Background

Pragmatic clinical trials are “primarily designed to determine the effects of an intervention under the usual conditions in which it will be applied”; they contrast with explanatory trials which “are primarily designed to determine the effects of an intervention under ideal circumstances” [1]. The United States (U.S.) National Institutes of Health (NIH) Healthcare Systems Research Collaboratory (the NIH Collaboratory) was established to advance large-scale pragmatic clinical trials through the conduct of pragmatic trial demonstration projects. These demonstration projects are being performed in large and diverse healthcare settings around the United States and allow exploration of best practices for appropriately designing pragmatic trials in addition to generating trial findings [2]. Each trial had a design phase for 1 year and a subsequent implementation phase if approved after a preliminary report. The NIH Collaboratory comprises the research teams that design and perform the individual trials; a coordinating center with expertise in design, biostatistics, bioethics, and electronic health data; and NIH scientists.

The purpose of this analysis was to measure the degree to which the NIH Collaboratory trials are pragmatic at both the design and implementation phases using a version of the Pragmatic Explanatory Continuum Indicator Summary framework (PRECIS-2) [3]. In addition, because all NIH Collaboratory trials begin with a yearlong planning phase to pilot test the intervention and evaluate feasibility of aspects such as outcome ascertainment methods and integration with workflow, we were able to study whether and how trial design changed from conceptualization to implementation. As a secondary goal, we sought to assess the usability of PRECIS-2 as a tool for assessing pragmatic features across studies and over time.

Methods

Ethics

The individual NIH Collaboratory Trials were approved by the relevant research ethics boards. This analysis did not require informed consent from raters nor ethical approval because the data sources were limited to the study protocols, not information about human subjects. In addition, all raters were co-authors of the paper rather than subjects of the research.

Setting

The U.S. NIH is the largest medical research agency in the world. Through funding from the NIH, the NIH Collaboratory seeks “to strengthen the national capacity to implement cost-effective large-scale research studies that engage health care delivery organizations as research partners. The aim of the program is to provide a framework of implementation methods and best practices that will enable the participation of many health care systems in clinical research” [4]. The NIH Collaboratory funded five pragmatic clinical trials at both a planning and implementation phase in 2012 and 2013, respectively. These trials are described in Table 1 and include 1) Active Bathing to Eliminate (ABATE) Infection, 2) Lumbar Image Reporting with Epidemiology (LIRE), 3) Collaborative Care for Chronic Pain in Primary Care (PPACT), 4) Strategies and Opportunities to Stop Colorectal Cancer (STOP CRC), and 5) Time to Reduce Mortality in End-Stage Renal Disease (TiME). Although additional trials have been funded through the Collaboratory, they are not included in this analyses as they had not been awarded funding for the implementation phase until completion of the analyses.

Table 1 Summary of included projects
Full size table

Participants/Raters

Raters were trial principal investigators (PIs) or other investigators from their team (n = 4), Coordinating Center staff (1), or NIH staff (6). Six of the raters had familiarity with all five trials either because they had participated in funding decisions or regular cross-project meetings. Two raters (both NIH staff) had limited knowledge of any of the projects prior to participating in the PRECIS 2 exercise. Raters were recruited based on their interest and availability. The six NIH staff rated all five trials. The PIs or other investigators, as well as the Coordinating Center staff, each rated only two trials, one of which was his/her own.

Rating procedures

To measure the pragmatic nature of the NIH Collaboratory trials, we used the PRECIS-2 toolkit (https://crs.dundee.ac.uk/precis). The CONSORT workgroup on Pragmatic Trials created the PRECIS criteria to help trialists design trials that are pragmatic across multiple domains [1, 5]. While not primarily intended to analyze trials post hoc, the original PRECIS scale was successfully used for this purpose [5]. Based on findings from the initial use of the tool, a team at the University of Dundee developed the second version [6], which reduces the number of domains rated from 10 to nine, makes comparisons to usual care without explicit rating of the control conditions, and considers external validity in the recruitment and setting domains.

The PRECIS-2 toolkit includes nine domains: (1) eligibility - who is selected to participate in the trial; (2) recruitment - how participants are recruited into the trial; (3) setting - where the trial is being done; (4) organization - what expertise and resources are needed to deliver the intervention; (5) delivery flexibility - how the intervention is delivered; (6) adherence flexibility - what measures are in place to make sure participants adhere to the intervention; (7) follow-up - how closely participants are followed-up; (8) primary outcome - how relevant is it to participants; and (9) primary analysis - to what extent all data all included. Each domain is scored on a five-point Likert scale from very explanatory (1) to very pragmatic (5), with a score of 3 indicating that a trial is equally pragmatic and explanatory. In addition, an overall composite score is reported for a given trial to characterize the overall pragmatic nature of the trial. We also calculated an overall mean score for each domain across trials, to illustrate for which domains the trials, in general, were more or less pragmatic.

Training

All raters received training in applying PRECIS-2. The training consisted of an orientation webinar by one author (RG) based on the PRECIS-2 toolkit, practice with a published protocol, and a second web conference to calibrate ratings by discussing ratings that differed among the individuals participating in the training. Following the training, each of the five demonstration projects was rated by eight raters and evaluated at two time points, using the initial grant application and a required progress report written at the end of a 1-year planning period that included changes to the protocol or implementation approach. Four trials included a rating by a trial team member. Each rater entered their two sets of ratings on a form, which included space for comments.

Analysis

We produced one PRECIS wheel for each time point (pilot/planning phase and implementation phase) for each of the five trials using the PRECIS-2 Toolkit, which calculates the median scores for ratings for a given trial. The data were analyzed using STATA 12.0 (StataCorp, College Station, TX, US) and SAS 9.3. We calculated the mean and median scores and the range of scores for each domain, trial, and time point. We also calculated the differences in scores between the two time points for each of the PRECIS-2 domains for each trial. To evaluate change over time, we examined the spread in these differences and determined the level of statistical significance (for a given domain for a given trial) using the sign test in Stata 12.0, a Wilcoxon nonparametric test of equality of matched pairs. Interrater agreement or reliability was calculated for each trial using Gwet’s AC1 statistic [7], a more robust version [8] of Fleiss’ Kappa [9]. Interrater agreement was also measured using the intraclass correlation coefficient. Additionally, we obtained each trial principal investigator’s impressions on the degree of congruence of the ratings with on-the-ground experience.

To inform our secondary goal of assessing PRECIS-2 usability, we reviewed the comments provided by the raters for each trial. Two authors (KJ and GN) organized the comments by domain and indicated study design aspects that raters considered in their scoring that were not specified in the PRECIS-2 toolkit, as well as rating challenges. All raters reviewed these results.

Results

Comparison of domains across trials

All five demonstration projects were rated to be more pragmatic than explanatory. The overall composite scores, calculated on the basis of the average score of the means and medians across the domains, are all greater than 3 on a scale where 3 signifies equally pragmatic/explanatory. Mean and median scores for each trial at the implementation phase for each domain are presented in Table 2. Whereas all five trials were more pragmatic than explanatory (that is, overall composite median rating > 3.0), TiME and LIRE were found to be the most pragmatic (4.5 overall composite median rating for both). The domains for which those two trials were most pragmatic (that is, mean of eight raters > 4.5) were recruitment (mean of eight raters = 4.8) and follow-up (4.9) for LIRE, and eligibility (4.8), recruitment (4.6), follow-up (4.6), and primary outcome (4.9) for TiME. The domains for which they were less pragmatic but were still more pragmatic than explanatory (that is, mean of eight raters < 4.5 but > 3.0), were organization (3.8), setting (4.0), primary outcome (3.6), and eligibility (4.3) for LIRE and delivery (4.1), setting (4.4), organization (4.4), and adherence (4.4) for TiME. The trials that were less pragmatic but still more pragmatic than explanatory were ABATE and PPACT (overall composite median rating = 3.5 and 3.6, respectively.) The domains that were rated as more explanatory than pragmatic (that is, mean of eight raters < 3.0) were organization (2.6) and delivery (2.1) for ABATE and organization (2.9) and adherence (2.8) for PPACT. However, those two trials were also found to be very pragmatic (that is, mean of eight raters > 4) on several dimensions including recruitment (4.5) and analysis (4.5) for ABATE and primary outcome (4.6) and analysis (4.9) for PPACT. STOP CRC fell in between, with an overall composite median rating of 3.9. In general, the domains along which trials were, on average, most pragmatic included primary analysis (mean = 4.7 range = 4.5 to 4.9)), recruitment (4.3 (3.6 to 4.8)), eligibility (4.1 (3.4 to 4.8)), setting (4.1 (4.0 to 4.4)), follow-up (4.1 (3.4 to 4.9)), and primary outcome (4.1 (3.5 to 4.9)). On average, the less pragmatic, although still more pragmatic than explanatory, the domains were organization (3.3 (2.6 to 4.4)), flexibility of delivery (3.5 (2.1 to 4.5)), and flexibility of adherence (3.8 (2.8 to 4.5)).

Table 2 PRECIS scores by trial (at implementation phase) and domain
Full size table

Reliability

At the implementation phase, we found modest but statistically significant interrater agreement for four of the five trials (AC1 statistic ranged from 0.23 to 0.40, p values ≤ 0.001; Table 2). Intraclass correlation coefficients ranged from 0.05 to 0.31. Ratings for the planning phase were similar: AC1 statistics ranged from 0.11 to 0.44 and all p values were ≤ 0.001 (data not shown). We examined outliers to see if differences existed among raters who were and were not previously familiar with the studies or for raters who were rating their own project. We did not observe any notable patterns. However, the principal investigators’ on-the-ground experience did not always match the rater-assessed examples of change, as discussed further below.

Change over time

Trial refinements over the course of the planning phase represented responses to logistical issues, stakeholder preferences, and input from the NIH Collaboratory members as the studies approached full implementation. Some examples are shown in Table 1. Figure 1 shows the PRECIS wheels for each of the five trials at the two time points. We found between one and three rater-assessed significant changes over time for each study; however, we did not note any consistency in terms of direction or domain. Additionally, the PIs agreed with the direction of the ratings for only one study. As an example, raters assessed that recruitment procedures became more pragmatic for one study, but the PI indicated that the story was more complex, with some aspects becoming more pragmatic but others less so. Study procedures were refined such that organizational approaches to patient prioritization did indeed more closely mirror everyday clinical care; however, the timing of patients receiving the intervention became more tied to study-specific provider randomization points.

Fig. 1
figure 1

PRECIS wheels as assessed by raters for each of the five trials at two time points. Ratings on a 1 – 5 scale indicate more explanatory to more pragmatic ratings. The dashed line indicates the planning phase. The solid line indicates the implementation phase

Full size image

Rater comments

Table 3 summarizes factors that raters noted they had considered in rating the study. These were largely consistent with the examples in the PRECIS-2 toolkit that were included in the training materials. However, the raters’ comments about additional considerations that they factored into assigning domain ratings highlight that the PRECIS-2 ratings are not necessarily conclusive but generate a starting point for discussion, as we describe in more detail below.

Table 3 Examples of rater explanations for ratings of individual trials
Full size table

Raters noted several challenges in applying the criteria in their commentary. For the eligibility domain, raters had to consider both eligibility of the facility and eligibility of the patient. For example, for most of the trials, participating facilities or clinics were selected by convenience and no information was available about inclusion criteria even though this was the unit of randomization. Facility willingness to participate was viewed by raters as an aspect of eligibility that was more explanatory than pragmatic.

The setting and organization domains proved particularly difficult to rate relative to usual care. Raters commented that there are many different aspects to consider when rating settings within the diverse U.S. health system, including geography, types of care provided, and financing (for example, fee for service versus managed care). Furthermore, institutions where these trials were occurring all had the resources and infrastructure to support a systems-change intervention, making it possible that these institutions had relatively high organizational resources to support complex quality improvement. Even if this infrastructure was not research-specific, this potential difference from usual care led to more explanatory ratings of the setting and organization domains.

For the flexibility of delivery domain, the determination of whether an intervention was relatively more restrictive than a strict quality control protocol in usual care was challenging. Similarly, it was challenging to rate the flexibility of adherence for an intervention relative to usual care because, if the intervention was successful, the adherence procedures could become usual care. Second, few studies documented extensively efforts undertaken to maintain “adherence” at the organizational level. For example, when leadership changes occur, the need arises for substantial discussion and planning to continually “engage” the stakeholders/leadership in the health system. Most of these activities are not planned but are undertaken ad hoc when health systems lose their leadership. To what extent these efforts to re-engage leadership in the conduct of the trial represented less pragmatic adherence is unclear.

Primary outcome is rated according to the extent to which it is relevant to participants, but raters struggled with how to rate outcomes that might be more important to health systems than to patients, for example, process efficiency. Raters also had to determine how to factor in criteria that pertained to multiple domains, and whether they should “be counted” more than once. For example, consent by patients or organizational willingness to participate pertained to multiple domains: eligibility, recruitment, setting. Thus, some raters attempted to provide an average score across unrelated subdomains; other raters may have only considered a single subdomain.

Discussion

The objective of this study was to analyze five pragmatic trials in order to characterize pragmatic versus explanatory design by PRECIS-2 domains and how design details changed over the course of a yearlong study-planning period. In five trials designed as pragmatic trials in diverse U.S. healthcare settings, we observed that trials were designed as more pragmatic than explanatory as measured by all PRECIS-2 domains.

Raters struggled to use the PRECIS system for this analysis, as illustrated by the comments in Table 3, PI discussions, and the ICC range from 0.05 to 0.31. For comparison, in a study by Glasgow et al., in which they studied three effectiveness trials of weight loss in obese patients with comorbid conditions [5], the ICC for individual items was 0.72, and the overall kappa interrater reliability on the composite PRECIS score was r = 0.88. The large difference in interrater reliability is surprising given that we used a similar training approach and had access to detailed study information.

Whereas the rating challenges limit our ability to draw conclusions about specific studies, some general observations emerge. Across studies and time points, the domains rated as most pragmatic were analysis and recruitment, whereas those that were closer to explanatory (average range 3 to 4) were organization, delivery, and adherence. This could reflect, in part, that it may be easier to be pragmatic for some domains than for others. For example, it is relatively easy to be pragmatic for patient eligibility by taking all comers; but it is often difficult to be pragmatic when trying to deliver an intervention.

It is important to note that explanatory elements of pragmatic design do not necessarily relate to study quality. Some trial aspects may need to be designed in a more explanatory manner in order to answer the study question. PRECIS-2 ratings provide guidance to researchers on the appropriate corresponding study procedures. For example, the more explanatory rating of the organization domain (how the resources, provider expertise and the organization of care delivery in the intervention arm of the trial compare to usual care) indicates that the study involves extra resources such as training. By noting this during the design phase, study teams can make sure they communicate with involved health systems about time and resource requirements. However, it is important to note that requiring additional training does not necessarily make the organizational domain more explanatory: the same approaches to training personnel to roll out a trial intervention could be the same as those approaches used by the organization to roll out a clinical or quality improvement intervention.

It is difficult to compare our overall findings to other reports that rated studies using PRECIS domains because they utilized an earlier version of the tool and had different study questions. However, one other study has used PRECIS criteria to examine change over time. Elder and Munk [10] used a modified PRECIS wheel to obtain input on study methodology while planning a new phase of research examining two complementary therapies for chronic low back pain. The study led to re-evaluation of the design of certain aspects, for example participant characteristics, that were rated as more explanatory than expected and could be made less restrictive. However, as in our experience, the authors concluded that having a more explanatory characteristic within a pragmatic trial may be appropriate depending on the research question.

This study generated insights that may be useful for future use or refinement of the PRECIS-2 tool. As per Table 3, raters struggled with how to apply ratings. In particular, comparing the intervention to usual care requires guidance about 1) how to handle domains such as recruitment that can pertain to health care settings or patients; 2) how to rate a systems-change intervention (which could become, but is not currently, standard care) and 3) what level of existing supports/standards, for example communication with leadership or use of electronic health record functionality, is considered typical in usual care. After this study was completed, the PRECIS-2 designers published a manuscript that fully described the PRECIS-2 tool [3]. Their explanation of the domains in detail highlights that the complexities we encountered resonated with theirs, including how to rate an intervention that is designed to change usual care and how local care nuances (for example, data systems) can influence ratings. The challenge of using the tool, especially for some criteria, suggests that the PRECIS-2 criteria may need to be further refined in order to have sufficient specificity to enable comparison of intervention to usual care in the context of a broad range of settings. The issue is not just tool development, but also clarification of how we characterize care and what components are “usual”. Having this more detailed understanding of usual care and guidance regarding how to characterize care would enable a more clear understanding of the degree to which the intervention differs from usual care and has practical utility to a health system, given the diversity of health systems that exist.

In addition, it would be helpful to have guidance about the amount of study information that teams need in order to best use the PRECIS-2 tool. If we had this guidance, the benefit would be that all trials could more easily capture information. In turn, this would enable better comparisons across trials and allow for analysis of a broader trial portfolio. Additionally, looking across trials and across time points would be useful, but we were limited by what information we had. Therefore, providing guidance on what information is needed to best apply this tool would better enable its utility.

The materials that we used to rate the study were the grant application and progress report, which contained many details that pertained to the PRECIS-2 domains but were not organized according to PRECIS-2 domains. The information in these documents may have contributed to low interrater reliability; however, they contained a substantial amount of implementation-oriented information. It is possible that ratings would have been even more difficult using the details typically available in a protocol or manuscript of study findings. It also is possible that the limited number of raters for each study could have contributed to the low interrater reliability. However, in both the Loudon et al. paper [3] and our experience, ratings ultimately benefited from local familiarity with the health system where the trial was being conducted and direct input from study team members helped prompt discussion and clarification about study details. As such, we do not necessarily see value in review by an external group, except for providing advice about how to best use the PRECIS-2 tool. The ability of the PRECIS-2 framework to support discussions about how to interpret and operationalize design decisions is helpful.

Conclusions

The raters participating in the process found it an informative way to learn about pragmatic design in general as well as about specific studies. This study demonstrates that PRECIS-2 can be used to rate protocols, as well as for study planning, and helps address the need for systematic approaches to reporting pragmatic studies [11]. However, results of analyses using the criteria post hoc should factor in the challenges encountered in our analysis. In addition, refinements would be helpful for raters. These could include creating additional rating criteria, linked to numerical rating, and exploring different formats for brief training of raters. Our results highlight that researchers should anticipate making changes to study protocols for pragmatic trials based on health system realities.