The eHealth readiness of Australia's medical specialists - Final Report

Clustering analysis

Page last updated: 30 May 2011

Purpose

We used cluster analysis to group practitioners with similar needs and attitudes. This allowed us to develop tailored interventions based on each cluster’s distinct motivations.

Through this process, we defined clusters that are:
  • Reachable – i.e. we can target specific interventions to each group
  • Interpretable – i.e. we can understand who they are and what matters most to each
  • Distinct – i.e. different from each other on key attitudes and behaviours.
  • We can easily define reachable clusters based on demographic criteria. However, by adhering to a strictly demographic approach, we lose the ability to incorporate the needs-based insights required to develop targeted interventions. In contrast, while a needs-only approach provides rich and robust insights on attitudes and behaviours, it falls short on the reachability dimension. Therefore, we applied a hybrid’ needs-plus’ clustering approach, which combines needs-based and demographic dimensions so that we can identify targeted interventions that can be effectively delivered to the relevant practitioners (Exhibit 29).
EXHIBIT 29
(D)
We need strong variation in at least some attributes or variables to define groups that are distinct. A ‘needs plus’ clustering approach enables us to identify much greater variation in responses, especially when compared with basic demographic measures such as age, gender and region.

Process to develop ‘needs-plus’ clusters

The ‘needs plus’ approach follows a two-step process, described below.

Step 1: Select variables and create clusters based on needs dimensions:
  • Identify attitude ‘themes’ – We used a factor analysis across all eHealth attitudinal variables to identify themes across attitudes
  • Identify and retain relevant ‘themes’ – We identified a set of themes (approximately 7–8) that were the most representative of eHealth attitudinal dimensions
  • Select final list of variables – We identified one or two variables that were strongly related to each theme and that also exhibited sufficient variability in the sample
  • Conduct 1st stage segmentation – We used hierarchical clustering (Ward’s method8) to fold similar respondents into increasingly larger groups – this approach yielded an initial ‘coarse’ solution to assess
  • Select ideal cluster solution – We ran multiple iterations of the hierarchical clustering analysis and identified the cluster solution that offered the most effective target groups.

Step 2: Refine needs-based clustering by adding behavioural and profile variables to create ‘needs plus’ clusters:
  • Improve needs-based cluster solution – We used k-means clustering on existing hierarchically derived clusters to improve on our solution by adding percent private as a variable to enhance reachability
  • Profile the final solution – We profiled the final solution on all variables to make a comprehensive, final assessment of its quality, relative to the segmentation objectives
  • Refine and ‘bring to life’ the profiling description – We used deep-structured interviews and stakeholder discussions and workshops to test and refine the profiles and ensure that they were both insightful and actionable (i.e. helpful to define relevant intervention strategies, meaningful and reachable).


8 For more detail, refer to Joe H Ward’s “Hierarchical Grouping to Optimise an Objective Function”, Journal of the American Statistical Association, Volume 58, Issue 301 (Mar 1963), 236–244.)