Does a frame have to be a list of people?

No; maps, geographic areas, and registers of other units can serve.

Can two frames be used?

Yes, with suitable handling of coverage and overlap.

Does weighting always fix missing coverage?

No; it requires relevant information and assumptions about the missing units.

AI Glossary / Statistics and probability / Intermediate

Sample frame

Definition

A sample frame, more commonly called a sampling frame, is the list, map, register, or other operational device used to identify and access units for sample selection. It connects the intended survey population to the selection procedure. It is not necessarily identical to the target population: missing units, duplicates, outdated entries, or ineligible records can create coverage problems.

What it means

The population describes whom or what a study wants to learn about. The frame describes how the study can find and select those units. A company may want to study all active customers while drawing its sample from a billing register that omits some accounts or includes closed ones.
Frames can contain more than names. Contact details, geographic identifiers, size measures, and classification variables can support stratification, selection, follow-up, and weighting. Their quality affects the entire survey process.

How it works

Define the target population and sampling unit before selecting a frame. A business establishment, enterprise, individual, household, and transaction are different units. A register of enterprises may not directly support a study of individual establishments without additional mapping.
Choose a frame suitable for the task. List frames enumerate identifiable units. Area frames divide geographic space into units that can support multistage selection. A complete list of final respondents need not exist at the first stage of an area-based design.
Assess coverage against relevant external information. Look for missing eligible units, ineligible entries, duplicate representations, incorrect classifications, and stale contact details. The number of records alone cannot reveal whether the frame adequately covers the population.
Treat undercoverage and nonresponse separately. An eligible unit absent from the frame may have no route into the sample. A selected unit that cannot be contacted or declines to participate creates a later response problem. The remedies and implications differ.
Maintain unique identifiers and a clear record of the selection version. Business births, closures, mergers, address changes, and changes in eligibility can alter the frame between periods. Reproducibility requires knowing which version was used at the time of selection.
If several frames are combined, address overlap explicitly. A unit listed in more than one source can have several opportunities for selection. Deduplication or a valid multiple-frame estimation method is needed; simply concatenating lists can change inclusion probabilities.
Use frame information carefully in stratification and weighting. Misclassified units or outdated size measures can weaken precision or distort procedures. Calibration to known totals may help with some discrepancies, but it does not automatically recover information about systematically missing groups.
Document the gap between the frame population and the target population. When the gap cannot be resolved, narrow the supported claims or state the additional assumptions needed. For AI evaluation, a convenient collection of available inputs is a frame-like source, but it may cover only part of the intended deployment environment.

Three practical examples

Customer surveys: analysts select accounts from a maintained register after checking eligibility and duplicates.
Household research: an area frame supports staged selection of locations and then households.
Business AI monitoring: researchers use an establishment register and check whether new digital-service firms are adequately included.

Think of it this way

Think of the frame as the address book used to send invitations. The guest population is the set of people you intend to reach; the address book is the tool for reaching them. A perfect random draw from an incomplete address book still cannot select a person whose entry is missing.

Worked example

A fictional target population contains 10,000 active establishments. A frame contains one record for each of 9,000 eligible establishments, plus 300 ineligible records and 200 extra duplicate records. The frame therefore has 9,500 records, but it covers only 9,000 distinct eligible units, or 90% of the target population.
If 950 records are selected by simple random sampling without replacement, each record has probability 950/9,500 = 0.10 of selection. That does not give every target establishment a 10% inclusion probability.
An omitted establishment has probability zero. An eligible establishment represented once has probability 0.10. One represented by two records has probability 1−[(8,550×8,549)/(9,500×9,499)], approximately 0.190, of at least one record being selected.
The example shows why record count, population coverage, and unit inclusion probability are different quantities. Frame repair or an appropriate design is needed before claiming equal-probability selection of establishments.

Illustrative case study

Fictional case: Brookfield Research prepares a survey of business AI use from a commercial contact list. A quality review finds duplicate branches, inactive firms, and poor coverage of recently formed enterprises.
The team updates identifiers and eligibility, supplements coverage using a documented source, and records the remaining limitations. It avoids presenting a random draw from the original list as a representative survey of every business in the market.

Common misconceptions and mistakes

A frame is not simply another word for the population.
A long list is not necessarily a complete or current list.
Random selection does not remove undercoverage because units outside the frame remain unavailable for selection.
Duplicates can affect probabilities as well as increase administrative work.
A census of every frame record does not automatically constitute a census of the intended population. Ineligible entries and missing units still matter even when no sampling is performed.

Business use

Frame quality determines whether the people or organizations relevant to a decision can enter the evidence base. Invest in maintenance, eligibility rules, and coverage assessment before increasing the survey budget. When an AI system is evaluated using a restricted source of cases, explain which deployment situations that source can and cannot represent.

Check your understanding

Answer: A survey randomly selects from a register that excludes all new firms.

Can new firms be represented directly?

Answer: No; they have no selection route through that register.

Why can two records for one establishment cause a problem?

Answer: They can give it more opportunities to be selected than a singly listed establishment.

Related terms

Nonresponse
Target population
frame population
undercoverage
overcoverage
sampling unit
inclusion probability
multiple-frame sampling

Learn next

Compare a proposed survey frame with the population definition. Count and investigate duplicates, stale entries, and known omissions before designing the selection and estimation steps.

Sources

Statistics Canada, Quality Guidelines, Coverage and frames: https://www150.statcan.gc.ca/n1/pub/12-539-x/2009001/coverage-couverture-eng.htm The coverage and duplicate-selection example and case are original fictional illustrations.

Previous: Sample covariance matrix Next: Sample mean