Dr Maria Alejandra Pinero de Plaza, PhD, CHIA
Dr Maria Alejandra Piñero de Plaza is a Research Fellow at the Caring Futures Institute, Flinders University, and Adjunct Fellow at the Medical School of Adelaide University. A Certified Health Informatician Australasia (CHIA) and an MIT-certified Applied AI and Data Scientist, she works at the intersection of implementation science, data science and consumer-led research. She leads the Frail, Homebound and Bedridden People (FHBP) program and develops PROLIFERATE and PROLIFERATE_AI, co-designed frameworks for evaluating innovation in complex health systems. Her methodological interests include co-occurrence and multilayer network analysis of qualitative data, Bayesian modelling and participatory evaluation — with a consistent focus on people whose circumstances exclude them from conventional care and research.
Session
Qualitative studies of multiply disadvantaged groups often finish where network analysis could begin: with a list of themes and counts. When open-text responses are coded at participant level, the resulting affiliation (bipartite) network of participants and issue codes can be projected into a one-mode issue co-occurrence network. We examine whether three separately coded subsets can be harmonised into one defensible co-occurrence network, and what such a network adds to conventional thematic reporting.
The data come from the co-designed, consumer-centric study "Making the invisible visible: Exploring the experiences of frail, homebound and bedridden people" (Flinders University Social and Behavioural Research Ethics Committee, Project No. 8557). We shared an online survey via social media and university press release channels; three researchers coded open-text responses from 140 participants, each working on mutually exclusive subsets (n=46, 47 and 47). Each coder extended the coding framework differently, and identical code numbers referred to different concepts, so we could not pool coder-specific labels directly. We derived a shared semantic core of eight issues comparable across all three codebooks, constructed a participant-by-issue binary matrix, and projected it into a weighted one-mode network with issues as nodes and co-reporting counts as edges. We treated four entries that reproduced the entire shared code list as unresolved in the primary analysis and retained them as authored in a sensitivity analysis. Because coders worked on disjoint subsets, whole-sample intercoder reliability cannot be estimated; we treat this as a structural feature of the design.
Issue coding was present for 122 of 140 participants; 80 had at least one evaluable shared-core issue and form the network denominator (Figure 1A; median 2 issues, mean 2.75, range 1-7). Health burden was most prevalent (48; 34.3%), followed by reduced participation (43; 30.7%), reduced contact (41; 29.3%) and inability to travel (28; 20.0%). The strongest raw co-occurrence was reduced participation with health burden (n=33). The eight-node network formed one connected component with density 0.964 (27 of 28 pairs; Figure 1B); the only absent pair was partner separation with hospitalisation. On a graph this dense, community detection and centrality are not interpretable: we make no community claim, and weighted betweenness — nonzero only because of the single absent edge — is reported as descriptive and shown to shift under sensitivity analysis.
Edge metric choice was decisive. Raw counts conflate prevalence with association. Re-expressing edges as lift (observed/expected under independence) and Jaccard similarity reordered the structure (Figure 1C): loss of paid work with inability to travel showed the highest lift (2.14) and second-highest Jaccard (0.53) despite ranking tenth by count (n=18), while reduced contact with health burden (n=24) co-occurred at chance levels (lift 0.98) and partner separation with health burden below chance (lift 0.56). Sensitivity analysis left prevalence rankings and strongest pairs unchanged.
For research, co-occurrence networks of coded qualitative data reveal structure that thematic reporting hides, but only if edges are modelled as association rather than frequency and interpretation avoids centrality claims on near-complete graphs. For policy, combinations identify a concrete target: a six-issue cluster (reduced contact, loss of paid work, reduced participation, financial hardship, inability to travel and health burden) reported by 8 participants (Figure 1D), a subgroup carrying social, financial and health burdens missed by single-issue responses. The study demonstrates a transparent, auditable path from multiply coded qualitative data to a reproducible network, and motivates a planned multilayer analysis of invisibility, healthcare exclusion, support needs and participant-proposed solutions.