How to Build a Sampling Frame Before Drawing Conclusions from Business Data

A sample can be selected carefully and still point to the wrong business reality.

The usual problem is not the random-number generator or the sample size. It is the list that came first. If the list leaves out relevant businesses, includes the wrong ones, duplicates others, or uses a different unit than the decision requires, the resulting sample cannot repair those flaws.

That list is the sampling frame: the operational set of units that had a chance to be selected. It is not automatically the same as the population you hope to describe.

Keep three terms separate:

• The target population is the full group your conclusion is meant to describe.
• The sampling frame is the list from which selection is actually possible.
• The sample is the subset selected from that list.

Random selection can support inference about the covered frame, subject to the design and other sources of error. It does not extend a finding to businesses that never appeared on the list.

Start with the claim, not the available file

Before assembling records, write one sentence describing the decision question:

> We want to estimate or compare ___ among ___ during ___.

This forces useful specificity. “How common is online booking among local salons?” needs definitions for local, salon, operating, and the period of interest. “What share of customers adopted feature X?” needs a definition of customer, adoption, and the date at which both are measured.

Words such as *all*, *typical*, *market*, *customers*, and *businesses* often conceal the most important choices. Replace them with rules someone else could apply consistently.

A good scope statement might read: “Active, public-facing hair salons with a physical location in the city limits as of June 30.” That is more restrictive than “local salons,” but it gives the analysis a visible boundary.

Decide what one row represents

A business can appear as a legal entity, parent company, operating location, account, contract, workspace, product, or transaction. None is universally correct.

The choice changes the question being answered. A procurement team examining spend concentration may reasonably use paid vendors from last year’s accounts-payable data. A team planning supply continuity may instead need every approved supplier, including those with no invoice during that year.

Official business statistics make a related distinction between a company and an establishment. An establishment is a single physical location where business is conducted; a company can own more than one establishment. [U.S. Census Bureau methodology](https://www.census.gov/programs-surveys/susb/technical-documentation/methodology.html) treats these as different concepts for a reason.

Write the unit into the frame specification. “One row per eligible operating location” is clearer than “one row per business.”

Turn eligibility into fields and rules

Build the frame using explicit inclusion and exclusion criteria. Depending on the decision, these may cover:

• geography or service area;
• industry or business category;
• operating status;
• size, revenue, employee, or transaction thresholds;
• relationship status, such as prospect, active customer, or approved supplier; and
• a reference date or reporting period.

Each rule should be traceable to a field, a documented review decision, or a stated limitation. If operating status comes from a directory label that is often stale, record that uncertainty rather than treating it as settled fact.

The reference date matters. A frame is a dated snapshot, not a permanent map of a business population. New businesses open, others close, ownership changes, and classifications lag. Business survey programs account for this by adding newly active businesses and removing identified inactive ones as their frames are maintained. [Census documentation for the Annual Survey of Manufactures](https://www.census.gov/programs-surveys/asm/technical-documentation/methodology.html) offers one example.

Audit the list before drawing from it

A practical frame audit does not need to certify that a list is perfect. It should make important weaknesses visible before they become conclusions.

First, inventory the source or sources. A CRM may represent known accounts. An invoicing system may represent vendors paid in a given period. A map directory may represent businesses that meet its listing and categorization rules. Each may be useful, but none automatically represents every relevant business.

Then check four common defects:

1. Undercoverage: eligible units are missing. 2. Overcoverage: records are included even though they are out of scope. 3. Duplicates: the same unit can be selected more than once. 4. Misclassification: a record is assigned the wrong category, status, location, or other eligibility attribute.

These are coverage issues, not minor cleanup inconveniences. Statistical guidance identifies omissions, erroneous inclusions, duplication, and misclassification as ways a frame can differ from its target population. [Statistics Canada](https://www150.statcan.gc.ca/n1/pub/12-539-x/2009001/coverage-couverture-eng.htm) also recommends comparing a frame, or part of it, with alternative sources where comparable sources exist.

Useful checks include counting records by key segment, inspecting a small random set for eligibility and duplicate entities, and comparing totals or segment mix with a second source. A comparison may expose a gap; agreement does not prove both sources lack the same blind spot.

If a directory likely misses home-based or newly opened salons, do not call the frame comprehensive. Add a complementary source if the overlap can be handled responsibly, narrow the claim to the directory-covered population, or describe the exercise as exploratory.

Freeze and document the frame

Once the rules are set and the audit is complete, preserve a version of the frame used for selection. Record the extraction date, source systems, eligibility rules, exclusions, unresolved cases, deduplication decisions, final record count, and sampling method.

A compact frame record can include: intended conclusion, target population, sampling unit, inclusion rules, exclusion rules, known gaps, frame date, final count, and selection method.

Only then is it time to choose a sample size and selection approach. A margin-of-error calculation applies to sampling error under a probability design; it does not erase coverage, measurement, processing, or nonresponse problems. [Statistics Canada’s survey guidance](https://www.statcan.gc.ca/en/customized-products-services/custom-surveys/questions-to-ask) makes that distinction explicit.

The final reporting rule is straightforward: state exactly who the findings can describe. If the frame covers active bill-to accounts linked to a workspace on a stated date, say that. Precision about the frame is not a weakness. It is what keeps a business conclusion from claiming more than the data can support.