASTD describes soft tissue disorders that result from repetitive movements or exposure to mechanical vibrations, such as bursitis, tendonitis, or carpal tunnel syndrome. Qaimants can also re-enter the system after wage loss is discontinued, due to a reaggravated injury or return-to-work complications. These claims are known as reactivations, or reopenings if additional wage loss payments are involved. The first point of contact is usually with one of four regional call centres, located in Richmond, Victoria, Kelowna, or Prince George. People with short-term claims (less than 3 three weeks) will usually stay within the Call Centre and Entitlement Units. At the Call Centre, questions about claim status are answered and entitlement decisions on straightforward claims are made. At the Entitlement Unit, decisions are made on straightforward and complex claims, and an Entitlement Officer would also manage straightforward cases. Figure 3: Service Delivery Locations C ran brook Vancouver Nelson 4 17 Service Delivery Locations (SDL) are located throughout the province, with each SDL grouped into a region. Case Management operations in each SDL manage longer-term, complex claims, as well as entitle and manage ASTD claims. There are three Regional Directors which oversee the Lower Mainland, Vancouver Island / Terrace, and the Interior locations. The primary resources in each SDL are the Case Managers, who are responsible for adjudicating and aciWnistering wage loss payments, managing return-to-work plans, and coordinating assessments with Medical and Nurse Advisors. There are between two to thirteen Case Managers in each SDL, structured into teams based on major employers and, more recently, ASTD specializations. Team Assistants support the case management process as well as assist Vocational Rehabilitation Coordinators that may be located at the SDL. Client Service Managers supervise the Case Managers and are responsible for their respective SDLs. Case Management handles approximately 30,000 short term disability (STD) claims received by the WCB. 1.2 Problem Description The primary objective of this study is to conduct an internal benchmarking study of case management operations to assess relative efficiencies of service delivery locations. In order to evaluate these efficiencies, we must compute and incorporate the factors influencing case management efficiency. This involves deterrruning the key performance outcomes and discovering the relationships that influencing factors have on these outcomes. Since Case Management incurs the most costs and highly specialized personnel, our efforts to understand factors driving performance outcomes would be best utilized in this business unit. 5 The analysis incorporates 109 case management desks in the 17 SDLs, for activity performed in the year 2000. Performance comparisons are limited to SDLs in British Columbia because the compensation system is different throughout Canada, so comparisons with out of province case management operations may not be appropriate due to different premiums, benefits, and services available in different Workers' Compensation Boards. This study was conducted under the auspices of the Centre for Operations Excellence (COE). The C O E is an applied research centre operating in the Faculty of Commerce at the University of British Columbia. The Centre integrates faculty, post-doctoral fellows, and M.Sc. students with industry partners through applied research projects. Project teams work with the industry partners to solve technically challenging business problems through the use of Management Science and Operations Research techniques. The C O E has been involved with the WCB in projects with the Call Centre (Sanegre, 1998), Risk Management (Urbanovich, 1999) and Case Management (Lin, 2000) using statistical and simulation analysis. This study leverages the knowledge and contacts gained through these projects. 2. LITERATURE REVIEW 2.1 Benchmarking Compensation Systems The Workers Compensation Research Institute (WCRT) has been analyzing performance, efficiency, and costs in workers' compensation systems in the United States since 1983. The WCRI is an independent, not-for-profit research organization that provides high-quality, objective information about public policy issues in workers' compensation. The WCRI has 6 several benchmarks for compensation system performance by state1. The Association of Workers' Compensation Boards of Canada2 has published similar comparative studies relating to compensation benefits and legislation differences between the 12 provinces and territories in Canada. These metrics include benefit costs, administrative and medical/legal expenses, timeliness, case closures, and medical provider utilization. Telles and Neal (1999) examined the trends of 75 performance indicators associated with these metric categories for Massachusetts, in conjunction with the CompScope multi-state benchmarking study published by the Workers Compensation Research Institute. The analysis highlights differences in legislation, medical resource availability and utilization, and litigation within Massachusetts based on claims evaluated over 12 and 36 month maturity levels (i.e. evaluating groups of claims 1 or 3 years after the claim was started). Although this study reports apparent trends in changing costs in different sectors of compensation systems, it does not offer predictive or other hypothetical models relating to the relationships behind compensation performance outcomes such as indemnity benefits and duration which drive these costs. Another key difference between this study and our research is that the WCRI study is treating mdernnity benefits, duration, and other cost drivers as controllable factors, whereas we consider these to be process outcomes. 1 http:// www, wcrinet. org/benchmarks .html 2 http://www.awcbc.org/ 7 2.2 Analysis of Workers' Compensation Board Performance Indicators Macfarlane and Weltz (1998) examined the set of performance indicators used by the Workers' Compensation Board of British Columbia, to detenriine the extent to which performance indicators are consistent with accepted criteria for "effectiveness", mcluding the extent to which they address relevant aspects of WCB services and outcomes, and the extent to which they provide clear, accurate and useful information to the WCB with respect to current performance and changes needed. This report studies key performance indicators at the corporate level, as well as indicators in place for the Prevention Division, the Rehabilitation and Compensation Division, and the Finance and Information Services Division. Macfarlane and Weltz report that although the existing set of performance indicators are a significant improvement over previous WCB reporting and outcomes, some performance indicators are too broad and cannot be tied to changes in program design. The study also points out data accuracy issues and a need for monitoring trends and interactions in this complex environment. Issues concerning effective measures, data, and reporting concerns are discussed in the Methodology section of this thesis, and trends and interactions are reported in the Results chapter. Macfarlane and Weltz state that a major gap in the information available is the general absence of a systematic approach to performance evaluation. This thesis addresses this concern, and attempts to provide such a framework through incorporating quantitative techniques such as multivariate and data envelopment analysis. 8 2.3 Data Envelopment Analysis Approaches There is an abundance of literature in statistics, economics, and operations research describing quantitative modeling approaches to efficiency measurement and setting performance targets. Farrell (1957) introduced productive efficiency as a production possibility set consisting of the conical hull of input-output vectors. Charnes, Cooper, and Rhodes (1978) expanded upon this concept to incorporate multiple outputs, introducing data envelopment analysis as a methodology for evaluating efficiency. Subsequent developments in data envelopment analysis are referenced in the Methodology section of this thesis. Although there is no documentation describing applications to workers' compensation systems, several examples of data envelopment analysis in health care and insurance industries can be applicable to this type of internal benchmarking problem, as the business models for these organizations are relatively similar. Siddharthan, Ahern, and Rosenman (2000) used data envelopment analysis to measure relative technical efficiencies of health maintenance organizations. This objective can be considered to be analogous to measuring service delivery location efficiencies, as both organizations incorporate durations and customer types (among other factors) when considering effective performance measures. Ozcan (1998) demonstrated significant cost savings for otitis media (ear inflammation) treatment by physicians, changing controllable input and output factors towards targets based on physicians' best practices. A similar application within the WCB could be determining controllable factors based on case management best practices for rehabilitation and return-to-work treatment plans. This thesis develops the framework for applying D E A 9 techniques to case management operations. The benefits of this approach are featured in the Results section of this thesis. 3. METHODOLOGY 3.1 Project Approach Figure 4: Project Components Project Definition I Brainstorming I Collect Data Identify Key Performance Metrics Performance Metric Benchmarking Overall SDL Efficiency Analysis I Development of Implementation Strategies Project Reporting Figure 4 displays the analysis and project reporting phases associated with this study. This project was initiated in April 2001, with the bulk of the research and analysis being conducted between June and August. The project schedule was managed and updated to accommodate WCB data and personnel availability, while acknowledging the dependencies inherent in the analysis and project reporting. 10 The dark diamond symbols in Figure 4 represent project updates and interim reports presented to senior management and Client Service Managers representatives from each region. The Project Definition phase involved working sessions with WCB senior management to clarify and finalize the project scope, objectives, timeline, resource requirements, and deliverables. Subsequent project components are described in detail throughout this chapter. 3.2 Brainstorming Sessions The Brainstorming phase consisted of meetings with case management personnel to understand how SDL performance is reported, and how this information is currendy used. The primary objective was to collect ideas concerning how available information could be modified to be more effective, identify gaps in reporting, and obtain input on information that is not currently captured that may be useful. A similar approach was taken when the WCB contracted with Nexus Actuarial Consultants (1996) to conduct a survey of Canadian Worker's Compensation jurisdictions. The key distinction is that our method surveyed internal resources on performance measures and relationship conjectures related to case management, as opposed to high-level measures for the entire WCB reported by non-British Columbia workers' compensation system personnel. A series of interviews and working sessions with Regional Directors, Quality Client Service Managers, and Client Service Manager representatives from across the province were held to identify and analyze measurement issues associated with effective case management. The C O E team facilitated meetings involving the majority of the Client Service Managers (CSMs) 11 from each of the regional groups (Lower Mainland, Vancouver Island, Interior). Surveys were used to follow up and verify feedback generated during these sessions. Figure 5: CSM Survey Sample - Ranking Performance Indicators Performance s McaMirc '-»L)i.Trnilion. •' i . " Data Source- • <* J ' • It.inkinu Volume Opened Number of Incoming Claims routed to the SDL CaRRs 1 2 3 4 5 6 7 Volume (loicd'i * Number of Claims Closed in lifile- 1 V>2--3 4 5 6 7 •BHnBMHHiEi Admin Costs Admin costs per claim CaRRs 1 2 3 4 5 6 7 C Liiim C'osN -STDCiv .U - H C Costs Dollars Paid on \ \ age 1 oss, and I lealth-f are Only ( lamv. on a " " pet claim average lilllllliH 1 2 3 4 5 6 7 Aged Claims Inventory Average Age of Wage Loss Claims * Avg. Number of Wage Loss Claims Compensation Systems Reports 1 2 3 4 5 6 7 Reopened ( lainis Niimbei ot Reactivated (liims fro in prev iously closed chinm 4 5 6 7 Income Continuity Percentage of Claims paid within 17 days of Injury Date Data Warehouse 1 2 3 4 5 6 7 \ppeals .1 he proportion appealed oKlanm denied at the SDL 1 . 2* 3 4 5 o -Successful Appeals The proportion of appealed claims that is successfully overturned Appeals Database 1 2 3 4 5 6 7 12 Figure 6: CSM Survey Sample - Ranking Influencing Factors Factors Definition Ranking Incoming Profile Percentage IrumCall Centre and Entitlement.Officc , The pioporlion of inuommi; claims, louted from the call center and entitlenient 1 2 3 4 5 0 7 Incoming Profile Claimant Age The amount of time the claim has already spent prior to arrival to SDL 1 2 3 4 5 6 7 Incoming Profile ( Lununi Ulucation 1 he level ofeducdlion ol Ihe worker •.l:y?2€£3 . 4 fs" 6- 7 •• Incoming Profile Clamant Gender The gender of the worker 1 2 3 4 5 6 7 Incoming Profile Seventy Index 1 he scvcnt> code ol the n.itun.' of injury 1 2" 1 4 5 f» 7 Incoming Profile Industry The industry code associated with the claim 1 2 3 4 5 6 7 Incoming Profile Industry profile of a region. Incoming Profile FYnploju-. 1 he enlplc^BcWelassbCialod with the Lmplover base ol the icmon Incoming Profile Geography The location of the work site 1 2 3 4 5 6 7 Incoming Profile Location of third party providers Incoming Profile Residence of the claimant Incoming Profile Migration patterns of claimants Incoming Profile \ \ . IgL RltC The WagFRate assouated with tK c l a i m ^ !l; \2' 3 4 5 c 7 Incoming Profile Other Demographics Area Population Employment Rate, Injury Rate 1 2 3 4 5 6 7 Incoming Profile Percentage without \ \ age The percentage of claims inventory without \ \ auc 1 O T S Figures 5 and 6 are sample pages of the questionnaire sent to the CSMs foUowing the initial working sessions. The rarddngs are ordinal, with a response of 1 representing not at all reflective of case management performance, and a response of 7 representing extremely reflective of case management performance. The results of this feedback can be found in Section 4.1. The investigation of issues raised in the CSM working sessions is limited by the available data. The analysis involved testing hypotheses where the data associated with potential influencing factors is accessible, quantified, and interpretable. Data source details 13 for entries listed in Figure 5 are described in the Data Extraction and Assembly section of this thesis. After the initial working sessions, CSM representatives from each area were contacted on a regular basis to offer additional hypotheses and anecdotal evidence of factors influencing performance measures. CSM representatives also provided criticisms of existing measurements and reviewed interim analysis results to provide face validity and the opportunity for feedback and corrective modifications to the modeling approaches. 3.3 Data Collection Representatives from Statistical Services, Compensation Systems, Compensation Finance, the Information Systems Division (ISD), Human Resources, and the Program Evaluation and Research Unit (PERU) were consulted to identify data sources, understand how information is currently distributed, and identify known measurement issues with respect to Case Management. Statistical Services computes claims statistics for reporting to the WCB Annual Report, the Association of Workers' Compensation Boards of Canada (AWCBC), and several other published reports and statistics on financial, injury, and industry related trends. Compensation Systems publishes monthly key performance indicators for case management, which are broken down by SDL and Case Management Desk. Compensation Finance summarizes administrative budget and expense information for the Compensation Division. ISD maintains the computer systems, and develops new systems to support service delivery, 14 administrative / decision support activities, and management information systems. Human Resources maintains the employee database, which includes work histories and experience levels for case management personnel. P E R U conducts studies on rehabilitation, work conditioning, and health care providers. P E R U also mamtains Overall Customer Satisfaction surveys administered to claimants by Ipsos-Reid, a third party market research firm. Following meetings with these department representatives, we developed a data request. Access to the Enterprise Data Warehouse was granted to facilitate the ongoing querying of detailed claim activity, employer profiles, and injury data. WCB personnel from the aforementioned departments provided us with data extracts from sources described in Table 1. Table 1 also describes the extraction and assembly methodology employed to assemble the analysis datasets for this study. 15 Table 1: Data Assembly Methodology Data Source Assembly Methodology Enterprise Data Warehouse (EDW) • Tables from the EDW were extracted in csv format using SQL and Crystal Reports, and assembled based on common claim numbers • Adjustments to claim records before merging tables were made to synchronize field formats (e.g. date vs. text) and field lengths (leading zeros) Compensation and Rehabilitation Reporting System (CARRS) • Claims workflow data for all SDLs was extracted by WCB personnel and provided as Access databases Payments Transaction Database • Short term disability (STD) and health care only (HCO) payments and associated workdays for all claimants were extracted by WCB personnel in csv format Appellant Database • Review board statistics for claimants appealing denied claims were provided by Compensation Systems in Access database format Human Resources PeopleSoft Database • Years of experience and days off for training, sick leave, and vacation for Case Management Resources were extracted from the PeopleSoft database maintained by Human Resources • Case Manager identities (names and employee id numbers) were scrambled to ensure anonymity and confidentiality • WCB representatives associated desk numbers to these scrambled case management identifiers Compensation Finance • HR Salary Costs for 1999 and 2000 were provided by Compensation Finance by area office • Annual Travel Expenses, by area office, were provided to be used as an approximation for geographic dispersion o These expense records were also used to validate approximate travel times based on employer location postal codes associated with employer tables in the Data Warehouse Data was summarized at both the claimant level and by case management desk, by merging records with common claim numbers and case management desk identifiers. Subsequent meetings were held with the data providers and CSMs to address issues regarding interpretability, sparseness, representativeness, and redundancy of reporting. These issues stem mainly from the fact that existing metrics are very broad, and not designed to specifically measure and compare case management operations. Once these issues were resolved the summarized datasets were used to analyze existing metrics, determine relationships between outcome measurements, and develop relationships between outcomes and influencing factors. 16 3.4 Identifying Key Performance Measures When comparing service delivery locations and case management desks, performance measures should be intuitive, tractable, and easy to communicate. Several indicators exist with the potential to be used as comparative performance measures. If a few performance outcomes can explain most of the variability in all related measurements, then these select outcomes should be used for the purpose of comparing Case Management Desks. Each observation is a case management desk's performance. If we use the SDL as the unit of analysis we would not have enough data points (only 17) to draw significant statistical conclusions. This project phase employed Principal Component Analysis, in conjunction with management feedback, to investigate the relationships between existing performance indicators and select a subset of key performance measures to be used as a basis for comparing case management outcomes. Principal Component Analysis is a multivariate technique that is concerned with explaixiing the variance-covariance structure of a set of variables through a few linear combinations of these variables (Johnson and Wichern, 1998, p. 458). These uncorrelated linear combinations are known as prmccpdarmponents. We analyze the first few principal components that explain the majority of the variability, and then select a summary variable from the principal component that can intuitively and statistically explains the behaviour of similar variables. The analysis is conducted in SAS using the FACTOR procedure (Hatcher and Stepanski, 1999). The results are reported in Section 4.2. 17 3.5 Benchmarking Key Performance Measures We seek to quantify the direction and strength that the potential factors (generated in the CSM working sessions) have on t^he key performance measures. We use regression analysis to estimate the relationship between performance indicators and the influencing factors. The goal is to discover factors that differentiate case management desks (in terms of claimant profiles), which also influence the summary performance outcomes. Some of the explanatory factors to be considered, such as industry sector, injury type, and body part are categorical variables. This means that at the claim-level dataset, the values describing these factors are non-numeric, in descriptive categories. For example, instead of numerical value quantifying the severity of an injury, the entry for a claimant would be a code representing the type of injury (strain, fracture, etc.). Table 2: Sample of Claim Records Desk Injury Type Industry Sector Claimant 1 Desk 1 Bursitis Service Claimant 2 Desk 1 Back Strain Manufacturing Claimant 3 Desk 1 Back Strain Primary Industry Claimant 4 Desk 2 Laceration Retail Claimant 5 Desk 2 Fracture Other Table 2 illustrates a simplified example of the categorical variables used to describe claimant and desk activity. The actual data tables have several categories for each claimant profile descriptor, for thousands of claimants across hundreds of case management desks. In order to convert these categorical variables into data that can be incorporated into the models, proportions of each category were summarized for each desk. Table 3 shows how the example data in Table 2 is now summarized at a desk level. 1 Table 3: Sample of Categorical Variables Summarized by Desk Deskl Desk 2 Proportion of Bursitis Claims 0.333 o.ooo;; Injury Category Proportion of Back Strain Claims 0.667 0.000 Proportion of Laceration Claims 0.000 0.500 Proportion of Fracture Claims 0.000 0.500 Proportion of Service Sector Claims 0.333 0.000 Proportion of Manufacturing Claims ' 0.333 0.000 Industry Category Proportion of Primary Industry Claims 0.333 0.000 Proportion of Retail Claims 0.000 0.500 Proportion of Other Industry Claims 0.000 0.500 The major difficulty with this approach is that instead of having one variable to describe an influencing categorical factor, we now have several. With multiple categorical variables describing injury, industry, and body part, these variables may confound each other and make the results of the regression analysis difficult to interpret. CSM representatives could not intuitively reconcile why one variable in a categorical factor would be significant, and the others in that category would not be. For example, if the model states that the most predictive variables are the proportion of back strains, and the proportion of retail sector claims, the CSM representatives would question why the other injury and industry variables were not deemed significant. The goal for this section of the analysis is to reduce the categorical factors into single numeric variables that summarize each nominal factor to facilitate interpretation while still describing activity at the desk. 19 Cluster analysis is used to group categorical variables summarized at the desk level. Cluster analysis searches the data structure, which in this case is the proportions of categorical variables for each case management desk, to find natural groupings based on sirnilarities (or oUssirnilarities) between case management desks. The CLUSTER procedure in SAS is used to implement the analysis, based on hierarchical clustering methods. These hierarchical clustering methods are based on correlations between pairs of variables (Johnson and Wichern, 1998). We develop two clusters for each categorical factor, so that one numerical summary value can be used to describe the activity at a desk. For example, if a desk has a 0.87 proportion of Cluster 1, its proportion of Cluster 2 is 0.13 as there are only two high level clusters used for each factor. The final clusters for each categorical variable are discussed in the Results section of this thesis. Details on the clustering algorithms utilized are in Appendix II. Stepwise selection in the multiple regression was used to determine the factors that best explained the behaviour of the performance outcomes. Stepwise selection chooses the influencing factors to be incorporated into the regression model by adding factors based on how much explanatory power each additional factor contributes. A n additional factor will not be incorporated into the regression model unless it adds significant explanatory power for the performance outcome, when used with the other factors. A l l combinations of variables not found to be significant by the stepwise procedure are incorporated into the regression model and compared to the stepwise output to ensure that interactions between seemingly insignificant variables are not substantial. 20 3.6 Overall Efficiency Analysis 3.6.1 Data Envelopment Analysis - Introduction Data envelopment analysis (DEA) is a linear programming-based technique for measuring the relative performance of Decision Making Units (DMUs) where the presence of multiple inputs and outputs makes comparisons difficult. The usual measure of efficiency, described as a ratio of output to input, is not always adequate since the multiple inputs and outputs are related to different resources, activities, and external factors. D E A is a weighted ratio of outputs to inputs, with the weights for each D M U optimized to maximize that particular DMU's efficiency score. Case Management Desks are the 109 DMUs in this study, with the output and input factors determined from the preceding multivariate analyses. D E A is appropriate because it is a nonparametric method of analysis, which does not impose unknown structures on the data. Since no research or generally accepted conjectures have been established with regards to the relationships between input and output factors in processing claims, we can use the mathematical optimization to suggest combinations that define efficiency and evaluate efficient units from an aggregate perspective of performance outcomes. D E A supplements the regression analysis by examining the interaction between input factors and multiple performance outcomes. D E A explicitly reveals examples of efficient and inefficient units, and generally provides more aggressive performance targets than regression-based approaches. Al l models are implemented in Microsoft Excel, using Frontline Solver for the optimization process. Visual Basic for Applications (VBA) code can be used to control Solver and iterate 21 optimizations to solve the D E A formulations and report results for all DMUs automatically. Albright (2000) has included V B A code for a series of decision support tools built using Excel. We have expanded upon this code to accommodate the scope of the problem and alternate D E A models, as well as to implement the dual as well as primal formulations to verify the robustness of the Solver. Proudlove (2000) also advocates using the developing power of spreadsheets to implement D E A , to make the technique more widely available and allow for greater flexibility in model exploration and specification. Figure 7: Sample DEA Optimization Input using Microsoft Excel > Set Target Cell: lambda^values^Bl^ 17: $1$ 117, $K$ 117, t h5ypject to the Constraints: Equal To: C Max <•< Min C Value of: |<L -By Changing Cells; — — — Guess $B$130:$I$130 = InputConstraints2 $K$130 = OutputConstraints2 OutputConstraint = 1 Add Change Delete Solve Close Options Reset All Help Figure 7 is a screenshot of the optimization input screens in one of the D E A formulations. Each D M U runs this optimization routine, with references for the constraint and objective cell ranges changing with each iteration. The input screen is hidden once data is entered and the D E A formulation type has been selected, to minimize processing time. Details on the D E A implementation can be found in Appendix III. The subsequent sections describe formulations for the standard D E A models. These models deliver different results based on constant or variable returns to scale, whether the 22 production objective is output-oriented (maximizing outputs given fixed inputs) or input-oriented (minimizing inputs given fixed outputs), and the allowable values that input and output weights can take. 3.6.2 The CCR Model The CCR model is also known as the Constant Returns to Scale Model. It was developed by Charnes, Cooper, and Rhodes (1978) as the first D E A model. The CCR method can be formulated in two ways: using an output-oriented approach or an input-oriented approach. A n output-oriented approach seeks to maximize outputs with a fixed level of inputs. A n input-oriented approach seeks to minimize inputs while mamtaining the same level of outputs. Figure 8: CCR Example: 2 Outputs, 1 Input Efficiency Frontier / CM 3 Q. F D* B* y y y A* • C _ , - ^ - f £ ' - — > E* y y s ' y y / *^.— 3 4 Output 1 23 Figure 8 shows a simplified example of Data Envelopment Analysis with an output-oriented approach and constant returns to scale. Units A-F are producing 2 Outputs with 1 Input. Units C and F are producing the most outputs for a given input out of all the DMUs being analyzed. We now construct an Efficiency Frontier, which envelops all other DMUs that are not at this level of peak efficiency. Units C and F receive efficiency scores of 100%, as they are outperforming the DMUs with respect to the 2 Outputs measured. Units A, B, D, and E are assigned target levels of performance, based on the radial distance from the origin to the Efficiency Frontier constructed based on the performance of Units C and F. These are denoted by A*, B*, D*, and E*. The efficiency scores of the inefficient units are computed by taking the ratio of the actual radial distance from the origin to the projected radial distance from the origin. In other words, if an inefficient unit is 20% away from reaching the efficient frontier, assuming constant returns to scale in radial measures, then that unit is 80% efficient. Input-oriented DEA approaches can be formulated in a similar manner: Minimize Inputs used, given fixed Outputs, with the efficiency frontier being created by those DMUs who consume the least amounts of inputs. The problem becomes difficult to formulate and solve graphically as the dimensions exceed 2 Outputs and 1 Input, or 2 Inputs and 1 Output. The fractional programming formulation (Cooper, Seiford, and Tone, 2000, p. 23) of this problem is as follows: 24 Maximize: 0 = u l y l o + u l y 2 o + . . . + usyso ^lo+vlx2o+... + vmymo Subject to: ^yXj+uxy2o+... + usysj vxxXJ+vxx2o+... + vmymJ v 1 , v 2 , . . . , v m > 0 u],u2,...,us >0 <1 0" = !,...,«) The notation for this model specifies n DMUs, m input items, and s output items. The input and output data for DMU/ are ( . , x2.,..., xm .) and (v,.y2.,..., ys}r) respectively. The input and output weights are denoted by V, where v goes from 1 to m inputs, and ur, where r goes from 1 to 5 outputs. The X ( 0 and yr0 notation describes the inputs and outputs associated with the DMU being optimized (DMUj = DMUo). Essentially this model states that for each DMU/, we are maximizing the weighted outputs over the weighted inputs by changing the input and output weights. The only constraints are that the input and output weights be nonnegative, and that no other DMU can be more than 100% efficient using the optimal input and output weights with that DMU's data. Another way of phrasing this is that every unit has optimal business rules computed that maximize its efficiency, such that no other unit can be more than 100% efficient with those rules in a different business environment. 25 Since this model is fractional, linear programming solution methods cannot be applied. A n equivalent linear program is developed by setting the denominator of the objective function equal to a constant value, and ensuring that weighted outputs are less than weighted inputs, for all DMUs (equivalent to the previous constraint of holding maximum efficiency at 100%). Maximize: 0 = uiylo+uly2o+... + usyso Subject to: v i x \ o + v i x 2 o + ••• + vmymo - ^ (1 * s arbitrary - this can be set at any constant, so long as it is the same for each DMU's optimization) u \ y \ j + u \ y i o +-+usysj + v i x 2 o + -+vmymj u = V l . V 2 » - , V r a > 0 ux,u2,...,us > 0 The efficiency score, 9, is deterrnined by the optimal weights computed in the D E A formulations described above. The projected performance targets for inefficient DMUs are found by first identifying the input excesses and output shortfalls associated with the D E A solution (Cooper et al., 2000, pp. 44-47). For an input-oriented formulation, these slack vectors can be defined as: Input Excesses: s" = Qx0 - XX Output Shortfalls: s+ = YA. - y 0 26 That is, the slack for an inefficient DMU's input variable is defined as the product of the efficiency score and the value for that input variable, minus the product of the input variables and a nonnegative vector X. The slack for an inefficient DMU's output variable is defined as the product of the vector of output variables and the nonnegative vector X, minus the actual value for that output variable. These projection formulations are based on the constraints of the dual formulation of the linear program, which require that the activity (0xO) y0) belong to the production possibility set, and that assumptions relating to the production possibility set dictate that (XX, YX) outperforms (6x0y0) for inefficient DMUs. The dual formulation of this model is provided in Appendix III. We use these slack vectors identified to compute target values for input and output factors in the D E A model. Before we can build projections based on the slack variables, we must first identify possible input excesses and output shortfalls by optimizing the total slack in the formulation, based on the optimal value of 9. That is, after solving for 0, u, and v, (the optimal efficiency score, the input weights, and the output weights) for all DMUs, we must maximize the total slack associated with that set of efficiency scores to obtain the total input excesses and output shortfalls for the inefficient units. The formulation is as follows: 27 Maximize: co = es~ +es + Subject to: S = Qx0 - XX S+ = YX-y0 X>0, S >0, S+>0 e is a vector of ones, so that the objective function is simply the total input excesses and output shortfalls. This optimization states that the maximum slacks can be computed by maximizing total slack subject to nonnegative slack vectors, and ensuring that the formulae derived for the slack vectors based on optimal efficiency scores and the dual CCR formulation still holds. Efficiency can be attained if the input values are reduced radially by the optimal efficiency score and the input excesses are eliminated, or if the output values are increased by the output shortfalls. That is, a D M U becomes efficient when the efficiency score = 1, and there are no slacks in production. This leads to the CCR projection, which identifies target input and output values to achieve efficiency: Target Input Value = Radially-Reduced Input Value minus Input Excess Target Output Value = Output Value plus Output Shortfall For the Output-Oriented Approach, we can develop projections based on increasing outputs, so that these formulae become: 28 Target Output Value = Radially-Augmented Output Value plus Output Shortfall. Target Input Value = Input Value minus Input Excess The Results section contains target values for the Case Management Desks, in addition to a list of the top desks mentioned in the Reference Sets of the inefficient DMUs. A Reference Set consists of all units that would perform better than the D M U being optimized, if the DMU's optimal weights were being applied to the input and output data of the units in the Reference Set. The top candidates for chssemination of best practices should come from the units in an inefficient DMU's Reference Set. Additional formulations can be found in Appendix III. 3.6.3 The BCC Model The BCC model is also known as the Variable Returns to Scale Model. It was developed by Banker, Charnes, and Cooper, (1984) to address Efficiency Frontiers where the Production Possibility Set exhibits variable returns to scale characteristics. Inefficiencies are measured by comparing observed performance with composite reference units, constructed as convex combinations of other observations. For an input-oriented approach, the model maximizes, for each separate D M U , the distance of the output of the evaluated unit from the composite level of a reference unit that consume less or equal input amounts (the converse measurement takes place for output-oriented approaches). The BCC efficiency frontier represents all convex combinations of observed input-output combinations that are not dominated by other convex combinations. The BCC model differs 29 from the CCR model only in that the sum of the X vectors is equal to 1 (Cooper et al., 2000, p. 88) to ensure these convex combinations. The convexity constraint in the model formulation ensures that the composite unit is of similar scale size as the unit being measured. The efficiency score obtained from this model gives a score that is at least equal to the score obtained using the CCR model. Figure 9 shows a simplified comparison of production frontiers for the CCR model and the BCC model (which is built on the convex hull of the observations), with one output and one input. Figure 9: Comparison of Efficiency Frontiers 0 1 2 3 4 5 6 Input 3.6.4 Slacks-Based Efficiency Measurement A slacks-based measure of efficiency (SBM) combines input-oriented and output-oriented approaches in a single model. This model also includes a dimension-free or units invariant measure, which gauges efficiency evaluation invariant to the units of measure used for the 30 different of outcome, and is monotone decreasing in each input and output slack (Cooper et al., 2000, pp. 96-97). A fractional program is formulated by solving for the minimum ratio of average input and output mix inefficiencies (excesses and shortfalls). The following formulation assumes that the input and output data is non-negative. In application, zero input values are deleted from the objective function. A D M U is efficient only if slacks are zero in all inputs and outputs. Minimize: p- m Subject to: Xg — XA. + S + y0 = YX-s" X>0, S">0, S+>0 For inefficient units, efficiency can be gained by simply deleting the slacks. Subtracting input excesses from each input variable and augmenting output shortfalls to each output variable will bring the D M U to 100% efficiency. Solving for the slacks directly eliminates the need to execute a two-stage D E A (as with the CCR and BCC formulations) in order to compute target projections. 31 SBM results and model comparisons are described in the Results section. Detailed calculations for transforming the fractional SBM into a linear model are described in Appendix III. 4. RESULTS 4.1 Client Service Manager Feedback The CSM bramstorrning sessions proved to be informative and provided a starting point for our data analysis and methodology development. Recommendations were made to senior management to hold regular opportunities for feedback and development of ongoing performance initiatives that encourage Case Manager and Client Service Manager participation. Performance indicators with respect to timelines, duration, and volumes were discussed. Timeliness is defined as the time claimants spend in the WCB system, duration is defined as the days that claimants are compensated for wage loss, and volumes refer to the number of claims directed to or processed by case managers. A few CSM groups were using performance indicators in place that were specific to their SDL, for example outstanding action items per desk. Other CSMs responded to our survey of important performance measures by claiming that case management operations perform to what they are measured on, implying that deciding upon important performance indicators have traditionally come from senior management, rather than being determined by peers. We classified the outcomes and indicators discussed into Low, Medium, and High levels of importance, as shown in Table 4. 32 Table 4: Performance Indicators and CSM Feedback Priority Performance Indicator Description Comments Low Client Surveys Score (on a scale of 1-10) from Overall Customer Satisfaction surveys administered by Ipsos-Reid, through random phone solicitation of WCB claimants CSMs felt that this survey was not designed to measure the quality of case management practices. Successful Appeals The proportion of denied claims brought to the Workers' Compensation Review Board (WCRB) that is appealed. If a claim is successfully appealed, this may indicate a poor adjudication decision by the case manager originally assigned to the claim. Volume of Complaints The total number of complaints brought to the SDL or Compensation Services. Not adequately captured in existing systems. Income Continuity The proportion of entitled claimants paid within 17 days of injury. Reflective of timeliness of adjudication decision. Early Intervention Early intervention strategies for safe and timely rehabilitation. Reflected in other timeliness and satisfaction indicators. Medium Claims Costs Total short-term duration and health care costs. Health care costs may be out of case management's control. Appeals The number of claims brought to the WCRB. Possibly indicative of poor adjudication decisions. Return-to-Work (RTW) The proportion of claimants who successfully return to work. High RTW implies strong case mgr. performance. Closure Rate The ratio of the volume of claims closed to the volume of new claims. Volume of work assigned may be dependent on current performance. Reopened Claims The proportion of closed claims that are reactivated with an associated payment. An acceptable time frame between initial closure and reactivation must be determined. Proportion of Claims on Wage Loss The proportion of active claims that are receiving wage loss benefits. More reflective on claimant profile than case mgr. performance. Administrative Costs The administrative costs associated with case management. Salary info, available through Comp. Finance. 28-Day List The proportion of claims entitled within 28 days of injury. Also reflects performance of Entitlement Units. High 85-Day List The proportion of claims with wage loss less than 85 days. Counts the total number of wage loss days paid, including reopenings. Volume Opened (Claims In) The number of claims assigned to an SDL. Volume of work assigned may be dependent on current performance. Aged Claims Inventory A weighted measure of claim volumes and duration. Maintained by Compensation Systems. Volume Closed (Claims Closed) The number of claims closed by an SDL. Indicates efficiency by tracking completed cases. Disallow Rate The proportion of pending claims disallowed for wage loss. Reflective of work styles or claimant profiles. 33 The highly ranked factors were similar in nature to a study conducted by Nexus Actuarial Consultants (1996) of different workers' compensation board jurisdictions. The Nexus study revealed that timeliness of payment/income continuity, duration of short term duration (STD) claims, percentage of STD claims reopened, costs of benefits, return-to-work rates, and recovery rates were deemed as important benchmarks for compensation systems performance. In addition to these outcomes, the CSMs identified how measurements relating to these outcomes are currently being recorded. We discovered that existing measurements could be improved through improving data accuracy, completeness, terminologies, and accountability. An example of accountability in reporting performance indicators is shown in Figure 10, which demonstrates the difference in the magnitude of wage loss duration when comparing aggregate wage loss duration distributions to the wage loss duration distribution incurred by Case Management, for claims closed in 2000. We see that the distribution of total wage loss duration is significantly greater than the wage loss days that case management should be accountable for. That is, the quartiles and outliers for total wage loss duration 34 Figure 10: Wage Loss Duration Incurred by Case Management W a g e Loss Durat ion: Total W a g e Loss Days and C a s e M a n a g e m e n t ( C M ) W a g e Loss 1 6 % —i 0 20 40 60 80 1 00 120 140 1 60 180 2 0 0 220 240 D a y s Case Management should only be measured on the wage loss duration it is accountable for, especially when comparing performances between desks. A summary of measurement-related recommendations can be found in the Conclusions chapter of this thesis. We also solicited a list of factors that may influence case management performance outcomes. These factors are tested against performance outcomes for significance through regression analysis, and results are shared with CSM representatives to ensure interpretability. The initial list of performance-influencing factors includes geography, the proportion of ASTD claims, claimant wage rates, medical resources, employer profiles, claimant profiles, injury types, case management experience, body part, gender, and caseload. We are limited to investigating factors where data is available and consistently recorded and interpreted, so this project could not address all concerns and potential initiatives identified by the CSM groups. The results of our analysis of influencing factors on outcomes are described in Section 4.3 of this thesis. 35 4.2 Principal Component Analysis Results When comparing service delivery locations and case management desks, performance measures should be intuitive, tractable, and easy to communicate. Several indicators exist with the potential to be used as comparative performance measures (see Table 4). If a few performance outcomes can explain most of the variability in all related measurements, then these select outcomes should be used for the purpose of comparing case management desks. We note from Table 4 that the indicators deemed to be important by the client service managers surveyed could be classified into three main areas: volumes, timeliness, and quality. For example, indicators that measure duration and other mdemnity figures can be grouped into timeliness, and the number of cases opened (claims in), the number closed (claims closed), or the closure rate describe volumes. We seek to find variables within the principal components that have the highest correlations with similar metrics in these categories. The purpose of the principal components analysis is not to use the principal components directly, but to identify the most influential variables in the eigenvectors that explain the majority of the variability. 36 Figure 11 displays the magnitude of the variables within the two most explanatory principal components, for case management desks with claims closed in the year 2000. These correspond to Eigenvectors 1 and 2 in Figure 12. The values shown on the vertical axis of Figure 11 have been multiplied by 100 and rounded to the nearest integer (as per SAS conventions of displaying the high-coefficient variables within each eigenvector). The magnitude of each variable may differ depending on which variables and time frames are used, but the same variables are consistently influential in each principal component group: the number of claims closed explains the behaviour of variables related to volumes, and the average number of days in case management explains the variation in other measures related to timeliness. Figure 12 is the scree plot of eigenvectors, which graphically represents the proportion of variation (on the vertical axis) accounted for by each principal component. The two principal components used comprise nearly 70% (0.6856) of the cumulative variations among the volumes and timeliness metrics variables. Using additional principal components would account for more of the variability amongst all indicators, however it would be difficult to interpret a summary variable for these additional eigenvectors. Variables representing quality are difficult to summarize and isolate to case management operations. Candidate indicators for the 'quality' outcome are overall customer satisfaction survey scores, the proportion of claims sent to the Appellant Board for review, the proportion of claims upheld by the Appellant Board, the proportion of successful return-to-work cases, and the proportion of claims reopened. 38 CSM feedback indicated that customer satisfaction surveys might not necessarily reflect good or bad quality in case manager performance, as the questions are vague and responses may be biased based on unfavourable (but perhaps appropriate) decisions regarding wage loss and return-to-work schedules. Statistics from the Appellant Board were not used due to the time frame: since we were interested in claims closed in 2000, sufficient time will not have elapsed between a claim's wage loss denial or possible premature end of wage loss benefits and its arrival at the WCB review board. For example, if a claim had been denied for wage loss in November 2000, and it took over a year for the claim to arrive at the review board, we would not have captured it during our data extraction phase. Return-to-work data was also difficult to capture and summarize by case management desk. We therefore decided to use the proportion of claims reopened as a proxy variable for quality. The allowable timeframe between a claim's initial closure and reopening was discussed with CSM representatives and senior management. It was determined that claims reopened within six months of initial closure would be considered a potentially avoidable reopening, where a case manager may have mitigated that reopening through perhaps extended or suspending wage loss instead of closing the claims, or executing some other proactive claims management strategy. If a claim was reopened after six months of initial closure, we can consider that to be a valid, uncontrollable reaggravation of the injury in most cases. 39 Figure 13: Distribution of Time Between Initial Closure and Reopening Time Between Closure and Reactivation 12% 10% o rz 13 cr 8% 6% -| 4% 2% -I 0% ppppphnrnninirfinrri-rr-rrir,--. 0 20 40 60 80 100 120 140 160 180 200 220 240+ Days Figure 13 shows the distribution of days between a claim's initial closure and case reactivations with associated wage loss payments, by case management desk. We note that the six-month cutoff for defining potentially avoidable reopenings is rather subjective, as there is no detectable spike in the distribution at 180 days, but rather a gradual decrease. Table 5: Case Management Outcomes and Interrelated Measurement Variables Case Management Outcomes Interrelated Measurement Variables Timeliness Average Time in Case Management Average Time in System Average Time from Injury to Payment Average Time from Registration to Payment Average Wage Loss Duration Wage Loss Duration - 95th percentile Average Time to Decision Volumes Claims Closed Claims In Claims Active Closure Rate Quality Reopenings Appellant Board - Proportion Appealed, Proportion Upheld Return-to-work metrics 40 Table 5 lists the interrelated measurement variables for each main performance outcome, and the variables we have chosen to represent these outcomes (in bold), based on how well they predict the behaviour of similar metrics in the principal components analysis, as well as subjective considerations incorporating feedback and intuitive validation by the Client Service Managers. 4.3 Multivariate Analysis Results 4.3.1 Cluster Analysis Feedback from CSM representatives indicates that the cluster analysis models grouped injuries that were common among desks (Injury Cluster 1) versus injuries that were less common (Injury Cluster 2). The cluster analysis results for industry sector and body part follow the same intuition. We refer to Injury Cluster 1 as the proportion of "Strains and-itis Claims", and use this numeric value to simplify the categorical Injury Type factor. Table 6: Injury Cluster Groups Injury Cluster 1 Injury Cluster 2 Other Strains Fractures Stress Amputations Hernias Lacerations Back Strains Contusions Tendonitis Dislocations Carpal Tunnel Other Bursitis 41 Figure 14: Dendogram of Clustered Injury Variables 1 . 2 5 A 1 . 00 v e r a 0 . 7 5 1 9 0 0 . 5 0 0 . 2 5 0 . 0 0 P r o P P r o P P r o P P r o P s h t e r r e n n n e I f r a c t LI r e s Name o f O b s e r v a t i o n o r C l u s t e r Figure 14 displays the average linkage hierarchical procedure, displayed i n a dendogram, a tree-diagram i n w h i c h the c o m b i n e d vertical distances represent the similarity between injury types. T h e shorter the vertical lines the m o r e similar the injury types are, and the longer the lines are the more different the injury types are. F o r example, a case management desk w i t h a high p r o p o r t i o n o f tendonits claims is l ikely to have a high p r o p o r t i o n o f carpal tunnel syndrome claims. T h i s makes intuitive sense, as b o t h o f these injuries can be classified as Activity-Related Soft Tissue Disease ( A S T D ) . Similar intuitive conjectures can be proposed for each o f the sub-clusters and main clusters. 42 The main clusters for the Industry Sector and Body Part categorical factors are reported below. Details on the clustering procedure and output can be found in Appendix II. Table 7: Industry Cluster Groups Industry Cluster 1 Industry Cluster 2 Primary Transportation Manufacturing Construction Trades Public Sector Service Other We refer to Industry Cluster 1 as the proportion of "Primary and Manufacturing Claims", and use this numeric value to simplify the categorical Industry Type factor. Table 8: Body Part Cluster Groups Body Part Cluster 1 Body Part Cluster 2 Knee Back Lower Extremity Hand Feet Upper Extremity Hips Elbow Neck Shoulders Trunk Systems Other We refer to Body Part Cluster 1 as proportion of "Lower-Body Claims", and use this numeric value to simplify the categorical Body Part factor. The numeric summaries derived from the cluster analysis (defined as the proportion of claims within a certain main cluster) are used as inputs for the regression models to determine significant factors that differentiate case management desks and influence performance outcomes. The descriptive capability of the regression using the numeric summaries versus the binary nominal variables is slightly lessened (the R-square decreases by 43 4% for the most predictive model), however using the summary variables makes the comparison of categorical factors between desks more intuitive for the CSM reps. 4.3.2 Multiple Regression This section contains results of the analysis between factors and key outcomes (as derived in Section 4.2). Figures 15 - 17 show the distribution of each key outcome by case management desk. We propose that the factors revealed in the multiple regression analysis comprise a portion of the variability in these distributions, and the variability unaccounted for by the regression formulae may be attributable to other factors that we were not able to capture and quantify, and differences in performance across desks. Figure 15: Distribution of Claims Closed 120 140 160 180 200 220 240 260 280 300 320 340 360 Claims Closed This performance outcome is somewhat bimodal - that is, there appears to two centres to the distribution of Claims Closed across desks. The right side of this frequency distribution is primarily associated with more urbanized areas. 44 Figure 16: Distribution of Average Time in Case Management 50 60 70 80 90 100 110 120 130 140 150 160 Time in Case Management (Calendar Days) Figure 17: Distribution of Reopenings 25% Proportion Reopened Within 6 Months 45 We first test quantifiable factors suggested by the CSMs against the outcomes considered using simple linear regression. Tables 9-11 display these results, which were verified through anecdotal descriptions offered by CSM representatives. Potential factors listed here not previous described are the Proportion of " Y " Claims, which is the coding used for ASTD cases; the Time Before Case Management (CM) for both Complex (C) and ASTD (Y) Qaims; 'Valid' Reopenings, which we define as the number of claims that were reopened after six months since initial closure (implying an injury reaggravation that may be difficult to manage); geography, quantified by annual travel expenses; "Non-Closure" activities, which we define as the proportion of decisions made by case management that did not have direct impact on closing claims (denying, suspending, or opening for investigation only), and Wage Rate, which is which is summarized by the median employee wage rate by case management desk. Table 9: Simple Regression of Factors on Claims Closed Claims Closed =/'(...) R-Square Parameter Estimate Proportion Y Claims 0.23 386.2 Time Before CM - C Claims 0.21 253.5 Body Part Type 0.16 -424.1 Caseload 0.16 1.7 Injury Type 0.14 270.3 Time Before CM - Y Claims 0.14 114.9 Geography 0.09 -0.7 Valid Reopenings 0.06 -1181.9 Annual Hours Worked 0.04 -0.1 SDL Experience 0.03 -2.0 Gender 0.03 -82.6 Industry Sector Not Significant -45.9 WCB experience Not Significant -0.4 Age Not Significant 21.9 "Non-Closure" Activites Not Significant -48.7 Wage Rate Not Significant 0.1 46 Table 10: Simple Regression of Factors on Avg. Time in Case Mgmt. Avg. Time in Case Mgmt = /(...) R-Square .' Parameter. Estimate . Time Before CM - C Claims 0.16 -78.2 Proportion Y Claims 0.12 -95.8 Time Before CM - Y Claims 0.12 -37.0 Injury Type 0.10 -77.6 "Non-Closure" Activites 0.06 44.3 Body Part Type 0.05 86.6 Geography 0.05 0.2 Wage Rate 0.03 -0.3 Industry Sector Not Significant -2.6 WCB experience Not Significant 0.1 SDL Experience Not Significant -0.1 Annual Hours Worked Not Significant 0.0 Gender Not Significant 2.9 Age Not Significant -7.9 Caseload Not Significant -0.1 Valid Reopenings Not Significant 196.0 Table 11: Simple Regression of Factors on Proportion Reopened Proportion Reopened = / ( . I .) R-Square . Parameter^Estimate Industry Sector 0.14 0.0502 Geography 0.09 0.0003 Wage Rate 0.05 0.0004 Caseload 0.04 0.0003 Valid Reopenings 0.04 0.3457 Body Part Type 0.04 0.0743 Gender 0.02 0.0278 Proportion Y Claims Not Significant 0.0062 Injury Type Not Significant 0.0085 WCB experience Not Significant 0.0000 SDL Experience Not Significant 0.0005 Annual Hours Worked Not Significant 0.0000 Age Not Significant 0.0352 "Non-Closure" Activites Not Significant -0.0157 Time Before CM - Y Claims Not Significant -0.0087 Time Before CM - C Claims Not Significant 0.0084 47 Before these factors were entered into stepwise multiple regression models, we first omitted Time Before Case Management as a potential predictive factor, due to the fact that this variable simply flags difficult claims, which is what we are in essence trying to predict for the outcome Average Time in Case Management - essentially we would be using an outcome to predict an outcome, which may not be accurate. Potential predictive factors were not normalized before being used as inputs to the regression model, as sufficient relationships were found without this transformation, transformed variables may limit straightforward explanations of the regression model to CSM representatives and senior management. Complete SAS output of the stepwise regression procedure can be found in Appendix II. The stepwise regression procedure for claims closed yielded the following model: Claims = 176.3 +397 x Proportion^ Closed Model R-Square = 50.3% V ASTD + 0.47 x Average Caseload 0.0012x ( Annual \ Travel Expenses Table 12: Factors Influencing Claims Closed Factor Parameter Data Range Proportion of ASTD Claims 397.3 0% - 38% Average Caseload 0.47 33-132 Annual Travel Expenses -0.0012 $3195-$87,185 The correlation coefficient, or the amount of variability explained by these factors, is 50.3%. This means that the remainder of the variability is attributable to case management performance, in addition to factors that are not captured in available data systems. The 48 parameter value reflects the direction and strength of the dependent variable, relative to its scale. For example, the high parameter value for Proportion of ASTD Claims does not mean that its influence is several times stronger than caseload, but that the scale of the ASTD variable is several times smaller than the caseload variable. The stepwise regression for Average Time in Case Management can be explained by: Time in 49.7-l lOx Proportion V CaseMgmt Model R-Square = 37.7% ASTD + 0.45 x J ( Average \ (Non - Closure} + 49.4x Caseload, y Activities j Table 13: Factors Influencing Time in Case Management Factor Parameter Data Range Proportion of ASTD Claims -110 0% - 38% Average Caseload 0.45 33 - 132 "Non-Closure" Activities 49.4 18%-69% The amount of variability explained by these factors is 37.7%. The remainder of the variability in this measurement across desks can be attributable to case management performance, in addition to factors that are not captured in available data systems. The following regression model describes the factors that best explain the distribution of reopenings: Proportion = -0.03+ 0.072 x Reopened Model R-Square = 20.0% f Proportion ^ Strains & "itis' Injury Type + 0.053 x Proportion ^ Prim & Mfg Industry Sector + 1.9E-7x Annual Travel Expenses j 49 Table 14: Factors Influencing Reopenings Factor Parameter Data Range Injury Type 0.072 8%- 100% Industry Sector 0.053 8% - 44% Annual Travel Expenses 1.9 x 10"' $3195-$87,185 The amount of variability explained by these factors, is 20.0%. Since this model is not as significant as the regressions for the other two outcomes, this may be indicative of the strong variation in case manager performance and work styles (not captured in the regression models) in terms of the outcome proportion reopened within six months of initial closure. The predicted values from the regression models can be used as performance targets for each SDL. Figure 18 shows Actual and Predicted performance for the outcome Claims Closed, for the three regions in the province. Actual and Predicted performance are displayed as box plots, with a diamond representing the median, the box surrounding the diamond representing the middle 50% of the distribution, and the lines extending from the box representing the tails of the distribution. Figure 18: Regression-Based Performance Targets for Claims Closed 400 350 "S 300 w g 250 w 200 | 150 O 100 50 0 I fllllflll o I I <j> I p^i, 1 ' \s i <> ^ Actual Predicted Actual Predicted Actual Predicted 50 Regional and SDL identifiers have been removed for confidentiality purposes. Charts similar to this figure were presented to senior management and CSM representatives for all three key performance outcomes, by SDL as well as by region. Figure 19: Regression-Based Performance Targets for Avg. Time in Case Mgmt. 160 E 0) 140 "GO >» 120 C O c 100 CD 80 E 60 i -d> 40 > < 20 0 I xH I • •?•: O A E o l ° Y i J I Actual Predicted Actual Predicted Actual Predicted Figure 20: Regression-Based Performance Targets for Proportion Reopened " O CD c CD Q . O CD LX. c o '•c o Q . O 10.0% 9.0% 8.0% 7.0% 6.0% 5.0% 4.0% 3.0% 2.0% 1.0% 0.0% i o i o I 1 Q • i I ~ -9s-o t - t - 1 I Actual Predicted Actual Predicted Actual Predicted 51 T h e regression models produced predicted values that were very close to actual performance, identifying regional differences i n case management operations. T h e multivariate analysis has successfully determined the factors that differentiate case management desks across S D L s and regions, and quantified the impact that these factors have o n key case management outcomes. T h e benefit o f applying performance targets based o n the regression analysis is that this relatively standard approach brings about measurement targets that are easy to understand and are effective for emphasizing improvement o n a single summary outcome. Limitat ions to this approach are that the m e t h o d does not consider the net performance of the desk, and that targets are already met for strong performers. 4.4 Data Envelopment Analysis Results W e formulate the data envelopment analysis models based o n the key outcomes and influencing factors detenriined i n the multivariate analysis. In D E A formulations, O u t p u t Factors are outcomes w e wish to maximize, and Input Factors are elements we w i s h to minimize . I n order to transform the outputs into targets we wash to maximize, the P r o p o r t i o n Reopened outcome is converted to the N u m b e r N o t Reopened. Average T i m e i n Case Management is transformed into an input: T i m e (Days) i n Case Management, to shift this outcome towards a target we w i s h to minimize . 52 The volumes and non-reopened claims outcome are measured as the total number of claims per year. Non-reopenings and volumes can be considered to be valid and distinguishable outcomes, as they capture different aspects of efficiency: throughput and quality. The distribution of these outcomes is displayed in Figures 15 and 17. Although timeliness is transformed into an input for this formulation, there can still be performance targets established in an output-oriented approach, obtained by subtracting the input excess (slack) from the actual average number of days in case management. The caseload input is measured as the average monthly caseload. The geography / travel expense variable is modeled as the annual travel expense accrued by case management, set as an input in the D E A formulation. The input variables describing a desk's claimant profile are captured as proportions for the claim type, injury type, and industry sector. The input "Non-Closure Activities" is the proportion of decisions made by case management where the decision was not associated claim-closure related tasks, such as denying a claimant for wage loss or suspending a claim. Although the input and output factors are measured in different scales, the validity of the D E A efficiency score still holds because the scale measurements are consistent for each D M U . The weights assigned by the optimization are based on comparisons between DMUs as opposed to the magnitude of variables within between a particular DMU's input and output vectors. Normalization of the input and output factors or other scale adjustments is not usually required in standard D E A approaches. 53 Figure 21 displays the distribution of overall efficiency scores, using the CCR Output-Oriented model. The median efficiency score for all SDLs is 88%, with the middle half of the efficiency scores being between 77% and 100%. Figure 21: Distribution of DEA Efficiency Scores 35% 30% 25% c a> a> 10% 5% 0% 60% 65% 70% 75% 80% 85% 90% Relative Efficiency Scores 95% 100% Nearly a third of the DMUs are efficient - this may indicate that too many input and output factors have been incorporated into the D E A formulation, or the need for further post-processing of D E A results to improve cUscrimination. However, having a large number of efficient case management desks proves to be advantageous when evaluating efficiency by SDL, as shown in Figure 22. 54 Figure 22: Distribution of Efficiency Scores by SDL CD i— O ro o or -o i t 111 100% 95% 90% 85% 80% 75% 70% 65% 60% 0 100% 95% 90% 85% 80% 75% 70% 65% 60% SDL Figure 22 shows the same distribution o f efficiency scores, by S D L (names have been removed for confidentiality). W e can see that there are strong performers at b o t h the S D L and province levels. T h i s implies that best practices can be shared to i m p r o v e case management efficiency w i t h i n an S D L , as wel l as w i t h i n a region, or across the province. A C S M may not be forced to l o o k outside his or her region for examples o f best practices. Table 15: Comparison of DEA and Regression Rankings Spearman's Correlation Coefficient Regression (\tolures) Regression (Timeliness) Regression (Reopenings) DEA(CCR-O) Regression (Nfolurres) 1.00 0.63 0.09 0.45 Regression (Timeliness) 1.00 0.24 0.27 Regression (Reopenings) 1.00 0.18 DEA(CCRO) 1.00 Number of Efficient Desks 54 23 55 Table 15 compares the ordinal rankings of the case management desks. The relative rankings of the D E A results (highlighted) are compared against the relative rankings of each of the key outcomes. The rankings of the key outcomes were computed by how much actual performance exceeds the predicted values. Table 15 also shows the number of efficient desks, for each measurement. We define efficiency for each of the univariate outcomes as meeting or exceeding the regression-based target. We can see that the D E A results are not highly correlated with any of the individual performance outcomes, which verifies the interactions between input factors and performance outcomes as modeled by the D E A formulation. We also note that approximately half of the desks are deemed to be efficient for the Volumes and Timeliness Outcomes. This is probably because regression aims to predict for the average value, so roughly half of the observations will be over the predicted values and half will be below (since that is how the residuals should appear in regression analysis). This implies that the regression for the Reopenings outcome is not as strong (confirming the lowR-square for this model, as described in Section 4.3). This also justifies use of the D E A models, as it would be more difficult to discriminate best practice SDLs when approximately half of the units can be described as strong performers. The D E A efficiency improvement projections, as described in Section 3.6.2 can be used to derive performance targets for each case management desk. Figure 23 displays the actual and target distributions (by desk) of Average Time in Case Management for a given SDL. 56 Figure 23: Comparison of Regression and D E A Based Performance Targets 110 100 E 90 CD 2, 80 CO •— 70 CD 1 60 50 40 I 0 O I 1 1 . 110 - 100 - 90 - 80 - 70 . 60 - 50 - 40 Actual Target 1 Target 2 For each SDL, Actual Performance, Target 1, and Target 2 display the range of performance values. Target 1 represents performance targets based on regression-predicted values, and Target 2 represents performance targets based on D E A radial projections. This allows Regional Directors and CSMs to be flexible in setting performance targets, understanding the drivers behind what differentiates case management desks and how optimal performance compares to their unit. There is no time limit or expectation for achieving either of these targets inherent in the models, so perhaps implementation can be done sequentially: first strive to attain the regression-based targets (which are computed based on a desk's dependent variables, or mcoming profile), then set desk level targets to perform at the level of the most efficient desk in the province, assuming that best practices (work styles and conditions) that facilitate peak efficiency can be transferred to inefficient DMUs over time. In other words, if a case manager can reach targets based on expected performance at my SDL for one or more outcomes, the case manager should then try to obtain "peak efficiency" (keeping in mind that efficiency measurements are peer-based). These overall efficiency improvement efforts should be focused towards the facilitation of best practices as opposed to focusing on one particular outcome. 57 Although the efficiency scores and targets shown are based on the CCR model, the analysis was also conducted using the B C C and Slacks-Based-Measurement formulations as well. Tables 16 and 17 show the correlations between D E A models, as well as the median efficiency scores and number of efficient units. Table 16: Ordinal Rank Comparison of D E A Formulations Spearman's Correlation Coefficient CCR-O BCC-O SBM CCR-O 1.00 0.94 0.87 BCC-O 1.00 0.87 SBM 1.00 Table 17: Efficient Units Comparison of D E A Models Number of Efficient Desks Median Efficiency Score Mean Efficiency Score CCR-O 31 88% 87% BCC-O 33 89% 87% SBM 27 72% 76% The Spearman's rank correlation is significant between all three formulations (Constant Returns to Scale, Variable Returns to Scale, and Slacks-Based Efficiency). The efficiency scores and targets reported to senior management are based on Constant Returns to Scale, since factors were chosen based on linear relationships (multiple regression). It could be also argued that since proportion data is present for some input factors, Variable Returns to 58 Scale should be used. Slacks Based Measurement focuses on improving inputs and outputs simultaneously, however most of the inputs associated with a case management desk (other than Time in Case Management) are non-controllable factors. In all D E A formulations, four SDLs (from different regions) were consistently in the top decile of performance scores. Recommendations were made to senior management to use these SDLs as model units from which to facilitate best practices. To further narrow the scope for where to draw best practices from, we can rank efficient desks by the frequency that they are used as peer units for inefficient SDLs. Figure 24 displays how frequently efficient SDLs can perform better than the DMU's efficiency score, i.e. how often the SDL can perform better with another SDL's business rules in another environment. Figure 24: Reference Set Frequency c -o "5 E UJ : 0 5 10 15 20 25 30 35 Number of Times in Reference Set 59 Figure 24 shows the number of times that an efficient D M U is in an inefficient DMU's reference set, with the frequency on the horizontal axis, and the list of DMUs on the vertical axis (masked for confidentiality). 80% of the reference sets include 14 case management desks. This implies that top practices from amongst the efficient DMUs would most effectively be solicited from these 14 case managers. These desks are located in different SDLs throughout the province. As an efficiency improvement scenario, we computed the system-wide improvement if each case management desk were to perform at the median predicted target level for its SDL. Claims Closed, Days in Case Management, and Reopened Claims would improve by 6%, 6%, and 16% respectively. These efficiency gains were also broken down and reported by region. Similar hypothesis can be applied to the D E A analysis as well. If every case management desk performs to at least the median target necessary for D E A efficiency, the median overall efficiency score would increase from 88% to 100%, with 54 desks performing at peak efficiency (a 74% increase from the 31 desks that are currently performing at optimal efficiency. The key to realizing these benefits lies in the successful identification and transference of best practices, and effective initiatives and procedures set by senior management. 60 5. C O N C L U S I O N S A N D S U M M A R Y O F R E C O M M E N D A T I O N S In this thesis we conducted a study of performance indicators, determined key metrics for comparing case management desks based on correlation analyses and CSM feedback, quantified the impact of influencing factors on these measurement outcomes, and determined relative efficiencies and targets based on regression and data envelopment analysis approaches. Results and recommendations were presented to various levels of Compensation Systems management. We provided a comprehensive document (Centre for Operations Excellence, 2001) summarizing the project methodology, data sources, interviews, and key fmdings -namely the quantification of influencing factors and identification of efficient SDLs and case management desks - targeted towards executive management, with sections of recommendations and SDL-specific measurements and improvement targets to be distributed to the Client Service Managers. Table 18 summarizes the recommendations delivered to WCB senior management, as outlined throughout this thesis. 61 Table 18: Summary of Recommendations Performance Sectors Recommendation Measurement Issues Identify a select group of fielded elements representing important activity, and implement a review process for ensuring data accuracy. Communicate to all information users which fields are considered accurate. Develop a process for adding fielded data elements over time. Monitor the completeness of important data elements. Perform stratified sampling of non-STD claims to capture workload related to all entitlement decisions. Develop and communicate guidelines for appropriate time ranges for use in activity reporting. Standardize definitions and usage of typical performance measurements at all levels of management. When reporting case management wage loss duration performance, exclude wage loss on reopenings greater than six months, and time spent in the call center or entitlement. Key Performance Measures Compare performance across time and across case management desks using the following performance measurements: • Claims closed per case manager • Average time in case management (or median and 95th percentile) • Proportion of claimants reopened within six months of first closure Exclude outlier cases (i.e. long duration) when reporting summary statistics such as the average. Conduct regular reviews of quantifiable outcomes to assess relationships between existing and potential performance indicators. Factors Affecting Outcomes Prior to comparing practices across offices and desks, incorporate the impact of the following influential factors: • Proportion of ASTD claims per desk, Caseload, Injury Type, Industry Sector, Geography (approximated by travel expenses), and "Non-Closure" Activities, representing the proportion of time spent denying claims for wage loss and other activities not actively contributing to case closures in the E-File / CARRS workflow systems. Regression-Based Performance Targets & Overall Efficiency Analysis When comparing an outcome measurement across desks or offices, use a regression analysis to calibrate the magnitude of the adjustment for each influencing factor. When comparing overall efficiency across desks or offices, use Data Envelopment Analysis to calibrate the magnitude of the adjustment for each influencing factor. Best practices should be transferable both within and between case management offices; strong performing offices are identified in this study at both the one-way (evaluating each outcome independently) and overall efficiency levels. 62 6. F U T U R E R E S E A R C H A N D A P P L I C A T I O N S The results of this project should be incorporated as part of a continuous improvement strategy and a change management process. Continuous benchmarking against best practices should ensure consistent efficiency gains across all units. From a research-oriented perspective, improvements to our models can be achieved through continual analysis over time, as well as incorporating new modeling approaches and data elements. Ongoing efficiency initiatives emphasizing team building, training, and other "soft skills" are required to fully identify and realize the system improvements outlined in this thesis. 6.1 Modeling Extensions Once more meaningful variables are captured (specifically indicators describing quality), they should be incorporated into the D E A analysis. Data elements from other WCB divisions, such as Prevention, may prove to have interesting relationships with case management efficiency. The migration to a more comprehensive data warehouse is currently in place, and would most likely facilitate the analysis of such information. Elkins and Lawrence (2000) propose combining D E A with forecasting models as a decision support tool over time. This approach would definitely have merit in our application, if Prevention Division information contains predictive capabilities for activity in Compensation Systems. Comparison of the standard D E A methods should again be utilized to ensure consistency among target units, or analyze the behaviour of the production possibilities set when new variables are considered. This process would be most effective if incorporated over time, using a window analysis to track D M U performance and shifts in relative efficiency among units, regions, and across the province. 63 Alternative modeling approaches could incorporate case manager or CSM preference information in determining target performance. The multipliers for input and output factors could be restricted based on assurance-region type constraints provide by CSM input (Cooper et al., pp. 152-154). Selection of input and output factors for the D E A could also incorporate CSM feedback (as opposed to linear regression based methods), through an Analytic Hierarchal Process type approach (Wang and Zhu, 1992). Further discrimination amongst the case management desks may be achieved through a variety of new methods, including using Principal Components Analysis in both the Input and Output factors to obtain a common set of weights for all case management desks (Adler and Golany, 2001). We may also wish to formulate this production set with undesirable outputs, as reopenings and time in case management are actually outcomes of the process. There are also stochastic D E A approaches (Sengupta, 1991) that may yield interesting results since the durations associated with injury recovery, employer compliance, and rehabilitation delays are definitely not deterministic, and may vary between SDLs. The results of this study also support further initiatives beyond performance measurement, including data mining, simulation, and addition to other total quality management techniques. 64 6.2 Ongoing Efficiency Initiatives Implementations of organizational strategies beyond the scope of this project are necessary to facilitate best practices. Simons (1995) recommends that D E A cannot be used in isolation, but as a tool within the cycle of performance management (see Figure 25). The stages of defining inputs and outputs, measuring performance, and setting targets must be related to the objectives of the organization. Regular review of performance indicators with respect to WCB, and specifically Compensation Systems' objectives would maximize the value of overall efficiency analysis as a performance improvement process. Figure 25: Cycle of Performance Measurement & Control 65 Total quality management tools such as affinity diagrams, prioritization matrices, activity network diagrams, and other Six Sigma improvement projects are worth investigating to achieve this goal. Although Compensation Services has regular CSM meetings within regions, interaction and best practices between regions is somewhat infrequent. Another initiative for improving group efficiencies is job-shadowing case management teams with different work styles and business environments. Conway and Forrester (1999) suggest comparing behaviour and practices of process-oriented versus organization-oriented teams to review group effectiveness using theories established in psychology, social network theory, and innovation studies. Although this would be time-intensive (and outside of the Centre for Operations Excellence's realm of expertise), it would be most beneficial in capturing the qualitative elements that drive high quality and peak efficiency, facilitating the quantitative benefits outlined in this thesis. 66 REFERENCES Adler, N . , Golany, B. (2001) Including Principal Component Weights to Improve Discrimination in Data Envelopment Analysis. Working Paper, Available by Request from http://pluto.huji.ac.il/ ~ msnic/adler-publications.htm. Albright, S. Christian (2001). VBA for Modelers: Developing Decision Support Systems Using Microsoft, Excel. Duxbury Resource Center. Charnes, A., Cooper, W.W. and Rhodes, E. (1978). Measuring the Inefficiency of Decision Making Units, European journal of Operational Research, 2, 429-444. Cooper, W.W., Seiford, L .M. , Tone, K. (2000). Data Envelopment Analysis. Kluwer Academic Publishers. Elkins, T.T., Lawrence, K . D . (2000) A Sales & Operations Planning Decision Support System for the Freight Service Industry. Working Paper, Presented at the San Antonio INFORMS Annual Meeting. Farrell, M.J. (1957). The Measurement of Productive Efficiency. Journal of the Royal Statistical Society Series A 120, 253-281. Hatcher, L„ Stepanski, E J . (1994). A Step-by-Step Approach to Using theSAS Systemfor Simple and Multkariate Statistics. Gary, N C : SAS Institute Inc. Johnson, R.A., Wichern, D.A. (1998). Applied Multkariate Statistical Analysis. Upper Saddle River, NJ: Prentice Hall. Lin, C. (2000). A Simulation of Case Management Operations at doe Workers' Compensation Board -A Decision Support Toolfor Htman Resource Allocation. Master's thesis. Vancouver: University of British Columbia. MacFarlane, V., Weltz, A . (1998). Royal Commission on Workers' Compensation in BC -Performance Indicators Final Report Victoria: Province of British Columbia. Nexus Actuarial Consultants Ltd. (1996). WCB - BC Survey of Canadian Workers' Compensation Jurisdictions Final Report. [Mimeograph], Workers' Compensation Board of British Columbia. Ozcan, Y A . (1998). Physician Benchmarking: Measuring Variation in Practice Behavior in Treatment of Otitis Media. Health Care Management Science, 1(1), 5-17. Proudlove, N . (2000). Using Excel for Data Envelopment Analysis. Manchester School of Management Working Paper Series 2000. Retrieved November 13, 2001, from http://info.sm.umist.ac.uk/wp/Papers/wp2007.htm 67 Sanegre, R. (1998). Scheduling Customer Service Representatives for the Workers' Compensation of British Columbia. Master's thesis. Vancouver: University of British Columbia. Sengupta, J.K. (1991). Robust Solutions in Stochastic Linear Programming. Journal of the OperationalResearch Society 42(10), 857-870. Siddharthan, K., Athern, M . , Rosenman R. (2000). Data Envelopment Analysis to Determine Efficiencies of Health Maintenance Organizations. Health Care Management Science 3(1), 23-29 Simons R.(1995). MP in Action - The Newsletter ofMathematical Programmingin IruJustry and Commerce. Retrieved November 13, 2001, from http://www.eudoxus.com/ mpac9504.pdf. Telles, C A . , Nells, T.L. (1999). Benobmarking the Performance of Workers' Compensation Systems: CompScope Measures for Massachusetts. Workers' Compensation Research Institute. Urbanovich, E. (1999). IdentifyingHigh-Risk Claims Within the Workers''Cbmpensation Boardof British Columbia's Claim Inuentary by using Logistic Regression Modeling. Master's thesis. Vancouver: University of British Columbia. Wang, L., Zhu, F., (1992). The Combination CfAHP andDEA And Its Application In Business Management Decision Making Proceedings of the 2nd Chinese Symposium on A H P , Beijing, China, 372-379. 68 APPENDICES Appendix I: Data Elements The following table provides detailed definitions of the terms and variables referenced in this thesis. Data Element Description Wage Loss Duration • The total short-term duration for claims that were closed in the year 2000. • The sum of the WORKDAYS field on all payment records for claims that were closed in the year 2000. Expected Duration Variance • The difference between the short-term disability (STD) and rehabilitation days paid quantity (a.k.a. STD work days lost) and the expected workdays lost for the primary injury of the claim. • The expected work days lost for an injury is based on a median value that is computed by Statistical Services for all injuries of the same type that occurred within a certain time period. Time Between Closure and Reactivation • The number of calendar days between a claim's first closure date and its first reactivation by a case manager. Case Management Wage Loss Duration • The wage loss duration only incurred under case management ownership; excluding wage loss days before a claim is transferred to the service delivery location and wage loss associated with reopenings greater than six months past the claim's first closure date. Average Time in Case Management • The average difference between the date routed to the SDL and the first closure date by case management, in calendar days. Average Time from Injury to First Payment • The average difference between the injury date the first payment date, in calendar days, for all claims closed by case management in 2000. Average Time to Decision • The average time to entitlement from routing to the SDL, in calendar days, for all claims closed by case management in 2000. Average Time in System • The average time from claim registration to first claim closure, for all claims closed by case management in 2000. Proportion of Claims Reopened • Proportion of claims reactivated with an associated payment, within six months of initial closure. Total Wage Loss Payments • The total amount of wage loss payments issued by a desk in 2000. Claims Closed • The total amount of wage loss payments issued by a desk in 2000. Gender • Claimant gender, summarized by the proportion of male claimants by desk Employer Size • Employer size, in terms of number of employees, associated with a claim Age • Claimant age, summarized by the average claimant age by desk Sector • The major employment sector classification groupings for claims closed in 2000. • First summarized by proportions in case management by desk, then clustered by similarities and differences between desks o Summarized at the desk level by proportion of claims from the primary and manufacturing sector 69 Appendix II: SAS Multivariate Procedures and Output Cluster Analysis Description 0ohnson and Wichern, 1998) Cluster analysis is an exploratory data analysis tool used to sort observations into groups, so that the degree of association is strong between members of the same cluster and weak between members of different clusters. In our application, the observations are the variables in each categorical factor (see Section 3.5). We use the average linkagp algorithm to group variables in each categorical factor together. The input to this algorithm is the correlation matrix of the variables. Subsequent notation will refer to this as the distance matrix D = { dik }, where is the correlation between variables i and k. First, the most similar (most highly correlated) pairs of variables are merged together, for example U and V are highly correlated and merged to form UV. Then, the distance between (UV) and other cluster Ware calculated using: SI''. d(uv)w = ~—* > where dik is the distance between object i in the cluster (UV) and object k in the cluster W, and N(uv) and Nw are the number of items in clusters (UV) and W, respectively. These steps are repeated a total of N - 1 times, with all clusters in a single cluster after the algorithm terminates. The identities and levels of the grouped variables and clusters where the mergers take place are recorded. 70 This cluster analysis algorithm is implemented using SAS code with the following structure (http://sasdocs.ats.ucla.edu/). PROC CLUSTER METHOD = name < options > ; BY variables ; COPY variables ; FREO variable ; ID variable ; RMSSTD variable ; VAR variables ; Only the PROC CLUSTER statement is required, except that the FREQ statement is required when the RMSSTD statement is used; otherwise the FREQ statement is optional. Usually only the VAR statement and possibly the ID and COPY statements are needed in addition to the PROC CLUSTER statement. The rest of this section provides detailed syntax information for each of the preceding statements, beginning with the PROC CLUSTER statement. The remaining statements are covered in alphabetical order. 71 Regression Procedure - SAS Code * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * . * CLOSED CLAIMS = f (ALL_SIGNIFICANT_PREDICTORS) ; ******************************************************************************** ** . proc reg data = lib_01.as_desk_08 outest= reg_est; model claims_closed = sdl_travel_costs prop_Y avg_c omplet e_c a s e1oad; output out= l i b _ l l . c c _ f i n a l p = yhat r=yresid; quit; * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * . * AVG TIME IN SDL = f (ALL_SIGNIFICANT_PREDICTORS) ; * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * . proc reg data = lib_01.as_desk_08 outest= reg_est; model avg_time_sdl = prop_Y disallow_rate avg_complete_caseload; output out= li b _ l l . t i m e _ f i n a l p = yhat r=yresid; quit ; * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * . * FINAL MODEL ; * PROP REACT LT 6 MON = f (ALL_SIGNIFICANT_PREDICTORS); * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * . proc reg data = lib_01.as_desk_0 8 outest= reg_est; model prop_lt6mon prop_prim_mfg prop_strains_itis sdl_travel_costs; output out= l i b _ l l . r e a c t _ f i n a l p = yhat r=yresid; quit ; 72 T - T - T - f -+-> O O O C O — o o o co o o o o A • V V V o CO CO < co CO o •r- o in Ll_ i— 3 co in CM T-O •-H A o CO CO CD CM o > • C_ D . V CO CO T 3 C_ CO CO CO CO 3 CO C_ o co T— i-H ca C- in CO cn co CO IT) co CO c_ CO i ^ - CT) i-> CO CM CO c U J o T— T - O J o co ca o •a U - in 4-> o o m o <D CO CM o in <0 o o o o r H o c a> i— o 1 ca i_ CO o cn CO CO o co E 3 0) co CO C T • H cr m CO r_ co C_ CO o in •si- CO CO CO ca i CO • H 0 0 T— C J o r H o 3 cc 4-» CO in 1— CO T-o C T CO E o •<* I"-a> CD CO •1—1 E • H CM o co 7 3 o T 3 CO ca 4-» co a> c cr < CO c_ CO co o o c CO ca +-> CO U J r- 1 cn 3 i— o • H ca a. y— co T 3 _ l i-H I_ CO U J o CO • H o Q 1 > 4 - CO CO T— CM CM • H o O CO o CO "* co CO co •<t U) c_ E 4 - c_ C J CO in CO T— U J a. • H O s CO h - CO CM O T— CO 3 3 in co T— (_ L L . y— y— T— 1— r H r H CO C/) C T i— i— CO CO Q CD O •H co o co CD +-> cc T 3 CO in T— CO O >. C J E CD S 0 ) i-H CO J Z i-H Cfl C_ T 3 1- .a c CO CO CO < 0_ o •iH r H !_ CO CO Ll_ CO m CO c CO CO > Q o o CO • H CO i— i— a> CO o o 1 C o CO a> +J C_ 1 • H T 3 U J c Cfl i-H CO c co CO > CO r H a> T 3 a. > a. C L C H— CO Cfl E CD • H CO 4 - o <_ > o Q O C L CO i-H !_ • H 1 o •—i o CO o CO CO 1 a. 1 CO cr a o JO • H r H O CD 3 s-3 o C/j I m *J 3 C/5 c o o co C - o U J U HH CO Q . < •a ca o CO CO Cfl CO O o 1 o 1 te r H CO CO r H CO C L > C L r H CO Cfl E n O !_ >- O ca c_ 1 O • H CO 1 C L | c_ • H r H O D ) CO c 7 3 C > > h H ( 0 C L < u r-T - 00 i -4-- o o o o — o o o o o o o o A . . . . V V O V o o o a> 3 CS > 00 CO CO "t o in in co in CD E I-a> 3 T -+J T3 _ l CO CD UJ >, O O CO o o t- s CO Q-< CO O rH LU CD CD DC ~a J= O h- CD S CD O T3 c_ LO CD 3 CvJ L_ o CM 1- 0> LO rH CS c_ i— CO CO CO CS i— CJ •<* 13 C- co i - CO CO > C\J 0) C UJ o CO CM o I-- LO CS U- co co •M o i 1- o rH co CVJ T-T3 o o CO I 1 CD c a> E CS r_ oo • H CD CS cn CM 4-1 3 CM CO CD c r 1 c r CM cn C_ CO C- CD CO CO CO LO CO CO CS a> 4-1 LO 1- CO CM > co 3 DC • H CB •<t O "fr LO CB LO c r CD E f - CO LO LO CD CM CO •<—i E •H co 1^ . CO O LO 1 T3 CO CS 4-1 •o c DC < CD t_ CO O ) cn cn o CO CS 4-* CB UJ •<t o 1 •H CS D_ i— CD c_ E • E CS •r4 •H > t - CO CM co 0) CO i— 4-* •W o CD 00 CM CO LO CO 1 H— c co cn CO o CO LU cn O E CB co co co co o LO > 3 3 CM o •>t LO C Ll_ f— y— 1— CS CO CO c r CD D •H co co co oo 4-> CO 1— co i — CD CD >. E rH rH CB JD CS f_ T3 CS C CB CB •H < 0 . O C_ rH CS CD > u_ co LO CO c: CO Q O o CS CB 4-> i — T— CD O c S CD 1 CD 4-* CD •a C_ CB 4-1 c UJ c CS t_ CD CD CO CD > 4-> 1 rH CL m T3 Q. 3 C L CD C CD o E a • H CD 4- O >- rH O O C L a> rH C_ 1 >-* o rH o CD o CD CD C L CB 1 CS DC a o n 4-1 O CO CO H a n < CD O o CD T3 CS O O CO (_ o UJ o +J rH CD CL C L rH CD o E J3 O >- rH O CB c_ 1 rH o •r4 CD Q. CB 1 t_ 4-1 O CO at CS C C- -H > > rH C L T3 < A t_ Q. o i T- i n (o CO O CO CO t O O CO <M O O O O V o o o o o CO CO CO 03 T - O CO o T- f CM CM CO co LO > o c_ CD CO CD CO o s. C o C3 1— LO ' Li. (M •r- cfl !_ i — CO CO LU T3 c_ C J i — CM 00 c o o C LU o o O i -CU CO • co a . •H o o o o o CD 0) o CO CM CU CO c 1— 1— CM t_ a> CO CO 1 S o CO CU c r CO C L c r o co C_ co O CO o CO 1 C_ o o DC C L o CT CD CO •1—1 L _ CD CO co r~ c O o 1 • a co CD + J LO CO T— 1 (D CD C C < a; •H CO LO co T- LU s t_ C L CS + J CU E CM LO t^- LO 03 3 T— o •H co E •H o o O LO + J T3 _ l CD C_ E CO 4-< • CO a) LU C_ CO I_ 03 o o o T ->> o Q 1 > 4- CO CO CO T— CO CO CO 4-" CO LU 1 co CO o O C L o <D LO CM CO o 1— CO (O a . CO c S o c_ o CO 0) CO LU CO a . c_ o S CO o co 1— •0- i — < C L 3 o o o o o c_ c o C3 r—i CO co CT CD LU CU •H co o o o o o o t-> CU CC • a a> CO CD J Z o i—t >. E L i . ! — , — 1- CU s . 0 i—I 10 Q J Z CO CO c_ 1- •H c CO C_ < a . CO > CO L i . co LO CO c > •H (0 Q o o CO 4-> 4-* C 1— 1— CD •H CO CU S D3 1 o T3 >l- CO o C 4-1 c_ E c a> LU c CO 1 •H <—t C L CO a> > •w E CO CU <u s " O C L •H t- > o c t - CD C 4-* CO + J CD >t- O C L 03 C_ o C L a> i-H C_ 1 1 4-> o CD o CD CD C L Q . cfl QC Q o JD 4^ O O i-H w CO C L. t- T3 o _1 n C L C L co o C O •D CD O o .a CO CO > • H 03 CO | O S - CO O E . C I I H .H E CO CD •H C_ > C- 4-1 CO a i t L I I 4->, C L C L I O O I C— C_ "O C L C L 00 Appendix III: Detailed DEA Formulations Sample VBA Code - DEA CCR-O Implementation Sub CalcFormulas() ' This sub calculates formulas for the Model, starting just below the changing cells from the previous sub. Dim i As Integer With Range("A5").Offset(NUnits + 4, 0) ' Set up constraints that input costs incurred must be greater than or equal to output values achieved. .Value = "Constraints that input costs must cover output values" .Offset(1, 0) = "Unit index" .Offset(1, 1) = "Input costs" .Offset(1, 3) = "Output values" 1 There is a constraint for each unit. For i = 1 To NUnits ' Labels in column A (1, 2, etc.) are needed for later on, to enable use of VLookup function. .Offsetd + i , 0) = i ' The input cost incurred for any unit is the sumproduct of the changing c e l l range (UnitCosts) and the appropriate input data row. The same goes for output value. Note how the appropriate row is specified. .Offsetfl + i , 1).Formula = _ "=Sumproduct(InputCosts," & Range("InputsUsed").Rows(i).Address & ") " .Offsetd + i , 2) = ">=" .Offsetd + i , 3) .Formula = _ "=Sumproduct(OutputPrices," & Range("OutputsProduced").Rows(i).Address & ")" Next 1 Name appropriate ranges. LTable is for later on with the VLookup function. Range(.Offset(2, 1), .Offset(NUnits + 1, 1)).Name = "InputValues" Range(.Offset(2, 3), .Offset(NUnits + 1, 3)).Name = "OutputValues" Range(.Offset(2, 0), .Offset(NUnits + 1, 3)).Name = "LTable" End With ' Set up constraint that the selected unit's total input cost is 1. With Range("A5").Offset(2 * NUnits + 7, 0) .Value = "Constraint that selected unit's input cost must equal a nominal value of 1" .Offsetd, 0) = "Selected unit's input cost" 76 ' Get the selected unit's total input cost with a VLookup. With .Offsetd, 1) .Formula = "=VLookup(B3,LTable,2)" .Name = "SellnputValue" End With .Offsetd, 2) = " = " .Offsetd, 3) = 1 .Offset(3, 0) = "Maximize selected unit's output value (to see i f i t is 1, hence efficient)" .Offset(4, 0) = "Selected unit's output value" ' Get the selected unit's total output value with a VLookup. It is the target c e l l for maximization. With .0ffset(4, 1) .Formula = "=VLookup(B3,LTable,4)" .Name = "SelOutputValue" End With End With End Sub Sub RunSolver() 'ReDim InputWeights(NInputs) ' Set up and run the Solver once for each unit, f i r s t placing i t s index (1, 2, etc.) in c e l l B3. Dim i As Integer, j As Integer For i = 1 To NUnits Range("B3") = i SolverReset SolverOk SetCell:=Range("SelOutputValue"), MaxMinVal:=1, _ ByChange:=Union(Range("InputCosts"), Range("OutputPrices")) SolverAdd CellRef:=Range("Inputvalues"), Relation:=3, FormulaText:="OutputValues" SolverAdd CellRef:=Range("SellnputValue"), Relation:=2, FormulaText:=1 SolverOptions AssumeLinear:=True, AssumeNonNeg:=True, MaxTime:=10000, Iterations:=10000, Precision:=0.0000 01 SolverSolve UserFinish:=True 1 Capture the quantities for the report in the TotallnputCost, TotalOutputValue, and EffIndex arrays. For j = 1 To NInputs TotallnputCost(i, j) = Range("InputCosts").Cells(j) * InputUsed(i, j) Next j For j = 1 To NOutputs TotalOutputValue(i, j) = Range("OutputPrices").Cells(j) * OutputProduced(i, j) Next j Efflndex(i) = Range("SelOutputValue") InputWeights(i) = Range("InputCosts") OutputWeights(i) = Range("OutputPrices") 77 Next i ' Reports the results in a separate worksheet. Worksheets("Results").Activate With Range("B4") For i = 1 To 109 .Offsetd, 0) = Efflndex(i) Next End With With Range("C4") For i = 1 To 109 Range (.Of f set (i , 1), .Offsetd, NInputs) ) .Value = InputWeights (i) Range (.Of f set ( i , NInputs + 1), .Offsetd, NInputs + NOutputs) ) = OutputWeights(i) Next i End With End Sub DEA Formulations (Cooper et al., 2000) CCR Input Oriented Dual Formulation Minimize: 9 (real variable) Subject to: 0xo - XX > 0 (input variable constraints corresponding to non-negative input weights) Y A , ^ ^ (output variable constraints corresponding to non-negative output weights) X > 0 (non-negative vector of variables, corresponding to constraint that no D M U can be more than 100% efficient) 78 CCR Output Oriented Dual Formulation Minimize: 9 (real variable) Subject to: 9x0 - XX > 0 (input variable constraints corresponding to non-negative input weights) ya - YX < 0 (output variable constraints specifying that target output performance must exceed existing performance) X > 0 (non-negative vector of variables, corresponding to constraint that no D M U can be more than 100% efficient) BCC Input Oriented Dual Formulation Maximize: z = uy„ - u0 Subject to: vx„= 1 -vX+ uY- u0e= 1 v>0, u>0 where z and u0 are scalars, and e is a vector of l's . 79 BCC Output Oriented Dual Formulation Maximize: z = vx„- v 0 Subject to: uy0 = 1 vX- uY- v0e= 1 v>0,u>0 where z and v 0 are scalars, with being the scalar associated with enforcing variable returns to scale. SBM Linear Formulation Define: ST = ts~ S =ts A = tX where t is a positive scalar. Then, Minimize: 1=1 Subject to: tx0=XA + sr ty0=YA-ST A >0,5^ >0,.S">0, t>0 80 


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items