This chapter examines how innovation capacity is distributed across the Estonian public sector. Using organisational-level results from the OECD Innovative Capacity survey and k-means clustering, it identifies three distinct organisational profiles: empowered front-runners, disconnected adopters and the constrained majority. These clusters differ in their levels of innovation activity and the organisational conditions that support it. These supports include leadership, skills, psychological safety, resourcing and technology use. The findings suggest that Estonia’s public sector does not operate as a single innovation system, but as a three-speed landscape shaped by diverse capacities and circumstances. The chapter recommends how Estonia can strengthen public sector innovation by leveraging empowered front-runner organisations, stabilising disconnected adopters and unlocking the potential of the constrained majority.
Strengthening the Innovative Capacity of the Public Sector of Estonia
4. Innovation in a three-speed public sector
Copy link to 4. Innovation in a three-speed public sectorAbstract
The pre-conditions of innovation tend to reinforce themselves over time. If innovation systems are not actively steered and developed, then existing structural elements and chance (e.g. having better leadership or resourcing over time) can play a large role in diverging capacity across the system. Organisations with psychological safety, slack resources and authorising environments accumulate experience, confidence and legitimacy to innovate, while those without them become progressively more risk‑averse and reactive. As a result, innovation systems tend to sort themselves into qualitatively distinct clusters if not counteracted by other levers. Understanding these clusters reveals where innovation is enabled, where it is fragile and where it is constrained.
Organisational clusters shape how innovation initiatives play out in practice. The same reform will be implemented differently depending on an organisation’s circumstances. In high‑capacity organisations, reforms may accelerate experimentation and scaling; in low‑capacity ones, they may increase stress, symbolic compliance or innovation theatre. There is no one-size-fits-all approach.
Innovation clusters are expected within the Estonian public administration due to its organisational heterogeneity and decentralised nature, with few organisational-level levers connected to innovation in place. The Estonian public sector contains digital front-runners and those relying on legacy technology, very large and very small agencies, as well as well-resourced municipalities alongside resource-constrained ones. Each organisation’s work is shaped by involvement in local, national and EU‑level projects, organisational slack and the priorities of its top leadership. Some organisations are operating close to a systemic innovation model, others are innovating despite the system, and many are stuck in a constrained middle where it is hard to go beyond incremental change.
The OECD’s innovative capacity (IC) survey results suggest that Estonia’s public sector contains clusters of different innovation types. These are empowered front-runners, disconnected adopters and the constrained majority. These clusters reflect structural and organisational differences. They suggest that there is no “average” public sector organisation, but three different tendencies when it comes to public sector innovation.
Mapping the organisational innovation landscape
Copy link to Mapping the organisational innovation landscapeTo map the public sector innovation landscape, the OECD applied a k-means clustering algorithm to organisational-level survey data. This machine learning technique groups organisations based on their similarity across multiple innovation variables, such as psychological safety, leadership support and Generative AI (GenAI) adoption. A detailed breakdown of the technical mechanics, mathematical logic and foundational literature for this approach is provided in Annex 4.A.
The analysis grouped Estonian public sector organisations into three distinct innovation clusters based on their similarity. In Figure 4.1 of standardised z-scores, the central vertical dashed line represents the average across all surveyed organisations. Points plotted to the right mean the cluster scores above average for that trait, while points to the left mean it scores below average. The three clusters reveal a clear spectrum of innovation maturity across Estonia’s public sector, moving from a lagging group to a highly advanced, tech-forward group. The identified clusters were named: 1) Disconnected adopters; 2) Empowered front-runners; and 3) Constrained majority.
Figure 4.1. Standardised z-scores for three innovation clusters in Estonia
Copy link to Figure 4.1. Standardised z-scores for three innovation clusters in Estonia
Note: 1) Disconnected adopters; 2) Empowered front-runners; and 3) Constrained majority.
Source: OECD based on the Public Sector Innovation Capacity Survey (2026).
Rather than viewing these groupings as rigid final verdicts, they are best understood through their “centroids” – the mathematical averages that define the "typical" profile of an organisation within that group (see further in Annex 4.A). The data revealed that organisations within the same cluster share highly specific structural and cultural traits. For example, a typical profile within one cluster might feature strong leadership support and high digital skills, paired with low institutional permission to experiment. While the algorithm assigns every organisation to exactly one cluster, individual entities within a group will naturally deviate from the average. An organisation might exhibit moderate rather than high leadership support, but it remains classified within the cluster it matches most closely.
Different clusters reflect different innovation profiles
Copy link to Different clusters reflect different innovation profilesThe clusters reflected different approaches to innovation and the kinds of innovation they delivered. Figure 4.2 reveals that while baseline innovation occurs across the entire Estonian public sector, the strategic nature and frequency of these initiatives diverge sharply by cluster. Empowered front-runners (Cluster 2) demonstrate the highest overall occurrence of innovation, followed closely by the constrained majority (Cluster 3), whereas the lower-capability disconnected adopters lag behind in translating their efforts into realised projects.
Across all three clusters, incremental enhancement innovation remains the most common approach, with even the lagging organisations (e.g., in Cluster 1) focusing their limited capacity almost exclusively on these operational improvements. However, a stark divide emerges regarding more complex, strategic initiatives. Empowered front-runners and the constrained majority (Clusters 2 and 3) exhibit a much stronger propensity for adaptive and mission-oriented innovations. Anticipatory, future-focused innovation remains the least common type nationwide. Ultimately, these output variations underscore that while absolute funding levels may remain consistent between groups, differences in internal organisational capacity, team dynamics, and resourcing models directly dictate whether an entity can move beyond basic enhancements into complex forms (e.g., mission-driven) public sector innovation.
Figure 4.2. The innovation clusters identified in the Estonian public sector
Copy link to Figure 4.2. The innovation clusters identified in the Estonian public sector
Source: OECD based on the Public Sector Innovation Capacity Survey (2026).
The clusters also reflected different levels of resourcing, team dynamics and ways of working (Figure 4.3). Some clusters reflected organisations that reported significantly lower team and organisational capacity for innovation. There was also a stark difference in the level of dedicated budget and external funding available to different organisations; however, ministry funding and funding amounts did not differ markedly between clusters. This is not to say that funding is high or low in absolute terms, but rather that clusters did not differ markedly from one another in terms of ministry funding and funding amounts (whether high, low or in between). For example, empowered front-runners seem to enjoy a distinct competitive advantage. While their baseline ministry funding does not set them apart, they score exceptionally high on external funding, dedicated innovation budgets and risk appetite (Figure 4.3). An overview of each cluster is provided below.
Figure 4.3. Resources, funding and enabling conditions per cluster
Copy link to Figure 4.3. Resources, funding and enabling conditions per cluster
Note: 1) Disconnected adopters; 2) Empowered front-runners; and 3) Constrained majority.
Source: OECD based on the Public Sector Innovation Capacity Survey (2026).
Cluster 1: Disconnected adopters
Disconnected adopters are typically very small organisations operating on the margins of the system. They lack institutional support, stable funding and psychological safety. They reported the highest levels of GenAI adoption, often using technology to compensate for staff shortages.
Disconnected adopters also scored lowest on psychological safety and permission to innovate. They reported large gaps between leadership expectations and staff capability, particularly around GenAI literacy. Innovation in these organisations tended to be reactive and survival-driven rather than strategic. There was evidence that the organisations were more prone to “innovation theatre”, where leadership reported a higher rate of innovation than did the staff under them.
However, disconnected adopters’ innovation rate was not significantly different from that of the constrained majority (see above), and their innovation outcomes were slightly lower, suggesting that their resource-constrained, innovation-by-necessity circumstances have not translated automatically into more successful innovation or better outcomes.
Surprisingly, disconnected adopters reported the greatest innovation skills gaps. They had the greatest distance between how important their organisational leadership said innovation skills were and how confident their staff felt delivering those same skills (see Figure 4.4 below). This gap was most pronounced among managers’ innovation skills, especially their ability to link innovation to public impact and to use GenAI – which is striking because disconnected adopters also reported the greatest use of GenAI of any cluster.
Figure 4.4. Innovation skill gaps (skill demand minus skill supply)
Copy link to Figure 4.4. Innovation skill gaps (skill demand minus skill supply)
Note: 1) Disconnected adopters; 2) Empowered front-runners; and 3) Constrained majority.
Source: OECD based on the Public Sector Innovation Capacity Survey (2026).
Readers should be cautious about generalising the findings of this cluster to other public sector organisations. Only six surveyed organisations were grouped as disconnected adopters, and they represented very small organisations by headcount (the average number of staff per organisation was close to four). There may be a small-numbers effect whereby extreme results are more likely to occur when the sample size is small. However, the magnitude of the difference in how these organisations approached innovation set them clearly apart from the other two clusters.
Cluster 2: Empowered front-runners
Empowered front-runners represent medium-to-large organisations that combine a strong innovation culture (the will) with stable funding and resources (the way), relative to the other clusters. Their innovation occurrence and outcomes were the highest of the clusters, and they scored highest on almost all the variables that contribute to innovation. They are a large cluster of 27 surveyed organisations.
Empowered front-runners score highest on psychological safety, risk appetite and perceived organisational capacity. Unlike other clusters, they show a skills surplus, meaning staff confidence meets or exceeds leadership demands. Their innovations produce the most diverse outcomes, particularly improvements in service quality and reductions in administrative burden. Dedicated budgets and ministry-level funding provide the stability needed to move beyond pilots toward implementation.
Empowered front-runners scored higher than disconnected adopters and the constrained majority on all clustering variables except GenAI adoption. The disconnected adopters outperformed the other clusters on GenAI adoption: the z-score for GenAI adoption was close to 0.2 for empowered front-runners, compared with nearly 0.5 for disconnected adopters and -0.2 for the constrained majority.
Nonetheless, the fact that empowered front-runners score consistently higher across such a wide range of innovation factors is noteworthy. This may reflect an “Anna Karenina” effect, whereby all successful endeavours share the presence of many diverse preconditions for success, while each unsuccessful endeavour is unsuccessful for lack of any one of these preconditions. Empowered front-runners may be more likely to possess all or most of the preconditions as a bundle, whose positive effect is disproportionate to their individual contributions.
Cluster 3: The constrained majority
The constrained majority is the largest cluster, representing much of the Estonian public service. Among surveyed organisations, 40 were grouped in the constrained majority. These organisations sit slightly below average on most innovation dimensions and are characterised by inertia rather than resistance.
The constrained majority’s reported innovation rate was low, only 23% vs 34% in the empowered front-runners. The constrained majority had the lowest GenAI uptake, scored low on innovation drivers and were strongly associated with having no dedicated innovation funding. Innovation tended to occur only when triggered by external mandates. They reported the most acute skills gaps at the managerial level, particularly in assessing public impact, using GenAI and integrating new knowledge into practice. The OECD validation mission confirmed that this group exemplifies the “frozen middle” of the Estonian public sector.
Potential drivers and systemic origins of cluster divergence
Copy link to Potential drivers and systemic origins of cluster divergenceWhile the empirical data from the innovation clustering does not allow us to make definitive causal assumptions regarding how these groupings originally formed, combining these profiles with the system-level barriers identified in Chapter 2 and prior international research on public sector innovation can offer a powerful explanatory framework. The divergence of the empowered front-runners strongly underscores a structural Matthew Effect that is actively accelerated by Estonia's funding architecture. The Matthew Effect (Merton, 1968[1]) describes a cumulative compounding advantage – in an administrative context, it means that organisations possessing even a slight initial advantage in capability find it easier to secure additional opportunities and resources, while those lacking capacity are increasingly marginalized. Because the state lacks dedicated central innovation budgets and relies on external funds for its innovation financing, organisations with higher initial administrative capacity have more potential to successfully monopolise these complex grant streams, concentrating also procurement capability so intensely that a tiny 10% slice of public entities executes 98.7% of all innovative procurement nationwide (Chapter 2). This is echoed by research in other public administration settings where external networks and non-standard funding streams can act as catalysts for public-sector innovation (Demircioglu and Audretsch, 2017[2]).
Crucially, in Estonia, this systemic environment curtails the ambition of even these front-runners; despite their superior capability and resource mix, they produce strikingly little radical or transformative innovation, remaining overwhelmingly confined to incremental, enhancement-oriented process digitalisation. Because project-based EU grants lack long-term financial sustainability and carry strict financial penalties for failure, even the most advanced organisations are heavily incentivised to de-risk their portfolios, choosing to digitise existing legacy processes rather than attempting the high-stakes experimentation required to fundamentally reinvent policy logic.
Conversely, the constrained majority reflects entrenched institutional inertia that is systematically enforced by the wider administrative ecosystem. This frozen middle may not be merely risk-averse by choice; rather, rigid budget planning cycles and aggressive efficiency drives have hollowed out the 'slack resources' and uncommitted time necessary for transformative thinking. This combines with a retrospective, fault-finding audit culture and procurement habits that favour lowest-price criteria, reflecting a broader government expectation that procurement should operate this way given the involvement of public money. These factors mean standard bureaucracies are effectively locked into a path-dependent cycle, defaulting to low-risk operational enhancements only when explicitly compelled by top-down legislative mandates. Meanwhile, the tiny cohort of disconnected adopters highlights the acute margin of this system, where severe staffing shortages in micro-organisations force a reactive, survival-driven reliance on tools like GenAI to maintain baseline operations, entirely bypassed by strategic managerial support or psychological safety.
These divergent clusters are anchored in cross-cutting structural realities, which warrant tailored, cluster-specific interventions rather than blanket solutions. Future policy frameworks at the system level are also crucial. They must focus on systemic, government-wide reforms (as described in Chapter 2) – such as establishing a centralised innovation procurement competence centre, building dedicated domestic risk-mitigation funds and restoring institutional slack – designed specifically to lower the legal and financial stakes for the frozen middle, while providing the structural safe spaces necessary to unlock genuinely transformative, anticipatory capacity among the front-runners without worsening existing ecosystem fragmentation.
Crucially, as highlighted by Pollitt and Bouckaert (2017[3]) in their analysis of public management reform, 'frozen middle' dynamics inevitably occur when structural reforms shift macro-level metrics but fail to alter internal cultural incentives and day-to-day managerial realities. This underscores that top-down systemic reform without targeted, organisation-level cultural supports will fail to gain traction in legacy bureaucracies, just as isolated internal agency supports will remain starved without wider systemic adjustments; the two must be deployed in tandem to successfully break the cycle of inertia. Hence, the following recommendations should not be divorced from recommendations from Chapter 2 and from other parts of the assessment.
Recommendations
Copy link to RecommendationsThe analysis indicates that innovation capacity building must be differentiated by cluster rather than applied uniformly. On the organisational level:
Disconnected adopters primarily need cultural and managerial enablers: psychological safety, autonomy and safe-to-fail environments. They also face large skills perception gaps between leadership and staff, reinforcing the risk of innovation theatre if expectations are not recalibrated or capability is not developed.
Empowered front-runners could also benefit from continued investment in innovation culture, but their binding constraints are different. They need greater evaluation capability (especially impact and value-for-money evaluations – see Chapter 6) as well as secure, flexible funding to support scaling rather than perpetual piloting. They also need to be encouraged to take up more transformative innovation practices and goals (connected to anticipation and missions).
The constrained majority requires motivation and permission to innovate, alongside practical enablers such as support to adopt GenAI and sludge audits to reduce administrative burden. Funding constraints are central, but the evidence suggests that resources alone are insufficient without autonomy and managerial capability.
Notably, both local and central government include empowered front-runners, demonstrating that high innovation capacity is possible at multiple administrative levels when culture and resources align.
Local government organisations are more frequently found among disconnected adopters, reflecting small organisational size, limited resources and high service demands. While some of this reflects small-number effects, validation interviews confirm that it also accurately reflects their operating reality.
Central government organisations are more heavily represented in the constrained majority, particularly policy-heavy ministries with limited dedicated innovation funding.
Recognising these clusters makes visible both Estonia’s latent strengths and its fault lines. It shows that advancing public sector innovation requires a multi-speed strategy: protecting and leveraging front-runners, stabilising disconnected adopters, and unblocking the constrained majority.
Disconnected adopters
Start with targeted diagnostics and visible leadership support. Disconnected adopters face deep enabling constraints. They report low psychological safety, weak permission to innovate, and low managerial capability. Government Office should start a series of targeted diagnostic reviews in these organisations. Top leadership in disconnected adopters should then set clearer (and realistic) expectations, signal permission to experiment, and support staff who already innovate under pressure. This approach will address their systemic constraints better than a generic capability-building programme.
Recalibrate expectations and build core skills. Disconnected adopters show large gaps between what leaders expect and what managers can do. The largest gaps concern GenAI literacy, public impact and orchestration. Leaders should align expectations with current capacity. Government (the Government Office and Ministry of Finance) should then provide targeted upskilling, shared tools and practical support. It should also consider seconding experienced innovation leaders to these organisations, where they can share skills and foster effective ways of working.
Target managerial capability where it matters most. Leadership should focus on building the skills, noted above, that most constrain performance: GenAI literacy, public impact, and orchestration. Training should help managers govern GenAI use, define success beyond outputs, and review innovation portfolios more deliberately. Shared prototyping support can also reduce the burden on very small organisations.
Make existing innovation visible and legitimate. Disconnected adopters already use GenAI out of necessity. They also already innovate at a similar rate to the constrained majority (but below that of empowered front-runners). Government Office should identify these innovation efforts, document them, and share them across the system. Visible recognition can reduce isolation, strengthen confidence, and help these organisations attract support. It can also shift perceptions from low capacity to constrained but adaptive performance.
Empowered front-runners
Diversify the innovation portfolio. Front-runners already innovate more than other organisations. The next step is to broaden the types of innovation they pursue. Current activity appears more focused on enhancing what already exists than on adaptive, anticipatory or mission-oriented work. The Government Office should create targeted mandates and funding for these less common forms of innovation (for example, by running targeted campaigns to take up the anticipatory grants proposed in Chapter 2). It should also link them to strategic priorities, including the next state planning cycle.
Use front-runners to build capability across government. Empowered front-runners combine strong culture, strong skills, and stable resources. Government should use this capacity more deliberately. Estonia already has secondments within its public administration system, but the Government Office and the Ministry of Finance could use this mechanism more strategically by creating a targeted innovation-capability secondment programme. This programme could place individuals, pairs or small teams from front-runner organisations in lower-capacity organisations for a fixed period, with a clear mandate to transfer skills, support real projects and strengthen organisational routines.
Support secondments with clear mandates. Government Office should lead the creation of a playbook for making innovation secondments work well. Secondments need formal backing to work well. Senior leaders should define the purpose of each placement, protect time for knowledge transfer, and expect receiving organisations to apply what they learn. This can help convert individual expertise into wider organisational capability.
The constrained majority
Set explicit innovation expectations. The constrained majority tends to innovate only when external pressure creates a reason to act. The Government Office should explore levers to set stronger innovation expectations in the State Plan 2028–2034. For this group, measures in the annual budget planning and performance budgeting frameworks could also be especially beneficial (as described in Chapter 2). This can create the external driver that these organisations currently lack. However, the Government Office should be careful not to let extrinsic organisational incentives to innovate translate into extrinsic motivators at the team or individual level that might extinguish existing intrinsic motivations to innovate. Therefore, expectations for organisations to innovate should be framed to staff as fostering an authorising environment for innovation, rather than as a pretext for setting individual-level innovation targets or making individual-level innovation incentive payments.
Create permission and support for intrapreneurship. The constrained majority needs stronger authorising environments. Government Office should work with senior leaders to communicate how innovation is (or will be) valued and legitimate in their organisation. They should back this message with small, protected budgets for staff-led pilots. This combination of permission and practical support can help move organisations beyond inertia.
Reintroduce targeted innovation sprints. The Government Office should reintroduce sprints and consider how it can prioritise conducting innovation sprints with constrained-majority organisations. These sprints can be a powerful motivational device to drive innovation, help staff apply innovation methods to real problems and build confidence through use.
Across clusters
Match funding to organisational maturity. A single funding model does not fit all three clusters. The Government Office, together with the Ministry of Finance, should consider small readiness grants for disconnected adopters and the constrained majority, or consider how block funding or organisational savings (described in Chapter 2) could be tied to innovation capacity targets (e.g. in skills development, innovation prevalence and culture). These grants can support basic governance, skills development and early experimentation. It should reserve larger competitive funds for front-runners that are ready to scale proven work. Clear criteria should guide access to each funding stream.
Create a tailored capability-building offer for different organisational groups. Government Office should oversee the creation of a tailored programme that responds to the specific needs of different organisational clusters. The programme should offer targeted masterclasses, short learning sessions and practical clinics based on each group’s main constraints and priorities. For example, disconnected adopters may need hands-on support to institutionalise safe GenAI use, public impact and innovation management, while the constrained majority may need support on intrapreneurship, experimentation and reducing administrative burden (e.g. sludge). Front-runners may benefit more from advanced sessions on scaling, evaluation and anticipatory innovation. A tailored offer can make support more relevant, improve uptake and help organisations apply what they learn to real problems.
References
[2] Demircioglu, M. and D. Audretsch (2017), “Conditions for innovation in public sector organizations”, Research Policy, Vol. 46/9, pp. 1681-1691, https://doi.org/10.1016/j.respol.2017.08.004.
[5] Hartigan, J. and M. Wong (1979), “Algorithm AS 136: A K-Means Clustering Algorithm”, Applied Statistics, Vol. 28/1, p. 100, https://doi.org/10.2307/2346830.
[8] Jain, A. (2010), “Data clustering: 50 years beyond K-means”, Pattern Recognition Letters, Vol. 31/8, pp. 651-666, https://doi.org/10.1016/j.patrec.2009.09.011.
[6] Kaur, M. et al. (2022), “Innovative capacity of governments: A systemic framework”, OECD Working Papers on Public Governance, No. 51, OECD Publishing, Paris, https://doi.org/10.1787/52389006-en.
[4] Lloyd, S. (1982), “Least squares quantization in PCM”, IEEE Transactions on Information Theory, Vol. 28/2, pp. 129-137, https://doi.org/10.1109/tit.1982.1056489.
[1] Merton, R. (1968), “The Matthew Effect in Science”, Science, Vol. 159/3810, pp. 56-63, https://doi.org/10.1126/science.159.3810.56.
[3] Pollitt, C. and G. Bouckaert (2017), Public Management Reform: A Comparative Analysis - Into the Age of Austerity, Oxford: Oxford University Press.
[7] Rousseeuw, P. (1987), “Silhouettes: A graphical aid to the interpretation and validation of cluster analysis”, Journal of Computational and Applied Mathematics, Vol. 20, pp. 53-65, https://doi.org/10.1016/0377-0427(87)90125-7.
Annex 4.A. Methodological overview of k-means clustering
Copy link to Annex 4.A. Methodological overview of k-means clusteringTo identify the organisational clusters, the OECD applied a k-means clustering algorithm to organisational-level survey data. K-means is a partitional, distance-based clustering algorithm. Given n organisations each described by a vector of numeric features (say, scores on leadership, experimentation, skills, data-sharing, risk appetite), it partitions them into k groups so as to minimise the total within-cluster sum of squares — the squared Euclidean distance between each organisation and its cluster's mean (centroid). It iterates by assigning each organisation to the nearest centroid, recomputing centroids, and repeating until assignments stop changing (Lloyd, 1982[4]).
In this case, the k-means algorithm clustered public sector organisations based on similarity across multiple variables related to innovation, such as innovation drivers, psychological safety, skills perception, GenAI adoption and resourcing. Each organisation’s variables were transformed to standardised scores (z-scores), indicating how far it sits above or below the mean for all organisations on that variable. The algorithm then iteratively assigned organisations to clusters, so that within-cluster similarity is maximised, and between-cluster difference is minimised. The algorithm is stochastic in that it picks a random starting point for each cluster and iterates towards the best grouping it can find based on its algorithm.
One virtue of k-means clustering is that it relies on a centroid which sits at the centre of a given cluster. The centroid thus represents a "typical" organisation in that group, e.g. high leadership support and skills but low institutional permission to experiment. The centroid acts as an interpretable shorthand for the cluster as a whole. Organisations assigned to that cluster may deviate from it on some variables, e.g. high leadership support and skills but moderate institutional permission to experiment; or moderate leadership support and skills but low institutional permission to experiment. However, each organisation is classified according to the centroid it matches best, and the centroid affords analysts an interpretable shorthand of what characterises that cluster.
Algorithms like k-means clustering also suit innovation-capacity analysis because it categorises organisations' observed profiles across multiple behavioural, cultural and capability dimensions simultaneously, rather than sorting them into categories defined a priori. Unlike a normative framework, it identifies configurations through a bottom-up analysis of the data. Granted, the algorithm does use a top-down framework to guide its iteration through the data (Hartigan and Wong, 1979[5]). In this case, a top-down framework that specifies which variables to use, e.g. innovation drivers, capacity, team dynamics, GenAI adoption, etc. This framework of variables is based on the OECD’s systemic framework of innovative capacity in government (Kaur et al., 2022[6]). The method then assigns organisations to a given cluster according to those variables using a bottom-up search algorithm. Finally, the analyst evaluates the number of clusters and how they reflect the chosen variables with standard validation tests, e.g. silhouette or elbow diagnostics, to determine the validity of the clusters (Rousseeuw, 1987[7]).
Another virtue of k-means is that it is simple, fast and scalable. This means that the same method can be used for clustering hundreds or thousands of public servants or dozens of organisations (Jain, 2010[8]). This allows public sectors to replicate these clustering techniques in their own work. Its flexibility also allows the model to be run across many different public sectors, with markedly different sample sizes and innovation profiles, while still retaining the core underlying method – supporting better comparability between innovation assessments.
However, readers should note some limitations of k-means clustering. First, determining the optimal number of clusters requires judgement. K-means algorithms attempt to maximise the ratio of intra-cluster to inter-cluster variation; however, the number of clusters that achieve that best ratio per number of clusters is a judgement call. To address this, analysts run silhouette and elbow diagnostics to help determine the optimal number of clusters. However, the diagnostics are not determinative, merely informative, and how many clusters is optimal depends on subject matter expertise. Second, every organisation is classified as belonging to exactly one cluster, even one sitting on a boundary. Innovation profiles are often genuinely mixed, so this can impose false sharpness. The solution to this is to not over-index on which cluster an organisation falls into and treat its cluster assignment as a starting point, not a final verdict. Third, k-means clustering starts its iterations from a random seed, and the random seed can influence the final allocation. This risk is mitigated by running the k-means clustering multiple times with multiple starting seeds, then taking the mean or most representative clusters as the final allocation.