DHS Household Semantic Measurements

Product status: IMPLEMENTED CODEBOOK-BACKED DERIVED MEASUREMENT
Authority: L3 REBUILT / RELEASE VALIDATION STILL REQUIRED
Protected real-data acceptance: NOT RECORDED IN THIS SITE
Experiment role status: NO TREATMENT / OUTCOME / COVARIATE ROLE ASSIGNED
Producing repository: fcv-empirical-data

What it is

The DHS stack can now move from opaque HR source variables to a deliberately small set of reusable empirical meanings backed by official DHS recode documentation.

content-hashed HR household Silver
        ↓
versioned DHS variable registry
        ↓
codebook-backed semantic transformation
        ↓
household semantic measurement
        ↓
MeasurementContract + QA + RunManifest

This is an empirical semantic layer. It is not an experiment-role layer.

The fact that a variable has a reusable empirical meaning does not make it an outcome, treatment, control, moderator, or estimator input automatically.

Current registry scope

The initial registry is intentionally small and DHS-VII / Household Recode specific.

DHS source variable	Current empirical measurement	Important boundary
HV206	dhs.household.electricity_access	Standard 0/1 coding; documented missing remains missing; unsupported codes remain unresolved.
HV270	dhs.household.wealth_quintile	Source-defined ordered quintile; explicitly survey-relative rather than an absolute cross-survey wealth scale.
HV201	dhs.household.drinking_water_source_code	Preserves the source category code; no automatic improved/safe-water harmonization.

Each definition requires explicit codebook provenance and survey-phase compatibility.

The registry is not generated broadly from labels and does not revive meanings merely because historical notebooks used a variable in a particular way. In particular, unsupported old interpretations of variables such as HV215 or HV040 do not become current source truth.

Grain and identity

The derived durable product is materialized at:

source_row_id × measurement_id

source_row_id is the physical observation identity inherited from HR Silver. This is deliberate: source household IDs may be missing or duplicated and those anomalies must remain visible rather than forcing a false uniqueness contract.

The conceptual observation remains a household observation within a survey.

Missingness and unresolved values

The measurement layer keeps distinct states such as:

• observed;
• missing_source_value;
• source_missing_code;
• unmapped_source_code.

Unknown or unsupported codes never become zero or a convenient documented category.

This distinction is part of the measurement semantics, not an estimator-side cleanup choice.

Comparability and timing

The product carries explicit temporal semantics and comparability status from the variable definition.

Examples:

• electricity access can have a standard semantic definition while still requiring release-specific validation before L4 research authority;
• wealth quintile is survey-relative and should not be interpreted as an absolute wealth scale across surveys;
• detailed drinking-water categories may be country-specific and are therefore preserved without over-harmonization.

Contract and provenance

The materializer:

• consumes a content-hashed HR Silver DatasetRef;
• verifies the exact HR Parquet content hash before derivation;
• versions the output using the registry hash;
• writes a durable surveys.dhs.hr_household_measurements product;
• emits MeasurementContract objects for each registered measurement;
• persists the versioned registry and its provenance;
• records QA and RunManifest lineage.

The measurement contract contains no treatment/outcome/covariate vocabulary.

What this enables

The DHS scientific frontier has moved one step downstream.

A future experiment no longer has to begin by asking what HV206, HV270, or HV201 mean at all. It can consume a documented empirical measurement and then make a separate scientific choice about whether and how that measurement enters a design.

For example:

dhs.household.electricity_access
        ↓
experiment decides role + timing + sample + estimator

not:

HV206
        ↓
implicitly becomes outcome because an old notebook used it that way

What is still missing

This product does not establish:

• which DHS measurement is the primary outcome for FCV research;
• which measurement is a control or moderator;
• household exposure to investment or conflict;
• household → cluster → geography experiment projection;
• displacement-aware treatment linkage;
• survey-weight / PSU / strata estimator choices;
• cross-survey harmonization beyond each definition's declared comparability;
• protected real-source acceptance;
• L4 research validation.

Technical reference

• DHS_INTEGRATED_SUBSTRATE.md