Abstract
Background
Previous observational studies comparing early versus delayed surgery for periprosthetic hip fractures may yield inaccurate estimates because they did not account for immortal time. Immortal time refers here to the period during which patients had to remain alive to be classified in the delayed surgery group. Patients who died before delayed surgery could not enter that group, potentially making delayed surgery appear safer than it truly was. Although early surgery is recommended for native hip fractures, whether it improves survival after periprosthetic hip fracture remains uncertain because prior null findings may reflect either no true effect or bias arising from immortal time. We therefore used a target trial emulation framework to estimate the effect of early versus delayed surgery strategies on short-term mortality.
Questions/purposes
(1) When immortal time is appropriately addressed, what is the estimated effect of an early surgery strategy (within 2 days after admission) compared with a delayed surgery strategy (3 to 10 days after admission) on 30- and 90-day mortality risks? (2) How do mortality estimates from a conventional analysis based on observed surgical timing, which is susceptible to immortal time, compare with those from the target trial emulation?
Methods
Using 2014 to 2024 data from a Japanese hospital administrative database, we identified 4626 potentially eligible patients who were hospitalized with periprosthetic hip fractures. The database captured admission dates, procedure dates, and mortality across many hospitals, enabling analysis of surgical timing and short-term mortality for this relatively uncommon fracture. We excluded 1.3% (62 of 4626): 0.3% (14) who were younger than 45 years, 0.2% (11) who had pathologic fractures, and 0.8% (37) who were treated at institutions where data submission was temporarily suspended during the 90-day follow-up window, leaving 4564 eligible patients. We used the clone-censor-weight approach to emulate a randomized trial and estimate the observational analog of the per-protocol effect. On day 0 (admission), each patient was cloned, and one clone was assigned to each strategy. Thus, both strategies included the same patients at baseline (mean age 84 years; 80% female), and all baseline characteristics were identical by design. Clones were artificially censored for nonadherence to the assigned strategy, and inverse probability weighting was used to adjust for selection bias from artificial censoring and loss to follow-up. Of 4564 clones assigned to each strategy, 11% (510) in the early surgery strategy and 34% (1564) in the delayed surgery strategy were not artificially censored for nonadherence. Loss to follow-up before 30 days occurred in 2.9% (132) and 7.5% (341), respectively; before 90 days, it occurred in 4.1% (189) and 11.8% (538), respectively. Weighted pooled logistic regression estimated mortality risks, risk differences, and risk ratios at 30 and 90 days; 95% confidence intervals (CIs) were obtained from 500 bootstrap samples. As a methodologic benchmark, we also estimated mortality using a conventional classification based on observed surgical timing (days 0 to 2 versus days 3 to 10), recognizing that this postbaseline classification was susceptible to immortal time.
Results
At 30 days, there was no clear evidence of a difference in mortality risk between the strategies, although the estimates were higher under the early strategy (1.72% versus 0.83%; risk difference 0.89% [95% CI -0.32% to 1.72%]; risk ratio 2.06 [95% CI 0.64 to 3.73]). At 90 days, there was no clear evidence of a difference, and the effect estimates were near the null (2.35% versus 2.31%; risk difference 0.04% [95% CI -1.08% to 1.24%]; risk ratio 1.01 [95% CI 0.59 to 1.68]). Complete-case estimates were similar, and subgroup estimates were imprecise, with no consistent evidence of effect modification by age or sex. Compared with the emulation, the benchmark analysis yielded lower estimated 30-day mortality risks in both surgical timing groups: 1.10% in the early surgery group and 0.48% in the delayed-surgery group (risk difference 0.62% [95% CI -0.20% to 1.80%]; risk ratio 2.27 [95% CI 0.74 to 14.10]); the estimated 90-day mortality risks were 2.08% and 2.05%, respectively (risk difference 0.03% [95% CI -1.55% to 1.83%]; risk ratio 1.01 [95% CI 0.36 to 2.22]).
Conclusion
Our findings do not support assuming that the 48-hour standard for native hip fractures applies to periprosthetic hip fractures solely to reduce short-term mortality. Surgeons may consider the time required for medical stabilization and operative planning, although our findings do not justify avoidable delay because other outcomes were not evaluated. The lower 30-day estimates obtained using observed surgical timing highlight potential bias from postbaseline classification, including immortal time and selection based on eventual surgery; the 90-day estimates were similar. Because causal interpretation relies on unverifiable assumptions, and residual bias from unmeasured factors remains possible, future observational studies should emulate target trials and incorporate fracture severity, operative complexity, and surgical resources to evaluate complications, reoperation, and functional recovery.
Level of Evidence
Level Ⅲ, therapeutic study.