2.2.1 Delphi panel recruitment and participant characteristics.

Following recommended protocols for Delphi studies, we will purposively recruit domain experts with demonstrated expertise in entrepreneurship, innovation, and resilience. Eligibility required active engagement in at least one of the following areas: academic research on entrepreneurship or resilience, entrepreneurial practice, ecosystem leadership (e.g., incubators, accelerators, investor networks), or policymaking related to SMEs. Experts will be identified through publications on entrepreneurial resilience, professional and policy networks, and targeted outreach to ecosystem organisations. Recruitment will aim to ensure geographic diversity across OECD regions (e.g., USA, Germany, and Australia).

The ideal Delphi sample size ranges from 15 to 23 participants [21]; however, for a multistakeholder analysis such as this, a larger panel improves replicability and diversity of perspectives [22]. We therefore aim for a panel of approximately 70–85 respondents, representing non-OECD contexts, gender diversity, institutional roles, and social dimensions highlighted in prior literature. We will aim for diversity in gender, geography, and disciplinary background, including experts from business, economics, public health, and related fields. Given typical attrition rates of 15%–40% in Delphi studies [21,23], we will recruit approximately 100–120 experts in Round 1, with the aim of retaining around 70–85 participants by Round 4.

2.2.2 Survey development and questions.

Survey questions were developed in alignment with previous Delphi frameworks [10] and qualitative research guidelines [24,25]. Questionnaires combine closed-ended Likert-type items with open-ended questions to elicit both structured ratings and richer qualitative justification.

• Round 1 will focus on conceptual foundations and initial item generation. Participants will first provide background information and an open-ended definition of entrepreneurial resilience, specifying what should be considered inside and outside the construct. They will then list up to five key drivers and five key drainers of resilience based on their experience, before reviewing a literature-informed category list and indicating any additional factors they regard as important.
• Building on synthesised Round 1 responses, Round 2 will ask participants to rate the shortlisted drivers and drainers on (i) their importance for entrepreneurial resilience and (ii) their actionability through policy, training, or support programs, using Likert-type scales. For each factor, respondents will also indicate whether it is context-sensitive (e.g., COVID-19, gender, youth, digital transformation, conflict, senior entrepreneurs), enabling us to tag items by setting and population. Free-text fields will allow participants to suggest additional drivers, drainers, or contextual notes.
• Round 3 will shift to measurement and outcomes. For each priority factor, participants will select preferred measurement approaches (e.g., validated psychometric scales, longitudinal surveys, theoretical or econometric modelling, mixed-methods designs) and may recommend additional methods. They will then choose and rank outcome metrics for entrepreneurial resilience (such as business survival/continuity, growth, adaptability, time-to-recovery, access to liquidity, digital capability indices, and workforce resilience), identifying those most relevant for policy and program design.
• In Round 4, participants will review a consolidated list of drivers and drainers derived from Rounds 1–3 and will indicate their level of agreement with including each in the final framework. They will also rate the importance and feasibility of key policy and program levers (e.g., startup funding, relief and liquidity access, digital infrastructure, mentoring and coaching, gender-inclusive policies) and will rank their top three priorities. A final open-ended question will invite additional recommendations on measurement tools, outcome metrics, and resilience-building strategies.

All Delphi rounds will be administered online using SurveyEngine, with each questionnaire designed to take approximately 15–20 minutes and to permit asynchronous completion. Full survey instruments for Rounds 1–4 are provided in Appendix B.

2.2.3 Data analysis.

Consensus will be assessed using a combination of central tendency and dispersion criteria, alongside Kendall’s coefficient of concordance (W) [26]. An item will be considered to have reached consensus if it achieves a median rating of ≥6 on the 7-point Likert scale and an IQR ≤1. Stability across successive rounds will be evaluated by examining changes in medians, IQRs, and Kendall’s W; items showing minimal change between rounds will be treated as stable, while unstable items will be re-presented or refined in subsequent rounds. Subgroup comparisons (e.g., by region or professional role) will be undertaken where relevant. Open-ended responses will be coded using an a priori codebook derived from prior literature and the study’s conceptual framework, with allowance for emergent themes. Two coders will independently code responses; inter-rater reliability will be assessed using Cohen’s [27]. Discrepancies will be resolved through discussion and adjudication. Iterative synthesis across rounds will track movement toward consensus.

2.3.1 Sampling and recruitment.

We will use purposive sampling [29] to recruit entrepreneurs and ecosystem actors whose experiences reflect the contextual lenses identified in prior literature and the Delphi rounds, including gender, age (youth versus senior entrepreneurs), digital intensity, and exposure to major shocks such as COVID-19 or conflict-related disruptions. Participants will be identified through business networks, incubators, professional associations, and existing project contacts, with limited snowball sampling to reach under-represented groups. We anticipate conducting approximately 15–25 in-depth interviews and, where appropriate, 4–6 small focus groups with 5–8 participants each. Qualitative formats will be selected based on feasibility, participant availability, and the suitability of the topic for individual versus group discussion. Stratification (e.g., by gender, age, digital intensity, or exposure to major shocks) will be used to capture variation in how drivers and drainers of resilience are experienced across contexts.

2.3.2 Data collection procedures.

In-depth interviews and focus groups will be conducted virtually (e.g., via Zoom or Microsoft Teams) to enable participation across regions and accommodate entrepreneurial schedules. Interviews are expected to last approximately 20–40 minutes and focus groups 50–70 minutes. Sessions will be facilitated by experienced qualitative researchers. With participants’ consent, discussions will be audio-recorded and transcribed verbatim. A semi-structured topic guide, informed by prior literature and Delphi findings, will ensure coverage of key domains while allowing for emergent issues [30].

Participants will be invited to describe recent adverse events or periods of uncertainty and their responses. Facilitators will use Delphi-derived prompts (e.g., self-efficacy, social support, institutional trust, digital readiness, financial buffers, gender norms) to explore which factors were most influential in their trajectories. Where focus groups are conducted, brief participatory exercises (such as card-sorting and scenario prioritisation) may be used to guide refinement of DCE attributes.

2.3.3 Link to analysis and DCE design.

Interview and focus-group transcripts will be analysed using the same directed content analysis approach applied to open-ended Delphi responses [31]. Coding will draw on the Delphi-derived codebook, with additional inductive codes created where new mechanisms or contextual nuances emerge. Two researchers will code transcripts independently using qualitative analysis software (e.g., NVivo), and inter-rater reliability will be assessed using Cohen’s [27]. Coding discrepancies will be resolved through discussion and consensus; where disagreement persists, a third senior researcher will adjudicate. Thematic matrices will be used to compare factors across groups and to identify convergent, complementary, or divergent patterns.

These qualitative findings will play a dual role: (a) deepening understanding of how drivers and drainers of entrepreneurial resilience operate in practice, including instances where the same factor functions as both an enabler and a constraint; and (b) informing the refinement of DCE attributes and levels to ensure that the experimental design reflects realistic, interpretable, and policy-relevant trade-offs for entrepreneurs.

2.4.1 Experimental design and choice tasks.

The DCE will be administered online using SurveyEngine. Each choice task will present respondents with two hypothetical policy support alternatives and a status quo (no additional support) option. Respondents will be asked to select their preferred option in each task. A D-efficient experimental design will be generated using NGene software. Initial priors will be obtained from the pilot survey and qualitative insights from the Delphi and qualitative phases. These priors will be updated following pilot estimation to refine the final experimental design for the main survey. The number of choice tasks per respondent will be limited to maintain cognitive feasibility (approximately 8–12 tasks), with blocking applied where necessary. Attributes will be coded using effects coding to facilitate interpretation and avoid confounding with the intercept. The design will ensure level balance and minimal correlation across attributes, while allowing estimation of key main effects and a limited set of theoretically motivated interactions (e.g., type of support by targeting).

2.4.2 Study sample and recruitment.

The DCE sample will consist of entrepreneurs, SME owner–managers, and key ecosystem actors such as incubator managers and investor representatives. SurveyEngine, an international panel provider with experience in recruiting specialised business samples, will administer the survey and manage recruitment across the three countries. Soft quotas will be applied to ensure diversity across gender, age (youth versus senior entrepreneurs), sector, and digital intensity. We aim to recruit approximately 300 respondents per country (total target sample of around 900 participants). This target aligns with established recommendations for estimating mixed logit models and preference heterogeneity in DCEs [32,33]. Final sample targets may be refined based on pre-test response rates, the number of attributes and levels included in the experimental design, and specific model identification requirements.

2.4.3 DCE analysis.

DCE data will be analysed within a random utility maximisation framework [34]. Each choice will be modelled as a function of the attributes and levels presented in the choice tasks. Mixed logit models [35,36] will be used to estimate preference parameters and to account for random taste heterogeneity across respondents. These models will quantify the trade-offs entrepreneurs are willing to make across policy and support characteristics.

From the estimated models, we will derive:

• the relative importance of each attribute in shaping support preferences;
• marginal rates of substitution between key attributes (e.g., mentoring versus liquidity support, or digital enablement versus processing time);
• predicted uptake of alternative policy packages under different design scenarios; and
• subgroup differences in preferences (e.g., by gender, youth status, digital intensity, or exposure to recent shocks).

Robustness checks will include alternative model specifications such as conditional logit or latent class models. Quantitative findings from the DCE will be triangulated with Delphi and qualitative insights from interviews and focus groups to identify which combinations of policy supports are both highly valued and contextually feasible, thereby informing evidence-based interventions to strengthen entrepreneurial resilience.