docs/data_dictionary.md

Data dictionary

Version 1.0.0. Row counts below come from the current build manifest; independent release validation is pending. Every table has a CSV and Parquet file under data/. The machine-readable schema records Arrow types, while this dictionary explains meaning, keys, units and missing values.

Conventions and joins

• post_id is an original public source-post ID, stored as a string. Read it as a string when loading CSV; do not convert IDs to floating point. It joins to posts.post_id.
• No source reply ID or account identifier is exported. observation_id is a zero-based sequential row number assigned in source CSV order after first-ID deduplication. It is local to this build and is not a source key or a chronological ordering.
• topic_id joins to topics.topic_id; phrase_id joins to phrases.phrase_id; delay_band joins to delay_bands.delay_band. These joins are many-to-one from feature/summary tables. Topic and phrase tables overlap by design.
• Dates use YYYY-MM-DD in UTC; months use YYYY-MM. Source-post timestamps retain the upstream UTC ISO 8601 timestamp. Reply timestamps are not exported.
• Counts are integers. Shares are fractions from 0 to 1, not percentages. Multiply by 100 for percent or 1,000 for matches per 1,000 records.
• CSV booleans are true/false. Blank CSV cells represent null; Parquet preserves typed nulls. Arrow fields are declared nullable even where this release populates every row. Do not interpret schema nullability as evidence of observed missing data.
• Dense post/pattern/band tables include explicit zero counts. Feed tables contain supplied days/months only; omitted dates are not asserted to be platform inactivity. Cohort detail tables are sparse positive counts within the captured window; see their descriptions below.
• Snapshot engagement fields use the builder's nonnegative-integer conversion: missing, empty, invalid or negative source values become zero. They do not distinguish observed zeros from those fallback cases and are not current engagement counts.

The reply component uses CC BY-NC 4.0. feed_daily, feed_monthly and feed_topics are marked component-specific in the manifest: the CNN / Matt Stiles archive declaration is CC0, and underlying post rights remain separate. Code is MIT-licensed.

reply_features: 377,622 rows

One retained captured reply per row. Primary key: observation_id; foreign keys: post_id, delay_band.

Field	Type / unit	Meaning
observation_id	int64	Zero-based local row number in retained source-file order. No source reply identifier is encoded.
post_id	string	Public selected parent-post ID; joins to posts.
reply_day_utc	string / UTC day	Calendar day of the captured reply.
delay_hours	float32 / hours	Reply timestamp minus source-post timestamp, cast to IEEE 754 Float32 before assigning the band.
delay_band	int8 / 0–10	Lower-inclusive, upper-exclusive band computed from the Float32 delay.
likes_at_collection	int64 / count	Upstream reply favourites_count after the conversion described above.
has_captured_text	bool	Whether HTML-entity-decoded, stripped content_clean is nonempty. The text itself is omitted.
topic_*	bool	At least one match to the corresponding topic regex; the suffix equals topic_id.
phrase_*	bool	At least one match to the corresponding phrase regex; the suffix equals phrase_id.

Topic columns: topic_data_centers, topic_ai, topic_power, topic_energy, topic_jobs, topic_water, topic_tariffs, topic_iran, topic_peace, topic_war, topic_crypto, topic_prices.

Phrase columns: phrase_thank_you, phrase_fake_news, phrase_nuclear_weapons, phrase_strait_of_hormuz, phrase_peace_through_strength, phrase_god_bless, phrase_total_victory, phrase_make_america_great_again, phrase_united_states, phrase_radical_left, phrase_ceasefire, phrase_iran, phrase_oil, phrase_peace, phrase_war, phrase_america.

Matching uses Python re.IGNORECASE with no language filter. Each flag counts presence, not occurrences. Empty captured text yields false flags. The source CSV contains 377,624 rows; retaining the first row per source comment ID removes two duplicates. IDs are used transiently and not serialized.

posts: 97 rows

One selected parent post per row. Primary key: post_id.

Field	Type / unit	Meaning
post_id	string	Original public post identifier.
posted_at_utc	string / UTC timestamp	Source-post creation timestamp.
source_url	string / URL	Public source-post URL with query and fragment removed.
captured_replies	int64 / replies	Deduplicated reply rows retained for this parent.
nonempty_reply_records	int64 / replies	Retained replies with has_captured_text=true.
likes_at_collection	int64 / count	Source post's retained favourites_count.
reposts_at_collection	int64 / count	Source post's retained reblogs_count.
reported_replies_at_collection	int64 / count	Upstream post replies_count, separate from the number of replies captured in this dataset.

Snapshot engagement counts use the conversion described under conventions. A reported reply total need not equal captured replies.

topics: 12 rows

Primary key: topic_id.

Field	Type	Meaning
topic_id	string	Stable pattern name and suffix of reply/feed topic columns.
label	string	Human-readable pattern label.
pattern	string / Python regex	Exact case-insensitive matching rule, not a classified topic or stance label.

The twelve IDs are listed with the reply-feature columns above. Some patterns match ambiguous terms such as AI, energy or water. Different topics may match the same record.

phrases: 16 rows

Primary key: phrase_id.

Field	Type	Meaning
phrase_id	string	Phrase text with spaces replaced by underscores; suffix of reply phrase columns.
phrase	string	Literal phrase being searched.
pattern	string / Python regex	Exact case-insensitive pattern, using word-character boundaries and flexible whitespace between phrase words.

delay_bands: 11 rows

Primary key: delay_band. All bounds are measured in hours.

Field	Type / unit	Meaning / missing value
delay_band	int64 / 0–10	Join key used by reply and delay-summary tables.
lower_hours_inclusive	float64 / hours	Included lower bound.
upper_hours_exclusive	float64 / hours	Excluded upper bound. Null for band 10 means no upper bound.

Band	Interval
0	0 to less than 1 minute
1	1 to less than 5 minutes
2	5 to less than 15 minutes
3	15 minutes to less than 1 hour
4	1 to less than 3 hours
5	3 to less than 6 hours
6	6 to less than 12 hours
7	12 to less than 24 hours
8	24 to less than 72 hours
9	72 to less than 168 hours
10	168 hours and above

Use the numeric bounds in the table, applied to the Float32 delay, for exact reconstruction. Rounding to another precision before assigning bands can change values near boundaries.

post_topics: 1,164 rows

Dense 97 × 12 combinations. Primary key: (post_id, topic_id).

Field	Type / unit	Meaning
post_id, topic_id	string	Parent-post and topic foreign keys.
source_contains_topic	bool	Selected source-post text matches the topic pattern.
matching_replies	int64 / replies	Captured replies matching that topic at least once.
captured_replies	int64 / replies	All retained replies to this post, including empty text.
reply_share	float64 / fraction	matching_replies / captured_replies; null if the denominator is zero.

Source-post matching uses the supplied cleaned source text after removing publisher sharing-token URL parameters. Reply text is entity-decoded and stripped before matching.

post_phrases: 1,552 rows

Dense 97 × 16 combinations. Primary key: (post_id, phrase_id).

Field	Type / unit	Meaning
post_id, phrase_id	string	Parent-post and phrase foreign keys.
source_occurrences	int64 / matches	Number of nonoverlapping phrase-regex matches in cleaned source-post text. Each occurrence contributes to the count.
matching_replies	int64 / replies	Captured replies matching the phrase at least once.
captured_replies	int64 / replies	All retained replies to the post, including empty text.
reply_share	float64 / fraction	matching_replies / captured_replies; null if the denominator is zero.

To calculate own-thread recurrence, sum matching_replies only where source_occurrences > 0. This is descriptive recurrence, not evidence of exposure or agreement.

post_delays: 1,067 rows

Dense 97 × 11 combinations. Primary key: (post_id, delay_band).

Field	Type / unit	Meaning
post_id	string	Parent-post foreign key.
delay_band	int64	Delay-band foreign key.
captured_replies	int64 / replies	All captured replies for this post whose Float32 delay falls in the band, including empty text.

Zero-count bands are retained.

post_topic_delays: 12,804 rows

Dense 97 × 12 × 11 combinations. Primary key: (post_id, topic_id, delay_band).

Field	Type / unit	Meaning
post_id, topic_id	string	Parent-post and topic foreign keys.
delay_band	int64	Delay-band foreign key.
matching_replies	int64 / replies	Replies matching this topic within the post/band.
captured_replies_in_band	int64 / replies	All captured replies in the post/band, including empty text; denominator for a within-band share.

Zero matches and zero denominators are retained. A share calculated with a zero denominator is undefined, not zero.

post_phrase_delays: 17,072 rows

Dense 97 × 16 × 11 combinations. Primary key: (post_id, phrase_id, delay_band).

Field	Type / unit	Meaning
post_id, phrase_id	string	Parent-post and phrase foreign keys.
delay_band	int64	Delay-band foreign key.
matching_replies	int64 / replies	Replies matching this phrase within the post/band.
captured_replies_in_band	int64 / replies	All captured replies in the post/band, including empty text; denominator for a within-band share.

Zero matches and zero denominators are retained. Join post_phrases on (post_id, phrase_id) and require source_occurrences > 0 for own-thread phrase recurrence by band.

participation_daily: 52 rows

One supplied UTC reply day per row. Primary key: date.

Field	Type / unit	Meaning
date	string / UTC day	Captured reply day.
captured_replies	int64 / replies	Retained replies on this day.
active_accounts	int64 / source accounts	Distinct source account IDs with at least one retained reply on this day, aggregated before export.
first_observed_accounts	int64 / source accounts	Accounts whose earliest captured reply day is this day.
returning_accounts	int64 / source accounts	Active accounts first observed on an earlier captured day: active_accounts - first_observed_accounts.
last_observed_accounts	int64 / source accounts	Accounts whose latest captured reply day is this day.

Counts refer to accounts, not verified unique people. Returning does not require activity on the immediately preceding day. Summing active or returning counts across days counts account-days, not unique accounts. First and last observation do not measure a platform join or exit.

cohorts: 52 rows

Primary key: first_day_utc, the first captured reply day defining the cohort.

Field	Type / unit	Meaning
first_day_utc	string / UTC day	Earliest captured reply day for all accounts in the cohort.
accounts	int64 / source accounts	Number of distinct source accounts in that cohort.
captured_replies	int64 / replies	All captured replies across the collection by accounts in this cohort, including replies after their first day.

Cohort account counts sum to 123,271. There are 50,622 accounts with more than one captured reply; that is not a multiday-participation count. Source account IDs are used transiently to build aggregates and never exported.

cohort_activity: 1,310 rows

Primary key: (first_day_utc, active_day_utc). Join first_day_utc to cohorts; join active_day_utc to participation_daily.date.

Field	Type / unit	Meaning
first_day_utc	string / UTC day	Cohort's first-observed day.
active_day_utc	string / UTC day	Captured day on which cohort members replied.
active_accounts	int64 / source accounts	Distinct accounts in this cohort with at least one reply on the active day.

Only positive-count combinations are exported. For a valid cohort/day within the captured window on or after its first day, an omitted combination means zero captured active accounts. Earlier days are outside the cohort's observed lifetime; dates outside the collection have no coverage. Repeated days cannot be summed to infer unique people or continuous participation.

cohort_last_observed: 1,261 rows

Primary key: (first_day_utc, last_day_utc). Join the first day to cohorts; join the last day to participation_daily.date.

Field	Type / unit	Meaning
first_day_utc	string / UTC day	Cohort's first-observed day.
last_day_utc	string / UTC day	Latest day on which these accounts have a captured reply in this collection.
accounts	int64 / source accounts	Accounts sharing the first/last captured-day pair.

Only positive-count pairs are exported. An omitted valid first/last pair within the captured window means zero accounts observed with that pair. Last day must be on or after first day. Collection endpoints censor these histories; last observed is not an unfollow, exit or political defection.

feed_daily: 1,646 rows

One supplied broader-feed UTC day per row. Primary key: date. This is a separate population from captured replies.

Field	Type / unit	Meaning
date	string / UTC day	Supplied day containing at least one retained feed record.
records	int64 / feed records	Feed records after first-ID deduplication and the November 4 cutoff, including empty text and shared material.
nonempty_records	int64 / feed records	Records with nonempty content after stripping HTML tags, decoding entities and trimming whitespace.
topic_*	int64 / feed records	Records matching the corresponding topic at least once; same twelve suffixes as reply_features. These columns are counts, not booleans.

Topic patterns overlap. Divide by records, including empty content, for the rate used by the monthly feed-topic summaries. Supplied days are retained; missing calendar days are not filled or labeled as inactivity.

feed_monthly: 57 rows

Primary key: month.

Field	Type / unit	Meaning
month	string / UTC month	Supplied feed month. The first and last months are bounded by the snapshot window.
records	int64 / feed records	All retained feed records in this month, including empty text and shared material.
nonempty_records	int64 / feed records	Records with nonempty cleaned content, using the same transform as feed_daily.

The 37,951-record feed spans February 14, 2022 through November 4, 2026. Partial endpoint months are not full-month measures. Later feed months do not extend the reply sample past April 20.

feed_topics: 684 rows

Dense 57 × 12 combinations over supplied months. Primary key: (month, topic_id). Join to feed_monthly and topics.

Field	Type / unit	Meaning
month	string / UTC month	Supplied feed month.
topic_id	string	Topic foreign key, using the same regex definitions as reply features.
matching_records	int64 / feed records	Records in the month matching the topic at least once.
records	int64 / feed records	All retained feed records in the month, including empty content.
share	float64 / fraction	matching_records / records; supplied months have positive denominators.

Zero matches are explicit. Topic shares overlap and must not be summed as mutually exclusive categories. Feed counts include shared material and do not isolate Trump's original wording.