Covariate information and statistical analysis
Comorbidity ICD diagnosis codes before the date of the hip fracture were collated from the National Patient Registry (with data from 1964 and onward) to calculate the number of comorbidities and Charlson's comorbidity index.[18, 19] In addition, we used this register to retrieve ICD code information about alcohol or drug abuse (yes/no) and any psychiatric disease (yes/no).
Supplementary information was assessed by a comprehensive computer-assisted telephone interview conducted between 1998 and 2002. Previously, the twin pairs participated in mailed questionnaire surveys in 1970 (twins born in 1896 to 1925) and 1973 (twins born in 1926 to 1958). The overall response rate for the old questionnaires was about 90%, and the response rate for the telephone interview was 84%. The interviews included a number of items on lifestyle behavior, diseases, and symptoms related to fracture risk and survival,[14, 22] including smoking status (never, former, current), physical activity level (low, medium, high; similar categories have been found to be predictive both of hip fracture and of mortality) marital status (married or cohabitant, widowed, single), visual impairment (yes/no), hearing aid (yes/no), use of estrogen-replacement therapy (never/ever use), any prescribed medication (yes/no), nonprescribed or prescribed medication use (yes/no), present use of corticosteroids (yes/no), body mass index (BMI, kg/m2, continuous), alcohol abstainer (yes/no), and an index for activity of daily living (AD). The ADL index was constructed from self-reported responses to the capacity of cooking/shopping, dressing/bathing, help with medication, help with household work, or other daily support. The questionnaire data information was time-updated when performing the analysis, ie, if the hip fracture event occurred before the telephone interview, we used the earlier questionnaire information, but if the hip fracture happened after the telephone interview, we used this latter information.
Statistical analyses were performed using SAS software version 9.3 (SAS Institute Inc., Cary, NC, USA), Stata 11.2 (Stata Corporation Inc., College Station, TX, USA), and R (R Foundation for Statistical Computing, Vienna, Austria, 2008). Missing questionnaire values were replaced by multiple imputation and the Markov chain Monte Carlo (MCMC) method and five imputations by PROC MI and MIANALYZE in SAS. In the imputation procedure, we used number of comorbidities, Charlson index, any psychiatric disease, alcohol or drug abuse, the interview/questionnaire information described above, age, sex, hip fracture status, and finally mortality status at end of follow-up. To compensate for the nonrandomized design of our observational study, we used propensity-score methods. The individual propensity scores, defined as the conditional probability of obtaining a hip fracture based on covariates (ie, those described above but excluding mortality status), were estimated with a multiple logistic-regression model.
Follow-up time was accrued from date of the hip fracture event, ie, the same date for each hip fracture discordant pair, until date of death or the end of the study period (December 31, 2010). We estimated sex-stratified age and propensity score-adjusted hazard ratios (HRs) of death by intrapair Cox proportional hazards regression (PROC PHREG, SAS) and their 95% confidence intervals (CIs) for the twin with a hip fracture compared with her co-twin without a hip fracture. Thus, in the analyses, the pair status variable was used as a stratum variable, fixing the baseline hazard within a matched pair. Because the rate of death varied by time of follow-up, we display results from piece-wise models with cut points by 1-year of observation. Adjustment for other types of osteoporotic fractures (eg, distal forearm, spine, and proximal humerus) that occurred during follow-up affected our estimates only marginally (data not shown).
Moreover, we estimated mean difference in survival time between the twin with a hip fracture and her co-twin up to 1 year and up to 10 years after hip fracture (also including estimates of sex difference in survival) and finally to the point of expected survival according to Official Statistics Sweden at the median age of the hip fracture event. Average expected survival was defined based on average Swedish population data during the study period. The number of years lost with 95% CIs was calculated using a bias corrected method and accelerated bootstrap CI derived from the draw of 10,000 bootstrap twin pairs. We expressed the difference in survival between the fractured and the nonfractured twins in years as well as a proportion relative to the expected survival time. In addition, the estimated total years lost because of the fracture were multiplied by the average annual number of hip fracture cases in Sweden during the study period based on data from the National Patient Register.
In the analysis of unrelated twins, we used flexible parametric models[28, 29] for continuous estimation of the sex-specific propensity score–adjusted HRs together with 95% CIs by time of follow-up. The 95% CIs were derived from 5000 bootstrap samples of twin pairs. The propensity score was based on the same model as the identical twin analysis.