Nguyen-Trung & Nguyen (2026) — Narrative-Integrated Thematic Analysis (NITA)

TL;DR: NITA is a six-stage non-coding approach to LLM-assisted thematic analysis that replaces the coding-categorizing step entirely with narrative construction, positioning LLMs as dialogic thinking partners rather than pattern-matching engines. Built on the PERFECT reflexive monitoring framework, it is explicitly designed for experienced qualitative researchers working in interpretive traditions that have rejected AI-driven coding.

Problem

The dominant models for AI-assisted qualitative research either (a) use AI as a coder — having it apply or generate codes, then measuring reliability — or (b) use AI as a dialogic partner in ways that still involve some form of category generation. Neither model is satisfactory for researchers committed to narrative and meaning-centered epistemologies. The former is incompatible with Big-Q research; the latter risks slipping back into coding-by-another-name.

Nguyen-Trung & Nguyen take as their starting point the Jowsey et al. (2025) ecosystem of concern: Braun & Clarke’s reflexive TA has explicitly rejected GenAI, and any NITA claim to compatibility with reflexive traditions would be disingenuous. But the authors observe that the range of TA traditions is broader than reflexive TA alone. The gap they address is for researchers who want narrative-centered, meaning-focused analysis — explicitly not code-and-count — but who are open to AI assistance for theme generation at scale.

Approach

NITA proceeds in six stages:

1. Planning PERFECT — the researcher works alone, completing the seven PERFECT components before any AI contact with data.
2. Preparation — data cleaned and prepared; familiarity established; analyst identity secured.
3. Generating candidate themes — LLM generates candidate themes from the prepared data corpus; researcher reviews, challenges, and revises these iteratively.
4. Constructing individual narratives — for each data source (interview, document), the researcher constructs a narrative account grounded in the data. AI may assist with draft language, but the narrative argument is the researcher’s.
5. Constructing meta-narratives — themes and narratives are integrated into overarching meta-narrative accounts of what the data means.
6. Writing up — final analytic story told as a coherent interpretive narrative, not a code-frequency report.

The absence of a coding step is deliberate and theoretically grounded. NITA draws on three methodological ancestors: Framework Analysis (Ritchie & Spencer), Matrix Analysis (Miles et al.), and Interpretive Thematic Integration (ITI). All three organize data through conceptual structures rather than code categories. Coding, the authors argue, creates the illusion of objectivity by reducing interpretation to classification — NITA bypasses that illusion by never entering the coding register.

The PERFECT Framework

PERFECT is the reflexive monitoring procedure embedded within Stage 1. Seven components:

• Purpose — researcher articulates what the research is for and why this method
• Envision — imagines what the final analysis might look like; surfaces anticipations and biases
• Realize — examines personal positionality relative to the topic and participants
• Formulate — develops the analytical approach collaboratively with LLM; first point of human-AI contact
• Experiment — tests the approach iteratively on sample data; refines prompts
• Check & Reflect — evaluates whether AI outputs are resonant, distorted, or missing; reflexive audit
• Tune — adjusts the workflow based on what was learned

The first three components (PER) constitute the researcher’s private space, completed before AI contact. The next two (FE) are the human-AI collaborative space. The final two (CT) return to the researcher as sole evaluator. PERFECT functions as a structured reflexive audit trail, not merely a checklist — the researcher’s responses to each component become part of the analytic record.

AI’s Role

AI is positioned as a thinking partner in a dialogic thinking mode — a new mode of analysis the authors add to Freeman’s (2016) taxonomy of qualitative thinking modes (categorizing, narrative, visual, etc.). The dialogic thinking mode treats AI outputs as perspective-challenging provocations rather than evidence or conclusions. The researcher is never downstream of AI; every AI output is upstream input to researcher judgment.

In the worked example (Down East dataset, same corpus used in the GAITA paper nguyen-trung-gaita-2025), the authors used a Custom GPT configured with RAG on the dataset and powered by GPT-4o. This enabled source-grounded responses that could be checked against raw data — reducing hallucination risk and maintaining traceability. The Custom GPT configuration is described in enough detail that it is reproducible.

Results: of six final NITA themes, five matched themes from the GAITA coding-based approach. One additional theme — “Resisting change” — was found by NITA but not by GAITA. The authors attribute this to NITA’s holistic narrative orientation, which can detect thematic patterns that resist coding because they appear as narrative posture rather than lexical frequency.

Epistemological Stance

Pragmatist, nonpositivist. The authors are explicit that NITA is not reflexive TA (Braun & Clarke’s approach, which has rejected GenAI) and does not claim Big-Q compatibility in that specific tradition. But they situate NITA in the broader range of TA traditions that include interpretive, narrative-centered, and constructionist approaches. The epistemological position is closer to what nicmanis-spurrier-ai-guide-2025 call the “middle ground” — not the positivist pole (reliability-focused, code-frequency analysis), but not the reflexive TA pole that requires distinctly human meaning-making.

The dialogic thinking mode is epistemologically significant: it redefines what AI assistance does from classifying to challenging. This shifts the epistemological burden from the AI (can it code correctly?) to the researcher (can they resist, revise, and challenge AI’s framings?).

Rigor and Trustworthiness

NITA builds rigor into structure rather than metrics. PERFECT provides a documented reflexive audit trail — the check & reflect and tune stages generate a record of how the researcher evaluated and modified AI outputs. This replaces reliability coefficients with process transparency, aligning with trustworthiness criteria in the qualitative tradition (validity-trustworthiness).

The same-corpus comparison with GAITA provides indirect validation: if a non-coding narrative approach produces largely convergent themes as a carefully validated coding approach, the non-coding results are credible. The one additional theme (“Resisting change”) is offered as evidence that NITA can find what coding misses — but this is a single worked example on a single dataset, not a systematic validation study.

Limitation the authors acknowledge: NITA requires experienced qualitative researchers. The reflexive audit procedures, the interpretive authority required to challenge AI outputs, and the narrative construction skills presupposed are not entry-level competencies. The authors explicitly state NITA is not for novices, and warn that its use by researchers without sufficient qualitative grounding would produce a hollow simulation of interpretive analysis.

Limitations Not Acknowledged

• The worked example compares NITA to the same researcher’s prior GAITA study on the same dataset — this is a validity check against the researcher’s own earlier work, not an independent validation.
• “Dialogic thinking mode” as a new addition to Freeman’s taxonomy is proposed, not established. Whether the field will adopt this framing is unresolved.
• The RAG-configured Custom GPT is described as a solution to hallucination, but no systematic hallucination testing is reported. The Jowsey et al. (2025) hallucination finding (58% fabricated quotes) under unguided AI conditions is not addressed.
• NITA’s relationship to the ongoing Jowsey/De Paoli/Greenhalgh debate is not directly engaged. The paper situates itself outside reflexive TA but doesn’t fully address whether the broader epistemological objections to GenAI apply.

Connections

• nguyen-trung-gaita-2025 — same first author, same dataset; GAITA is the coding-based precursor; NITA is the non-coding evolution. Reading these together shows a methodological trajectory.
• friese-caai-framework-2026 — CAAI also replaces coding with dialogic interaction. NITA and CAAI are the two most developed non-coding AI-assisted TA frameworks in the corpus; their comparison is productive.
• wise-et-al-2026-ai-not-the-enemy — AI-in-the-loop analysis similarly enlists AI for qualitative commitments rather than replacing them. Wise et al. engage LLM architecture explicitly in a way NITA does not.
• jowsey-et-al-2025-we-reject — the ecosystem of concern that motivates NITA’s explicit distance from reflexive TA.
• qualitative-ai-methods — NITA sits in the dialogic partner category, with the important distinction that it explicitly avoids the coding register.
• human-ai-collaboration — PERFECT framework is a detailed model of human-AI labor division that warrants inclusion.
• validity-trustworthiness — PERFECT as reflexive audit trail; process transparency as trustworthiness criterion.
• prompt-engineering — ACTOR prompting from GAITA carries over; the PERFECT Formulate/Experiment stages are prompt development stages.
• epistemic-flattening — the dialogic thinking mode is designed precisely to counter epistemic flattening; the researcher’s challenge and revision role is the structural defense.