Latent Coding: interpreting the meaning beneath the words

Framings and metaphors: the recurring figurative structures through which a topic is thought — illness as battle, customers as traffic, time as debt.

Presuppositions: what utterances take for granted — “even the night staff were consulted” presupposes a hierarchy of consultation-worthiness nobody asserted.

Absences and euphemisms: what is systematically unsaid, renamed, or passed over, detectable only against the corpus’s pattern.

Subject positions: who speakers cast themselves and others as — expert, victim, realist — beneath the propositional content.

In Braun and Clarke’s framework this is the latent level of theme development; in content analysis, Graneheim and Lundman’s “latent content”. Both traditions insist on the same discipline: the underlying meaning must be argued from textual evidence, not intuited past it.

Braun and Clarke’s Using thematic analysis in psychology distinguishes semantic from latent themes and ties the latent level to constructionist framings — the analysis examines the ideas and assumptions that shape the semantic content.

Graneheim and Lundman’s Qualitative content analysis in nursing research frames latent content as the interpreted, underlying meaning of text, and supplies the quality vocabulary — credibility, dependability, transferability — that latent claims are judged by.

For the interpretive craft itself, the discourse-analytic tradition is the deep well: latent coding borrows its attention to what language does while remaining within a thematic architecture.

• When the question concerns assumptions rather than reports: how do clinicians conceptualise “difficult” patients? what does “merit” mean inside this hiring process?
• When the surface is known to be curated — stigmatised topics, institutional talk, anywhere impression management is structural.
• When semantic results keep converging on something unexplained: surface themes that co-occur oddly are often one latent structure wearing several outfits.
• Not for questions stakeholders must audit line-by-line — latent findings are arguable interpretations, and a programme board wanting verifiable barriers should get semantic analysis.
• Not without interpretive resources. Latent coding needs a theoretical sensibility to interpret with; run bare, it produces free association.

An organisation asked why burnout persisted despite a celebrated wellbeing programme. Semantic analysis of 300 staff comments had already listed workload, staffing, and “resilience training fatigue”; the latent pass asked what conceptions of responsibility structured the talk.

Trace coding found three recurrent patterns: workload appeared overwhelmingly in agentless grammar (“the demands have grown” — no one demanding); wellbeing appeared as personal inventory (“my resilience”, “top up my reserves”), a possession employees carry rather than a condition work produces; and management appeared as weather — anticipated, endured, never addressed as decision-makers (“when the restructures blow through”). Counter-evidence was hunted and found bounded: union representatives’ comments broke all three patterns, naming actors and decisions — the exception that displayed the rule’s reach.

The latent theme — responsibility privatised: the corpus conceptualises wellbeing as an individual asset depleted by an agentless environment — was reported with its lens (a governmentality reading), its trace distributions, and the union counter-case. The practical upshot differed sharply from the semantic report’s: not more resilience sessions, but the finding that the programme’s own language taught staff to blame their reserves. The organisation rewrote its wellbeing communications; that is what latent analysis is for.

• Interpretation without traces. Themes announced from analytic authority, with quotes as decoration rather than evidence — the failure that discredits the level.
• Semantic themes with latent titles. Renaming “staff report high workload” as “discourses of overwork” changes the register, not the analysis.
• One lens applied as truth. Reporting the reading as what the data “really mean”, rather than as an interpretation from a named position.
• Ignoring reflexivity. The analyst’s own categories are part of the instrument at this level; a latent analysis with no reflexive account is missing a methods component.
• Depth as a status move. Choosing latent because it sounds more sophisticated, on a question the surface would have answered — the commonest and least admitted error.

Latent findings are constitutively contestable: different lenses yield different readings of the same corpus, and the method’s integrity lies in making the contest possible, not in winning it by fiat.

Reliability tools lose purchase — inter-coder agreement on interpretations of the unsaid is neither achievable nor, in the reflexive traditions, sought. Quality shifts to credibility: trace evidence, counter-evidence handling, and the coherence of the interpretive chain.

And the level tempts overreach into minds and motives. Latent coding licenses claims about meaning structures in the data; it does not license psychoanalysis of participants, and write-ups that slide from “the talk assumes” to “they secretly believe” have left the method.

Semantic coding is the surface-level counterpart — often the right first pass whose organised results the latent pass then interrogates.

Reflexive thematic analysis houses both levels and supplies the reflexive obligations latent work depends on.

Discourse analysis offers the fullest toolkit for language-level traces; latent thematic work borrows from it constantly.

Critical discourse analysis commits to a normative standpoint where latent coding may remain analytic.

The trace layer is where tooling contributes: finding every agentless construction around “workload”, every possessive framing of wellbeing, across 300 comments is retrieval-and-coding work, and Evidano handles it with each coded trace linked to its passage — so the distribution a latent theme rests on is demonstrable, and the counter-evidence hunt is a search rather than a hope. Semantic first passes to organise the corpus come cheap the same way.

The interpretation is not delegable, and this is a design fact rather than a disclaimer: a model can surface candidate patterns in the said; deciding what unsaid structure they evidence — through which lens, against which counter-readings — is the analyst’s argument to make and defend.