In a new Medical Care article published ahead of print, I worked with colleagues at UCLA’s Fielding School of Public Health to explore whether Medicare expenditures associated with fall-related injuries (FRI) depend on how FRIs are identified in claims data.
We identified an area in which epidemiological and financial data for fall injuries may be inadequate as currently collected and adapted a recently developed algorithm to evaluate whether it would accurately identify fall injuries in Medicare claims data.
We then compared the incidence and Medicare-related expenditures of fall injuries using this newly adapted algorithm with several more commonly used methods. The findings showed that prior methods may either substantially overestimate or underestimate fall injury expenditures, with implications for the cost-effectiveness of fall injury prevention programs.
Altogether, the findings may influence future research in this area and possibly have an impact on national policymakers debating the appropriateness of a Medicare falls benefit.
And now to the specifics.
Past studies produce inconsistent estimates of such expenditures. Estimates range from $2,000 to $26,000 per faller, $1,000 to $10,000 per fall, and $5,600 to $43,000 per fall-related hospitalization. This likely has something to do with methodological choices among researchers—specifically, choice of method to identify FRIs.
Most studies use a method of identifying FRIs that probably undercounts FRIs drastically. This method, which uses ICD-9 external cause-of-injury codes (or e-codes) (so the authors call it e-code only, or “ECO”) establishes the place and mechanism of injury. Well-known problems with how often hospitals and emergency departments accurately report such codes give reason to believe that only the most severe injuries attributable to falls are coded with this method. The problem this creates when estimating expenditures is that only the most expensive FRIs are likely counted while less expensive ones are left off. (This also undercounts the prevalence of FRIs.)
Because of that issue, a second method was developed which additionally includes a number of ICD-9 diagnostic categories (so the authors call it e-codes plus diagnostic codes, or “ECDC”). These include fractures, dislocations, sprains, and strains, and contusions. In other words, if an older adult is treated at the hospital for a hip fracture, that is included as an FRI under this method; but, so is a contusion of a finger. Yet, it’s not clear that all (or even most) bruises are due to accidental falls. In fact, even some fractures could be due to factors unrelated to falls (such as auto accidents or pathologic fractures). The concern with this approach is that it could “find” more FRIs in claims data—an undoubted improvement over the traditional method using e-codes only—but also count as FRIs injuries unrelated to falls. This would have important implications for prevalence and expenditure estimates—it could overcount the number of FRIs in the claims data while underestimating expenditures (because it would include in the average expenditures less serious injuries that may not be fall-related).
Given issues with the sensitivity and specificity of the above two methods, a recently created algorithm used e-codes, ICD-9 diagnostic codes, and (CPT) procedure codes for things like casting, splinting, and x-rays (which might suggest the presence of an FRI). Development of the algorithm, created by UCLA/RAND clinical researchers (not this study’s authors) (we call it “AUR” for adapted UCLA/RAND method), was guided by prior work on the incidence of Medicare fractures and an injury classification scheme developed by Mary Tinetti. The authors of the present study adapted this algorithm. They assumed that the method improved on the sensitivity and specificity of the other methods, which might result in prevalence and expenditure estimates between those of the other two methods.
Here is what we found:
Total annual per-faller expenditures ranged from $5,648 to $12,171. The first method, ECO, had the highest estimate ($12,171), the second method, ECDC, had the lowest estimate ($5,648), and the third method, AUR, had an estimate in between the other methods ($9,389).
Similarly, total annual Medicare expenditure estimates ranged broadly. The ECO method had the lowest estimate ($4 billion), the ECDC method had the highest ($25 billion), while again the AUR method had an estimate in between the others ($13 billion). (Take with a grain of salt –the confidence intervals are wide.)
A wide range of estimates was also obtained for: expenditures by service component (e.g., hospital, nursing facility, outpatient) following an FRI; estimated out-of-pocket costs; and expenditures by the initial treatment setting (inpatient, ED-only, or outpatient).
All of this means that method matters. The traditional EOC method estimates expenditures that are quite high but prevalence that is quite low. The ECDC method (with e-codes and a broad set of ICD-9 codes) estimates much lower average expenditures but much higher prevalence (resulting in 6 times higher total annual Medicare expenditures). The ECDC method is likely a considerable improvement over the traditional ECO method because it produces much more credible prevalence estimates—yet, we found that it may locate too many “FRIs” (not actual FRIs) in outpatient settings—underestimating expenditures.
The newer AUR method’s middle-of-the-road estimates are probably the most accurate—assuming one agrees that (unlike the assumption built into the second method) not all fractures, sprains, and bruises are due to falls. However, all of the methods have been and will continue to be important as researchers strive to improve their ability to “find” FRIs using imperfect data.
Nonetheless, researchers and policymakers should use caution when interpreting estimates that come from the ECO method and possibly the ECDC method—as those methods will likely produce upper and lower ends of a confidence interval. When considering cost-effectiveness for fall prevention efforts or the plausibility of a Medicare fall benefit, the use of estimates in the middle of that confidence interval such as obtained using the UCLA/RAND algorithm may be most appropriate.