Osteoporosis-related fracture case definitions for population-based administrative data

Administrative data were from the central Canadian province of Manitoba, which has a population of approximately 1.2 million. Like other Canadian provinces, Manitoba has a system of universal healthcare. The Manitoba Centre for Health Policy Research Data Repository houses computerized databases that contain records of virtually all contacts with the health care system. The databases can be linked via an anonymized personal identification number. Ethics approval for this research was received from the Manitoba Health Research Ethics Board and permission to access the study data was provided by the Manitoba Health Information Privacy Committee.

Hospital and physician billing claims data were used to construct the fracture case definitions. Diagnoses in hospital data are recorded using the International Classification of Diseases, 9^th revision, Clinical Modification (i.e., ICD-9-CM), up to and including the 2003/04 fiscal year (a fiscal year runs from April 1 to March 31) and the International Classification of Diseases, 10^th revision, Canadian version (i.e., ICD-10-CA), for subsequent years. A maximum of 16 ICD-9-CM codes and 25 ICD-10-CA codes are recorded on each record. Physicians who are paid on a fee-for-service basis submit billing claims to the provincial health ministry; these claims capture almost all outpatient services, including those provided in hospital emergency and outpatient departments. While a small percentage of physicians are salaried (i.e., about 7% of family physicians [20], approximately 90% submit parallel billing claims for administrative purposes. Physician billing claims contain a single ICD-9-CM code as well as service codes, which are defined in a fee schedule (manitoba.ca/health/manual). Diagnosis codes in hospital and physician data have been validated and used extensively in Manitoba [21‐23] and in other Canadian jurisdictions [24] for research about chronic conditions, and they have been used in multiple studies about osteoporosis-related fractures [9, 11, 25].

Validation data were from the Canadian Multicentre Osteoporosis Study (CaMos) [26]. CaMos is an ongoing, population-based cohort study designed to provide national estimates of the prevalence and incidence of osteoporosis and osteoporosis-related fractures. The study population is composed of non-institutionalized individuals who reside within a 50-km radius of one of nine study centres located across Canada. These geographic areas encompass approximately 40% of the Canadian population and include both rural and urban residents. Households in each area were selected by random draws of listed telephone numbers; one randomly selected household member greater than 25years was asked to participate. A total of 9423 participants entered the study (2884 males and 6539 females) during the 18-month recruitment period. The characteristics of study participants have been described previously [27, 28]. A refusal questionnaire was used to assess selection bias. For osteoporosis prevalence, the primary study outcome, selection bias was not observed in any of the study age groups, with the exception of a small amount of bias in the 80+ age group [29]. Informed consent was obtained from participants and the study received approval from the institutional review board at each participating centre. Data on incident fractures was compiled from CaMos participants for the period from 1996 to 2006. These data were collected using annual postal questionnaires and/or in-person interviews [27]. A detailed interviewer-administered questionnaire was used at years 3, 5, and 10. At baseline, year 5, and year 10, lateral lumbar and thoracic spine X-rays and bone mineral density tests were performed. At years 1, 2, 4, and 6 through 9, a detailed two-page questionnaire was mailed to participants asking about hospitalizations and fractures within the past year. Participants who reported having a fracture were asked to provide consent for study staff to collect additional data about the fracture diagnosis from the treating physician and/or hospital record (i.e., verification by radiology report). We limited the study to fractures occurring in individuals 50years of age and older, with age defined at the fracture index date.

Table 1 lists the 35 administrative data case definitions that were selected for investigation. These were selected based on a review of published studies [3, 12, 30], recommendations from clinical co-investigators with expertise in fracture ascertainment in administrative databases (WDL, SM), and the authors’ previous experience ascertaining other chronic diseases, include osteoporosis, in administrative data [31‐33]. The case definitions were differentiated by: (a) source of data, (b) number of records with the relevant diagnosis code(s), (c) type of diagnosis in hospital data, (d) presence of service codes in physician billing claims, and (e) duration of the fracture-free period. With the exception of one hip fracture case definition, all site-specific definitions used the same ICD-9-CM and ICD-10-CA diagnosis code(s). For hip fracture, we considered ICD-9-CM 820 (fracture of neck of the femur) and 821 (fracture and other unspecified parts of the femur) because some hip fractures may be assigned a less precise diagnosis code [34]. Case definitions were based on hospital data only (hip) or hospital and physician claims data, in keeping with previous research [3, 15]. For the latter, case definitions requiring one or at least two records with the specified diagnosis code(s) were considered. Service codes capture radiologic and magnetic resonance imaging services for incident clinical vertebral fracture, immobilization or fixation services for wrist fracture, and surgical repair and fixation procedures for hip fracture. Service codes have also been used in previous studies to improve fracture ascertainment [35]. Fracture-free periods of zero, six or twelve months were considered, using the site-specific fracture index date to establish the end-point of the fracture-free period.

To illustrate the interpretation of the case definitions, H1 identifies hip fractures using hospital records with ICD-9-CM 820 or 821 (ICD-10-CA S72.0, S72.1, or S72.2) as the most responsible (i.e., primary) diagnosis; it does not use physician service codes nor does it require a fracture-free period. In contrast, case definition H13 identifies hip fractures from hospital records with ICD-9-CM 820 (ICD-10-CA S72.0, S72.1, or S72.2) in any diagnosis field. A physician service code was present within the hospitalization period and a 12-month fracture-free period was adopted. For wrist fracture, case definition W1 identifies fractures using hospital or physician billing records with ICD-9-CM 813 (ICD-10-CA S52) in any diagnosis field. This case definition requires a physician service code to accompany the diagnosis code and does not adopt a fracture-free period.

The fracture index date was the date of the first diagnosis or service code for a fracture event. Pathologic fractures were included because they represent a small proportion of all fractures and their exclusion can lead to underestimation of the fracture burden due to osteoporosis [36]. For each case definition, the number of incident fractures was generated for the Manitoba population 50years of age and older for fiscal years 1997/98 to 2006/07. Age, which was defined using the fracture index date, was obtained from health insurance registration files. For hip fracture, counts of incident fractures were generated both including and excluding residents of long-term care (i.e., nursing home) facilities [37]; the CaMos data excludes residents of these facilities and this may affect comparability of estimates. Residence in a facility was determined from nursing home files containing admission and separation dates.

Crude and sex- and age-specific fracture rates were calculated for administrative case definitions and for the CaMos data. Generalized linear models (GLMs) were used to test for differences in the estimates from administrative and CaMos data [38]. For each case definition, sex-specific models that contained the main effects of age (five-year groupings from 50‒54years to 80+ years) and source (i.e., administrative; reference: CaMos) were fit to the data, along with a model that combined the data for males and females; this latter model contained the main effects of age, source, and sex. A negative binomial distribution was adopted to model the data due to the presence of extra-Poisson variation in the case counts. Fit was evaluated using the ratio of model deviance to degrees of freedom, which should be close to one for a well-fitted model. A Wald ² statistic was used to test the difference in the estimates from administrative and CaMos data. Each test was conducted at the =.01 significance level in order to control the familywise error rate, the probability of committing at least one Type I error. The relative rate (RR) of incident fractures for administrative data case definitions was obtained by exponentiation of the regression coefficient for the source variable.

GLMs with generalized estimating equations were used to test for differences among the case definitions [39]. Annual fracture counts (i.e., for 1997/98 to 2006/07) for population strata defined by age (five-year groupings from 50‒54years to 85+ years) and sex were modeled assuming a negative binomial distribution. An exchangeable structure, which assumes constant correlation in successive years, accounted for dependence among the incident fracture counts. Models containing the main effects of age group, case definition, sex, year, and the year x case definition interaction were fit to the data for each fracture site. The interaction was used to test for differences in the linear trend across case definitions. If the interaction was statistically significant, linear contrasts were used to test for differences amongst selected pairs of case definitions. A significance level of =.05 was adopted for the omnibus test and each of the contrast tests was conducted at =.01. If the interaction was not significant, the main effects of year and case definition were investigated.