SASIGNAL ATLASCross-industry intelligence / Research desk
SIGNAL ATLAS / RESEARCH DESK

Retail sales data is also a survey, just of firms not people

Census's own methodology page shows a third of monthly retail sales estimates are imputed, not reported.

The signal

Filed to Signal Atlas's method desk, this entry compares two ways of measuring the same behaviour: asking people what they did, and counting what a firm reports it sold. Pew Research Center's own account of why survey answers can be wrong states that "not all survey questions are answered accurately," and gives a documented example: more people report having voted in a past election than voting records show actually cast a ballot, because, in the Center's words, voting "is generally considered a socially desirable behavior" and such behaviors "are overreported," while socially undesirable behaviors are underreported.

The evidence

The natural alternative is a transaction-based series, such as the Census Bureau's quarterly retail e-commerce sales estimate. But the Bureau's own page describing how the underlying Monthly Retail Trade Survey is collected shows this is not a raw transaction feed either: it comes from a "stratified probability sample of approximately 13,000 employer firms," each "asked to report their sales and inventory data for the month just ending," with roughly a third of monthly sales estimates filled in by imputation where a firm does not respond. The contrast, then, is not "survey, therefore unreliable" against "transaction data, therefore true." It is a contrast between two different units of report, an individual describing past behaviour and a firm reporting aggregate sales, with two different failure modes: social desirability and memory on one side, non-response and imputation on the other.

Timeframe and confidence

Editorially, the Pew finding on overreported voting is well documented and long-standing, and the Census disclosure of imputation is a direct, dated methodological statement rather than an outside criticism. Confidence that both kinds of series carry a specific, named source of error is high; confidence that either kind is simply "the truth" against which the other should be checked is not supported by either document.

What would change the reading

Confidence in a specific figure would improve if the publishing body disclosed the imputation rate or the direction of self-report bias for that particular question, rather than for the method in general, since both biases described here vary by topic and respondent.

Neither a survey nor an official sales estimate is a direct photograph of behaviour. Each is a measurement with a documented, named source of error, and a trend claim should name which one it is using and what that error looks like.

Source trail

  1. Frequently Asked Questionswww.pewresearch.org · Source publication: not established · Retrieved 2026-09-16

    States that survey answers are not always accurate and documents the overreporting of voting as a socially desirable behaviour.

  2. Monthly Retail Trade - Monthly Retail Trade Survey Methodologywww.census.gov · Source publication: not established · Retrieved 2026-09-16

    Describes the Monthly Retail Trade Survey's sample of about 13,000 firms and states that roughly a third of monthly sales estimates are imputed.

  3. Quarterly Retail E-Commerce Sales Reportwww.census.gov · Source publication: not established · Retrieved 2026-09-16

    Illustrates the transaction-adjacent series built from the survey described above.

Event date
No single event
First source date
Unknown
Source-record publication
Not supplied — draft retained
Preparation
2026-09-16

Read across the evidence

Platforms & distribution · All 100 historical records