Styxis
Company
Knowledge

Can a labour-market net flow survive a smaller linked sample?

Assess missing wave links through population coverage and flow denominators, not a row-count check alone.

Why a small net flow can move sharply

An employment net flow is the difference between larger inflows and outflows. A modest reduction in linked records can therefore produce a much larger relative movement in that difference. A wave-link dataset is not adequately assessed because its total row count looks plausible. Even with Truthound, the team must first define whose links are missing and which flow is being measured.

The quality analysis ONS published

The ONS October 20, 2026 LFS quality update describes a May 2026 collection-operations error and roughly 4,000 fewer linked records. Its 10,000 simulations examine limited changes in broad patterns alongside reduced precision, particularly for net flows. They are not a reconstruction of identified missing records or corrected estimates.

Keep unlinked people visible

Retaining only successfully linked rows hides the characteristics of those who did not link. Use the earlier wave's eligible population as a reference and retain link status, observed or imputed status, and relevant review groups. Choose those groups within privacy access and the statistical design. Arbitrarily multiplying small subgroups can create more volatile comparisons.

Give the hypothesis a population

Make the hypothesis falsifiable: for example, that missing links were distributed similarly across the eligible population. Group-level linkage rates and earlier-period ranges can reveal evidence against that assumption. Similar rates do not prove an absence of non-response bias. A rule should not assume the true later status of people who were never successfully linked.

Definitions before grouped SQL

Truthound Core Python SDK 3.3.4, operating on a prepared Polars table, provides QueryGroupAggregateValidator for caller-supplied grouped SQL, expected groups and numeric bounds. CompletenessRatioValidator checks a specified minimum non-null ratio. Both need external definitions of eligibility, successful linkage, expected group ranges and the meaning of null. Otherwise the query can correctly validate the wrong population.

A pandas join answers a different question

An existing pandas 3.1.1 merge(validate=...) workflow remains useful for a required 1:1 relationship between wave keys. Uniqueness of joined keys and sufficient coverage of the eligible sample are different conclusions. Keep the pandas relationship check where it belongs and use a separate coverage table to explain the population left out of the join. pandas can match null keys on both sides; do not count a respondent without a key as a successful wave link.

An absent group has two meanings

No eligible people in a group is different from a group whose data never arrived. Do not turn a missing aggregate row into a 0% linkage rate automatically. Keep the expected group catalogue beside its observability state, so coverage checks do not invent failures for populations outside the actual sample.

What sensitivity analysis cannot restore

Show gross inflows and outflows before interpreting their net difference. This makes the amplification of a small residual visible. Sensitivity analysis explores possible effects; it does not restore missing observations. Removing links from an illustrative test dataset can examine a pipeline assumption, but cannot establish how missingness occurred in the actual survey.

Leave an interpretable decision

A useful hold decision should be explainable beyond a single linkage percentage. If one group's coverage changes, retain enough context for statistical owners to examine collection and definition changes. Rather than one quality-pass field, record the populations and flow interpretations for which the evidence remains adequate. That decision can be revisited when the next wave arrives.

Conditions

Define the eligible population, linkage states, review groups and permitted access before aggregation.

Alternatives

Keep pandas 3.1.1 1:1 join validation while assessing eligible-population coverage separately.

Limitations

  • Linkage checks and sensitivity analysis do not reconstruct missing observations or prove an absence of non-response bias.

Evidence

Event checked
Subject checked
Modified