Empirical Product Catalog

Document status: CURRENT PRODUCT CATALOG
Last verified: 2026-08-23

This page answers:

What empirical objects can I actually use today, and what do they mean?

The catalog deliberately separates four ideas:

• implemented — the producing code and contract shape exist;
• synthetically accepted — the path is exercised on controlled fixtures;
• materialized / accepted on real source data — a durable real source artifact has been run through that path and recorded;
• experiment-used — a declared scientific experiment has consumed that current product.

Those are different evidence levels. A product can be implemented and synthetically accepted without yet having a canonical real-data experiment run.

For authority rules, see Current Empirical Data Authority. For experiment readiness, see Experiment Surface Catalog and Validation Status.

At a glance

Product / foundation	Natural grain	Geography	Coverage semantics	Current use status
ACLED source-native violence	event Silver; geo_uid × period_id × native_event_type Gold	shared analytical geography via explicit membership relation	sparse Gold; absent row = unknown by default	implemented; contracted harness projection exists; canonical real current-artifact run not recorded
AidData GeoGCDF	source project-geometry Silver; contracted area-period commitment measurement	shared point / areal geography relation; polygons may span units	verified structural zero only inside declared source/geography/time support and strict resolution conditions	implemented reference investment measurement; real current experiment not recorded
World Bank Projects API	one source project per exact WB id	no current source-native spatial exposure product asserted by this catalog	project-row absence has no no-project/control meaning	source-native Silver implemented; current WB-only contracted spatial experiment still needs a measurement path
AidData CLG-LMIC	relational source records plus child tables	no current source-native spatial exposure product asserted by this catalog	row absence has no no-project/control meaning	source-native Silver implemented; not itself a treatment surface
Survey-native substrate	household / person / respondent / cluster / EA / other declared grain	explicit geography relations with ambiguous/unmatched states	survey coverage is not coerced into area-period zero semantics	reusable substrate implemented and used by DHS verticals; Afrobarometer ingestion still absent
DHS Household Recode (HR)	conceptual household observation; durable physical key source_row_id	cluster identity preserved; analytical geography is separate	source missingness remains missing; no implicit area-period coverage semantics	source-native Silver v2 implemented; truthful grain contract; protected real-data acceptance and experiment use not recorded
DHS Geospatial Covariates (GC)	survey × cluster; optional long survey × cluster × source_variable	cluster-associated measurement; no GID aggregation	cluster availability; absent/missing never zero; temporal meaning registry-driven	cluster covariate Silver implemented and hardened; current experiment use not recorded
DHS GE/GPS Geography	supplied GPS cluster row + separate geography-membership relation	reported-coordinate membership with displacement/ambiguity metadata	coordinate/linkage states remain explicit; no true-location claim	cluster Silver + reported-coordinate geography implemented; current experiment use not recorded
DHS integration QA	support/provenance report across HR + GC + GPS	compares cluster identity/support; does not assign household geography	source-only clusters and unresolved ID normalization remain explicit	implemented synthetic QA evidence; not an analysis dataset
DHS Household Semantic Measurements	source_row_id × measurement_id	inherits household/cluster identity; no exposure assigned	household observation coverage; missing/unmapped codes remain explicit	initial DHS-VII codebook-backed registry implemented; no experiment roles assigned

DHS is now an integrated empirical family

The DHS Empirical Stack now has five distinct capabilities around one survey identity:

SurveyCatalog
├─ HR  → household Silver
├─ GC  → cluster covariate Silver + temporal semantics
├─ GPS → displaced reported-coordinate Silver
│        → reported-coordinate geography relation
├─ integration QA → survey / grain / cluster-support evidence
└─ HR registry → household semantic measurements

This is an important change from both earlier states: first when only the generic survey substrate existed, and later when HR/GC/GPS existed but were documented mainly as parallel verticals.

The integration report still does not create a canonical joined DHS table. It verifies that the separate products can be related without silently losing source-only clusters, confusing DHSID with DHSCLUST, normalizing IDs such as 001 → 1, or accepting a false DatasetRef.grain claim.

The semantic registry also remains deliberately upstream of scientific use. A measurement such as dhs.household.electricity_access has an explicit empirical meaning; whether it becomes an FCV outcome or control is still an experiment decision.

How to read a product card

Every card uses the same questions:

1. Source authority — what registered source or release is authoritative?
2. Natural grain — what does one row or observation naturally represent, and what key is actually guaranteed unique?
3. Geography — how is spatial identity linked without hiding ambiguity?
4. Time / coverage — what temporal support is actually asserted?
5. Important facts — which fields are source/empirical measurements rather than scientific roles?
6. Contract artifacts — which durable contracts and manifests make the product reproducible?
7. Known limitations — what must not be inferred from the product?
8. Experiment history — has a current contract-backed experiment actually consumed it?

Where the producing repository does not publish a stable cross-release date range, this catalog says so rather than inventing one from a historical file.

Product authority rule

The product cards summarize current producing repositories; they do not replace them.

human product card
      ↓
fcv-empirical-data product / source vertical
      ↓
DatasetRef + contracts + RunManifest + durable artifact

If a card and a producing artifact disagree, investigate the producing artifact and update this catalog. Do not silently choose the more convenient interpretation.

What is deliberately not a product here

This catalog does not list the following as empirical products:

• treated/control flags;
• experiment outcomes or covariates;
• matching samples;
• causal exposure dates;
• estimator-ready balanced panels;
• pooled cross-source investment totals;
• jobs_any or similar labels interpreted as treatment;
• a household-level DHS mega-table with GC/GPS values replicated by convenience;
• a de-displaced DHS true-location table;
• a silently normalized cluster crosswalk that rewrites 001 and 1 as the same source identity;
• historical variable meanings promoted without current codebook provenance.

Those are experiment-use, uncertainty-modeling, harmonization, or derived-annotation concepts and require their own declared provenance and scientific rules.

Next products expected

The survey-side extensions are now narrower and more concrete:

• additional codebook-backed DHS household measurements only where definitions are authoritative and useful;
• additional DHS recode families such as PR/IR/KR only when a named scientific question needs them;
• a displacement-aware candidate-geography product if an experiment requires uncertainty beyond reported-coordinate membership;
• Afrobarometer source-native respondent/EA verticals;
• broader cross-survey harmonization only when its comparability assumptions can be stated explicitly.

The product catalog should add those only when the producing architecture actually exists.