Table 1 Summary of estimands
Estimand | TB-specific Estimand | Composite Estimand | Assessable Estimand | Per-Protocol Estimand |
|---|---|---|---|---|
Defining characteristic | Focuses on the effect of the treatment exclusively on TB disease outcomes. | Cautious, assuming a worst-case scenario. Lack of evidence of cure is assumed to be absence of cure. | A middle group between the TB-specific and composite estimands, providing a link to previous trials | Targets the treatment effect among the (unknown) sub-group of TB patients that take an adequate treatment course. |
Intention | To disaggregate efficacy events from non-TB-related adverse events and other events. | To provide a cautious estimate of the treatment effect assuming many intercurrent events are indicative of absence of cure, following a strict “intention-to-treat” approach. | To provide a treatment effect estimate more closely aligned with previous trial analyses. | To provide the treatment effect in the group of participants who comply with key components of the protocol including treatment adherence. |
Potential target stakeholders | Product developers and asset owners, research scientists, guidelines development groups, clinicians seeking to understand the efficacy of TB treatments. | National and regional TB programs, researchers seeking to compare results against previous clinical trial data, patients seeking a practical treatment effect when deciding whether to take a given TB treatment. | Researchers seeking to compare results against previous clinical trial data. | Clinicians deciding whether to prescribe a given TB treatment. |
Uses in previous TB clinical trials | Similar to Failure or Relapse analysis in STREAM stages 1 [8] and 2 [9] | Often labeled “Strict ITT” [7] or Microbiologically Eligible [3] | Similar to a “Modified Intention-to-treat” [7, 10, 11] or “Assessable” [3] analyses. | A similar target to per-protocol analyses in previous trials, but a different estimation method more correct to the “per-protocol” label [12]. |
Estimand attributes | ||||
1. Treatments being tested and compared | Trial-specific (should align with the experimental and control/standard of care regimens offered to participants through the trial) | |||
2. Target population | Trial-specific (may be shaped by the trial inclusion and exclusion criteria) | |||
3. Participant-level endpoint | Trial-specific (protocol-defined definition to determine favorable or unfavorable long-term clinical efficacy) | |||
4. Specification of ICEs and associated handling strategies | See Table S1 for complete list of ICEs and associated handling strategies. Handling strategies: • Treatment policy • Composite • Hypothetical | See Table S1 for complete list of ICEs and associated handling strategies. Handling strategies: • Treatment policy • Composite | See Table S1 for complete list of ICEs and associated handling strategies. Handling strategies: • Treatment policy • Composite • Hypothetical | See Table S1 for complete list of ICEs and associated handling strategies. Handling strategies: • Treatment policy • Composite • Hypothetical Principal Stratum |
Estimation methods with assumptions, particularly in handling of missing or censored outcomes. | 1. Multiple imputation - Assumes the ICE occurred at random given the observed data and covariates used in the model (data are “missing at random”) 2. IPCW - No unmeasured confounders associated with censoring - Censoring is not associated with outcome determination conditional on the covariates used in the model | 1. No outcomes are missing or censored, estimation based on simple proportion. | 1. Multiple Imputation - Assumes the ICE occurred at random given the observed data and covariates used in the model (data are “missing at random”) 2. IPCW - No unmeasured confounders associated with censoring - Censoring is not associated with outcome determination conditional on the covariates used in the model | 1. Bayesian framework (for principal stratum strategy) - Applies multiple imputation for hypothetical strategy ICEs. o Assumes the ICE occurred at random (data are “missing at random”) - 3 standard assumptions for partial identifiability of estimand (joint exchangeability, monotonicity, consistency) - Specification of prior distribution reflective of model assumptions |
5. Population Summary Measure and estimation method | Risk difference - Cochrane Mantel Haenszel o Naïve approach assuming hypothetical ICEs are durable cure o MI - Kaplan-Meier o Naïve approach censoring participants with hypothetical ICEs at the time of the ICE occurrence o MI o IPCW | Risk difference - Cochrane Mantel Haenszel - Kaplan-Meier | Risk difference - Cochrane Mantel Haenszel o Naïve approach assuming hypothetical ICEs are durable cure o MI - Kaplan-Meier o Naïve approach censoring participants with hypothetical ICEs at the time of the ICE occurrence o MI o IPCW | Risk difference - Bayesian Framework o Individual membership within principal strata is based on the unobservable distribution of ICEs given the observed and counterfactual intervention assignments. The estimand is not fully identifiable, but inference is obtained by placing Bayesian priors reflective of the necessary assumptions on the probability model. |