Four behavioural personas
1,014 Londoners · YouGov omnibus · ELWA/NLWA commission · Studio Zao analysis
circular-minded High engagement,
circular-minded Low engagement,
passive disposal High engagement,
passive disposal
X — Clothing engagement: How actively and frequently a persona acquires clothing. Derived from social/relational engagement (RC1) and buying frequency. High = frequent, multi-channel, trend-responsive. Low = occasional, replacement-driven.
Y — Disposal mindset: How circular-oriented their disposal behaviour and intent is. Derived from absence of inertia (−RC3), sustainability values contribution (RC4), self-identified recycler rate, and inverse general waste rate. High = pro-circular, low barriers, good outcomes. Low = passive, reactive, higher waste leakage.
The matrix separates two distinct problems: consumption volume (how much enters the wardrobe) and disposal orientation (how consciously it leaves). C1 and C4 share the right half — both are more active consumers — but sit in opposite quadrants on Y. C4's 8% general waste rate despite active consumption is the key counterintuitive finding: they have the right mindset, just not the frequency. C2 and C3 share the left half but are separated by values: C3's sustainability orientation (RC4 = +2.02) is the defining distance from C2.
Cluster Comparison
Four behavioural personas across 1,014 London respondents
| Stage | C1 Accumulator | C2 Pragmatic Replacer | C3 Ethical Keeper | C4 Seasonal Clearer |
|---|
Personas are behaviourally defined. Demographics are contextual — not defining characteristics.
Relevant for intervention channel targeting — C1's platform profile is markedly different from all other clusters
Factor structure — R3 Varimax rotation
PCA on 87 behavioural variables (n=1,014). 26.9% variance explained across 7 factors. All eigenvalues >1.0 (Kaiser criterion). Robustness-tested against polychoric PCA and Gower k-medoids.
Standard PCA with Pearson correlations was chosen over polychoric PCA for two reasons: the dataset is predominantly binary and ordinal with limited scale range, and polychoric estimation on 87 variables with n=1,014 introduced instability in the correlation matrix. Polychoric was tested as a robustness check and produced a near-identical factor structure with marginally different loadings, confirming the standard approach was appropriate.
An R4 refactoring attempt (8 factors) was explored to test whether the RC1 social/relational engagement factor could be decomposed further. The additional factor failed to achieve eigenvalue >1.0 and produced cross-loadings that reduced interpretive clarity. R3 (7 factors) was retained as the most parsimonious solution with clean separation between constructs.
Cluster count was selected at k=4 via silhouette analysis (0.18, modest but stable). k=5 was tested and produced a micro-cluster (n=47) that fragmented C1 without adding interpretive value — it separated a high-digital subgroup that was better captured as a within-cluster dimension. k=3 merged C3 and C4 into an undifferentiated low-engagement group, losing the critical values/inertia distinction.
Gower distance with k-medoids (PAM) was tested as an alternative clustering approach that handles mixed variable types without distance assumptions. It produced a broadly similar 4-cluster solution (Adjusted Rand Index 0.61 with the PCA-based clusters) but with less clean separation on the sustainability and disposal factors. The PCA + k-means pipeline was retained on both silhouette score and interpretive grounds.