Lecanemab for Patients With Early Alzheimer Disease

Key Points

Question

The US Food and Drug Administration has promoted innovation in clinical trial design via the bayesian approach; does that make clinical trials more efficient?

Findings

In this bayesian analysis of a randomized trial, the bayesian design assigned most patients to better performing doses. Appropriately accounting for missing data our analyses found that the most effective dose of lecanemab nearly doubles efficacy at 18 months relative to placebo in comparison with restricting to patients who completed follow-up.

Meaning

These results suggest that innovations associated with using the bayesian approach improve the efficiency of drug development and the accuracy of clinical trials, even when there is substantial data missingness.

This bayesian analysis of a clinical trial examines the effectiveness of bayesian design for adaptively assigning patients to effective dosage regimes.

Abstract

Importance

Bayesian clinical trial designs are increasingly common; given their promotion by the US Food and Drug Administration, the future use of the bayesian approach will only continue to increase. Innovations possible when using the bayesian approach improve the efficiency of drug development and the accuracy of clinical trials, especially in the context of substantial data missingness.

Objective

To explain the foundations, interpretations, and scientific justification of the bayesian approach in the setting of lecanemab trial 201, a bayesian-designed phase 2 dose-finding trial; to demonstrate the efficiency of using a bayesian design; and to show how it accommodates innovations in the prospective design and also treatment-dependent types of missing data.

Design, Setting, and Participants

This study was a bayesian analysis of a clinical trial comparing the efficacy of 5 lecanemab 201 dosages for treatment of early Alzheimer disease. The goal of the lecanemab 201 trial was to identify the effective dose 90 (ED90), the dose achieving at least 90% of the maximum effectiveness of doses considered in the trial. This study assessed the bayesian adaptive randomization used, in which patients were preferentially assigned to doses that would give more information about the ED90 and its efficacy.

Interventions

Patients in the lecanemab 201 trial were adaptively randomized to 1 of 5 dose regimens or placebo.

Main Outcomes and Measures

The primary end point of lecanemab 201 was the Alzheimer Disease Composite Clinical Score (ADCOMS) at 12 months with continued treatment and follow-up out to 18 months.

Results

A total 854 patients were included in trial treatment: 238 were in the placebo group (median age, 72 years [range, 50-89 years]; 137 female [58%]) and 587 were assigned to a lecanemab 201 treatment group (median age, 72 years [range, 50-90 years]; 272 female [46%]). The bayesian approach improved the efficiency of a clinical trial by prospectively adapting to the trial’s interim results. By the trial’s end more patients had been assigned to the better-performing doses: 253 (30%) and 161 (19%) patients to 10 mg/kg monthly and 10 mg/kg biweekly vs 51 (6%), 52 (6%), and 92 (11%) patients to 5 mg/kg monthly, 2.5 mg/kg biweekly, and 5 mg/kg biweekly, respectively. The trial identified 10 mg/kg biweekly as the ED90. The change in ADCOMS of the ED90 vs placebo was −0.037 at 12 months and −0.047 at 18 months. The bayesian posterior probability that the ED90 was superior to placebo was 97.5% at 12 months and 97.7% at 18 months. The respective probabilities of super-superiority were 63.8% and 76.0%. The primary analysis of the randomized bayesian lecanemab 201 trial found in the context of missing data that the most effective dose of lecanemab nearly doubles its estimated efficacy at 18 months of follow-up in comparison with restricting analysis to patients who completed the full 18 months of the trial.

Conclusions and Relevance

Innovations associated with the bayesian approach can improve the efficiency of drug development and the accuracy of clinical trials, even in the context of substantial data missingness.

Trial Registration

ClinicalTrials.gov Identifier: NCT01767311

Introduction

The clinical development of lecanemab and other amyloid-targeting drugs for treating Alzheimer disease has been controversial.^1,2,3,4 The bayesian and adaptive aspects of the lecanemab phase 2b dose-ranging clinical trial 201 (ClinicalTrials.gov identifier NCT01767311)^5,6 have been misunderstood and misinterpreted, resulting in improper conclusions about the trial’s clinical implications.^7,8,9,10

This study’s goal is to explain the foundations, interpretations, and scientific justification of the bayesian approach. Lecanemab trial 201 provides an excellent example. In this article, we evaluate the efficiency of using a bayesian design and how it accommodates innovations in the prospective design of clinical trials. We also demonstrate how the bayesian design of lecanemab handled missing data, and the advantages of doing so properly.

One criticism leveled against lecanemab trial 201 was that it “utilized bayesian statistics, a type of statistical methodology that allows researchers to make changes to a study in the middle.”⁷ This incorrect statement is badly misleading. It suggests casualness that is not appropriate in clinical trial design and conduct. It is true that the design adapted to results that were accruing in the trial. It is also true that the future conduct of the trial depended on these adaptations. But bayesian designs are prospective. Once the design was set, none of the researchers could make changes “in the middle.” Moreover, none of the researchers were privy to information accruing in the trial that would have allowed them to make such changes. The trial’s Independent Monitoring Committee (IMC) had access to results in the trial but they too were unable to modify the design. The IMC’s purpose was to ensure that the design was carried out per protocol specifications. The design algorithm that was programmed to carry out the protocol completely prescribed how and what changes could be made without human intervention.

Bayesian designs are expensive in a sense. The researchers have to work diligently at the design stage to build an algorithm that will take actions that most efficiently achieve the trial’s objective. In the case of lecanemab 201 this was “to establish the effective dose 90% (ED90) on the Alzheimer’s Disease Composite Clinical Score (ADCOMS) at 12 months of treatment in patients with early Alzheimer’s disease.”⁵ The ED90 is the minimum dose that provides at least 90% of the maximum benefit among the dose regimens included in the trial.

An automaton could have run trial 201 (with 1 exception indicated below). The automaton must be able to simulate a clinical trial as designed. This makes it possible to evaluate its operating characteristics, including type I error rate and statistical power.

The principal rationale for taking the bayesian approach in adaptive design is that it provides tools for building a learning algorithm. Bayes rule enables updating what is known about the various treatments continually throughout a trial. Such updating is not possible in the frequentist approach.¹³ The bayesian approach also allows for calculating predictive probabilities regarding whether the trial will successfully achieve its goal. If this probability is low then the trial may stop, or otherwise have its future course modified.^{11,12,13,14,15}

Bayesian adaptive designs control type I error, typically requiring simulations.^11,12 Virtual patients are accrued, and their outcomes are generated by computer according to assumed benefits of experimental therapy. For example, type I error is the proportion of positive trials among many thousands of trials generated assuming the drug has no effect.⁵

The US Food and Drug Administration (FDA) has promoted innovations in drug trials, including through the concept of “complex innovative design” (CID). A working definition of CID is one that requires simulation to evaluate the trial’s operating characteristics.¹² The FDA’s CID guidance focuses on phase 3 registration trials, including bayesian trials. There have been few such trials, as commentators have remarked.⁸ Among the most important “firsts” was a pathbreaking bayesian randomized adaptive catheter ablation trial that led to FDA approval in paroxysmal atrial fibrillation and high-profile publication in 2010.¹⁶

The first CID that led to a drug approval was the initial phase 3 trial of dulaglutide in type 2 diabetes, called AWARD-5 (NCT00734474).¹⁷ The design, developed with FDA as part its Critical Path Initiative,¹⁸ was wholly bayesian and included adaptive randomization to 7 positive doses in its first stage, with randomization probabilities updated every 14 days. AWARD-5’s first stage was similar to lecanemab 201. Bayesian predictive probabilities determined when and whether the trial would shift seamlessly into a fixed randomization second stage, further evaluating 2 doses that would eventually be marketed. The algorithm chose the 2 doses that maximized a clinical utility index. Both doses selected were in the middle of the range of doses that were being evaluated.

The bayesian approach is rare in Alzheimer research, but it is not new. We know of 1 other Bayesian clinical trial in Alzheimer disease. Between 2007 and 2009, Abbott conducted a phase 2 dose-finding trial (NCT00555204) with a bayesian design, one that was similar to that of lecanemab 201. The Abbott trial was negative.¹⁹ Also, in 2011 and 2012, AlzForum reported on meetings that promoted the bayesian adaptive approach in Alzheimer disease.^20,21,22

Lecanemab 201’s methods and results have been described elsewhere.^5,6 In this article we present important and novel aspects of the trial and describe the underlying statistical models used in its design. The trial provides an efficient model for demonstrating dose-response relationships and, in turn, a drug’s efficacy. We show how the interim results of the trial drove its adaptations. In particular, the design quickly learned that the lower doses of lecanemab were ineffective. The design greatly lowered the probability of assigning patients to the smaller doses and increased allocation probabilities to the 2 highest doses. This improved the precision of the efficacy estimates for the higher doses. Improving the treatment of trial participants is sometimes the objective of adaptive randomization.²³ Although not a goal in a dose-finding trial, focusing on doses likely to be the ED90 can improve the overall treatment of trial participants.

We also describe how the trial’s bayesian design accommodated an unanticipated adaptation while the trial was ongoing, one that led to substantial missing data. The trial’s prospective bayesian approach utilized multiple imputation,²⁴ which turned out to be more valuable than the designers anticipated. Handling data missingness is always a challenge.²⁴ Trial 201 presented a new challenge when dose-specific dropouts were mandated.

Methods

Detailed methods for the design and analysis of lecanemab 201, including its CONSORT diagram, have been published.^5,6 Here we provide an overview of the trial’s methodology and descriptions of methodology we used in addressing our study’s objectives.

Lecanemab 201 was an 18-month, multicenter, double-masked, placebo-controlled, 6-arm trial. There were 5 lecanemab regimens: 5 mg/kg monthly, 10 mg/kg monthly, 2.5 mg/kg biweekly, 5 mg/kg biweekly, and 10 mg/kg biweekly. The primary analyses were comparisons with placebo. All patients received biweekly infusions of placebo or the assigned dose of lecanemab in masked fashion.⁶ The trial was approved by each center’s institutional review board or independent ethics committee. All participants provided written informed consent.

The primary end point was ADCOMS at 12 months. ADCOMS is a novel, composite clinical outcome designed to be sensitive in early Alzheimer disease.²⁵ The maximum duration of protocol therapy was 18 months. ADCOMS at 18 months was a key secondary end point. Efficacy and safety data were collected at trimonthly visits from 3 to 18 months postrandomization.

Statistical Analysis

The basis of inference in the bayesian approach is the posterior probability distribution of the various parameters, including that of the primary end point ADCOMS, for each lecanemab dose vs placebo. This distribution varies over time and is updated at each interim analysis. It is posterior in the sense that it is based on all currently available information, whether during the trial or after the trial ends, which is the final posterior distribution.

Trial 201’s protocol specified a super-superiority threshold (SST) of 0.03 on the ADCOMS scale for lecanemab doses vs placebo. The probability of super-superiority for a dose (Pr(>SST)) is the probability the dose has lower ADCOMS than placebo minus 0.03.

The sample sizes in trial 201 were adaptive in that they depended on the trial’s interim results. Accrual would stop for futility if Pr(>SST) below 5% for the most likely ED90 at any of the first 3 interim analyses or below 7.5% at any subsequent analysis. Accrual could not stop for success until the fourth interim analyses. Stopping from analysis 4 and onward required Pr(>SST) of 80% or above for the most likely ED90.

The trial also adapted randomization probabilities to favor doses that would be most informative about the dose-response relationship. The first interim analysis was set to occur when the 196th patient accrued. For these patients, the randomization probability equaled 1/7 for the 5 lecanemab doses and 2/7 for placebo. The second and future interim analyses occurred after every additional cohort of 50 patients, up to a maximum of approximately 800 patients. After full accrual, trimonthly interim analyses occurred until all patients had 18 months follow-up. At all interim analyses, randomization probabilities were proportional to the doses’ probabilities of being the ED90 weighted by its sample size. The probability of assignment to the placebo group was set equal to that of the dose most likely to be the ED90. Specifics for interim analyses are summarized in eTable 1 in Supplement 2.

To calculate the design’s operating characteristics, we simulated the trial under many scenarios. Some assumed the null hypothesis of no drug effect. Others assumed various positive dose-response relationships.⁵

The lone exception to the prospective nature of the trial was when in midtrial a non-US regulatory authority required that the highest dose regimen (10 mg/kg biweekly) be dropped for patients who were apolipoprotein E4 (APOE4) carriers, whether homozygous or heterozygous. Furthermore, protocol therapy was stopped for APOE4 carriers who had received dose 10 mg/kg biweekly for less than 6 months. This requirement led to high rates of missing data for dose 10 mg/kg biweekly.

Dropping dose 10 mg/kg biweekly for APOE4 carriers had implications on the trial’s design and conclusions. First, having many missing visits on 10B and many fewer APOE4 carriers randomized to 10 mg/kg biweekly, the eventual ED90, meant much lower power for the trial than planned. In addition, having fewer APOE4 carriers assigned to the 10 mg/kg biweekly dose raised the possibility that an interaction between dose and APOE4 status would affect the trial’s conclusions and interpretations. We address both concerns.

The required revisions meant the algorithm of assignment probabilities had to be adjusted, which we did staying as close as possible to the original design. The trial continued without pausing accrual. The original design used bayesian multiple imputation for missing data.²⁴ This method appropriately estimates missing observations despite patients’ different reasons for missingness. Imputation methodology such as “last observation carried forward” estimate a particular value to replace the missing value. These are biased and inappropriately increase precision.²⁴ The bayesian statistical model imputes each missing observation as a probability distribution.²⁴ This distribution is based on results available for all patients. Having many missing visits inflates the standard deviation of these distributions, appropriately representing the greater uncertainty.

Members of the chartered IMC that conducted interim analyses were a software engineer and 2 independent statisticians. The IMC ran the prospective algorithm at each interim analysis and created a report summarizing the results of the analyses (eFigure 3 in Supplement 2). Analysis code was custom code run on Fortran Compiler version 19.0.2.187 (Intel). The software was programmed and validated in advance, with unambiguous rules for all adaptive decisions. The IMC could not modify the prospective design. All members of the trial sponsor were masked to interim trial results and adaptations, except for the adaptation required by the regulatory requirement cited above.

Results

Overall, 854 patients were randomized and treated in lecanemab 201 between December 2012 and November 2017. Follow-up continued through May 2019. A total of 238 patients were included in the analysis for the placebo group (median age, 72 years [range, 50-89 years]; 137 female [58%]); 587 patients were assigned to a lecanemab treatment group (median age, 72 years [range, 50-90 years]; 272 female [46%]) (eTable 2 in Supplement 2).⁶

Patient 196 was randomized on January 30, 2014, triggering the first interim analysis. Bayesian calculations and the IMC report were finalized on February 3, 2014. The rightmost panel of eFigure 3A in Supplement 2 shows updated randomization probabilities for use until the next interim analysis. eFigure 2 in Supplement 2 shows cumulative enrollment over time and by dose group. eFigure 1 in Supplement 2 shows the randomization probabilities over time. By the trial’s end many more patients had been assigned to the better-performing doses: 253 (30%) and 161 (19%) patients to 10 mg/kg monthly and 10 mg/kg biweekly vs 51 (6%), 52 (6%), and 92 (11%) patients to 5 mg/kg monthly, 2.5 mg/kg biweekly, and 5 mg/kg biweekly, respectively.

Figures 1A and 1B and Table show final 12-month and 18-month analyses. At 18 months the mean improvement for the 3 lowest doses was 0.013 compared with 0.039 for the 2 highest doses. For equal randomization, and with about 122 patients per lecanemab dose, the estimated improvement vs placebo in the trial would have been 0.024. This compares with 0.029 for the adaptive randomization in the actual trial, a 21% improvement.

Figure 1. Final Results by Dose Cohort for the Final 12-Month and 18-Month Analyses.

In panels B and E, whiskers indicate 95% probability intervals for the bayesian model. In panels C and F, the probability that each dose is the ED90 sums to 1. Probability that the benefit on Alzheimer Disease Composite Clinical Score (ADCOMS) is greater than the super-superiority threshold (SST) is also indicated for each dose, with probabilities not constrained to sum to 1. The super-superiority and futility boundaries refer to the probability of being greater than the SST (Pr(>SST)). If the Pr(>SST) for the most likely ED90 exceeds the success boundary, this would indicate that the analysis would recommend going to phase 3. If Pr(>SST) for the most likely ED90 is below the futility boundary, then the analysis would recommend stopping accrual (if still accruing) and not going to phase 3. The third plot on the rightmost panel, Pr(>PBO), is the probability that the dose’s benefit on ADCOMS is greater than that on placebo.

Table. Bayesian Model-Based Conclusions for ADCOMS, Month 12, and Month 18.

Group	No. (%)	Changes on ADCOMS from baseline, mean (95% CI)a			Posterior probabilities
		Change from baseline	Difference from placebo	Ratio to placebo	ED90	Superiority	Super-superiorityb
After 12 mo of follow-up
Placebo	247 (29)	0.113 (0.090 to 0.137)	NA	NA	NA	NA	NA
5 mg/kg monthly	51 (6)	0.117 (0.077 to 0.160)	0.004 (−0.043 to 0.053)	1.046 (0.654 to 1.526)	0.036	0.441	0.078
10 mg/kg monthly	253 (30)	0.084 (0.061 to 0.107)	−0.029 (−0.062 to 0.003)	0.750 (0.515 to 1.034)	0.385	0.962	0.481
2.5 mg/kg biweekly	52 (6)	0.134 (0.087 to 0.181)	0.020 (−0.032 to 0.073)	1.194 (0.743 to 1.728)	0.010	0.220	0.029
5 mg/kg biweekly	92 (11)	0.116 (0.085 to 0.149)	0.003 (−0.036 to 0.043)	1.046 (0.712 to 1.442)	0.009	0.439	0.050
10 mg/kg biweekly	161 (19)	0.077 (0.048 to 0.105)	−0.037 (−0.074 to 0.000)	0.684 (0.411 to 1.00)	0.561	0.975	0.638
After 18 mo of follow-up
Placebo	247 (29)	0.172 (0.142 to 0.202)	NA	NA	NA	NA	NA
5 mg/kg monthly	51 (6)	0.156 (0.108 to 0.206)	−0.016 (−0.073 to 0.041)	0.912 (0.608 to 1.265)	0.129	0.719	0.320
10 mg/kg monthly	253 (30)	0.142 (0.113 to 0.171)	−0.031 (−0.072 to 0.011)	0.829 (0.623 to 1.072)	0.194	0.927	0.513
2.5 mg/kg biweekly	52 (6)	0.156 (0.101 to 0.210)	−0.017 (−0.079 to 0.045)	0.911 (0.573 to 1.291)	0.138	0.702	0.333
5 mg/kg biweekly	92 (11)	0.165 (0.127 to 0.205)	−0.007 (−0.056 to 0.043)	0.965 (0.705 to 1.279)	0.019	0.622	0.183
10 mg/kg biweekly	161 (19)	0.126 (0.090 to 0.160)	−0.047 (−0.093 to −0.001)	0.736 (0.504 to 0.994)	0.520	0.977	0.760

Abbreviations: ADCOMS, Alzheimer Disease Composite Clinical Score; CI, (bayesian) Credibility Interval; ED90, effective dose 90; NA, not applicable.

^a Smaller values are better in all 3 comparisons.

^b Probability of super-superiority is the probability dose-effect is more than SST of 0.03 greater than placebo. This was the trial’s primary analysis.

Figure 2 shows the superiority and super-superiority comparisons against placebo over time. For example, the probability of superiority for the 10 mg/kg weekly regimen at 18 months is 0.977.

Figure 2. Posterior Probabilities Over Time Since Randomization.

The super-superiority and futility boundaries refer to the probability of being greater than the super-superiority threshold (Pr(>SST)). If Pr(>SST) for the most likely ED90 exceeds the super-superiority boundary (95%) at analysis 4 or later, then analysis would recommend going to phase 3. If the Pr(>SST) for the most likely ED90 is below the futility boundary (5% for first 3 analyses and 7.5% thereafter), then analysis would recommend stopping accrual (if still accruing) and not going to phase 3. Pr(>PBO) is the probability that the dose’s benefit on Alzheimer Disease Composite Clinical Score is greater than that on placebo. Pr(>PBO) played no role in the design of the trial.

The most likely ED90 at 12 and 18 months was 10 mg/kg biweekly, which in the final posterior distributions had a mean effect size at 12 months (primary) of 0.037 (Figure 3). This estimate was larger than the SST of 0.03, but its posterior probability of exceeding 0.03 was only 0.638, less than the targeted 80%. The posterior probability of 10 mg/kg biweekly superiority to placebo at 12 months was 0.975. One minus this quantity, 0.025, is the posterior probability that 10 mg/kg biweekly is inferior to placebo (Figure 2). This is the bayesian analogue of a frequentist unadjusted 1-sided P value.

Figure 3. Final Posterior Distributions of the Benefit in Alzheimer Disease Composite Clinical Score (ADCOMS) for Dose 10 mg/kg Biweekly, the Most Likely Effective Dose (ED90), at 12 and 18 Months.

SST indicates super-superiority threshold. Benefits of the ED90 dose were compared with placebo results.

The most likely ED90 at 18 months was again 10 mg/kg biweekly. Its estimated reduction in disease progression was 0.047. The posterior probability of an effect size in comparison with placebo that is greater than SST was 0.760, which again was less than 80%. The posterior probability of 10 mg/kg biweekly’s superiority to placebo is 0.977, with analogous unadjusted 1-sided P value = .023.

Figure 3 facilitates comparing the 12- and 18-month analyses. Despite longer follow-up, the posterior distribution for the latter evinces greater spread and therefore greater uncertainty. The reason is that there is more missing data at month 18, including most of the long-term follow-up information on dose 10 mg/kg biweekly in APOE4 carriers.

Figure 4 addresses the bayesian analysis’s handling of missing data. It explains why the estimated benefit of the 10 mg/kg biweekly dose increased between months 12 and 18 despite the mandated stopping therapy for many 10 mg/kg biweekly patients. Figure 4A compares the placebo group with patients receiving the 10 mg/kg biweekly dose who did not complete treatment at their last visit. At all 5 last visits before dropping out, the mean ADCOMS was numerically better for dose 10 mg/kg biweekly. This reflects a tendency for better responders to therapy to stick with their treatment assignment unless mandated to stop it. When comparing patients who completed treatment with those who did not, the bayesian model recognizes that patients on 10 mg/kg biweekly performed better than placebo patients after they stopped therapy (Figure 4B). This was because many of the former were required to drop based on APOE status and not because of poor performance. Information available for the prospective bayesian model was more positive for 10 mg/kg biweekly than for placebo, and almost double that when restricting to patients who completed treatment.

Figure 4. Box-Whisker Plots Demonstrating Differential Effects of Multiple Imputing Values by Dose Group for Patients Who Dropped Out Before 18 Months.

In panel A, a comparison of the change from baseline in ADCOMS over time from randomization to 18 months for patients receiving placebo and 10 mg/kg biweekly suggests different reasons that patients dropped out. Dose 10 mg/kg biweekly had greater rates of administrative missingness and patients on placebo who went missing had greater levels of disease progression. In panel B, the status of all completers is shown as a horizontal black line with the box-whisker plots showing the distributions of those patients with missing values based on 10 000 imputations. The mean of the 10 000 imputed sample means of patients with imputed values is the horizontal black line within the orange box. The green box-whisker plot shows the 10 000 combined means of the completer and the imputed means. The mean in the box-whisker plot for patients in the placebo group is greater than the mean for completers, while the opposite is true for dose group 10 mg/kg biweekly. This shows the importance of bayesian modeling. The bayesian comparison of dose 10 mg/kg biweekly with placebo shows almost double the advantage as compared with restricting the sample to patients completing treatment.

Discussion

Bayesian approaches offer the ability to use accruing information in a clinical trial to most efficiently address the trial’s scientific questions. For dose-finding trials, this means focusing on doses that demonstrate benefit and are candidates for investigations in a subsequent trial or a continuation of the original trial. Lecanemab 201 results showed early that the 2 highest doses, 10 mg/kg monthly and 10 mg/kg biweekly, were outperforming the 3 lower doses, 2.5 mg/kg biweekly, 5 mg/kg biweekly, and 5 mg/kg monthly. The design algorithm adapted by assigning higher proportions of subsequent patients to the 2 better-performing doses—the placebo group was available to anchor the lower part of the dose-response curve.

Consequently, 414 of the 609 patients (68%) assigned on lecanemab doses were on the 2 higher doses. So, 2 of the 5 doses (40%) were assigned to 68% of the patients. This enabled more precise estimates where greater precision was most desirable. Although not an explicit trial goal, the average patient receiving lecanemab doses had better efficacy in comparison with equal randomization.

Adaptive designs tend to be smaller than those with fixed designs. However, they may be larger when additional information is required to achieve the trial’s objective. The trial reached its maximum sample size because it had not achieved its goal with smaller sample sizes.

The final bayesian analyses are based on the posterior distributions of change from ADCOMS for dose 10 mg/kg biweekly over placebo (Figure 3). For dose 10 mg/kg biweekly, the respective probabilities of superiority and super-superiority are 0.975 and 0.638 at 12 months and 0.977 and 0.760 at 18 months.

Limitations

This study had several limitations. As regards the clinical trial, there are limitations associated with the regulatory requirement to stop assigning dose 10 mg/kg biweekly to APOE4 carriers. Despite the substantial missing data the original prospective design and analysis were preserved to the extent possible. But although the trial’s bayesian design was robust and able to ride out the storm, attributing the proper amount of uncertainty with conclusions, there is no way to completely make up for missingness. In addition, had there been no regulatory requirement to stop dose 10 mg/kg biweekly in APOE4 carriers, the design would have greatly increased the sample size for dose 10 mg/kg biweekly, the most likely ED90, and its posterior precision. This meant the company had to make a go-to-phase-3 decision with a less powerful phase 2 trial than they had planned.

As regards our bayesian modeling, all models have limitations.²⁶ This is especially so when modeling missing data.²⁴ The trial’s bayesian design assumes that known observations can predict future observations (with uncertainty). Estimates of disease progression from previous observations are based on patients who have future observations, irrespective of treatment and irrespective of whether missingness is voluntary or mandated. Because relevant evidence is missing, it is impossible to know whether these assumptions are correct.

Conclusions

The findings of this study suggest that innovations associated with the bayesian approach can improve the efficiency of drug development and the accuracy of clinical trials, even in the context of substantial data missingness. The regulatory requirement to drop dose 10 mg/kg biweekly, the most likely ED90, had a dramatic effect on the trial. But the prospective bayesian analysis model that included multiple imputations of missing data applied unchanged. The model adjustments nearly doubled the estimated benefit of dose 10 mg/kg biweekly vs placebo in comparison with analyses restricted to trial completers. This enabled the decision to conduct the positive phase 3 trial that has recently been announced (NCT04468659), with benefits on Alzheimer disease very consistent with the results of lecanemab phase 2 trial 201.

Supplement 1.

Trial Protocol

Supplement 2.

eFigure 1. Randomization Probability by Analysis Over Time

eFigure 2. Cumulative Enrollment Status by Analysis Number

eFigure 3. Results for All Interim Analyses

eTable 1. Interim Analyses and the Adaptive Actions Available at Each

eTable 2. Baseline Characteristics

Supplement 3.

Data Sharing Statement