We used purposive sampling where individuals were invited based on a specific criterion, namely their role or alleged specific expertise associated with the stakeholder group. This is a sampling technique intended to identify individuals that are assumed to be representatives of their role and that can provide detailed information. However, the concern with this criterion approach is that the expertise is assumed and that based on a role description, one may fail to identify those who are most knowledgeable or those with the greatest level of expertise [21]. We tried to alleviate this concern by first inviting individuals that the first and last author knew to be knowledgeable through their participation in conferences, seminars, and workshops over the last five years. Where this regarded stakeholders that the authors have worked with, or are currently working with, we denote this with an asterisk (“*”). After each interview, we asked the interviewee for recommendations of other stakeholders with specific expertise, an approach called snowballing. Finally, we also invited individuals based on internet-searches (cold calling). In total, we invited 49 individuals to participate.

Participants first received an invitational email with a link to the information letter, informed consent form, and protocol (see S1 File). We sent one reminder a week after. When participants agreed to participate, we scheduled an appointment for an online interview through Microsoft Teams. Participants received a link to the dashboard and a tutorial explaining the dashboard (see here and S2 File). They signed the informed consent form prior to the interview.

We conducted the interviews online between April and June 2021. Interviews lasted between 30 and 50 minutes. One team member led the interview (TH, female (Dr), MH, male (PhD candidate) or DS, male (Prof), who are all trained in qualitative research methods and familiar with the responsible research literature), whilst another team member observed and made notes on the interview’s process and its content. After a brief introduction of the interviewer and the interviewee, interviews were conducted using a topic guide (semi-structured) that was based on a literature review and various internal discussions (see S3 File). The topic guide was pilot tested (n = 3) for comprehensibility using cognitive interviewing [22]. After each interview, the team got together for reflection (peer-debriefing). If they wanted, interviewees received a brief written summary of the interview (member check) with the option to comment or correct. The interviews were transcribed by a transcription company under a data processing agreement. We used the COREQ checklist to write up our findings [23]. For a more elaborate description of our procedure, see study protocol (https://osf.io/z7bsg/).

The aim of our study is exploratory and descriptive. To get a broad understanding of stakeholders’ views, we applied content analysis that involved deductive and inductive elements [24]. We structured stakeholders’ views by looking for Strengths, Weaknesses, Opportunities, and Threats (SWOTs), see Table 1. Subcodes within each part of the SWOT were derived from the data. These subcodes pertain to the general approach of an institutional dashboard and responsible metrics, i.e., suggestions for improvement by interviewees that were specific to the proof-of-principle version are not featured in the results where they have been implemented successfully (e.g., the dashboard interviewees saw displayed a sample of publications, the final dashboard that will be published shortly includes all publications).

Our analysis consisted of three phases. First, two team members (TH and MH) read and coded 5 interviews independently using MAXQDA 2020 (Release 20.3.0, VERBI GmbH, Germany). They then each proposed a code tree, exchanged these, and resolved discrepancies (mostly whether something was best classified as a strength or an opportunity, a weakness or a threat) through discussion. With this revised code tree, the two team members coded another 5 interviews. This process was repeated until saturation was established [25], meaning that when coding additional interviews, no new codes were identified. The code tree that resulted was then presented to the full research team (together with the interviews). We analyzed all interviews. Yet, reviewing the final three interviews resulted in minor modifications to the wording of the code tree only (SI 4), which formed the basis of our results.

Overall, interviewees were pleased to see that the dashboard focused on specific measures that relate to concrete behaviors. Many interviewees thought the dashboard was helpful in showing institutions where they stand when it comes to responsible research practices. The dashboard approach allowed for creating a baseline, which was considered essential before talking about change or improvement on the metrics. Some expressed curiosity about the extent to which the dashboard would match internal data on self-evaluation processes.

Various interviewees indicated this was the first time they saw such a dashboard. They thought it fulfilled an important need and some interviewees felt these measures could serve as an alternative to publication output, journal impact factors, third-party funding, or other widespread metrics of institutional performance. Many interviewees considered the included topics timely, as they related to ongoing discussions in the field regarding responsible research and contemporary science policy.

Many interviewees indicated they appreciated the transparent overview. They liked the interactive features of the dashboard (e.g., switching between percentages and absolute numbers, hover-over fields with limitations) and appreciated the fact that it gave them a good overview whilst not overburdening the viewer with information.

Several participants indicated that they missed an overall narrative that would justify the choice of metrics included. In addition, it was unclear to some interviewees how the metrics related to each other, or to the quality of research. By presenting these metrics, it seemed as if these metrics were the standard metrics for institutional assessment, and various participants indicated that they did not consider that accurate as the metrics were not backed by an overall legal or regulatory framework. This made participants question who decides what good metrics for responsible research are. Without such a consented framework, participants feared that the dashboard would not be accepted as a viable complementary or alternative approach to evaluate research institutions.

Several interviewees expressed difficulty in interpreting the methods by which the metrics were compiled and the limitations inherent to these methods. This included the fact that the metrics may change over time, e.g., the numbers for Green Open Access might change once more publications are made available. Various participants stressed that the way particular metrics were operationalized was unclear. Finally, participants found it difficult to understand the metrics’ denominators, indicating it was unclear whether all metrics pertained to the same dataset or to different (parts of the) dataset(s), and whether the dashboard used all relevant available data sources (e.g., all eligible study registries), to provide a less biased measurement.

Some participants were concerned that metrics on robustness of animal research or Open Science stemming from publications in 2018 would present an outdated picture. They pointed to various developments in those areas in recent years and interviewees expected that publications from 2019 and 2021 would show higher robustness and Open Science scores. Although the fact that the data were taken from 2018 was a feature of the proof-of-principle dashboard, interviewees’ concerns underscore the need to continuously integrate new information into the dashboard, ideally in an automated manner.

The dashboard was considered a good tool to start an evidence-based conversation on how to improve responsible research practices and about a more holistic evaluation of research institutions, i.e., assessing different institutional facets and not merely citations or impact factors. Some interviewees remarked that the dashboard could provide more information about how one could induce change, and aid in the development of concrete interventions. The dashboard would then, in their view, need to be presented to organizational leadership on a regular basis to fuel a discussion about efforts to improve. In addition, participants thought the dashboard could promote collective improvement efforts (or ‘healthy competition’) that would enable institutions to learn from each other.

Several interviewees stressed that the dashboard should include the possibility to benchmark oneself over time. Whereas this was possible for some indicators, many interviewees believed that having this data over time for all metrics could help institutions to evaluate whether interventions to improve their performance on particular metrics were successful.

When thinking about how to prevent possible detrimental consequences of benchmarking, many interviewees stressed that it would be most responsible if the dashboard would be used for internal purposes only. The dashboard’s information would not be accessible to anyone outside the institution, except if the institutional leadership chose to publicize some information. Some interviewees suggested that the dashboard should go public with a 2- or 3-year delay, which would give institutions the opportunity to improve first.

Another opportunity that various interviewees raised was tailoring the dashboard to the institutional strategy. This would mean that different institutions would have different dashboards with different metrics, allowing institutions to create focus areas for a specific timeframe (e.g., in 3 years’ time, at least 60% of our publications should have open data). Related to this, some interviewees remarked that it would be helpful to link to more fine-grained levels of information, e.g., on a particular center within the research institution or a research field level. They remarked that they would appreciate the opportunity to correct or at least comment on the numbers (e.g., providing a reason why an institution performs badly on trial reporting), or to withdraw from the dashboard.

Various interviewees discussed additional metrics they thought might be included in the dashboard. Two of the most often mentioned suggestions were the percentage of publications that were first posted as a preprint and conflict of interest statements in publications. Interviewees involved in clinical trials thought a metric on the overall number of clinical trials conducted could provide useful context. They also suggested metrics related to patient engagement in clinical research. In relation to animal research, some interviewees thought a metric for preregistration of animal studies could be useful, while others mentioned the choice of the model system as an important metric for the generalizability and translation of research. More broadly speaking, various interviewees mentioned metrics related to diversity, such as the percentage of female researchers or female principal investigators at a particular institution, or metrics related to the societal value of research, such as uptake of research by the media, policy documents, clinical guidelines, or patents.

Different interviewees remarked that they missed statistics on metrics’ performance. They would like to see uncertainty indicators such as confidence intervals, so that they could assess when an institute would perform better or worse than all institutions together. They also asked for numbers on the sensitivity and specificity of the sampling and classification approaches.

Many interviewees feared that making the dashboard public in its current form would put institutions in a bad light. Whereas they thought some form of benchmarking could be helpful, they worried that the dashboard could be used as a tool to name and shame institutions, which could in turn have detrimental consequences for funding or for an institution’s reputation among the public. Some interviewees also mentioned that institutions should not be blamed for not meeting a metric’s requirement in the absence of sufficient infrastructure in place (e.g., a data sharing platform) that would allow institutions to comply. A few interviewees also pointed out that there is no inherent need for a comparison, since it is already self-evident that institutions should do their best to score high on these metrics.

Various interviewees were concerned that the metrics would be misinterpreted, whether intentionally or unintentionally. Especially under precious time, stakeholders might not appreciate the metrics’ limitations and instead take them at face value. This might result in inflating minor differences and risk the resource-heavy implementation of policies based on little understanding of the metrics. Additionally, some interviewees wondered whether other German UMCs were the right comparator, as there are large differences between the institutions.

Finally, various interviewees thought that the dashboard and the indicators therein could be gamed. Researchers at the respective institutions might feel that they must score high on these metrics and might get creative in the ways to score high, without living up to the ideals included in the respective metrics. More generally, some interviewees felt that boiling down these important issues to numeric indicators was not without risks.

There are a variety of frameworks available that are all applicable to the topic of responsible research practices. Broad examples include the TOP guidelines [27] that include, beyond data and code transparency, study preregistration and openness about materials, and focus on what journals incentivize. Another broad example, this time focused on individual researchers, are the Hong Kong Principles [7], where explicit attention is paid to peer review and, more recently, diversity [28]. More specific frameworks include ARRIVE for animal research [29], or the Cochrane Risk of Bias tool for clinical trials [30].

Some of our metrics find their origin in regulatory frameworks, such as the mandated registration of drug trials in EudraCT [31]. In addition, the World Health Organization has specified that results of clinical trials should be published in a relevant journal within two years of trial completion [32, 33]. The Declaration of Helsinki [34] calls for all research involving human subjects to be prospectively registered and published.

That said, no framework for responsible research seems broad enough or has received universal or consented support. It is questionable whether such a framework will be developed, and it raises the question of who should oversee its development. A potential starting point would be the UNESCO recommendations of Open Science [35]. An alternative approach, on a more local level, could be to involve a group of UMCs in a consensus-building procedure to find a set of indicators for responsible research that these particular UMCs agree upon (see [36]).

We witnessed a surge in dashboards for responsible research and related topics in recent years. Some have a broad focus, such as the European Commission Open Science Monitor, project TARA within the context of DORA or the French Open Science Monitor that display broad disciplinary fields [37–39]. Others focused on (and ranked) individual researchers in terms of transparency, such as the Curate Science Transparency leaderboard [40]. Some of our interviewees questioned whether an entire UMC is the right level for developing such a dashboard. It is interesting in this context that the DFG’s revision of the Guidelines for Safeguarding Good Research Practice [41] includes explicit chapters on responsibilities for heads of research organizations. We believe that a dashboard approach could support heads of research organizations to empirically monitor the implementation of the revised code of conduct.