Chapter 2: Methods in Real-World Evidence Generation — Study Design

1. Target Trial Paradigm

Author: Bradley MC

What is it used for?

Causal questions on the comparative effectiveness and safety of COVID-19 treatments and vaccines are ideally addressed in well-designed and well-conducted RCTs using protocol-based outcome ascertainment and adjudication. However, if an RCT is not feasible, ethical, or timely, or is cost prohibitive, data from observational studies using best pharmacoepidemiologic practices1,2 might offer a valuable alternative. To illustrate these study design practices, analogies to RCTs have been made: the target trial paradigm applies design principles from RCTs to the analysis of observational data to explicitly emulate the hypothetical RCT that would have been conducted. This hypothetical RCT is referred to as the target trial.3

The target trial is specified using a structured process to refine the causal goal of the research, formulate meaningful questions, and evaluate appropriate observational data sources and causal inference analysis strategies. If the target trial is specified as precisely as possible for the question of interest, it informs the design of the observational study. The inferences drawn from the observational study are then based on the principles applied in the target trial to assess a causal effect.4

A protocol specifying the target trial and detailing the emulation in observational data should include seven key domains:5

  1. Eligibility criteria
  2. Treatment strategies
  3. Treatment assignment
  4. Start and end of follow-up
  5. Outcomes
  6. Causal contrasts (intention-to-treat effect, per-protocol effect)
  7. Statistical analysis

Eligibility criteria for the observational study should be informed by the target trial, in most cases a hypothetical pragmatic trial.6 Pragmatic trials are designed to compare effectiveness of interventions in real-life practice. In contrast, typical RCTs are explanatory trials aimed to test whether the intervention works under optimal conditions.

Key considerations in emulating the target trial, which can help identify and minimize potential bias and confounding in the observational study, include:

  1. Emulation of random assignment of treatment at baseline; all relevant confounding factors should be balanced to ensure comparability (exchangeability) of the treatment cohorts.
  2. Proper specification of time zero; time zero is the point at which follow-up for outcomes begins, and it should be synchronized with determination of eligibility and assignment of treatment strategies. In the observational study, emulation of the time zero of the target trial is usually achieved by defining it as the time when an eligible patient initiates an intervention.7

Observational studies designed according to a framework that allows causal inference such as the target trial paradigm can estimate causal effects, rather than associations.8 Specifying the target trial can help with methodologic evaluation of observational studies. It can highlight deviations from ideal study conditions and allow identification of potential biases in the observational study — for example, when certain information required for emulation is not available5 — as well as any changes in eligibility criteria or other parameters as a result.9

What kinds of questions can be addressed?

The target trial paradigm can be used to evaluate important causal questions on COVID-19 treatment effectiveness and safety, and on COVID-19 vaccine effectiveness in real world settings. Target trial emulations using RWD are designed to overcome limitations seen in the COVID-19 vaccine RCTs, such as inability to perform adequate subgroup analysis due to sample size limitations, restrictive inclusion criteria, and a highly controlled setting that is difficult to replicate in real world mass vaccination programs.10 Using RWD may allow inclusion of groups often excluded from traditional RCTs, such as older adults and patients with comorbidities. Also, RWD can address questions that might be outside of traditional RCT conduct , such as duration of vaccine protection against infection, transmission, severe disease, death, and comparative effectiveness of different vaccines.11The following illustrates the use of the target trial paradigm to inform pharmacoepidemiologic studies of COVID-19 treatments.

Dagan et al.11,12 used RWD from a large health care organization in Israel to emulate a target trial evaluating the effectiveness of the BNT162b2 mRNA vaccine. Evaluated outcomes included: infection with SARS-CoV-2, symptomatic COVID-19, and COVID-19–related hospitalization, severe illness, and death among 596,618 vaccinated persons and 596,618 controls. The findings suggested high effectiveness of the BNT162b2 vaccine for preventing symptomatic COVID-19 in the real world, and high effectiveness for preventing outcomes such as hospitalization, severe illness, and death; similar to the vaccine efficacy reported in the RCT.13 The large observational study (1,192,236 participants compared to 21,720 in the RCT) allowed examination of outcomes in specific subpopulations that the RCT was not sufficiently powered to evaluate, such as those with multiple coexisting conditions. The finding that vaccine effectiveness might be slightly lower in those with more coexisting conditions is notable. However, a limitation to the use of RWD was the speed of vaccination, which contributed to frequent censoring of matched unvaccinated controls, especially among older individuals, and a reduction in the average follow-up time.

In the absence of data from RCTs, Gupta et al.14 used RWD from a large multi-site cohort, the Study of the Treatment and Outcomes in Critically Ill Patients With COVID-19 (STOP- COVID),15 to emulate a target trial. The study estimated the effect of early treatment with tocilizumab on mortality in COVID-19 patients in the intensive care unit (ICU). Again, the large size of the emulated study allowed subgroup analyses not always possible in the RCTs that were subsequently conducted. The results showed that patients treated with tocilizumab in the first 2 days of ICU admission had an almost 30% lower risk of death compared with those not treated with tocilizumab. The United Kingdom RECOVERY RCT later reported similar results.16 However, other RCTs did not show the same mortality benefit.17–19

Similarly, in the absence of evidence from RCTs, Al-Samkari et al.20 emulated an RCT as a target trial using RWD on COVID-19 patients admitted to ICUs in the US. In this study, survival was compared in patients who received therapeutic anticoagulation in the 2 days after ICU admission to those who did not receive therapeutic anticoagulation. Those who received anticoagulation in the first 2 days of ICU admission had similar in-hospital survival compared with those who did not. These results supported the recommendations of several professional societies against broad empirical therapeutic anticoagulation in patients with COVID-19.21,22

Gatto et al. used HealthVerity data to emulate a hypothetical target trial to examine the effectiveness of dexamethasone in preventing 28-day mortality among hospitalized patients with moderate to severe COVID-19.23 While the RECOVERY RCT already examined this question in the UK, the authors sought to emulate RECOVERY as a hypothetical target trial with varied design elements, and tailored for US practice and US patients.23

Using data from routine care in France, the target trial paradigm was also used to inform an observational study on the effectiveness of hydroxychloroquine for patients hospitalized with a COVID-19 infection and requiring supplemental oxygen. The findings suggested that hydroxychloroquine treatment at 600 mg/day added to standard of care was not associated with a reduction of admissions to ICUs or death 7 days after hospital admission compared to standard of care alone.24,25 These findings were consistent with data from RCTs that found hydroxychloroquine to be ineffective for treatment of COVID-19.26

What are the benefits and limitations of this design?

Emulating a hypothetical target trial can improve the conduct of observational studies, decreasing some common biases and confounding by emulation of random assignment and correct specification of time zero. For a robust observational study, other issues such as attrition, changes in treatment, and misclassification of outcomes should be accounted for, but these are discussed elsewhere. Misspecifying time zero can lead to selection and time-related biases such as immortal time bias.27 For example, an observational study on postmenopausal hormone therapy and heart disease reported >30% lower risk of the outcome in current users vs. never users,28 while an RCT reported >20% higher risk.29 Prevalent hormone therapy users were included in the observational design, whereas the RCT only followed incident users. As prevalent users were depleted of patients who were susceptible to thromboembolic risk imposed by postmenopausal hormone therapy, which would manifest early during exposure, they were less likely to experience the outcome.30 When the observational data were re-analysed using a new-user design (resembling a trial), the findings aligned with those of the RCT.31   Incorrectly specifying time zero and including prevalent users33,34 can also cause postbaseline information (collected during treatment) to be used in defining baseline characteristics.7 Constructing causal diagrams when specifying the target trial helps identify baseline confounders (non-colliders [See Chapter 3]) that should be balanced and colliders that should be left uncontrolled to prevent collider bias.35 The process of specifying a target trial helps to explicitly define the causal question, better understand the data that are required, and generally improve the quality of observational studies.5

Target trials are pragmatic, and therefore cannot always emulate key elements of an ideal RCT. For example, RWD cannot be used to emulate a placebo-controlled trial (an active comparator is generally needed), a trial with blinding (patients are typically aware of their treatment), treatment strategies that do not exist in the real world (enforcing adherence to a treatment), or any other type of monitoring that is not reflective of real-world practice.3 However, the pragmatic nature of the emulated target trial is not necessarily considered a limitation when the aim is to compare treatments to reflect use in the real world. RWD needed for the target trial emulation might not be available in the chosen data source (e.g., availability of clinical laboratory results for a specific disease biomarker). Data availability might also limit the ability to adequately balance comparison groups with regard to critical outcome risk factors that are associated with treatment assignment, thus acting as confounders.5