SSTARLET Trial: Bayesian TB Prevention

• SSTARLET is a Bayesian adaptive platform trial that evaluates shorter, safer TB preventive regimens in people with latent infection.
• The trial employs a non-inferiority design with cluster-randomized methods, robust MAP priors, and outcome-specific interim dropping to enhance regimen comparisons.
• It introduces an innovative simulation-based sample-size determination method to efficiently calibrate trial operating characteristics under complex adaptive rules.

SSTARLET is an ongoing Bayesian adaptive platform trial for tuberculosis preventive therapies in people with latent TB infection, formally expanded as “Shorter and Safer Treatment Regimens for Latent Tuberculosis” and registered as ClinicalTrials.gov ID NCT06498414. In the design literature, it is described as a cluster-randomized trial conducted in Canada and four additional countries with moderate to high TB incidence, with design calculations performed at the individual level because clustering is expected to be minimal in a similar population in which most clusters were single-member households. The trial’s stated purpose is to evaluate shorter and safer regimens for tuberculosis preventive therapy through a Bayesian adaptive non-inferiority platform using a common control and the possibility of arm addition and arm dropping (Hagar et al., 16 Jul 2025).

1. Clinical setting and trial identity

SSTARLET is situated in the setting of latent TB infection and tuberculosis preventive therapy (TPT). Its motivation is explicitly operational and clinical: existing preventive regimens are described as effective but often too long and poorly tolerated, which impedes global scale-up. The trial is therefore designed to compare alternative regimens intended to be shorter and safer while preserving acceptable performance relative to standard care (Hagar et al., 16 Jul 2025).

The trial is presented as a platform trial, meaning that multiple experimental regimens are evaluated against a common control group, with the possibility of modifying the set of active experimental arms during conduct. In the SSTARLET design paper, this platform structure is not treated as an abstract methodological preference; it is tied to concrete requirements, including efficient comparison of multiple regimens, mid-trial addition of a new experimental arm, interim dropping of inferior regimens, accommodation of a fixed delay between interim enrollment and outcome availability, and use of historical borrowing through robust MAP priors (Hagar et al., 16 Jul 2025).

A common misconception is to treat SSTARLET as a completed efficacy trial. The available arXiv material instead describes it as an ongoing trial and focuses on its design and calibration rather than reporting final clinical results. This suggests that the primary scholarly importance of SSTARLET, in the current preprint record, lies in its role as a real-world motivating case for Bayesian platform-trial methodology.

2. Regimens, control structure, and endpoints

The design paper specifies several self-administered daily regimens, with one control and multiple experimental arms (Hagar et al., 16 Jul 2025).

Component	Specification	Role
4R10	4-month daily rifampin	Control
2R20	Double-dose rifampin for 2 months	Experimental arm
1LP	Levofloxacin and rifapentine for 1 month	Experimental arm
Third experimental arm	Added mid-trial	Experimental arm

The primary and secondary outcomes are defined with unusual specificity for a platform-design paper. The primary outcome is the incidence of severe treatment-related adverse events (AEs) leading to permanent discontinuation, adjudicated by a blinded panel. The secondary outcomes are treatment completion, defined as taking at least $80\%$ of doses within $150\%$ of intended time, and tolerability, based on self-reported symptoms not severe enough to stop treatment (Hagar et al., 16 Jul 2025).

The trial is formulated as a non-inferiority platform. The non-inferiority margins are explicitly outcome-specific: $\delta_{m,1} = 0.04 \quad \text{for AEs},$

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$

These margins establish that SSTARLET is not optimizing a single scalar notion of regimen quality. Instead, it separates severe safety, completion behavior, and tolerability, while making the formal final non-inferiority declaration on the AE endpoint alone in the design formulation. A plausible implication is that the platform treats severe treatment-related discontinuation as the decisive endpoint for confirmatory inference, while using the other endpoints more prominently in interim adaptation.

3. Adaptive platform architecture

SSTARLET begins with an unequal allocation ratio of

$1:2:2$

across control 4R10, 2R20, and 1LP, respectively. The paper attributes this unequal allocation to greater prior information for 4R10. When the third experimental arm is added, 50\% of new participants are assigned to the new arm and the remaining 50\% are split equally among the existing arms. After interim decisions become available, allocation associated with dropped original experimental arms is redistributed among the remaining eligible arms (Hagar et al., 16 Jul 2025).

The interim architecture is driven by a fixed outcome lag. The motivating example is that interim analysis is triggered at about 700 enrolled, but the analysis is available only after another 300 participants are enrolled, giving results at about 1000 enrolled. In the formal notation, the interim analysis uses outcomes from the first

$n_1 = n$

participants, while the final analysis occurs at

$n_2 = c_2 n.$

In the simulation study,

$c_2 = 2.5.$

This lag is central to the design problem because arm 3 is added when interim enrollment is met, not when interim results are available. Consequently, there is a period in which enrollment continues under an expanded arm set before dropping decisions for the earlier experimental arms can be implemented. The paper identifies three SSTARLET-specific complexities arising from this framework: arm addition at interim trigger, response-adaptive randomization induced by arm dropping, and the fixed outcome lag (Hagar et al., 16 Jul 2025).

The active set of retained original experimental arms after interim is denoted

$$\mathcal{S} \in \left\{ \emptyset, \{1\}, \{2\}, \{1,2\} \right\}.$$

Arm 3 is added regardless of interim data and enters the final comparison set. This formulation makes clear that the platform is not merely multi-arm; it is explicitly state-dependent, with post-interim inference conditioned on the realized active set.

4. Bayesian decision framework

The inferential structure is built on pairwise comparisons of each experimental arm against the common control, indexed by outcome. Arms are indexed by

$j \in \{0,1,2,3\},$

with $150\%$0 as control, and outcomes by

$150\%$1

with $150\%$2 for AEs, $150\%$3 for non-completion, and $150\%$4 for non-tolerability. For each arm-outcome combination, $150\%$5 denotes the true event probability (Hagar et al., 16 Jul 2025).

The non-inferiority hypotheses are

$150\%$6

At interim, the posterior inferiority probability is

$150\%$7

An original experimental arm $150\%$8 is dropped if, for any endpoint $150\%$9,

$\delta_{m,1} = 0.04 \quad \text{for AEs},$0

The calibrated interim thresholds are reported as

$\delta_{m,1} = 0.04 \quad \text{for AEs},$1

At final analysis, the key posterior probability for non-inferiority on the primary endpoint is

$\delta_{m,1} = 0.04 \quad \text{for AEs},$2

and non-inferiority is declared when

$\delta_{m,1} = 0.04 \quad \text{for AEs},$3

The selected thresholds are

$\delta_{m,1} = 0.04 \quad \text{for AEs},$4

with the practically decisive final threshold

$\delta_{m,1} = 0.04 \quad \text{for AEs},$5

The paper states that $\delta_{m,1} = 0.04 \quad \text{for AEs},$6 was chosen to control FWER, whereas $\delta_{m,1} = 0.04 \quad \text{for AEs},$7 and $\delta_{m,1} = 0.04 \quad \text{for AEs},$8 were heuristic choices for desirable operating behavior (Hagar et al., 16 Jul 2025).

Outcome models use a binomial likelihood and Beta prior. For most arm-outcome pairs,

$\delta_{m,1} = 0.04 \quad \text{for AEs},$9

For AE rates in the control and 2R20 arms, the design uses a robust MAP prior,

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$0

with overall borrowing weight reported as

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$1

The historical borrowing is said to be informed by four previous clinical trials and to be particularly consequential for 2R20 in the simulations (Hagar et al., 16 Jul 2025).

5. Sample-size determination and operating characteristics

The SSTARLET design paper addresses a specific computational problem: regulatory agencies require Bayesian trial designs to be evaluated by estimating the sampling distribution of posterior probabilities via Monte Carlo simulation, but exhaustive simulation across all candidate design configurations is expensive in a platform setting with multiple endpoints, arm addition, arm dropping, variable allocation, and informative priors (Hagar et al., 16 Jul 2025).

The methodological contribution is an efficient approach that models the joint sampling distribution of posterior probabilities across multiple endpoints and trial stages using simulations conducted at only two sample sizes. The theoretical justification is that, under the stated asymptotic regime,

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$2

and the logit of posterior-probability quantiles evolves approximately linearly in $$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$3. The corresponding limit statement is

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$4

In practical terms, this permits interpolation or extrapolation of operating characteristics across candidate sample sizes without rerunning the full trial simulation at each value (Hagar et al., 16 Jul 2025).

The design targets are

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$5

corresponding to FWER and power targets. The efficient sample-size determination method is anchored at

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$6

with direct simulation on the grid

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$7

Both the efficient method and the full simulation use

$$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$8

replicates, and posterior probabilities are approximated using 1000 posterior draws (Hagar et al., 16 Jul 2025).

The paper’s concrete recommendation for SSTARLET is: $$\delta_{m,2} = \delta_{m,3} = 0.10 \quad \text{for non-completion and non-tolerability}.$$9 with

$1:2:2$0

These are described as the recommended sample sizes with observed outcomes at the interim and final analyses. The selected design is calibrated under a conservative scenario in which all three experimental arms are Clearly Acceptable, since few or none are then dropped at interim and information is spread across the most active arms. Within that conservative setting, 1LP is identified as the limiting case because it receives less allocation than the newly added arm 3 and does not benefit from the informative MAP prior (Hagar et al., 16 Jul 2025).

The computational comparison is also explicit: each estimated gray curve via the efficient method required about 22 minutes, whereas each black dotted direct-simulation curve over nine values of $1:2:2$1 required about 1 hour 40 minutes. This suggests that the main innovation is not a different inferential target, but a more tractable way to calibrate a complex Bayesian platform under real logistical constraints.

6. Interpretation, scope, and relation to adjacent literature

The present arXiv record supports a precise but narrow characterization of SSTARLET. It identifies the disease area, regimen structure, outcome definitions, Bayesian decision rules, and recommended design parameters, but it does not provide completed trial results, comparative efficacy findings, or a post hoc clinical interpretation of observed outcomes. It also narrows attention to a two-analysis window of what could in principle be a continuing platform, so the “final analysis” in the design paper is the final analysis of that modeled window rather than necessarily the terminal analysis of the platform as a whole (Hagar et al., 16 Jul 2025).

Two adjacent papers are relevant mainly for context rather than direct trial description. The paper “SECRET: Semi-supervised Clinical Trial Document Similarity Search” develops a protocol-level retrieval method that could be used to find historical trials similar to SSTARLET on the basis of title, disease, intervention, keywords, primary outcome measures, and eligibility criteria, but it explicitly contains no trial-specific findings about SSTARLET itself (2505.10780). A plausible implication is that SECRET could support protocol benchmarking or analog-based risk review for SSTARLET without altering the trial’s formal design.

Similarly, “Sample-targeted clinical trial adaptation” presents a Bayesian method for sample size adjustment by deciding which participants should be removed so that trial distortion is minimized, using subgrouping based on treatment assignment and participant belief about assignment. That paper is conceptually relevant to adaptive-trial methodology, but it does not explicitly use the name “SSTARLET” and does not define an acronym matching SSTARLET (Arandjelovic, 2014). It should therefore not be cited as direct evidence about SSTARLET itself.

Taken together, the current preprint literature portrays SSTARLET as a concrete, ongoing Bayesian adaptive non-inferiority platform trial in latent TB prevention whose main scholarly visibility derives from the methodological demands it imposes: common-control multi-arm comparison, outcome-specific interim dropping, delayed implementation of interim decisions, arm addition, and robust historical borrowing. The published design work presents SSTARLET less as a source of completed clinical findings than as a statistically intricate platform whose architecture motivated a new approach to Bayesian sample-size determination and operating-characteristic evaluation (Hagar et al., 16 Jul 2025).