Section 1 of 5

What Content Analysis Is

History, manifest vs. latent content, and codes as variables.

Origins

Lasswell and the propaganda studies

Harold Lasswell's wartime work at the Library of Congress operationalised content analysis: counting symbols, tracing their distribution, and drawing inferences about communicative intent.

His framing question still organises the field:

Who says what, to whom, in which channel, with what effect?

Codification to synthesis

Berelson (1952) to Krippendorff (2018)

Berelson, 1952

First textbook definition: objective, systematic, quantitative description of manifest content. Deliberately restricted to the literal surface.

Krippendorff, 1980–2018

Expanded to latent meaning; formalised reliability requirements; developed the alpha statistic. Now the canonical reference.

Bernard, Wutich, and Ryan (2017) position content analysis as part of a continuum of text-analytic methods, which is the framing this course uses.

The key distinction

Manifest vs. latent content

Contemporary content analysis blends both. The analyst declares which is which and demonstrates reliability for latent codes.

The defining move

Codes as variables

Thematic analysis

A code is a label attached to a passage. Its job is to organise interpretation for one study.

Content analysis

A code is a variable that takes a value (0/1, or a count) for every recording unit. It is a column in a data matrix.

Once codes are variables, frequency tables, cross-tabulations, and chi-squared tests follow naturally. Both the qualitative and quantitative halves are load-bearing.

Carry forward

When content analysis is the right choice

• Your question asks how often, or whether subgroups differ in what they say.
• The codes-as-variables move is what separates content analysis from thematic work.
• Both qualitative and quantitative halves are necessary; neither is decoration.

A later section covers the design decisions that make a content analysis defensible.

Introduction and Overview

Content analysis is the oldest systematic method for analysing text in the social sciences, and it is also the bridge between the qualitative and quantitative traditions of this course. Where the thematic analysis you learned earlier in the course stops once you have a defensible set of themes, content analysis keeps going: it counts the themes, distributes the counts across subgroups, and tests whether the differences are larger than chance. The method is qualitative in its first move (identifying what the relevant categories are) and quantitative in its second (counting their occurrence). It is the technique you reach for when your research question asks both what kinds of loneliness do people describe and how often, and does that distribution differ by age, gender, or caregiver status.

This section traces how content analysis became the workhorse it is today. We start with Harold Lasswell's WWII analysis of Axis propaganda, watch Bernard Berelson codify the method in 1952, follow Klaus Krippendorff's modernisation into the canonical 2018 monograph, and end with the contemporary synthesis offered by Bernard, Wutich, and Ryan (2017, Ch. 11). Along the way we will pull apart the distinction between manifest and latent content, the line that separates "what the text literally says" from "what it means", and we will explain the single operational move that turns content analysis into the hybrid method it is: treating codes as variables. By the end of the section you will be able to say what content analysis is and why it is methodologically different from the thematic work you have already done.

1.1 A Short History: From Lasswell to Krippendorff

The modern history of content analysis begins with Harold D. Lasswell, the political scientist who in 1927 published Propaganda Technique in the World War, the first systematic effort to study communication content as a window onto political intent. Lasswell's question was not "what does this propaganda mean to its readers?" but "what categories of appeal does it use, and in what proportion?" The methodological commitment was to systematic counting of explicit features, mentions of leaders, mentions of enemies, deployment of symbols, rather than impressionistic close reading. The pay-off was that two analysts working independently could produce comparable numbers.

During the Second World War, Lasswell led the Experimental Division for the Study of War-Time Communications at the U.S. Library of Congress. The Division analysed German, Italian, and Japanese propaganda in volume that no individual reader could have made sense of impressionistically. By systematic coding of explicit features, the frequency with which Axis broadcasts mentioned specific Allied generals, the proportion of broadcast time devoted to economic versus military themes, the rise and fall of named enemies week by week, the Division produced intelligence assessments that helped guide both counter-propaganda efforts and broader policy decisions. The work was content analysis as inference about a producer: from the text, back to the propagandist's strategic state of mind. The method got its first major public-policy validation in this period (Lasswell, Lerner, & Pool, 1952).

The post-war codification belongs to Bernard Berelson. His Content Analysis in Communication Research (1952) gave the field its first textbook and its most-cited definition: content analysis is "a research technique for the objective, systematic, and quantitative description of the manifest content of communication" (Berelson, 1952, p. 18). Each of the four adjectives in that definition matters. Objective meant that the analysis was repeatable by other analysts. Systematic meant that the same procedure was applied across the entire sample. Quantitative meant counting, not impression. Manifest meant the literal surface of the text, what was said, rather than what the analyst imagined the author meant. Berelson's stance was, in effect, the high-modernist version of content analysis: positivist, quantitative, and resolutely uninterested in latent meaning.

The next major recodification came from Klaus Krippendorff, whose Content Analysis: An Introduction to Its Methodology (1980, 2004, 2013, and now in its 4th edition, 2018) brought the method into the contemporary social-science mainstream. Krippendorff's most important moves were two. First, he redefined content analysis as "a research technique for making replicable and valid inferences from texts (or other meaningful matter) to the contexts of their use" (Krippendorff, 2018, p. 24), broadening Berelson's "objective description" to "inference about context," which made room for latent content. Second, he gave the field its statistical centre of gravity by developing Krippendorff's alpha, the reliability statistic that has since become the methodological standard for content analysis (we will use it in a later section). Krippendorff's text remains the field's most-cited methodological reference and the standard against which contemporary content-analytic studies are judged.

The synthesis you are reading in this course comes from Bernard, Wutich, and Ryan (2017), who present content analysis as part of a continuum of text-analytic methods rather than as a standalone enterprise. Their stance is that the strict separation Berelson drew between qualitative and quantitative analysis was always more rhetorical than real, and that contemporary content analysis is best understood as a method that blends the two: qualitative judgement determines the coding scheme; quantitative procedures determine how the codes are distributed and whether the differences are reliable. This is the position that Hsieh and Shannon (2005) articulated influentially in the health-research literature when they distinguished conventional, directed, and summative content analysis, a typology that has since structured most qualitative-content-analytic work in nursing, health services research, and public health.

1.2 Manifest Versus Latent Content

A frequently asked methodological question about content analysis is: are you coding what the text literally says, or what it means? Bernard, Wutich, and Ryan's answer, and the answer of the modern field, is that the choice is yours, but it must be explicit.

Manifest content is the literal, surface-level content of the text. If your code is "mentions of pets," a manifest coder counts every appearance of the words pet, dog, cat, Rufus, Marie's cat, parrot, and so on. The advantage is that two manifest coders working from a clear word-list will produce nearly identical counts; the disadvantage is that manifest content misses everything the participant is doing with the text other than literal naming.

Latent content is the underlying meaning that requires interpretation to surface. If your code is "loneliness framed as the cost of having loved," a latent coder reads a passage and decides whether the participant is articulating the idea, regardless of whether the specific words "cost" or "love" appear. Linda's account of Bill's empty chair (P05) is a latent expression of "loneliness-as-residue-of-marriage", she does not use that phrase, but the meaning is unmistakable to a competent reader. Latent coding captures more of the meaning but is harder to do reliably.

Berelson's classical definition (1952) restricted content analysis to manifest content. Krippendorff and the subsequent generation rejected the restriction, on the grounds that inference about context, the contemporary purpose of content analysis, almost always requires reading for latent meaning. The modern compromise is that content analysis routinely codes both, but the analyst must declare which is which and must demonstrate that latent codes can be applied reliably.

1.3 Codes as Variables: The Operational Move

Here is the single move that makes content analysis what it is. In thematic analysis (an earlier module), a code is a label you attach to a passage; the code's job is to organise interpretation. In content analysis, the same code becomes a variable: a column in a data frame, with a value for every unit in the corpus. Once codes are variables, the entire apparatus of descriptive and inferential statistics is available.

The reframing is subtle but transformative. Consider the code loneliness-as-residue-of-marriage. In thematic analysis, the code is a finding: "this is one of the kinds of loneliness participants describe." In content analysis, the code becomes a variable that takes the value 1 for every transcript in which the theme appears and 0 elsewhere. Now you can ask: in what proportion of transcripts does the theme appear? Does that proportion differ between widowed and non-widowed participants? Is the difference larger than chance? Does the proportion grow over the course of the interview, that is, do participants need to talk for a while before they articulate the theme? Each of these is a statistical question that the codes-as-variables move makes available.

Bernard, Wutich, and Ryan are explicit that this move is what places content analysis in the hybrid position it occupies. It is not "qualitative analysis with numbers attached" (which is what Berelson thought it was), nor "a quantitative method applied to text" (which is what nineteenth-century philology was). It is a method in which qualitative judgement is required to define the variables, and quantitative procedures are used to analyse them. Both halves are necessary; neither is optional.

1.4 When Content Analysis Is the Right Tool

Not every qualitative research question calls for content analysis. The method's distinctive payoff is in questions that involve distributional comparison: how often something appears, whether subgroups differ in their use of a code, whether something changes over time. Where the question is genuinely interpretive, what does this single passage mean, or how does this participant make sense of their experience, thematic analysis (an earlier module), schema analysis (a later module), or narrative analysis (also a later module) will give you more analytic traction than content analysis will.

Concretely, you should reach for content analysis when one or more of the following are true:

• You have a comparison question. Caregivers versus non-caregivers; younger versus older; immigrant versus native-born; pre-pandemic versus post-pandemic. Content analysis is built for these.
• Your corpus is large enough that close reading alone would be impractical. Twenty transcripts is on the lower end, you can certainly do content analysis on a corpus this size, and this lesson does, but for corpora of 50, 200, or 5,000 documents, content analysis is one of the few options that scales.
• You need replicable counts that other researchers can audit. Where policymakers or journal reviewers will demand "how many" rather than "how rich," content analysis is the methodological answer.
• Your research question is longitudinal. Trends, shifts, before/after comparisons, content analysis is the standard tool for them in the qualitative literature.
• You are writing for a quantitative audience. Public-health journals routinely ask for the kind of distributional evidence that content analysis produces. Where thematic analysis would be rejected as too impressionistic, content analysis can be persuasive.

Introduction and Overview

An earlier section framed content analysis methodologically. This section is about the design decisions that determine whether your content analysis will be defensible. There are three of them, in order: what counts as the text to be analysed (sampling), what counts as the unit being coded (recording units and context units), and how the codes are applied with sufficient consistency that the resulting counts mean something (reliability). Get any one of these wrong and the analysis is suspect; get all three right and you have a piece of content-analytic work that a sceptical quantitative reviewer will accept.

Bernard, Wutich, and Ryan are unusually explicit on these matters because content analysis, more than any other qualitative method, can be done badly in ways that produce numbers that look authoritative but are not. The history of the field is littered with frequency tables built on inconsistent coding of poorly specified units sampled from non-representative corpora. The remedy is procedural: you commit, in writing, to a sampling rule, a unit-definition rule, a coding scheme, and a reliability target, before you begin coding. You then execute the procedure transparently, report what happened, and let the reader audit it.

2.1 Sampling Text: The Three Kinds of Units

Krippendorff's most enduring methodological contribution after the alpha statistic is the distinction between three levels of unit in any content-analytic design:

• Sampling units, the chunks of text that are drawn from the population. In a corpus of 20 loneliness transcripts, each transcript is a sampling unit. In a corpus of newspaper coverage of the overdose crisis, each article is a sampling unit. In a Twitter dataset, each tweet is a sampling unit.
• Recording units, the unit you actually code. The recording unit is where the variable takes its value. Recording units may be the same as sampling units (you code each transcript as a whole) or smaller (you code each sentence, each paragraph, each turn-at-talk, or each named theme-bearing passage).
• Context units, the surrounding text the coder is permitted to consult when deciding how to code a recording unit. If your recording unit is a sentence and your context unit is the paragraph, the coder reads the sentence in the context of the paragraph but assigns the code to the sentence alone.

The trio matters because the choice of each one shapes both what the analysis can show and how reliable it will be. A study that uses the whole transcript as the recording unit produces a frequency-per-transcript table; a study that uses the sentence as the recording unit produces a frequency-per-sentence table. These two analyses can give different answers to the same surface question, and the difference is methodological, not substantive.

2.2 Developing a Coding Scheme

The coding scheme, the codebook, in the language of an earlier module, is the spine of any content analysis. In an earlier module you built one for thematic analysis; here you adapt it for the variable-treatment of content analysis. Three properties matter more in content analysis than they did in thematic analysis.

Exhaustive coverage. Your coding scheme should cover the content you care about. For content analysis specifically, exhaustive coverage means that every recording unit can be assigned at least one code (or the explicit "not-coded" residual category). The classical Berelson position is that codes should be mutually exclusive, each unit gets exactly one code, on the grounds that overlapping codes make the resulting frequencies hard to interpret. Krippendorff and the contemporary field reject the mutual-exclusivity requirement: a passage can simultaneously instantiate "loneliness-as-existential-fact" and "loneliness-coped-with-by-pet-companionship" and there is no good reason to force the coder to choose. The modern compromise is multi-coded passages are permitted if the codebook says so and the multi-coding is consistent.

Clear operational definitions. Each code in your scheme needs a definition that another coder could apply without consulting you. The definition has three parts: a brief substantive description, an inclusion rule (what counts), and an exclusion rule (what doesn't count). Inclusion and exclusion rules are where reliability is won or lost. "Mentions of loneliness" is a definition without inclusion/exclusion rules. "Any statement in which the participant attributes loneliness to a specific event or relationship (inclusion); excluding generic statements about loneliness in the population or society at large (exclusion)" is a definition that can be applied reliably.

A priori versus emergent. Content-analytic codebooks come from one of two directions, or (usually) both. A priori codes come from the literature, your conceptual framework, or your interview guide; you decide before you read the data that you will code for, say, "stigma," "social-comparison," and "coping-by-substance-use" because the literature on loneliness says these are the right categories. Emergent codes come from the data, you read the corpus, notice what is there, and let the codebook grow to fit. Hsieh and Shannon (2005) call the first directed content analysis and the second conventional content analysis; their summative third type starts from word counts and works upward. Elo and Kyngäs (2008) and Mayring (2000) offer complementary process descriptions widely cited in nursing and European health-research traditions.

2.3 Inter-coder Reliability: Stricter Than Thematic Analysis

Reliability is the test of whether two coders, working independently with the same codebook, would assign the same codes to the same passages. In thematic analysis (an earlier module), reliability matters but it is one consideration among several; in content analysis, reliability is load-bearing, the frequencies you report are only as good as the consistency with which the codes were applied. If two coders applying the same codebook produce different counts, the counts are noise.

The field's standard reliability statistic is Krippendorff's alpha (Krippendorff, 2004, 2018; Hayes & Krippendorff, 2007). Alpha has three properties that recommend it over the older Cohen's kappa: it accommodates any number of coders (where Cohen's kappa handles two); it accommodates any level of measurement (nominal, ordinal, interval, ratio); and it handles missing data gracefully. The arithmetic compares the observed disagreement among coders to the disagreement expected by chance, and produces a coefficient that runs from 1 (perfect agreement) through 0 (chance) to negative values (worse than chance). Put simply, alpha asks how much better than random guessing the coders did: when they never disagree the observed disagreement is zero and alpha is 1, and when they agree only as often as chance would predict the two disagreements are equal and alpha is 0.

The thresholds are convention, not law, and they vary slightly between sources. The 0.80/0.67 dichotomy comes from Krippendorff's own writing and is the most commonly cited in the field. Some health-research applications adopt a stricter 0.70 floor and a 0.80 publication target. Whatever you adopt, declare it in your methods section before you compute it.

Reliability is computed on a subset of the corpus, typically 10–20%, that two coders code independently. The remaining 80–90% is then coded by one coder alone, with the assurance that the reliability of the system is documented. The reliability subset is usually drawn randomly from the corpus to ensure that the reliability estimate is generalisable.

2.4 The Workflow End-to-End

A complete content-analytic study, in Bernard, Wutich, and Ryan's framing, follows these steps in order:

1. Formulate the research question, explicitly distributional or comparative.
2. Define the population of texts, what corpus is the analysis drawn from?
3. Sample the texts, if the population is small enough, the entire population may be sampled; otherwise, define a sampling frame and a sampling procedure.
4. Decide on sampling, recording, and context units, declare them in writing.
5. Develop the codebook, a priori, emergent, or hybrid; with operational definitions for every code.
6. Train coders, on a practice subset that does not enter the final analysis.
7. Compute reliability on a 10–20% subset, Krippendorff's alpha; revise codebook if needed.
8. Apply the final codebook across the corpus, one coder for the remaining 80–90% is typical for a small-corpus study.
9. Analyse the resulting frequency table, descriptive statistics, then inferential tests (a later section).
10. Report transparently, corpus, sampling, units, codebook (in an appendix), reliability, analysis, interpretation.

Introduction and Overview

Once your coding is done and your reliability is acceptable, you have a coded matrix: rows are recording units (transcripts, paragraphs, or sentences), columns are codes, and the cell values are either binary indicators (1 if the code applied, 0 otherwise) or counts (the number of times the code applied to that unit). This matrix is the input to the inferential half of content analysis. This section walks through what you can do with it: descriptive frequency tables, cross-tabulation by subgroup, chi-squared tests for distributional differences, and trend analyses where the corpus is longitudinal. It then introduces dictionary-based content analysis, LIWC and sentiment dictionaries, and previews the computational content analysis that a later module will develop in depth.

The intellectual point of this section is that content analysis is a hypothesis-testing method when it wants to be. Whatever you tested in your earlier regression work, you can test on content-analytic codes: differences in proportions, associations, trends, interactions. The variables are codes rather than survey items, but the inferential machinery is the same. This is what Bernard, Wutich, and Ryan mean when they say content analysis is the bridge between the qualitative and quantitative traditions.

3.1 Frequency Tables: The Starting Point

Every content-analytic study starts with descriptive frequencies. For the loneliness corpus, a typical first-pass frequency table looks like the one below: each of the eleven codes in the codebook supplied with the dataset, the count of transcripts in which the code appears (out of 20), and the percentage. These are the numbers the supplied export produces; a different codebook would produce different ones.

Already this table is doing analytic work. The most prevalent code, loneliness-as-embodied-ache, appears in 18 of 20 transcripts, suggesting that loneliness is described in bodily terms (ache, weight, exhaustion, sleeplessness) by almost everyone, regardless of who the participant is. The least prevalent, loneliness-as-untranslatable-word, appears in 3 transcripts, and you can already predict who those three participants are: the ones who came to Canada from elsewhere (Aarav, Amira, Chen). The frequency table tells you both what is shared across the corpus and what is concentrated.

3.2 Cross-Tabulation by Subgroup: The Comparison Move

Content analysis's real value emerges when you cross-tabulate the codes by a participant-level variable: age band, gender, caregiving status, immigration status, life-stage. Here is the same code distribution disaggregated by caregiver status in our worked corpus (5 caregivers, 15 non-caregivers):

Now the analytic action begins. The loneliness-as-sole-responsibility code appears in 4 of the 5 caregivers and in none of the 15 non-caregivers, a categorical difference that is the kind of finding content analysis is built to produce. The loneliness-as-existential-fact code shows the reverse pattern: 10 of 15 non-caregivers (the older, more settled participants like Helen and Frank) against 1 of 5 caregivers, whose loneliness is tied to a situation they expect to end. The loneliness-and-technology code is close to invariant by subgroup (80% against 73%), supporting the earlier hypothesis that this is a near-universal feature of contemporary loneliness regardless of life-stage. Note also the code that does not move: loneliness-hidden-by-shame sits at 60% in both groups, so the intuition that caregivers are the ones who hide their loneliness is not supported here. These patterns are descriptive; the next move is to ask whether they are larger than chance.

3.3 Chi-Squared Tests on Codes × Subgroup

The classical inferential test for a 2 × 2 (or larger) contingency table of categorical data is the chi-squared test of independence. Applied to a content-analytic frequency table, the chi-squared test asks: is the distribution of this code statistically different between the subgroups? The mechanics are familiar from earlier courses; here we apply them to qualitative codes treated as variables.

Take the loneliness-as-sole-responsibility code from the table above: 4 of 5 caregivers vs. 0 of 15 non-caregivers. The chi-squared test on this 2 × 2 table gives χ² = 10.42 with 1 degree of freedom, p = 0.001 (Fisher's exact p = 0.001; with cells this small Fisher is the number to report). The difference is statistically significant at conventional thresholds. Practically, and this is the part the qualitative analyst contributes, the analysis suggests that caregiver participants in this corpus describe a loneliness of being the only adult responsible for a dependent person, in a way that non-caregivers never do. The substantive interpretation might be that caregiving concentrates decision-making in one person, and that the absence of anyone to share the load is itself the lonely part. The chi-squared test is the warrant for the claim; the qualitative reading is the claim.

3.4 Trend Analysis: Codes Over Time

Content analysis is at its strongest in longitudinal applications. Where your corpus has a time dimension, coverage of a topic across years of newspaper articles, posts on a forum across the pandemic, or transcripts collected at multiple time points, the codes can be plotted over time and tested for trend. The classical example is Pool's (1952) analysis of editorial coverage of the Soviet Union across decades; the contemporary example is the proliferation of computational content analyses of social-media discourse before, during, and after specific events (elections, crises, vaccine roll-outs).

The loneliness dataset is cross-sectional, so trend analysis is not the dominant move, but you may have a quasi-temporal variable. For example, you could ask whether participants who were interviewed earlier in the corpus (the first 10 transcripts) differ systematically from those interviewed later (the last 10) in any code, as a way of probing whether your codebook was over-fitted to early transcripts. Or you could examine whether participants who lived through the pandemic as adults (versus those who were teenagers at the time) describe loneliness differently. These are pseudo-trend analyses on a cross-sectional corpus, but the inferential logic is the same.

3.5 Dictionary-Based Content Analysis

So far we have assumed that humans are applying the codes. Dictionary-based content analysis is the variant in which a pre-specified word list (a "dictionary") is applied to the corpus by a computer, and the resulting counts are treated as content-analytic variables. The approach is best understood as the manifest-content extreme of content analysis: it counts what is explicitly there, fast and at scale, at the cost of any latent-meaning sensitivity.

The best-known dictionary in the social sciences is the Linguistic Inquiry and Word Count (LIWC) dictionary, developed by James Pennebaker and colleagues from the early 1990s (Pennebaker, Boyd, Jordan, & Blackburn, 2015; Boyd, Ashokkumar, Seraj, & Pennebaker, 2022). LIWC is a curated set of about 90 categories, positive emotion, negative emotion, anxiety, sadness, body, health, family, social processes, cognitive processes, and so on, each defined by a list of words and word stems. Run LIWC over a transcript and it returns the percentage of words in each category. A transcript that scores high on "sadness" contains a high proportion of words like sad, lonely, grief, cry, miss, lost; one that scores high on "social processes" contains a high proportion of pronouns referring to other people.

LIWC has been used extensively in health-related text analysis: predicting depression from writing samples (Pennebaker, Mehl, & Niederhoffer, 2003), characterising trauma narratives, distinguishing the writing of patients on different medications, predicting suicide risk from social-media posts. It is the methodological ancestor of contemporary sentiment-analysis systems and remains in active use; LIWC-22 is the current version.

Sentiment dictionaries are the lighter-weight cousin of LIWC. The best-known are Bing Liu's Bing lexicon (about 6,800 words classified as positive or negative), the AFINN lexicon (a scored version, with words assigned valence from -5 to +5), and the NRC Word-Emotion Association Lexicon (which classifies words into Plutchik's eight basic emotions plus positive/negative). All three are freely available word lists that text-analysis software can apply to a transcript corpus automatically.

3.6 Computational Content Analysis: A Preview of a later module

Computational content analysis is the scalable extension of the dictionary approach, plus the addition of unsupervised methods that discover categories rather than counting pre-specified ones. The three families you will meet later in the course are: topic models (Latent Dirichlet Allocation and its successors, which discover thematic structure in large corpora without a codebook); word-embedding methods (which represent words as vectors in a high-dimensional space, enabling semantic similarity and analogy operations); and large-language-model-based coding (GPT-4-class systems applied as zero-shot or few-shot coders).

The methodological status of computational content analysis is still being worked out. The advantages are obvious: speed, scale, replicability of the computational pipeline. The disadvantages are real: opacity (LDA and LLMs cannot show their work the way a human coder's audit trail can), brittleness (a topic-model solution is sensitive to hyperparameter choices in ways the analyst rarely audits), and the recurring failure to validate computational outputs against human-coded gold standards. Bernard, Wutich, and Ryan's stance, and the stance of this course, is that computational content analysis is a useful complement to human content analysis, not a replacement for it. A later module will give you the operational machinery; this module is establishing the framework that machinery extends.

Introduction and Overview

Earlier sections framed content analysis methodologically. This section turns operational. You will see, end to end, how to take a Taguette codebook, apply it across the 20 loneliness transcripts, transform the resulting export into a codes × cases matrix, compute frequency tables and visualisations, run chi-squared and Fisher's exact tests on codes × subgroup, and conduct a keyness analysis that surfaces which words differ most between two subcorpora.

The workflow that follows starts from an exported Taguette CSV with one row per (passage, code) pair, columns for the document filename, the tag (code), and the passage content. A first coding pass usually covers 3–5 transcripts; a content analysis extends the codebook to 8–12 transcripts or more. The volume jump is intentional: you need enough cases for the frequency tables and chi-squared tests to be analytically informative, and 8–12 transcripts is the floor for a defensible content analysis of this dataset.

4.1 Step 1: Load the Taguette Export and Add Participant Variables

Your Taguette export is a CSV with one row per highlighted passage. Each row contains the document filename (which encodes the participant ID), the tag (your code), and the passage content. The first step is to open it as a table (a spreadsheet works well) and add the participant-level variables (age, gender, caregiver status, etc.) needed for subgroup comparison.

The course data bundle, HSCI_841_loneliness_data.zip, contains a worked Taguette export (taguette_export_week8.csv), the codebook with its 11 codes (codebook_week8.csv), and the participant table (participant_metadata.csv). The participant identifier is the part of each document filename before the pseudonym (for example, P01 in P01_Maya.txt), and it is the key that links each coded passage to its row in the participant table. After the join, each row is one highlighted passage, with columns for the code (tag), the passage text (content), and the participant-level variables (age, gender, life stage, caregiver status, immigration status) needed for subgroup analysis. Before going further, confirm that every participant in the export has been matched to a row in the participant table.

4.2 Step 2: Reshape Into a Codes × Cases Matrix

The codes-as-variables move is now a literal data-shape operation: convert the long-format export into a wide-format matrix where rows are participants and columns are codes. The cell values are typically binary (1 if the code applied to that participant at all, 0 otherwise) but can be counts (the number of times the code applied within the transcript).

A spreadsheet pivot table with participants as rows and codes as columns produces the count version directly; replacing every non-zero count with 1 gives the binary version. The participant attributes (age, gender, life stage, and so on) are then added as further columns. The result has one row per participant, 20 rows for the full corpus or 8–12 for a coded subset, and it is the analytic-ready frame for everything that follows.

4.3 Step 3: Frequency Tables

The overall and by-subgroup frequency tables in Sections 3.1 and 3.2 come from simple counts on the binary matrix. The overall frequency of a code is the number of participants with a 1 in that code’s column; the by-subgroup frequency is the same count taken separately within each subgroup and divided by the size of that subgroup.

4.4 Step 4: Visualise the Subgroup Comparison

A bar chart of code frequencies by subgroup is the standard published display for a content-analytic study, and any spreadsheet or statistics program can draw one.

The usual layout is a horizontal bar chart with the codes on the vertical axis, sorted so that the longest bars sit at the top, the percentage of each subgroup with the code present on the horizontal axis, and two bars per code, one for caregivers and one for non-caregivers. The chart shows percentages because the subgroups differ in size (5 and 15). In the worked corpus the bars pull apart most sharply for loneliness-as-sole-responsibility (80% against 0%) and loneliness-as-existential-fact (20% against 67%), and they sit level for loneliness-hidden-by-shame (60% in both groups). Codes for which the two bars differ markedly are the ones worth subsequent chi-squared testing.

4.5 Step 5: Chi-Squared and Fisher's Exact Tests

The inferential half of the analysis. For each code that showed a meaningful descriptive difference between subgroups in step 4, run the appropriate contingency-table test.

4.6 Step 6: Keyness Analysis

Keyness is the manifest-content cousin of the chi-squared analysis above. It asks: which words appear disproportionately in one subcorpus compared to another? The classical implementation is the log-likelihood ratio test on word frequencies between two subcorpora, and corpus-linguistics software computes it for every word in the vocabulary.

For the loneliness dataset, the natural keyness contrast is one of the subgroup contrasts you have already been working with: caregiver vs. non-caregiver, immigrant vs. non-immigrant, or under-40 vs. over-65. The output is a ranked list of words, each with a chi-squared (or log-likelihood) statistic and a sign indicating which subcorpus the word over-occurs in.

4.7 What a Content Analysis Reports

A content analysis of a coded subset ends in a frequency table and at least one defensible quantitative comparison. A complete report of it includes four artefacts: the coded transcripts (Taguette export), the frequency table (as a spreadsheet or CSV file), the chi-squared or Fisher's exact comparison with its odds ratio, and an interpretive memo of about 500 words that places the quantitative findings back into a qualitative reading.

Bringing It All Together

This lesson has done two related things. First, it has located content analysis historically, from Lasswell's WWII propaganda studies through Berelson's 1952 codification to Krippendorff's contemporary synthesis, and clarified what makes it methodologically distinctive. Content analysis is the qualitative-quantitative bridge: it begins with the same kind of qualitative judgement that drives thematic analysis (an earlier module) but turns the resulting codes into variables and applies the inferential machinery of earlier courses. The single operational move, treating codes as variables, is what lets a 20-transcript qualitative corpus answer questions that thematic analysis alone cannot: how often, how distributed, larger than chance.

Second, the lesson has given you the operational machinery to do content analysis on qualitative data such as the loneliness transcripts. A codebook is the input; the Taguette export is the bridge; the six-step workflow turns long-format coding data into wide-format codes × cases matrices, computes descriptive and inferential statistics, and produces publication-quality visualisations. The worked analysis, a coded subset, a frequency table, a chi-squared or Fisher's exact comparison, and an interpretive memo, is the course's first piece of hybrid quantitative-qualitative analysis.

What you take away from this lesson sets up the lessons that follow. Schema and narrative analysis moves in a different direction: not counting codes across many transcripts but tracing the cognitive structures inside individual transcripts. Discourse analysis deepens the attention to language and power. Analytic induction, QCA, and decision models return to systematic comparison with a different toolkit. Computational text and LLM analysis extends today's dictionary-based and keyness moves to the scale of millions of documents.

The final reflection below asks you to step out of method-mode and articulate where content analysis fits in your own methodological identity. There is no single right answer; the goal is to leave the lesson with a defensible stance on when this hybrid method belongs in your toolkit.

Final Knowledge Assessment

This glossary collects the key concepts, people, and methodological stances introduced in this lesson. Use it as a reference while you work through the material, or as a review before the final assessment. Type in the search box to filter entries.