Table S2.

Results of sensitivity analyses: alternative methods to adjust for confounding by indication

Approach	POCON dataset ICP monitoring OR (95% CI)	Tirilazad dataset Intracranial operation OR (95% CI)	EBIC dataset Primary referral OR (95% CI)
Alternative propensity score adjustment approaches
Propensity score adjustment†	0.80 (0.66–0.96)	0.86 (0.57–1.28)	0.92 (1.28–0.65)
Inverse probability weighting‡	0.73 (0.45–1.17)	1.11 (0.73–1.70)	0.90 (0.70 –1.16)
Alternative hospital-level adjustment approach
Random-effect preference-based approachⱡ	1.45 (0.81–2.97)	1.95 (0.67–4.56)	0.63 (0.37–1.06)
Random-effect model with random intercept for the interventionⱠ	2.35 (0.76–8.84)	6.61 (0.21–142.7)	0.64 (0.35–1.24)

Notes:

^†The natural logarithm of the propensity score was added to the analytic model.

^‡Results are presented for a 95% winsorized cohort.

^ⱡHospitals were divided into two groups (high preference for the intervention versus low preference for the intervention) based on the mean percentage of patients receiving the intervention. The OR represents the odds of a higher score on the GOS for high-preference hospitals in comparison to low-preference hospitals adjusted for observed patient-level confounders to increase statistical power.

^ⱠFor each center, the random intercept of exposure to the intervention was estimated in a random-effect model with the intervention of interest as outcome variable and all IMPACT variables and a random intercept for center as predictors. The random intercepts were subsequently added as predictors, together with the IMPACT variables, to increase statistical power. Results are presented for a 10% change.

Abbreviations: EBIC, European Brain Injury Consortium; GOS, Glasgow Outcome Scale; ICP, intracranial pressure; IMPACT, International Mission for Prognosis and Analysis of Clinical Trials; POCON, Prospective Observational Cohort Neurotrauma.