Codebook

This codebook documents CLARA (Communication of Local German Party Associations, Recorded Annually). CLARA records what local party chapters put on their own websites, so it is a panel of observable local party web presence rather than a census of local party activity, personnel, or issue positions. The unit of observation is the county-level party association, the Kreisverband, which is a party's association for one county (Kreis). Kreisverbände account for 2,591 of the 2,635 chapters and for 2,316 of the 2,319 chapters in the recommended analysis scope, which is the analysis_scope_included filter defined in the roster below. The small remainder are city- or town-level chapters, identified by org_level. A Kreisverband typically contains subordinate town-level chapters; those are not separate observations. Every substantive field is derived from website content, so a value exists only where the chapter had a website and the Wayback Machine archived it in that year. Condition analyses on observability_status, which is defined with the panel below (see also the data report and the coverage tables).

This release contains nine files, with identifiers stable across all of them:

• chapter_roster.csv — one row per chapter (identity + geography).
• chapter_year_panel.csv — one row per chapter-year: features with per-feature status columns for missingness-aware analysis.
• chapter_year_panel_enriched.csv — the panel plus derived measures, web-activity measures, and issue-topic presence flags. It keeps the per-feature status columns and adds status columns for the derived measures.
• chapter_year_topic.csv — long-format per-family issue-block counts, with a cap_major_topic column.
• chapter_year_news.csv — one row per chapter-year that has archived news articles: the article count and the content-label shares.
• chapter_year_news_topic.csv — long-format per-family article counts: one row per chapter-year and topic family.
• chapter_year_provenance.csv — one row per chapter-year: non-sensitive source-level provenance (see below).
• topic_family_cap_crosswalk.csv — maps each issue family to a Comparative Agendas Project (CAP) major topic.
• kreis_wahlkreis_crosswalk.csv — maps each county to the Bundestag electoral districts it falls in, with population weights, for the 2017, 2021, and 2025 federal elections.

chapter_roster.csv (one row per chapter)

chapter_year_panel.csv (one row per chapter × year)

The panel is rectangular: every one of the 2,635 chapters has a row in every year from 2015 to 2025, including years in which nothing was observed. A missing chapter-year in an analysis is therefore always a filter you applied rather than a row the release omits. Identifiers and geography (chapter_id, party, org_level, org_name, ags, municipality, kreis, kreis_id, bundesland, population) repeat the roster values on every row, so the panel can be used without joining the roster. That includes kreis_id, the recommended merge key for county-level covariates.

Observability

Because has_wayback_coverage and usable_content_threshold_met are recodes of observability_status, they carry no information it does not already carry. Do not enter them in a model as separate covariates alongside it.

Release coverage taxonomy

For most purposes, filter first on analysis_scope_included = 1 and then pick one of two coverage rules. The conservative rule keeps only coverage_basis = 'full'. The inclusive rule keeps every non-blank coverage_basis, which is the full researcher-usable denominator. Running an analysis under both rules, with coverage_basis as the split, is a natural robustness check.

Features and their status columns

Every feature X in this file comes with a companion column X_status explaining why the feature has the value it has. observed_positive: extraction ran and found a positive value. observed_zero: extraction ran and found none, so the zero is a measured zero rather than a blank. coverage_available_not_extracted: pages that could have carried the feature were archived, but they were never parsed, so the gap is in processing rather than in the archive. missing_insufficient_coverage: there was no usable archived coverage to extract from in the first place. A feature value is informative only under the first two statuses; under the other two the value is missing, not zero.

Interpreting the panel

• A chapter-year whose observability_status is no_archive_coverage has no observed features at all for that year, because no snapshot exists to observe. Its blank feature values are missing, not zeros, and coding them as zeros would record an archiving gap as an absence of party activity. Nothing further can be collected for these chapter-years: an archive that never captured the page cannot be revisited.
• For analyses that need feature values to be reliably observed, either filter to usable_content_threshold_met = 1 or keep the wider sample and condition on the per-feature status columns.

chapter_year_panel_enriched.csv (core features + derived measures)

This file has one row per chapter × year, the same rows as chapter_year_panel.csv. It keeps the panel's columns, per-feature status columns included, apart from the redundant recode has_wayback_coverage, and adds derived measures computed from the underlying extracted records. Those derived measures carry four status columns of their own: issues_present_status, chair_gender_status, chapter_head_gender_status, and web_activity_status. Each is documented below alongside the measure it governs. This file is therefore the one to use for missingness-aware work on the derived measures, not just on the core features. Every value is an aggregate: no names, emails, or text appear anywhere in it. A measure is null when the records it is built from were not observed for that chapter-year.

Board composition and professionalization

Issue emphasis

These fields measure issue salience, meaning which topics a chapter writes about at all. They do not measure position or stance: whether a chapter is for or against something is not coded.

The long file chapter_year_topic.csv (chapter_id, year, topic_family, n, cap_major_topic) gives the full per-family issue-block counts for all 26 families. Its cap_major_topic column maps each family to a Comparative Agendas Project (CAP) major topic (see topic_family_cap_crosswalk.csv) so the data can be used alongside CAP-coded sources. The mapping is best-fit, and a few families (tourism, sport, family) involve documented judgment calls.

Issue-topic presence flags

Wide 0/1 flags for the 15 most common policy families, derived from the same topic classification as the counts above. topic_<family>_present = 1 if the chapter-year has at least one issue block in that family, 0 if issue content was observed but none of it fell in that family, and null if no issue content was observed at all. Families: governance, economy, environment, education, mobility, social, urban_development, culture, finance, health, integration, housing, digital, security, elections. Per-family counts for these and the remaining families are in chapter_year_topic.csv.

Web activity and mobilization

These are structural signals read off the archived pages by deterministic rules, with no page text stored.

The seven link flags (donation_, join_party_, newsletter_, contact_form_, program_, event_calendar_, volunteer_) are keyword searches over the page markdown. Links of this kind usually sit in a site's persistent header or footer, so a positive flag means the pathway exists somewhere on the site, at the level of the site template. It does not mean the chapter was actively campaigning or fundraising that year. Positive rates are correspondingly high and vary by party. The event_calendar_present keywords exclude the ubiquitous "Termine", so that flag discriminates rather than saturating near 90%. These features operationalize the information-provision, participation, and mobilizing functions of Gibson & Ward's (2000) party-website scheme. Flags are null where no page was archived for the chapter-year. Use them, along with n_news_pages and n_pages_total, conditional on observability_status.

Year-to-year reliability

Feature values are measured from whatever the Internet Archive captured of a site that year, and capture varies from year to year for the same chapter. This adds noise to comparisons of a chapter with itself over time, but not to comparisons across chapters. I measure it two ways, both computed only where a feature's status column marks the value as informative.

For counts and shares I use the autocorrelation of first differences. A persistent quantity measured without error gives about 0, noise as large as the true annual changes gives about −0.33, and pure noise gives −0.5. The release values are −0.42 for issue_block_count, −0.42 for official_count, −0.31 for n_social_platforms, and −0.25 for news_count. Computed by party, the same statistic ranges from −0.18 to −0.50 for issue_block_count, so cross-party comparisons of within-chapter dynamics carry unequal amounts of noise.

For binary flags I count how often a value differs from both its neighbors while the neighbors agree, divided by how often the value changes at all. Independent noise gives 0.5, a clean one-time adoption gives 0. The issue-topic presence flags run 0.40 to 0.44. The social platform flags run 0.12 to 0.31, so their changes are mostly real adoption. The flags in between behave as though roughly one chapter-year in ten is misclassified.

Part of this noise is systematic rather than random. The year-to-year change in issue_block_count correlates 0.28 with the change in log page count, and official_count 0.17. In a chapter fixed-effects regression, adding log page count cuts the annual trend in issue_block_count by about three quarters, turns a flat trend in official_count into a declining one, and leaves n_social_platforms almost unchanged. Page counts partly reflect real activity, so this control can absorb signal along with noise. Report estimates with and without it and read the difference as a bound.

Noise of this kind costs precision when a measure is the outcome and attenuates the coefficient when it is a regressor, with differencing making the attenuation worse. Averaging over multiple years helps issue_block_count and n_social_platforms but does little for official_count, so it is not a general remedy.

Method and caveats

• Gender is inferred from first names by a classifier running on a locally hosted open-weight model (Qwen/Qwen3-30B-A3B-Instruct-2507-FP8, the model behind every model-generated layer in this release), with results cached so a given name always receives the same label. Inferring gender from first names is a standard method (benchmarked at ~94% accuracy by Santamaría & Mihaljević 2018), but it is a proxy: it errs on unisex names and on non-German or transliterated names, and it does not capture nonbinary identities. Checks on the released files are reassuring: the mean board female share orders as GRÜNE > LINKE > SPD > CSU > CDU > FDP > AfD. Still, treat any single chapter-year value as an estimate and condition on gender_coverage. Two mechanics deserve particular care. First, pages that print names as Nachname, Vorname can feed the surname to the classifier instead of the first name. Second, female_share is computed for boards of any size, with no minimum number of officials: on a board with one or two gender-classified officials the share can only take extreme values such as 0, 0.5, and 1, so it is very noisy there. Where the share carries analytic weight, restrict the sample to chapter-years with enough officials, using official_count for board size or n_female + n_male for the share's own denominator.
• Topic family and role category likewise come from the local model, which assigns each item one label from a fixed list (issue families are listed in config/issue_topic_crosswalk.csv; roles are the standard board functions). For topic families, a 54-entry dictionary resolves the most common topic strings by exact match and the model classifies the remaining topics from the heading text alone; role categories are assigned by the model throughout. No per-topic ground-truth accuracy check was run, which makes this the largest unverified assumption behind every topic field. dominant_topic_family additionally skews toward chapters with fuller issue pages, since a chapter needs several blocks before one family can stand out.
• officials_source records which page type supplied most of the board records. The extracted board pools three different groups of people: the chapter's internal party officers, electoral candidates, and elected representatives on the council or in the party's Fraktion. Condition board size, gender, and board_completeness on officials_source. A Fraktion roster, for example, has no treasurer or secretary to list and mixes in council members, so it scores low on board_completeness for a structural reason rather than a substantive one.
• chair_female means "at least one chair is female," not "the chair is female." Chapters with a co-chair structure (Doppelspitze, two people sharing the chair) therefore get two chances to trigger the 1, and that structure is more common in GRÜNE and LINKE, so read the measure as an upper-bound signal.
• The strict chapter_head_* fields are the research-facing leadership measure. They use focal party, organization level and name, role title, page type, and source URL to exclude faction chairs, affiliates, deputies, honorary and former chairs, other organizational levels, committees, and non-organizational spokespersons. The broad chair_* fields are left unchanged for reproducibility and comparison.
• Archival density confounds the substantive measures. Whether content is observed at all tracks how many pages the archive happened to capture: chapter-years where content is observed average about 100 archived pages, while those where it is not average about 20. Usable coverage also rises over time, from about 35% in 2015 to about 73% in 2025, and it differs across parties. A trend or a party gap in any feature can therefore be a trend or gap in archiving. Condition longitudinal and cross-party comparisons on observability_status rather than only noting it.
• photo_share operationalizes personalization of the chapter's web presence in the sense used in the German candidate-website literature (Zittel & Schweitzer): how far the site presents identifiable individuals rather than only the organization.
• Social-presence flags exclude national party accounts (e.g. facebook.com/CSU) via a handle filter. CSU has the lowest observed rate on every platform despite high page coverage. That is a joint product of the filter and CSU's heavy use of party-provided national templates, both of which suppress chapter-specific handles rather than inflate them, so low CSU social rates are expected and not an artifact to correct. A flag means a link was present on an archived snapshot, not that the account was active that year, since dead links persist in footers.
• n_social_platforms is the deterministic sum of the six platform flags. Entering it alongside those flags produces perfect collinearity, and its variation is dominated by Facebook and Instagram, the two common platforms.

Relation to standard coding schemes

The feature set maps onto established website- and text-coding frameworks. Gibson & Ward's (2000) five party-website functions (information provision, resource generation, networking, participation, campaigning) correspond to the program/contact, donation/join, social-link, event/volunteer, and news/activity fields. Issue families crosswalk to Comparative Agendas Project (CAP) major topics via topic_family_cap_crosswalk.csv. Issue fields measure salience, not position.

chapter_year_news.csv (one row per chapter-year with archived articles)

News volume and content-label shares per chapter-year, built from the deduplicated article records. Every article was classified by the same locally hosted open-weight model into a topic family, an article type, and a geographic focus; the human validation of those labels is reported in the quality section of the README. The file holds one row per chapter-year with at least one archived article, 14,733 rows in all. A chapter-year absent from this file had no article to summarize.

The share columns describe the composition of what a chapter posted and the archive captured, not how much it posted; a chapter-year with three archived articles can show a share of 1.0. Combine the shares with news_count whenever volume matters. The campaign share tracks the electoral calendar, peaking in 2017, 2019, 2021, 2024, and 2025.

chapter_year_news_topic.csv (long format, news topics)

Per-family article counts per chapter-year (90,668 rows), mirroring chapter_year_topic.csv but built from classified news articles: chapter_id, year, topic_family, n, cap_major_topic. Issue-page topics come from static program content and change little within a chapter, whereas news topics move year to year, so this file is the one to use for issue-attention dynamics. Articles labeled other or unclear are excluded.

chapter_year_provenance.csv (one row per chapter-year)

Source-level provenance for each chapter-year, recorded so that any aggregate in the release can be audited without publishing page text or personal data. Keyed on chapter_id and year.

Use these to gauge how much source material underlies a chapter-year's measures and to condition on capture timing.

kreis_wahlkreis_crosswalk.csv (one row per county × election × district)

Counties and Bundestag electoral districts (Wahlkreise) do not nest. For the 2025 election 293 counties lie entirely inside one district while 107 span two or more, and a district can equally cover several counties. District-level election results therefore cannot be joined onto the panel by county alone. This file supplies the weights that make the merge possible for the 2017, 2021, and 2025 federal elections. For each county and election it lists every district the county falls in, together with the share of the county's population living in the part of the county assigned to that district. Join on kreis_id and the election you need, then aggregate district results using population_share as the weight. The file has 1,640 rows and is unique on county, election year, and district.

The assignments come from the Bundeswahlleiterin's official municipality-to-district lists for each election, combined with municipality populations taken from the edition of the Destatis municipality register closest to that election's territorial status.

The weights come with two limits. First, where a single municipality is itself split across several districts, the published assignment gives no sub-municipal population figures, so that municipality's population is divided equally among the districts it touches. This affects 18 large cities, and for them population_share is an approximation rather than a measurement; in the file it shows up as exactly equal shares across a county's districts, as for Köln in 2025, which carries 0.25 in each of its four districts. Second, the file covers every German county, which is a superset of the 399 counties in the sampling frame. It also carries Cochem-Zell, which has no municipality of at least 5,000 residents and therefore falls outside the frame, and Eisenach, which was a county in its own right for the 2017 and 2021 elections before being merged into the Wartburgkreis. Rows for counties outside the frame find no match when the file is joined to the panel.