Cognitive-Affective Mapping (CAM) Overview
- CAM is a network representation where individuals map their thoughts and emotions with evaluative valence, capturing structured experiences.
- It employs a standardized visual grammar using distinct colors, shapes, and link styles to denote positive, negative, neutral, and ambivalent affect.
- CAM supports both qualitative case analysis and scalable quantitative network analysis, making it applicable in diverse research contexts.
Cognitive-Affective Mapping (CAM) denotes the use of cognitive-affective maps to represent how a person thinks and feels about a topic by depicting a network of concepts together with their affective valence. In the formulation introduced for direct empirical assessment, CAMs are not merely visualization devices but instruments for capturing individual experience, connotation, and belief structure, combining the openness of qualitative elicitation with the comparability of quantitative network data (Reuter et al., 2022). CAMs operationalize attitudes as structured systems of evaluatively loaded concepts: a topic is represented not as an isolated item or rating, but as an interconnected constellation of thoughts, events, emotions, and factual knowledge.
1. Origins and theoretical basis
CAMs originate in Paul Thagard’s work on mental networks. In Coherence in Thought and Action (2000), thought is treated as involving networks of connected representations that fit together coherently or incoherently via constraint satisfaction: positive constraints indicate that representations cohere or mutually support one another, whereas negative constraints indicate conflict or incompatibility. In Hot Thought (2006), this framework was extended by incorporating emotion through the HOTCO model of “hot cognition,” in which cognition and emotion are intertwined and network elements carry valence. CAMs, introduced in 2010 through the EMPATHICA system, synthesize these ideas into a visual format (Reuter et al., 2022).
The resulting theoretical commitment is that attitudes, beliefs, and experiences are not isolated items but interconnected constellations. A CAM therefore represents experiential structure as a network rather than a list. Concepts in the map can be “thoughts, events, emotions, or factual knowledge,” and their affective valence indicates whether they are experienced as positive, negative, neutral, or ambivalent. The emphasis is on connotation and coherence: linked elements may reinforce one another, conflict with one another, and differ in emotional intensity. This makes CAMs a practical visual instantiation of the view that belief systems and attitudes are structured as cognitive-affective networks.
A recurrent misconception is to treat CAMs as equivalent to ordinary concept maps or mind maps. The crucial distinction is that CAMs make evaluative polarity intrinsic to the representation rather than leaving it implicit or inferential. What is mapped is not only semantic association, but cognitive-affective organization.
2. Visual grammar and representational structure
A CAM is a network composed of nodes and edges. Nodes are textual concepts, and their valence is encoded using both color and shape. The visual code described for direct-assessment CAMs is:
- green ovals = positive
- red rectangles = negative
- yellow rectangles = neutral
- purple superimposed oval and rectangle = ambivalent
Positive and negative nodes each have three intensity levels, indicated by thicker borders for stronger affective connotation. Neutral and ambivalent nodes do not have intensity levels (Reuter et al., 2022). The use of both color and shape provides redundancy, which supports interpretability, and allows valence to be read at a glance.
Edges represent relations between concepts. Solid lines indicate positive correlation, agreement, or mutual reinforcement; dashed lines indicate negative correlation, disagreement, or conceptual opposition. In Thagard’s original model, edges were undirected and noncausal. More recent implementations optionally allow solid or dashed arrows, adding causal directionality so that CAMs can be analyzed as directed Markov graphs. Figure-based descriptions in the method paper also refer to weak, moderate, and strong connections and to directed connection variants, but no formal weighting formula for edge strength is provided (Reuter et al., 2022).
Structurally, CAMs encode five types of information: which concepts are present; how they are evaluated affectively; how strongly they are felt in some cases; whether concepts support or oppose one another; and, optionally, whether one concept influences another directionally. This representational scheme is central to CAM’s distinctiveness. It embeds semantic content and emotional structure in a single graph, rather than separating topic content from affective judgment.
3. Direct elicitation, software, and research workflow
A major methodological shift in CAM research is the move from researcher-drawn maps to participant-constructed maps. Earlier studies often involved researchers drawing CAMs from interviews, memoirs, or texts, which was useful for visualization but labor-intensive and vulnerable to interpretive bias because researchers determined what concepts and relations to include. The browser-based open-source tool Valence changed this by allowing participants to construct their own maps directly, thereby supporting larger-scale empirical work and reducing researcher-imposed structure (Reuter et al., 2022).
The elicitation procedure is standardized. Participants are first instructed in how CAMs work and how to use the software. The paper presents example instruction slides based on the neutral topic “shopping at the farmers’ market,” explaining how to begin with a concept, add more factors, assign valence, interpret the colors and shapes, use thicker borders for stronger positive or negative intensity, and distinguish solid supportive links from dashed conflicting links. Participants are then asked to draw a CAM on the substantive target topic. In the COVID-19 application, the prompt was:
“We are interested in capturing your experience, the events, thoughts, and feelings resulting from the current coronavirus outbreak. Using the mapping tool, please draw everything that comes to mind concerning your experience with the coronavirus. Think about what matters in the current coronavirus outbreak, and please do your best to draw everything that comes to your mind concerning the coronavirus”
Participants can work in field, lab, or online settings, and attention checks can be used to improve response validity. Valence supports two elicitation modes: open generation of concepts by participants, and arrangement of predefined concepts supplied by researchers. The first mode is suited to idiosyncratic experience; the second supports more standardized comparison.
Valence records each graph’s network properties and allows export as .png image files for qualitative inspection and .csv files for quantitative or network analysis. The tabular export includes each node and its valence or strength, link information between nodes, strength and possible directionality of links, and the x- and y-positions of nodes in the drawing space. The method paper explicitly notes that physical distances between nodes are not recorded as analytically relevant measures (Reuter et al., 2022). Two online versions are available, one operated by the Cascade Institute and one by the University of Freiburg; the Freiburg version supports arrows and directedness.
In practical terms, CAM research proceeds by defining the focal topic or prompt; providing standardized instructions and examples; having participants construct maps in Valence; exporting image files for qualitative reading; exporting tabular graph data for quantitative analysis; analyzing structural properties, semantic content, or both; and optionally combining CAMs with questionnaires, experiments, or longitudinal designs.
4. Analytical strategies and network parameters
CAM analysis spans qualitative interpretation and quantitative network description. On the qualitative side, image exports can be read case by case to interpret how a person organizes experience, which concepts appear central, and how affect is distributed. Researchers can also conduct qualitative content analysis analogously to established approaches such as Mayring’s, with inductive or deductive coding. One example involved two raters independently forming categories to characterize within-subject change between pre- and post-CAMs, including categories such as “positive development” and “similar to before,” after which ratings were summarized by group and compared. Thematic analysis is another option: researchers can generate a list of all concepts appearing in CAMs and categorize them into themes automatically or manually, including automated categorization with word databases followed by manual completion. Suggested text-level tools include LegislatoR, Lexicoder, LIWC, and Python (Reuter et al., 2022).
On the quantitative side, CAM exports permit calculation of latent structural and valence properties. The paper does not provide formal equations for these measures, but it does define exemplary parameters verbally.
| Parameter | Definition | Scale |
|---|---|---|
| Average Valence | Mean value of all node valences of a CAM | -3 to 3 |
| Valence Percentages | Percentages of the individual valence options | 0 to 1 |
| Central Node Valence | Valence of the most central node | -3 to 3 |
| Centrality | Number of links on a node normalized by total possible links | 0 to 1 |
| Density | Number of a CAM’s links divided by total possible links | 0 to 1 |
| Diameter | Maximum distance from one node to another | — |
| Number of Nodes | Total number of nodes | — |
| Number of Links | Total number of links | — |
| Simpson’s Diversity | Extent to which a network employs heterogeneous properties | 0 to 1 |
| Triadic Closure/Transitivity | Total number of triangles divided by total possible triangles | — |
A methodological note attached to these measures is especially important: “Neutral nodes are usually counted as zero, while ambivalent nodes could be counted as zero (one value) or as -1,5 +1,5 (zero in sum, but two values)” (Reuter et al., 2022). This unresolved coding choice directly affects quantitative treatment of ambivalence.
CAM measures can be compared across treatment and control groups, across time in within-subject designs, or correlated with external questionnaire variables. The method paper states that “classical significance tests, especially correlation and regression analysis, can be used.” It also mentions other possible analyses, including dependency of the network on a specific emotion, diversity of emotional properties and connections, structural diversity, and analysis of directed graphs when arrows are used.
5. Applications, methodological position, and scope
The direct-assessment literature emphasizes that CAMs are especially useful when a study seeks to know not only what people think about a topic, but how their thoughts and feelings are organized into a structured system (Reuter et al., 2022). The most detailed example concerns experiences with the coronavirus pandemic. Participants were asked to draw CAMs about “everything that comes to mind” concerning the outbreak, and the sample map included concepts such as “Hundreds of thousands of deaths,” “Social distancing,” “Working from home,” “Additional strains on mental health,” “Virtual hangouts with friends,” “Had to postpone my wedding,” and “Getting to spend time with new nephew.” The example demonstrates that CAMs can represent heterogeneous and emotionally mixed experience within a single integrated network.
A second example concerns pre- and post-intervention CAMs about the coronavirus pandemic, where the intervention was a one-hour walk. Structural changes in the maps were used to assess effects of the intervention. This shows that CAMs can be used in experimental or quasi-experimental designs and for measuring change over time. Earlier work reviewed in the same source used expert-drawn CAMs to visualize Camp David negotiations, conflicting stakeholder perspectives in social disputes and water policy, ideological and political belief structures, scientific values and ethical conflicts, cross-cultural worldviews, allegory and analogy structures, and psychotherapy.
This range of applications supports two broad uses. One is representational: CAMs as tools for expert synthesis of complex belief structures from existing materials. The other is elicitative: CAMs as direct participant-generated data. The latter is the distinctive methodological contribution of the contemporary direct-assessment framework.
Methodologically, CAMs are positioned as a bridge between qualitative and quantitative research. Relative to interviews, they preserve participant-generated content while providing analyzable structure that can be compared more objectively across cases. Relative to surveys, they avoid many instrument biases associated with wording, response scales, and question order. Participants can express themselves freely rather than choosing among predefined response options. The method paper also argues that CAMs are faster to administer at scale than structured interviews and that their visual “face validity” tends to increase acceptance and willingness to participate (Reuter et al., 2022).
CAMs should not, however, be conflated with all other mapping paradigms. Compared with cognitive maps, they add explicit affective valence. Compared with mind maps, directed acyclic graphs, and similar mapping techniques, they integrate emotional connotation directly into the network. Compared with fuzzy cognitive maps, they are distinguished here by their applicability to individual-level differences and direct assessment of personal experience. The authors do not recommend replacing interviews or surveys, but rather using CAMs as a complementary method, especially in mixed-methods and triangulation designs.
6. Limitations, validation, and adjacent computational developments
The principal limitation emphasized in the direct-assessment paper is representational compression. CAMs reduce affect to four broad categories—positive, negative, neutral, and ambivalent—and therefore lose nuance relative to multidimensional emotion theories that include arousal or other appraisal dimensions. CAMs capture basic affective connotation rather than the full phenomenology of emotion (Reuter et al., 2022). Another unsettled issue concerns ambivalent nodes, whose quantitative coding is not standardized. A related concern is participants’ understanding of links, especially inhibitory or dashed links; further validation is needed to ensure that researcher interpretations match participant intent.
Reliability and validity are framed largely as open research agendas. The paper cites an initial validity assessment in which a Bayesian inference algorithm was used to estimate the probability of randomly replicating obtained CAMs, with results suggesting that the maps reflected intentional participant choices rather than random imputation. Even so, the same source stresses the need for further validation, especially concerning ambivalence and inhibitory links (Reuter et al., 2022). Best-practice recommendations follow from these limitations: provide clear standardized instructions and examples, consider attention checks, decide explicitly how ambivalent nodes will be coded, be cautious about assumptions regarding inhibitory links, combine qualitative reading with structural analysis where appropriate, consider within-subject designs when maps vary strongly between individuals, and supplement CAMs with questionnaires or experiments.
Several recent computational papers are relevant to CAM by analogy rather than by direct implementation. “Conceptual Cognitive Maps Formation with Neural Successor Networks and Word Embeddings” develops a successor-representation-based semantic mapping model in which semantically related items cluster and novel inputs are placed near learned conceptual neighborhoods, but it includes none of the affective dimensions standard CAM requires; its relevance is therefore methodological and indirect, as a possible conceptual backbone for a later affective layer (Stoewer et al., 2023). “Modeling Cognitive-Affective Processes with Appraisal and Reinforcement Learning” is likewise not a symbolic CAM system, but it supplies computational definitions of suddenness, goal relevance, goal conduciveness, and power, suggesting one route by which CAM-like relations could be grounded dynamically in reward processing and temporal-difference learning (Zhang et al., 2023). “DualMind: Towards Understanding Cognitive-Affective Cascades in Public Opinion Dissemination via Multi-Agent Simulation” explicitly separates a slowly evolving cognitive state from a rapidly fluctuating affective state in LLM-driven agents and couples them through memory and diffusion dynamics; it is best interpreted as a dynamic latent-state analogue of CAM rather than an explicit concept-affect graph (Huang et al., 28 Jan 2026).
Taken together, these adjacent developments suggest a widening computational landscape around CAM. A plausible implication is that future work may increasingly combine the explicit, interpretable graph structure of participant-generated cognitive-affective maps with dynamic appraisal, memory, or diffusion mechanisms. The core CAM contribution nevertheless remains distinct: the direct rendering of subjective experience as a structured, evaluatively coded concept network that is simultaneously vivid, comparable, and analytically tractable.
