Beyond integration: a coherence and transparency framework for mixed methods research in the era of generative AI

A conceptual analysis in psychology proposes that mixed methods research cannot demonstrate quality merely by combining qualitative and quantitative evidence; the complete study must form a coherent, transparent and defensible research architecture. The proposed Mixed Methods Coherence and Transparency Framework, or MMCTF, organizes that architecture into five interconnected domains: epistemic coherence, theoretical and design coherence, integrative coherence, inferential and ecological coherence, and methodological transparency. Developed by Rune Johan Krumsvik and Marius Ole Johansen, the framework extends rather than replaces existing approaches to integration quality, joint displays, meta-inference and reporting. It is a theoretically derived, proposition-generating conceptual architecture rather than an empirically validated scoring instrument, so its usability, discriminant value and inter-rater applicability remain subjects for future research.

The central premise is that integration is necessary but insufficient for mixed methods coherence. Integration concerns the purposeful interfaces through which strands of evidence are connected, built upon, merged, embedded, transformed or brought together during interpretation. Coherence is broader: it asks whether the investigated phenomenon, philosophical assumptions, theoretical framework, research questions, purpose and design of mixing, methods, analyses, integration mechanisms and resulting meta-inferences collectively form an explicit and defensible whole. A study can therefore include an obvious integration point, such as a joint display, while retaining unexplained tensions among theory, sampling, analysis or interpretation. Conversely, rigorous qualitative and quantitative components may be defensible in their own terms yet remain weakly connected. The analysis accordingly distinguishes component quality, integration quality and overall coherence as related but separate dimensions of methodological quality.

The authors also separate coherence from alignment, fit and convergence. Alignment describes correspondence between particular elements, such as whether research questions fit the selected methods. Fit concerns the relationship between findings, including confirmation, expansion and discordance. Coherence is the higher-order property that emerges when these alignments and interfaces collectively form a defensible architecture. Importantly, disagreement is not evidence of incoherence. Qualitative and quantitative results may confirm, complement, expand or informatively challenge one another, depending on the declared purpose of mixing and the relationships established between the components. The proposed diagnostic sequence begins with an epistemic diagnosis of the kinds and levels of knowledge being claimed, followed by an assessment of theoretical and design compatibility. Researchers then specify how findings should interact, interpret what their relationship reveals about heterogeneity, context or ecological boundaries, and document the decisions, unresolved tensions and uncertainties that shape the meta-inference.

Psychological research illustrates why this distinction matters. Standardized symptom scales may show a statistically significant average improvement after an intervention, while interviews reveal that some participants experienced little meaningful benefit, unintended effects or benefits in domains omitted from the instrument. Such a discrepancy could indicate an unjustified mismatch in constructs, timing, sampling or operationalization, requiring methodological correction. It could instead represent productive discordance arising from different levels of analysis: a population-level estimate of predefined symptoms and person-specific accounts of experience need not be converted into interchangeable evidence. Under the MMCTF, researchers might link symptom trajectories to interview accounts in a joint display, purposefully examine cases that diverge from the aggregate pattern, or use qualitative findings to motivate theoretically justified subgroup analyses. The defensible conclusion would then be bounded rather than artificially convergent—for example, average symptom improvement alongside heterogeneous experiences and benefits not fully captured by the scale.

Vertical and horizontal relationships make the proposed architecture more precise. Vertical coherence follows the pathway from the phenomenon and philosophical assumptions through theory, questions, design, methods, analyses and integration to the meta-inference. Horizontal coherence concerns relationships among components working at comparable levels, including qualitative and quantitative findings, emic and etic perspectives, individual cases and aggregate patterns. The two are connected but not equivalent. A scale score and a person’s account of anxiety are not interchangeable, and their differences may be part of the reason for combining them. The framework is also sensitive to the qualitative tradition involved. Phenomenological inquiry must preserve the meaning and structure of lived experience; grounded theory requires that integration not prematurely constrain theoretical emergence; constructivist approaches must retain reflexivity and co-construction; and situational analysis must avoid reducing relational complexity and multiple positions to simplified variables. The authors call this requirement tradition-sensitive coherence: integration must preserve the defining commitments of the qualitative approach while establishing a defensible relationship with quantitative evidence.

Generative AI creates new methodological interfaces because it can participate in coding, summarization, classification, comparison, exploration of discordant cases, preliminary joint displays and candidate integrative interpretations. Predictive systems may instead identify risk profiles, trajectories, subgroups, anomalies or other patterns in structured or high-dimensional data. The authors emphasize a functional distinction: predictive output is not explanation, and generative output is not interpretation by default. AI-generated categories, summaries or explanations should therefore be treated as candidate analytical outputs whose evidential status depends on empirical grounding, theoretical consistency, contextual adequacy, sensitivity to alternative interpretations, treatment of negative cases and human evaluation. AI may assist researchers in noticing possible relationships, but researchers must return to the underlying qualitative material, quantitative results, theoretical framework and contextual evidence before turning those suggestions into substantive claims. Human researchers retain responsibility for the interpretations and knowledge claims ultimately advanced by the study.

Transparency is central to this process because readers need to reconstruct how AI affected the pathway from evidence to knowledge. The proposed human–AI division of analytical labour distinguishes computational or analytical support from interpretive authority and epistemic responsibility. Researchers should identify the system or model where relevant, the purpose and stage of use, data involved, consequential human–AI interactions, the use or rejection of outputs and the human evaluation applied to them. Documentation should be proportional to influence: minor linguistic or administrative assistance requires less detail than AI-assisted coding, classification, integration or interpretation. This distinction goes beyond disclosure. Saying that AI was used does not establish how it shaped evidence selection, transformation, comparison, findings or the final meta-inference. Nor can a model’s fluent synthesis substitute for methodological fit, because generated explanations may appear coherent while remaining insufficiently grounded in the data. The article accordingly warns against treating AI-generated coherence as methodological coherence.

Methodological reviews cited in the analysis indicate why transparency cannot be assumed. In a systematic methodological review of 57 empirical mixed methods studies in educational psychology, 35.1% provided no identifiable rationale for using mixed methods, while 31.6% neither discussed mixed methods nor referenced the mixed methods literature. These observations concern reporting rather than proving that the underlying studies were incoherent, but they show why readers cannot fully evaluate relationships that remain undocumented. Another review cited in the article found that 64% of mixed methods studies in leading educational technology journals did not identify themselves as mixed methods, while 68% of those that did failed to specify a core mixed methods design. Against this background, the authors position the MMCTF as a potential form of methodological resistance to “performative integration,” in which methods, joint displays or AI tools are added because they signal sophistication rather than because the research problem requires them. Complexity itself is not evidence of quality: a complex design can be incoherent, while a simple two-phase design can be coherent.

The framework’s immediate practical implication is to reverse technology-first reasoning. A defensible sequence proceeds from the research problem to its epistemic and theoretical rationale, then to the purpose and design of mixed methods inquiry, the need for integration, the analytical procedure and only afterward the choice of technology. Its domains are intended as critical prompts, not a compliance checklist or additive score, and the appropriate degree of philosophical explanation depends on the study’s complexity and claims. Future work could apply the framework retrospectively to studies across psychological subfields and methodological traditions, use it prospectively during design, and test whether independent reviewers recognize the same coherence problems and productive tensions. The authors also identify AI-mediated integration validity as an open agenda: future studies must determine when AI-assisted connection, transformation or synthesis adds defensible value, how such systems affect negative cases and discordance, and what documentation permits evaluation without recording every prompt or computational operation. Human researchers, the analysis concludes, retain responsibility for the knowledge claims their systems ultimately advance.

Subject of Research: A conceptual framework for evaluating coherence and transparency in mixed methods psychological research using generative AI.

Article Title: Beyond integration: a coherence and transparency framework for mixed methods research in the era of generative AI

Article References: Krumsvik, R. J., & Johansen, M. O. (2026). Beyond integration: a coherence and transparency framework for mixed methods research in the era of generative AI. Frontiers in Psychology, 17, Article 1971387. https://doi.org/10.3389/fpsyg.2026.1971387

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1971387

Keywords: mixed methods research, methodological coherence, research transparency, generative AI, psychological science, meta-inference, methodological pluralism, AI-assisted analysis, integration quality, Beyond, integration, coherence