Data dictionary

Every column of the Magellan Longevity Open Evidence Dataset, with its type, meaning and — the part that matters for reuse — exactly which source field it came from. Machine-readable copies of this page: JSON, Markdown, Frictionless datapackage.

Encoding is UTF-8 throughout. CSVs follow RFC 4180: comma-delimited, " escaped by doubling, one header row, LF line endings. Multi-value cells use ; with no spaces. An empty cell means the source does not record that value — it is never a zero, never a placeholder, and nothing is imputed.

Table 1 — products

130 rows × 32 fields. Primary key product_id. Files: products.csv, products.json, products.jsonl.

Table 2 — studies

2692 rows × 32 fields. Primary key study_key. Files: studies.csv, studies.json, studies.jsonl.

Controlled vocabularies

evidence_tier_key

grade_source

reported_direction

How reported_direction is computed

reported_direction is produced by a deterministic, published rule cascade, not by a human or a language model. Unfavorable cues are checked first, then no-difference cues, then favorable cues; a negation guard suppresses a cue when it is preceded within 45 characters by no / not / without / free of / lack of / absence of / neither. Rows where an unfavorable cue co-occurs with any other cue, or where a no-difference cue co-occurs with a favorable cue, are labelled mixed. Rows where nothing fires are labelled unclear and are NOT counted as null results. For observational biomarker papers the "direction" is the direction reported for the biomarker, not for an intervention. Every row carries direction_evidence so the classification can be audited or discarded.

It is applied to the write-up conclusion plus takeaway where a structured write-up exists (314 rows), and to the verbatim catalogue quote otherwise (2378 rows). That difference is recorded per row in direction_basis, because a fragment quoted from an abstract supports a weaker inference than a full conclusion sentence.

If you disagree with the classifier, you can rebuild the column yourself: outcome_statement, conclusion and takeaway are all shipped verbatim.

Joining the tables

products.study_keys        --split(";")-->  studies.study_key
studies.linked_product_ids --split(";")-->  products.product_id

Table 2 contains 46 rows that repeat a PubMed identifier already present under another internal key. Filter is_primary_record_for_pmid == true to reduce it to 2646 unique papers before counting anything.

import pandas as pd
p = pd.read_csv("products.csv")
s = pd.read_csv("studies.csv")
unique_papers = s[s.is_primary_record_for_pmid]          # 2646 rows
strong = p[p.evidence_tier_key == "strong"]              # 11 rows
null_findings = s[s.reported_direction == "no_difference"]  # 144 rows

More