UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Data quality and outcomes analysis on administrative health data Anderson, Marianne 2004

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

Item Metadata

Download

Media
831-ubc_2004-0353.pdf [ 35.14MB ]
Metadata
JSON: 831-1.0091735.json
JSON-LD: 831-1.0091735-ld.json
RDF/XML (Pretty): 831-1.0091735-rdf.xml
RDF/JSON: 831-1.0091735-rdf.json
Turtle: 831-1.0091735-turtle.txt
N-Triples: 831-1.0091735-rdf-ntriples.txt
Original Record: 831-1.0091735-source.json
Full Text
831-1.0091735-fulltext.txt
Citation
831-1.0091735.ris

Full Text

DATA QUALITY AND OUTCOMES ANALYSIS ON ADMINISTRATIVE HEALTH DATA by MARIANNE ANDERSON B.Sc. (Statistics) Simon Fraser University, 1998 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE IN BUSINESS ADMINISTRATION IN THE FACULTY OF GRADUATE STUDIES (Sauder School of Business) We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA August 2004 © MariAnne Anderson, 2004 JUBCl mm THE UNIVERSITY OF BRITISH COLUMBIA FACULTY OF G R A D U A T E STUDIES Library Authorization In presenting this thesis in partial fulfillment 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. MariAnne Anderson 13/08/2004 Name of Author, (pfease print) Date (dd/mm/yyyy) Title of Thesis: DATA QUALITY AND OUTCOMES ANALYSIS ON ADMINISTRATIVE HEALTH DATA Degree: MScB Year: 2004 Department of Sauder School of Business The University of British Columbia Vancouver, BC Canada grad.ubc.ca/forms/?formlD=THS page 1 of 1 last updated: 17-Aug-04 Abstract Administrative data are collected operationally on functional units through the process of normal business operations. In the case of health organisations, administrative data can include record-level patient demographics such as age and other physical characteristics, injury and treatment dates, codes, and their associated descriptions, as well as treatment fees and other costs. Outcomes research on such data may allow businesses to identify best practices, and to reduce costs by using existing demographic, injury and treatment data as predictors. The impact of poor data quality on an attempt to perform outcomes analysis uncovered the need to examine and address data quality issues and their impact on decision-making and research applications. To demonstrate the effective use of administrative and outcome-measurement data, a custom-application database was constructed and populated with existing (paper) records and with new administrative and outcome data for a pain treatment called Pulsed Signal Therapy. These data were massaged into a form that would allow for statistical analysis and possibly the identification of factors and trends that could expedite experimental design to test any hypotheses which may arise. Analysis of these data indicates that while the initial scores are consistently strong predictors across all measures of outcome, differences between groups appear over time immediately after treatment. Although it appears that older people are more likely to see an improvement than younger clients, age group and injury type are strongly associated. For Pain Frequency, we see not only a main effect based on the age of injury, but an apparent interaction between the age of the injury and the initial Pain Frequency score. Although data at the low end of initial scores are sparse, at the higher end, older injuries are less likely to see an improvement than more recent injuries. Finally, Restriction of Movement at one year post-treatment shows us that patients in the older age group are less likely to see an improvement from baseline Restriction of Movement. Although these are not experimental data, the trends identified through this type of analysis may be useful in guiding experimental design for hypothesis testing. Table of Contents ABSTRACT...! II TABLE OF CONTENTS Ill LIST OF TABLES: V LIST OF FIGURES VI ACKNOWLEDGMENTS VIII 1 INTRODUCTION 1 1.1 T H E W O R K E R S ' C O M P E N S A T I O N B O A R D O F B R I T I S H C O L U M B I A M A N D A T E 2 2 LITERATURE REVIEW 3 2 . 1 Q U A L I T Y O F A D M I N I S T R A T I V E H E A L T H - D A T A 3 2 . 2 O U T C O M E S R E S E A R C H A N D A D M I N I S T R A T I V E D A T A 5 2 . 3 I N F O R M A T I O N T E C H N O L O G Y A N D A D M I N I S T R A T I V E H E A L T H D A T A 8 2 . 4 M E A S U R E M E N T 9 2 . 5 A N A L Y S I S M E T H O D S 1 2 2 . 6 I N F O R M A T I O N S H A R I N G A N D E D U C A T I N G H E A L T H - C A R E P R O V I D E R S O N O U T C O M E S 1 2 3 ADMINISTRATIVE DATA 13 3.1 T H E P R O B L E M 1 3 3 . 2 Q U A L I T Y I S S U E S W I T H A D M I N I S T R A T I V E D A T A 1 3 3 . 3 C H A L L E N G E S 1 5 • 3 . 4 A P P R O A C H E S F O R C O P I N G W I T H T H E S E P R O B L E M S 1 6 4 OUTCOMES RESEARCH ON HEALTH INSURANCE DATA AND THE WCB 17 4 . 1 M O T I V A T I O N F O R O U T C O M E S R E S E A R C H O N H E A L T H I N S U R A N C E D A T A 1 7 4 . 2 S T E P S T O S U C C E S S F U L Q U A L I T Y O F C A R E A N A L Y S I S 1 8 4.2.1 A comparison group 18 4.2.2 A definition of quality 18 4.2.3 Outcome metrics based on this definition of quality 19 4.2.4 A model to compare outcomes 19 4.2.5 Adequate data upon which to build this model 20 4 . 3 B E N E F I T S O F A R E L I A B L E D A T A S E T 2 0 4 . 4 D I S S E M I N A T I O N O F R E S U L T S 2 0 4 . 5 W C B B A C K G R O U N D 2 1 4 . 6 A D M I N I S T R A T I V E D A T A C H A L L E N G E S F A C E D B Y T H E W C B 2 2 4 . 7 D I S C U S S I O N O F C U R R E N T W C B A N A L Y S I S C A P A B I L I T I E S 2 2 4.7.1 Internal analysis capabilities 22 4.7.2 Externally-developed purchased products 23 4 . 8 D I S C U S S I O N O F P R O V I D E R C O M P A R E ™ A T T H E W C B 2 3 4.8.1 ProviderCompare: provider benchmarking on a minimum-cost criterion 23 5 OUTCOMES-RESEARCH ATTEMPT ON WCB ADMINISTRATIVE DATA 25 5.1 M E A S U R E M E N T A N D A N A L Y S I S 2 5 5 . 2 D A T A I S S U E S 2 5 5.2 .1 Specific data issues. 26 5 . 3 U S A B L E D A T A 2 6 5.3.1 Data cleansing 27 5 . 4 T H E E F F E C T O F D A T A C O N T A M I N A T I O N A T T H E W C B 2 8 iii 6 DISCUSSION: RECOMMENDATIONS FOR COPING AND RESOLVING PROBLEMS WITH ADMINISTRATIVE DATA 30 6 .1 T H E F U T U R E V E R S U S T H E P A S T 3 0 6 . 1 7 Caseload • 3 7 6.1.2 Quality control 31 6.1.3 Accuracy incentives 31 6.1.4 Classification granularity 31 6.1.5 Dedicated coding staff 32 7 APPLICATION (PROOF OF CONCEPT): PULSED SIGNAL THERAPY (PST) 33 7.1 P R O B L E M : 3 3 7.1.1 Pulsed Signal Therapy 33 7.1.2 The treatment 33 7.2 M E T H O D S 3 4 7.2.1 Specific business requirements 34 7.2.2 Planning the database 34 7.2.3 Database development 34 7.2.4 PST Research Objective: 35 7 . 3 R E S U L T S ' - 3 6 7.3.1 Descriptive analysis of PST client data 36 7.3.2 Definition of Pain Scale 38 7.3.3 Preliminary analysis 38 7.3.4 Logistic modelling 40 7.3.5 The models 41 7.3.6 Examination of candidates for model building 42 7.3.7 Evaluation of models 43 8 DISCUSSION 64 9 WORK CITED • 65 10 APPENDICES 68 1 0 . 1 A P P E N D I X 1 - P S T D A T A B A S E R E L A T I O N S H I P D I A G R A M 6 8 1 0 . 2 A P P E N D I X 2 - R E L A T I O N S H I P P R O P E R T I E S 6 9 1 0 . 3 A P P E N D I X 3 - U N I V A R I A T E A N A L Y S I S 7 0 10.3.1 Predictors 7 0 7 0 . 3 . 2 Response Variables 72 1 0 . 4 A P P E N D I X 4 - B I V A R I A T E A N A L Y S I S , C R O S S T A B S , I N T E R A C T I O N S 7 3 7 0 . 4 . 7 Cross-tabs: Diagnosis by Sex 7 5 10.4.2 Cross-tabs: Diagnosis by BMI 7 5 10.4.3 Cross-tabs: BMI by Sex 7 5 7 0 . 4 . 4 Cross-tabs: Diagnosis (OA or STI) by Age Group 7 5 7 0 . 4 . 5 Correlations of initial Pain Intensity, Pain Frequency, and Restriction of Movement scores 76 1 0 . 5 A P P E N D I X 5 - R E L A T I O N S H I P S B E T W E E N O U T C O M E S A N D P R E D I C T O R S - 7 7 7 0 . 5 . 7 Logistic Regression 7 7 7 0 . 5 . 2 Right After Treatment 7 9 7 0 . 5 . 3 Six Weeks After Treatment 94 10.5.4 Six Months After Treatment 7 0 0 7 0 . 5 . 5 One Year After Treatment 7 7 3 1 0 . 6 A P P E N D I X 6 - P I V O T T A B L E S . . . : 1 2 3 iv List of Tables Tab le 1 - C r o s s Tabulat ion - D iagnost ic C l a s s by S e x 36 Tab le 2 - C r o s s Tabulat ion - BMI by S e x 36 Tab le 3 - C r o s s Tabulat ion - BMI by Diagnost ic C l a s s 37 Tab le 4 - C r o s s Tabulat ion - D iagnost ic C l a s s by A g e and S e x 37 Tab le 5 - Base l i ne S c o r e s 38 Tab le 6 - S c o r e s at 6 W e e k s Post -Treatment 38 Tab le 7 - Proport ion with At Leas t O n e Unit of Improvement at O n e Y e a r P o s t Treatment 39 Tab le 8 - S u m m a r y of cand ida tes for model bui lding: Signi f icant main effects and interact ions (Whi te=Candidates for mode l bui lding; Shaded=Resu l t s ; P-va lues in parentheses) 42 v List of Figures Figure 1 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Injury Type and Age Group Observed data; Main Effects fitted model; Interaction fitted model : 44 Figure 2 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Age Group Observed data; Main Effects fitted model; Interaction fitted model 45 Figure 3 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model 45 Figure 4 - Initial Pain Frequency vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model 47 Figure 5 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Age of Injury Observed data; Main Effects fitted model 48". Figure 6 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model 49 Figure 7 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Injury Type and Age Group Observed data; Main Effects fitted model ; 49 Figure 8 - Initial Pain Intensity vs. Probability of improvement Six Weeks After Treatment by Injury Type Observed data; Main Effects fitted model 51 Figure 9 - Initial Pain' Frequency vs. Probability of improvement Six Weeks After Treatment by Sex Observed data; Main Effects fitted model.. 52 Figure 10 - Initial Restriction of Movement vs. Probability of improvement Six Weeks After Treatment by BMI Observed data; Interaction fitted model 53 Figure 11 - Initial Pain Intensity vs. Probability of improvement Six Months After Treatment by BMI Observed data; Interaction fitted model 54 Figure 12 - Initial Pain Intensity vs. Probability of improvement Six Months After Treatment by Age of Injury Observed data; Main Effects fitted model 55 Figure 13 - Initial Pain Frequency vs. Probability of improvement Six Months After Treatment by Sex Observed data; Main Effects fitted model 56 Figure 14 - Initial Pain Frequency vs. Probability of improvement Six Months After Treatment Observed data; Main Effects fitted model 57 Figure 15 - Initial Restriction of Movement vs. Probability of improvement Six Months After Treatment by Age of Injury Observed data; Main Effects fitted model 58 Figure 16 - Initial Pain Intensity vs. Probability of improvement One Year After Treatment by Age Observed data; Interaction fitted model; Main Effects fitted model 59 Figure 17 - Initial Pain Intensity vs. Probability of improvement One Year After Treatment Observed data; Main Effects fitted model 59 vi Figure 18 - Initial Pa in F requency vs. Probabi l i ty of improvement O n e Y e a r After Treatment by A g e of Injury O b s e r v e d data; Interaction fitted mode l ; Ma in Effects fitted model .61 F igure 19 - Initial Restr ict ion of Movemen t vs . Probabi l i ty of improvement O n e Y e a r After Treatment by A g e Obse rved data ; Interaction fitted mode l ; Main Effects fitted mode l '. 62 Figure 20 - At Leas t T w o Units Of Improvement Overa l l 124 Figure 21 - At Leas t O n e Unit Of Improvement Overal l 125 F igure 22 - At Leas t O n e Unit Of Improvement, Ma les On ly 126 Figure 23 - At Leas t O n e Unit Of Improvement, F e m a l e s On ly 127 Figure 24 - At Leas t O n e Unit Of Improvement, Al l Wi th Soft T i s s u e Injuries 128 Figure 25 - At Leas t O n e Unit Of Improvement, Al l Wi th Osteoarthri t is 129 Figure 26 - At Leas t O n e Unit Of Improvement, Ma les Wi th Soft T i s s u e Injuries 130 Figure 27 - At Leas t O n e Unit Of Improvement, F e m a l e s With Soft T i ssue Injuries.... 131 F igure 28 - At Leas t O n e Unit Of Improvement, M a l e s Wi th Osteoarthri t is 132 Figure 29 - At Leas t O n e Unit of Improvement, F e m a l e s with Osteoarthri t is 133 vii Acknowledgments I would like to thank Doug Tailing for the many discussions we had on logistic modelling, Nicole Hansen for the invaluable assistance she provided in editing several revisions of this document, Darren Stanley for helping me with many, many formatting issues, and Martin Puterman for giving me the time to complete this work. I would also like to thank Drs. Craig Martin and Greg Franklin and the helpful clerical staff at the Workers' Compensation Board of British Columbia, and Dr. Cecil Hershler for the insightful discussions we had over the course of developing the database and the ensuing modelling which comprise the latter part of this work. Finally, I would like to thank my good friend Rick Smith for helping me learn the intricacies of Access database programming. viii 1 Introduction T h e impact of poor data quality on an attempt to perform ou tcomes ana lys is uncovered the need to examine and address data quality i ssues and their impact on dec is ion -making and research appl icat ions. The project evo lved from appl ied research for the purpose of optimizing the quality of care on exist ing administrat ive data, to a project plan on how to ach ieve this end . Administrat ive data are col lected operat ional ly on functional units through the p rocess of normal bus iness operat ions. In the c a s e of health organisat ions, administrat ive data can include record- level patient demograph ics such as age and other physica l character is t ics, injury and treatment dates , c o d e s , and their assoc ia ted descr ip t ions, a s well a s treatment fees and other costs . O u t c o m e s research may al low bus inesses to identify best pract ices, and to reduce costs by using exist ing demograph ic , injury and treatment data as predictors. A l though administrat ive data do not typically contain ou tcomes measures , proxies such as the duration of a c la im or the total costs assoc ia ted with a c la im may be helpful in determining treatment eff icacy. A l though the quality of administrat ive data may be sufficient for normal bus iness appl icat ions, c ross -use for research can be problemat ic b e c a u s e of incons is tenc ies within non-f inancial f ields: incomplete records or unavai lable data hinder detai led analys is . E v e n worse , inaccurate data may lead to inappropriate conc lus ions . Administrat ive data have a role to play in opt imizing the quality of patient care , to set benchmark s tandards of care , and to reduce operat ing costs . Th is thesis provides the necessa ry requirements and condit ions under which administrat ive data can be success fu l l y and r igorously employed to this purpose. Th is thesis has three parts: 1. A d i scuss ion of administrat ive health data. 2. A d iscuss ion of our attempt to perform ou tcomes research on Worke rs ' Compensa t i on Board ( W C B ) administrat ive data, and the w e a k n e s s in terms of quality in the W C B data. 3. App l ied "proof of concept" : a cus tom da tabase speci f ical ly des igned to col lect ou tcomes data, and an examp le of the type of analys is that can be performed on such data. Wha t fol lows is an examinat ion of data quality i ssues that may ar ise in heal th- insurance da tabases , the types of research that these organisat ions might like to undertake, and the problems they may face in using exist ing data for these purposes. A p p r o a c h e s for cop ing with these problems are d i s cussed . Cha l l enges presented by the current envi ronment are il lustrated with an ana lys is attempt using W C B administrat ive data. Final ly, there is a d iscuss ion of app roaches to coping with and resolv ing these prob lems, wh ich conc ludes with an appl ied da tabase that has been deve loped for the purpose of conduct ing ou tcomes research . 1 1.1 The Workers' Compensation Board of British Columbia Mandate The Workers' Compensation Board is an administrative agency that operates under the authority of the Workers Compensation Act. The WCB is dedicated to the safety, protection, and health of workers. Its employees: • Monitor and promote occupational health and safety practices through enforcement of the Occupational Health & Safety Regulation, worksite inspections, education and consultation. • Provide return-to-work and clinical rehabilitation, compensation and vocational training to workers who are injured or suffer from an occupational disease. • Provide compensation to dependants of workers who have died as the result of a work-related injury or occupational disease. • Provide compensation and assistance to victims of criminal acts (under the authority of the Criminal Injury Compensation Act).1 Workers who are injured in the course of their duties as emp loyees have their medica l cos ts re imbursed by the W C B rather than by the Med ica l Se rv i ces P l a n ( M S P ) of Brit ish Co lumb ia . Compensa t i on for w a g e loss due to disabil ity, whether permanent or temporary, is a lso paid by W C B . W C B is an insurance provider, and premiums paid by employers are set on the bas is of industry c lass and safety per formance. T h e s e premiums, plus income from a variety of investments, are the sources that fund W C B . B e c a u s e W C B is a medica l insurance provider, it emp loys phys ic ians as medica l adv isors who consul t on pol icy dec is ions with regard to medica l concerns . A m o n g other duties at W C B , these adv isors are involved in project deve lopment (such as using administrat ive data for care or cost object ives) speci f ic to the way injured c la imants are treated. 1 Ava i l ab le : <ht tp : / /www.worksafebc.com/corporate/about /goals /c le fau l t .asp>. 2 2 Literature review 2.1 Quality of administrative health-data Lorence & Ibrahim (2003), Re imer (1999), Ch ipman (1999), Kang & K im (1998), Dobrzykowsk i , (1998), and S c h u r m a n (1990) d i scuss quality i ssues that ar ise when gathering administrat ive data. Lo rence and Ibrahim d iscuss how the accu racy in cod ing unf ielded text can vary accord ing whether trained coders or health care providers are entering these data. In particular, they contrast the coding of unfielded text by trained coders , with cod ing done by heal th-care providers. Interestingly, Lo rence and Ibrahim find that dedicated coding staff make fewer errors than do health care providers. They contend that health care practit ioners may not only lack the extensive training of ded icated coders , but may feel that performing this detai led coding takes va luable t ime away from patient care. Re imer examines the trade-off between accu racy and productivity in administrat ive data that a re gathered a s a by-product of the gather ing agency ' s core bus iness . In d i scuss ing Insurance Corporat ion of Brit ish Co lumb ia ( ICBC) administrat ive data, Re imer notes that many factors affect data quality, including such bus iness object ives as productivity inc reases , and the personal exper iences of those col lect ing and recording the data. Re imer asser ts that s ince evaluat ions of data convers ion focus disproport ionately on quantity (i.e., the amount of work accompl ished) rather than quality, data accuracy , and comp le teness may be compromised . Thus , bias and/or incons is tenc ies in the data may be over looked. S c h u r m a n s t resses the need for careful ly des igned information technology (IT) and wel l trained data entry staff in col lect ing administrat ive data. S h e e m p h a s i z e s the importance of having management understand both this p rocess , and the va lue of accurate and representat ive data-report ing for management dec is ion-mak ing . In particular, S c h u r m a n notes the importance of adequate training, not only for those entering administrat ive data, but s o that management understands what is involved in doing so . Th is e c h o e s concerns ra ised by Re imer that evaluat ions of data convers ion focus disproport ionately on quantity. S c h u r m a n asser ts that the regular reporting of accurate and representat ive data provides an important aid for management dec is ion-mak ing , but s ince those lacking formal training in statist ics may not appropriately scrut inize reported f igures, great care must be taken to ensure that these f igures are representat ive. C h i p m a n d i s c u s s e s how the lack of consistent definit ions and criteria among agenc ies can introduce bias based on the col lect ing party's a g e n d a and personal criteria, noting that what information gets recorded may depend upon individual guidel ines or criteria. For example , it may be the c a s e that only the most severe of multiple injuries will be recorded. Th is could mean that "complete" data represent only a b iased subset of all appl icab le records. Fur thermore, b e c a u s e agenc ies ' priorities differ, the amounts and types of information col lected ac ross these agenc ies will a lso differ. Cons ide rab le interest exists in the c ross -use of administrat ive patient records for ou tcomes research . Kang and K im indicate the need for consistent coding structures and da tabase architecture amongs t agenc ies col lect ing administrat ive health data for the purpose of compar ing ou tcomes. 3 They examined the accuracy of e lectronic patient records ( E P R s ) in anticipation of future research . They sugges t that al though the records themse lves (both paper cop ies and electronic forms) were most likely to hold cons is tenc ies , at issue w a s a lack of c o m m o n coding or da tabase architecture s tandards between hospi ta ls. Fur thermore, paper records may be incomplete, contain non-standard abbreviat ions, or have legibility concerns . In determining whether a record holds enough information to be va luab le as a research tool, it w a s conc luded that the lack of overal l s tandards impeded further ana lys is . They recommend that s tandard ized abbreviat ions and terms to be used wheneve r poss ib le . Echo ing concerns ra ised by K a n g and K im, Ch ipman brings to light the lack of consistent definit ions and criteria among agenc ies . Th is can result in d isagreement as to what consti tutes a particular c lassi f icat ion. Therefore, there is a need for the deve lopment of a consistent set of required informat ion. 2 S c h u r m a n d i s c u s s e s the need for adequate outcome data to support all levels of heal thcare dec is ion makers in the a s s e s s m e n t of rehabil itative care . S c h u r m a n d i s c u s s e s data quality improvement strategies that were deve loped a s a serv ice to ass is t rehabilitation facil it ies in develop ing functional a s s e s s m e n t and program evaluat ion sys tems . Instrumental in the deve lopment of these sys tems is the ability for participating facil it ies to use the s a m e sys tem, which ultimately a l lows for benchmark ing and per formance compar isons between facil i t ies. Cons is tent ou tcome data ana lys is and reporting, educat ion, and research are all supported through these strategies. C h i p m a n warns that without communicat ion between agenc ies , a s s e s s m e n t of pol ic ies and programs will lack clarity, and therefore accuracy . T h e deve lopment of a consistent protocol for gathering and document ing data is essent ia l to the effective measurement of r isks and benefits, and cooperat ion among agenc ies is needed in order to meet this end . Data col lect ion and management may lag behind the current capabi l i t ies of information technology. Re ime r holds that a l though much information ex is ts , b e c a u s e it is gathered a s a by-product of the gathering agency ' s core bus iness , the extraction of useful information is often problematic. Re ime r sugges ts that I C B C ' s information sys tem needs to be refined with respect to the convers ion of raw data into accura te and appl icable information. S c h u r m a n s t resses that in addit ion to consistent cod ing , effective and reliable patient tracking and ou tcome reporting requires painstaking precis ion in the deve lopment of information technology that will not only support the ana lys is required for dec is ion -mak ing, but be flexible enough to adapt to changes over t ime. Foreshadowing both Shu rman ' s endorsement of careful ly des igned IT for the col lect ion of data which will be used for reporting, and Kang and K im 's concern with cons is tency in both electronic and paper patient records, Dobrzykowsk i i l lustrates the importance of preparat ion and of thoroughly understanding data availabil i ty and limitations before 2 In the W C B case, inconsistent identifications could refer to ICD9 codes not being reliably coded with the same formats, such as leading and/or trailing zeros. Similarly, the fact that procedure codes change from year to year is indicative of a lack of consistency. Additionally, the lumping together of fees for multiple procedures under single codes makes it virtually impossible to trace backwards to identify the individual procedures performed on claimants. 4 attempting to perform ou tcomes research . In particular, Dobrzykowsk i sugges ts evaluat ing the fol lowing factors before attempting ou tcomes research on archival data: • Subscr ibe rs • Da tabase s ize • O u t c o m e s sca les /measu res • Data col lect ion methods • Data integrity • Repor ts • Training and support serv ices • Rater reliability • Contractual requirements • C o s t s 2.2 Outcomes research and administrative data G o e r g e & Lee (2002), B e c k m a n (1999), Schuber t (1999), Sz ick , A n g u s , N icho l , Harr ison, P a g e , & Moher D (1999), R y a n (1998), B rookes (1997), Wi l l iams & Y o u n g (1996), Parente , Weiner , Garn ick , R ichards , Fow les , Lawthers, Chand le r & P a l m e r (1995), Wiener , Parente , Garn ick , Fowles , Lawthers & Pa lme r (1995), and Mant & Hicks (1995) d i scuss aspec ts of ou tcomes research with administrat ive data. In a conversat ion with B e c k m a n , the University of Brit ish Co lumb ia Rehabi l i tat ive S c i e n c e professor lamented that "nothing good " had been done in the a rea of ou tcomes research on administrat ive health data. At first g lance, this s e e m e d like an opportunity to break ground. But upon further study of this prob lem, it b e c a m e c lear that this has not been for lack of trying. Wi l l i ams and Y o u n g note that a l though health serv ices researchers use administrat ive data for ou tcomes research , such data may not be appropr iate, complete , accurate , and of high enough quality to contribute to effective research . They note that al though health serv ices researchers in C a n a d a and e lsewhere draw on information der ived from administrat ive da tabases for ou tcomes research , these da tabases were originally organ ized for documentat ion and payment purposes. Wi l l iams and Y o u n g identify three sou rces of administrat ive data often s e e n in current ou tcomes research : demograph ics such as age and sex , d i sease a s coded using the International Class i f icat ion of D i s e a s e (ICD-9), and the procedures (and their assoc ia ted codes and costs) performed by health care providers. T h e fact that administrat ive da tabases were not originally intended to be used for research purposes ra ises for concern the issue of data quality, a s well as patient pr ivacy and consent . Schuber t (1999,15) notes: the motor vehicle insurance, workers' compensation and disability-management industries have placed more importance on the measurement of the economic consequences of claims than on the human and social costs of injury. This is illustrated by the fact the information gathered by these institutions is predominantly financial, relating to the length of time that claim files remain open, and to the type (medical cost or wage-loss) and amount of payments. 5 Schuber t further d i scusses the need to identify and col lect the pat ient-centred measurements that lead to improved ou tcomes , as ide from the usual cost and duration of c la im data that are typically measu red . A s ana lys is of huge datasets b e c o m e s more feasib le, there will be increasing d e m a n d for this type of research b e c a u s e of its potential to contribute to restructuring and f inancial sav ings . At its best, administrat ive data can be a va luable source of information to researchers and pol icy-makers, enabl ing them to improve quality of care and predict future health care requirements whi le reducing costs assoc ia ted with inefficient and ineffective care. However , in order that this type of research is r igorous, the data must be high quality. Schuber t asser ts that with the reasons for heal th-serv ices research ranging from pure economic necess i ty to the grander ideals of furthering knowledge and improving the quality of patient care, the importance of trustworthy and representat ive data is becoming paramount. Improvements in the quality and compatibi l i ty of the da tabases must be performed to real ise these goa ls , and this will necessar i l y involve a t remendous investment in quality control measu res to improve data accuracy . G o e r g e and L e e extol the virtues of administrat ive data for program-part ic ipat ion research , but lament incons is tenc ies in the col lect ion and documentat ion of s u c h data. In spite of the many drawback involved, G o e r g e and Lee contend that administrat ive data are general ly super ior to other forms of data for program-part ic ipat ion research . They further sugges t that because administrat ive data are col lected on "whole populat ions" (i.e., as a census) , they provide the researcher with the possibi l i ty to study low inc idence phenomena that are expens ive to uncover using surveys . G o e r g e and Lee indicate that administrat ive data provide information about which respondents would unlikely be truthful in an interview. However , they indicate that the biggest drawback to us ing administrat ive date is not necessar i ly in the data themse lves , but in the lack of documentat ion about the quality of these data. Their recommendat ions with respect to c leaning and organiz ing administrat ive data concentrate on the determinat ion of internal incons is tenc ies, a s wel l as the methods of col lect ion, p rocess ing , and main tenance used before del ivering data to the researcher . Addit ional ly, they sugges t that researchers famil iar ize themse lves with related datasets , and get to know the operat ions and serv ice del ivery methods of the program under ana lys is in order to understand the incons is tenc ies that will inevitably ar ise. In an effort to gain this type of knowledge on the W C B ' s sys tem of adjudicat ion and appea ls p rocess , the Roya l C o m m i s s i o n into the Worke rs ' Compensa t i on Board (1998) requested an integrated set of W C B data so that c la ims could be t racked from incident report through adjudicat ion and appea l . Th is longitudinal study of c la imants who registered a new c la im in 1994, 1995 or 1996 b e c a m e known as "The Cohor t Project". The cohort of roughly 600 ,000 c la ims w a s fol lowed through the end of 1997. In analys ing these data, R y a n performed an a s s e s s m e n t of the eff iciency of the sys tem to cope with its own bus iness . R y a n notes that data that are col lected through the different s tages of c la imant life cyc les lack cons is tency or integrity, and that avai lable data reflect the "s i lo ing" which occurs between departments. Cohor t development , identif ication, and extract ion, and the ana lys is itself were each co lossa l undertakings, requiring months of co-operat ive effort on the parts of the Roya l C o m m i s s i o n and var ious departments of the W C B . T h e W C B provided the Cohort 6 Project with severa l large legacy (first generat ion) da tabases containing over 100 tables and mil l ions of records. Further data, relating to Workers ' Compensa t i on Rev iew Board appea ls , were obta ined f rom the Rev iew Boa rd . T h e s e were not relational da tabases . Whi le these sundry data sou rces al lowed for c la ims and appea ls data to be l inked, the result ing dataset w a s far f rom integrated. R y a n demonst ra tes that identifying, locat ing, and extracting the information relating to three cohort groups of injury c la ims involved a t remendous commitment of W C B resources , particularly s ince they were g leaned f rom board information sys tems that were not des igned to provide an integrated v iew of the life of a c la im: The various individuals, departments and agencies involved with different stages of the claim's life cycle collect information as needed for their own purposes and there is often little consistency or integrity between the various data sets. In short, the available data mirrored the compartmentalized, snap-shot perspective of the business (11). Sz ick et al. address many of the problems assoc ia ted with any comparat ive study of health care ou tcomes as they relate to the compar i sons of the quality of care. They asser t that an objective operat ional definition of "quality" is necessary . Cand ida tes for research could be the trade-off between total disabil ity days and medica l cos ts , or be tween total disabil ity days and total c la im cost. (It is interesting to note that this is exact ly what P rov ide rCompare tried to do. Th is will be d i scussed below.) However , they note that both the length of the disabil ity and the pay rate earned by the injured worker are reflected together in the total c la im cost. They sugges t building a proxy out of a s tandard ised income model so that the amounts could be -added in a more meaningful way, rather like a weighted s u m of total disabil i ty days and medica l costs . We ine r et al. p ropose that the profiling of exist ing administrat ive c la ims data can be used as an ongoing method to support qual i ty- improvement activit ies through the deve lopment and d isseminat ion of pract ice guidel ines. Motivation for this type of research c o m e s from mounting pressures within the Med icare program to help monitor and control the costs assoc ia ted with caring for the health needs of an ag ing populat ion. In this appl icat ion, We ine r et al. examined the billing records of a number of phys ic ians using a ser ies of d iabetes-tracking procedures. In doing so , they were ab le to determine that elderly patients with d iabetes did not appear to be receiving optimal care. B e c a u s e quality ou tcomes are not measured in administrat ive da tabases , proxies for these m e a s u r e s are somet imes built out of exist ing f ields, s u c h a s the durat ion and cost of a c la im. Mant and H icks warn of the pitfalls in building proxy measu res for ou tcome. For example , they d i scuss the inappropr iateness of an "eff iciency index", which encourages increased activity for the s a m e health care expendi ture, but fails to cons ider the benefi ts or adverse effects of the measured activity on health. T h e suggest ion is that ou tcomes measurements can be comparat ive ly insensit ive tools for a s s e s s i n g the quality of care when compared with relatively short audits of p rocess of care. Looking to longer-term assessmen t , Schuber t ci tes the current emphas i s that S a s k a t c h e w a n Government Insurance (SGI) has on rehabil i tation. H e asser ts that the expedit ion of recovery t imes would be benef ic ial in that it could both help the injured person and society as a whole, and reduce the cost of automobi le insurance. S G I contracted with the University of Saska t chewan ' s Institute for Health and O u t c o m e s R e s e a r c h ( IHOR) to help S G I redes ign the way and the type of data that is gathered, 7 and to include the c la imant in this col lect ion. SG I fol lows up with c la imants ' information regarding their health and quality of life for a year, and I H O R is analys ing the data to s e e which factors are important to recovery. A l though the new Appl icat ion for Benef i ts is five t imes as long as the old one, compl iance s e e m s to be good s ince patients are provided with sufficient explanat ion as to its role. T h e new form captures information about such demograph ics a s marital status and gender , but it a l so conta ins a s tandard ised pain d iagram and a descr ipt ion of pain intensity and related disabil i ty that is filled out by the claimant. Other quest ions address the impact of the injury on day- to-day activit ies, and even get at certain aspec ts of the c la imant 's personal i ty factors, such as a proclivity toward dep ress i on . 3 M u c h of this type of data col lect ion is f ocussed on identifying individuals who are at risk of develop ing chronic condit ions. Parente et al. have produced a detai led methodology of how to transform administrat ive health data into a form useful for both research and management a ims , including quality of care analys is . Their d i scuss ion s p a n s c la ims-data architecture, contents and accuracy , methods, and costs assoc ia ted with convert ing c la ims data into ana lys is data, technical innovat ions in health pol icy studies, and patient confidentiality. Us ing health insurance (Medicare) data, Parente et al. deve loped a high-quali ty da tabase through the employment of strategies they deve loped from studying the contents, architecture, and accu racy of c la ims data. In doing so , patient-data that had been col lected from var ious sou rces (in this c a s e , hospital admiss ions , phys ic ian office serv ices , outpatient serv ices , nursing homes , and home health serv ices) were gathered, c l eansed , manipulated, and matched on a record level. Parente ef al. determined that with these strategies, health insurance data can be turned into a rich source of information for a s s e s s i n g the "quality of care del ivered". Final ly, B rooks has deve loped what should be necessa ry reading for anyone attempting to do ou tcomes research using an Amer ican-s ty le Heal th Managemen t Organisat ion ( H M O ) mode l . Under this parad igm, primary care providers are the ga tekeepers of patient care, burdened with the immediate responsibi l i ty of keeping individual patient cos ts low. They are held f inancial ly responsib le in c a s e s where their patients use too many "downst ream" serv ices such as specia l is ts and medica l rehabil i tation. In the C a n a d i a n context and many other jur isdict ions, primary care providers face no such barrier to patient care , and it is important to dist inguish between the two sys tems of health care before d iscuss ing health care da tabases . Never the less , the author feels that even with their imperfect ions, these da tabases can and should be used immediately to enhance understanding where poss ib le . For example , Intermountain Health C a r e Sys tem 's Cl in ical P rogram Monitor ing uses time ser ies to model cos ts over t ime and lengths of c la ims over t ime - a reas where data quality are likely to be of better quality than they will be for non-f inancial measu res . 2.3 Information technology and administrative health data E a c h of these sources Who ley , P a d m a n , Hamer & Schwar tz (2000), and S tausbe rg , Ko ike & Albrecht (1998) emphas i ze s o m e aspec t of IT for administrat ive data. Depression is associated with increased perception of pain and longer recovery times (Marano 2002). 8 The authors contrast the deve lopment of administrat ive and cl inical health information sys tems . T h e administrat ive uses can al low tracking of health-plan enrolment, c la ims t ransact ions, and physic ian profil ing, while the cl inical aspec t revolves around patient health records, wh ich c l in ic ians use to m a n a g e the del ivery of care . W h o l e y et al. d iscuss the integration of administrat ive and cl inical health information sys tems as a tool to support health management initiatives. Efficient and cost-effect ive health care may require the integration of these funct ions. Who ley et al. purport that al though Heal th Ma in tenance Organ izat ions ( H M O s ) use information technology for both administrat ive and cl inical pu rposes , its use is still mainly limited to managing c la ims and enrolment information, and to generat ing reports. They note that the effective use of IT to support providers in the del ivery of health management initiatives is in its infancy. B e c a u s e data wa rehouses organ ize data from different sou rces in a logical way, they al low for a smoother transition from data to information. By providing a c o m m o n portal to var ious data sou rces , the cos ts assoc ia ted with s u c h types of ana lys is a s phys ic ian profiling and quality measurement can be reduced. S tausberg et al. d i scuss var ious opt ions for coping with the d e m a n d s of performing ou tcomes ana lys is through legacy sys tems , noting that most hospital information sys tems are cons idered to be legacy sys tems. Whi le such sys tems can provide c lose integration of different administrat ive functionalit ies with their charac ter -based user interfaces, they lack the user-fr iendly familiarity of the now-common graphical user interface (GUI). Furthermore, these sys tems may be unable to support Structured Query Language ( S Q L ) , making it difficult to produce ad -hoc quer ies for the ana lys is of cl inical data. Opt ions d i scussed include replac ing, capsulat ing, or complement ing exist ing legacy sys tems. The re is a trade-off between the up-front cos ts of building the data wa rehouse proposal by W h o l e y et al., and the continuing costs of awkward a c c e s s to ana lys is datasets . S ince the costs to replace legacy sys tems are prohibitive, and the addit ion of ana lys is tools can make the overal l sys tem unworkable, an alternative needed to found. For compi l ing information and subsequen t ana lys is , S tausberg et al. r ecommend developing a network of s i te-speci f ic P C s that is l inked to a central sys tem for data storage. 2.4 Measurement C a s s i d y (1999), Pa rsons (1999), Law, K ing, Russe l l , M a c K i n n o n , Hurley, & Murphy (1999), St raker (1999), Davenpor t & Dennis (1996), G range r & Brownsche id le (1995), Dav ies & Cromb ie (1995), Donabed ian (1988), and C a r e y & P o s a v a c (1982) d i scuss i ssues surrounding the measurement of treatment ou tcomes. Affect ing all aspec ts of data col lect ion, the issue of measurement is surpr is ingly compl icated when it c o m e s to health data. F rom establ ishing basel ine injury statist ics, to choos ing between patient-centred and cl in ic ian-centred measu res of treatment ou tcome, what we measu re and what we use to measure it are the tough quest ions that prec lude data col lect ion. Echo ing concerns vo iced by many in the industry, P a r s o n s s p e a k s to the importance of establ ishing an accurate descr ipt ion of injury in C a n a d a . Th is would al low for the deve lopment of basel ine national injury statist ics, which could then be used to evaluate 9 injury-prevention initiatives. T o ach ieve this end , P a r s o n s e m p h a s i z e s the need for an integrated and systemat ic approach to data col lect ion. Pat ient-centred measu res pertain to the patient's wel l -being, functional status, and health-related quality of life. Cl in ic ian-centred measu res pertain to things that can be examined easi ly by c l in ic ians, such as the degree of degenerat ion, or the range of motion. C a s s i d y holds that a s s e s s m e n t of ou tcomes should emphas i se pat ient-centred measu res , al lowing that whi le primary measurements may be useful for guiding cl inical care, they do not necessar i ly add ress the needs of the patient in that they may not be valid predictors of recovery. C a s s i d y holds that a s s e s s m e n t of ou tcomes should e m p h a s i s e patient-centred measu res . The author notes that there are psychometr ical ly sound measurements of patients' own accounts of their health status: both the 36- i tem Short Fo rm Heal th Survey (SF-36) the S i c k n e s s Impact Profi le (SIP) have a l ready shown success fu l results in compar ing ou tcomes for patients with chronic condit ions. T h e s e techniques, when monitored over t ime, may be an effective way to al low causa l in ferences about factors that may be most responsib le for final patient status. Th is technique would be simi lar to the quality control methods used in industry to monitor and correct or improve p rocess , and could thus be used to d e c r e a s e variabil ity in care by providing feedback to practit ioners and other s takeholders . C a s s i d y indicates that the argument against pat ient- focussed measurement is that it is subject ive, which tends to imply that it is a softer form of measurement . However , c l in ic ian-centred methods are a lso subject ive, and may not all be particularly rel iable. L a w et al. d i scuss the deve lopment of a dec is ion protocol as to which ou tcome measu res should be used in client and program evaluat ion. Donabed ian m a k e s an important point regarding the trade off between implicit and explicit measuremen ts for quality of care , noting that the quality of care is proportional to its ef fect iveness. Wh i le implicit measurements are ultimately f lexible, they lack a cr isp definition that would al low for the deve lopment of guidel ines used for quantif ication, and hence general izat ion. Converse ly , explicit measurements are categor ical , but may be too coa rse to al low the level of distinction that may be required to account for variability within each category. C a r e y and P o s a v a c d i scuss var ious aspec ts of LORS- I I (Rev ised Level of Rehabi l i tat ion Sca le ) , such as how to use it to deve lop expectat ions for patient ou tcomes , and to compare actual ou tcome levels from groups of providers. A s s e s s m e n t s are made using a four-point sca le by Occupat iona l Therap is ts , Phys ica l Therap is ts , S p e e c h Therap is ts , and Regis tered Nurses . Straker offers a review of var ious body discomfort a s s e s s m e n t tools. Whi le the focus of this article is on ergonomic assessmen t , there are other poss ib le appl icat ions to these measu res , such as pain measures . A theoretical d iscuss ion of discomfort a s s e s s m e n t pre faces a descr ipt ion of the var ious tools avai lable, with the intent of ass is t ing in the dec is ion of when to use discomfort a s a measure and which tool to use . There are a large number of possib le subject ive sca les , which the author has grouped into the fol lowing broad c l asses : verbal rating sca les , v isual ana logue sca les , numer ic rating sca les , and graph ic rating sca les . T h e v isual ana logue discomfort sca le is probably the most widely appl icab le of these tools, and s p e a k s to C a s s i d y ' s call for pat ient-centred measures . 10 Straker warns that it is not discomfort itself that is being measu red ; rather, it is a correlat ion that is a s s u m e d between the measure taken, and discomfort. Furthermore, pain researchers have found that different people interpret terms differently. For example , numbness may s e e m worse than stiffness to one person, whi le the opposi te may be true for another. T h e author recommends the use of a V isua l Ana logue Discomfort S c a l e or Verba l Numer ica l Rat ing S c a l e for the a s s e s s m e n t of intensity. Wi th regard to repeated measures , Donabed ian m a k e s an important note about the psycholog ica l effects of cal l ing patients back to be re-examined when more prec ise information on their ou tcomes is needed : this very act ion can change ou tcomes measu res in a manner simi lar to that demonstrated in the famous Hawthorn c a s e . 4 Grange r and Brownsche id le d i scuss the need for program evaluat ion a n d uniformity in ou tcome measurement in the context of medica l rehabilitative procedures. T h e Funct ional Independence Measu re (FIM) and the Uniform Data S y s t e m for Med ica l Rehabi l i tat ion (created at University, of N e w York , Buffalo) are d i s c u s s e d , a long with the importance of continuity in the reporting on injury severi ty and ou tcomes . Dav ies and Cromb ie sugges t that b e c a u s e the reliability, validity, and relative lack of b ias involved in m e a s u r e s us ing wel l -def ined p r o c e s s e s of care (what is done to patients, where, when , and how) on speci f ic patient groups, confounding c a u s e d by the case-m ix can be less intrusive than with measur ing ou tcomes. T h e authors d i scuss a major w e a k n e s s in ou tcomes measu res - s ince these are based on observat ional s tudies, compar isons with historical results can easi ly be confounded by di f ferences in the c a s e mix. Dav ies and Cromb ie further note that attempts at adjust ing for these di f ferences have met with only limited s u c c e s s - b e c a u s e adjustment techn iques are difficult to use , they are rarely appl ied. At i ssue is the fact that identifying and routinely col lect ing relevant prognost ic factors is necessa ry in order for the appropriate adjustments to be made , and that this may not be s o easy to ach ieve . Th is sa id , Davenport and Dennis point out that measur ing the p rocess of care may only be useful when the c a s e mix is well def ined and homogeneous , s ince the p rocesses of care will vary accord ing to the severity of injury. 5 Donabed ian identif ies two aspec ts of heal th-care per formance, one technica l , and the other, interpersonal. The author points out that the quality of technical care is proportionate to the ef fect iveness of that care, where ef fect iveness is def ined a s the real ized fraction of heal th-outcome improvement when compared with the improvements in health status that are currently poss ib le and avai lable. Not ing that this compar ison is a lways against those improvements in health status that the current health care pract ices have made poss ib le , the quality of care may change over t ime: as techniques improve, s o do expectat ions. However , quality of care can still appea r consistent over t ime if we adjust for what is known at the t ime e a c h of these quality measurements is a s s e s s e d . 4 In the famous Hawthorn experiment, researchers were trying to prove that working conditions, such as good lighting, had a direct effect on productivity. As we now know, they found that productivity improved no matter what visible change was made. They deduced that people perform better when they feel cared for and when something is being done to help them. 5 This may be a noteworthy factor in the case of the P S T data, where injuries range from degenerative aging diseases to injuries in young, fit, professional athletes. 11 Donabed ian asser ts that the c o n s e q u e n c e s of care depends both on the patient's percept ions, and the considerat ion of monetary cost. There is a dif ference between optimally and maximal ly effective care, in that opt imum care can be thought of a s the derivative of the curve def ined by Benef i ts over Cos t , where maximal ly effective care is s imply the highest point on the benefits curve. 2.5 Analysis methods Horton & Lipsitz (1999) cons ider multivariate methods deve loped for analys is of t ime ser ies data, where the observat ions are not independent , may be appl ied to other si tuat ions where the assumpt ion of independence may be violated. In the c a s e of treatment paths, an individual who rece ives physiotherapy as part of his or her surgical after-care has not received this therapy independent ly of this surgery. Th is situation resembles t ime ser ies in the s e n s e that it is a s e q u e n c e of treatments, and s o the techniques descr ibed in the article may apply. 2.6 Information sharing and educating health-care providers on outcomes Freemant le , Harvey, Wolf, G r imshaw, Grilli & Bero (1999), and Bas insk i (1993 & 1995) cons ider the communicat ion of ou tcomes research . Freemant le et al. speak to the issue of determining effective m e a n s of d isseminat ing and communicat ing information amongs t health care practi t ioners, noting that these m e a n s should enab le this information to be received as well as poss ib le . Bas insk i notes that al though many guidel ines have been deve loped in an effort to improve and maintain the quality of health care, little has been done in terms of their strategic implementat ion. D isseminat ion of guidel ines through journal publication or product ion of educat ional materials d o e s not ensure that they will be put into pract ice. In order to improve the implementat ion of research f indings, the author states that a variety of implementat ion strategies must co inc ide with d isseminat ion. Success fu l implementat ion of guidel ines a lso requires that all partners (from patient to polit ician) in the health care sys tem work together. Further research and evaluat ion of implementat ion strategies are required, as are the resources necessa ry for conduct ing such investigat ion. Only after these needs are met can providers begin to work towards coordinated and comprehens ive implementat ion programs. C o m p a r e d with the current focus on the deve lopment of cl inical pract ice guidel ines, the effort devoted to their evaluat ion is meagre when we cons ider that the ultimate s u c c e s s of such guidel ines depends on their routine evaluat ion. In later work, Bas insk i noted the importance of evaluat ing guidel ines before they are implemented, and of evaluat ing the health care programs in which these guidel ines play a central role. 12 3 Administrative Data "Administrat ive data" refers to data that are col lected operat ional ly on funct ional units through the p rocess of normal bus iness operat ions. In the c a s e of health organisat ions, administrat ive data can include record- level patient demograph ics such a s age and other physica l character ist ics, injury and treatment dates, codes , and their assoc ia ted descr ipt ions, a s well as treatment fees , and other costs . 3.1 The Problem O u t c o m e s research may al low bus inesses to identify best pract ices and to reduce cos ts by using exist ing demograph ic , injury, and treatment data as predictors. A l though administrat ive data do not typically contain ou tcomes measu res , proxies such as the duration of a c la im or the total costs assoc ia ted with a c la im may be helpful in determining treatment eff icacy. A l though the quality of administrat ive data may be sufficient for normal bus iness appl icat ions, c ross -use for research may be problematic. Incomplete or unavai lable data may hinder detai led analys is . E v e n worse , inaccurate data may lead to inappropriate conc lus ions . There is a need to examine and address administrat ive data quality i ssues , and their impact on dec is ion-making and research appl icat ions. During a project that sought to ana lyze W C B claimant data, it b e c a m e c lear that there w a s a need to examine and address data quality i ssues and their impact on dec is ion -making and research appl icat ions in industr ies using administrat ive data. T h e project evo lved from appl ied research for the purpose of opt imizing the quality of care on exist ing administrat ive data, to a project plan on how to ach ieve this end . T h e utility of this thesis c o m e s from the desire and the potential to opt imize the quality of care , to set benchmark s tandards for health-care providers and c a s e managers , and to possib ly reduce costs due to ineff iciencies in the del ivery of health care. Cos t sav ings from such research could a lso be real ised from fraud reduct ion: in c a s e s where projected ou tcomes fall terribly short of observat ions, f lags could be set for further investigat ion of these c la ims. However , before benchmark ing can lead to "best" pract ice guidel ines, the quality and cons is tency of data col lect ion must improve. Rea l is ing potential ga ins from ou tcomes ana lys is depends upon c lear and appropriate communicat ion of goa ls and capabi l i t ies between those analys ing the data, and those manag ing it. 3.2 Quality Issues with Administrative Data There are multiple appl icat ions for administrat ive data: administrat ive needs , f inancial requirements, and , in s o m e c a s e s , research studies, and statistical ana lyses . T h e third is not usual ly the concern of people who set up da tabases . Administrat ive data are col lected for different reasons by different depar tments with different priorities. B e c a u s e , up until recently, comput ing speed and storage have been expens ive , different departments within the s a m e organisat ion may e a c h have separa te , 13 stand-a lone da tabases . In many c a s e s , these s tand-a lones are legacy sys tems , wh ich , while inefficient, are very expens ive to rep lace. Departmenta l da tabases may lack the defaults, error check ing , and enforced f ields that could improve data comp le teness and accuracy . Fur thermore, with severa l s tand-a lone sys tems that may contain over lapping information, updating one da tabase does not c a s c a d e to update them all . Al l bus inesses keep information on their cl ients. In the c a s e of la rge-sca le organizat ions, such as health and other types of insurance, record- level information is often very detai led, including far more than just payment information. It would be advan tageous to ana lyse and use these exist ing data to reduce operat ion costs . In the c a s e of health insurance providers, administrat ive data are often gathered from employers , c la imants, and heal th-care practi t ioners, and may be handled internally by c a s e managers , entit lement off icers, da ta entry technic ians, account ing personne l , risk a s s e s s o r s , information location teams, medica l adv isors , and many others. Da tabase sys tems that have deve loped independent ly among these var ious depar tments may have different definit ions, s tandards, needs , and time f rames. Th is results in a narrow focus, which hinders cross-depar tmenta l ef fect iveness. For example , C h i p m a n (1999) writes that the Insurance Corporat ion of British Co lumb ia ( ICBC) is facing this type of prob lem: "Different agencies have different priorities..." (7). "...Close cooperation is required to obtain the measurements we need in order to document the characteristics of road crashes, to assess the effectiveness of safety measures, and to reveal the safety effects of measures whose primary outcome was something other than safety" (7). Re imer adds : ICBC's information system needs considerable refinement to convert raw data into accurate information that action can be based on (4). S a s k a t c h e w a n Government Insurance has had to dea l with this problem. A l s o , Utah 's Workers ' Compensa t i on A g e n c y has undertaken to use a product cal led P rov ide rCompare (which will be d i scussed later) after spend ing severa l yea rs remedying their data w o e s (Huffman 1999). Furthermore, Wi l l iams (1996) writes: As health services research and the data on which it relies become increasingly important to health reform and the evaluation of efficiencies of interventions, the existing data sources must be improved. Substantial benefit could be gained from an investment to improve the quality and compatibility of databases. A considerable investment is required in quality control, however, to improve data accuracy. A problem introduced by a narrow, intra-departmental focus (often referred to a s "siloing") is that it can make it difficult for any one department to take the initiative to work on address ing these issues . Th is level of data reform is a formidable problem, one that involves and affects every department in its own way, and that s e e m s to grow in magni tude upon c loser inspect ion. 14 Responsib i l i ty for manag ing the f inancial and other resources required to accomp l i sh this miss ion will span many (if not all) departments of the organisat ions involved. T h e crux of the problem is that a group of depar tments individually trying to max imize their operat ing ef fect iveness is not necessar i ly the best way to opt imize overal l corporate ef fect iveness. 3.3 Challenges W h e n ^information is col lected from different bus iness units and/or sys tems , attempts to perform ana lys is may be frustrating. Data quality i ssues need to be add ressed before any meaningful c ross -purpose ana lys is can be performed. Exist ing ana lys is tends to revolve around f inancial information such a s c la im duration or total c la im cost. Th is is not s o much b e c a u s e it is the most useful type of ou t comes research , but rather, because f inancial information tends to contain the most rel iable data for ana lys is (Wil l iams, 1996). Th is is an artefact of administrat ive da tabase des ign -such da tabases exist for the documentat ion and payment information assoc ia ted with health c la ims. C la imants get paid for t ime off of work due to injury, and health care providers get paid for the treatments they provide to injured c la imants. Incorrect payments to either of these s takeholders are more likely to be not iced than other data incons is tenc ies b e c a u s e the bus iness operates to perform this funct ion, and b e c a u s e peop le not ice w h e n they have not rece ived adequate remunerat ion. B e c a u s e these s takeholders expect to be paid, they act as external quality control for payment data. Schuber t (1999) notes that the information gathered by health insurance providers tends to revolve around the f inancial information assoc ia ted with c la im durat ion and payment types. A c o n s e q u e n c e of the reasons driving col lect ion of these data is that a c c e s s to them tends to be largely unidirect ional, namely from the record down to the detai l . Trac ing backward f rom, or filtering by these detai ls may not a lways be poss ib le b e c a u s e s o m e fields may be incomplete, the da tabase may not be set up for it, relevant data may not be f ie lded, or there may s imply be too many data-entry errors. A s an examp le in the heal th-care paradigm, procedures that are repeated on a s ingle c la imant may be grouped together as if they occurred on the s a m e date. Th is s a v e s data-entry t ime for billing c lerks, but does not al low for t ime-ser ies ana lys is without manual ly re-coding (from paper c la imant files) the procedures under their correct dates. S i n c e this level of c la imant information may only be avai lable in unf ie lded form for s o m e records, an untrained layperson may not be ab le to perform this task. It is only recently that comput ing power and data storage have been sufficient to perform the des i red types of ana lyses . Data that had been col lected without this in mind are usual ly rife with incons is tenc ies that would not necessar i ly pose a prob lem on a case -by -c a s e bas is . However , these incons is tenc ies can c a u s e prob lems that make in ferences and general izat ions about injury types and treatment paths next to impossib le . A n y type of analys is to identify best pract ices or to f lag inappropriate serv ice can only work well if the data are accurate. M iss ing f ields may translate into b iased ana lys is if the reason for their a b s e n c e is systemat ic . S u c h analys is shou ld , and undoubtedly will be (appropriately) cr i t ic ized. 15 3.4 Approaches for coping with these problems Appropr iate ana lys is involves gaining a better understanding of the bus iness pract ices that generate the data. Improving data reliability involves educat ing all of the players to ensure that everyone concerned understands what is involved in consistent ly recording, stor ing, and retrieving the data required to ach ieve this end . Th is may include case load adjustments, quality control measu res , and accuracy incent ives. W h e n extremely detai led classi f icat ion structures are involved, it may a lso include a trade-off between granularity and accuracy of data. T h e i ssues that need to be add ressed have two distinct e lements : historical data need to be c leansed and made as usable as poss ib le , and management initiatives must co -ordinate ac ross departments to form a cohes ive and useful knowledge base for future ana lys is undertakings. Th is inc ludes re-thinking how and what data are co l lected, s tored, and ana lysed . Data col lect ion should be driven not only by administrat ive concerns , but a lso by the type of analys is that is des i red. Data need to be col lected with the purpose of ou tcomes ana lys is in mind, and potential ana lys is should be documented before building such a da tabase. Due to historical pract ices, archived data will a lways be less useful than data that have been col lected after implement ing improvements to data col lect ion and storage. If s takeholders take a proactive s tance on data quality before undertaking to perform ou tcomes ana lys is , researchers will be more likely to uncover useful f indings than if they cont inue to make do with the inaccurac ies that currently exist. S ince historical data problems can never be complete ly reso lved, the sooner the p rocess is ref ined, the sooner research goals can be ach ieved . Wi th appropr iate structures in p lace, not only will the explorat ion of ou tcomes , t racking, and pract ices be poss ib le , but a lso , the research generated will g ive t remendous visibility and credibil ity to organisat ions in terms of s takeholder interest. 16 4 Outcomes Research on Health Insurance Data and the WCB O u t c o m e s research on health data requires an ability to extract c la imant- level data on the bas is of injury/i l lness type, affected region, severi ty level , age , gender , industry, and other f ields, and to compare s o m e measurab le aspec ts of their ou tcomes on the bas is of their treatments or treatment s t reams. 4.1 Motivation for outcomes research on health insurance data W h e r e poss ib le , the ability to determine best treatment pract ices could al low for provider and treatment benchmark ing. Benchmark ing would enab le health insurance providers to set guidel ines for treatments that result in the best c la imant ou tcomes, and opt imize the quality of care. T h e s e guidel ines could then be communica ted to potential information users such a s medica l adv isors , and health care providers. Opt imal treatment patterns such as these, though initially cost ly, may actual ly reduce overal l expendi tures should these treatments reduce t ime off of work for injured c la imants. In s o m e c a s e s , individual health care providers may perform better than others when it c o m e s to the care of particular types of injuries. For example , it is poss ib le that a phys ic ian in an industrial town, where a certain type of work-related injury is endemic , has deve loped an excel lent course of treatment for this injury. Benchmark ing would al low researchers to identify such an individual, and to dist inguish the procedures that are being performed. Th is information could then be conveyed to health care practit ioners in other regions, which could serve to help all health care providers in the del ivery of current or lead ing-edge pract ices. A l though minimising c la im cost is important to workers ' compensa t ion agenc ies , the goa l may be to opt imize the quality of care. Whi le this may appear well in tended, quantifying "quality of care" is no s imple task: quality means different things to different people. F rom an employer 's point of v iew, it could mean getting the emp loyee back on the job as quickly as poss ib le , with little or no re lapse due to the initial injury. Th is could be the s a m e goal for the injured worker, al though the employee may p lace a stronger emphas i s on getting the most current (and somet imes most costly and most t ime-consuming) treatment. It could a lso m e a n a more complete return to normal function than would be general ly cons idered adequate by consult ing health care practit ioners or departmental policy. It could even m e a n patient sat isfact ion with the quality of care received. F rom the point of v iew of the health practit ioner, quality of care may m e a n using methods that he or she is most comfortable us ing. However quantif ication is ultimately def ined, quality of care can involve identifying the best treatments (based on s o m e combinat ion of ou tcome measures ) for injuries, and may involve identifying the practit ioners providing these treatments, and the establ ishment of a "gold s tandard" for which practit ioners can strive. A l though this could ultimately reduce total indemnity cos ts , there may be a trade-off between the treatment cost and wage replacement: if an injured worker receives the best poss ib le care , and if for certain types of injuries this general ly results in shorter recovery t imes, many workers will be back on the job sooner than if care w a s not opt imal. In such c a s e s , the cost of treatment may be higher, but offset by reduced wage compensa t ion , resulting in reduced total c la im costs . Converse ly , the best course of treatment may involve keeping patients with certain types of injuries off work for a longer period of t ime than has been done in 17 the past, which would incur higher initial c la im costs . However , if the injured worker is less likely to become d isabled from the s a m e injury at s o m e later date, or if the treatment method is less invasive, the total indemnity will still be reduced. 4.2 Steps to successful quality of care analysis Optimiz ing for quality of care depends upon first defining the parameters. In order to perform quality of care ana lys is on health data, we need the fol lowing: 1. A compar ison group 2. A definition of quality 3. Ou t comes measurements based on this definition of quality 4. A model to compare ou tcomes on the bas is of treatment path 5. Adequa te data upon which to build this model 4.2.1 A compar ison group Th is construct would initially need to be appl ied to a narrow s c o p e , al though it is poss ib le that resulting methodolog ies would be appl icable to other injuries with distinct treatment paths. A candidate for this prel iminary ana lys is is Anter ior Cruc ia te L igament ( A C L ) tears, more common ly known as knee tears. Treatment for this injury fol lows two distinct paths for the purpose of compar ison : surgical and non-surg ica l . 4.2.1.1 Description of treatment methodology for A CL tears In both c a s e s , physiotherapy is a lmost a lways in the treatment s t ream. A l though surgeons general ly repair these tears, there is s o m e quest ion as to when and for w h o m these repairs should be performed. Sever i ty and/or impact of the injury, gender , age , occupat ion , and other demograph ics may be mitigating factors, and need to be exp lored. Bes t health care pract ices might involve a trade-off between medica l cos ts , and indemnity (wage replacement) costs . 4.2.2 A definition of quality A definition of quality is one e lement that will be required to perform this ana lys is . In an article from the Institute for Cl in ica l Evaluat ive S c i e n c e s ( ICES) , "Heal th C a r e Del ivery in C a n a d a and T h e United States: A re there Re levant Di f ferences in Heal th C a r e O u t c o m e s ? " , Sz i ck et a l . (1988) quote a J A M A article by Donabed ian . Donabed ian categor ised quality of care in terms of the information needed for making a s s e s s m e n t s , and from which a s s e s s m e n t s can be drawn: structure, p rocess , and ou tcomes . Donabed ian 's premise is that there may be causa l relat ionships between structure and p rocess , and between p rocess and outcomes. Sz ick et a l . acknowledge that an objective and operat ional definition of "quality" is necessary . O n e such definition could be the trade-off between Total Tempora ry Disabil i ty (TTD) days and medica l costs , or even between T T D s and total c la im cost. Here , the problem is that the length of the disability and the pay rate earned by the worker both inf luence total c la im cost . A proxy might be built out of a s tandard ised income model s o that these two amounts could be added in a meaningful way , rather like a weighted s u m of T T D s and medica l costs . Mult ivariate techn iques, s u c h a s Pr incipal Componen t s Ana lys i s or Factor Ana lys is may prove useful here, perhaps 18 al lowing the determinat ion of a "cost-quali ty" index and a "life-quality" index that could be looked at separate ly or jointly. 4.2.3 Ou tcome metr ics based on this definition of quality S o m e measurements based on this definition of quality are required for model bui lding. Donabed ian (1988) notes that identifying the quality of care on the bas is of ou tcomes is severe ly confounded by di f ferences in the c a s e mix. Donabed ian warns : "Before assessment can begin we must decide how quality is to be defined" (1743) Separat ing out the cost from the quality aspec t of care is problemat ic - a l though quality of care is the fundamental goal of health care , it is often partially or complete ly def ined in terms of cost-ef fect iveness, largely because those paying for care are not the ones receiving it or providing it. Donabed ian (1988) identifies three major componen ts in this a s s e s s m e n t : 1. The serv ice provided must be needed , competent , cost-effect ive, t imely, consistent with current knowledge and presents a minimal risk to the patients 2. this serv ice must be provided to an individual or group that has the capaci ty to improve 3. a des i red outcome must be real ised. Bui lding on the above three components , Donabed ian provides the fol lowing as an operat ional definition for quality of care: "the degree to which health serv ices ... increase the l ikel ihood of des i red health ou tcomes that are consistent with current knowledge" (1748). 4.2.3.1 Outcome measurement Measur ing ou tcomes in t hemse l ves is riddled with prob lems, s ince there is somet imes an extended lag from the care ep isode to the time a c la im is cons idered comple ted , and intervening events may obscure this link. Typical ly, there is a lso no ment ion of the c la imant 's sat isfact ion with his or her treatment. Th is is likely to have an impact on how quickly and complete ly function (which W C B does not measure , either) is regained fol lowing the injury date. 4.2.4 A model to compare ou tcomes Ana lys is to determine the most appropriate course of treatment using administrat ive data is observat ional rather than exper imenta l . For this reason , t remendous care must be taken to control as much variability a s poss ib le within compar ison groups, espec ia l ly for such measu res as gender , d iagnos is , severity, age , recreat ional activi ty- level, and occupat ional duties that are likely to affect c la imant ou tcomes. Mult ivariate techn iques may prove useful here, but it is important to note that for univariate normal , we have two parameters: u and a. For bivariate normal , we have two us and two a s , plus a covar iance, s o 5 parameters. For p-dimensional data, we have p m e a n s , p var iances , , and p(p-1)/2 covar iances , for a total of (p 2 + 3p)/2 parameters. W e would need a s izeab le samp le for large p, s ince p = 10 g ives 65 parameters ! A n ideal compar ison group keeps equal all other relevant var iab les but the treatments, and s ince this can 19 ser iously reduce the s ize of the groups, it is imperative that these data f ields are populated as complete ly as poss ib le . It is a lso important to note, once aga in , the need for an unb iased dataset: in the c a s e where d iagnos is codes are only recorded accurate ly for the most severe injuries, our ana lys is will not be representat ive for all severi ty levels, wh ich limits general izabi l i ty. 4 .2.5 Adequa te data upon which to build this model Fol lowing the S a s k a t c h e w a n Governmen t Insurance (SGI) parad igm, measurements for such c la imant-centred data as "return to funct ion" and "sat isfact ion with treatment" are des i rab le , a long with measures of c la im reactivat ion, and reliable information on severi ty of injury (Schubert 1999). Measu remen ts like these would a l low us to track c h a n g e s over t ime, evaluate trends, give snapsho ts of the present, and indicate ext remes on many criteria. 4.3 Benefits of a reliable dataset Cla imant data, which are col lected through f inancial and account ing da tabase sys tems , do not a lways lend themse lves eas i ly to research appl icat ions. C r o s s - u s e of administrat ive data for analys is appl icat ions may be difficult b e c a u s e of comp le teness and accu racy problems that do not affect those entering and manag ing these data for their own purposes . Whi le administrat ive da tabases such as those at the W C B contain a weal th of data, on a c a s e - b y - c a s e bas is these data may be too incomplete to al low for interesting compar ison groups. B e c a u s e the comp le teness of a record may depend on the severi ty of the injury, b ias can interfere with data representat iveness, and , from an ana lys is point of v iew, this lack of representat iveness may somet imes be invisible without a thorough understanding of how and why the information has been col lected and m a n a g e d . B u s i n e s s p ressures and time constraints often minimize attention to data i ssues and pred ispose dec is ions . Unfortunately, if the dec is ion is made to proceed with ana lys is on data that are not representat ive, at best, the results may not be genera l izab le . At worst, the l ikelihood of making embar rass ing and costly mis takes inc reases . Ana lys i s of accurate data provides the ability to correctly identify t rends, to establ ish workable guidel ines, and , perhaps most signif icantly, to engender trust on the part of shareho lders and others relying on the ou tcomes of ana lys is . Th is trust needs to be built, and by educat ing and communicat ing with all those involved, the p rocess will lead to improved co-operat ion. Long-term resolution of this problem will involve co-operat ion and co-ordinat ion of sen ior management teams, data analysts , and researchers , and frontline data-gather ing personne l . W h e n this work is meaningfu l , involved part ies are more likely to accept the p rocess , and this in turn will serve to inc rease the utility of data gathered. 4.4 Dissemination of results In particular, the results of ou tcomes research based on factual ev idence will be credible to health care providers themse lves , wh ich will further serve to create interest. However , d isseminat ion of the gu ide l ines—through journal publ icat ion or product ion of educat ional mater ia ls—does not ensure that the guidel ines will be put into pract ice. T h e provider must s e e these materials as interesting, e a s y to interpret, and valuable. For example , upon the request of any health care provider in B C , the Brit ish Co lumb ia Med ica l 20 Assoc ia t ion sends out a "Pat terns of Pract ice Mini-Prof i le" report. W C B medica l adv isors prefer this format to the Prov ide rCompare report card , s ince it is more sel f -explanatory, and leans more heavi ly upon graphical presentat ion of data. It is a lso famil iar to B C providers. Borrowing from this format could lead to higher readership of circulated materials. In order to improve the accep tance of research f indings, a variety of implementat ion strategies (and perhaps s o m e targeted marketing) must co inc ide with d isseminat ion: The effects of printed educational materials compared with no active intervention appear, at best, small across studies and of uncertain clinical significance (Freemant le e r a / . 1999, 1). The additional impact of more active interventions produced mixed results: audit and feedback and conferences/workshops did not appear to produce substantial changes in practice, although the observed effects in the evaluations of educational outreach visits and opinion leaders were larger and likely to be of practical importance. However, the impact of these strategies cannot be assessed reliably from this review, as they are a small subset of the available estimates of their effectiveness. None of the studies included full economic analyses, and thus it is unclear to what extent the effects of any of the interventions may be worth the costs involved. (Freemant le et al. 1999, 2). Successful implementation requires all partners in the health care system, from patient to politician, to work together. Further research and evaluation of implementation strategies is required, as are sufficient resources to conduct the investigation. Only then can providers work towards co-ordinated, comprehensive implementation programs (Bas insk i 1993, 755). Determining an effective means of d isseminat ing and communicat ing information a m o n g health care practit ioners in such a way that it optimally received is an important considerat ion. However if the results are not val id, all other efforts will be moot (Freemant le 1999). 4.5 WCB Background T h e W C B (2004) is ded ica ted to the safety, protect ion, and health of workers . The i r goa l is the c la imant 's early, sa fe , and lasting return to work. W h e n a claimant has an accepted c la im with W C B , W C B will pay approved medica l expenses and wage- loss benefits, plus any necessa ry rehabil itation serv ices , to return the c la imant to a productive life. W C B a lso provides pens ions where there is a permanent disabil ity, and pays toward funeral costs in the event of death. W C B provides pens ion benefi ts to dependants of workers who have been killed on the job. In addit ion to the employer and the var ious health care practi t ioners the claimant s e e s in establ ishing a c la im, a claimant may c o m e in contact with as many a s nine different W C B departments for a s ingle c la im. Cla i re Lin (2000) performed a thorough study of the fol lowing e lements in her thesis, which examined W C B client c a s e managemen t a s it relates to human resource al locat ion: 21 • C la imant • Health care provider • P a p e r forms and records • Data entry, which may be done by more than one department, and for more than one purpose (for examp le , Health Serv i ces , account ing, adjudicators) • Mainf rame record • C leans ing algori thms T h e W C B account ing staff who issue wage- loss cheques and who pay health care practit ioners may or may not a lways be the s a m e individuals who are entering c la imant data into the var ious computer sys tems. A l though the players work as a team, e a c h of them has speci f ic needs for c laimant information. Th is may affect the overal l comp le teness of data col lected for the da tabase . 4.6 Administrative data challenges faced by the WCB Compl icat ing the usability of c laimant data are the facts that there are no consistent appl icat ions of injury descr ipt ions, and that bus iness pract ices can change during the lifetime of a claimant, in s o m e c a s e s leaving data f ields containing two or more different vers ions of procedure codes and/or fees . For example , in 1997, the Med ica l Se rv i ces P lan ( M S P ) of Brit ish Co lumb ia changed their fee-code format from three digits to five digits, which will be further d i scussed below. A l so , fees charged vary from year to year for s o m e , but not al l , p rocedures performed by heal th-care practit ioners. In s o m e c a s e s , this part icular prob lem c a n introduce b ias : in c a s e s where c la ims have more complete data than others, the underlying reason could be pol icy changes in data col lect ion within the lifetime of the c la im. T h e comp le teness could a lso be related to the severity and/or the duration of the injury, and this could go either way: s ince more severe injuries tend to cost more to the sys tem, more care may be taken in recording the information col lected from a claimant. Converse ly , with more procedures being performed, there may be more opportunity for errors to be made in the data col lect ion for f ields that are not directly related to bil l ing, s u c h as the format of the fee code or the date on which the procedure w a s admin is tered. 4.7 Discussion of current WCB analysis capabilities 4.1 A Internal ana lys is capabi l i t ies T o the W C B , a s in any bus iness , minimising costs is a worthwhile object ive s o long a s this does not compromise quality of care. To this end , current ana lys is being performed at W C B includes a model for forecast ing and f lagging for further invest igat ion, c la imants who are at risk for "convert ing" to long-term disabil ity status (Urbanovich 1999). A l s o , W C B is making attempts to examine the most effective treatments for such psycholog ica l work-related disabil i t ies as post-traumatic s t ress. T h e s e types of research often involve ad -hoc quer ies of exist ing data that are pul led together on an a s - n e e d e d bas is . Quer ies of this nature may be difficult and/or t ime-consuming to perform, and can require cons iderab le background knowledge of the var ious sys tems that hold different levels of client data. 22 4.7.2 External ly-developed purchased products In addit ion to these and other exist ing ana lyses , W C B tested products such as H N C ' s Prov iderCompare™, and Ver iComp™ for the purposes of provider benchmark ing , and fraud detect ion, respect ive ly . 6 For a number of reasons , Prov iderCompare™ fai led to ach ieve the des i red objective. 4.8 Discussion of Provider Compare™ at the WCB 4.8.1 P rov ide rCompare : provider benchmark ing on a min imum-cost criterion Wi th this product, health ca re providers are either ranked or compared on the bas is of median cost or median total t ime off of work of their c la imants. In s o m e c a s e s , these measu res may, through serendipi ty, be minimised s imul taneously , but in many c a s e s , there may be trade-offs between them. A product cal led Prov iderCompare™ should be able to provide adequate compar i sons among and between health care providers. T h e structure of this product depends on lowest cos ts and shortest t ime off work a s indicators of "best treatment". T h e reporting structure avai lable performs its provider benchmark ing ana lys is separate ly on the bas is of lowest cost or shortest length of disabil ity, but not jointly. Opt imiz ing quality of care is a worthy goa l , but it is doubtful that this can be identified one measure at a time. S imply returning an injured worker back to work in a timely fashion might not indicate that his or her heal th-care provider opt imized care. Implicit in any ana lys is are trustworthy data. T h e W C B is an insurance provider, and as s u c h , kept track of payments to c la imants and health care providers s ince its inception in 1917. There are mil l ions of c la imant records on file, but this d o e s not necessar i ly translate into usab le data at a detai led level. Th is def ic iency in administrat ive health data has been noted e lsewhere : Certainly much data exists, but often not in formats useful to statisticians. Just because data are recorded, it doesn't necessarily follow that researchers will be able to extract useful information from them (Reimer 1999, 5). S o m e of the data sou rces over lap, and many suffer f rom varying degrees of incomple teness . Altogether, in the initial f ive-year batch of data, there were about 11.6 million medica l detail t ransact ion records, 19,600 provider records, and 5.6 million c la ims r e c o r d s . 7 O f these, only 6.8 mill ion of the medica l detail records (59%), 16,000 provider records (82%), and 722,000 c la ims records (13%) had sufficiently complete information to be used with this product. Unfortunately, this meant that P rov ide rCompare would be able to ana lyze fewer than 1 3 % of all W C B c la imants ' records for the f ive-year per iod spann ing 1994 to 1998, which ra ised ser ious concerns about b ias. Data col lect ion is 0 HNC stands for 'The Hecht-Nielsen Neuro Computing Company'. Robert Hecht-Nielsen was the founder of the company. 7 For details, please see "Corporate Data Inventory - System Information", "Data Inventory -Table | File Information" and "Data Inventory - Column | Field Information". Documents prepared by Maureen Charron in the context of the Royal Commission into the Workers' Compensation Board in 1998. 23 much more thorough at present, but as in so many c a s e s , data col lect ion by agents and providers may have been less complete in the years before its use for other appl icat ions had been cons idered , and may have been more complete for more ser ious c la ims. In the United Sta tes , where P rov ide rCompare w a s deve loped , the motivation behind implement ing the product c o m e s from the Amer i can Health Ma in tenance Organisat ion (HMO) mode l . A n H M O is a group that contracts with medica l faci l i t ies, phys ic ians, employers , and somet imes , individual patients to provide medica l care to a group of individuals. Th is care is usual ly paid for by an employer at a f ixed per patient pr ice, and is general ly a for-profit corporat ion with responsibi l i t ies to its s tockholders to keep costs down and profits up. In an H M O , Pr imary C a r e Prov iders act as treatment ga tekeepers , and are accountab le for downst ream costs incurred from referred providers on their c la imants . 8 It is on the bas is of this construct that cost compar i sons are made between Prov iders when using P rov ide rCompare . T h e C a n a d i a n model for health care del ivery does not conduct bus iness under this parad igm. E v e n when phys ic ians give referrals to other health care providers in C a n a d a , the referring physic ian is not cons idered to be responsib le for the costs incurred from this referral. In retrofitting the construct of "primary care provider" for serv ices del ivered in C a n a d a , it is not a lways c lear which provider should be charged with the responsibi l i ty of directing such a patient's care. In the c a s e of acute injury, the attending physic ian at the emergency ward may del iver most of the health care required-for an injury, but the c la imant 's family doctor may be the one giving out referrals for rehabil itative care. T h e construct of "primary care practit ioner" does not fit for a C a n a d i a n appl icat ion, and al though Prov ide rCompare al lows the user to bypass this logic for s o m e reports, overr iding this default e l iminates a great dea l of the purpose this product w a s des igned to accomp l i sh . If it is of any utility at al l , P rov ide rCompare appears better sui ted to the H M O ' s cha l lenge in identifying high medica l or total indemnity costs . In the H M O mode l , the chief objective is cost reduct ion. W C B ' s interest in opt imizing care may involve a trade-off between higher medica l and/or total indemnity cos ts for more appropriate care . There may be no way to evaluate this potential trade-off by using off-the-shelf software on incomplete data. There is clearly a need for an ana lys is appl icat ion on the C a n a d i a n Worke rs ' Compensa t i on model to fill in this void, but there appears to be no exist ing product that will perform what is required. Primary Care Providers are the Doctors, Chiropractors, and Physiotherapists who are charged with managing the course of treatments for an individual patient's care. 24 5 Outcomes-Research Attempt On WCB Administrative Data S i n c e an off-the-shelf analys is product proved inadequate for the task of opt imizing the quality of care for the W C B , an internal attempt w a s made to extract c la imant data to try to identify optimal treatment s t reams for a se lec ted injury type. 5.1 Measurement and analysis The first step of any per formance ana lys is is measurement . Th is appl icat ion requires complete and accurate descr ipt ions of the particular injuries, the treatments used , and a breakdown of the costs involved. A l s o needed are objective measuremen ts of the ou tcomes from these interventions. 5.2 Data issues S o m e of these var iables exist, and al though there are ample data in the var ious sys tems at W C B , c la imant records are s e l d o m comple te enough or c lean enough to be usab le for the purpose of ou tcomes analys is . A long with other information, W C B claimant data contain the fol lowing information: Type of industry/job duties Reg ion S e x A g e Injury date C la im length Type of disabil i ty (permanent, temporary) W a g e rate Affected region T y p e of injury Type of treatment Prov iders P rocedu res F requency of treatment C o s t s per c la im Med ica l (procedure) cos t W a g e loss Unfortunately, on a record level , data f ields are often incomplete, and contaminated with inaccurate or spur ious data, making usage , and any conc lus ions drawn, very unrel iable. 9 S o m e claimant-sat isfact ion measu res are surveyed, but are not avai lable on a record level: a l though a voluntary survey of client sat isfact ion is taken of c la imants who are treated through W C B ' s "Work Condi t ioning Program" , this survey is confidential and anonymous , and so cannot be related back to the individual c la im record for compar ison or model building with its assoc ia ted administrat ive data. 9 For details, please see "Corporate Data Inventory - System Information", "Data Inventory -Table | File Information" and "Data Inventory - Column | Field Information". Documents prepared by Maureen Charron in the context of the Royal Commission into the Workers' Compensation Board in 1998. 25 5.2.1 Spec i f i c data i ssues P rob lems encountered with W C B claimant data included miss ing f ields, format changes , inconsistent c leans ing algor i thms, unfielded data, and mult i -purpose f ields that must be read in context to be unders tood. Health care providers record ICD9 (International Class i f icat ion of D i s e a s e , vers ion 9) codes when they submit their bil l ings to insurance providers for payment. T h e s e codes are required for fee payments to be p rocessed , or the provider does not receive payment for serv ices rendered. The difficulty with these codes is that there are s o many of them: two thick vo lumes are required to list them al l . Furthermore, providers have no incentive whatsoever to ensure that these codes are accura te or speci f ic , and s ince few practit ioners are motivated to look up speci f ic codes for f ree, they often enter one of the dozen or s o codes (often referred to as a "cheat sheet") , which are encountered the most frequently. For example , an A C L tear might get recorded as "general symptoms" (780.90) or "contusion of lower leg" (924.10), espec ia l ly if it w a s accompan ied by a spra ined foot. Th is problem is compounded when there are multiple injuries, in which c a s e somet imes only the most severe of these has its ICD9 code recorded. C h a n g e s in bus iness pract ices often contribute to data incons is tenc ies . For example , M S P Procedure F e e codes changed in 1997 from a 3-digit to a 5-digit format. Never the less , fees invoiced on paper to the W C B from heal th-care providers somet imes still use old codes . A l s o , pre-1997 cla imant records a l ready in the da tabase are still in the old format. Date sensit ivity built into the da tabase could reduce this problem: classi f icat ion s c h e m e s that integrate old and new coding s c h e m e s al low multiple- year ana lys is ac ross coding s c h e m e s . Th is , however, has not yet been done at the W C B . Prox ies for measu res that do not exist, such as impact, severity or return to function, can be der ived by using measures that do exist, such as occupat ion, medica l cost , and time off of work. However , b e c a u s e der ived f ields are only a s accura te a s the underly ing data, this may prove more daunting than it s e e m s . 5.3 Usable data Al though defining quality of care in quantif iable terms using exist ing measu res is one problem that must be add ressed (since any model or compar ison method can only be descr ibed in quantitative terms with quantitative data) f inding usable data is another (Schurman 1990; Donabed ian 1988). T h e initial goal of this research fai led b e c a u s e the data quality w a s insufficient for so granular a level of ana lys is . T h e goal w a s s imply too ambit ious. S ince the data integrity of certain f ields is somet imes quest ionable, a dec is ion w a s made to deve lop an ana lys is model on a c leaned and checked subset of the data. O n c e the model had been deve loped , it would then be appl ied to raw (uncleaned) administrat ive data s o as to compare the results. D iscuss ion of the two results would be an important tool in demonstrat ing why inaccurac ies in the data c a n be s o problematic. Th is could have accomp l i shed the fol lowing: the c lean data-subset might have y ie lded an important result, and the raw data would have either backed up or refuted the result using the c leaned subset . 26 If the raw data refuted the' subse t result, it would have demonst ra ted the w e a k n e s s of the exist ing data. Model l ing on the raw data could generate no or spur ious results, which may be refuted upon check ing with a c leaned data subset . However , if analys is on raw administrat ive data backed up the c leaned-subse t result, it would have demonstrated the robustness of the mode l . 5.3.1 Data c leans ing W C B quer ied their administrat ive data to produce a data extract for this attempt, but even the ad -hoc query produced a dataset that w a s riddled with prob lems. F e e - c o d e s (along with their assoc ia ted formats and fees charged) can change from year to year. T h e s a m e field that stores the fee code stores the number of days in hospi tal , but the data extract ignored the mult ipurpose nature of this f ield, and imposed a five-digit format to it. Th is a d d e d leading or trail ing ze ros to s o m e records , wh ich complete ly changed the interpretation of these fee codes . A l so , it w iped out post -1997 records that incorrectly conta ined the pre-1997 codes . Most signif icantly, neither the data extract nor da tabase itself updates the old codes into the new ones , which m a k e s quer ies on a s ingle procedure type imposs ib le without knowing all the codes that have been assoc ia ted with the procedure. Us ing a data extract re leased to U B C by W C B ' s Da ta Gua rd ians , 212 c la ims were c o d e d with the nearest ICD9 code for A C L tear, "(844.20) Cruc ia te L igament Knee S p r a i n " . 1 0 Only 12 were cand idates for ou tcomes ana lys is because : 1. A l though the procedure codes only changed from a 3-digit to a 5-digit format in 1997, the f ive-year (1994 - 1998, inclusive) da tabase query for our extract w a s set for a 5-digit p rocedure-code field. Th is extract is how Prov ide rCompare got the data as wel l . Fur thermore, the field used at W C B to record the procedure c o d e s is a mult i -purpose field that a lso records the number of days of hospital isat ion, if there were any. In the c a s e of one day of hospital isat ion, this field can appear to contain the procedure code "00100" , which is n o n s e n s e . 1 1 Th is field must be read in context f rom the mainf rame sys tem at W C B . T h e effect is that pre-1997 procedure codes contained in the extract are t runcated, miss ing , or contain spur ious ze ros . A n y c la ims from before 1997 must therefore be tr immed out of the samp le . S ince health care providers somet imes forgot that the codes had changed , s o m e of the post-1996 five-digit f ields are incorrectly coded with the old, pre-1997 three-digit codes . On ly 27 of the 212 "(844.20) Cruc ia te L igament K n e e Spra in " c la ims are from 1997 to the present date. 2. O f these 27, only 12 appeared as "c losed" , or inactive, c la ims. For ou tcomes research , c la ims must be c losed in order to investigate best pract ices, s ince open c la ims do not yet have all their assoc ia ted costs and ou tcomes. 1 0 Using the M S P Fee Guide, the closest ICD9 code for A C L tear is 844.20 "Cruciate Ligament Knee Sprain". 1 1 Procedure code 00100 is in reality a type of office visit: VISIT IN OFF ICE (AGE 0 - 74). 27 A data audit w a s attempted on these c la ims, which were doub le -checked with Heal th-Se rv i ces staff on the mainf rame sys tem. In s o m e c a s e s , four different s o u r c e s were checked for data validity and comp le teness : the P rov ide rCompare ' s da tabase in Irvine, the U B C extract f rom the data guard ians, the paper file, and the mainf rame. (The mainframe vers ion and the data guardian vers ion should have been identical, s ince they ostensib ly c o m e from the s a m e source , but the format change in 1997 precluded this from being the case . ) T h e data audit of these 12 c losed c la ims uncovered another problem: the paper and electronic f i les do contain information pertaining to the procedure codes , but not all of this information is f ie lded. A t this point, it is necessa ry to make the distinction between f ielded and unfielded data. Whi le all the information col lected about a c la imant should be in the paper or e lectronic fi le, not all of this information is f ie lded into code . For examp le , when a physic ian is asked to write a letter reviewing a c la imant 's health status, he or she is paid a set rate for the serv ice , which has its own fee code . T h e paper file will l ikely contain a copy of this letter, but may not contain a written record of the fee paid to the Phys ic ian for this serv ice , much less a fee code. T h e s e will s h o w up on the "List Paymen ts " sc reen of W C B ' s mainf rame sys tem, but there is no s ingle location where all the claimant information is f ie lded, and stored. Data audit ing of the paper and electronic fi les is extremely inefficient. Th is m a k e s verif ication of even a subset of data extremely labour- intensive. It requires t ime and resources from filing c lerks and from the emp loyees who handle, col lect and enter the data , s ince these are the peop le who are famil iar with the bus iness p r o c e s s e s that surround claimant information. Whi le exper ienced W C B personnel can read such prose-type information in context (on a c a s e - b y - c a s e basis) , unfielded data cannot be used in any pract icable way for ou tcomes research . 5.4 The effect of data contamination at the WCB Unfortunately, even this modif ied goal fai led. A l though c leaning and verifying a subset of data is necessa ry before performing compar isons of treatment s t reams, b e c a u s e the c leaned subset w a s s o tiny, this could not be performed with any statistical power. Contaminat ion of the data limits their applicabil i ty for research , but this prob lem is not restricted to W C B . In fact, the search for good examp les of publ ished ou tcomes research based on health care data is what exposed the magni tude and spread of such data i ssues . I C B C ' s own Recove ry authors have been writing a great dea l about the need to add ress this problem, Utah's Worke rs ' Compensa t i on A g e n c y has been act ively working on their data for severa l yea rs , and Saska t chewan Governmen t Insurance (SGI) has a l ready redes igned how, and what, information they collect. There is no shor tage of information about prob lems with the data: in the context of the Roya l C o m m i s s i o n into the Worke rs ' Compensa t i on Board in 1998, data prob lems were identified at W C B in the a n a l y s i s . 1 2 W C B has a Dec is ion Suppor t Se rv i ces program to store and m a n a g e col lected data, but in many c a s e s , b e c a u s e of the l imitations within these data, the appl icat ion of ana lys is mode ls will be fraught with hazards . Whether with 1 2 "Corporate Data Inventory - System Information", "Data Inventory - Table | File Information" and "Data Inventory - Column | Field Information". Documents prepared by Maureen Charron in the context of the Royal Commission into the Workers' Compensation Board in 1998. 28 or without software like P rov ide rCompare , attempting to ana lyse such data without considerat ion for these limitations will produce unrel iable results. C h i p m a n (1999) writes of the need for consistent identif ications. In our appl icat ion, reliability of the ICD9 field is unstable both in content, and in format: leading ze ros in s o m e but not all records changes the interpretation in these c a s e s . 1 3 Unfortunately, at the time of this writ ing, ICD9 codes are the only bas is for extracting a compar ison group of c la imants for any particular injury. It is unfortunate they are not recorded with sufficient accuracy or f requency to be used to extract a sufficient samp le of c la imants with a particular injury. It is possib le that those W C B ICD9 codes that are correct are a b iased subset consist ing of the more ser ious injuries, which would adverse ly affect the ability to extract genera l ized information from the study. Furthermore, there were s o few c la imants listed with the nearest code for A C L tear within the f ive-year data extract, that statistical power would have been very low. 1 3"latent syphilis"(92.4); contusion of lower leg" (924.10). 29 6 Discussion: Recommendations For Coping And Resolving Problems With Administrative Data If any meaningful ou tcomes ana lys is is to be done on administrat ive health data , these data must first be representat ive. E n h a n c e d communicat ion between and among depar tments is imperat ive if the f low of information is to run smoothly. Internal pol ic ies must be put into p lace to ensure ongoing communicat ion and accu racy with respect to data. The first step toward understanding how to so lve this problem is to ana lyse the bus iness p rocesses that generate the data of interest. Th is means examin ing the f low of information a s it g o e s from claimant to mainf rame, including what is lost and/or what is ga ined in terms of accuracy and comp le teness a long the way. Implicit in this is the need to a s s e s s and control the quality of da ta at e a c h s tage that information is t ransferred. A l s o implicit in this is the need to understand who is involved in this flow, and what the main s teps are between these e lements . 6.1 The future versus the past Look ing backward tells us where we have b e e n , and it is important if we are to know where we want to go. Certainly, internal c leans ing algor i thms can accomp l i sh a great dea l toward the end of garner ing usable data from what exists. However , this is not the ideal path to fol low for future data integrity. S i n c e the near past is far more interesting and useful than the distant past, the immediate focus needs to be on improving the management of data now so that high quality ou tcomes research can be performed within a few years on the most recent data . Improving the managemen t and content of older data is important, but this can be done anyt ime. T h e sooner current data col lect ion and management are opt imized, the less work will be involved in manag ing an integrated and accurate da tabase . A l though initially frustrating, the data problems and their impact on research capabi l i t ies provide an opportunity for learning during this difficult evolut ion. Improvements in the col lect ion and management of data are imperative if useful research is to be performed. If an organizat ion w ishes to predict the future based on the past, they must be prepared to invest cons iderab le time and effort in c leaning historical data. A m o n g the necessa ry tasks are the populat ion of miss ing f ields where poss ib le , apply ing date-sensi t ive lookup tables where appropriate, and normal iz ing tables. Ideally, this will involve an integrated approach , setting up a sys tem where everyone has a c c e s s to the s a m e information, and where all information is updated at once in a s ingle relational da tabase ac ross the organisat ion. If there were a desi re to accurate ly evaluate per formance from this point forward, the first task would be to accept that in order to do this, changes must occur in their data col lect ion methodology. W h a t it is that the organisat ion wou ld like to be ab le to study, ana lyse , and better understand should drive both the type of information to be gathered, and how it should be organ ised. 30 Looking to what others have done and are doing is critical at this step, both to s e e what has worked , and what has not. T h e literature review of this thesis provides an overv iew of such research . 6.1.1 C a s e l o a d Management -dr iven staff-performance metr ics often p lace greater weight on p rocess ing time than accuracy . Cler ica l staff who enter client data often feel cons iderab le pressure to adhere to t ime constraints, and may suffer repercuss ions when dead l ines are not met. Simi lar sanc t ions might not be meted out for lack of quality. Addi t ional i s sues that may contribute to ensu ing data quality i ssues include the stressful envi ronment these workers may work under. Addit ional staff may be necessa ry to ensure that case loads do not become unmanageab le . 6.1.2 Qual i ty control Data-entry quality control may be virtually non-existent for all but f inancial da ta . S i n c e cl ients and those who provide them with serv ices are highly motivated to ensure they are paid, the built-in error-checking character ist ic of f inancial data ensures that they are usual ly quite accurate. Qual i ty control techn iques could eas i ly be used on administrat ive data to al low for a reporting mechan i sm to ensure data quality. Th is mechan i sm would ideally be in a format that would be e a s y for both management and staff to use , such a s a graphic indicating target and historical data complet ion and accuracy , both overal l , and field-by-field 6.1.3 A c c u r a c y incent ives S i n c e the weight ing of per formance metr ics must shift to include accu racy as wel l a s s p e e d , per formance incent ives to staff could include rewards for both accu racy and s p e e d . Th is could greatly improve the ownership and buy-in of data-entry staff. 6.1.4 Class i f icat ion granularity Data-entry staff may not know why non-account ing information is co l lected, and may not understand how their work contr ibutes to research. Furthermore, these personnel may lack the necessa ry incentive to assu re the appl icat ion of detai led classi f icat ion sys tems . To save t ime, staff performing administrat ive data-entry may take short cuts. A n examp le of this is the w idespread use of "cheat shee ts " - abbreviated code tables for broadly def ined and frequently encountered c lassi f icat ions, wh ich may inc lude a "ca tch al l" code such as "Genera l Symptoms" or "Other". Even when such data are accurate, extremely detai led classi f icat ion s c h e m e s reduce the power of statistical analys is to such a degree that only enormous datasets can withstand the loss. Furthermore, this level of granularity may not be pract ical b e c a u s e it may simply cost too much to do. A less granular but eas ier - to-use classi f icat ion s c h e m e may be far more useful than a n extremely detai led one to wh ich few peop le adhere . A computer ized form with drop-down menu items may be a useful tool to encourage accu racy with this type of c lassi f icat ion. 31 6.1.5 Ded icated cod i ng staff In research organizat ions, data are stored as unfielded text, which is subsequent ly coded by expert, dedicated coders . W h e r e record- level data are miss ing , unclear, or inconsistent, the coder will contact the researcher for clarif ication before proceeding with data entry. S i n c e using administrat ive data for ou tcomes research adds this addit ional role to the organizat ion, this may be the best method of cod ing for these research p rocesses a s wel l . A l though this type of coding is s low, and the learning curve for new coding staff is long, there are many advan tages to using dedicated coding staff, such as finer c lassi f icat ion granularity, and improvements in both data accu racy and cons is tency. 32 7 Application (Proof Of Concept): Pulsed Signal Therapy (PST) 7.1 Problem In spite of all the problems inherent to administrat ive data, performing really useful ou tcomes research holds great promise, espec ia l ly a s data quality ca tches up with the quantity of data now avai lable in administrat ive sett ings. i In anticipation of improvements in administrat ive data col lect ion, one way of preparing for such research would be to perform ana lys is and model-bui ld on ou tcomes data that have been speci f ical ly col lected for the purpose of ou tcomes research . Hav ing deve loped a success fu l model and reporting format, these methods would then be appl ied to the improved administrat ive health data. Th is is an examp le of ou tcomes ana lys is methodology on data col lected for the purpose of ou tcomes research . 7.1.1 Pu l sed S igna l Therapy Pu l sed S igna l Therapy ( P S T ) is a relatively new treatment for chron ic pain. A l though used extensively in Europe , it is only now getting attention in North A m e r i c a , and al though it is c l assed as an alternative therapy in C a n a d a , is not yet approved in the United States. B e c a u s e P S T practi t ioners would like to bring this treatment into the mains t ream, ou tcomes research is paramount in defending the worth of this method of pain management . Dr. Cec i l Hershler , a V a n c o u v e r physic ian who operates a P S T clinic, required a cus tom da tabase appl icat ion to col lect and extract patient outcome data. Histor ical data existed on paper f i les for approximately 400 patients, and the appl icat ion needed to be built to house these and new patient records. 7.1.2 T h e treatment Pat ients receiving P S T are typically treated on 9 or 12 consecut ive days with one-hour appl icat ions of the Pu lsed S igna l , which is admin is tered by m e a n s of a coil apparatus into which the affected joint or region is p laced. This ser ies of appl icat ions is cons idered to be one complete treatment. Dur ing appl icat ion, treatment is undetectable. Del ivery of the treatment p roduces neither sound nor movement , and if not for the p resence of an indicator light, neither the patient nor the administrator would be aware that treatment w a s underway. In v iew of this, if an affected region or injury appears to be particularly respons ive to treatment, a doub le-blind test could easi ly be set up to test this hypothesis. 33 7.2 Methods 7.2.1 Spec i f i c bus iness requirements The P S T cl inic col lects patient demograph ics such as age and sex , as wel l as d iagnost ic and a s s e s s m e n t data. The cl inic a lso col lects s o m e non-treatment data, which are used for market ing the product, s u c h a s how the cl ient heard of P S T . At issue is the fact that each patient can be treated for more than one affected region, and the s a m e affected region may have more than one injury or injury type, and more than one complete course of treatment. There is a lso recidiv ism without re-injury, which can require further treatment in the form of "boosters" . 7.2.2 P lann ing the da tabase After a thorough examinat ion of both the bus iness requirements and the p rocess of care assoc ia ted with the P S T clinic, the da tabase w a s des igned to avoid s o m e of the problems we have s e e n in administrat ive health da tabases . In particular: Reco rd level data must be complete and accurate. Th is may not be poss ib le for the historical (paper) records, but new records will be entered electronical ly, al lowing for enforced f ields where des i red (for examp le , sex , injury type, affected region), and error-check ing for such things a s imposs ib le birth dates and injury dates. D iagnos is must be recorded accurately. T h e compromise is that there will be fewer and less-spec i f ic d iagnos is cho ices , but this will encourage greater accu racy within this f ield. T h e cho ices will be limited to a list in a drop-down menu. T h e region(s) of the body that is(are) affected must a lso be recorded accurately, with a simi lar compromise made by limiting the number of cho ices to those contained in a drop-down list. E a c h treatment will be recorded on the date it was admin is tered, and not consol idated under a c o m m o n date. It must be poss ib le to extract patient data by region affected, injury type, sex , weight, or any other potential predictor of ou tcome that may be of interest. B e c a u s e patients may need subsequen t treatments for the s a m e or for different affected regions, it must be possib le to dist inguish between courses of treatment, and between affected regions. Th i s is particularly important for establ ish ing i ndependence amongs t observat ions, where only the first course of treatment on a new patient may be cons idered . 7.2.3 Da tabase deve lopment A da tabase appl icat ion w a s written in Microsoft A c c e s s 2000 , and is documented below (p lease s e e append ices) . T h e P S T da tabase a c c o m m o d a t e s the cl ient's needs for ou tcomes research in the fol lowing w a y s : • Data required for ana lys is are f ie lded. • Affected regions are limited to a list of cho ices that do not have to be looked up in a manua l . 34 • Broad c l a s s e s of d iagnost ics are used to streaml ine injury types, and are a lso limited to a list of cho ices . • T h e structure a l lows for more than one affected region being affected at different t imes for a s ingle patient. • The structure al lows for one affected region having more than one complete injury a s s e s s m e n t , d iagnos is , and course of treatment. • Data entry staff thoroughly understand that these data are speci f ical ly required for research , and are g iven sufficient t ime to properly enter patient data. T h e cl inic a l ready t racks the fol lowing ou tcome measurements before, immediately after, s ix w e e k s , six months, and one year fol lowing treatment: • Pa in Intensity • Pa in F requency • Restr ict ion of Movemen t • Swel l ing • Warming • Discolour ing • Pares thes ia • Medicat ion C h a n g e T h e s e measu res are col lected using a five-point sca le , with 0 = none or never, 1 = slight or se ldom, 2 = moderate or somet imes , 3 = severe or often, and 4 = ext reme or a lways. Addit ional ly, the Activi t ies of Dai ly Living measu res (hours s leep ing , minutes s tanding, wa lk ing , sitting, and driving) a re col lected before treatment, but have not historically been appl ied to the fol low-up evaluat ions. Upon suggest ion to the client, it w a s dec ided that that these measu res may be used in fol low-up evaluat ions for future patients. A n est imate of Sever i ty may a lso be a s s e s s e d in the future, and has level cho i ces of "Low", "Med ium" , and "High" . 7.2.4 P S T R e s e a r c h Object ive: Of the ou tcome measures currently t racked, the client feels that the most objective are "Pa in Intensity", "Pa in Frequency" , and "Restr ict ion of Movement" . Th is may account for the fact that of the ou tcomes measures historically t racked, these are the most consistent ly populated fields on paper historical records. Track ing these changes over t ime will be the first step. Th is will be done by looking at these measu res individually, and a lso by looking at their d i f ferences from the pre-treatment sco res at the four post-treatment evaluat ions: immediately fol lowing treatment, six w e e k s , six months, and one year post-treatment. Th is will be broken down to further granularity by looking at these f ields and their d i f ferences c rossed by d iagnost ic c lass (Osteoarthrit ic vs . Soft T i s s u e Injury); gender ; Body M a s s Index; and duration of the injury (this can be approx imated by the date of injury subtracted from the date of first treatment). 35 7.3 Results All analysis is performed using S P S S for Windows (10.0.7). 7.3.1 Descriptive analysis of PST client data PST clients may come in for more than one treatment on more than one area. To maintain independence between cases, for these clients, only the first course of treatment is considered. Subsequent treatments of the same or different affected regions have not been considered in these results. There are 599 patients, 221 (37%) male and 378 (63%) female. Of the 599 eligible patient records available, 491 can clearly be categorised as either Osteoarthritis (OA) or Soft Tissue Injuries (STI). The average PST patient is 52.6 years of age, and the average age of injuries seen in the clinic is 7.8 years. About half of all patients are at a healthy weight for their height according to National Heart, Lung and Blood Institute (US) BMI groupings. 1 4 Very few are underweight. Sex and injury type are associated (chi-square=14.831, d.f.=2, p<0.001). In describing injury type by sex, males are more likely to present with O A injuries, and females are more likely to present with soft tissue injuries (or combination injuries). In fact, 56% of the, males seen at the PST clinic presented with O A injuries, while half (50%) of the females came in for treatment of soft tissue injuries, and a further 11% came in for treatment of combination STI-OA injuries. Table 1 - Cross Tabulation - Diagnostic Class by Sex O A STI Both A L L Male 107 75 10 192 Female 136 173 37 346 A L L 243 248 47 538 CHI-SQUARE = 14.831 WITH D.F. = 2, p<0.001 We see that male patients are more likely to be overweight than females: Table 2 - Cross Tabulation - BMI by Sex Underweight Normal Overweight Obese Total Male 5 86 93 22 206 Female 22 204 87 32 345 Total 27 290 180 54 551 CHI-SQUARE = 27.451 W TH D.F. = 3, p<0.001 1 4 B o d y M a s s Index (BMI) is an indicat ion of the appropr ia teness of o n e s weight to one ' s height. It is compu ted f rom the formula BMI = 0 .45359237*we igh t (in pounds)/ [height ( inches)* .0254] 2 36 Arthritis research has estab l ished a link between obesi ty and osteoarthr i t is . 1 5 Th is link is suppor ted by our data: more than half (53%) of those with osteoarthrit is have a higher-than-average BMI . In contrast, only about a third (34%) of the cl inic 's STI cl ients have a BMI in the "overweight" or "obese" category: Table 3 - Cross Tabulation - BMI by Diagnostic Class Underweight Normal Overweight Obese Total O A 5 99 87 32 223 STI 16 135 65 12 228 OA/STI 1 22 15 5 43 Total 22 256 167 49 494 C H I - S Q U A R E = 24.250 W I T H D.F. = 6, p<0.001 There are approximately as many patients under age 50 a s there are 50 and over. Not surprisingly, ST Is are more common amongst younger patients, whi le O A is more c o m m o n amongst o lder patients. However , this dif ference is even more pronounced when looking at sex : amongst o lder patients, about % of both ma les and fema les c o m e to the cl inic with O A injuries, whi le younger fema les are far more likely to arrive at the cl inic due to STI pain than O A pain: 8 3 % of younger female cl ients have STIs , vs . 5 9 % of younger ma les . Table 4 - Cross Tabulation - Diagnostic Class by Age and Sex O A or STI but not both O A STI Total Male Under 50 34 49 83 Female Under 50 27 132 159 Male 50 and over 70 25 95 Female 50 and over 108 41 149 Total 239 247 486 CHI -SQUARE = 123.4 WITH D.F. = 3, p< 1001 T h e most frequently s e e n injuries occur in the neck (C-Sp ine) , lumbar (L-Spine) , knee, or hip. O f the 491 records that c a n be clear ly determined to be either O A or STI , 399 occur in these affected regions. Note that in the C - S p i n e there are more STI records than O A records, possib ly due to whip lash injuries. T o summar ize , O A injuries s e e n in the P S T clinic are general ly well es tab l i shed, and are more likely to be s e e n in o lder patients, ma les , or those who are overweight, while ST Is are general ly more recent, and are more likely to be s e e n in younger patients, fema les , or normal-weight patients. 1 5 Being overweight is a risk factor for developing OA: there is a clear link between obesity and the development of OA of the knee in women (Anderson and Felson 1988; Felson and Anderson 1988; Felson and Chaisson 1997). 37 7.3.2 Definition of Pa in S c a l e Pat ients arriving at the P S T clinic for treatment are asked to rate their Pa in Intensity, P a i n Sever i ty, and Restr ict ion of Movemen t on a s c a l e of 0 to 4 , where 0 = "none", 1 = "slight", 2 = "moderate", 3 = "severe" , and 4 = "extreme". T h e s e sco res are measured before, right after, six w e e k s , s ix months, and one year post-treatment. In e a c h c a s e , patients are asked to provide this pain rating at its worst. 7.3.3 Prel iminary analys is W e s e e below the distribution of these sco res at base l ine, and s ix -weeks post-treatment: Table 5 - Baseline Scores Baseline Scores Initial Pain Intensity Initial Pain Frequency Initial Restriction of Movement 0 3 4 37 1 11 7 23 2 55 65 90 3 264 200 150 4 256 313 273 Grand Total 589 589 573 Table 6 - Scores at 6 Weeks Post-Treatment Scores at 6 weeks post-treatment Initial Pain Intensity Initial Pain Frequency Initial Restriction of Movement 0 19 23 59 1 90 50 56 2 188 139 169 3 169 189 142 4 95 160 122 Grand Total 561 561 548 A n interesting observat ion is the relat ionship between initial "Pa in Intensity" sco res and the groupings " S e x " and "STI /OA" : females arrive at the cl inic with higher initial pain sco res than males , and STI patients c o m e in with higher pain sco res than O A patients. W e have s e e n that STI patients tend to be younger than O A patients. STI patients a lso tend to have high initial pain sco res . Converse ly , O A patients tend to be older, and typically arrive at the cl inic with a lower pain score that they have lived with for longer. With typically higher initial pain sco res , fema les exempl i fy the initial pain sco re dif ference between O A and STI patients, while amongs t ma les the dif ference is not statistically signif icant. O A injuries tend to be older injuries: the m e a n age of an O A injury is 9.4 years o ld, whereas the m e a n age of STIs is about 5.4 years (t = 4.9, p <0.001). Th is may not be 38 surprising, considering that initial Pain Intensity scores for STIs tend to be higher - more unbearable pain may drive patients to get help sooner. Also, STIs are often acute injuries, while OA is degenerative. (Note that it is more difficult to determine a clear start date for osteoarthritis than it is for soft tissue injuries.) Also, and perhaps not surprisingly, patients with O A injuries are significantly older than patients with STIs: the mean age for O A patients is 61, for STIs: 43. (Please refer to Appendix 6 - pivot tables for overall changes over time, then broken down by sex and injury type.) One year post-treatment, 67% of patients improved by at least one point on the Pain Intensity scale. This approximately replicates results seen experimentally. 1 6 Only 2% experience degradation, and 31% experience no change in their condition. When looking at a stronger improvement of at least two points, about 37% show improved condition, and the remainder show a less than two-point improvement. Table 7 - Proportion with At Least One Unit of Improvement at One Year Post Treatment Count P a i i ^ itensity One Y e a r After Grand Pa in Intensity Before • 0 1 2 3 4 Total 0 2 1 3 1 4 4 2 8 10 15 1 34 3 14 35 50 47 5 151 4 12 26 31 41 37 147 Grand Total 36 76 96 89 42 339 % Better 67.0%| % No Change 31.0%) % Worse 2.1%| The post-treatment improvement patterns of males and females are different. For Pain Intensity and Pain Frequency, females are more likely to show improvement than males, a difference that becomes more pronounced through to six months post-treatment. It is interesting to note that STIs improve less dramatically than do O A injuries initially, but this difference diminishes after six weeks post-treatment. This may be a reflection of the higher initial pain scores found in STIs presenting for pain treatment. When looking at males only, we see that initial Pain Intensity scores for STIs and OAs are fairly similar, but that OAs respond better to treatment from the start: by the end of the initial treatment period, almost half of males with O A injuries show improvement, compared with less than a third of males with STIs showing improved condition. This 69.5% of treated patients and 45.1% of placebo patients showed clinically meaningful improvement, with a chi squared p-value of 0.003 (Trock, Bollet, and Markoll 1994, 1910). 39 difference persists through to one year post-treatment, with a lmost 6 0 % of the STIs and a lmost 7 0 % of the O A injuries showing improvement. For fema les , there appears to be a tendency for STIs to present as signif icantly more painful than O A , but aga in , O A injuries improve more dramatical ly and quickly. W e s e e the s a m e initial jump immediately fol lowing treatment: 5 0 % of female O A patients s e e immediate improvements as compared with only 3 5 % of female STI patients. Th is di f ference d imin ishes by the sixth week after treatment. Both settle down to slightly fewer than 7 0 % improving by one year post-treatment. A l though STIs tend to be fairly recent, and are typically s e e n in younger patients, these injuries do not appear to be as likely to respond quickly to P S T a s does O A pain. 7.3.4 Logist ic model l ing G iven that this is not an exper iment, but rather, an observat ional study, predictive mode l -building is fraught with di f ferences in the c a s e mix. Furthermore, there is sel f -select ion bias b e c a u s e these data are based on cl ients who elected to c o m e in for P S T of their own volition and pay for this procedure out of pocket. For these reasons , mode ls built from such administrat ive data may be treated as the bas is for des igned exper iments, not as predictive mode ls in their own right. By using exist ing data in this way, we can narrow the s c o p e of our search when des ign ing exper iments for the purpose of predict ive model-bui ld ing, and compare administrat ive post -hoc results to those d iscovered experimental ly. T h e logistic mode ls will use d ichotomized post-treatment di f ferences in observed Pa in Intensity as ou tcomes, and age group, A g e of Injury Group , body m a s s index group, sex , and injury type as predictors. Note that patients who arrive at the cl inic with very high initial " P a i n " and "Restr ict ion of Movement " sco res may be different from the rest, and s o this var iable will be control led for in all predict ive mode ls cons idered . Injury-related covar iates for this model include injury type ( O A or STI), and the age of the injury. Outcome measures • T h e 12 outcome measures of improvement in Pa in Intensity, Pa in Frequency , and Restr ict ion of Movement have been d ichotomized a s "Not Improved" or "Improved by at least one unit" as compared with basel ine measu res for each of the four t ime per iods: " immediately after", "six w e e k s after", "six months after", and "one year after" treatment. Th is d ichotomy has been c o d e d a s 0 for "no improvement" and 1 for "at least one unit of improvement" for ana lys is by logist ic regress ion. Predictors • S e x is Ma le (coded as 0) or Fema le (coded as 1) • A g e has been c lassi f ied into two groups: o under 50 years of age (coded as 0) o 50 + (coded a s 1) • Body M a s s Index (BMI) has been c lassi f ied into two groups: o Not overweight (< 24.99) (coded as 0) o Overweight (25 +) (coded as 1) 40 • Injury type: o Osteoarthri t is (OA) (coded a s 0) o Soft T i ssue Injury (STI) (coded as 1) • A g e of injury has been classi f ied into two groups: o L e s s than 4 years old (coded as 0) o 4 + (coded a s 1). 7.3.5 T h e mode ls H o s m e r and L e m e s h o w (2000, page 95) recommend that "Any var iable w h o s e univariable test has a p-value < 0.25 is a candidate for the mult ivariable mode l , a long with all var iab les of known cl inical importance." T h e twelve ou tcomes var iables will be model led as fol lows: Univariate analysis Improvement in Sco re = univariable predictors Controlling for initial score Improvement in Sco re = signif icant (p<0.25) univariable predictor (aside from initial score) run individually whi le controll ing for initial score Full main-effects model Improvement in S c o r e = Final s tep using backward s tepwise regress ion on the s u m of signif icant var iab les from the univariable mode ls tested Full interaction model Improvement in S c o r e = Signif icant var iables from the full no-interaction mode l , p lus two-way interactions, tested against main-effects model Final interaction model Improvement in Sco re = final s tep of backward s tepwise el imination on the full interaction mode l , entry method to capture all avai lable data Final model T h e most reasonab le of the main effects and interactions mode ls Al l var iables are entered using backward s tepwise (p-in: 0.10; p-out: 0.25) to produce the full and reduced main effects mode ls . The signif icant var iables from the final step of backward s tepwise are then re-entered using entry method to produce the final mode ls . Th is is done to produce mode ls that are based on more of the avai lable da ta , wh ich may produce a model with a better fit and with smal ler conf idence intervals for the regress ion coeff icients. The categor ical contrasts are sex , age group, body m a s s index group, injury type, and A g e of Injury Group . P l e a s e s e e Append ix 5, sect ion 8.5.1.4 for categor ical var iab les codings. A summary of model-bui ld ing candidates is shown on the next page. Detai ls of these mode ls are presented in append ices . D iscuss ion of the mode ls will fol low this chart: 41 Table 8 - Summary of candidates for model building: Significant main effects and interactions (White=Candidates for model building; Shaded=Results; P-values in parentheses) P-values given in parentheses Improvement in score (from initial baseline) Right After Six Weeks Six Months One Year Pain Intensity (PI) Univariable models Sex (0.24) Age (<0.01) Injury Type (<0.01) Initial PI (<0.01) Sex (=0.05) Injury Type (=0.08) Initial PI (<0.01) Sex (<0.01) Age (=0.12) BMI (=0.16) Age of Injury (=0.03) Initial PI (<0.01) Age (=0.15) Initial PI (<0.01) Pain Intensity (PI) Main Effects (with Initial PI) Age (<0.01) Injury Type <0.01) Sex (=0.22) Injury Type (=0.02) Sex (=0.06) BMI (=0.08) Age of Injury (<0.01) Age (=0.22) Pain Intensity (PI) Final Model And Possible Interactions Injury Type + Initial Pain Intensity Injury Type + Initial Pain Intensity Age of Injury + Initial Pain Intensity (Noteworthy interaction model with BMI and Initial Pain Intensity) Initial Pain Intensity Pain Frequency (PF) Univariable models Age (<0.01) BMI (=0.20) Injury Type (<0.01) Initial PF (=0.03) Sex (=0.03) Initial P F (<0.01) Sex (=0.05) Age of Injury (=0.17) Initial PF (<0.01) Age (=0.08) Injury Type (=0.15) Age of Injury (=0.09) Initial PF (=0.15) Pain Frequency (PF) Main Effects (with Initial PF) Age (<0.01) BMI (=0.17) Injury Type (<0.01) Sex (=0.05) Sex (=0.08) Age of Injury (=0.13) Initial P F (<0.01) Age (=0.11) Injury Type (=0.24) Age of Injury (=0.07) Pain Frequency (PF) Final Model And Possible Interactions Injury Type + Initial Pain Frequency Sex + Initial Pain Frequency Initial Pain Frequency Age of Injury* Initial Pain Frequency + Age of Injury x Initial Pain Frequency Restriction of Movement (ROM) Univariable models Age (=0.18) Injury Type (=0.03) Age of Injury (=0.12) Initial R O M (O.01) BMI (=0.22) Initial ROM(<0.01) Age (=0.23) Age of Injury (=0.13) Initial R O M (<0.01) Sex (=0.17) Age (=0.03) Initial R O M (<0.01) Restriction of Movement (ROM) Main Effects (with Initial ROM) Age (=0.16) Injury Type (=0.02) Age of Injury (0.11) Initial ROM(<0.01) Age (=0.24) Age of Injury =(0.04) Sex (=0.24) Age (=0.03) Initial R O M (<0.01) Restriction of Movement (ROM) Final Model And Possible Interactions Injury Type + Age of Injury + Initial Restriction of Movement BMI + Initial ROM + BMI x Initial ROM Age of Injury + Initial Restriction of Movement Age + Initial Restriction of Movement 7.3.6 Examination of candidates for model building Across all three outcome measures, and modelling at each of the four post-treatment intervals, the initial score appears as a significant predictor. Because of this, the initial score will be included as a control in all predictive models tested on the categorical contrasts Sex, Age Group, Injury Type, Body Mass Index Group, and Age of Injury Group. 42 T h e s e remaining contrasts are run in logistic models against the ou tcome var iab les to determine their eligibility as predictors. T h o s e with p-values of 0.25 or under are cons idered as candidates for further test ing. For the purpose of logistic regress ion ana lys is , we are treating the ordinal measu re of initial score as an interval var iable. B e c a u s e of this, and b e c a u s e it is such a strong predictor of ou tcome through every time period when measuremen ts are taken (right after, s ix w e e k s , six months, and one year post-treatment), it will be used as a control in all predict ive mode ls that we cons ider . In all c a s e s , mode ls may be read directly f rom the tables provided. Improvement in Pa in Intensity S c o r e right after treatment will be used as a detai led example . Al l further mode ls may be interpreted in the s a m e way from the tables. (P lease refer to Append i ces for tables.) 7.3.7 Evaluat ion of mode ls 7.3.7.1 Right After Treatment Improvement in Pain Intensity Score Univariate Analysis With predictors run individually, the candidate var iab les are A g e Group , Injury T y p e , Initial Pa in Intensity, and perhaps S e x . Sex There is weak ev idence (p=0.24) that fema les may be more likely to improve right after treatment than males : O . R . (odds ratio) = 1.23 (95% C l : 0.87 to 1.73). W e wiil s e e this factor emerge in s igni f icance in the next t ime interval. Age Group There is strong ev idence that A g e Group is a predictor of ou tcome (p<0.01). T h o s e aged 50 or more are 1.89 t imes more likely than those under 50 (95% C l : 1.35 to 2.65) to have improved their Pa in Intensity sco re by at least one unit right after treatment. Injury Type Having a soft t issue injury, rather than O A , drops the probability of a post treatment improvement by about half: O . R . = 0.53 (95% C l : 0.37 to 0.77, p<0.01), indicating that those with STIs are about half as likely to exper ience this initial post treatment improvement as those with O A . Initial Pain Intensity There is good ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.62 t imes (95% C l : 1.28 to 2.06) right after treatment with every one point increase in their initial Pa in Intensity score . Controlling for Initial Pain Intensity W h e n controll ing for Initial Pa in Intensity, S e x is not an important predictor of ou tcome (p=0.55), but A g e Group and Injury Type remain promis ing. 4 3 There is strong ev idence that A g e Group is a predictor of ou tcome: those aged 50 or more are 2.18 t imes more likely than those under 50 (95% CI : 1.54 to 3.09, p<0.01) to have improved their Pa in Intensity score by at least one unit right after treatment. For the model based on Injury type controll ing for Initial P a i n Intensity, having a soft t issue injury rather than O A drops the probability of a post treatment improvement by about half: O .R . = 0.48 (95% CI: 0.33 to 0.70, p<0.01), aga in indicating that those with STIs are about half a s likely to exper ience this initial post t reatment improvement a s people with O A . Note that A g e Group and Injury Type are strongly assoc ia ted - o lder patients are more likely to have osteoarthrit is, whi le ST Is are more c o m m o n amongs t younger patients. B e c a u s e both Injury T y p e and A g e G r o u p were reasonab le predictors w h e n control l ing for Initial Pa in Intensity, a variety of "full" models were tested using these predictors. W h e n entered together using backward s tepwise, all but A g e G r o u p and Initial Pa in Intensity drop off of the reduced mode l . N o signif icant improvement is noted with the addit ion of interaction terms, al though it may be reasonab le to cons ider interactions between Injury T y p e and Initial P a i n Intensity, and Injury T y p e and A g e Group . Th is possibi l i ty w a s examined in detail . Wh i le there appeared from the observed data to be a poss ib le interaction (younger O A cl ients and older STI cl ients appear to respond differently depend ing on their Initial Pa in Intensity score) , this relat ionship may be spur ious - s ince most o lder patients have osteoarthrit is and most younger patients have STI , there are proport ional ly fewer patients in these "atypical" groups, and -2 log l ikelihood is only 612 .2 with 4 d.f.: Injury Type & Age Group (Observed Data) 2 3 Initial Pain Intensity • O A 50 and Over •STI 50 and Over • O A Undergo •STI Under Injury Type & Age Group (Main Effects Model) - O A 50 and Over •STI 50 and Over Injury Type & Age Group Interaction 1 (Main+lnjType*IPI) Figure 1 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Injury Type and Age Group Observed data; Main Effects fitted model; Interaction fitted model Data at the lower end of the Initial Pa in Intensity s c o r e s are s p a r s e (p lease s e e Append ix 10.5.2.2), mak ing prediction of outcome at these lower levels problematic. However , mathemat ical ly, the best predictive model for these data is the addit ive model 44 of A g e Group control l ing for Initial Pa in Intensity - it has the higher -2 log l ikelihood of the mode ls controll ing for Initial Pa in Intensity (782.5, d.f. = 3), smal ler s tandard error on the predictor (in this c a s e , A g e Group) , and visual ly, ignoring the smal l deviat ion from linearity seen in the observed data for those aged under 50 with initial P a i n Intensity = 2, the main effects model appears to follow the observed data reasonab ly wel l (no benefit is obtained with the inc lusion of an interaction term): 0.6 i Age Group (Observed Data) 50 and over Under age 50 1 2 3 Initial Pain Intensity 0.6 -| Age Group O (Main Effects Model) / 0.5 0.4 0.3 -0.2 C 0.11 o 50 and over •M^—Under age 50 0.0 1 2 3 4 Initial Pain Intensity 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Age Group (Interaction Model) -50 and over •Underage 50 Initia? Pain Intensity Figure 2 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Age Group Observed data; Main Effects fitted model; Interaction fitted model 0.6 -, Injury Type (Observed Data) 0.5 -0.4 -0.3 -0.2 -• STI 0.1-o.o| 1 2 3 I . U I . I r > : ~ i 4 4 0.1 \ 0.0 Injury Type (Main Effects Model) 2 3 Initial Pain Intensity Figure 3 - Initial Pain Intensity vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model 45 This , however, is problematic: it m a k e s no s e n s e logical ly that o lder people would be more likely to exper ience an improvement (presumably due to s o m e aspec t of heal ing) than younger people, who general ly are in better health and hence more likely to respond to medica l and other interventions. B e c a u s e age and injury type are so strongly related, in spite of the fact that A g e Group provided the better mathemat ical mode l , at this point, the only reasonab le model for considerat ion is the addit ive model based on Injury Type when controll ing for Initial Pa in Intensity - those with soft t issue injuries appear to be more likely to exper ience an improvement in Pa in Intensity immediate ly after treatment (no benefit to the predictive ability of the model is obtained by including the interaction term). Th is f inding is consistent with publ ished results on exper imenta l P S T data. In summary , right after treatment, when controll ing for Initial Pa in Intensity, patients with STIs are about half as likely to exper ience post treatment improvement a s people with O A : O . R . = 0.48 (95% C l : 0.33 to 0.70, p<0.01). Improvement in Pain Frequency Score Univariate Analysis With predictors run individually, the candidate var iab les are A g e Group , Injury Type , Initial Pa in Frequency , and perhaps BMI group. Age Group There is strong ev idence that A g e Group is a predictor of ou tcome (p<0.01). T h o s e aged 50 or more are 1.81 t imes more likely than those under 50 (95% C l : 1.27 to 2.57) to have improved their Pa in F requency score by at least one unit right after treatment. BMI Group There is weak ev idence that BMI Group is a predictor of ou tcome (p=0.20). Overweight individuals are 1.27 t imes more likely than those with a BMI under 25 (95% C l : 0.88 to 1.82) to have improved their Pa in F requency score by at least one unit right after treatment. Injury Type Having a soft t issue injury, rather than O A , drops the probabil ity of a post treatment improvement by about half: O . R . = 0.47 (95% C l : 0.32 to 0.69, p<0.01), indicating that those with STIs are about half as likely to exper ience this initial post treatment improvement as people with O A . Initial Pain Frequency There is good ev idence (p=0.03) subjects will exper ience improvement inc reases of 1.30 t imes (95% C l : 1.03 to 1.64) right after treatment with every one point inc rease in their initial pain f requency score . Controlling for Initial Pain Frequency Having descr ibed in detail the results for improvement in Pa in Intensity above , subsequen t results can be read in the s a m e way f rom the tables provided in the append ices . W h e n controll ing for Initial Pa in Frequency , the predictors that retain s igni f icance include A g e Group (p<0.01), Injury Type (p<0.01), and , aga in to s o m e extent, BMI Group (p=0.17). Aga in , o lder people are more likely than younger people to note an improvement, overweight individuals are more likely to improve than normal or 46 underweight patients, and those with O A are more likely to respond than those with STIs . B e c a u s e the 9 5 % conf idence interval for A g e Group conta ins 1 (95% CI: 0.89 to 2.16, p=0.14), its predictive power is quite limited. W h e n consider ing combinat ions of these predictors in a full mode l , the best mathemat ica l fit is the model based on Injury Type controll ing for Initial Pa in Frequency. Injury Type (Observed Data) 2 3 Initial Pain Frequency 0.6 0.5 A Injury Type (Main Effects Model) 0.0 2 3 Initial Pain Frequency Figure 4 - Initial Pain Frequency vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model Al though the model based on A g e Group , Injury Type and Initial P a i n F requency holds s o m e predictive power, the -2 log l ikelihood drops while the deg rees of f reedom increase. Aga in , nothing is ga ined with the addition of interaction terms. Improvement in Restriction of Movement Score Univariate Analysis With predictors run individually, the candidate var iables are Injury T y p e , Initial Restr ict ion of Movement , and to a lesser extent, A g e of Injury Group and A g e Group . Age Group There is weak ev idence that A g e Group is a predictor of ou tcome (p=0.18). T h o s e a g e d 50 or more are 1.29 t imes more likely than those under 50 (95% CI : 0.89 to 1.85) to have improved their Restr ict ion of Movement score by at least one unit right after treatment. Injury Type In the univariable mode l , having a soft t issue injury rather than O A drops the probabil ity of a post treatment improvement by about a third: O . R . = 0.64 (95% CI : 0.43 to 0.95, p=0.03), indicating that those with STIs are less likely to exper ience this initial post treatment improvement as people with O A . 47 Age of Injury Group There is s o m e ev idence that A g e of Injury Group is a predictor of ou tcome (p=0.12). T h o s e with injuries that are at least 4 years old are less likely than those with more recent injuries to s e e an improvement in outcome: O . R . = 0.75 (95% CI : 0.52 to 1.08) to have improved their Restr ict ion of Movement score by at least one unit right after treatment. Initial Restriction of Movement There is strong ev idence (p<0.01) subjects will exper ience improvement inc reases of 1.56 t imes (95% CI: 1.29 to 1.88) right after treatment with every one point inc rease in their initial Restr ict ion of Movement score . Controlling for Initial Restriction of Movement W h e n controll ing for Initial Restr ict ion of Movement , the predictors that retain s igni f icance include A g e Group (p=0.16), Injury Type (p=0.02), and A g e of Injury G r o u p (p=0.11). Aga in , o lder people are more likely than younger people to note an improvement, those with O A are more likely to respond than those with ST Is , and older injuries are less likely to improve than more recent injuries. It is important to note here that not only is it the c a s e that o lder peop le are more likely than younger people to have O A , but that O A injuries tend to be o lder injuries than STIs in patients treated at the P S T clinic. Th is is interesting, b e c a u s e al though O A s e e m s more likely to respond than STI and older patients s e e m to respond better than younger patients (likely due to the strong correlation we have s e e n above) , w e s e e a higher l ikel ihood of recovery in more recent injuries than we s e e in o lder injuries. A g e of injury (Observed Data) 0.1 0.0 -4 + years •< 4 years 1 2 3 4 Initial Restr ict ion of Movement A g e of Injury (Main Effects Model ) 1 2 3 4] Initial Restr ict ion of Movemen t Figure 5 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Age of Injury Observed data; Main Effects fitted model 48 Injury Type (Observed Data) 0.0 1 2 3 4 Initial Restriction of Movement Injury Type (Main Effects Model) 0.0 1 2 3 4! Initial Restriction of Movement Figure 6 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Injury Type Observed data; Main Effects fitted model Injury Type & Injury Age Group (Observed Data) 0 1 2 3 Initial Restriction of Movement - O A 4 + years •ST I 4 + years • O A <4 •ST I <4 year year Injury Type & Injury Age Group (Main Effects Model) 1 2 3 Initial Restriction of Movement - O A 4 + years •ST I 4 + years • O A <4 •ST I <4 yeai year Figure 7 - Initial Restriction of Movement vs. Probability of improvement Right After Treatment by Injury Type and Age Group Observed data; Main Effects fitted model Al though the model b a s e d on Injury Type , A g e of Injury G r o u p and Initial Restr ict ion of Movemen t is not as strong mathematical ly as the ones based on Injury T y p e and Initial Restr ict ion of Movemen t or A g e Group and Initial Restr ict ion of Movemen t , it would appear from the data that older O A injuries do not respond a s well as more recently 49 d iagnosed O A , and that whi le ST Is in genera l do not respond as well as O A , more recent STIs respond more quickly than do older STIs . Furthermore, controll ing for A g e of Injury Group reduces the standard error in predicting the effect of Injury Type. In cons ider ing a predict ive mode l f rom exper imenta l data ca re shou ld be taken to control for this factor, and it will be left in the final model for this purpose. Note that there is no reason to suspec t interaction between these terms - the model including interaction does not include injury type except a s an interaction term, and has a lower -2 log l ikel ihood than the main effects mode l . 7.3.7.2 Six Weeks After Treatment I m p r o v e m e n t i n P a i n In tens i ty S c o r e Univariate Analysis With predictors run individually, the candidate var iables are S e x , Injury Type , and Initial Pa in Intensity. Sex There is good ev idence (p=0.05) that f ema les may be more likely to improve s ix w e e k s after treatment than males : O . R . = 1.42 (95% C l : 1.00 to 2.00). Injury Type Having a STI , rather than O A , drops the probability of a post-treatment improvement: O . R . = 0.72 (95% C l : 0.50 to 1.04, p=0.08). O n c e aga in , we s e e that O A injuries are more likely to improve after treatment with P S T than are STIs . Initial Pain Intensity There is good ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.77 t imes (95% C l : 1.40 to 2.25) six w e e k s after treatment with every one point increase in their initial Pa in Intensity score . Controlling for Initial Pain Intensity W h e n control l ing for Initial P a i n Intensity, S e x loses m u c h of its s ign i f icance a s a predictor of ou tcome (p=0.22), but Injury Type remains promising (p=0.02). For the model based on Injury Type controll ing for Initial Pa in Intensity, having a STI rather than O A drops the probabil ity of a post-treatment improvement by about two thirds: O . R . = 0.64 (95% C l : 0.44 to 0.94, p=0.02), aga in indicating that those with ST Is are less likely to exper ience this initial post treatment improvement a s people with O A . Note that controll ing for Initial Pa in Intensity actual ly improves the predictive strength of this var iable. Injury Type is a reasonab le predictor when controll ing for Initial Pa in Intensity, s o it w a s tested in a variety of models a long with S e x , even though S e x did not appea r to have much predictive strength when combined with Initial Pa in Intensity. 50 Figure 8 - Initial Pain Intensity vs. Probability of improvement Six Weeks After Treatment by Injury Type Observed data; Main Effects fitted model W h e n entered together us ing backward stepwise, all but Injury T y p e and Initial Pa in Intensity drop off of the reduced model . No signif icant improvement is noted with the addit ion of interaction terms. T h e best fit for a predict ive model for improvement in Pa in Intensity at s ix w e e k s post treatment is Injury Type controll ing for Initial Pa in Intensity. Improvement in Pain Frequency Score Univariate Analysis With predictors run individually, the candidate var iab les are S e x and Initial Pa in Frequency . Sex There is strong ev idence that S e x is a predictor of ou tcome (p=0.03). F e m a l e s are 1.49 t imes more likely than ma les (95% CI: 1.05 to 2.12) to have improved their P a i n F requency score by at least one unit, six w e e k s after treatment. Initial Pain Frequency There is strong ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.68 t imes (95% CI : 1.33 to 2.14) s ix w e e k s after treatment with every one point increase in their initial Pa in F requency score . 51 0.6 -i Sex (Observed Data) 0.6 -i Sex (Main Effects Model) 0.0 o.i ^ 0.2 A 0.3 A 0.4 A 0.5 A 0.0 2 3 4 2 3 4 Initial Pain Frequency Initial Pain Frequency Figure 9 - Initial Pain Frequency vs. Probability of improvement Six Weeks After Treatment by Sex Observed data; Main Effects fitted model Controlling for Initial Pain Frequency W h e n controll ing for Initial Pa in Frequency , S e x remains a strong predictor of ou tcome. In fact, it is the only model with any predictive ability at this t ime point. Improvement in Restriction of Movement Score Univariate Analysis With predictors run individually, the only candidate var iab les are Initial Restr ict ion of Movemen t and , to a lesser extent, BMI Group . There is weak ev idence that BMI Group is a predictor of ou tcome (p=0.22). Overweight patients are 1.25 t imes more likely than non-overweight patients (95% C l : 0.88 to 1.78) to have improved their Restr ict ion of Movement score by at least one unit s ix w e e k s after treatment. Initial Restriction of Movement There is strong ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.85 t imes (95% C l : 1.54 to 2.21) six weeks after treatment with every one point increase in their initial Pa in Intensity score . Controlling for Initial Restriction of Movement W h e n controll ing for Initial Restr ict ion of Movement in the addit ive mode l , the predictor BMI group loses its s igni f icance (p=0.62), al though interestingly, there appea rs to be an interaction between BMI group and Initial Restr ict ion of Movement . BMI Group 52 At the lower initial Restr ic t ion of Movemen t sco res , overweight pat ients appea r more likely to note an improvement than non-overweight patients, whi le at the higher end of the R O M sca le , these g roups do not appear different in the observed da ta . O n e could specu la te that at the lower Initial Restr ict ion of Movemen t sco res , patients who are not overweight may be more act ive, and hence , find even a low-level of Restr ict ion of Movement intrusive. A l though the -2 log l ikelihood is slightly lower for this mode l , and the standard error for BMI group is very large in this parad igm, the mode l nonethe less has s o m e visual appea l . 0.7 - BMI (Interaction Model) 0.6 -0.5 -0.4 -0.3 -0.2 -o.-r - O — Overweight (BMI 25 +) • Not overweight (BMI < 25) u.u -1 2 3 4 Initial Restriction of Movement Figure 10 - Initial Restriction of Movement vs. Probability of improvement Six Weeks After Treatment by BMI Observed data; Interaction fitted model 7.3.7.3 Six Months After Treatment Improvement in Pain Intensity Score Univariate Analysis With predictors run individually, the candidate var iab les are S e x , A g e of Injury Group , and Initial Pa in Intensity, and perhaps A g e Group and to a lesser extent, BMI Group . Sex There is good ev idence (p<0.01) that fema les may be more likely to improve S i x Mon ths after treatment than ma les : O . R . = 1.72 (95% CI: 1.16 to 2.55). Age Group Older individuals may be less likely to s e e an improvement at s ix months: O . R . = 0.74 (95% CI: 0.50 to 1.09, p=0.12). Importantly, from this point forward we s e e older people responding less wel l than younger people - a complete reversal of trend as compared with earl ier evaluat ion points. 53 BMI Group A l s o of interest, at this t ime we s e e that heavier individuals may not be a s likely to improve as normal or underweight patients: O .R . = 0.75 (95% C l : 0.50 to 1.12, p=0.16). Th is result, whi le not particularly strong, nonethe less m a k e s better s e n s e than what we s a w in earl ier t ime f rames. Age of Injury Group Older injuries may be less likely to respond to P S T at six months than more recent injuries: O . R . = 0.65 (95% C l : 0.44 to 0.96, p=0.03). Th is f inding appea rs to be fairly consistent throughout all t ime f rames. Initial Pain Intensity There is good ev idence (p<0.01) that subjects will exper ience improvement inc reases of 2.02 t imes (95% C l : 1.53 to 2.66) s ix months after treatment with every one point increase in their initial Pa in Intensity score . Controlling for Initial Pain Intensity W h e n controll ing for Initial Pa in Intensity, the var iables S e x , BMI G r o u p , and A g e of Injury Group have predict ive power on treatment ou tcome at s ix months. For the mode l based on S e x when controll ing for Initial P a i n Intensity, f ema les may be more likely to note a post-treatment improvement in Pa in Intensity: O . R . = 1.47 (95% C l : 0.98 to 2.22, p=0.06). Figure 11 - Initial Pain Intensity vs. Probability of improvement Six Months After Treatment by BMI Observed data; Interaction fitted model BMI group a lso has s o m e predict ive power when control l ing for Initial P a i n Intensity. Here , we s e e that heav ier individuals may not respond a s well as non-overweight 54 patients: O . R . = 0.69 (95% CI: 0.45 to 1.04, p=0.08). Interestingly, there once again appears to be an interaction between BMI group and initial Pa in Intensity: Overweight patients at low initial P a i n Intensity respond better than non-overweight patients, but non-overweight patients respond better than overweight patients at high initial Pa in Intensity scores . Figure 12 - Initial Pain Intensity vs. Probability of improvement Six Months After Treatment by Age of Injury Observed data; Main Effects fitted model A g e of Injury G r o u p may be a reasonab le predictor w h e n control l ing for Initial P a i n Intensity. Injuries o lder than 4 years are less likely to s h o w an improvement at this t ime point: O . R . = 0.56 (95% CI : 0.37 to 85, p<0.01). Runn ing these predictors together in full mode ls with and without interaction terms d o e s not appear to improve the predictive power of the mode l . For this t ime f rame, the interaction model b a s e d on A g e of Injury Group and Initial P a i n Intensity may be the best fit. Improvement in Pain Frequency Score Univariate Analysis With predictors run individually, the candidate var iables are S e x , Initial P a i n F requency , and , perhaps , A g e of Injury Group . Sex There is signif icant ev idence that S e x is a predictor of ou tcome (p=0.05). F e m a l e patients are 1.49 t imes a s likely to have improved their Pa in F requency sco re by at least one unit s ix months after treatment (95% CI: 1.00 to 2.20), echo ing results we have s e e n before. 55 Age of Injury Group There is weak ev idence that when run a lone a s a mode l , A g e of Injury Group is a predictor of ou tcome (p=0.17). Pat ients with older injuries are only 0.77 t imes as likely to have improved their Pa in F requency sco re by at least one unit s ix months after treatment (95% CI: 0.52 to 1.12). Initial Pain Frequency There is strong ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.53 t imes (95% CI : 1.18 to 1.98) six months after treatment with every one point increase in their initial Pa in F requency score . Controlling for Initial Pain Frequency 1 2 3 Initial Pain Frequency Figure 13 - Initial Pain Frequency vs. Probability of improvement Six Months After Treatment by Sex Observed data; Main Effects fitted model W h e n controll ing for Initial Pa in Frequency , S e x remains a modest predictor of ou tcome (p=0.07), with fema les o n c e aga in more likely to s e e an improvement s ix months after treatment (95% CI : 0.97 to 2.14). Nothing is ga ined with the addit ion of an interaction term. A g e of Injury Group improves its predict ive ability somewha t when control l ing for Initial Pa in Frequency . A l though it still appears that older injuries respond less wel l than more recent injuries, this relat ionship is tenuous (95% CI : 0.50 to 1.07, p=0.11), and aga in , nothing is ga ined with the addit ion of an interaction term. 56 0.6 -0.5 0.4 0.3 0.2 -0.1 -0.0 Figure 14 - Initial Pain Frequency vs. Probability of improvement Six Months After Treatment Observed data; Main Effects fitted model Full mode ls built f rom combinat ions of var iables fail to improve substant ial ly upon the predict ive ability of Initial Pa in F requency a lone, al though the addit ive model of S e x with Initial Pa in F requency may be worth cons ider ing. Note the dearth of observat ions in the lowest Initial Pa in F requency score . Improvement in Restriction of Movement Score Univariate Analysis With predictors run individually, the candidate var iab les are Initial Restr ict ion of Movement and , to a lesser extent, A g e of Injury Group and , perhaps, A g e Group . Age Group There is weak ev idence that A g e Group is a predictor of ou tcome (p=0.23). O lder patients are only 0.79 t imes as likely to have improved their Restr ict ion of Movemen t sco re by at least one unit, s ix months after treatment (95% C l : 0.54 to 1.16). Age of Injury Group There is modest ev idence that when run a lone as a mode l , A g e of Injury Group is a predictor of outcome (p=0.13). Pat ients with older injuries are only 0.74 t imes as likely to have improved their Restr ict ion of Movement score by at least one unit, s ix months after treatment (95% C l : 0.50 to 1.09). Initial Restriction of Movement There is strong ev idence (p<0.01) that subjects will exper ience improvement inc reases of 2.09 t imes (95% C l : 1.71 to 2.55) six months after treatment with every one point increase in their Initial Restr ict ion of Movement score . 57 Controlling for Initial Restriction of Movement Age of injury (Observed Data) 1 2 3 4 Initial Restriction of Movement Age of Injury (Main Effects Model InjAge + I ROM) •4 + years •< 4 years 1 2 3 4 Initial Restriction of Movement Figure 15 - Initial Restriction of Movement vs. Probability of improvement Six Months After Treatment by Age of Injury Observed data; Main Effects fitted model Control l ing for Initial Restr ict ion of Movement actually improves the s ign i f icance of A g e of Injury Group as a predictor (p=0.04): O . R . = 0.65 (95% C l : 0.42 to 0.99). A g e Group , however, b e c o m e s slightly less signif icant than when cons idered a lone. The re does not appear to be any signif icant interactions with Initial Restr ict ion of Movemen t for either of these predictors. In fact, testing var ious full mode ls with both A g e Group and A g e of Injury Group (with and without interactions), the only mode l worth cons ider ing for predict ive model ing is A g e of Injury Group when controll ing for Initiation Restr ict ion of Movement . 7.3.7.4 One Year After Treatment Improvement in Pain Intensity Score Univariate Analysis With predictors run individually, the candidate var iables are Initial P a i n Intensity, and , perhaps , A g e Group . Age Group There is weak ev idence (p=0.15) that when compared with younger pat ients, those over age 50 are less likely to improve at one year post-treatment: O . R . = 0.71 (95% C l : 0.45 to 1.13). Initial Pain Intensity There is good ev idence (p<0.01) that subjects will exper ience improvement inc reases of 1.89 t imes (95% C l : 1.38 to 2.58) right after treatment with every o n e point inc rease in their initial Pa in Intensity sco re . 58 Controlling for Initial Pain Intensity When controlling for Initial Pain Intensity, Age Group loses some of its already weak predictive ability (p=0.22). The addition of an interaction term does nothing to improve this relationship. Age Group (Observed Data) 2 3 Initial Pain Intensity Age Group (Main Effects Model) -50 and over "Under age 50 2 3 Initial Pain Intensity Age Group (Interaction Model) 50 and over Under age 50 2 3 Initial Pain Intensity Figure 16 - Initial Pain Intensity vs. Probability of improvement One Year After Treatment by Age Observed data; Interaction fitted model; Main Effects fitted model In fact, the only strong predictor of improvement in Pain Intensity one year post-treatment is the Initial Pain Intensity score: 0.8 Inititial Pain Intensity (Observed Data) 2 3 Initial Pain Intensity Inititial Pain Intensity (Best fit for full and interaction models) • A l l Patients o.o 2 3 Initial Pain Intensity Figure 17 - Initial Pain Intensity vs. Probability of improvement One Year After Treatment Observed data; Main Effects fitted model 5 9 Improvement in Pain Frequency Score Univariate Analysis With predictors run individually, a l though none of these is a strong predictor, the candidate var iables are A g e Group , Injury Type , A g e of Injury Group , and Initial Pa in Frequency . Age Group There is s o m e ev idence (p=0.08) that when compared with younger patients, those over age 50 are less likely to improve at one year post-treatment: O . R . = 0.68 (95% CI: 0.44 to 1.05). Injury Type There is weak ev idence that Injury Type is a predictor of ou tcome (p=0.15). T h o s e with STIs are 1.43 t imes more likely than those with O A (95% CI: 0.88 to 2.34) to have improved their Pa in F requency sco re by at least one unit, one year after treatment. Note that this is much different f rom the results we s a w earl ier at s ix months post-treatment. Note a lso that younger patients are more likely to improve at this evaluat ion point (this trend having a lso turned around at six months). Reca l l that Injury Type and A g e are strongly assoc ia ted - this result is not surpr is ing. Age of Injury Group There is s o m e ev idence that A g e of Injury Group is a predictor of outcome (p=0.09). C o m p a r e d with those with recent injuries, those with older injuries are 0.69 t imes less likely to have improved their Pa in F requency sco re by at least one unit, one year after treatment (95% CI: 0.44 to 1.06). Initial Pain Frequency There is weak ev idence (p=0.15) that subjects will exper ience improvement inc reases of 1.23 t imes (95% CI: 0.93 to 1.64) one year after treatment with every one point increase in their initial Pa in F requency score . Controlling for Initial Pain Frequency W h e n controll ing, for Initial Pa in Frequency , Injury Type loses a great deal of predictive ability. A g e Group a lso loses s o m e predictive ability, a l though we s e e an interesting pattern in the observed data: at lower Initial Pa in F requency S c o r e s , older patients are more likely to improve than younger patients, a trend which reverses for the higher Initial Pa in F requency sco res . 60 0.8 -, Age of injury (Observed Data) 0.0 2 3 Initial Pain Frequency 0.8 Age of Injury (Main Effects Model) 0.0 2 3 Initial Pain Frequency 0.8 Age of Injury (Interaction Model) 0.0 " 4 + years -< 4 years 2 3 Initial Pain Frequency Figure 18 - Initial Pain Frequency vs. Probability of improvement One Year After Treatment by Age of Injury Observed data; Interaction fitted model; Main Effects fitted model Age of Injury Group enjoys an improvement in predictive ability with the addition of Initial Pain Frequency, and here again, there seems to be an interaction: those with older injuries at the lower Initial Pain Frequency scores are more likely to note an improvement one year post-treatment than those with higher Initial Pain Frequency scores, and the opposite appears to be the case with more recent injuries. Although the interaction model has a lower -2 log likelihood, the graph of the model appears to follow the observed data reasonably well, and the p-values for the coefficients are low (although the standard errors are high). Univariate Analysis With predictors run individually, although none of these is a strong predictor, the candidate variables are Age Group, Initial Restriction of Movement, and, perhaps, Sex. Sex There is only weak evidence that Sex is a predictor of outcome (p=0.17). Females are .72 times more likely than males (95% CI: 0.45 to 1.15) to have improved their Restriction of Movement score by at least one unit one year after treatment Age Group There is strong evidence that Age Group is a predictor of outcome (p=0.03). Patients age 50 and over are 0.61 times less likely to have improved their Restriction of Movement score by at least one unit one year after treatment as younger patients (95% CI: 0.39 to 0.96). Initial Restriction of Movement There is strong evidence (p<0.01) that subjects will experience improvement increases of likelihood that subjects will experience improvement one year after treatment doubles 61 (95% CI : 1.61 to 2.49) right after treatment with every one point i nc rease in their initial Restr ict ion of Movemen t score . Controlling for Initial Restriction of Movement W h e n controll ing for Initial Restr ict ion of Movement , S e x loses predict ive ability, but A g e Group d o e s not. Furthermore, the paucity of data in the low Initial Restr ict ion of Movemen t sco res m a k e s the deviat ion f rom what we s e e in the mode l based on A g e Group and Initial Restr ic t ion of Movemen t less worr isome. Nothing appea rs to be ga ined with the addit ion of an interaction term, and the addit ive mode l based on A g e Group and Initial Restr ict ion of Movemen t is the only candidate for a reasonab le predict ive model . 0.9 Age Group (Observed Data) 1 2 3 Initial Restriction of Movement 0.9 Age Group (Main Effects Model) 0.0 50 and over Under age 50 1 2 3 . Initial Restrict ion of Movement 0.9 Age Group (Interaction Model) 0.0 50 and over Under age 50 1 2 3 Initial Restr ict ion of Movement Figure 19 - Initial Restriction of Movement vs. Probability of improvement One Year After Treatment by Age Observed data; Interaction fitted model; Main Effects fitted model 7.3.7.5 Summary of results over time O u t c o m e s shortly after treatment ( immediately after and s ix w e e k s after) appear to fol low a somewha t different pattern a s compared with ou t comes at s ix months and at one year. Whi le the initial sco res are consistent ly strong predictors ac ross all measu res of outcome, d i f ferences between groups appear over t ime. Immediately after treatment, al though it appears that older people are more likely to s e e an improvement than younger patients, age group and injury type are strongly assoc ia ted . It is c lear that patients suffering from osteoarthrit is and having a high initial score are the most likely to s e e an improvement. At six w e e k s , this is still the c a s e for Pa in Intensity, a l though we s e e s o m e indication that w o m e n with high initial P a i n F requency are more likely than men to note an improvement, and that Restr ict ion of Movemen t may be more likely to improve for overweight individuals than for non-overweight individuals, at least for lower levels of initial Restr ic t ion of Movement . It is important to note here that s e x and injury type are strongly l inked - osteoarthri t is is s e e n more often in female patients than in male patients so S e x may be con founded with Injury Type . Th is is a lso the c a s e for BMI: osteoarthrit is is s e e n more often in overweight patients than in non-overweight patients, and the interaction between BMI and Initial 62 Restr ict ion of Movement , whi le interesting, may be spur ious. Th is is important, b e c a u s e there is no logical reason why overweight individuals are more likely to improve than non-overweight individuals, as ide from the fact that those who are overweight may be less agi le to begin with, and s o a smal ler improvement may be more not iceable than it would be to a very act ive person. The story changes considerably at the s ix-month mark: for Pa in Intensity and Restr ict ion of Movement , the strongest mode ls at s ix months are based on Initial S c o r e with A g e of Injury Group . For these mode ls , those with more recent injuries are more likely to note an improvement. Th is is interesting b e c a u s e for the first t ime it is the more recent injuries that maintain a lasting improvement from basel ine, despi te the fact that osteoarthrit is (which tends to be an older injury) general ly responds better than soft t issue injuries. For Pa in F requency , it is only the initial score that remains a strong predictor of ou tcome at the six month mark. At one year, we s e e this response pattern for initial Pa in Intensity - it is only the initial sco re that is the strong predictor of ou tcome at this t ime interval. For Pa in F requency , we s e e not only a main effect based on the age of injury, but an apparent interaction between the age of the injury and the initial Pa in F requency score . A l though data at the low end of initial sco res are spa rse , at the higher end , older injuries are less likely to s e e an improvement than more recent injuries. Final ly, Restr ict ion of Movemen t at one year post-treatment shows us that patients in the older age group are less likely to s e e n an improvement from basel ine Restr ict ion of Movement . In earl ier t ime intervals, the type of injury confounded this effect. It is only at one year post treatment that we s e e that older people are less likely to maintain the improvement in Restr ict ion of Movemen t than younger peop le , regard less of injury. S o , in the short run, osteoarthrit is responds better than STI , and in the long run, younger people with more recent injuries are more likely to retain symptomat ic relief. 63 8 Discussion There were many concess ions made in preparing the P S T data for ana lys is . Noteworthy w a s the sca le used for measur ing the outcome sco res - whi le they were treated as cont inuous ratio data for the purpose of logist ic regress ion , there are, in fact, p rob lems in doing so . For example , the dif ference between an initial pain sco re of 1 or 2 may not be the s a m e as the difference between an initial pain score of 3 or 4. Another problem is the paucity of data in the low end - relatively few patients present with low initial pain sco res , s o these results may be quite b iased toward patients who suffer more signif icant pain. T h e s e patients may be quite rel ieved to s e e an improvement of only one unit on this sca le , whe reas patients w h o are not in s u c h seve re dist ress my notice less of an impact from a one unit drop. S p a r s e data in the lower-end of initial sco res may be particularly problemat ic for Restr ict ion of Movement - a younger patient suffering from an acute-onset soft t issue injury may be less sat isf ied with a one unit improvement than would an older patient suffering from a degenerat ive injury such a s osteoarthrit is, which c o m e s on slowly and gradual ly impairs movement . B e c a u s e this treatment is not covered by health insurance, patients may only try this treatment either when they can afford it (which is less likely for younger patients) or when severe pain dr ives them to seek alternative treatments. T h e s e are observat ional , not exper imental data. Furthermore, Dr. Hersh ler has conf i rmed that most of his cl ients feel that their condit ion has not improved for an extended period before seek ing P S T . There may be more suitable measu res to monitor for improvement in condit ion. T h e three main ou tcomes measu res were c h o s e n primarily b e c a u s e Dr. Hersh ler sugges ted that they were the most rel iable measu res within his da tabase , and because they would yield results that would be comparab le to exist ing research on P S T . A s data col lect ion improves in large-sca le administrat ive da tabases , the methodology outl ined above may b e c o m e more feas ib le to implement on a large sca le , which will a l low for more granular levels of ana lys is with regard to afflicted regions - with more data, there will be more statistical power when analyz ing these subse ts of the patient populat ion. Live graphical reporting of these data could add an effective m e a n s of monitoring trends for those who may be in a posit ion to direct research on a large sca le . In the P S T c a s e , the client fee ls that patients with shou lder injuries may compr ise a particularly respons ive group. A s more data are col lected, it would be of interest to s e e if this is the c a s e , and if it is, to try to identify factors that may be assoc ia ted with a favourable outcome. This could identify a group that would be wel l sui ted to double-bl ind exper imentat ion. 64 9 Work Cited A n d e r s o n , J . and D.T. Fe l son . "Factors assoc ia ted with osteoarthrit is of the knee in the First Nat ional Health and Nutrition Examinat ion. " Amer i can Journa l of Ep idemio logy 128 (1988): 179-189. Bas insk i , A . S . H . "Evaluat ion of cl inical pract ice guidel ines." C a n a d i a n Med ica l Assoc ia t i on Journa l 153.11 (1995): 1575-1581. Bas insk i , A . S . H . " Implementing pract ice guidel ines." C a n a d i a n Med ica l Assoc ia t ion Journa l 148.5 (1993): 753-755. B e c k m a n , C . Pe rsona l Interview. Summer , 1999. B rookes , P . H . A s s e s s i n g Med ica l Rehabi l i tat ion Prac t ices : T h e P romise Of O u t c o m e s R e s e a r c h . Bal t imore, M A : P . H . B rookes ,1997 . Ca rey , R . G , and E . J . P o s a v a c . "Rehabi l i tat ion program evaluat ion using a revised level of rehabilitation sca le (LORS-I I ) . " A rch ives of Phys ica l Med ic ine and Rehabi l i tat ion. Augus t 63.8 (1982): 367-370. C a s s i d y , J . D . "Measur ing Ou tcomes . How can we ensure that heal th-care serv ices del iver the most for our m o n e y ? " Recove ry 10.1 (1999): 15-17. C h i p m a n , M. "In sea rch of Cohe rence . " Recove ry 10.1 (1999): 4-7. Davenport , R . J . and M . S . Denn is . " A s s e s s i n g the Qual i ty of C a r e : Measur ing the p rocess of care is not a lways straightforward." Brit ish Med ica l Journa l 312 (1996): 185. Dav ies , H.T. and I.K. C romb ie . " A s s e s s i n g the Qual i ty of C a r e : Measur ing wel l suppor ted p rocesses may be more enl ightening than monitoring ou tcomes. " Brit ish Med ica l Journa l 311 (1995): 766. Donabed ian , A . "The Quali ty of C a r e . How C a n It B e A s s e s s e d ? " T h e Journa l of the Amer i can Med ica l Assoc ia t ion 260.12 (1988): 1743-1748. Fe l son , D.T., A n d e r s o n , J . J . , Naimark, A . , Wa lker , A . M . , and R .F . M e e n a n . "Obesi ty and knee osteoarthrit is: T h e F ramingham study." Anna ls of Internal Med ic ine 109 (1988): 18-24. Fe l son , D.T. and C . E . C h a i s s o n . "Unders tand ing the relat ionship between body weight and osteoarthrit is." Bai l l ieres Cl in ical Rheumato logy 11 (1997): 671-681 . Freemant le , N, E.L. Harvey, F. Wolf , J . M . Gr imshaw, R. Gri l l i , and L.A. Bero . "Pr inted educat ional mater ials: effects on profess ional pract ice and health ca re ou tcomes. " C o c h r a n e Library Issue 2 (1999): (needs page numbers) . Dobrzykowsk i , E. A . (Ed.). Essent ia l Read ings in Rehabi l i tat ion O u t c o m e s Measurement : Appl icat ion, Methodology, and Techno logy . Gai thersburg , M D : A s p e n Publ ishers 1998. 65 G o e r g e , R . M . and B . J . Lee . "Match ing and C lean ing Administrat ive Data , " S tud ies of Wel fare Populat ions: Data Col lect ion and R e s e a r c h Issues. M.Ver P loeg , R.A. Moffitt and C . F . Citro. (Eds.) . Wash ing ton , D .C . : Nat ional A c a d e m y P r e s s (2002): 197-198, 217. Granger , C . V . and C M . Brownsche id le . "Ou tcome measurement in medica l rehabil i tation." International Journa l of Techno logy A s s e s s m e n t in Health C a r e . Spr ing 11.2 (1995): 262-268. Horton, N .J . and S . R . Lipsitz. "Rev iew of Software to Fit Genera l i zed Est imat ing Equat ion Reg ress i on Mode ls . " T h e Amer i can Stat ist ic ian. M a y 53 (1999): 160-169. Hosmer , D.W. and S . Lemeshow . App l ied Logist ic Reg ress i on . N e w York . J o h n Wi ley & S o n s , 2000. Huf fman, J . Pe rsona l Interview, 1999. K a n g , S . and K.A. K im. "The Survey on the Comp le teness of the Med ica l R e c o r d s as the Bas i s for Produc ing Va luab le Heal th Information." M E D I N F O E d . Cesn ik , B., A . M c C r a y , and J . R . Scherrer . 9.1 Ams te rdam: IOS P r e s s , 1998. 75-79. Law, M. , G . K ing , D. Russe l l , E. M a c K i n n o n , P. Hurley, and C . Murphy. "Measur ing ou tcomes in chi ldren's rehabil i tation: a dec is ion protocol." A rch ives of Phys ica l Med ic ine and Rehabi l i tat ion June 80.6 (1999): 629-636. L in, C . " A Simulat ion of C a s e Managemen t Operat ions at the Worke rs ' Compensa t i on Board - a Dec is ion Suppor t Too l For H u m a n R e s o u r c e Al locat ion. " Dissertat ion, U B C , 2000. Lo rence , D .P . , and I.A. Ibrahim. "Dispari ty on Cod ing C o n c o r d a n c e : Do Phys ic ians and C o d e r s A g r e e ? " Journa l of Heal th C a r e F inance 29.4 (2003): 43 -53 . Mant, J . and N. Hicks. "Detect ing Dif ferences in Qual i ty of C a r e : T h e Sensit iv i ty of P r o c e s s and Ou tcome in Treat ing Acu te Myocard ia l Infarction." Brit ish Med ica l Journa l Sep tember 311 (1995): 793-796. Marano , Hara Estroff. "When Depress ion Hurts the Mind" Ju ly /Augus t Psycho logy Today (2002) h t tp : / /www.psvcho logv todav .com/h tdocs /prod /PTOAr t ic le /PTO-20030806-000016 .asp . A c c e s s e d July 28 , 2004. Parente , S . , J . P . Weiner , D.W. Garn ick , T . M . R ichards , J . Fow les , A . G . Lawthers, P. Chand le r and R . H . Pa lmer . "Deve lop ing a Quali ty Improvement Da tabase Us ing Health Insurance Data: A Gu ided Tour with Appl icat ion to Med icare 's Nat ional C la ims History Fi le" Amer i can Journa l of Med ica l Qual i ty 10.4 (1995): 162 P a r s o n s , D. "Quanti fying T rauma. " Recove ry 10.1 (1999): 8-9. Re imer , C . "Bui lding G o o d Information." Recove ry 10.1: (1999): 4-6. R y a n , K.A. "The Cohor t Project Report . A Longitudinal Ana lys i s of N e w C l a i m s Regis tered with the W C B of B C , 1994-1996." The Queen ' s Printer Nov 13 (1998). ht tp: / /www.gp.gov.be.ca/rcwc/research/ index.htm A c c e s s e d D e c e m b e r 6, 2000 . Schuber t , J . "Rehabi l i tat ion Focus . " Recove ry 10.1 (1999): 15-17. 66 S c h u r m a n , B.A. "Durable data. A rehabilitation center col lects and distr ibutes information for functional a s s e s s m e n t and program evaluat ion." Heal th Progress . November 71.9 (1990): 52-54. S tausberg , J . , O. Ko ike , and K. Albrecht . "Us ing a Compu te r -based Patient Reco rd for Qual i ty Managemen t in Surgery." M E D I N F O . E d . Cesn i k , B., A . M c C r a y , and J . R . Scherrer . 9.1 Ams te rdam: IOS P r e s s , 1998. 80-84. Straker, L .M. "Body discomfort a s s e s s m e n t tools." T h e Occupat iona l E rgonomics Handbook. Karwowsk i , W . and W . S . Mar ras . (Eds.) . S a l e m , M A : C R C P r e s s L L C , (1999): 1239-1251. Sz ick , S . , D. A n g u s , G . Nicho l , M. Harr ison, J . P a g e , and D. Moher . "Heal th C a r e Del ivery in C a n a d a and T h e United Sta tes : A re there Relevant Di f ferences in Health C a r e O u t c o m e s ? " I C E S (The Institute for Cl in ical Evaluat ive S c i e n c e s ) Publ icat ion # 9 9 - 0 4 - T R . 3 J u n e (1999): h t tp : / /www.n lcahr .mun.ca/dwnlds/ ICES 1999a.pdf. A c c e s s e d Apri l 2, 2004. Trock, D.H. , A . J . Bollet and R. Markol l . "The Effect of Pu l sed E lect romagnet ic F ie lds in the treatment of Osteoarthri t is of the Knee and Cerv ica l Sp ine . Repor t of R a n d o m i z e d , Double B l ind, P l a c e b o Control led Trials", Journa l of Rheumato logy . 21 (1994): 1903-11. Urbanov ich , E. "Identifying H igh-R isk C la ims Within the W C B of B C ' s C la im Inventory by Us ing Logist ic Reg ress ion Model ing . " Dissertat ion, U B C , 1999. W C B Mandate P a g e , ht tp: / /www.worksafebc.com/about us/defaul t .asp. A c c e s s e d Apri l 3, 2004. Who ley , D.R., R. P a d m a n , R. H a m e r and S . Schwar tz . "The Diffusion of Information Techno logy A m o n g Heal th Ma in tenance Organ izat ions. " Heal th C a r e Managemen t Rev iew 25.2 (2000): 24-28. Wiener , J . P . , S T . Parente , D.W. Garn ick , J . Fow les , A . G . Lawthers, and R . H . Pa lmer . "Variat ion in Of f i ce -Based Quali ty: A C l a i m s - B a s e d Profi le of C a r e Prov ided to Med icare Pat ients with Diabetes. " Journa l of the A m e r i c a n Med ica l Assoc ia t ion 273.19 (1995): 1503-1507. Wi l l iams, J.I. and W . Y o u n g . "Accu racy of Canad ian Heal th Administrat ive Da tabases . " I C E S (The Institute for Cl in ical Evaluat ive Sc iences ) Publ icat ion # 9 6 - 0 3 - W P . 3 J u n e (1996): 1-43. ht tp: / /phru.medicine.dal .ca/br iefs/admindata.htm. A c c e s s e d Apri l 2 , 2004. 67 10 Appendices 10.1 Appendix 1 - PST database relationship diagram TblPurirnlbiio PatientID LastName FirstName MiddleName SM Birthdate Heightlnches WeightLbs Street City Province PostaKode Country HomePhone WorkPhone CellPhone FexN umber EmailName NetureOflnterestlnPST ReferredByPhysiciariYN RefemngPhysKiarJD ExaminingPhysiciarJD XRaysYN XraysDate XraysSureOfDate XraysNotes Notes DemographicNotes Radio TV YellowPages Newspaper Articles Relative Friend PreviousPatient HealthCanProvider ChiucPhysician NotesAbotfHowHeard TblP hvsician P k y i k t a r J D Tnk DoctorLastName DoctorFirstName Street City PostaKode Province Country Phor*Nurnber FaxH umber EmuIAoVlress Notes I*bIJoin tOi Ren ion JointOrRegionID JointOtRetionNeme HTJSiiniD^^^H Title DoctorLastName DoctorFirstName Street City PostalCode Province Country PhoneN umber FaxN umber EmailAddress Notes TblPitirnlJouit PatientID JointOrRegionJD PreviousOperationYN OperationNotes Physio YN Cairo YN Massage YN AccupunctureYN PreviousOtherTherapyY TherapyNotes PrevBusInjuryYN NSAID.YN Antidepressants YN Analgesic YN MuscleRelaxantYN MedicationNotes [ I ' l D i t ' i i p r Diagnosis ID DiagnossDescnption OA or STI T M P a tie n t Join t Ev J lu iri on EvaluationID PatientID JointOrReponlD SpecialJnstiuctions_Foi_Patient Handling CommentsLog InjuiyDate DateAdmitted DiagnosklD Severity HistoryAndPhysKalF ladings MechmelD RecornmendedNumberOfTreat merits Additional_Numbei_Of_Treatments NumberOfSessions MedicationType MedxetionName Medxationlnterval MedxtionDose Before Evaluat«5n_Date BeforeExanuningPhysicianlD BeforePauiJntensity BefcrePam_Freq<jency BeforeRestr«:t»n_Of_Movement BeforeS welling BeforeWarmmg BeforeDiscolounng BeforeParesthetic BeforeChenge_Of_Med»:ation BeforeCommenls_By_Physician BeforeComments By Patient BeforeALD_Notes Before_Sleepmg_ Duration Before_Standing_Duration BeforeWalking BeforeS ittmg BtfoieDnvmg RightAfterEvaluaton^Date RightAfteiExanuningPhysciarJD RighlAfterPainJntensity RightAfteiPam Frequency PoghtAfterResirrtionJDf_Moveinent RlghtAfterS welling RightAfterWarming RuhtAfterDiscolounns 68 10.2 Appendix 2 - Relationship properties tblDiagnosistblePatientJointEvaluation TblDiagnosis tblPatientJointEvaluation DiagnosisU) 1 oo DiagnosisID Attributes: Enforced, Inherited, Cascade Updates, Cascade Deletes Relationship Type: One-To-Many (External) TblJointOrRegiontblPatientJoint TblJointOrRegion tblPatientJoint JointOrRegionID 1 oo JointOrRegionID Attributes: Enforced, Inherited, Cascade Updates, Cascade Deletes Relationship Type: One-To-Many (External) tblPatientJointtblPatientJointEvaluation TblPatientJoint tblPatientJoint PatientID 1 oo PatientID JointOrRegionID 1 oo JointOrRegionID Attributes: Enforced, Inherited, Cascade Updates, Cascade Deletes Relationship Type: One-To-Many (External) TblPhysiciantblPatientlnfo TblPhysician tblPatientlnfo PhysicianID 1 oo ReferringPhysicianID Attributes: Enforced, Inherited, Cascade Updates, Cascade Deletes Relationship Type: One-To-Many (External) 69 10.3 Appendix 3 - Univariate analysis 10.3.1 Predictors Variable N Missing Mean Standard Deviation Min Q1 Median Q3 Max Age at Admission 592 7 52.6 16.6 13.8 40.4 51.9 65.9 100.1 Age of Injury 592 8 7.8 9.6 0.0 2.5 4.3 9.9 56.4 Body Mass Index 551 48 24.6 4.2 15.7 21.6 24.2 26.6 44.3 Variable Values N % % Valid Sex Male 221 36.9% 36.9% Female 378 63.1% 63.1% Total 599 100.0% Age at Admission (Categories) 10-19 11 1.8% 1.9% 20-29 44 7.3% 7.4% 30-39 91 15.2% 15.4% 40-49 132 22.0% 22.3% 50-59 110 18.4% 18.6% 60-69 98 16.4% 16.6% 70-79 77 12.9% 13.0% 80+ 29 4.8% 4.9% Missing 7 1.2% 1.2% Total 599 100.0% Age of Injury (Categories) <1 Years 36 6.0% 6.1% 1-4 Years 301 50.3% 50.8% 5-9 Years 109 18.2% 18.4% 10-14 Years 63 10.5% 10.6% >15 Years 83 13.9% 14.0% Missing 7 1.2% 1.2% Total 599 100.0% BMI Group (National Heart, Lung and Blood Institute (US) groupings) Underweight - under 18.5 27 4.5% 4.9% Acceptable weight -18.5 to 24.99 290 48.4% 52.6% Overweight - 25.0 to 29.99 180 30.1% 32.7% Obese - 30 or more 54 9.0% 9.8% Missing 48 8.0% 8.7% Total 599 100.0% 70 Variable Values N % % Valid Osteoarthri t is (OA) 243 4 0 . 6 % 4 5 . 2 % Soft T i ssue Injury (STI) 248 4 1 . 4 % 4 6 . 1 % Diagnos is Both (OA and STI) 47 7 .8% 8 .7% Uncer ta in /Miss ing 61 10 .2% Total 599 100 .0% O A d isc bulge/protrusion 16 2 .7% 2 . 9 % O A d isc degenerat ion/spondy los is 47 7 .8% 8 .6% O A d isc herniation 15 2 . 5 % 2 .7% O A joint s p a c e narrowing/cart i lage degenerat ion/spurs 19 3 .2% 3 .5% (other) osteoarthrit is 146 2 4 . 4 % 2 6 . 6 % STI cart i lage injury 1 0 .2% 0 .2% STI l igament injury 12 2 . 0 % 2 . 2 % D iagnos is STI men iscus injury 2 0 .3% 0.4% Descr ipt ion STI muscu lar injury 7 1.2% 1.3% STI muscu lo- l igamentous 97 16 .2% 17 .7% STI rotator cuff /capsule injury 6 1.0% 1.1% STI tendon injury 9 1.5% 1.6% other STI 114 19 .0% 2 0 . 8 % C o m b i n e d O A and STI 47 7 .8% 8 .6% Tinnitus 10 1.7% 1.8% Other /Uncer ta in 51 8 .5% Total 599 100 .0% Ankle(s) 8 1.3% 1.3% Back , Lower (L-spine) 137 2 2 . 9 % 2 2 . 9 % Back , Upper (T-spine) 28 4 . 7 % 4 . 7 % Ear(s) 15 2 . 5 % 2 . 5 % Elbow(s) 4 0 .7% 0 .7% Foot /Feet 7 1.2% 1.2% Hand(s) 4 0 .7% 0 .7% Joint or Hip(s) 61 10 .2% 10 .2% Reg ion N a m e Knee(s) 61 10 .2% 10 .2% Neck (C-spine) 212 3 5 . 4 % 3 5 . 4 % Pelv is 6 1.0% 1.0% Sacro i l iac (SU) (L/R/Both) 26 4 . 3 % 4 . 3 % Shoulder(s) 25 4 . 2 % 4 . 2 % T M J (L/R/Both) 3 0 .5% 0 .5% Wrist(s) 2 0 .3% 0 . 3 % Total 599 100 .0% 71 10.3.2 R e s p o n s e Var iab les Variable N Missing Mean Standard Deviation Min Q1 Median Q3 Max Pa in Intensity Before Treatment 598 1 3.3 0.8 0 3 3 4 4 Pa in F requency Before Treatment 598 1 3.4 0.8 0 3 4 4 4 Restr ict ion of Movement Before Treatment 582 17 3.0 1.2 0 2 3 4 4 Pa in Intensity Right After Treatment 589 10 2.7 1.0 0 2 3 3 4 Pa in F requency Right After Treatment 589 10 3.0 1.0 0 2 3 4 4 Restr ict ion of Movemen t Right After Treatment 573 26 2.6 1.3 0 2 3 4 4 Pa in Intensity S ix W e e k s Post -Treatment 561 38 2.4 1.1 0 2 2 3 4 Pa in F requency S ix W e e k s Post -Treatment 561 38 2.7 1.1 0 2 3 4 4 Restr ict ion of Movement S ix W e e k s Post -Treatment 548 51 2.4 1.2 0 2 2 3 4 Pa in Intensity S ix Months Post -Treatment 444 155 2.3 1.1 0 2 2 3 4 Pa in F requency S ix Months Post -Treatment 444 155 2.6 1.2 0 2 3 4 4 Restr ict ion of Movemen t S ix Months Post -Treatment 431 168 2.1 1.3 0 1 2 3 4 Pa in Intensity O n e Y e a r Post -Treatment 339 260 2.1 1.2 0 1 2 3 4 Pa in F requency O n e Y e a r Post -Treatment 339 260 2.5 1.3 0 2 2 4 4 Restr ict ion of Movement O n e Y e a r Post -Treatment 326 273 1.9 1.4 0 1 2 3 4 72 10.4 Appendix 4 - Bivariate analysis, crosstabs, interactions Age at Admission by Sex, Diac nosis, Diagnosis Description and Joint or Region Name by Level N Missing Mean Standard Deviation Min Q1 Median Q3 Max X 0) Male 215 6 52.8 16.8 13.8 39.5 52.8 65.8 100.1 CO Female 377 1 52.5 16.5 16.2 40.9 51.6 66.0 99.9 w Osteoarthritis (OA) 239 4 61.3 14.8 21.6 49.7 64.8 71.7 100.1 <n o c Soft Tissue Injury (STI) 247 1 43.3 13.6 13.8 34.2 43.4 51.4 99.9 ro Both (OA and STI) 47 0 57.0 12.6 33.0 49.2 56.9 67.4 84.4 b Uncertain/Missing 59 2 53.2 17.7 18.6 41.5 53.3 66.0 88.0 O A disc bulge/protrusion 16 0 44.8 11.1 24.9 36.2 46.3 50.9 65.7 O A disc degeneration/spondylosis 47 0 60.1 15.6 21.6 48.8 61.7 71.0 90.9 O A disc herniation 14 1 51.7 15.3 31.0 39.6 46.4 68.1 76.2 O A joint space narrowing/cartilage degeneration/spurs 19 0 57.3 13.5 32.0 49.1 55.3 68.7 80.1 g (other) osteoarthritis 143 3 65.0 13.1 24.4 57.0 67.4 73.3 100.1 C L STI cartilage injury 1 0 34.6 * 34.6 * 34.6 * 34.6 w STI ligament injury 12 0 52.0 13.1 38.7 42.3 47.3 57.7 79.7 )sis D STI meniscus injury 3 0 52.2 14.7 35.7 35.7 56.9 63.9 63.9 )sis D STI muscular injury 6 0 32.2 10.9 20.4 22.1 30.8 41.3 49.8 c .ro t—* STI musculo-ligamentous 97 0 42.2 11.0 13.8 34.4 43.1 49.0 66.4 STI rotator cuff/capsule injury 6 0 52.4 14.8 36.2 38.1 51.5 65.5 74.1 L J STI tendon injury 9 0 52.1 17.0 18.3 43.5 56.0 58.7 80.1 other STI 113 1 42.6 14.6 16.3 31.6 42.2 50.9 99.9 Combined O A and STI 47 0 57.0 12.6 33.0 49.2 56.9 67.4 84.4 Tinnitus 9 1 57.5 16.9 18.6 51.3 61.3 68.4 75.2 Other/Uncertain 50 1 52.4 17.9 19.1 37.9 52.9 64.7 88.0 Ankle 8 0 56.4 3.4 52.6 52.9 56.4 59.9 60.8 Back, Lower (L-spine) 136 1 53.6 17.7 16.3 39.1 50.3 68.5 90.9 Back, Upper (T-spine) 28 0 50.3 16.7 19.4 37.8 49.6 65.1 82.0 Ear(s) 14 1 63.2 10.4 44.5 54.5 63.2 71.8 80.6 Nam Elbow(s) 4 0 46.1 16.9 30.8 31.2 44.9 62.0 63.6 Nam Foot/Feet 7 0 51.8 13.5 29.5 44.5 52.9 59.9 73.1 c o Hand(s) 4 0 55.5 6.1 49.7 50.0 55.3 61.1 61.4 CD Hip(s) 60 1 61.9 15.2 19.7 52.7 65.4 71.7 86.4 or Knee(s) 61 0 61.6 15.9 18.3 53.3 65.2 72.6 100.1 O —^» Neck (C-spine) 208 4 48.2 14.7 13.8 37.3 47.0 58.4 85.3 c o Pelvis 6 0 36.1 13.4 22.7 24.3 32.3 50.5 55.3 Sacroiliac (SU) (L/R/Both) 26 0 44.3 11.7 21.3 39.6 45.6 49.0 72.6 Shoulder(s) 25 0 51.0 17.8 22.3 36.2 50.9 62.6 80.1 TMJ (L/R/Both) 3 0 26.9 1.3 25.5 25.5 27.1 28.1 28.1 Wrist(s) 2 0 68.8 43.9 37.8 * 68.8 * 99.9 73 Distribution of BMI by Sex, Diagnosis, Diagnosis Description and Joint or Region Name by Level N Missing Mean Standard Deviation Min Q1 Median Q3 Max X CD male 206 15 25.7 3.8 17.4 23.7 25.1 27.2 44.3 w female 345 33 23.9 4.3 15.7 20.7 23.3 26.4 40.4 S|SOU Osteoarthritis (OA) 223 20 25.6 4.3 16.2 23.1 25.4 27.9 44.3 S|SOU Soft Tissue Injury (STI) 228 20 23.7 3.7 15.7 21.0 23.3 25.8 37.4 O) CC Both (OA and STI) 43 4 24.9 3.9 17.8 21.8 24.5 28.0 34.2 b Uncertain/Missing 57 4 23.9 5.2 17.2 19.7 23.6 25.4 40.4 O A disc bulge/protrusion 16 0 24.8 3.5 20.2 0.9 24.6 20.2 34.0 O A disc degeneration/spondylosis 44 3 25.4 4.0 16.6 0.6 25.4 16.6 35.1 O A disc herniation 14 1 25.0 3.5 20.3 0.9 25.3 20.3 32.3 c O A joint space narrowing/cartilage degeneration/spurs 17 2 26.9 4.5 20.3 1.1 26.1 20.3 37.8 o (other) osteoarthritis 132 14 25.7 4.6 16.2 0.4 25.1 16.2 44.3 escrip STI cartilage injury 1 0 20.5 * 20.5 * 20.5 20.5 20.5 escrip STI ligament injury 11 1 24.0 3.0 18.0 0.9 24.8 18.0 28.3 a STI meniscus injury 3 0 22.0 4.6 16.8 2.6 23.8 16.8 25.4 '(Jl ,—i STI muscular injury 5 1 22.8 2.9 20.1 1.3 22.2 20.1 25.8 w C D) CD , STI musculo-ligamentous 91 6 23.8 3.7 15.7 0.4 23.2 15.7 37.4 STI rotator cuff/capsule injury 6 0 24.1 3.4 17.5 1.4 24.8 17.5 26.7 L_J STI tendon injury 7 2 25.4 5.8 20.9 2.2 23.7 20.9 36.7 other STI 104 10 23.5 3.6 17.4 0.4 23.0 17.4 32.9 Combined O A and STI 43 4 24.9 3.9 17.8 0.6 24.5 17.8 34.2 Tinnitus 7 3 24.9 0.7 24.2 0.3 24.9 24.2 25.8 Other/Uncertain 50 1 23.7 5.6 17.2 0.8 23.3 17.2 40.4 Ankle 8 0 28.4 5.0 22.1 24.7 27.0 33.7 36.7 Back, Lower (L-spine) 129 8 25.0 4.0 16.6 22.1 24.9 27.3 39.6 Back, Upper (T-spine) 25 3 24.5 4.0 17.4 22.3 23.8 26.8 33.1 CD Ear(s) 11 4 24.7 3.0 17.7 24.2 24.9 25.8 29.4 £ CO Elbow(s) 4 0 22.3 1.0 21.7 21.8 22.2 22.8 23.0 z Foot/Feet 7 0 28.8 6.0 23.6 25.0 26.6 30.1 40.4 1 c o Hand(s) 4 0 24.0 6.0 18.9 19.2 22.3 30.4 32.5 'co CD Hip(s) 54 7 24.8 4.0 16.2 21.5 24.7 27.4 35.1 LY. Knee(s) 56 5 26.0 5.0 16.8 23.0 25.4 28.9 41.2 o Neck (C-spine) 195 17 23.8 4.0 15.7 21.0 23.6 25.8 44.3 cz 'o Pelvis 6 0 23.9 4.0 19.4 19.7 24.8 26.6 28.5 Sacroiliac (SU) (L/R/Both) 23 3 24.4 3.0 19.0 21.0 24.9 26.5 31.1 Shoulder(s) 24 1 23.7 4.0 17.5 19.6 24.7 26.7 29.3 TMJ (L/R/Both) 3 0 22.9 1.0 22.2 22.2 23.1 23.3 23.3 Wrist(s) 2 0 21.8 3.0 20.0 * 21.8 * 23.6 74 10.4.1 Cross - tabs : D iagnos is by S e x Obse rved counts O A STI Both A L L Ma le 107 75 10 192 F e m a l e 136 1 7 3 37 346 A L L 243 248 47 538 C H I - S Q U A R E = 14.831 W I T H D.F. = 2 p = 0.0006 10.4.2 Cross - tabs : D iagnos is by BMI Obse rved counts underweight normal overweight obese A L L O A 5 99 87 32 223 STI 16 135 65 12 228 O A / S T I 1 22 15 5 4 3 A L L 22 256 167 49 494 C H I - S Q U A R E = 24.250 W I T H D.F. = 6 p = 0.0005 10.4.3 Cross - tabs : BMI by S e x Obse rved counts underweight normal overweight obese A L L Ma le 5 86 93 22 206 F e m a l e 22 204 87 32 345 A L L 27 290 180 54 551 C H I - S Q U A R E = 27.451 W T H D.F. = 3 p < 0.0001 10.4.4 Cross - tabs : D iagnos is ( O A or STI) by A g e Group Obse rved counts Under 40 40-591 60+ A L L O A 24 76 139 239 STI 103 119 25 247 A L L 127 195 164 486 C H I - S Q U A R E = 137.773 W I T H D.F. = 2 p < 0.00001 10.4.5 Corre lat ions of initial Pa in Intensity, Pa in Frequency , and Restr ict ion of Movement sco res Initial Pa in Initial Pa in Initial Restr ict ion Intensity F requency of Movemen t Initial Pa in P e a r s o n Correlat ion 1 .392 .238 Intensity S ig . (2-tailed) .000 .000 N 598 598 582 Initial Pa in P e a r s o n Correlat ion .392 1 .266 Frequency S i g . (2-tailed) .000 .000 N . 598 598 582 Initial Restr ict ion of Movement P e a r s o n Correlat ion .238 .266 1 S i g . (2-tailed) .000 .000 N 582 582 582 10.5 Appendix 5 - Relationships between Outcomes and Predictors 10.5.1 Logist ic Regress ion 10.5.1.1 Logistic model E(Y|x) = 0 0 + 01 x 77(X) = E(Y|x) 77(x) = exp [00 + £1 x] / (1 + exp [00 + 01 x]) 10.5.1.2 The logit transformation g( X ) = In [77(X)/ (1 - 77(X))] 10.5.1.3 Dependent variable encoding Original Value Internal Value Not Improved 0 Improved by at least one unit 1 10.5.1.4 Categorical variables codings (Parameter coding of zero indicates compar ison group) Paramete r coding (1) Body M a s s Index Group Not Overweight (BMI <25) 0 Overweight (BMI >25) 1 A g e Group Under 50 0 50+ 1 Injury Type O A 0 STI 1 A g e of Injury Group Under 4 years 0 4 or more years 1 S e x Male 0 Fema le 1 10.5.1.5 Univariable models - predictors run one at a time Improvement in Score = Sex, Age Group, BMI Group, Age Of Injury, Injury Type, Initial Score (note: all but Initial Score are dichotomized categorical contrasts) g(x) = 0O + 01 (Sex) g(x) = 0o+01 (Age Group) g(x) = 0O + 01 (Body Mass Index) g(x) = 0O + 01 (Age of Injury Group) g(x) = 0o + 0i (Injury type) g(x) = 0o + 0i (Initial Score) 10.5.1.6 Significant single categorical contrasts, controlling for initial score Improvement in Score = Individual significant (p>0.25) contrasts from univariable models, plus Initial Score. This will be run with and without the two-way interaction between the contrast and the initial score. The predicted outcomes will be compared graphically to observed data. 10.5.1.7 Full main-effects models Improvement in Score = Sum of all significant (p>.25) univariable predictors. Final step of backward stepwise elimination will produce the reduced model, which will be re-run using entry method to capture all available data. The predicted outcomes will be compared graphically to observed data. 10.5.1.8 Full interaction models Improvement in Score = Sum of significant variables from the reduced main effects model, plus their two-way interactions. Final step of backward stepwise elimination will produce the reduced model, which will be re-run using entry method to capture all available data. The predicted outcomes will be compared graphically to observed data. 10.5.1.9 Final model Thorough evaluation of the above models will lead to the determination of the final model used for each outcome variable at each stage of patient evaluation. 78 10.5.2 Right After Treatment 10.5.2.1 Univariable models - each predictor modeled separately Pain Intensity Right after treatment -Univariable Models n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 589 0.21 0.17 0.24 1.23 0.87 1.73 Age Group (Age 50+) 582 0.64 0.17 <0.01 1.89 1.35 2.65 BMI Group (Overweight 25+) 541 0.06 0.18 0.73 1.06 0.75 1.50 Injury Type (STI) 484 -0.63 0.19 <0.01 0.53 0.37 0.77 Age of Injury Group (4+ years) 582 0.02 0.17 0.89 1.02 0.74 1.43 Initial Pain Intensity (raw score zero if no symptoms present, higher as symptoms worsen) 589 0.48 0.12 <0.01 1.62 1.28 2.06 Pain Frequency Right after treatment -Univariable Models n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 589 0.05 0.18 0.78 1.05 0.74 1.50 Age Group (Age 50+) 582 0.59 0.18 <0.01 1.81 1.27 2.57 BMI Group (Overweight 25+) 541 0.24 0.19 0.20 1.27 0.88 1.82 Injury Type (STI) 484 -0.75 0.20 <0.01 0.47 0.32 0.69 Age of Injury Group (4+ years) 582 -0.15 0.18 0.40 0.86 0.61 1.22 Initial Pain Frequency (raw score zero if no symptoms present, higher as symptoms worsen) 589 0.26 0.12 0.03 1.30 1.03 1.64 Restriction of Movement Right after treatment -Univariable Models n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 573 -0.00 0.19 0.98 1.00 0.68 1.45 Age Group (Age 50+) 567 0.25 0.19 0.18 1.29 0.89 1.85 BMI Group (Overweight 25+) 528 0.16 0.19 0.40 1.17 0.81 1.71 Injury Type (STI) 484 -0.45 0.20 0.03 0.64 0.43 0.95 Age of Injury Group (4+ years) 566 -0.28 0.18 0.12 0.75 0.52 1.08 Initial Restriction of Movement (raw score zero if no symptoms present, higher as symptoms worsen) 573 0.44 0.10 O.01 1.56 1.29 1.88 79 10.5.2.2 Significant single contrasts, controlling for Initial Pain Intensity Observed one unit improvement in PI Right after treatment, n = 589 Initial Pain Intensity 1 2 3 4 Female Proport ion with improvement 25% 33% 36% 52% Cel l count 4 30 157 181 Male Proport ion with improvement 0% 20% 45% 40% Cel l count 7 25 107 75 Improvement in Pain Intensity Right after treatment Main Effects Model Sex + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 589 0.11 0.18 0.55 1.11 0.78 1.58 Initial Pain Intensity 0.47 0.12 O .01 1.60 1.26 2.04 Constant -1.97 0.42 <0.01 0.14 -2Loglikelihood: 782.5 DF = 3 Improvement in Pain Intensity Right after treatment Interaction Model Sex + PI + Sex*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 589 -0.11 0.84 0.90 0.90 0.17 4.65 Initial Pain Intensity 0.43 0.19 0.02 1.54 1.06 2.24 Sex* Initial Pain Intensity 0.07 0.25 0.79 1.07 0.66 1.74 Constant -1.85 0.63 <0.01 0.16 -2Loglikelihood: 782.4 DF = 4 Comparison of observations with predictions 0.6 Sex (Observed Data) 2 3 Initial Pain Intensity 0.6 Sex (Main Effects Model) 2 3 Initial Pain Intensity 0.6 Sex (Interaction Model) 2 3 Initial Pain Intensity 80 Pain Intensity with Age Group Observed one unit improvement in PI Right after treatment, n = 582 Initial Pain Intensity 1 2 3 4 50 and over Proport ion with improvement 11% 26% 51% 56% Cel l count 9 38 140 120 Under age 50 Proport ion with improvement 0% 27% 26% 41% Cel l count 2 15 122 133 Improvement in Pain Intensity Right after treatment Main Effects Model Age Group + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 589 0.78 0.18 O.01 2.18 1.54 3.09 Initial Pain Intensity 0.57 0.13 <0.01 1.77 1.38 2.27 Constant -2.69 0.46 O.01 0.07 -2Loglikelihood: 751.8 DF = 3 Improvement in Pain Intensity Right after treatment Interaction Model AgeGroup + PI + AgeGroup*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 589 0.72 0.93 0.44 2.05 0.33 12.73 Initial Pain Intensity 0.56 0.22 0.01 1.75 1.14 2.69 Age Group* Initial Pain Intensity 0.02 0.27 0.95 1.02 0.60 1.72 Constant -2.65 0.77 <0.01 0.07 -2Loglikelihood: 751.8 DF = 4 Comparison of observations with predictions 0.6 - Age Group (Observed Data) 0.5 0.4 -0.3 0.2 -0.1 ^ ^ ^ ^ — 5 0 and over • • ^ ^ U n d e r age 50 0.0| 2 3 4 Initial Pain Intensity 0.6 0.0 Age Group (Main Effects Model) 2 3 Initial Pain Intensity 0.0 Initia i? Pain Intensity 81 Pain Intensity with InjuryType Observed one unit improvement in PI Right after treatment, n = 484 Initial Pain Intensity 1 2 3 4 O A Proport ion with improvement 13% 43% 47% 55% Cel l count 8 30 109 91 STI Proport ion with improvement 0% 5% 32% 40% Cel l count 3 19 107 116 Improvement in Pain Intensity Right after treatment Main Effects Model InjuryType + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 484 -0.74 0.19 <0.01 0.48 0.33 0.70 Initial Pain Intensity 0.50 0.14 O.01 1.64 1.26 2.14 Constant -1.66 0.45 O.01 0.19 -2Loglikelihood: 627.6 DF = 3 Improvement in Pain Intensity Right after treatment Interaction Model InjuryType + PI + lnjuryType*PI n (out of 599) Beta S .E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 484 -1.63 0.97 0.09 0.20 0.03 1.32 Initial Pain Intensity 0.40 0.17 0.02 1.49 1.07 2.07 InjuryType* Initial Pain Intensity 0.26 0.28 0.35 1.30 0.75 2.26 Constant -1.34 0.56 0.02 0.26 -2Loglikelihood: 626.7 DF = 4 Comparison of observations with predictions o.oi Injury Type (Observed Data) 1 2 3 Initial Pain Intensity 0.6 n Injury Type (Main Effects Model) 1 2 3 Initial Pain Intensity 0.6 Injury Type (Interaction Model) 2 3 Initial Pain Intensity 82 10.5.2.3 Testing full main effects and interaction models: Pain Intensity Pain Intensity with Injury Type and Age Group Observed one unit improvement in PI Right after treatment, n = 479 Initial Pain Intensity 1 2 3 4 O A 50 and Over Proportion with improvement 0% 14% 38% 49% Cell count 7 7 21 81 O A Under 50 Proportion with improvement 0% 0% 57% 39% Cell count 1 1 7 28 STI 50 and Over Proportion with improvement 0% 0% 9% 60% Cell count 2 2 11 30 STI Under 50 Proportion with improvement 0% 0% 0% 20% Cell count 1 1 8 76 Improvement in Pain Intensity Right after treatment Main Effects Model Age Group + Injury Type + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 479 0.61 0.22 O.01 1.84 1.19 2.83 Injury Type -0.46 0.22 0.03 0.63 0.41 0.97 Initial Pain Intensity 0.55 0.14 <0.01 1.73 1.32 2.28 Constant -2.30 0.52 <0.01 0.10 -2Loglikelihood: 612.2 DF = 4 Improvement in Pain Intensity Right after treatment Interaction Model # 1 AgeGroup + InjuryType + PI + lnjuryType*PI n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 479 0.64 0.22 <0.01 1.89 1.22 2.93 Injury Type -1.68 0:98 0.09 0.19 0.03 1.28 Initial Pain Intensity 0.41 0.17 0.02 1.51 1.08 2.12 InjuryType* Initial Pain Intensity 0.37 0.29 0.20 1.44 0.82 2.53 Constant -1.88 0.60 <0.01 0.15 -2Loglikelihood: 610.6 DF = 5 Improvement in Pain Intensity Right after treatment Interaction Model # 2 AgeGroup + InjuryType + PI + AgeGroup*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 479 0.85 1.00 0.40 2.34 0.33 16.76 InjuryType -0.47 0.22 0.03 0.63 0.41 0.96 Initial Pain Intensity 0.60 0.24 0.01 1.82 1.14 2.88 Age Group* Initial Pain Intensity -0.07 0.29 0.81 0.93 0.53 1.65 Constant -2.46 0.83 <0.01 0.09 -2Loglikelihood: 612.2 DF = 5 83 Improvement in Pain Intensity Right after treatment Interaction Model # 3 AgeGroup + InjuryType + PI + AgeGrou p*l nj u ryType n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 479 0.44 0.31 0.15 1.55 0.85 2.83 Injury Type -0.64 0.31 0.04 0.53 0.28 0.98 Initial Pain Intensity 0.56 0.14 <0.01 1.75 1.33 2.30 AgeGroup*lnjuryType 0.34 0.43 0.43 1.41 0.60 3.29 Constant -2.20 0.54 <0.01 0.11 -2Loglikelihood: 611.6 DF = 5 Improvement in Pain Intensity Right after treatment Interaction Model # 4 InjuryType + InitialPainlntensity + AgeGroup*lnitialPainlntensity + lnjuryType*lnitialPainlntensity n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 479 -1.93 0.99 0.05 0.15 0.02 1.02 Initial Pain Intensity 0.26 0.18 0.14 1.30 0.92 1.84 AgeGroup*lnitial Pain Intensity 0.19 0.07 <0.01 1.21 1.06 1.37 Injury Type*lnitial Pain Intensity 0.44 0.29 0.13 1.55 0.88 2.76 Constant -1.37 0.57 0.02 0.25 -2Loglikelihood: 610.6 DF = 5 84 Comparison of observations with predictions 85 10.5.2.4 Significant single contrasts, controlling for Initial Pain Frequency Observed one unit improvement in PI Right after treatment, n = 582 Initial Pain Intensity 1 2 3 4 50 and over Proportion with improvement 17% 24% 43% 42% Cell count 6 42 107 151 Under age 50 Proportion with improvement 0% 15% 24% 29% Cell count 1 20 92 159 Improvement in Pain Frequency Right after treatment Main Effects Model Age Group + PF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 582 0.67 0.18 O.01 1.95 1.36 2.79 Initial Pain Intensity 0.35 0.12 <0.01 1.42 1.11 1.81 Constant -2.29 0.46 <0.01 0.10 -2Loglikelihood: 717.22 DF = 3 Improvement in Pain Frequency Right after treatment Interaction Model Age Group + PF + Age Group *PF n (out of 599) Beta S .E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 582 0.75 0.96 0.44 2.12 0.32 13.96 Initial Pain Frequency 0.37 0.22 0.10 1.45 0.93 2.24 Age Group * Initial Pain Freq'cy -0.02 0.27 0.93 0.98 0.58 1.66 Constant -2.35 0.81 <0.01 0.10 -2Loglikelihood: 717.21 DF = 4 Comparison of observations with predictions 0.5 Age Group (Observed Data) 1 2 3 Initial Pain Frequency 0.5 Age Group (Main Effects Model) •50 and over •Underage 50 2 3 Initial Pain Frequency 0.5 Age Group (Interaction Model) 1 2 3 Initial Pain Frequency 86 Pain Frequency with BMI Group Observed one unit improvement in PF Right after treatment, n = 541 Initial Pain Frequency 1 2 3 4 Not overweight (BMI < 25) Proport ion with improvement 0 % 1 8 % 3 3 % 3 2 % Ce l l count 3 33 102 168 Overweight (BMI 25 +) Proport ion with improvement 2 5 % 3 2 % 3 0 % 4 0 % Cel l count 4 28 80 119 Improvement in Pain Frequency Right after treatment Main Effects Model BMI Group + PF n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 541 0.25 0.19 0.17 1.29 0.89 1.86 Initial Pain Frequency 0.28 0.13 0.02 1.33 1.04 1.70 Constant -1.81 0.45 <0.01 0.16 -2Loglikelihood: 673.98 DF = 3 Improvement in Pain Frequency Right after treatment Interaction Model BMI Group + PF + AgeGroup*PF n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 541 0.29 0.88 0.74 1.34 0.24 7.54 Initial Pain Frequency 0.29 0.17 0.09 1.34 0.96 1.87 BMI Group* Initial Pain Freq'cy -0.01 0.25 0.96 0.99 0.60 1.62 Constant -1.83 0.60 <0.01 0.16 -2Loglikelihood: 673.98 DF = 4 Comparison of observations with predictions BMI (Observed Data) 1 2 3 Initial Pain Frequency 2 3 Initial Pain Frequency 87 Pain Frequency with InjuryType Observed one unit improvement in PF Right after treatment, n = 484 Initial Pain Frequency 1 2 3 4 O A Proportion with improvement 20% 26% 43% 48% Cell count 5 35 86 111 STI Proportion with improvement 0% 15% 30% 24% Cell count 0 20 82 143 Improvement in Pain Frequency Right after treatment Main Effects Model InjuryType + PF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper InjuryType 484 -0.83 0.20 <0.01 0.44 0.29 0.65 Initial Pain Frequency 0.29 0.14 0.03 1.34 1.03 1.74 Constant -1.28 0.46 <0.01 0.28 -2Loglikelihood: 597.32 DF = 3 Improvement in Pain Frequency Right after treatment Interaction Model InjuryType + PF + lnjuryType*PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 484 0.72 1.00 0.47 2.05 0.29 14.47 Initial Pain Frequency 0.44 0.17 <0.01 1.55 1.11 2.17 Injury Type* Initial Pain Freq'cy -0.45 0.28 0.12 0.64 0.37 1.12 Constant -1.77 0.58 <0.01 0.17 -2Loglikelihood: 594.89 DF = 4 Comparison of observations with predictions 0.6 i 0.0 Injury Type (Observed Data) 2 3 Initial Pain Frequency 0.6 n Injury Type (Main Effects Model) 0.0 2 3 Initial Pain Frequency 0.6 i Injury Type (Interaction Model) 0.0 1 2 3 Initial Pain Frequency 88 10.5.2.5 Testing full main effects and interaction models: Pain Frequency Observed one unit improvement in PF Right after treatment, n = 479 Initial Pain Frequency 1 2 3 4 OA 50 and Over Proport ion with improvement 0% 25% 23% 49% Cel l count 4 4 26 63 OA Under 50 Proport ion with improvement 0% 0% 14% 23% Cel l count 1 1 7 22 STI 50 and Over Proport ion with improvement 11% 40% Cel l count 0 0 9 25 STI Under 50 Proport ion with improvement 18% 26% Cel l count 0 0 11 57 Improvement in Pain Frequency Right after treatment: Main Effect Main Effects Model Age Group + Injury Type + PF n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp< C.I. for Beta) Lower Upper Age Group 479 0.33 0.23 0.14 1.39 0.8S 2.16 Injury Type -0.66 0.23 <0.01 0.52 0.33 0.80 Initial Pain Frequency 0.35 0.14 0.01 1.43 1.09 1.87 Constant -1.76 0.52 <0.01 0.17 -2Loglikelihood: 586.88 DF = 4 Improvement in Pain Frequency Right after treatment:Interaction AgeGroup + InjuryType + PF + lnjuryType*PF n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 479 0.30 0.23 0.19 1.35 0.87 2.10 InjuryType 0.43 0.14 <0.01 1.53 1.16 2.03 Initial Pain Frequency -0.21 0.06 <0.01 0.81 0.72 0.92 Injury Type* Initial Pain Freq'cy -1.96 0.52 O.01 0.14 <0.01 <0.01 Constant 0.30 0.23 0.19 1.35 -2Loglikelihood: 585.03 DF = 5 Comparison of observations with predictions Injury Type & Age Group (Observed Data) 1 2 3 Initial Pain Frequency -OA 50 and Over •STI 50 and Over •OA Under 50 -STI Under 50 Injury Type & Age Group (Main Effects Model) 2 3 Initial Pain Frequency "OA over 50 -STI over 50 -OA under 50 •STIunder 50 Injury Type & Age Group Interaction Model (AgeGrp+IPF+lnjTyp*IPF) 1 2 Initial Pain Frequency - O A over50 •STI over 50 •OA under 50 •STI under 50 89 10.5.2.6 Significant single contrasts, controlling for Initial Restr'n of Movement Observed one unit improvement in ROM Right after treatment, n = 567 Initial Restriction of Movement 1 2 3 4 50 and over Proportion with improvement 20% 22% 37% 38% Cell count 10 50 70 145 Under age 50 Proportion with improvement 25% 21% 25% 33% Cell count 12 38 79 126 Improvement in Restr. of Move't Right after treatment Main Effects Model Age Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 567 0.26 0.19 0.16 1.30 0.90 1.89 Initial Restriction of Movement 0.44 0.10 <0.01 1.55 1.28 1.88 Constant -2.42 0.35 <0.01 0.09 -2Lc glikelihood: 659.77 DF = 3 Improvement in Restr. of Move't Right after treatment Interaction Model Age Group + ROM + Age Group *ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 567 -0.03 0.67 0.97 0.97 0.26 3.65 Initial Restriction of Movement 0.39 0.15 <0.01 1.47 1.11 1.96 Age Group * Initial RestOfMove't 0.09 0.20 0.65 1.09 0.74 1.60 Constant -2.25 0.50 <0.01 0.11 -2Loglikelihood: 659.57 DF = 4 Comparison of observations with predictions 0.4 -Age Group (Observed Data) 0.3 -0.2 C ) 1 1 0.1 -—O—50 and over Under age 50 0 0 1 Initial Restriction of Movement 4 90 Initial Restriction of Movement with Injury Type Group Observed one unit improvement in ROM Right after treatment, n = 484 Initial Restriction of Movement 1 2 3 4 O A Proport ion with improvement 3 3 % 1 9 % 3 4 % 4 1 % Cel l count 12 4 3 62 109 STI Proport ion with improvement 14% 2 8 % 2 4 % 2 5 % Ce l l count 7 40 66 119 Improvement in Restr. of Move't Right after treatment Main Effects Model Injury Type + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 484 -0.49 0.21 0.02 0.61 0.41 0.92 Initial Restriction of Movement 0.36 0.10 <0.01 1.43 1.17 1.76 Constant -1.84 0.36 <0.01 0.16 -2Lc glikelihood: 556.85 DF = 3 Improvement in Restr. of Move't Right after treatment Interaction Model Injury Type + ROM + Injury Type* ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 484 0.34 0.71 0.63 1.40 0.35 5.63 Initial Restriction of Movement 0.47 0.15 <0.01 1.60 1.21 2.13 lnjuryType*lnitialRestOfMove't -0.25 0.21 0.22 0.78 0.52 1.17 Constant -2.20 0.49 <0.01 0.11 -2Loglikelihood: 555.39 DF = 4 Comparison of observations with predictions 0.0 Injury Type (Observed Data) 1 2 3 4 Initial Restriction of Movement Injury Type (Main Effects Model) 1 2 3 Initial Restriction of Movement Injury Type (Interaction Model) 1 2 3 4 Initial Restriction of Movement 91 Restriction of Movement with Age of Injury Group Observed one unit improvement in ROM Right after treatment, n = 566 Initial Restriction of Movement 1 2 3 4 4 + years Proport ion with improvement 14% 1 5 % 3 3 % 3 2 % Ce l l count 14 46 82 142 < 4 years Proport ion with improvement 3 3 % 3 2 % 2 9 % 4 0 % Ce l l count 9 41 66 129 Improvement in Restr. of Move't Right after treatment Main Effects Model Age of Injury Group + ROM n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 484 -0.30 0.19 0.11 0.74 0.51 1.07 Initial Restriction of Movement 0.43 0.10 <0.01 1.54 1.28 1.86 Constant -2.08 0.34 <0.01 0.12 -2Loglikelihood: 662.23 DF = 3 Improvement in Restr. of Move't Right after treatment Interaction Model Age of Injury Group + ROM + Age of Injury Group*ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 484 -0.46 0.67 0.49 0.63 0.17 2.32 Initial Restriction of Movement 0.41 0.13 <0.01 1.51 1.16 1.95 InjAgeGrp* InitialRestOfMove't 0.05 0.19 0.79 1.05 0.72 1.54 Constant -2.00 0.45 <0.01 0.13 -2Loglikelihood: 662.16 DF = 4 Comparison of observations with predictions Age of injury (Observed Data) 1 2 3 4 Initial Restriction of Movement Age of Injury (Main Effects Model) 1 2 3 4 Initial Restriction of Movement Age of Injury (Interaction Model) 1 2 3 4 Initial Restriction of Movement 92 10.5.2.7 Testing full main effects and interaction models: Restr'n of Movement Observed one unit improvement in ROM Right after treatment, n = 478 Initial Restriction of Movement 1 2 3 4 O A < 4 years Proport ion with improvement 0% 50% 25% 42% Cel l count 4 4 16 19 O A 4 + years Proport ion with improvement 0% 25% 17% 30% Cel l count 8 8 24 43 STI < 4 years Proport ion with improvement 0% 25% 38% 22% Cel l count 4 4 24 41 STI 4 + years Proport ion with improvement 0% 0% 13% 30% Cel l count 3 3 16 23 Improvement in Restr. of Move't Right after treatment Main Effects Model InjAgeGrp + InjType + IROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 478 -0.59 0.22 <0.01 0.55 0.36 0.84 Age of Injury Group -0.46 0.21 0.03 0.63 0.42 0.96 Initial Restriction of Movement 0.36 0.10 <0.01 1.43 1.16 1.75 Constant -1.53 0.38 <0.01 0.22 -2Loglikelihood: 548.84 DF = 4 Improvement in Restr. of Move't Right after treatment Interaction Model # 1 InjAgeGrp + InjType + IROM + lnjType*IROM n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group -0.46 0.21 0.03 0.63 0.41 0.96 Initial Restriction of Movement 0.44 0.11 O.01 1.55 1.25 1.92 Injury Type*lnitialRestOf Move't -0.19 0.06 <0.01 0.83 0.73 0.94 Constant -1.79 0.37 <0.01 0.17 -2Loglikelihood: 547.68 DF = 4 Comparison of observations with predictions Injury Type & Injury Age Group (Observed Data) 0 1 2 3 Initial Restriction of M o v e m e n t " O A 4 + years • S T I 4 + years • O A < 4 years • S T I < 4 years Injury Type & Injury Age Group (Main Effects Model) 1 2 3 Initial Restriction of Movement - O A 4 + years •STI 4 +years - O A <4 years •STI < 4 years Injury Type & Injury Age Group Interaction Mode l (I n j A g e Gr p+l ROM + lnjTyp*IROM) 1 2 3 Initial Restriction of M o v e m e n t - O A 4 + years • S T 14 + years • O A < 4 years • S T I < 4 years 93 10.5.3 S ix Weeks After Treatment 10.5.3.1 Univariable models - each predictor modeled separately Pain Intensity Six Weeks after treatment -Univariable Models n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 561 0.35 0.18 0.05 1.42 1.00 2.00 Age Group (Age 50+) 554 0.06 0.17 0.71 1.06 0.76 1.49 BMI Group (Overweight 25+) 514 -0.03 0.18 0.86 0.97 0.68 1.38 Injury Type (STI) 465 -0.33 0.19 0.08 0.72 0.50 1.04 Age of Injury Group (4+ years) 554 0.02 0.17 0.93 1.02 0.73 1.42 Initial Pain Intensity (raw score zero if no symptoms present, higher as symptoms worsen) 561 0.57 0.12 <0.01 1.77 1.40 2.25 Pain Frequency Six Weeks after treatment -Univariable Models n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 561 0.40 0.18 0.03 1.49 1.05 2.12 Age Group (Age 50+) 554 0.06 0.17 0.71 1.07 0.76 1.49 BMI Group (Overweight 25+) 514 0.01 0.18 0.94 1.01 0.71 1.44 Injury Type (STI) 465 -0.16 0.19 0.39 0.85 0.59 1.23 Age of Injury Group (4+ years) 554 -0.08 0.17 0.65 0.93 0.66 1.30 Initial Pain Frequency (raw score zero if no symptoms present, higher as symptoms worsen) 561 0.52 0.12 <0.01 1.68 1.33 2.14 Restriction of Movement Six Weeks after treatment -Univariable Models n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 548 0.17 0.18 0.34 1.19 0.84 1.69 Age Group (Age 50+) 542 0.15 0.17 0.38 1.16 0.83 1.63 BMI Group (Overweight 25+) 504 0.22 0.18 0.22 1.25 0.88 1.78 Injury Type (STI) 465 -0.11 0.19 0.56 0.90 0.62 1.29 Age of Injury Group (4+ years) 541 -0.15 0.17 0.38 0.86 0.61 1.21 Initial Restriction of Movement (raw score zero if no symptoms present, higher as symptoms worsen) 548 0.61 0.09 <0.01 1.85 1.54 2.21 94 10.5.3.2 Significant single contrasts, controlling for Initial Pain Intensity Observed one unit improvement in PI Six Weeks after treatment, n = 561 Initial Pain Intensity 1 2 3 4 Female Proportion with improvement 25% 46% 54% 68% Cell count 4 26 150 174 Male Proportion with improvement 0% 29% 59% 54% Cell count 7 24 101 72 Improvement in Pain Intensity Six Weeks after treatment Main Effects Model Sex + PI n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 561 0.22 0.18 0.22 1.25 0.88 1.79 Initial Pain Intensity 0.55 0.12 <0.01 1.74 1.37 2.21 Constant -1.68 0.41 <0.01 0.19 -2Loglikelihood: 741 .4 D F = 3 Improvement in Pain Intensity Six Weeks after treatment Interaction Model Sex + PI + Sex*PI n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 561 0.16 0.82 0.84 1.18 0.24 5.89 Initial Pain Intensity 0.54 0.19 <0.01 1.72 1.19 2.48 Sex* Initial Pain Intensity 0.02 0.25 0.94 1.02 0.63 1.65 Constant -1.64 0.61 <0.01 0.19 -2Loglikelihood: 741 .4 D F = 4 Comparison of observations with predictions 2 3 Initial Pain Intensity 2 3 Initial Pain Intensity 95 Pain Intensity with Injury Type Observed one unit improvement in PI Six Weeks after treatment, n = 465 Initial Pain Intensity 1 2 3 4 O A Proportion with improvement 13% 56% 56% 74% Cell count 8 27 104 90 STI Proportion with improvement 0% 22% 59% 54% Cell count 3 18 103 111 Improvement in Pain Intensity Six Weeks after treatment Main Effects Model Injury Type + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 465 -0.44 0.19 0.02 0.64 0.44 0.94 Initial Pain Intensity 0.53 0.13 <0.01 1.70 1.32 2.21 Constant -1.23 0.44 <0.01 0.29 -2Loglikelihood: 614.6 DF = 3 Improvement in Pain Intensity Six Weeks after treatment Interaction Model InjuryType + PI + lnjuryType*PI n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 465 0.56 0.89 0.53 1.76 0.31 10.09 Initial Pain Intensity 0.67 0.18 <0.01 1.95 1.37 2.79 InjuryType* Initial Pain Intensity -0.31 0.27 0.25 0.74 0.44 1.24 Constant -1.66 0.59 <0.01 0.19 -2Loglikelihood: 613.27 DF = 4 Comparison of observations with predictions Injury Type (Observed Data) — O — O A • STI 2 3 Initial Pain Intensity 0.8 0.7 0.6 0.5 0.4 0.3 I 0.2 0.1 0.0 Injury Type (Main Effects Model) 2 3 Initial Pain Intensity Injury Type (Interaction Model) 2 3 Initial Pain Intensity 96 10.5.3.3 Testing full main effects and interaction models: Pain Intensity Pain Intensity with Injury Type and Age Group Observed one unit improvement in PI Six Weeks after treatment, n = 465 Initial Pain Intensity O A Fema le Proport ion with improvement 0% 33% 53% 47% Cel l count 15 O A Male Proport ion with improvement 0% 0% 58% 67% Cel l count 12 STI Fema le Proport ion with improvement 0% 0% 40% Cel l count 59% 10 STI Male Proport ion with improvement 0% 0% 0% Cel l count 59% 8 37 Improvement in Pain Intensity Six Weeks after treatment Main Effects Model Sex + Injury Type + PI n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% C.I. for Exp(Beta) Lower Upper Sex 0.14 0.20 0.47 1.16 0.78 1.71 Injury Type -0.46 0.20 0.02 0.63 0.43 0.93 Initial Pain Intensity 465 0.52 0.13 <0.01 1.69 1.30 2.19 Constant -1.27 0.44 <0.01 0.28 -2Loglikelihood: 614.08 DF = 4 Improvement in Pain Intensity Six Weeks after treatment Interaction Model IPI + Sex*lnjury Type + Injury Type*PI n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% C.I. for Exp(Beta) Lower Upper Initial Pain Intensity 0.62 0.14 <0.01 1.86 1.41 2.44 Sex*Injury Type Injury Type* Initial Pain Intensity Constant 465 0.49 0.29 0.09 1.63 0.92 2.86 -0.24 0.08 <0.01 0.79 0.67 0.92 -1.48 0.44 <0.01 0.23 -2Loglikelihood:i 610.80 DF = 4 Comparison of observations with predictions Injury Type & Sex (Observed i F e m a l e -STI F e m a l e Injury Type & A g e Group Interaction 1 (Main+lnjType*IPI) 2 3 Initial Pain Intensity " O A F e m a l e " S T I F e m a l e - O A M a l e •STI M a l e 97 10.5.3.4 Significant single contrasts, controlling for Initial Pain Frequency Observed one unit improvement in PI Six Weeks after treatment, n = 561 Initial Pain Frequency 1 2 3 4 Female Proport ion with improvement 2 5 % 1 7 % 4 8 % 5 3 % Ce l l count 4 35 118 197 Male Proport ion with improvement 0% 2 9 % 3 6 % 4 3 % Ce l l count 3 24 77 99 Improvement in Pain Frequency Six Weeks after treatment Main Effects Model Sex + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 561 0.35 0.18 0.05 1.42 0.99 2.03 Initial Pain Frequency 0.51 0.12 <0.01 1.66 1.31 2.11 Constant -2.19 0.44 <0.01 0.11 -2Loglikelihood: 745.82 DF = 3 Improvement in Pain Frequency Six Weeks after treatment Interaction Model Sex + PI + Sex*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for 3eta) Lower Upper Sex 561 -0.08 0.87 0.93 0.92 0.17 5.07 Initial Pain Frequency 0.43 0.19 0.03 1.54 1.05 2.25 Sex* Initial Pain Frequency 0.13 0.25 0.61 1.13 0.70 1.84 Constant -1.93 0.67 <0.01 0.14 -2Loglikelihood: 745.57 DF = 4 Comparison of observations with predictions Sex (Observed Data) 2 3 Initial Pain Frequency Sex (Main Effects Model) 1 2 3 Initial Pain Frequency 98 10.5.3.5 Significant single contrasts, controlling for Initial Restriction of Movement Observed one unit improvement in ROM Six Weeks after treatment, n = 504 Initial Restriction of Movement 1 2 3 4 Overweight (BMI 25 +) Proport ion with improvement 38% 32% 60% 53% Cel l count 25 58 119 Not overweight (BMI < 25) Proport ion with improvement 0% 28% 49% 59% Cel l count 11 54 72 126 Improvement in Restr. of Move't Six Weeks after treatment Main Effects Model BMI Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 504 0.09 0.19 0.62 1.10 0.76 1.59 Initial Restriction of Movement 0.65 0.10 <0.01 1.91 1.57 2.31 Constant -2.23 0.34 <0.01 0.11 -2Loglikelihood: 640.71 DF = 3 Improvement in Restr. of Move't Six Weeks after treatment Interaction Model BMIGrp + ROM +BMIGrp*ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 504 1.34 0.67 0.05 3.80 1.02 14.15 Initial Restriction of Movement 0.82 0.14 <0.01 2.27 1.72 2.99 BMIGrp*lnitialRestOfMove't -0.38 0.20 0.05 0.68 0.46 1.01 Constant -2.78 0.47 <0.01 0.06 -2Loglikelihood: 637.00 DF = 4 Comparison of observations with predictions BMI (Observed Data) 1 2 3 4 Initial Restriction of Movement BMI (Main Effects Model) 1 2 3 Initial Restriction of Movement BMI (Interaction Model) -Overweight (BMI25+) •Not overweight (BMI < 25) 1 2 3 4 Initial Restriction of Movement 99 10.5.4 Six Months After Treatment 10.5.4.1 Univariable models - each predictor modeled separately Pain Intensity Six Months after treatment -Univariable Models n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 444 0.54 0.20 <0.01 1.72 1.16 2.55 Age Group (Age 50+) 437 -0.30 0.20 0.12 0.74 0.50 1.09 BMI Group (Overweight 25+) 407 -0.29 0.20 0.16 0.75 0.50 1.12 Injury Type (STI) 358 0.10 0.22 0.64 1.11 0.72 1.69 Age of Injury Group (4+ years) 437 -0.43 0.20 0.03 0.65 0.44 0.96 Initial Pain Intensity (raw score zero if no symptoms present, higher as symptoms worsen) 444 0.70 0.14 O.01 2.02 1.53 2.66 Pain Frequency Six Months after treatment -Univariable Models n (out of 599) Beta S .E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 444 0.40 0.20 0.05 1.49 1.00 2.20 Age Group (Age 50+) 437 -0.02 0.19 0.93 0.98 0.68 1.43 BMI Group (Overweight 25+) 407 -0.15 0.20 0.44 0.86 0.58 1.27 Injury Type (STI) 358 0.07 0.21 0.74 1.07 0.71 1.63 Age of Injury Group (4+ years) 437 -0.27 0.19 0.17 0.77 0.52 1.12 Initial Pain Frequency (raw score zero if no symptoms present, higher as symptoms worsen) 444 0.42 0.13 <0.01 1.53 1.18 1.98 Restriction of Movement Six Months after treatment -Univariable Models n (out of 599) Beta S . E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 431 0.13 0.20 0.52 1.14 0.76 1.71 Age Group (Age 50+) 425 -0.24 0.20 0.23 0.79 0.54 1.16 BMI Group (Overweight 25+) 396 -0.04 0.21 0.86 0.96 0.64 1.44 Injury Type (STI) 358 0.07 0.21 0.76 1.07 0.70 1.62 Age of Injury Group (4+ years) 424 -0.30 0.20 0.13 0.74 0.50 1.09 Initial Restriction of Movement (raw score zero if no symptoms present, higher as symptoms worsen) 431 0.73 0.10 <0.01 2.09 1.71 2.55 100 10.5.4.2 Significant single contrasts, controlling for Initial Pain Intensity Observed one unit improvement in PI Six Months after treatment, n = 444 Initial Pain Intensity 1 2 3 4 Female Proportion with improvement 0% 41% 58% 78% Cell count 3 22 119 139 Male Proportion with improvement 20% 38% 59% 53% Cell count 5 21 82 51 Improvement in Pain Intensity Six Months after treatment Main Effects Model Sex + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex AAA 0.39 0.21 0.06 1.47 0.98 2.22 Initial Pain Intensity 0.66 0.14 <0.01 1.94 1.47 2.57 Constant -1.96 0.47 <0.01 0.14 -2Loglikelihood: 563.87 DF = 3 Improvement in Pain Intensity Six Months after treatment Interaction Model Sex + PI + Sex*PI n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 444 -1.50 0.94 0.11 0.22 0.04 1.39 Initial Pain Intensity 0.33 0.21 0.11 1.39 0.93 2.09 Sex* Initial Pain Intensity 0.59 0.28 0.04 1.80 1.03 3.15 Constant -0.93 0.66 0.16 0.39 -2Loglikelihood: 559.6 DF = 4 Comparison of observations with predictions 2 3 Initial Pain Intensity Sex (Interaction Model) 2 3 Initial Pain Intensity 101 Pain Intensity with Age Group Observed one unit improvement in PI Right after treatment, n = 437 Initial Pain Intensity 1 2 3 4 50 and over Proportion with improvement 14% 33% 55% 71% Cell count 7 27 102 96 Under age 50 Proportion with improvement 0% 50% 62% 71% Cell count 1 14 97 91 Improvement in Pain Intensity Six Months after treatment Main Effects Model Age Group + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 437 -0.23 0.20 0.27 0.80 0.53 1.19 Initial Pain Intensity 0.70 0.14 <0.01 2.01 1.52 2.66 Constant -1.71 0.50 <0.01 0.18 -2Loglikelihood: 556.48 DF = 3 Improvement in Pain Intensity Six Months after treatment Interaction Model AgeGroup + PI + AgeGroup*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 437 -1.09 0.98 0.26 0.34 0.05 2.28 Initial Pain Intensity 0.54 0.22 0.02 1.72 1.11 2.66 Age Group* Initial Pain Intensity 0.26 0.29 0.37 1.30 0.73 2.31 Constant -1.19 0.75 0.11 0.30 -2Loglikelihood: 555.66 DF =4 Comparison of observations with predictions Age Group (Interaction Model) 2 3 Initial Pain Intensity 102 Pain Intensity with BMI Group Observed one unit improvement in PI Six Months after treatment, n = 407 Initial Pain Intensity 1 2 3 4 Not overweight (BMI < 25) Proport ion with improvement 0 % 3 3 % 6 3 % 7 5 % Ce l l count 5 21 115 89 Overweight (BMI 25 +) Proport ion with improvement 3 3 % 4 1 % 5 1 % 6 5 % Cel l count _ 3 J 17 77 78 Improvement in Pain Intensity Six Months after treatment Main Effects Model BMI Group + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 407 -0.37 0.21 0.08 0.69 0.45 1.04 Initial Pain Intensity 0.75 0.15 <0.01 2.11 1.58 2.84 Constant -1.86 0.50 <0.01 0.16 -2Loglikelihood: 518.16 DF = 3 Improvement in Pain Intensity Six Months after treatment Interaction Model BMI Group + PI + BMI Group*PI n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 407 0.99 1.01 0.33 2.68 0.37 19.40 Initial Pain Intensity 0.94 0.21 <0.01 2.55 1.69 3.85 BMIGroup * Initial Pain Intensity -0.42 0.30 0.17 0.66 0.36 1.19 Constant -2.45 0.68 O . 0 1 0.09 -2Loglikelihood: 516.25 DF = 4 Comparison of observations with predictions -Overweight (BMI 25 +) •Not overweight (BMI < 25) 2 3 InitiaI Pain Intensity BMI (Main Effects Model) -Overweight (BMI 25+) • Not overweight (BMI < 25) 1 2 3 Initial Pain Intensity 0.1 0.0 •Overweight (BMI25 +) •Not overweight (BMI < 25) 1 2 3 Initial Pain Intensity 103 Pain Intensity with Age of Injury Observed one unit improvement in PI Six Months after treatment, n = 437 Initial Pain Intensity 1 2 3 4 < 4 years Proportion with improvement 33% 50% 60% 82% Cell count 3 22 100 72 4 + years Proportion with improvement 0% 21% 57% 63% Cell count 4 19 100 115 Improvement in Pain Intensity Six Months after treatment Main Effects Model Age of Injury Group + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 437 -0.57 0.21 <0.01 0.56 0.37 0.85 Initial Pain Intensity 0.76 0.15 <0.01 2.14 1.60 2.86 Constant -1.73 0.49 <0.01 0.18 -2Loglikelihood: 551.62 DF = 3 Improvement in Pain Intensity Six Months after treatment Interaction Model InjAgeGrp + PI + lnjAgeGrp*PI n (out of 599) Beta S .E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 437 -0.23 0.98 0.82 0.79 0.12 5.48 Initial Pain Intensity 0.82 0.23 <0.01 2.27 1.46 3.54 InjAgeGrp * Initial Pain Intensity -0.11 0.30 0.72 0.90 0.50 1.61 Constant -1.92 0.72 O.01 0.15 -2Loglikelihood: 551.49 DF = 4 Comparison of observations with predictions 0.9 -0.8 -Age of injury (Observed Data) 0.7 -0.6 -0.5 -0.4 -0.3-0.2 -0.1 -- ' = — 4 + years ^ ^ ™ < 4 years 0.0^ 2 Initia I P 3 3in Intensity 4 2 3 Initial Pain Intensity Age of Injury (Interaction Model) 2 3 Initial Pain Intensity 104 10.5.4.3 Testing full main effects and interaction models: Pain Intensity Observed one unit improvement in PI Six Months after treatment, n = 400 Initial Pain Intensity 1 2 3 4 < 4 years, not overweight Proportion with improvement 0% 0% 50% 67% Cell count 2 1 2 10 4 + years, not overweight Proportion with improvement 0% 0% 10% 58% Cell count 2 1 2 10 < 4 years, overweight Proportion with improvement 0% 100% 40% 48% Cell count 1 0 1 10 4 + years, overweight Proportion with improvement 0% 0% 33% 52% Cell count 2 2 6 44 Improvement in Pain Intensity Six Months after treatment Main Effects Model BMI Group + InjAgeGrp + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 400 -0.34 0.22 0.11 0.71 0.46 1.08 Age of Injury Group -0.55 0.22 0.01 0.58 0.38 0.89 Initial Pain Intensity 0.81 0.16 O.01 2.24 1.65 3.05 Constant -1.77 0.52 <0.01 0.17 -2Loglikelihood: 503.71 DF = 4 Improvement in Pain Intensity Six Months after treatment Interaction Model: BMIGrp + InjAgeGrp + BMIGrp(lnjAgeGrp + PI) n (out of 599) Beta S .E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper BMI Group 400 -0.81 0.30 <0.01 0.44 0.25 0.79 Age of Injury Group 0.90 0.17 <0.01 2.47 1.78 3.43 Initial Pain Intensity 0.61 0.44 0.17 1.84 0.78 4.35 InjuryType* Initial Pain Intensity -0.22 0.10 0.03 0.80 0.66 0.98 Constant -1.92 0.52 <0.01 0.15 -2Loglikelihood: 501.15 DF = 5 Comparison of observations with predictions Initial Pain Intensity -4 + years, overweight -4 + years, not overweight •< 4 years, overweight •< 4 years, not overweight -4 + years, overweight » 4 + years, not overweight •< 4 years, overweight »<4 years, not overweight 1 . 0 0 . 9 H 0 . 8 0 . 7 0 . 6 0.5 0 . 4 H 0 . 3 0 . 2 0 . 1 0 . 0 BMI & A g e of Injury Inte raction (I njAge+IPI + BMI*lnjAge+BMI*IPI) Initial P a i n Intensity -4 + years, overweight - 4 + years, not overweight •< 4 years, overweight •<4 years, not overweight 105 10.5.4.4 Significant single contrasts, controlling for Initial Pain Frequency Pain Frequency with Age of Injury Observed one unit improvement in PF Six Months after treatment, n = 444 Initial Pain Frequency 1 2 3 4 All Patients Proportion with improvement 0% 29% 49% 51% Cell count 5 48 150 239 Improvement in Pain Frequency Six Months after treatment Main Effect Model (no interaction-models survived) Initial PF as sole predictor n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Pain Frequency 0.42 0.13 <0.01 1.53 1.18 1.98 Constant -1.54 0.46 <0.01 0.21 -2Loglikelihood = 603.73: DF = 2 Comparison of observations with predictions 106 Pain Frequency with Sex Observed one unit improvement in PF Six Months after treatment, n = 444 Initial Pain Frequency 1 2 3 4 Female Proportion with improvement 0% 28% 53% 56% Cell count 3 29 91 160 Male Proportion with improvement 0% 32% 44% 43% Cell count 2 19 59 79 Improvement in Pain Frequency Six Months after treatment Main Effects Model Sex + PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 444 0.36 0.20 0.07 1.44 0.97 2.14 Initial Pain Frequency 0.41 0.13 <0.01 1.51 1.16 1.96 Constant -1.73 0.48 <0.01 0.18 -2Loglikelihood: 600.44 DF = 3 Improvement in Pain Frequency Six Months after treatment Interaction Model Sex + PF + Sex*PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 444 -0.40 0.95 0.67 0.67 0.10 4.30 Initial Pain Frequency 0.27 0.21 0.21 1.31 0.86 1.99 Sex* Initial Pain Frequency 0.23 0.27 0.41 1.25 0.73 2.14 Constant -1.25 0.74 0.09 0.29 -2Loglikelihood: 599.77 DF = 4 Comparison of observations with predictions 0.6 -, Sex (Observed Data) 1 2 3 Initial Pain Frequency 0.6 -, Sex (Main Effects Model) 1 2 3 Initial Pain Frequency 0.6 Sex (Interaction Model) 0.1 A 0.0 •Female •Male 1 2 3 Initial Pain Frequency 107 Pain Frequency with Age of Injury Observed one unit improvement in PF Six Months after treatment, n = 437 1 2 3 4 < 4 years Proport ion with improvement 0 % 2 8 % 4 9 % 5 9 % Ce l l count 0 25 74 98 4 + years Proport ion with improvement 0 % 3 3 % 4 8 % 4 6 % Ce l l count 5 21 73 139 Initial Pain Frequency Improvement in Pain Frequency Six Months after treatment Main Effects Model Age of Injury Group + PF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 437 -0.31 0.20 0.11 0.73 0.50 1.07 Initial Pain Frequency 0.44 0.14 <0.01 1.55 1.19 2.02 Constant -1.42 0.48 <0.01 0.24 -2Loglikelihood: 591.72 DF = 3 Improvement in Pain Frequency Six Months after treatment Interaction Model InjAgeGrp + PF + lnjAgeGrp*PF n (out of 599) Beta S . E . p-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 437 0.69 0.95 0.47 1.98 0.31 12.85 Initial Pain Frequency 0.60 0.21 <0.01 1.82 1.21 2.74 InjAgeGrp * Initial Pain Frequency -0.29 0.27 0.28 0.75 0.44 1.28 Constant -1.98 0.72 <0.01 0.14 -2Loglikelihood: 555.66 DF = 4 Comparison of observations with predictions 0.7 Age of Injury (Observed Data) 1 2 3 4 Initial Pain Frequency 0.7 Age of Injury (Main Effects Model) 1 2 3 Initial Pain Frequency 0.7 Age of Injury (Interaction Model) 2 3 Initial Pain Frequency 108 10.5.4.5 Testing full main effects and interaction models: Pain Frequency Observed one unit improvement in PF Six Months after treatment, n = 437 Initial Pain Frequency 1 2 3 4 <4 years Fema le Proport ion with improvement 0 % 2 0 % 5 2 % 6 5 % Cel l count 0 15 44 71 <4 years Male Proport ion with improvement 0 % 4 0 % 4 3 % 4 4 % Cel l count 0 10 30 27 4+ years Fema le Proport ion with improvement 0 % 3 8 % 5 1 % 4 8 % Cel l count 3 13 45 87 4+ years Ma le Proport ion with improvement 0 % 2 5 % 4 3 % 4 2 % Ce l l count 0 2 8 28 Improvement in Pain Frequency Six Months after treatment: Main Effects Model Sex + InjuryAgeGrp + IPF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp| C.I. for Beta) Lower Upper Sex 437 0.35 0.20 0.08 1.43 0.96 2.13 Age of Injury Group -0.30 0.20 0.13 0.74 0.5C 1.09 Initial Pain Frequency 0.42 0.14 <0.01 1.53 1.17 1.99 Constant -1.62 0.49 <0.01 0.20 -2Loglikelihood: 588.68 DF = 4 Improvement in Pain Frequency Six Months after treatment: Interaction IPF+Sex*IPF + In jAgeGrplPF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Pain Frequency 437 0.41 0.15 <0.01 1.51 1.13 2.01 Sex* Initial Pain Frequency 0.11 0.06 0.07 1.11 0.99 1.25 InjAgeGrp* Initial PainFrequency -0.09 0.06 0.10 0.91 0.81 1.02 Constant -1.56 0.47 <0.01 0.21 -2Loglikelihood: 587.71 DF = 4 Comparison of observations with predictions 0 . 7 - Sex & Inj Age Group (Observed Data) 0 . 6 -0 . 5 -0 . 4 -0 . 3 -0 . 2 -—C>—4 + y e a r s F e m a l e 0 . 1 - -^>—4 + y e a r s M a l e < 4 y e a r s F e m a l e o.oC • < 4 y e a r s M a l e • i i 1 2 3 4 Initial Pain Frequency 0 . 7 Sex & Inj Age Group (Main Effects Model) 0 . 3 0 . 2 0 . 1 0 . 0 4 + y e a r s . F e m a l e •4 + y e a r s , M a l e '< 4 y e a r s . F e m a l e '< 4 y e a r s , M a l e 2 3 Initial Pain Frequency 0 7 Sex & Inj Age Group Interaction Model 4 + y e a r s , F e m a l e 4 + y e a r s , M a l e < 4 y e a r s , F e m a l e < 4 y e a r s , M a l e 1 2 3 Initial Pain Frequency 109 10.5.4.6 Significant single contrasts, controlling for Initial Restr'n of Movement Restriction of Movement with Age Group Observed one unit improvement in ROM Six Months after treatment, n = 425 Initial Restriction of Movement 1 2 3 4 50 and over Proport ion with improvement 33% 39% 58% 68% Cel l count 9 38 50 112 Under age 50 Proport ion with improvement 43% 41% 65% 73% Cel l count 7 29 54 98 Improvement in Restr. of Move't Six Months after treatment Main Effects Model Age Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 425 -0.25 0.21 0.24 0.78 0.51 1.18 Initial Restriction of Movement 0.73 0.10 <0.01 2.07 1.69 2.53 Constant -1.79 0.35 O.01 0.17 -2Loglikelihood: 514.60 DF = 3 Improvement in Restr. of Move't Six Months after treatment Interaction Model Age Group + ROM + Age Group *ROM n (out of 599) Beta S .E . P -value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 425 -0.07 0.68 0.91 0.93 0.25 3.49 Initial Restriction of Movement 0.76 0.15 <0.01 2.14 1.58 2.88 Age Group * Initial RestOfMove't -0.06 0.21 0.78 0.94 0.63 1.41 Constant -1.88 0.50 <0.01 0.15 -2Loglikelihood: 514.52 DF = 4 Comparison of observations with predictions 110 Restriction of Movement with Age of Injury Group Observed one unit improvement in ROM Six Months after treatment, n = 424 Initial Restriction of Movement 1 2 3 4 4 + years Proportion with improvement 27% 36% 63% 62% Cell count 11 33 59 114 < 4 years Proportion with improvement 50% 42% 59% 80% Cell count 6 33 44 96 Improvement in Restr. of Move't Six Months after treatment Main Effects Model Age of Injury Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 424 -0.44 0.22 0.04 0.65 0.42 0.99 Initial Restriction of Movement 0.75 0.10 <0.01 2.13 1.74 2.60 Constant -1.78 0.35 <0.01 0.17 -2Loglikelihood: 509.34 D F = 3 Improvement in Restr. of Move't Six Months after treatment Interaction Model Age of Injury Group + ROM + Age of Injury Group*ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 424 0.56 0.69 0.41 1.75 0.46 6.73 Initial Restriction of Movement 0.92 0.16 <0.01 2.52 1.84 3.43 InjAgeGrp* InitialRestOfMove't -0.32 0.21 0.13 0.73 0.48 1.10 Constant -2.28 0.51 <0.01 0.10 -2Loglikelihood: 506.98 D F = 4 Comparison of observations with predictions 0.8 -I 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Age of injury (Observed Data) 1 2 3 4 Initial Restriction of Movement 0.8 0.7 0.6 0.5 0.4 °' 3| 0.2 0.1 0.0 Age of Injury (Main Effects Model) 1 2 3 4 Initial Restriction of Movement 0.8 0.7 0.6 0.5 0.4 0.3 0.2| 0.1 0.0 Age of Injury (Interaction Model) 1 2 3 4! Initial Restriction of Movement 111 10.5.4.7 Testing full main effects and interaction models: Restr'n of Movement Restriction of Movement with Age of Injury Group Observed one unit improvement in ROM Six Months after treatment, n = 424 Initial Restriction of Movement 1 2 3 4 4 + years Proport ion with improvement 27% 36% 63% 62% Cel l count 11 33 59 114 < 4 years Proport ion with improvement 50% 42% 59% 80% Cel l count 6 33 44 96 Improvement in Restr. of Move't Six Months after treatment Main Effects Model Age of Injury Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 424 -0.44 0.22 0.04 0.65 0.42 0.99 Initial Restriction of Movement 0.75 0.10 <0.01 2.13 1.74 2.60 Constant -1.78 0.35 O.01 0.17 -2Loglikelihood: 509.34 DF = 3 Improvement in Restr. of Move't Six Months after treatment Interaction Model Age of Injury Group + ROM + Age of Injury Group*ROM n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Restriction of Movement 424 0.84 0.11 <0.01 2.31 1.85 2.88 InjAgeGrp* InitialRestOfMove't -0.16 0.07 0.02 0.85 0.75 0.97 Constant -1.99 0.34 <0.01 0.14 -2Loglikelihood: 507.65 DF = 3 Comparison of observations with predictions Age of injury (Observed Data) 1 2 3 4 Initial Restriction of Movement Age of Injury (Main Effects Model njAge + IROM) 1 2 3 4 Initial Restriction of Movement Age of Injury (Interaction Model IROM + lnjAge*IROM) 1 2 3 4 Initial Restriction of Movement 112 10.5.5 One Year After Treatment 10.5.5.1 Univariable models - each predictor modeled separately Pain Intensity One Year after treatment -Univariable Models n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 339 0.23 0.24 0.34 1.26 0.79 2.00 Age Group (Age 50+) 334 -0.34 0.24 0.15 0.71 0.45 1.13 BMI Group (Overweight 25+) 303 0.17 0.25 0.48 1.19 0.73 1.94 InjuryType (STI) 265 -0.08 0.26 0.75 0.92 0.55 1.54 Age of Injury Group (4+ years) 333 -0.05 0.23 0.81 0.95 0.60 1.50 Initial Pain Intensity (raw score zero if no symptoms present, higher as symptoms worsen) 339 0.64 0.16 <0.01 1.89 1.38 2.58 Pain Frequency One Year after treatment -Univariable Models n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 339 0.05 0.23 0.83 1.05 0.67 1.63 Age Group (Age 50+) 334 -0.39 0.22 0.08 0.68 0.44 1.05 BMI Group (Overweight 25+) 303 -0.25 0.24 0.29 0.78 0.49 1.24 Injury Type (STI) 265 0.36 0.25 0.15 1.43 0.88 2.34 Age of Injury Group (4+ years) 333 -0.38 0.22 0.09 0.69 0.44 1.06 Initial Pain Frequency (raw score zero if no symptoms present, higher as symptoms worsen) 339 0.21 0.14 0.15 1.23 0.93 1.64 Restriction of Movement One Year after treatment -Univariable Models n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex (Female) 326 -0.33 0.24 0.17 0.72 0.45 1.15 Age Group (Age 50+) 322 -0.50 0.23 0.03 0.61 0.39 0.96 BMI Group (Overweight 25+) 293 -0.11 0.24 0.66 0.90 0.56 1.45 Injury Type (STI) 265 0.14 0.25 0.59 1.14 0.70 1.88 Age of Injury Group (4+ years) 320 -0.22 0.23 0.33 0.80 0.51 1.26 Initial Restriction of Movement (raw score zero if no symptoms present, higher as symptoms worsen) 326 0.69 0.11 <0.01 2.00 1.61 2.49 10.5.5.2 113 10.5.5.3 Significant single contrasts, controlling for Initial Pain Intensity Observed one unit improvement in PI One Year after treatment, n = 334 Initial Pain Intensity 1 2 3 4 50 and over Proportion with improvement 0% 55% 60% 73% Cell count 3 22 75 78 Under age 50 Proportion with improvement 0% 45% 72% 76% Cell count 1 11 74 67 Improvement in Pain Intensity One Year after treatment Main Effects Model Age Group + PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 334 -0.29 0.24 0.22 0.75 0.46 1.20 Initial Pain Intensity 0.63 0.16 <0.01 1.88 1.37 2.58 Constant -1.19 0.55 0.03 0.30 -2 Log likelihood: 405.85 DF = 3 Improvement in Pain Intensity One Year after treatment Interaction Model AgeGroup + PI + AgeGroup*PI n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 334 -0.28 1.09 0.80 0.75 0.09 6.43 Initial Pain Intensity 0.63 0.26 0.02 1.89 1.13 3.15 Age Group* Initial Pain Intensity -0.00 0.33 0.99 1.00 0.52 1.91 Constant -1.20 0.87 0.17 0.30 -2Loglikelihood: 405.85 DF = 4 Comparison of observations with predictions Age Group (Observed Data) 50 and over Under age 50 2 3 Initial Pain Intensity Age Group (Main Effects Model) -50 and over •Under age 50 2 3 In itia I Pa in Intensity Age Group (Interaction Model) 50 and over Under age 50 2 3 Initial Pain Intensity 114 10.5.5.4 Testing full main effects and interaction models: Pain Intensity Pain Intensity as sole predictor Observed one unit improvement in PI One Year after treatment, n = 336 Initial Pain Intensity 1 2 3 4 All Patients Proportion with improvement 0% 53% 66% 75% Cell count 4 34 151 147 Improvement in Pain Intensity One Year after treatment Main Effects Model (no interaction-models survived) Initial PI as sole predictor n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Pain Intensity 479 0.64 0.16 <0.01 1.89 1.38 2.58 Constant -1.35 0.52 <0.01 0.26 -2Loglikelihood: 412.79 DF = 2 Comparison of observations with predictions Inititial Pain Intensity (Observed Data) 2 3 Initial Pain Intensity Inititial Pain Intensity (Best fit for full and interaction models) 0.2 -All Patients 2 3 Initial Pain Intensity 115 10.5.5.5 Significant single contrasts, controlling for Initial Pain Frequency Observed one unit improvement in PI One Year after treatment, n = 334 Initial Pain Intensity 1 2 3 4 50 and over Proportion with improvement 50% 45% 61% 49% Cell count 2 22 62 92 Under age 50 Proportion with improvement 22% 66% 64% Cell count 0 9 59 85 Improvement in Pain Frequency One Year after treatment Main Effects Model Age Group + PF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for 3eta) Lower Upper Age Group 334 -0.36 0.22 0.11 0.70 0.45 1.08 Initial Pain Intensity 0.19 0.15 0.21 1.21 0.90 1.61 Constant -0.17 0.54 0.75 0.84 -2Loglikelihood: 452.56 DF = 3 Improvement in Pain Frequency One Year after treatment Interaction Model Age Group + PF + Age Group *PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 334 0.91 1.09 0.40 2.49 0.30 20.92 Initial Pain Frequency 0.43 0.25 0.09 1.54 0.94 2.52 Age Group * Initial Pain Freq'cy -0.37 0.31 0.23 0.69 0.37 1.27 Constant -1.01 0.89 0.26 0.37 -2Loglikelihood: 451.11 DF = 4 Comparison of observations with predictions 0.8 n Age Group (Observed Data) 0.0 50 and over Unde rage 50 1 2 3 Initial Pain Frequency 0.8 0.7 Age Group (Main Effects Model) 0.2 0.1 0.0 -50 and over •Under age 50 1 2 3 Initial Pain Frequency 0.8 0.7 Age Group (Interaction Model) 0.2 0.1 0.0 •50 and over •Under age 50 2 3 Initial Pain Frequency 116 Pain Frequency with InjuryType Observed one unit improvement in PF One Year after treatment, n = 265 Initial Pain Frequency 1 2 3 4 O A Proport ion with improvement 0 % 4 2 % 6 3 % 5 3 % Ce l l count 1 19 46 60 STI Proport ion with improvement 3 8 % 6 4 % 6 5 % Ce l l count 0 8 50 80 Improvement in Pain Frequency One Year after treatment Main Effects Model Injury Type + PF n (out of 599) Beta S .E . P-value Exp (Beta) 95.0% Exp( C.I. for 3eta) Lower Upper Injury Type 265 0.30 0.25 0.24 1.35 0.82 2.23 Initial Pain Frequency 0.25 0.18 0.16 1.29 0.91 1.82 Constant -0.65 0.61 0.29 0.52 -2Loglikelihood: 354.90 DF = 3 Improvement in Pain Frequency One Year after treatment Interaction Model InjuryType + PF + lnjuryType*PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Injury Type 265 0.02 1.27 0.99 1.02 0.08 12.34 Initial Pain Frequency 0.22 0.22 0.33 1.25 0.80 1.93 InjuryType* Initial Pain Freq'cy 0.08 0.37 0.82 1.09 0.53 2.23 Constant -0.55 0.75 0.47 0.58 -2Loglikelihood: 354.84 DF = 4 Comparison of observations with predictions Injury Type (Observed Data) 2 3 Initial Pain Frequency 0.7 Injury Type (Main Effects Model) 0.3 0.2 0.1 0.0 O A - S T I 1 2 3 Initial Pain Frequency Injury Type (Interaction Model) 0.3 •] 0.2 J 0.1 0.0 O A - S T I 1 2 3 Initial Pain Frequency 117 Restriction of Movement with Age of Injury Group Observed one unit improvement in ROM One Year after treatment, n = 333 Initial Restriction of Movement 1 2 3 4 < 4 years Proport ion with improvement 3 3 % 6 2 % 7 0 % Ce l l count 0 15 61 69 4 + years Proport ion with improvement 5 0 % 44% 6 5 % 48% Cel l count 2 16 57 110 Improvement in Restr. of Move't One Year after treatment Main Effects Model Age of Injury Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age of Injury Group 333 -0.42 0.23 0.07 0.66 0.42 1.03 Initial Restriction of Movement 0.25 0.15 0.09 1.28 0.96 1.72 Constant -0.34 0.52 0.51 0.71 -2Loglikelihood: 449.84 DF = 3 Improvement in Restr. of Move't One Year after treatment Interaction Model Age of Injury Group + ROM + Age of Injury Group*ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Restriction of Movement 333 2.12 1.08 0.05 8.29 1.00 68.92 Age of Injury Group 0.68 0.24 <0.01 1.97 1.23 3.18 InjAgeGrp* InitialRestOfMove't -0.75 0.31 0.02 0.47 0.25 0.87 Constant -1.76 0.82 0.03 0.17 -2Loglikelihood: 443.82 DF = 4 Comparison of observations with predictions 0.8 Age of injury (Observed Data) 4 + years < 4 years 2 3 Initial Pain Frequency 0.8 Age of Injury (Main Effects Model) 0.0 1 2 3 Initial Pain Frequency 0.8 -| Age of Injury (Interaction Model) 0.0 • 4 + years •< 4 years 1 2 3 Initial Pain Frequency 11 8 10.5.5.6 Testing full main effects and interaction models: Pain Frequency Observed one unit improvement in PI One Year after treatment, n = 334 Initial Pain Intensity 1 2 3 4 50 and over Proportion with improvement 50% 45% 61% 49% Cell count 2 22 62 92 Under age 50 Proportion with improvement 22% 66% 64% Cell count 0 9 59 85 Improvement in Pain Frequency One Year after treatment Main Effects Model Age Group + PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Seta) Lower Upper Age Group 334 -0.36 0.22 0.11 0.70 0.45 1.08 Initial Pain Intensity 0.19 0.15 0.21 1.21 0.90 1.61 Constant -0.17 0.54 0.75 0.84 -2Loglikelihood: 452.56 DF = 3 Improvement in Pain Frequency One Year after treatment Interaction Model Age Group + PF + Age Group *PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Initial Pain Frequency 334 0.26 0.15 0.08 1.30 0.97 1.74 Age Group * Initial Pain Freq'cy -0.12 0.06 0.07 0.89 0.78 1.01 Constant -0.40 0.51 0.43 0.67 -2Loglikelihood: 451.82 DF = 3 Comparison of observations with predictions 0.8 Age Group & Initial PF (Observed Data) 1 2 3 Initial Pain Frequency 0.8 Age Group & Initial PF (Main Effects Model) 0.8 1 2 3 Initial Pain Frequency Age Group & Initial PF (Interaction Model IPF+AgeGrp*IPF) 1 2 3 Initial Pain Frequency 119 10.5.5.7 Significant single contrasts, controlling for Initial Restr'n of Movement Restriction of Movement with Sex Observed one unit improvement in PF One Year after treatment, n = 326 Initial Restriction of Movement 1 2 3 4 Female Proport ion with improvement 22% 61% 60% 71% Cel l count 9 33 57 95 Male Proport ion with improvement 40% 50% 77% 77% Ce l l count 5 18 26 60 Improvement in Restriction of Movement One Year after treatment Main Effects Model Sex + PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 326 -0.31 0.26 0.24 0.74 0.44 1.23 Initial Restriction of Movement 0.69 0.11 <0.01 2.00 1.61 2.49 Constant -1.40 0.39 <0.01 0.25 -2Loglikelihood: 386.50 DF = 3 Improvement in Restriction of Movement One Year after treatment Interaction Model Sex + PF + Sex*PF n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Sex 326 -0.05 0.75 0.95 0.95 0.22 4.16 Initial Restriction of Movement 0.75 0.19 O.01 2.11 1.45 3.08 Sex* Initial RestrOfMovement -0.09 0.24 0.72 0.92 0.58 1.46 Constant -1.57 0.61 0.01 0.21 -2Loglikelihood: 386.37 DF = 4 Comparison of observations with predictions Sex (Observed Data) 1 2 3 4| Initial Restriction of Movement 0.9 i 0.8 -0.7 0.6 0.5 ^ 0.4 0.3 0.2 0.1 0.0 Sex (Main Effects Model) 2 3 4 Initial Restriction of Movement Sex (Interaction Model) 1 2 3 4 Initial Restriction of Movement 120 Restriction of Movement with Age Group Observed one unit improvement in ROM One Year after treatment, n = 322 Initial Restriction of Movement 1 2 3 4 50 and over Proportion with improvement 29% 52% 59% 67% Cell count 7 27 39 82 Under age 50 Proportion with improvement 29% 61% 72% 79% Cell count 7 23 43 71 Improvement in Restr. of Move't One Year after treatment Main Effects Model Age Group + ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 322 -0.54 0.25 0.03 0.58 0.36 0.95 Initial Restriction of Movement 0.70 0.11 <0.01 2.02 1.62 2.52 Constant -1.35 0.38 <0.01 0.26 -2Loglikelihood: 378.55 DF = 3 Improvement in Restr. of Move't One Year after treatment Interaction Model Age Group + ROM + Age Group *ROM n (out of 599) Beta S . E . P-value Exp (Beta) 95.0% Exp( C.I. for Beta) Lower Upper Age Group 322 -0.15 0.73 0.84 0.86 0.21 3.59 Initial Restriction of Movement 0.78 0.17 <0.01 2.17 1.55 3.05 Age Group * Initial RestOfMove't -0.13 0.23 0.56 0.88 0.56 1.37 Constant -1.56 0.54 <0.01 0.21 -2Loglikelihood: 378.21 DF = 4 Comparison of observations with predictions Age Group (Interaction Model) 1 2 3 4 Initial Restriction of Movement 121 10.5.5.8 Testing full main effects and interaction models: Restr'n of Movement Observed one unit improvement in ROM One Year after treatment, n = 322 Initial Restriction of Movement 1 2 3 4 50 and over Female Proportion with improvement 25% 50% 48% 63% Cell count 4 16 27 51 50 and over Male Proportion with improvement 33% 55% 83% 74% Cell count 3 11 12 31 Under age 50 Female Proportion with improvement 20% 71% 70% 80% Cell count 5 17 30 44 Under age 50 Male Proportion with improvement 50% 33% 77% 78% Cell count 2 2 6 13 Improvement in Restr. of Move't One Year after treatment Main Effects Model AgeGrp + IROM n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Seta) Lower Upper Age Group 478 -0.54 0.25 0.03 0.58 0.36 0.95 Initial Restriction of Movement 0.70 0.11 <0.01 2.02 1.62 2.52 Constant -1.35 0.38 <0.01 0.26 -2Loglikelihood: 378.55 DF = 3 Improvement in Restr. of Move't One Year after treatment Interaction Model # 1 IROM + lnjType*IROM n (out of 599) Beta S.E. P-value Exp (Beta) 95.0% Exp( C.I. for Seta) Lower Upper Initial Restriction of Movement 0.71 0.11 <0.01 2.04 1.63 2.56 Injury Type*lnitialRestOfMove't -0.77 0.26 <0.01 0.47 0.28 0.78 Constant -1.41 0.37 <0.01 0.24 -2Loglikelihood: 374.71 DF = 3 Comparison of observations with predictions 122 10.6 Appendix 6 - pivot tables T h e following tables show the percent of patients who s a w improvement, no change , or who got worse , at different intervals fol lowing P S T therapy. Pa in Intensity, Pa in Frequency , and Restr ict ion of Movement initial sco res are compared with sco res immediately after treatment, s i x -weeks after, s ix months after, and one year post-treatment. These tables are as follows: • At least two units of improvement o Overal l • At least one unit of improvement: o Overa l l o Males only o F e m a l e s only o All with osteoarthrit is o All with soft t issue injury o Males with osteoarthrit is o Males with soft t issue injury o F e m a l e s with osteoarthrit is o F e m a l e s with soft t issue injury 123 IB U >-> "<r "j- E r~- cn n in v to PHI* n K - «- CM O E CM JoT 3? !8 3 £ 3 8 w a N O r (M W ' 1 -to |«j D-|o T - CM CO • X 00 — - Q ••- CO ••-w in m r r CD I - J I T - CM ( O N -v N 01 m 10 >-S is 3,2 £ < < £ >> c o c 30-2 | a o f OW CD h-o CN • a O- CO =- cl a = Q 03 = I T ; CO .E C O 0-^ 1— O C") Ol O O r ffl N W ^ E E l a> o> =, § a < g • o o 0) o a: O Q i i o £ <0. tO Ci O K - o 2 3 ™ CO I*. CO mm Figure 20 - At Least Two Units Of Improvement Overall 124 ra o >-o c O - =• a — o S co CM co o O) n N <D « •t- *- CO pas CO CN O CM T - m CO I 8 c o s c •-l"l I a g • on • to 10 lit ' 8 Si *5 si o w w v O <u o >< in « < <» o T T J— 01 to co T -1 to Ol O) CO «f CM tO cJ! o to a Ol CO e CM CI f O cn T - in co m — if) co in oo CM OJ If) o i - M n • I Q O Oil ) O i - CO CN co in Q. |o ••- CNJ co -q- Cj| WW 1 o s * * —• C o I i f I £ i i s < a s « o ct o o ° I § § I f. CO UU - I Figure 21 - At Least One Unit Of Improvement Overall 1 2 5 n o >-o> c O o s CM n o o s *• CD "* CM CM |co in U b : T- N n v CD N Ui CO ID O CD lo lcN «D m I . P I * i to « o CN f- CO 03 n|o T- N n ^ O •o c • M W ^ O 03 _ H I n o l o CM O F 2 < iP a 1 o r*- o> 5 ID] CM N O T-i- CM » JO £ to CD « O O) 2 i O z 1 E 03 > o S c £ = 03 _ 0) 5 & O £ E i i ™ ro o * P i " «. to K ^ ffl _ • CD O) CD i • n TT a • I. , « , O CM CO (D\ Figure 22 - At Least One Unit Of Improvement, Males Only 126 > a c O o x 55 Figure 23 - At Least One Unit Of Improvement, Females Only 127 to o > CD c O y- I*- «0 o M n Tt O |N O N. W CM CD • CM CO O K O C N C O l C c O S x 53 CM (O s s CO I i w CD CM CO T- CM [Oft CO • CN CO CM 8R! CO CO T" CD r» CO CM o * * eo <*- LO C£ iity BE c o iin Int and 1 CL o •>- CM CO TT 1 1 (3,5 •>- CD O CO RI CO K KPT ^ CO CO CO CN I 5 ^ f N •* * O p a 1 co CO LO § •) 1 i n i 5 SS Z • <5S § N i l > o to to ml •C CD 1- COl - CM] CC |p r- CM W ' 0) O) o c a a = 3 ~ tr 2 * o 0. o co S RI = 1 CO CN ^ a : | o i - CM co • Figure 24 - At Least One Unit Of Improvement, All With Soft Tissue Injuries 128 |r- m to en cn i lo o «- co esi '-l —• r n n o f i (3 ( E g 9 a < > V O O OC S O Q §1 to O CL U : l o ^ I N fo ' Figure 25 - At Least One Unit Of Improvement, All With Osteoarthritis ^ oo cn] mm CM CO O) O N - co CM • CN i 5 s 1. lo i ~ CM CO f O i r r- o Tj- CM X O T- CM CO • •>- CN 1 cn CM cn r- CM Ivy^ l oo co )s PS" ™ cn fi' O T - tM CO • O *- CM CO • I F i - I ICNltN CO ico lw coUo o 0. u E a o c Q, 3 O O t- CM CO • CM co 0) r^ . i n I Figure 26 - At Least One Unit Of Improvement, Males With Soft Tissue Injuries 130 ts o > c O f CO CO *r r r CO o led CN 0> |<D IP 1^ -•r- CO i-l m > io K • Ico bdl I i - l o ^ H n y i O s x io E _ CO h-LL. CO i - >r- co m i - co co -CN-3T • CN co rr C?| tu o s & o CM CN CO CM m CM r-o £ g 3-1 <D co co co rr t T i n o i JcT ri- ( W W r-l O r- CM CO rr C3| if CM CM # o ~ CO CO CM III © i. P 1 1 . co O rr s 0 9 i l l r- i _ O l~ —I. » . « = IS i - CN CO rf Q\ Figure 27 - At Least One Unit Of Improvement, Females With Soft Tissue Injuries 131 CO a >-o c O o S (0 0) O) Figure 28 - At Least One Unit Of Improvement, Males With Osteoarthritis 132 Figure 29 - At Least One Unit of Improvement, Females with Osteoarthritis 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

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"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0091735/manifest

Comment

Related Items