|Home | About | Journals | Submit | Contact Us | Français|
We identified hip fracture risks in a prospective national study.
Baseline (1993–1994) interview data were linked to Medicare claims for 1993–2005. Participants were 5,511 self-respondents aged 70 years and older and not in managed Medicare. ICD9-CM 820.xx (International Classification of Diseases, 9th Edition, Clinical Modification) codes identified hip fracture. Participants were censored at death or enrollment into managed Medicare. Static risk factors included sociodemographic, socioeconomic, place of residence, health behavior, disease history, and functional and cognitive status measures. A time-dependent marker reflecting postbaseline hospitalizations was included.
A total of 495 (8.9%) participants suffered a postbaseline hip fracture. In the static proportional hazards model, the greatest risks involved age (adjusted hazard ratios [AHRs] of 2.01, 2.82, and 4.91 for 75–79, 80–84, and ≥85 year age groups vs those aged 70–74 years; p values <.001), sex (AHR = 0.45 for men vs women; p < .001), race (AHRs of 0.37 and 0.46 for African Americans and Hispanics vs whites; p values <.001 and <.01), body mass (AHRs of 0.40, 0.77, and 1.73 for obese, overweight, and underweight vs normal weight; p values <.001, <.05, and <.01), smoking status (AHRs = 1.49 and 1.52 for current and former smokers vs nonsmokers; p values <.05 and <.001), and diabetes (AHR = 1.99; p < .001). The time-dependent recent hospitalization marker did not alter the static model effect estimates, but it did substantially increase the risk of hip fracture (AHR = 2.51; p < .001).
Enhanced discharge planning and home care for non–hip fracture hospitalizations could reduce subsequent hip fracture rates.
HIP fracture among older adults in the United States is a frequent, serious, and expensive health event that results in substantially increased morbidity and mortality (1–4). Of the 316,000 hip fractures that occurred in 2005, 89.3% were in people aged 65 and older (5), 95% resulted from falls (6), and 75% were in women (5). Up to 20% of older hip fracture patients died within 1 year (7), and only half were able to return home or live independently after hospital discharge (8,9). Survivors were much more likely to experience significant disability and reduced quality of life (10–12). Hip fracture expenditures were $20 billion in 2006 (13). Thus, it is not surprising that Healthy People 2010 (14) established national goals of reducing annual hip fracture rates to 4.7 per thousand for men and to 4.2 per thousand for women. Current hip fracture rates (5.6 per thousand for men and 11.1 per thousand for women), however, underscore how far we are from meeting those goals (5).
To facilitate progress toward the Healthy People 2010 (14) hip fracture rate goals, all potential risk factors must be identified and their effects accurately calibrated so that appropriate intervention points can be targeted (6). Although the literature on hip fracture risk among older adults is extensive (15–18), few studies have simultaneously considered the roles of sociodemographic, socioeconomic, place of residence, health behavior, disease history, and functional and cognitive status risk factors in large, nationally representative, prospective studies. In this article we use data from the Survey on Asset and Health Dynamics Among the Oldest Old (AHEAD) linked to Medicare claims to do this. Our study has five strengths. First, AHEAD is a large, nationally representative sample whose participants were aged 70 and older at their baseline interviews in 1993–1994. Second, their Medicare claims were available for up to 12 years postbaseline surveillance. Third, the baseline AHEAD interviews provided an extensive array of measures within each of the above categories of risk factors. Fourth, using the claims data we constructed a dynamic (time-dependent) “health shock” measure reflecting the increased risk of hip fracture during postdischarge transition periods following hospitalization for something other than hip fracture. Finally, we used propensity score methods (19–22) to adjust for potential selection bias.
The AHEAD study has been described elsewhere (23–26). All analyses reported here were initially weighted to adjust for the unequal probabilities of selection due to the multistage cluster and oversampling of African Americans, Hispanics, and Floridians. The baseline response rate was 80.4% (N = 7,447). Of the 7,447 baseline participants, 802 (10.8%) could not be linked to their Medicare claims, another 604 (8.1%) were in managed Medicare at baseline, and another 530 (7.1%) were proxy-respondents, leaving 5,511 AHEAD self-respondents in our analytic sample who could be linked to their Medicare claims for calendar years 1993–2005 and who were not enrolled in managed Medicare plans at baseline (74.0% of the total AHEAD sample). Proxy-respondents were excluded because cognitive and psychosocial data were not ascertained for them. Participants in managed Medicare plans at baseline were excluded because these plans do not have the same data reporting requirements as fee-for-service Medicare plans (27). Participants were censored at the time of either of these two competing risks—death or enrollment into managed Medicare—whichever came first.
To adjust for the potential selection bias introduced by these three exclusions, we used propensity score methods to reweight the data (19–22). Among all 7,447 participants, we estimated a multivariable logistic regression model of inclusion in the analytic sample using baseline interview data including all the risk factors identified below, as well as others (complete list available on request). The fit of this model was good (C statistic = .72; Hosmer-Lemeshow statistic = .15) (28,29). Within each propensity score (predicted probability) decile, we determined the average participation rate (i.e., inclusion in the analytic sample, or P) and used the inverse (1/P) to reweight the data (i.e., to adjust the existing weights traditionally used in analyzing the AHEAD data for the potential selection bias). This gives greater influence to participants in the analytic sample most like those not included. Finally, the propensity score weights were rescaled so that the final weighted N was equal to the actual number of participants in the analytic sample (i.e., 5,511).
Baseline interview dates were used to individually mark the beginning of the surveillance periods. To identify hip fractures we used ICD9-CM principal admitting diagnostic codes 820.xx. Hip fractures had to occur at least 1 day after the participant's baseline interview, and censoring occurred at the time of either of two competing risks—death or enrollment into managed Medicare—based on data from the Medicare claims denominator file. Multivariable proportional hazards regression with competing risks (30) was used to model time to hip fracture. Model development and evaluation followed standard procedures (31,32).
Within each group of risk factors, we included traditional measures, as well as novel indicators (to broaden the potential net) when these were available in the AHEAD. Sociodemographic factors included age, sex, race, living alone, marital status, and the importance of religion. Socioeconomic factors included education, income, total wealth, the number of health insurance policies, and perceived neighborhood safety. Place of residence was measured by population density, geographic region, and dwelling type. Health behaviors included the body mass index, smoking history, and current alcohol consumption.
Disease history included whether the participant had ever been told by a physician that he/she had angina, arthritis, cancer, diabetes, a heart attack, heart disease other than a heart attack, a previous hip fracture, hypertension, lung disease, psychological problems, or a stroke, as well as a comorbidity index. We also included whether the participant was hospitalized in the year prior to baseline, and the number of physician visits during that period as indirect indicators of otherwise unmeasured disease burden. Functional status was measured by self-rated health, counts of the number of difficulties with activities of daily living (ADLs) and instrumental activities of daily living (IADLs); the number of reported depressive symptoms; falling in the year prior to baseline; reports of bothersome pain; the ability to pick up a dime; and self-rated vision, memory, and urinary incontinence. Cognitive status included immediate and delayed word recall, and the Telephone Interview to Assess Cognitive Status (TICS) (33).
Although extensive, our list of baseline risk factors was not complete. Bone mineral density (BMD) was a major omission (8,9,15). It has recently been shown, however, that although BMD contributes substantially toward the prediction of hip fracture risk among postmenopausal women, it does not appreciably alter the risk estimates obtained for other covariates in multivariable models (34). Thus, the omission of BMD is unlikely to have biased our parameter estimates.
We introduced a dynamic element to the analysis by constructing a health shock marker using the postbaseline Medicare claims for each participant. This time-dependent covariate was switched “on” at any time prior to censoring when the participant was admitted to a hospital for any primary ICD9-CM diagnosis other than hip fracture. It stayed “on” for n days after discharge and was then switched “off.” It could subsequently be switched back “on” at the onset of another precensoring hospital admission for something other than hip fracture. This measure captures the transition period when older adults are especially vulnerable to adverse effects associated with both their underlying health shock (i.e., the reasons for their admission) and the consequences of their treatments (e.g., delirium, disorientation, and increased risk for falling), especially in fragmented health care delivery systems) (35–40). Sensitivity analyses were conducted to determine which of several values of n was most predictive of hip fracture—14, 30, 60, 90, and 120 days.
Table 1 contains the percentages (or means) for each of the risk factors that were considered (as well as their crude hazards ratios [HRs]). In the analytic sample, the mean age was 77 years, 38% were men, 10% were African American, 4% were Hispanic, and 41% were widowed. One fourth had only been to grade school, and mean income was $25,417. One fourth had arthritis, 9% had angina, 13% had cancer, 12% had diabetes, 46% had hypertension, 5% had previously fractured a hip, and 7% had psychological problems. The mean number of ADL and IADL difficulties was 0.34 and 0.41, respectively. Health shocks were experienced by 81%, and hip fractures were experienced by 8.9%. The total number of person-years of surveillance was 39,112 with a mean of 7.1 per person.
Table 2 contains the adjusted hazard ratios (AHRs) obtained from a trimmed static model that included only those baseline risk factors independently significant at the .05 level or beyond (consolidated from forward, backward, and stepwise modeling; Column 1) and the AHRs obtained from a trimmed static model that included only those baseline risk factors and the dynamic health shock marker that were independently significant at the .05 level or beyond (Column 2). Given space constraints, our focus is on the AHRs shown in Table 2 (rather than the crude HRs shown in Table 1) because these identify prevention intervention targets and demonstrate the value of dynamic modeling. We note here that further analysis (not shown) revealed that the loss of statistical significance and effect size diminishment that was observed among the crude HRs for most risk factors in the adjusted models stems from their intervening effects on falling, which is the major precipitating cause of hip fracture. This was especially the case among the disease history markers.
The risk of hip fracture increased with age and peaked among those aged 85 or older who had a 391% greater risk than participants who were 70–74 years old. African Americans were 63% less likely and Hispanics were 44% less likely than whites to have fractured their hips. Participants living in mobile homes had a 41% greater risk of hip fracture than those living in either single- or multiple-story nonmobile homes. Those who were obese or overweight were less likely than normal weight participants to have fractured their hips, whereas those who were underweight were more likely to fracture their hips (60% and 23% lower risks, 73% higher risk, respectively). Smoking history increased the risk of hip fracture by 52% for former smokers and by 49% for current smokers. AHEAD participants with diabetes or psychological problems were 99% and 42% more likely to fracture their hips than older adults without those conditions, whereas those with heart disease were 23% less likely to fracture their hips. Those reporting poor health were 43% more likely to fracture their hips than were other participants. Individuals who had fallen in the year prior to baseline were 35% more likely to have fractured their hips. Participants who scored below the median on or refused to answer the immediate word recall test had 39% and 69% greater risks of hip fracture, whereas those who had better scores on the TICS were 27% less likely to fracture their hips. When the health shock variable (crude HR = 2.88, AHR = 2.51 at 14 days; AHR 2.00 at 30 to 120 days; all p values <.001) was added to the model (Column 2), it did not diminish the effect estimates in the trimmed static model (Column 1), but it did substantially increase both the overall model fit (χ2 improvement in the log likelihood ratio = 19.07 at 1 df; p < .005) and the risk of hip fracture by 151%.
Three points from our analysis warrant further discussion, the first of which is by far the most promising. It involves the importance of the dynamic health shock measure, which was indirectly measured in this study using the time-dependent recent hospitalization indicator. There are two aspects of this effect to keep in mind—it was quite large, and it did not mediate the effects of the baseline hip fracture risk factors. What this dynamic health shock measure captures is the transition period when older adults are especially vulnerable to adverse effects associated with both their underlying health shock (i.e., the reasons for their recent hospital admission) and the consequences of their treatments (e.g., delirium, disorientation, and increased risk for falling), especially in fragmented health care delivery systems (35–40).
When calibrated at 14 days, which is when the effect size peaked, the health shock measure increased the risk of hip fracture by 151% and substantially improved the fit of the model (χ2 improvement in the log likelihood ratio = 19.07 at 1 df; p < .005). Although the effect size diminished to a 100% risk increase at 30 days, it then remained stable throughout 120 days. This early (14 day) risk peak, followed by a somewhat reduced risk by 30 days that plateaus for at least 4 months, reflects the initial concentration of risk during the immediate postdischarge period (i.e., short-term risks associated with delirium, disorientation, and increased risk for falling), followed by the enduring elevated risk (i.e., long-term risks associated with reduced systemic and reserve capacities) associated with hospitalization among older adults. As such, this suggests that postdischarge planning and monitoring for at least the first 2–4 weeks following hospitalization for something other than a hip fracture could reduce the risk of subsequent hip fracture.
Clearly, the introduction of the health shock (i.e., recent hospitalization) measure is a very promising development that underscores the need to shift from static to dynamic risk modeling approaches. Further research is needed, however, that explores the health shock measure in order to clarify what the underlying etiologic mechanisms might be. Such research should include whether restrictions to surgical versus medical admissions, shorter versus longer stays, or other decompositions would identify particular hospitalization subsets that pose the greatest risks for subsequent hip fracture.
The second discussion point is that there were no surprises in the identification of either the static baseline risk factors or the magnitudes of their risks. That is, our risk estimates are remarkably consistent with the extant literature (15). Of the 5,511 AHEAD participants in our analysis, 495 (8.9%) suffered a postbaseline hip fracture. In the static baseline model, hip fracture risk was significantly higher for older adults, whites, women, those who lived in mobile homes, were normal or underweight, current and former smokers, those with diabetes, had poor self-rated health, were fallers, or had lower cognitive status. The greatest of these risks involved age (AHRs of 2.01, 2.82, and 4.91 for 75–79, 80–84, and ≥85 year age groups vs those aged 70–74 years; p values <.001), sex (AHR = 0.45 for men vs women; p < .001), race (AHRs of 0.37 and 0.46 for African Americans and Hispanics vs whites; p values <.001 and <.01), body mass (AHRs of 0.40, 0.77, and 1.73 for obese, overweight, and underweight vs normal weight; p values <.001, <.05, and <.01), smoking status (AHRs = 1.49 and 1.52 for current and former smokers vs nonsmokers; p values <.05 and <.001), and diabetes (AHR = 1.99; p < .001).
Some might wonder why the risk associated with falling was much smaller by comparison (AHR = 1.35; p < .01). This is likely due to two reasons. The first is the length of our surveillance period. We had postbaseline Medicare claims for up to 12 years, with a mean of 7.1 years. As the surveillance period lengthens, the relevance of falling in the year prior to baseline diminishes. Sensitivity analyses (not shown) in which we arbitrarily censored the analyses at 2 and 5 years bear this out. A second possible reason for the relatively modest effect of falling is that although it precipitates 95% of all hip fractures among older adults, only a small percentage of falls results in a hip fracture (2–6). That is, falling is a necessary but not sufficient condition for hip fracture.
The third discussion point involves the public health policy implications from our static risk factor model. We found that hip fractures were frequent (8.9%) and that the greatest risks for them were either immutable or very hard to change in older adults—age, sex, race, body mass, smoking status, and diabetes. This does not bode well for intervention identification and successful implementation (1,6,15). Nonetheless, these major risk factors could and should be used by health care providers to target older adults for increased monitoring and evaluation, as well as the delivery of applicable preventive interventions.
In concluding this article, we note that our study is not without limitations. Four warrant mention here. First, BMD was not available for inclusion in the analysis. Although it has recently been shown that BMD does not appreciably alter risk estimates obtained for other covariates in multivariable models (34), and therefore its omission is unlikely to have biased our parameter estimates, we were nonetheless unable to establish either its additive or relative effects. Second, AHEAD did not contain any performance-based tests, such as the Short Portable Physical Battery, which would have allowed us to evaluate the especially relevant roles of balance, lower body strength, and gait speed to hip fracture risk (41,42). Third, although the AHEAD is rich in self-reported data and linked to Medicare claims for more than a decade, detailed clinical histories were not available, restricting our study to an epidemiologic versus etiologic analysis. Finally, we relied solely on baseline (i.e., static) risk factors from the AHEAD self-reports, even though several of them (such as ADLs, IADLs, and self-rated health) were repeated at most follow-ups. Although this approach is consistent with much of the hip fracture risk literature (1–4,15), it is suboptimal and creates the potential for the sole dynamic measure (the health shock indicator) to appear to be more important. Inclusion of the repeated self-reports, however, would have created a myriad of complexities associated with missing data, selection bias, and correlated error structures.
Supported by National Institutes of Health grant R01 AG-022913 to F.D.W.
F.D.W. is the Co-PI of the Center for Research in the Implementation of Innovative Strategies in Practice (CRIISP) at the Iowa City VAMC, G.E.R. is the PI of CRIISP, M.P.J. is the Senior Statistician at CRIISP, and K.K.R. is a CRIISP Analyst. CRIISP is funded through the Department of Veterans Affairs, Veterans Health Administration, Health Services Research and Development Service (HFP 04-149). The opinions expressed here are those of the authors and do not necessarily reflect those of any of the funding, academic, or governmental institutions involved.