Summary findings

The original plan aimed for 900 RescueCo matched to 900 non-RescueCo patients, providing 90% power. Due to the strike and a slower enrollemnt rate, the study is projected to contact 900 RescueCo, but with a 60% follow-up rate and a 75% match rate (conservatively based on baseline analysis), the sample size enrollment will be 405 per arm. This sample size is underpowered based on both analyzing a binary classification of the GOSE score and a simulation of the 1-8 GOSE score. It would be powered for the minimal detectable effect of a 0.675 prevalence ratio between RescueCo and matched control patients.

Using SuperLearner-estimated propensity scores matching with a 0.2 caliper distance, 80% of RescueCo patients were successfully matched (80 out of 97). The matching was based on data from 97 RescueCo patients and 1309 non-RescueCo patients collected through June, and baseline variables only as the 6 month survey data is not available yet. The most important variables contributing to differences in propensity scores were: Mechanism of injury, Care at the scene, and Distance between the injury location and the hospital.

Sample size calculations

Reminder (from analysis plan): Assuming the Rescue.co group will have 0.75 times the risk of having a GOSE score indicating severe disability relative to the control group (22% expected proportion) then a study with the proposed sample size of 900 Rescue.co matched to 900 non-Rescue.co patients (assuming conservatively a 0.1 correlation of outcomes within matched pairs) will have 80% power at a type I error rate of 5%. There was an assumption of a 60% 6-month response rate, and 60-80 Rescue.co patients a month

Enrollment to date is from 02/2024 to 09/2024, and will continue till the end of 08/2025. From IPSOS master sheet: 226 completed, 8 not yet entered, 21 still in Ward, for a total of 255 enrolled. 120 who were not excluded were lost to follow-up.

Projected enrollment under different scenarios altering from analysis plan

All assume a 60% followup rate and a conservative 75% match percentile (82% were matched through June) - see matching section below for details

  • Scenario 1: future monthly average matches past averages: 255 + (255/7.5) * 12 = 663 patients. (Note assumption September collection halfway through). 663 * 0.6 * 0.75 = 300 recontacted and matched
  • Scenario 2: future monthly average matches past post-strike averages: 255 + 40 * 12 = 735 patients, 735 * 0.6 * 0.75 = 331 recontacted and matched
  • Scenario 3: future monthly average matches past post-strike averages, and 50% of lost patients can be recontacted (with future recapture of the 50%): 255 + 60 * 12 = 975 patients, 975 * 0.6 * 0.75 = 439 recontacted and matched
  • Scenario 4 (planned): With additional recruitment, still meet the 900 (assume 60% LTFU) 900 * 0.6 * 0.75 = 405 recontacted and matched

Updated power calculations under different scenarios

Note when I redid the analysis plan calculations assuming 900 per arm, using the R function power.paired.test, I got a power of 0.9, not 0.8.

First, I calculate power across different sample sizes under the original MDE (RR=0.735) and a reduced MDE powered for 405 per arm (RR=0.675) using a paired unconditional exact tests and a binary outcome of GOSE > 4.

Then, I calculate power across different MDE (relative risk) scenarios with a 405-person per arm sample size.

I simulated the power of analyzing a continious GOSE outcome (that still has a paired correlation of 0.1 and a prevalence of 22% when using a binary cutoff of GOSE>4):

N Power
405 0.638
439 0.676
800 0.920

Notes on power calculations:

  • Secondary analyses using targeted machine learning for estimation using all available data may be more efficient than matching.
  • Matching N:1 will improve power
  • I expected the simulated power calculations to have more power, instead of the observed less power, compared to the binary outcome, but that could be a function of the binary outcome prevalence or how I simulated the GOSE levels within a simulation bounded to give a 22% prevalence of GOSE > 4 in the control and a 0.1 correlation between arms. I ma

Initial check on study matching balance.

From the analysis plan: For an initial check on study matching balance, we propose using accumulated data for the first four months to examine the overlap of the distribution of matching covariates to ensure we have a reasonable balance. This assessment is outcome blind, so does not impact our final inferences. We can do so by constructing propensity score (PS) models (treating participation in FLARE as the outcome and the potential confounders/matching variables are the predictors) and then looking at 1) the overlap of the distribution of PS in the FLARE vs. non-FLARE groups, and 2) using our proposed matching algorithm to estimate the potential probability of a FLARE subject being unmatched.

Currently matching only on available baseline variables, as followup survey is not avaliable. Calculating the propensity score using:

Age, sex, marital status, education, morbidity_vars (medical history, chronic illness, prior surgery), previous injury, distance from hospital the injury occurred, injury place, injury activity, alcohol use, injury mechanism, intent, penetrating wound, scene care, prior care, signs of life, general AIS, eiss AIS, serious AIS, arrival date, injury date, payment method, and whether the patient has a mobile number.

The propensity score was calculated using the 97 RescueCo patients through June and the 1309 non-RescueCo patients collected through the end of June (First quarter).

Matching results

Using a nearest neighbor matching on the PS, with a caliper distance of 0.2, 80% of RescueCo patients are matched

Control Treated
All 1309 97
Matched 80 80
Unmatched 1229 17
Discarded 0 0

Distribution of propensity scores

Baseline variable contributions to the propensity score

The most important variables leading to differences in propensity scores are the mechanism of injury, care at the scene, and distance between the injury location and the hospital.

After propensity score matching, the injury distance and scene care are the variables that are the most residually unbalanced: