Background
Stratified randomization is commonly used to achieve balance in randomized controlled trials (RCTs) but may not always lead to balanced treatment groups, especially when there are multiple baseline covariates that are relevant prognostic factors [
1‐
4]. Treatment group imbalance across baseline covariates could lead to erroneous estimation and interpretation of treatment effect sizes reported in RCTs both at interim and final analyses [
5]. Although primary analyses that adjust for imbalance in baseline covariates remains a popular method for dealing with baseline covariate imbalance [
6‐
8], ignoring imbalance across important baseline prognostic factors during randomization may result in significant loss in statistical power to detect treatment differences, even if these factors are adjusted for in analysis. In addition, balance at interim analysis, when sample sizes are necessarily small, is important to permit confident decision-making by the Data Safety and Monitoring Committee. As a result of increased awareness of the potential for loss of power, balancing treatment allocation for influential covariates has become more important in modern clinical trials [
9].
Covariate adaptive randomization designs have been proposed as an alternative to stratified randomization. This class of randomization design is advantageous in that it is able to achieve between-group balance across baseline covariates even in small-sample-size RCTs [
10‐
14]. Covariate adaptive randomization designs are particularly useful in stroke trials where treatment groups may vary on multiple but influential baseline covariates; these include treatment with intravenously administered alteplase and baseline stroke severity among others. Commonly used adaptive randomization strategies to balance baseline covariates include stratified constrained randomization and minimization [
15,
16]. However, both strategies can only balance on few categorical baseline covariates and may lead to imbalance especially in continuous, mixed continuous/discrete covariates [
17]. Minimal sufficient balance randomization is a recent, more robust covariate adaptive randomization strategy for use in RCTs with several multiple baseline covariates [
18,
19]. This technique is efficient in sequential RCT designs where baseline covariates are expected to be balanced at each interim monitoring.
In this study, we describe the implementation of minimal sufficient balance randomization for the “Endovascular treatment for Small Core and Anterior circulation Proximal occlusion with Emphasis on minimizing CT to recanalization times” (ESCAPE) trial, a multicenter RCT that enrolled 316 acute-ischemic-stroke subjects [
20,
21]. We provide recommendations about the potential use of minimal sufficient balance for sequential designs in stroke trials.
Results
At the interim analysis of the actual data (
N = 243), 183 subjects received intravenously administered alteplase, while 59 subjects did not receive it. There were no significant differences between both arms of the trial on the baseline covariates in subjects who received intravenously administered alteplase and those who did not receive it, suggesting an overall balance in minimal sufficient balance patient allocation into both treatment arms (see Table
2). In the group that did not receive intravenously administered alteplase, the proportion of subjects that had true random assignment (after burn-in) was 87.2%, while the minimal sufficient balance algorithm exhibited true randomness (i.e., there was no voting on any of the balancing covariate) in dynamically allocating the last 30 subjects. On the other hand, 97.5% of the subjects who received the intravenously administered alteplase were truly randomly assigned (after burn-in) and there was no voting on any of the balancing covariate for the last 66 subjects randomized before interim analysis.
Table 2
Treatment differences on baseline covariates by intravenously administered alteplase status at interim analysis (N = 243)
Age (mean, SD) | 68.82 (12.24) | 68.69 (14.78) | 0.95 | 73.56 (13.92) | 70.26 (14.97) | 0.40 |
NIHSS score (mean, SD) | 16.54 (5.41) | 17.28 (5.75) | 0.38 | 17.83 (5.81) | 15.78 (5.21) | 0.16 |
Sex (% female) | 47.92% | 51.72% | 0.61 | 52.78% | 56.52% | 0.78 |
ASPECTS score (% (8–10)) | 68.75% | 68.97% | 0.98 | 66.67% | 73.91% | 0.56 |
Baseline occlusion location (% MCA M1) | 70.83% | 70.11% | 0.92 | 72.22% | 73.91% | 0.89 |
At the end of the trial, 237 subjects received the intravenously administered alteplase, while 78 subjects did not receive it. A permutation test of overall group differences revealed no evidence of imbalance on each baseline covariate when the trial was stopped for overwhelming efficacy (see Table
3). In the group that did not receive intravenously administered alteplase, 91.4% of the subjects were truly randomly assigned (after burn-in) at the end of the trial, while there was no voting on any of the balancing covariates in randomizing the last 49 subjects at the end of the trial. For subjects who received intravenously administered alteplase, 98.2% of the subjects had true random assignment at the end of the trial and there was no voting on any of the balancing covariate in dynamically allocating the last 120 subjects at the end of the trial.
Table 3
Treatment differences on baseline covariates by intravenously administered alteplase status at interim analysis (N = 316)
Age (mean, SD) | 69.16 (12.32) | 68.43 (14.89) | 0.68 | 69.76 (16.31) | 69.45 (13.19) | 0.93 |
NIHSS score (mean, SD) | 16.38 (5.35) | 16.69 (5.75) | 0.66 | 16.93 (5.60) | 15.73 (5.43) | 0.34 |
Sex (% female) | 50.00% | 51.28% | 0.844 | 57.78% | 57.58% | 0.99 |
ASPECTS score (% (8–10)) | 66.67% | 70.09% | 0.572 | 60.00% | 72.73% | 0.24 |
Baseline occlusion location (% MCA M1) | 73.33% | 72.65% | 0.906 | 73.33% | 75.76% | 0.81 |
Discussion and conclusion
This study describes the implementation of minimal sufficient balance randomization, a covariate adaptive randomization strategy, in the ESCAPE trial. Our results demonstrate that overall group balance was achieved in dynamically allocating subjects to treatment arms in the ESCAPE trial both at the time of the interim analysis and at the time that the trial was stopped. This suggests that this covariate adaptive randomization strategy can assist data and safety monitoring boards to make reliable decisions about futility or efficacy in a randomized controlled trial.
Similar to all adaptive trial design approaches, the integrity of this approach relies upon the immediate availability and accuracy of data used, and, in this case, data on the covariates used to implement the randomized minimization approach. Because of the nature of stroke treatment, fast treatment decisions are necessarily taken without certainty on all preclinical data. For example, a patient’s age might be determined just upon the year of birth, but later require correcting by 1 year because the complete date of birth becomes known several hours later. We instituted a policy of correcting the data based upon the main trial database, which was considered the source of truth. This introduces a small degree of imprecision in the application of the approach since randomization will depend upon what values are in the trial randomization database. It is expected that the importance of this kind of inaccuracy will wane over time as the trial accrues more and more patients.
It was particularly important to us, in planning the ESCAPE trial, to ensure that we maintained balance on key variables at the time of interim monitoring. Because the trial was relatively small in sample size, the use of the minimal sufficient balance algorithm served as potential insurance against the possibility of random imbalance on key prognostic variables. This provided some confidence to the Data Safety and Monitoring Board that the results would not be confounded by covariate imbalance at the interim look. These results show that minimal sufficient balance randomization is able to achieve a high rate of true random allocation compared to other commonly used randomization schemes such as permuted blocks.
This use of a minimal sufficient balance algorithm in this trial demonstrates the potential advantage of this covariate adaptive design in achieving balance of key prognostic variables in a small-sampled RCT with pre-specified interim analyses thresholds. In permuted block designs, the last subject(s) in a given block size will have a deterministic treatment assignment, with the proportion with true random allocation varying by block size. In comparison, the degree of true random allocation was very high in the ESCAPE trial, preserving the fundamental nature of the RCT. Further research is needed to demonstrate how several parameters of the minimal sufficient balance randomization (e.g., such as the number and nature of baseline prognostic factors, number of anticipated interim analyses, tolerance limits, and biased coin probability) can aid the determination of the expected sample size for a trial design.
This study has limitations. The minimal sufficient balance algorithm as proposed by Zhao et al. was implemented in the ESCAPE trial by stratifying on a prognostic factor (intravenously administered alteplase) while balancing on the remaining five baseline covariates within each strata. This stratified minimal sufficient balance randomization was adopted because of variation in standard of care clinical use. Future research will use computer simulations to examine the statistical properties of this variant of minimal sufficient balance randomization scheme under a variety of data analytic conditions.
In summary, this study describes the implementation of the minimal sufficient balance, a covariate adaptive randomization scheme, in the ESCAPE trial with results from empirical data confirming balance between treatment arms across six baseline prognostic factors. The minimal sufficient balance randomization scheme is advantageous for achieving group balance in small-sampled trials. We therefore recommend minimal sufficient balance randomization for use in clinical trials where balance on multiple baseline covariates is important.