Skip to content

Glossary

Terms defined once here, linked from every Reference and Guide page that uses them. If you find yourself re-explaining one of these inline, link here instead.

CDR (Curated Data Repository) The versioned BigQuery dataset containing all All of Us research data. Each release is identified by a version string (e.g. C2024Q3R9). Access it via the WORKSPACE_CDR environment variable on the Researcher Workbench: os.environ["WORKSPACE_CDR"].

Registered Tier The default data access level. Contains demographics, conditions, procedures, measurements, drug exposures, and survey data. Does not include genomics or unshifted dates (dates are shifted by a random offset per participant).

Controlled Tier Elevated access requiring additional IRB approval. Adds genomic data (short-read WGS), unshifted dates, and geographic data below state level.

Index date The reference date for a participant in a study — typically the date of first diagnosis for cases. All temporal features (labs, medications, conditions) are anchored relative to this date to prevent temporal leakage.

Pseudo-index date A synthetic index date assigned to controls who lack the defining event. Common approaches: random date within the control's EHR span, or matched to the case's index date in a paired design. Required for any temporal analysis involving controls.

Carrier A participant who carries at least one variant in a specified gene or gene panel, as determined from short-read WGS data in the Controlled Tier.

P/LP (Pathogenic / Likely Pathogenic) ClinVar classification indicating a variant is disease-causing or probably disease-causing. The standard clinical-significance filter for germline predisposition analyses.

FDR / q-value False Discovery Rate correction for multiple hypothesis testing. The q-value is the FDR-adjusted p-value. Use Benjamini-Hochberg (statsmodels.stats. multitest.multipletests(method='fdr_bh')) as the default correction.

Dry run A BigQuery feature that estimates bytes processed without executing the query or incurring cost. Call client.query(sql, job_config=QueryJobConfig (dry_run=True)) and read job.total_bytes_processed.

Bytes billed The actual storage scanned by a BigQuery query, rounded up to the nearest 10 MB minimum. AoU Workbench credits are consumed based on bytes billed, not bytes returned.

OMOP CDM (Common Data Model) The data model underlying AoU's CDR. Tables like condition_occurrence, measurement, drug_exposure, and person follow OMOP conventions but with AoU-specific extensions and naming in the cb_ (Cohort Builder) tables.

cb_ tables vs. OMOP tables AoU provides both standard OMOP tables (e.g. condition_occurrence) and Cohort Builder tables (e.g. cb_condition). The cb_ tables are pre-joined/denormalized for common queries. Reference pages specify which table family they use.

concept_id The OMOP integer identifier for a clinical concept (condition, drug, lab test, etc.). Mapped from source codes (ICD-10, LOINC, RxNorm) via the concept and concept_relationship tables.

person_id The unique participant identifier in AoU. All clinical tables join on person_id. Always cast as INT64 in BigQuery queries.

observation_period The OMOP table recording each participant's span of data availability. observation_period_start_date and observation_period_end_date define when AoU has records for a participant — not necessarily when they first entered the healthcare system.

visit_occurrence_id Foreign key linking clinical events (conditions, measurements, drugs) to a specific visit/encounter. Often NULL in AoU data — inner joins on this column silently drop unlinked events.

type_concept_id A family of columns (condition_type_concept_id, measurement_type_concept_id, etc.) indicating the provenance of a record: EHR, self-reported survey, physical measurement at enrollment, etc. Critical for distinguishing data sources.

ds_survey A denormalized survey table available in newer CDR versions. Provides pre-joined survey questions and answers, simpler to query than the raw observation table for survey data.

Quasi-separation / complete separation A logistic regression failure mode where a predictor perfectly (or nearly perfectly) predicts the outcome — common with rare variants where all carriers are cases. Produces inflated odds ratios with enormous confidence intervals. Use Firth's penalized regression when carrier counts are low.

Temporal leakage Information from after the prediction point (index date) contaminating features that should only use pre-index data. Produces models that look accurate in development but fail on prospective data.