Can records be sampled even when they were collected earlier?

Yes; that is the record-based sense highlighted by the OECD glossary.

Is every target unit equally likely to appear?

Not generally. Its links and the frame sampling design determine its opportunity to be represented.

AI Glossary / Statistics and probability / Intermediate

Indirect sampling

Definition

Indirect sampling obtains information about a target population through records or other linked units rather than sampling and observing the target units directly. The OECD glossary uses the term specifically for sampling documents or records. In modern survey methodology, it also describes sampling from a different, linked population and constructing estimates for the target population through those links.

What it means

The units available for selection may differ from the units about which a conclusion is wanted. A business might have a list of support tickets but want to describe customers. A statistical agency might have an establishment list but need enterprise-level estimates.
This distinction matters because one target unit can generate several records or have several links. Sampling records uniformly can give frequent record-generators more opportunities to appear. A sample that is random at the record level is not automatically an equal-probability sample of people or businesses.

How it works

State the target population, available frame, selected unit, and measurement unit separately. In the OECD's record-based sense, the selected material could be a completed form containing information already collected about a population member. That does not necessarily involve drawing a fresh sample of people for direct interviews.
In the linked-population framework, describe how a selectable frame unit connects to target units. Links can be one-to-one, one-to-many, or many-to-one. The design and estimator need to recognize this structure rather than treating every observed link as a different target unit.
A known probability design provides selection probabilities for frame units. Target-unit probabilities or suitable estimation weights must then be derived from the design and linkage structure. They are not generally obtained by copying a frame weight onto every discovered target record.
One approach is the generalized weight share method. At a conceptual level, target quantities are allocated across relevant frame links, and the resulting contributions are estimated using the sampling design. The allocation must preserve the intended target total under its assumptions.
The simple equal-sharing example below is a transparent special case. More complicated designs require their own formulas and variance calculations. It is not a universal instruction to divide every record by a duplicate count.
Coverage is essential. If a target unit has no link to any selectable frame unit, the design cannot discover it through that frame. Weighting the observed units more heavily does not automatically recover information about an entirely unrepresented subgroup.
Link errors also matter. A false match can attach data to the wrong target, while a missed link can distort multiplicity or coverage. Validate the linkage procedure and examine how uncertain matches affect conclusions.
Distinguish nonresponse from failure to identify a link. A selected unit might refuse to respond, or the relationship needed to reach a target might be missing. Each problem can require different adjustments and assumptions.
Retain enough design information for uncertainty estimation. Treating a linked sample as independent, equally weighted target observations can misstate precision. Document what population the final estimate covers and what limitations remain after weighting.

Three practical examples

An agency samples stored survey forms to produce an early estimate before processing every form.
A business samples transactions to investigate customers, explicitly accounting for customers with different transaction counts.
A survey team reaches enterprises through sampled establishments and derives weights using the relationships between those units.

Think of it this way

Think of selecting numbered doors that lead to rooms. Some rooms have several doors, while others have one or none. Randomly choosing doors does not give every room the same chance of being reached. You need the door-to-room map to interpret the resulting room sample.

Worked example

An original example has four selectable records. Records 1 and 2 both link to target A, whose value is 10. Record 3 links to B, whose value is 20, and record 4 links to C, whose value is 30. The target total is 60.
Select exactly one record uniformly, so each record has selection probability 1/4. Share A's value equally across its two links: each receives 5. B and C each have one link, receiving 20 and 30.
Estimate the total by multiplying the selected link's contribution by four. The four possible estimates are 20, 20, 80, and 120. Their equally weighted average is 60, confirming design unbiasedness in this toy design.
If the full target value were attached to every link without sharing, the possible estimates would be 40, 40, 80, and 120. Their average is 70, overstating the total because A is counted through two links.
The correct estimator is still highly variable when only one record is selected. Unbiasedness across repeated samples does not guarantee that an individual estimate is close to 60.

Illustrative case study

In a fictional case, Beacon Services wants to estimate customer-level workload from a ticket archive. Analysts initially treat sampled tickets as if each represents a different customer.
A review reveals that high-activity customers appear repeatedly. The team defines the customer population, checks whether customers without tickets belong in the target, and designs a weighting procedure around the record-to-customer links. It reports coverage and sampling uncertainty separately. The case illustrates why the target unit matters; it does not claim that a simple duplicate-removal step solves every linked-sampling problem.

Common misconceptions and mistakes

Indirect sampling does not necessarily mean nonrandom sampling. Conversely, random record selection does not make the target sample representative without appropriate reasoning.
Do not confuse an indirect sampling design with indirect measurement of a latent characteristic. Do not assume deduplication alone fixes unequal selection probabilities, missing links, or target units absent from the frame.

Business use

Indirect sampling can make existing administrative records useful when a complete target-unit list is unavailable or costly to construct. It can support audits and business surveys, but the record system's purpose may differ from the analytical objective. Budget for linkage validation, coverage assessment, design-aware weighting, and uncertainty estimation before promising population-level findings.

Check your understanding

Why does the naive worked-example estimator average 70 rather than 60?

Answer: Target A's value is represented twice through two frame records. Sharing its contribution across those links restores the correct expectation in this specific design.

Related terms

Record linkage
Coverage error
Sampling frame
Target population
Inclusion probability
Multiplicity
Generalized weight share method

Learn next

Next, study probability sampling and Horvitz-Thompson estimation. Then examine weight-sharing methods, linkage errors, and variance estimation for the particular linked-population design being considered.

Variance estimation

Sources

OECD Glossary of Statistical Terms, Indirect sampling and Direct sampling: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf Statistics Canada, Indirect sampling and the GWSM: https://www150.statcan.gc.ca/n1/pub/12-001-x/2013001/article/11829/section2-eng.htm Statistics Canada, Register-based sampling for household panels, Inclusion weights: https://www150.statcan.gc.ca/n1/pub/12-001-x/2016001/article/14544/03-eng.htm

Previous: Indicator function Next: Individual Bias