Directed Content Analysis: coding with theory up front

The method uses theory twice. As instrument: concepts become operational codes with definitions and inclusion rules, giving the analysis a running start no inductive study has. As defendant: the coded corpus becomes a structured test of the theory’s reach, because the procedure requires tracking everything the categories could not accommodate.

That second use is what separates directed content analysis from box-ticking. The uncodable residue — passages relevant to the phenomenon but homeless in the framework — is not noise; it is the study’s most valuable output, the raw material for extending or challenging the theory.

The stance also carries a known bias to manage: analysts primed with categories find them. The countermeasures are procedural — inclusion rules tight enough to be falsifiable per passage, a mandatory residue category, and coding passes that ask “what does not fit?” as seriously as “what fits where?”.

The naming source is Hsieh and Shannon’s Three approaches to qualitative content analysis, which distinguishes conventional (inductive), directed (deductive, theory-driven), and summative approaches, and specifies the directed procedure: derive categories from theory, code, and treat non-fitting data as candidates for new or revised categories.

For a stepwise elaboration researchers can follow like a protocol — sixteen steps across preparation, organisation, and reporting — Assarroudi and colleagues’ Directed qualitative content analysis is the most operational treatment available.

The broader content-analysis quality vocabulary (credibility, dependability, and the manifest/latent distinction) comes from the same tradition as the conventional approach; the two branches share standards and differ only in where categories originate.

• When a credible framework exists and the question is its application: does this adoption model describe how these clinics actually adopted the tool? Do these complaint narratives fit the established taxonomy?
• When comparability matters across studies or sites — shared frameworks make findings cumulative in a way fresh inductive categories never are.
• When time is short and the phenomenon is not virgin territory. A validated codebook is months of category development you do not repeat.
• Not when the field lacks usable theory — forcing a weak framework onto rich data produces neither test nor discovery.
• Not when the goal is to characterise the phenomenon afresh; that is conventional content analysis or reflexive thematic work.

A service research team coded 160 crisis-chat transcripts against an established five-stage help-seeking model (problem recognition, decision to seek, source selection, disclosure, evaluation). The codebook operationalised each stage with inclusion rules; the pilot on 15 transcripts forced one rule revision and taught the coders that stage order could not be assumed.

Full coding found the model’s stages present but its architecture wrong for this medium: 61% of chats opened mid-disclosure, with problem recognition performed retrospectively — users narrating how they had come to realise the problem while already disclosing it. The residue category accumulated 214 passages, and its largest cluster was testing the service (probing the responder’s humanness, judgement, and speed before substantive disclosure) — behaviour the model had no stage for.

The published analysis reported prevalence per stage, the sequence finding, and a proposed model revision: a trust-testing stage preceding disclosure in anonymous digital channels, evidenced by the residue cluster. The framework left the study changed — which is directed content analysis succeeding, not failing.

• No residue category. Everything forced into the framework guarantees the theory survives — and voids the study’s evidential value.
• Codebook categories without rules. Concept names alone (“self-efficacy”) invite drift; rules make disagreement — and therefore reliability — possible.
• Confirmation reporting. Presenting only the passages that fit, with the framework re-described as findings.
• Frequency worship. Category counts without analysis of how the concept manifested, varied, and failed.
• Framework-shopping after the fact. Choosing the theory once the data are known converts a test into an illustration; the direction must be declared in advance.
• Mislabelled inductive work. Deriving categories from the data and citing the directed approach — the reverse of the usual mislabelling, and equally corrosive.

The method sees with the theory’s eyes: phenomena the framework does not name arrive only through the residue, and a thin residue discipline makes the analysis exactly as blind as its instrument.

Its tests are also asymmetric — presence of theorised categories is easy to show; absence is entangled with data type, instrument wording, and coder priming. Claims that the theory “failed” need the same care as claims that it held.

And directed analysis inherits its framework’s politics: coding service-user talk with an institutional taxonomy reproduces the institution’s way of seeing. Sometimes that is the point; it should always be a choice made knowingly.

Conventional qualitative content analysis is the inductive sibling — categories from the data — for phenomena without usable theory.

Summative content analysis starts from word-level counting and moves to contextual interpretation.

Framework analysis offers a matrix architecture that pairs naturally with directed codebooks in applied team research.

Straussian grounded theory is the right label when the goal is generating theory rather than testing it.

Directed analysis is the qualitative method most directly accelerated by AI, because its heart is a codebook applied at scale. Evidano takes the operationalised categories and codes the full corpus against them — each application tied to its supporting quote — so reliability checking, prevalence analysis, and residue review start from a complete, auditable first pass; published comparisons of codebook-driven coding of this kind report 92–96% agreement with expert human coders.

The framework’s operationalisation, the residue’s interpretation, and the verdict on the theory are the researcher’s work. The tool makes the test cheap to run; what the test means is not its call.