When the Risk Model Misses Something
A fracture-risk estimate can be more informative than a T-score alone and still be wrong for the situation in front of the clinician. The important question is how to recognize that boundary: which inputs are simplified or missing, how fixed thresholds can distribute decisions unevenly, and what limited evidence means for testing and treatment. Readers who can see those failure surfaces are less likely to confuse a precise-looking number with a complete account of osteoporosis risk.
A model cannot use information it never receives
Every risk model is a selective description of a person. FRAX uses age, sex, weight, height and several clinical factors, with the option of femoral-neck BMD. That selection makes the estimate usable, but it also defines what the estimate cannot know. The USPSTF identifies several limitations: glucocorticoid use and alcohol exposure are entered in binary form rather than as fully quantified doses; lumbar-spine BMD and trabecular bone score are not included; falls and frailty are not collected; some of the cohort data used for race-specific fracture incidence are decades old; mortality estimates have not been updated since 2004; and conditions such as diabetes are omitted even though they may signal increased risk.
An omitted factor does not automatically invalidate the estimate. It changes the claim that the estimate can support. A calculator can still rank risk reasonably well in the populations and circumstances in which it was studied. But the result should not be treated as though it had measured every clinically relevant feature. If a person has recurrent falls and the model does not ask about them, the output is conditional on the model’s blindness. If glucocorticoid exposure is reduced to a yes-or-no field, the output does not express the full difference between doses and durations.
This distinction becomes important when a number is used near a decision threshold. A threshold turns a continuous estimate into an action category: below one side, one response may be considered; above the other, another response may be considered. The threshold can be useful for consistency, but it cannot repair information that was never entered. A person just below a threshold is not necessarily meaningfully different from a person just above it, especially when the model leaves out a factor that matters in both cases.
The USPSTF gives race-specific FRAX calculations as a concrete example of this problem. In the United States, the versions calibrated to observed fracture-incidence data produce lower predicted risk for Asian, Black and Hispanic people than for White people with the same age, sex, weight, BMD and listed clinical risk factors. The statement also notes uncertainty about which version to use for people who are multiracial or who immigrated from another country.
The concern is not that a calculation becomes acceptable simply because it is consistent. It is that a calibration choice can change who appears to cross a fixed threshold even when the listed clinical profile is identical. A numerical rule may therefore reproduce differences in background fracture rates without establishing that those differences apply to every individual. The output can look neutral while depending on assumptions about which population a person most resembles.