Content analysis: a complete guide, with a worked example

Content analysis is a method for systematically categorising communication and describing what it contains. The material can be interview transcripts, open-text survey responses, policy documents, newspaper coverage, social media posts, clinical notes or broadcast media — anything recorded that can be read.

Its defining feature is systematic categorisation. Every unit of the material is examined against the same set of categories, applied by the same rules, so the description that results is a property of the material rather than of whoever happened to read it. That is why content analysis is often the method of choice where a finding has to be reproducible.

It has a longer history than most qualitative methods, originating in communication research in the first half of the twentieth century, and it retains the quantitative instincts of that origin: content analysis is comfortable counting, in a way that most interpretative methods are not.

Every content analysis sits somewhere on a spectrum between describing what is literally there and interpreting what it means.

Manifest content

The surface of the material: the words actually used, the topics explicitly raised, the categories a reader could identify without inference. Coding manifest content is closer to measurement — two competent coders should agree most of the time, and where they do not, the coding frame is usually at fault.

Example: coding whether each policy document mentions equality impact assessment. Either the phrase and its concept appear, or they do not.

Latent content

The underlying meaning: what is implied, what tone is being struck, what is conspicuously absent. Coding latent content requires interpretation, which means agreement between coders is harder to achieve and reliability statistics need to be read more carefully.

Example: coding whether each policy document treats equality as a compliance obligation or as a purpose. No single phrase settles it; the coder is making a judgement about the document as a whole.

This is the most common question we are asked about qualitative method selection, and the honest answer is short: does frequency answer your question?

A concrete test. If your finding would be reported as “equality was mentioned in 34 of 51 documents, most often as a compliance requirement”, you are doing content analysis. If it would be reported as “equality functioned as a language of reassurance rather than a commitment”, you are doing thematic analysis. Both are defensible; only one answers each question.

Hsieh and Shannon’s 2005 typology is the one most UK examiners expect to see cited, and naming which of the three you used is now close to mandatory in health and social research.

The distinction matters most for directed content analysis, because its characteristic weakness is confirmation. If your categories come from a framework, the analysis will find the framework. The safeguard is to report explicitly what did not fit — material that falls outside an a priori frame is frequently the most interesting result in the study, and the studies that discard it silently are the ones reviewers criticise.

Content analysis is often applied to material that already exists, which makes sampling a design decision rather than a recruitment problem. It is also the decision most often left undescribed.

Where the corpus is large, a random or systematic sample is defensible and should be described as such. Where it is small enough to analyse whole, say that you analysed the population rather than a sample — it is a stronger claim and it costs nothing to state.

The most common sampling weakness in published content analyses is convenience framed as completeness: analysing the documents that happened to be online, and reporting percentages as though they described all documents of that type. If your corpus is what you could find, say so.

The coding frame is the instrument. In content analysis it does the work that a questionnaire does in survey research, and it deserves the same care.

Where categories come from

Deductive frames are built from theory, prior research or a policy framework before the data is read. Inductive frames are built from a subset of the data and then applied to the rest. Most real studies are hybrid: a prior frame, revised after a pilot on 10–20% of the material.

What each category needs

Categories must also be exhaustive and, in most designs, mutually exclusive. If a unit can fall into two categories at once, either the categories overlap and should be merged, or the design should allow multiple codes per unit — which is fine, but it changes how you report frequencies and must be stated.

The unit of analysis

Decide before coding whether your unit is the word, the sentence, the paragraph, the speaking turn, or the whole document. This decision determines what your counts mean, and changing it midway invalidates everything coded before the change.

A study examined how 51 local authority strategy documents discussed digital exclusion. The unit of analysis was the paragraph; a paragraph could receive more than one code.

What the counts show

Access framing dominates: it appears in 86% of documents and accounts for more than half of all coded paragraphs. Design framing — the possibility that services themselves create exclusion — appears in fewer than one document in five. Only 10% of strategies contain a measurable target.

That pattern is the finding, and it is a finding frequency can carry. The contrast between how often the problem is described as the citizen’s (access, skills, choice: 479 paragraphs) and how often it is described as the service’s (design: 37 paragraphs) is a quantitative claim about a body of policy text, and it would be considerably weaker stated as an impression.

What the counts do not show

They do not show that access framing is wrong, or that authorities using design framing perform better. Content analysis describes the material; it does not evaluate it. Studies that slide from describing what documents say to judging what authorities do are making a claim their design does not support.

Because content analysis claims replicability, it usually has to demonstrate it. A subset of the material — commonly 10–20% — is coded independently by a second coder, and agreement is calculated.

Percentage agreement is the figure most often reported and the least informative. Two coders assigning a category that occurs 90% of the time will agree 90% of the time by chance alone. Chance-corrected statistics exist precisely for this, and reviewers know it.

Methods

Results

Report counts in a table rather than in prose — six categories described in sentences is unreadable, while the same six in a table is scannable. Give both the number of documents and the number of coded units where they differ, because a category appearing many times in one document tells a different story from one appearing once in many.

Illustrate categories with short extracts even in a quantitative content analysis. Counts establish the pattern; extracts establish what the category actually looks like in the material, and without them the reader has to take the frame on trust.