We used data files from Medicare provider analysis and review part A to identify beneficiaries who underwent total hip replacement (n=483
970) or total knee replacement (n=873
125) from 2001-5. Medicare is an insurance programme operated by the US government and serves as the primary insurance for all Americans aged 65 or older. Patients were identified using the ICD-9CM (international classification of diseases, ninth revision, clinical modification) procedure codes (81.51 and 81.53 for primary and revision total hip replacement, 81.54 and 81.55 for primary and revision total knee replacement).5 6 20
The part A files contain a range of data collected from discharge abstracts for all fee for service Medicare enrolees admitted to hospital, including patient characteristics, postal code of primary residence, ICD-9CM codes for primary and secondary diagnoses and procedures, admission source (for example, emergency department or transfer from outside hospital), admission and discharge dates, discharge disposition (for example, home, another acute care hospital, death), death occurring up to three years after discharge, each patient’s unique Medicare beneficiary number allowing for identification of readmissions, and each hospital’s unique six digit identification number. We obtained postal code level median household income for each patient by linking the postal code of residence in the Medicare provider analysis and review files to postal code level income data available from the 2000 US census.21
Comorbid illnesses were identified using algorithms described previously22
and updated by other researchers,23
which consider 30 specific conditions and exclude comorbidities that may represent complications of care or that are related to the primary reason for admission to hospital. Additional high risk conditions specific to joint replacement surgery (previous hip or knee replacement, acute fracture, and active joint infection) were identified using methods defined in previous studies using administrative data to assess orthopaedic outcomes, and were considered as additional comorbid conditions.5 24 25
As is customary in studies using administrative data to assess orthopaedic outcomes, we excluded several patient populations from our primary analysis, including patients who underwent joint replacement after transfer from another acute care hospital (n=3764), those with acute fractures (n=41
192), those who had partial hip replacements (n=465), and those who had received multiple joint replacements during the index admission (n=2831). These patient populations are generally excluded from studies assessing orthopaedic outcomes because they are usually heterogeneous and are at especially high risk of adverse outcomes in ways that may not be well captured using administrative data.5 6 17
We excluded patients that had missing data on race (n=5583) and on postal codes as this precluded us from obtaining socioeconomic measures for these patients through linkage with the US census data (n=31
447). We calculated primary and revision total hip replacement and total knee replacement volume as well as aggregate annual joint replacement volume for each hospital by summing the number of procedures done in each hospital in the Medicare provider analysis and review data during each year. By linking the Medicare data to the 2006 American Hospital Association annual survey26 27
we obtained additional hospital characteristics, including whether each hospital was a major teaching centre, whether each hospital was or was not owned by physicians, the total number of annual admissions, and the number of hospital beds. We excluded 24 hospitals (n=1577 patients) that could not be linked to the survey. In total, application of all exclusion criteria resulted in the exclusion of 84
014 patients (6% of the initial sample). The sum of the listed exclusions does not equal the total reported exclusions because some patients had more than one type of missing data.
Measurement of hospital orthopaedic specialisation
We calculated the degree of each hospital’s orthopaedic specialisation by building on methods developed for identification of physician owned specialty hospitals.16 17 28 29
Specifically, for each hospital we calculated the percentage of Medicare admissions classified as major diagnostic criteria 8 (diseases of the musculoskeletal system) during 2001-5; each hospital’s specialisation could range from 0 (no admissions for major diagnostic criteria 8) to 100 (all admissions classified as major diagnostic criteria 8). We used graphical techniques and univariate methods to examine the distribution of orthopaedic specialisation among all US hospitals carrying out major joint replacement. We then stratified the hospitals into fourths, fifths, and 10ths containing equal numbers of hospitals based on their degree of orthopaedic specialisation, with fifths serving as the basis for our primary analyses.
Firstly, we evaluated trends in personal characteristics, socioeconomic status, and prevalence of comorbid illness of patients admitted to hospitals across fifths of orthopaedic specialisation (lowest fifth, least specialised; highest fifth, most specialised). We used logistic regression for dichotomous patient characteristics, such as percentage with diabetes, and linear regression for continuous variables, such as patient’s age, while controlling for clustering of patients within hospitals. Secondly, we used similar statistical methods to compare the characteristics of less specialised with more specialised hospitals. Specifically we compared the annual number of Medicare admissions, annual orthopaedic (major diagnostic criteria 8) admissions, annual joint replacement volumes, number of beds, hospital ownership status, and hospital teaching status between each fifth of orthopaedic specialisation.
Thirdly, we compared rates of adverse outcomes and length of hospital stay across fifths of orthopaedic specialisation for patients treated in hospitals. In particular we examined rates of six separate adverse outcomes occurring during the index admission or within 90 days of surgery that have been used in previous studies using administrative data to assess orthopaedic outcomes, including5 6 25 30 31
pulmonary embolism, deep vein thrombosis, postoperative haemorrhage, deep wound infection, and death, as well as myocardial infarction during the index admission. The primary outcome was a composite representing the occurrence of one or more of these adverse outcomes. The secondary outcome of interest was death within 90 days of the index surgery.
Fourthly, we used multivariable generalised linear models with a logit link to compare the odds of both the composite outcome and mortality with hospital specialisation, from the least specialised hospitals (lowest fifth) to the most specialised hospitals (highest fifth the reference category).32 33
We used these models to compare both the unadjusted odds of adverse outcomes and the adjusted odds of adverse outcomes with hospital specialisation after accounting for the differences in patient and hospital characteristics, and hospital procedural volume. We further accounted for the clustering of patients within hospitals with models using random effects. In developing the multivariable models we specified that key covariates, including patient characteristics and procedure type (primary or revision total hip replacement or total knee replacement), be included in our models; whereas other variables such as comorbid illnesses were included in the model only if they were significantly associated with the outcome at an α ≤0.15. The final model for the composite outcome contained 40 variables (see web extra appendix 1). For each multivariable model we carried out a test for trend for the hospital specialisation variable (P<0.01). Model discrimination was evaluated using the c statistic (mortality 0.80 and composite 0.70) and calibration was assessed using the Hosmer-Lemeshow statistic.34 35
All analyses were done using SAS 9.0.
To ensure the robustness of our findings we carried out several secondary analyses. Firstly, we repeated our analyses while including patient populations excluded in our primary analyses (for example, those admitted after hospital transfer, those with acute fractures) (see web extra appendix 2). Secondly, we repeated our analyses using alternative definitions of hospital specialisation including stratifying hospitals into fourths (see web extra appendix 4) and 10ths (see web extra appendix 5) of orthopaedic specialisation and including hospital specialisation as a continuous measure rather than a categorical one. Thirdly, we repeated our regression analyses to examine the association between fifth of hospital specialisation and each of the individual components of the composite outcome, such as deep vein thrombosis, infection, or myocardial infarction. Fourthly, we divided hospitals into low (<25 total hip replacement and total knee replacement procedures annually), medium (26-100), and high (>100) volume strata and repeated our multivariable analyses, to explore whether the relation between hospital specialisation and outcomes was similar among hospitals with similar joint replacement volumes. Finally, we carried out separate analyses for patients receiving each of the four types of procedure (primary and revision total hip replacement and total knee replacement) (see web extra appendix 3).