Future Directions in Clinical Trials and Intention-To-Treat Analysis: Fulfilling Admirable Intentions Through the Right Questions

ABSTRACT

We call for clinical trials researchers to carefully consider questions about use of intention-to-treat (ITT) analysis and per protocol analysis. We discuss how questions about efficacy and mechanisms of efficacy are appropriately answered through the application of per protocol analysis. ITT analysis is well-suited and appropriate for addressing questions related to treatment effectiveness, typically adherence to the treatment with respect to an outcome. While guided by admirable intentions, ITT analysis is often not guided by the right questions, leading to ITT misapplication. We address additional misconceptions that often lead to ITT misapplication, including issues relating to treatment noncompletion and violation of random assignment. We further highlight future directions and implications, particularly that future clinical child and adolescent research trial designs will be increasingly characterized by hybrid trials that combine elements of efficacy, effectiveness, and implementation research, where ITT and per protocol analysis will be appropriately applied to answer the right questions.

Asking the Right Question

Most readers are likely familiar with the adage that failure to learn from the past will lead to repeated mistakes in the future. With this in mind, we begin our Future Directions article with brief reflections back from whence we came and lessons learned with respect to per protocol and intention-to-treat (ITT) analysis. Our reflections then return us back to the future, encapsulated by the less familiar quote of psychoanalyst Carl Jung, “To ask the right question is already half the solution to a problem” (1981, p. 23). This quote is pertinent in understanding why per protocol and ITT analysis have been (mis)applied in clinical child and adolescent randomized control trials.

Before proceeding, it is important we note what each of these two analytic approaches is intended to accomplish (if the right question is asked). Per protocol analysis aims to determine the effect on an outcome of receiving the assigned treatment as specified in the protocol (Tripepi et al., 2020). It achieves this goal by including in analysis patients who received and completed the assigned treatment (i.e., patients who received a full dose of treatment). In ITT analysis, the aim is to determine the effect on an outcome of assigning a given treatment by including in analysis all patients assigned to treatment, regardless of whether they received or completed the assigned treatment (Tripepi et al., 2020).

Both types of analyses address important and valid questions, but they are different questions. Per protocol analysis is appropriate for addressing questions related to treatment efficacy and efficacy mechanisms, albeit like any analytic method, subject to limitations. ITT analysis, in contrast, is well-suited and appropriate for addressing questions related to treatment effectiveness, particularly adherence to the treatment with respect to an outcome. A main takeaway of this article is that ITT analysis is often not appropriate for addressing questions related to treatment efficacy, including randomized controlled clinical efficacy trials, especially mechanism-based clinical efficacy trials. Correct identification of one’s research questions (efficacy/mechanisms vs. effectiveness) points toward the appropriate approach to analyses. It is thus “already half the solution” to moving closer in fulfilling one’s admirable intentions.

If the question is whether a treatment has efficacy – does the treatment work when implemented with integrity and patients receive the “full dose,” then the randomized controlled clinical trial (RCT) needs to be designed to answer that distinct question and analyzed per protocol. An important caveat: ITT can be appropriate for efficacy trials when adherence is high and given faithful treatment implementation (Jaccard, 2024). However, this is rare in clinical science RCTs including child and adolescent RCTs. If the question is about mechanisms of efficacy – how does the treatment work? – then the trial needs to be designed to answer that distinct question and analyzed per protocol. If the question is about effectiveness – does the treatment work under real world circumstances wherein patient adherence and clinical variables are not rigorously controlled? – then the trial needs to be designed to answer that distinct question and analyzed with ITT. Lastly, if researchers aim to address more than one of these questions within the same study (i.e., efficacy, mechanisms of efficacy, and/or effectiveness), then the trial needs to be designed to address each respective question and analyzed with the appropriate methods. That may require the use of both per protocol and ITT analyses.

The future of clinical trials research in child and adolescent mental health will likely be increasingly characterized by hybrid designs. Such designs combine elements of efficacy, effectiveness, and implementation research (e.g., Landes et al., 2020; Onken, 2019), where both ITT and per protocol analysis need to be appropriately applied to answer the respective (right) questions. In our view, it is incumbent on clinical child and adolescent trials’ researchers to build a science around these distinct research questions. It is also our view that clinical child and adolescent psychological science has much to gain from a continued focus on the distinct questions addressed by efficacy trials and effectiveness trials and that the advances to be gained by hybrid designs are exciting. From a future directions’ perspective, moving the science in this way would permit drawing valid inferences and conclusions and synthesizing new knowledge accordingly. However, this is not the current situation. Rather than asking the right (distinct) questions, researchers often have muddled the questions leading to a clinical child and adolescent RCT research literature that is often muddled as well.

We elaborate below on these ideas by illustrating how we have strived to distinguish efficacy and efficacy-based mechanism research questions and not to muddle them with effectiveness. Our illustration focuses on our 30+ years program of research that aimed to answer the question about whether and how cognitive-behavioral therapy (CBT) outcomes for anxiety disorders in children and adolescents (youth) may be enhanced with parent involvement. While our focus is on clinical child and adolescent randomized clinical trials, particularly our anxiety trials, the points we make and the conclusions and implications we draw are pertinent to randomized trials in other populations (e.g., adults) and disciplines (e.g., psychiatry).

Structure and Organization (And Some Irony)

We present first some historical background regarding why it was considered important in the first place to determine whether psychological treatments produce positive outcome, that is, are efficacious. Next, we discuss the constructive shift away from focusing solely on whether therapeutic change is produced (efficacy) and to how change is produced (mechanisms). However, concomitant with the constructive shift or step forward, it is a step backward. The step backward entailed evaluating questions of both efficacy (i.e., whether the treatment works) and mechanism-based efficacy (i.e., how does the treatment work) with ITT analysis in studies contaminated by non-adherence. We refer to our research on studying the role of parents in child anxiety CBT to illustrate these points. The principles and concepts we illustrate through our research can be readily applied to other efficacy and mechanism-based efficacy trials, as well as implementation trials that test the effects of an intervention on implementation fidelity (e.g., Beidas et al., 2012). The latter represent special cases of efficacy trials that ask whether the intervention designed to increase implementation fidelity produces an effect on fidelity when patients receive the intervention as intended.

It is worth noting the irony that in 2024 we need to recommend per protocol analysis in a Future Directions piece. It is ironic because in the early stages of efficacy trials research, virtually all child and adolescent RCTs used per protocol analysis, not ITT analysis (see Silverman et al., 2008). For reasons that are difficult to nail down, sometime around the 2000s, ITT largely replaced per protocol analysis in statistical evaluation of most clinical psychology trials including clinical child and adolescent efficacy trials, mechanism-based efficacy trials, and effectiveness trials (the latter being appropriate application of ITT). This replacement is a major catalyst for our writing this article.

The Beginning of Asking Whether Psychological Treatments are Efficacious

The authors are not old enough to have read at the time of publication the seminal literature reviews about the (non)effects of adult psychotherapy by Eysenck (1952) and child psychotherapy by Levitt (1957). We are old enough though to have these articles assigned as required readings in our graduate courses, and the feelings of self-doubt and confusion they evoked. Here, we were – learning to help those who suffer from mental health problems while learning concomitantly that the benefits of psychotherapy are no better than that which occurs naturally over time with no therapy whatsoever. Childhood problems especially were viewed as often transient phenomena that would be outgrown over development.

Under such circumstances, what is the point of learning about child, adolescent, and adult mental health problems if there are no viable psychotherapeutic approaches to alleviate these problems? For some, this question could be resolved by leaving the profession all together. For others, it could be resolved by concluding psychotherapy is more art than science and is too complicated to study scientifically. Yet for others, it could be resolved with the creation of a body of empirical research studies that would refute these claims.

We oversimplify, but it does provide historical background for a major paradigm shift that occurred in clinical science research, especially psychotherapy research. The paradigm shift led to a generation of clinical scientists conducting thousands of well-controlled therapy outcome studies using the medical model’s experimental gold standard, the RCT. These outcome studies appeared in major journals in clinical psychology, including Journal of Consulting and Clinical Psychology and this journal (e.g., as well as summaries of these empirical studies through the journal’s Special Issues that supported use of various treatments for specific adult and child and adolescent psychological disorders; see Chambless & Ollendick, 2001; Silverman & Hinshaw, 2008). The cumulative data and knowledge generated underscore the treatments that have been studied most, the strength of the evidence for these treatments, and under what conditions and for whom are some treatments better suited than other treatments in clinical settings (e.g., Kazdin, 2011).

We count ourselves among those who believe the field is currently in a stronger and better place with the cumulative knowledge and advances that have been made through the evidence-based treatment movement compared with when we first read Eysenck (1952) and Levitt (1957). We further appreciate the rise of implementation science, which involves introducing evidence-based practices into community and clinical settings, as well as the targeted distribution of information (i.e., dissemination; Silverman et al., 2004). The rise of implementation science has facilitated these critical efforts to disseminate knowledge about and training in evidence-based treatments (e.g., Substance Abuse and Mental Health Services Administration; https://www.samhsa.gov/resource-search/ebp; Society of Clinical Child & Adolescent Psychology; https://effectivechildtherapy.org/).

Despite these advances, there is no denying the legitimate critique that the majority of RCT samples contain an inadequate number of patients from historically marginalized backgrounds, raising questions about the generalizability of RCT findings across diverse samples (e.g., Pina et al., 2003). Beyond questions about generalizability, there exists a long and painful history of systemic health disparities that still make the availability of evidence-based treatments nearly impossible for far too many. We are optimistic nonetheless that current efforts to improve equity and representation will continue in the future, yielding as well a more inclusive and diverse clinical science in the future (see e.g., National Institute of Mental Health’s [NIMH] Strategic Framework for Addressing Youth Mental Health Disparities). Our optimism is indicated, too, by our next section.

Directions for Future Research: Slowly but Surely the Best is Coming

Almost two decades ago, in 2006, the first author presented her Division 53 Presidential Address entitled, The Best is Yet to Come: Advancing Research and Practice in Clinical Child and Adolescent Psychology, at the American Psychological Association convention in New Orleans. The presentation and its title (a shoutout to Frank Sinatra) underscored the tremendous progress that had been made in developing and evaluating randomized controlled efficacy trials, largely stimulated by Eysenck (1952) and Levitt (1957). This progress was applauded and reflected the zeitgeist of the 1990s, which was to address the critical, initial, and right question about efficacy (i.e., whether a treatment produces change).

The Address emphasized that to continue to make progress and advance child and adolescent RCT research, a shift was required. This shift would require the field to move from randomized controlled efficacy trials to randomized controlled explanatory or mechanism-based efficacy trials. Mechanism-based efficacy trials provide insights into the relationships between outcomes and potential determinants of those outcomes while at the same time providing insights into changing those determinants, all without the clutter of non-adherence and poorly implemented protocols.

The Address further delineated novel design and implementation strategies needed to achieve explanatory trials including shifting from measuring outcomes exclusively (as per efficacy trials) to measuring putative mediators or mechanisms of change. This shift would require tandem work as well on theory construction (Jaccard & Jacoby, 2020) and training in and understanding of complex statistical analyses. The resultant knowledge from mechanism-based efficacy trials would lead to leaner, more efficient, more beneficial, and more precise psychotherapeutics. Such knowledge would allow for empirical establishment of the specific components that ought to be included in each treatment to produce therapeutic improvement, and those components that ought to be excluded because they are superfluous.

Thus, mechanism-based efficacy trials have more than theoretical and scientific utility. They are key to facilitating dissemination and implementation of treatments in community settings. This is because the training of clinicians could be targeted/focused on what is essential for therapeutic change. Altogether, this would yield a “less is more” perspective, of which we and others are fans (Pettit et al., 2017; Price et al., 2017; Schleider et al., 2020; Silverman et al., 2016). Of course, we are fans when the “less” is grounded in clinical science theory and empirical data.

Of further note, emphasis on mechanisms became a cornerstone of NIMH funding priorities in the 2000-teens. For example, in a 2014 article in JAMA Psychiatry entitled, NIMH Clinical Trials: New Opportunities, New Expectations, Insel (Director at the time) and Gogtay (2014) stated that moving forward, “Every trial will need to include a mediator that tests the mechanism of action” (or “engaging targets”).” NIMH around the same time made significant changes in clinical trials’ grant funding mechanisms, which laid out the different stages for identifying and confirming “targets” or mechanisms (Insel, 2015). While there is likely high agreement among scientists on the importance of understanding mechanisms that produce positive therapeutic outcome, we believe it is safe to say that agreement drops when it comes to the specific mechanisms that scientists choose to emphasize and the resultant measurement approaches used. From our perspective, if there is compelling theory and data to support targets or mechanisms and measurement approaches then these hold promise and therefore are worth pursuing. There are many different scientific pathways (literally; metaphorically) to advance clinical science including clinical child and adolescent psychology.

Strategic Use of per Protocol Analytic Strategies

In the 2006 Presidential Address, current knowledge about parent involvement in their child’s CBT for anxiety was offered as a case example of how the best is yet to come, and how to bring the best about. A directional shift in zeitgeist was recommended. Instead of conducting more efficacy or “horse-race” trials that pit treatments against one another, it was time to shift to mechanism-based trials. Such trials would ask: How might parents be involved in their child’s CBT to enhance outcomes relative to individual youth anxiety CBT? These questions are about both efficacy (does parent involvement enhance outcomes?) and mechanisms (how does parent involvement enhance outcomes?).

If one wanted to develop and test the efficacy of involving parents in CBT and to isolate/identify the mechanism by which parent involvement impacts youth anxiety, does it make sense to have a research design where significant numbers of patients in the experimental arm do not receive the intervention per protocol due to non-adherence or treatment dropout? We think not. The goal is to answer the questions regarding whether involving parents reduces youth anxiety (efficacy) by changing the parent behaviors targeted in the intervention (mechanism). ITT analyses cannot address such questions, except under rare cases in which adherence is very high.

By contrast, per protocol analyses can address such questions, albeit given certain assumptions (just as ITT analyses require assumptions). We highlight our recent youth anxiety RCT that evaluated two CBT arms with parent involvement and an individual youth CBT arm as an example of the application of per protocol analyses in a mechanism-based efficacy trial (Silverman et al., 2022). Our study was driven by a stubborn, challenging conundrum: Meta-analyses showed that parent involvement in their child’s CBT failed to show enhanced improvement compared with no parent involvement (e.g., Breinholst et al., 2012; In-Albon & Schneider, 2007; Reynolds et al., 2012; Silverman et al., 2008; Spielmans et al., 2007; Thulin et al., 2014). Adding to the conundrum is Peris and colleagues’ (Peris et al., 2021) even more recent meta-analysis conclusion that CBT with parent involvement “did not confer advantage over individual CBT” (p. 287).

We thus used a dismantling design to address how parental involvement influences outcome by investigating distinct parenting components contained within each of two distinct CBT + parent arms as mediators. One arm trained parents in reinforcements skills and the other arm trained parents in relationship skills. Our findings revealed that on average youths in either one of the two CBTs with parent involvement had lower anxiety than those in individual youth CBT. The dismantling design allowed us to further demonstrate efficacy mechanisms that mediated youth anxiety reduction (e.g., decreases in parent use of negative reinforcement targeted in the reinforcement arm; see Silverman et al., 2022 for full details). We mention this trial here not only because of its novel findings that two distinct CBTs with parent involvement enhanced individual youth CBT, but because it represents an RCT that asked the right question to derive a solution about mechanisms of parent involvement to enhance CBT youth anxiety outcome.

We recognize that the use of per protocol analysis in our trials relied on traditional per protocol analysis that often is referred to as “naïve” per protocol analysis rather than more modern per protocol analysis strategies. Naïve per protocol frameworks analyze data from treatment completers without considering variables that may influence treatment noncompletion. Modern per protocol extends the approach by identifying variables, a priori, which are expected to influence treatment noncompletion, gathering data on these variables, and including these variables in the modeling process to adjust for effects of potential compromises to randomization. We refer readers to Jaccard (2024) for a description of the more modern per protocol approaches. They include direct covariate modeling or principal stratification (Funk et al., 2011; B. Wang et al., 2019), inverse probability treatment weighting (Austin & Stuart, 2015; Robins, 1986), complier average causal effect (CACE) analysis (Ashworth et al., 2020; Sobel & Muthén, 2012), G computation (Snowden et al., 2011; Vansteelandt & Keiding, 2011), and targeted maximum likelihood analysis (Luque‐Fernandez et al., 2018; Schuler & Rose, 2017). If we knew then what we know now we would have taken advantage of these approaches. Nonetheless, we were careful to document potential biasing effects due to treatment dropout and concluded that treatment noncompletion was mostly random and did not systematically differ across arms. We also included a small set of covariates to minimize potential compromises to randomization. As noted though, we did not identify these variables in a rigorous a priori sense that is normative for modern per protocol approaches.

Strategic Avoidance of ITT Strategies in Efficacy Research

While the field has moved forward by asking questions about mechanisms that underlie efficacy and not simply the horseraces, many current efficacy and mechanism-based trials use ITT analysis. ITT analysis is so entrenched in the field that we have seen multiple journal and grant reviews deem any trial that did not use ITT analyses as fatally flawed and warranting rejection.

In the sections that follow, we provide the reasons why this entrenched view, though of admirable intention, leads to failure to fulfill intentions. Here are the three main reasons. One, which we noted above but elaborate on below because it is so important: ITT analyses are appropriate for effectiveness trials, not for efficacy trials and especially mechanism-based efficacy trials (unless adherence and implementation are high). Two, ITT analysis often requires having posttest and follow-up data on treatment noncompleters. Three, critiques of per protocol analysis as violating random assignment are often over-stated and there are reasonable methods to address the problem. We elaborate on each of these three points below. Again, we refer readers to Jaccard (2024, Chapter 27) for detailed delineation of statistical approaches used to address each point.

ITT Analyses Often are Suboptimal for Efficacy and Mechanism-Based Efficacy Trials; They are Appropriate for Effectiveness trials

As the well-known statistician Dallal (2012) put it colloquially regarding medication trials, the efficacy question asks, “what happens if people take this stuff?” and the effectiveness question asks, “what happens if you hand this stuff to people and [ask]/tell them to take it?” Underlying Dallal’s (2012) point is that ITT analyses typically underestimate the effect of a treatment on outcomes regarding efficacy, tend to yield biased estimates of the efficacy-based effects of a treatment on mechanisms or mediators, and can yield biased estimates of the effects of a mediator on an outcome. All this led Dallal to conclude, “intention-to-treat analysis, as it is often used, is a fraud.”

The essence here is that generalizing to a per protocol population is not the same as generalizing to an ITT population, though both are important. This is because ITT analyses confound efficacy and adherence to treatment protocol. As elaborated in Jaccard (2024), the ultimate effect of a treatment on an outcome can be thought of as a multiplicative function of (1) treatment efficacy, (2) proper clinic implementation of the treatment, and (3) proper adherence to the treatment protocol on the part of patients. If one’s interest is in understanding efficacy, then it is important to remove the noise and biases due to improper clinic implementation and clients’ lack of adherence to the protocol. ITT analyses confound three types of mechanisms: (1) those that make a treatment efficacious, (2) those associated with implementation fidelity, and (3) those associated with adherence/completion. If the focus of a study is on efficacy and the mechanisms that are responsible for efficacy, it is important to eliminate the “noise” created by the latter two mechanisms. Per protocol analysis is appropriate for efficacy trials, including mechanism-based efficacy trials, precisely because it controls for these sources of noise.

Our CBT plus parent involvement studies demonstrated this: Parent-based CBTs changed mediators/mechanisms they were hypothesized to change to reduce youth anxiety and these mechanisms were related to youth anxiety reduction, in an efficacy sense (Silverman et al., 2022). ITT analyses typically cannot address such questions

Appropriate ITT Analysis Requires Having Posttest and Follow-Up Data on Treatment Noncompleters

Much research claims to conduct ITT analyses yet they often lack the data necessary to do so appropriately. This is likely due in part to the entrenched belief that one must conduct ITT analyses to be a legitimate RCT (i.e., if you randomize, you must analyze) and because the limitations of ITT methods are not widely recognized. Principled ITT analysis requires that posttest and follow-up data be collected on treatment noncompleters (S. Wang & Hu, 2022). Few studies do this and if they do, the amount of follow-up data on treatment noncompleters, including in clinical child and adolescent trials, is typically insufficient.

Perhaps researchers do not collect posttest data on treatment noncompleters in part because it can be exceedingly challenging, sometimes impossible, to obtain data from individuals who have discontinued their participation in an RCT. Another challenge (and disincentive) is that requests on grant applications to obtain these data can increase substantially an application’s budget, which already is usually tight. As a result, researchers instead attempt to conduct ITT analyses by applying modern missing data algorithms, such as multiple imputations, full information maximum likelihood (FIML), or hot deck methods. Although well-intentioned, the application of ITT analyses in such instances can be problematic. For example, if we apply FIML to an observed data set that only includes treatment completers, then how treatment noncompleters respond at posttest cannot be part of the model estimation process (Little et al., 2012).

Matters become more complicated if, without posttest data on treatment noncompleters, researchers “guess” outcome scores of noncompleters and use these “guesses” in their data analysis via strategies like last observation carried forward (LOCF), which make unverifiable and questionable assumptions (Newgard & Lewis, 2015). For example, LOCF assumes all posttest scores of treatment noncompleters are identical to their scores obtained before dropping out (often their baseline scores but could be mid-treatment, or some other session for others), an assumption that is unrealistic and can produce biased estimates of standard errors (Little et al., 2012). These methods become even more problematic in mechanism-based trials where when there are hypothesized mediators. When there are mediators, it is necessary to consider how treatment noncompleters would score not only on the outcomes but on the mediators, too. This all needs to be considered without introducing bias into estimates of the structural relationships between these variables.

The upshot is that as soon as there is posttest missing data of any kind but especially for treatment noncompleters, ITT analyses require assumptions that can undermine inferences. Moreover, missing data at posttest and/or followup may be due to treatment noncompletion, but may be due to other factors as well, such as a treatment completer missing the followup evaluation because of illness. The appropriate methods for handling missing data vary according to the extent of the missing data and the circumstances under which the missingness arises. Jaccard (2024) discusses the various methods and circumstances for interested readers. This includes further details about when the circumstance involves mediators (and moderators) and if one seeks to answer efficacy-based questions. Overall, though, muddledness can be avoided when the focus is efficacy-based by using modern per protocol analyses and by restricting conclusions to efficacy concerns rather than making effectiveness claims.

Critiques of per protocol analysis as violating random assignment can overstate the problem

A frequent critique of per protocol analyses is that one cannot be assured that treatment and control arms are equated on variables that matter via random assignment because random assignment could be compromised. This critique is valid but it ignores the fact that (a) sometimes the bias is irrelevant to the question at hand, and (b) often there are statistical corrections that can be made for bias. As noted, ITT analysis shifts the research question from efficacy to effectiveness. The problem of compromised random assignment is not solved by switching to ITT analysis and thereby ignoring or changing the question of interest. Again, treatment noncompletion does not undermine random assignment unless there is non-trivial bias among those patients who do not complete treatment versus those who remain in treatment and in ways that compromise between-arm comparisons.

The extent to which the per protocol sample compromises random assignment can be empirically determined for variables measured at baseline. Sound trial design involves anticipating possible sources of consequential imbalance due to dropout and then measuring those sources so that they can be explicitly addressed using modern methods of per protocol analysis. It is of course always possible that non-trivial bias exists on a variable that was not measured, which might undermine a study’s inferences. However, such variables can sometimes be addressed by using instrumental variables (Angrist & Pischke, 2009). In many cases, imbalance will only create bias when the amount of bias is substantial (Strube, 1991).

Summary and Implications for Future Directions

To realize the prediction made by the first author’s Presidential Address in 2006 that the best is yet to come, we exhort researchers to carefully consider precisely what question they seek to answer and then select the appropriate research design and analytic approach relative to that question. We hope we established in this Future Directions piece that questions about the combined impact of adherence, implementation, and effectiveness are appropriately answered through the application of effectiveness trial designs and ITT analysis. Questions about efficacy and mechanisms of efficacy are appropriately answered through the application of per protocol analyses. This is true for clinical outcomes per se, as well as for efficacy trials of interventions designed to increase adherence and implementation fidelity. Per protocol analyses are not perfect, but every analytic approach has its issues. Our main point in this Future Directions article is that the issues surrounding per protocol design and analysis are workable every bit as much as issues surrounding ITT design and analyses. The demand to conduct ITT analyses no matter what the focus of a trial, though having admirable intentions, is not the right approach and can result in ITT misapplication. For a focus on efficacy and on efficacy-based mechanisms, modern per protocol analysis is preferable (except under rare circumstances when adherence is high and treatment is implemented faithfully).

Our emphasis in this article has been on the importance of distinguishing efficacy and effectiveness, and per protocol and ITT analysis, and we believe each can continue to advance the field. As we noted, the future of clinical trials research in child and adolescent mental health will likely be characterized by increasing use of hybrid trials. Such trials combine elements of efficacy and effectiveness, as well as implementation research. Hybrid designs hold promise for maintaining scientific rigor and internal validity while also accelerating the translation of interventions to routine care settings. In such trials, application of both ITT and per protocol analyses usually will be appropriate to address different research questions. For example, this would entail first applying modern per protocol analysis to answer questions related to efficacy and efficacy-based mechanisms (does this treatment work and work through the theorized mechanisms when adherence is high?). Per protocol analysis would be followed by ITT analysis to determine whether findings are consistent with the per protocol results when adherence is not rigorously controlled (i.e., moving toward effectiveness). If treatment noncompletion is low, the findings likely will be consistent across per protocol and ITT analyses. If, however, the effect sizes obtained with ITT are smaller than with per protocol analysis, then this would suggest that there are issues relating to adherence and/or implementation that weaken the per protocol effects of treatment and that efforts should be directed toward addressing these issues to enhance outcomes in routine care. Whichever scenario unfolds, by asking the right questions and selecting the appropriate analytic approach to answer those questions, clinical trials researchers can fulfill admirable intentions and ensure that the “best has come.” We look forward to and are excited about these exciting future directions.

Disclosure Statement

No potential conflict of interest was reported by the authors.