-
Notifications
You must be signed in to change notification settings - Fork 0
Human Sense making and triangulation
CRISP-T operationalizes triangulation—a cornerstone of rigorous qualitative inquiry—by structuring processes through which textual themes and numeric patterns iteratively inform one another. Human sense-making remains central: computational outputs (topics, clusters, coefficients) are treated as interlocutors in theorizing rather than authoritative conclusions.
| Dimension | Textual Artifact | Numeric Artifact | Integrative Action |
|---|---|---|---|
| Thematic Convergence | Topics/Keywords | Feature Importance | Compare high-importance numeric predictors with frequent thematic occurrences |
| Pattern Confirmation | Summaries/Sentiment | Regression Coefficients | Validate directional hypotheses from narrative tone against quantitative signs |
| Structural Alignment | Narrative Variability | TDABM / PCA Structure | Map narrative categories to topological or latent component regions |
| Category Refinement | Memos/Codes | Cluster Profiles | Use cluster-specific narrative sets to collapse or differentiate codes |
| Theoretical Saturation | Exemplar Documents | Performance Metrics | Assess if new numeric evidence shifts category boundaries |
- Data import and preprocessing unify documents and DataFrame under a corpus.
- Exploratory textual analysis surfaces emergent concepts (e.g., "time pressure").
- Numeric analyses (regression, decision tree) quantify associations among variables.
- Semantic filtering isolates document subsets aligned with numeric signals.
- Memo writing integrates both modalities: "High travel_time cluster narratives emphasize schedule fragmentation—supports resource strain category."
- Iteratively revisit coding with contradictory or ambiguous quantitative findings.
- Constant Comparison: Semantic search plus filtering enables rapid juxtaposition of new document subsets.
- Theoretical Sampling: Numeric outliers or topological boundary balls indicate where further qualitative examination may yield novel categories.
- Selective Coding: Predictive model outputs guide consolidation around a core explanatory construct.
In CRM support analysis, an emergent theme "integration complexity" arises. A decision tree highlights api_errors_rate and regression confirms its negative association with renewal_probability. Narrative excerpts retrieved semantically from high-error accounts deepen the category, triangulating descriptive and inferential layers.
Patient narratives about "poor sleep" align with low sleep_balance and high fatigue_score clustering. TDABM shows a connected region of such cases, strengthening a selective code on sleep-mediated fatigue. Contrasting narratives with normal sleep but high fatigue prompt theoretical refinement—exploring inflammatory markers.
| Element | Description |
|---|---|
| Category | Name (e.g., Temporal Resource Strain) |
| Qualitative Evidence | Exemplar quotes, topic prevalence |
| Quantitative Evidence | Coefficient magnitudes, decision tree splits |
| Integrative Proposition | Formal statement linking modalities |
| Counterevidence | Divergent cases prompting revision |
- Avoid algorithmic determinism—numeric confirmation does not override narrative nuance.
- Maintain participant context; anonymize sensitive identifiers.
- Document analytic decisions for transparency and auditability.
- Automated memo scaffolding from metadata.
- Consensus modeling (ensemble interpretations) to stabilize triangulation.
- Mixed-effects regression for hierarchical qualitative structures.
- Mettler et al. (2025) — Computational text analysis reflection.
- Rudkin & Dlotko (2024) — Structural analysis via Ball Mapper.