Language was not associated with regional anesthesia use in a propensity-score-matched sensitivity analysis (OR 1.23) for total joint arthroplasty

Description #

In a sensitivity analysis using 1:1 propensity-score matching (1,010 patients matched on age, sex, race, ethnicity, insurance, procedure type, ASA, BMI, surgical year, and surgery patient class), preferred language remained non-significantly associated with regional anesthesia use: OR 1.23 (95% CI, 0.51–2.97; P = .6537), with the CI crossing 1. Covariate balance was achieved on all matched variables (standardized differences). This corroborates the primary multivariable null.

"In our sensitivity analysis using propensity score matching, we successfully matched 1010 patients according to baseline characteristics. Covariate balance was assessed using standardized differences and met for all variables. There was no significant difference in RA use among language (English/LEP) using a logistic regression model (OR, 1.23 [95% CI, 0.51–2.97, P = .6537])." (Sabra, 2025, p. 995)

Methods Context #

What? #ⓘ

The observable: regional anesthesia utilization by preferred language (English/LEP), re-estimated in a propensity-matched cohort as a robustness check on the primary model.

"We calculated propensity scores for RA use with covariates of age, sex, race, ethnicity, insurance, procedure type, ASA level, BMI, surgical year, and surgery patient class." (Sabra, 2025, p. 993)

How? #ⓘ

Propensity-score matching at a 1:1 ratio using nearest-neighbor matching, with standardized differences to assess balance, followed by a logistic regression model on the matched sample.

"Patients were matched using nearest-neighbor matching at a 1:1 ratio. Standardized differences were used to assess balance between groups. After matching, a logistic regression model was run to measure the association between LEP and RA use." (Sabra, 2025, p. 993, spanning pp. 993–995)

Who? #ⓘ

1,010 propensity-matched THR/TKR patients drawn from the full HSS 2016–2023 cohort of 58,918.

"In our sensitivity analysis using propensity score matching, we successfully matched 1010 patients according to baseline characteristics." (Sabra, 2025, p. 995)

Other Notes #

Directionally the matched OR (1.23) points opposite to the unmatched adjusted OR (0.93), but both are non-significant with wide CIs — consistent with no detectable language effect and limited precision (see power/ceiling caveats).

Caveats #

• Retrospective single-institution EMR study with provider-bias variability and unmeasured cognitive impairment The findings come from a single-institution, retrospective EMR-based study at one specialized orthopedic hospital (HSS), which limits generalizability to centers with different patient populations, payer mixes, and RA practices. The authors also note variability in provider bias, and that they could not exclude patients with cognitive impairment because that variable was unavailable in the record — a potential uncontrolled confounder of both language documentation and anesthetic decision-making. Interpreter use and modality were captured only as documented in the EMR, which may undercount actual language-service use.
• Underpowered to detect the small observed language effect despite large sample (post hoc power) Despite the large overall sample (58,918 patients), the study is by the authors' own post hoc assessment underpowered to detect the small, non-significant language effect it observed. The design targets clinically meaningful differences of ≥2%; a true difference smaller than that (or concentrated in the tiny LEP subgroup) could be overlooked. This is the central interpretive caveat on all of the null language findings: absence of a detected association is not evidence of no association. The authors caution the results should be interpreted cautiously and note that post hoc power calculations are not recommended for definitive conclusions.
• Near-universal regional anesthesia creates a ceiling effect with only 13 non-RA LEP patients [Inferred:] Because regional anesthesia was near-universal at this institution (98.8% of 58,918 patients), the "did not receive RA" outcome group is only 707 patients, and the LEP patients within it number just 13. With so few LEP non-RA events, there is almost no outcome variance in which a language-based disparity could be detected — a ceiling/floor effect that limits both the univariate and adjusted analyses regardless of the total sample size. The null language association is therefore weak evidence of true equivalence and may not transfer to institutions where RA is not already the default. (This is my inference from the Table's cell counts, complementary to the authors' own power caveat.)