Participants and setting
The Global Longitudinal Study of Osteoporosis in Women (GLOW) is an observational study designed to explore risk factors for and health consequences of fragility fractures in 60,393 women aged 55 years and older who had consulted their physician in the past 24 months, involving 723 physician practices at 17 sites in 10 countries (Australia, Belgium, Canada, France, Germany, Italy, Netherlands, Spain, UK, and US). Each site obtained local ethics committee approval to participate in the study. This has been described in detail previously . Briefly, based on the GLOW Hamilton cohort, a sample of approximately 4,000 participants were enrolled between May 2008 and March 2009 and stratified according to age strata such that two-thirds of participants were aged no less than 65 years. Women were eligible for inclusion if they had no language barriers or cognitive impairment, and were not too ill or institutionalized to complete the study survey .
Participants were surveyed annually with mailed questionnaires. Telephone interviews were performed if the participant needed assistance with finishing the survey or did not return mailed questionnaires. Surveys used in the GLOW study were designed to be self-administered by participants and covered the domains as follows: participant characteristics and risk factors, perception about fracture risk and osteoporosis, medication use, co-morbidities, health care use and access, physical activity, physical function and quality of life . This study only used data in the Hamilton site, Canada, while the other 16 sites were not involved in this study. Specifically, our study was a longitudinal analysis of the 3-year GLOW cohort of women in Hamilton. Written informed consent was obtained from all participants, and the study was reviewed and approved by the Western Institutional Review Board.
Construction of the FI at baseline
We constructed our FI based on the standard procedure and framework suggested by Searle and Rockwood [1, 19]. In creating a FI, the selection of variables should satisfy four basic criteria: biologically sensible, accumulate with age, do not saturate too early, and collectively cover a range of systems . In previous studies, 30 to 70 deficits had been used to construct the FI [1, 20]. The deficit accumulation approach does not necessarily require the exact same variables or the same number of deficits, to form the FI [20, 21]. Searle, however, recommended that the FI should include at least 30 to 40 total deficits . The finalized FI consisted of 34 variables at baseline, which included co-morbidities (n = 15), activities of daily living (ADL) (n = 12), symptoms and signs (n = 6), and healthcare utilization (n = 1). However, no measure of on social support deficits could be included in the FI because no such variable was recorded in the GLOW study. Regarding each variable, the consensus on its eligibility among authors was reached before it was included to construct the FI. As suggested by Rockwood and Searle [7, 19], each deficit variable was dichotomized or polychotomized and mapped to the interval 0–1 (e.g., for self-rating of health, ‘Excellent’ was coded as 0, ‘very good’ as 0.25, ‘good’ as 0.5, ‘fair’ as 0.75 and ‘poor’ as 1) to represent the frequency or severity of the deficit. None of the included deficit variables had more than 5% missing values.
The FI was calculated by adding up the values of deficits and dividing by the total number of items (n = 34), with the FI ranging from 0 to 1. For instance, if an individual had three deficits with each score of 1 point, two deficits with each score of 0.5 point and the other 29 deficits with each score of 0, her cumulative values of deficits would therefore be 4 divided by 34 giving a FI = 0.12.
The outcomes included falls, fractures, death and overnight hospitalizations. All the outcomes were self-reported and information from medical records was not available.
The primary outcome in this study, falls, were measured at baseline and each year of follow-up. Participants reported number of incident falls (none, one time, more than one time) in the prior 12 months on the annually mailed questionnaires.
Women were identified as having baseline self-reported fractures that included fractures of the clavicle, upper arm, wrist, spine, rib, hip, pelvis, ankle, upper leg or lower leg since the age of 45 years. Incident fractures and the dates of the fractures were reported on the 1-, 2- and 3- year follow-up surveys. At baseline, participants categorized their number of overnight hospitalizations according to the options on the questionnaire (none, one time, two times, more than two times). At each follow-up year, they were asked to record the total number of hospitalizations and total number of nights spent in hospital. Death was ascertained through contact with participants’ spouses, friends or family members and through electronic searches of obituaries. Some spouses and family members notified us of the participant’s death when they received mailings from our office, or when we called the homes of those participants who did not mail back their annual questionnaire. If we were unable to contact the household of the non-responders, we searched electronic databases of obituaries for entries which matched the participant’s full names and dates of birth.
The continuous FI was reported as mean and standard deviation (SD). Comparison of the categorized falls status (none, one time, more than one time) at baseline was examined using Chi-square tests for categorical variables and analysis of variance (ANOVA) for continuous variables. The rate of deficit accumulation per year at baseline was calculated on the basis of the mean FI with age. To make the results comparable with other studies [1, 19, 20], the rate of deficit accumulation was reported on a crude scale as well as on a log scale.
The relationship between each individual variable included in the FI and risk of falls during the third year of follow-up was also examined, after adjusted for baseline age. Binary logistic regression (i.e., had falls versus no falls) was performed to assess the association between baseline FI and risk of falls if the proportional odds assumption for ordinal logistic regression (i.e., no falls, one fall, more than one fall) was not met, taking women with no falls as the reference category.
Because the dates for falls were not available, unless otherwise emphasized, analyses of the relationship between baseline FI and risk of falls was conducted only using the data on the falls during the third year of follow-up. Baseline age-adjusted binary logistic regression models and fully-adjusted multivariable logistic regression models were performed and compared, where fully-adjusted models were adjusted for age, body mass index (BMI), smoking, drinking, education and baseline falls, to analyze the association between baseline FI and risk of falls. Receiver operating characteristic curves (ROC) were used to calculate the areas under the curve (AUC), which could judge the discriminability of the FI. A bootstrap analysis resampling 1,000 times with replacement from the original sample was conducted to assess the internal validation of the relationship between the FI and falls . Moreover, a sensitivity analysis was performed to investigate the relationship between the FI and the incident falls during the 3-year follow-up period, in which participants were dichotomized as having new incident falls (i.e., without baseline falls) and having recurrent falls (i.e., with baseline falls).
Secondary outcomes were fractures, death and overnight hospitalization. Similarly, age-adjusted models and fully-adjusted models were carried out and compared. Because the dates for death and overnight hospitalization were unavailable, logistic regression was applied to death during the follow-up, while Poisson regression was used to analyze the relationship between the FI and overnight hospitalizations during the third year of follow-up given the number of overall nights spent in the hospital as count data. Cox proportional hazards regression based on time-to-event was used to assess the associations between the FI and fractures after adjusting for age, BMI, smoking, drinking, education, baseline fractures and family history of fractures, where both a statistical test of proportional hazards assumption and a graphical examination using Schoenfeld residuals were performed .
To compare the results with other studies’ findings [8, 19, 24], all the statistics on the associations between the FI and falls, fractures, death and overnight hospitalization were reported based on an increase of 0.01 on the FI. The statistics were also measured and presented by per 1-SD increment of the FI.
Given that the findings on participants aged <65 years were scanty , subgroup analyses were conducted using the cut-off point of 65 years for age. For missing data, if <10% of observations on a variable were missing, the mean or median of the variable in its group was imputed . If ≥10% of data were missing, assuming they were missing at random, we conducted multiple imputations by including other relevant variables that were judged by clinical knowledge [26, 27]. All statistical tests were two-sided using an alpha level of 0.05 and all analyses were conducted with the software package SAS, version 9.3 (SAS Institute, Inc., Cary, NC).