Event Data Analysis
- Introduced by
- Conflict-studies and computational-social-science traditions (McClelland, Schrodt, King)
- Type
- Automated coding and analysis of who-did-what-to-whom event records
- Family
- Content analysis / computational text
What does it do?
Event data analysis converts streams of news reports into structured records of political interactions — who did what to whom, when — and aggregates them into time series of cooperation and conflict between actors.
Each event is coded as a source actor, an action type drawn from an ontology such as CAMEO, a target actor, and a date. Modern systems extract these events automatically from millions of news stories, enabling near-real-time measurement of interstate and intrastate behavior for forecasting and analysis.
The intuition behind the method
International and domestic politics unfold as a constant flow of discrete actions: a state condemns another, troops clash, leaders meet, a ceasefire is signed. Buried in the daily news is a record of all these interactions, but in unstructured prose. Event data analysis turns that prose into a database where every action is tagged with who initiated it, what kind of action it was, and who it targeted. Once each action is scored on a cooperation–conflict scale, the daily noise becomes a measurable time series of how hostile or cooperative two actors are toward each other, which can be charted, modeled, and forecast.
History
Political event data began with Charles McClelland's WEIS and Edward Azar's COPDAB projects in the 1960s and 1970s, hand-coded from news. Philip Schrodt and colleagues pioneered automated coding with systems like KEDS and TABARI and developed the CAMEO ontology, while King and Lowe's 2003 study demonstrated that automated extraction could match human coders. Schrodt's 2012 review surveyed the field's renaissance driven by web-scale text and machine coding, and large near-real-time datasets such as ICEWS and GDELT, exemplified by Boschee and colleagues' ICEWS release, made event data central to conflict forecasting.
When to use it
Use event data analysis when you need fine-grained, high-frequency measures of interactions between political actors over time — interstate conflict and cooperation, protest and repression, or crisis escalation — and when the relevant behavior is reported in news media. It is well suited to forecasting, early warning, and dynamic analysis of dyadic relations. It is less appropriate when the behavior of interest is poorly covered by media, when media bias systematically distorts which events are reported, or when the research question requires deep contextual interpretation that structured codes cannot capture.
Limitations and considerations
- Subject to media coverage and reporting bias: under-reported regions and events are systematically missing.
- Automated extraction makes coding errors, misidentifying actors, actions, or targets in ambiguous text.
- Duplicate reporting can inflate counts if deduplication is imperfect, distorting intensity measures.
- Structured codes strip context, so the same code can conflate substantively different actions.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Produces high-frequency, fine-grained time series of political interactions at a scale impossible for hand coding.
- Standardized ontologies (CAMEO) and intensity scales make events comparable across actors, places, and time.
- Automated coding enables near-real-time, continuously updated datasets for early warning and forecasting.
- Validated against human coders, with automated systems shown to match trained coders on conflict data.
Applications
- Forecasting interstate and intrastate conflict and crisis escalation for early-warning systems.
- Measuring dyadic cooperation and conflict between states over time to test international-relations theories.
- Tracking protest, repression, and contentious politics within countries at high temporal resolution.
- Analyzing the dynamics of mediation, negotiation, and de-escalation in conflicts.
- Building real-time situational-awareness dashboards from continuously coded global news.
How it works
How it works
A large corpus of news reports — from international wires and local outlets, often translated to English — is ingested continuously. Deduplication and filtering remove repeated stories and irrelevant content, since the same event reported many times must not be counted many times.
Natural-language processing parses each story to identify the actors and the action linking them, producing a structured tuple of source actor, action, target actor, and date. Actor dictionaries map names and roles to standardized codes (countries, governments, groups), and the verb phrase is matched to an action.
Formula
e = (\text{source actor}, \text{action}, \text{target actor}, \text{date})Each extracted action is mapped to a code in a standardized ontology such as CAMEO (Conflict and Mediation Event Observations), and each code carries a Goldstein-type intensity score placing it on a cooperation-to-conflict scale. This turns heterogeneous verbs into comparable, quantitative measures of interaction type and intensity.
Formula
\text{action} \mapsto \text{CAMEO code} \mapsto \text{Goldstein score}Coded events are aggregated by actor dyad and time period: for each pair of actors and each day, week, or month, the events are summed or intensity-weighted to produce a net cooperation–conflict score and event counts. These dyadic time series are the analytical unit for most studies.
Formula
C_{ab}(t) = \sum_{e \in t} g(\text{action}_e)\,\mathbb{1}\{src=a, tgt=b\}Because automated coding is imperfect, the data are validated against human coders and known benchmarks; King and Lowe showed automated extraction can match trained human coders, especially using rare-events evaluation. The resulting time series feed forecasting models, escalation analyses, and tests of international-relations theories.
Common pitfalls
- Failing to deduplicate, so heavily covered events are counted many times and dominate the series.
- Ignoring media-source bias and treating event frequency as a direct measure of underlying activity.
- Comparing data from different coders or ontologies without harmonization, since CAMEO and earlier schemes differ.
- Over-trusting automated codes for sensitive or ambiguous events without human validation of a sample.
Frequently asked questions
What is CAMEO and how does it relate to Goldstein scores?
CAMEO (Conflict and Mediation Event Observations) is a standardized ontology of political actions — verbal cooperation, material conflict, and so on — used to code event verbs into a common hierarchy. Goldstein-type scores assign each action code a numeric value on a cooperation-to-conflict scale, so that, for example, signing an agreement scores strongly cooperative and an armed attack scores strongly conflictual. Together they let heterogeneous news verbs be turned into comparable, quantitative intensity measures for aggregation.
How reliable is automated event coding compared to human coders?
King and Lowe's 2003 study, using a rare-events evaluation design, found that a well-built automated extraction system could match the performance of trained human coders on international conflict data at far lower cost. Reliability nonetheless varies with text quality, actor disambiguation, and the ontology, so best practice validates automated output against human-coded samples and known benchmarks, and treats coding error as a measurement issue to be modeled rather than ignored.
What is media bias in event data and why does it matter?
Event data are only as complete as the news that reports the events. Regions, actors, and event types that receive little media attention are systematically under-recorded, and source outlets may emphasize certain kinds of events. This reporting bias means event counts reflect both real activity and coverage patterns. Analysts mitigate it by using diverse and local sources, modeling under-reporting, and interpreting trends rather than treating raw counts as a census of behavior.
Related methods
Sources
Schrodt, P. A. (2012). Precedents, Progress, and Prospects in Political Event Data. International Interactions, 38(4), 546–569.
Open DOI referenceCopied.King, G., & Lowe, W. (2003). An Automated Information Extraction Tool for International Conflict Data with Performance as Good as Human Coders: A Rare Events Evaluation Design. International Organization, 57(3), 617–642.
Open DOI referenceCopied.Boschee, E., Lautenschlager, J., O'Brien, S., Shellman, S., Starz, J., & Ward, M. (2015). ICEWS Coded Event Data. Harvard Dataverse.
Open DOI referenceCopied.