Secondary research

In depth

Official statistics

Official statistics are figures published by government statistical offices and agencies, such as population counts, business counts and household surveys. They are usually the largest and best-documented sources available, and their documentation is where to start. Three examples show what to read for.

Business counts. The U.S. Census Bureau’s County Business Patterns is an annual series of the number of establishments, their employment and their payroll by industry, for areas down to counties and ZIP codes. An establishment, in the Census Bureau’s definition, is a single physical location at which business is conducted: one supermarket company with forty stores is forty establishments. Employment is counted for the pay period that includes 12 March. And the series covers only establishments with paid employees; businesses with none appear in a separate product, Nonemployer Statistics, so a coffee cart run by its owner alone is invisible in County Business Patterns.

Population surveys. The Census Bureau’s American Community Survey publishes its full 1-year tables only for areas with 65,000 or more people; simplified 1-year supplemental tables reach areas of 20,000 or more; and 5-year estimates, built from 60 months of data, cover every area, down to census tracts and block groups. The Bureau’s guidance states the trade-off plainly: the 1-year estimates are the most current, while the 5-year estimates rest on the largest sample and are the most reliable. Both are sample estimates, so read them with their margins of error, which the Census Bureau publishes at the 90% confidence level rather than the 95% that B04 used. And compare like with like: a 1-year figure for a city against a 5-year figure for a town mixes two periods.

Household spending. The UK’s Office for National Statistics measures household spending with the Living Costs and Food Survey, a voluntary survey of private households in which members aged 16 and over record their daily spending for two weeks; its results appear in Family spending in the UK. Its quality guide is candid about the limits. The achieved sample in the financial year ending 2025 was 4,998 households, the response rate was 26% in Great Britain and 28% for the UK, spending is self-reported and may be under-reported (reported spending on tobacco and alcohol appears lower than other sources suggest), and communal establishments such as student halls of residence are left out.

None of this makes these sources poor. It makes them sources with documented definitions, coverage and accuracy, which is exactly what lets you judge how far they are from your question. A source that publishes nothing about its method is harder to use, not easier.

Trade associations and the trade press

A trade association is an organisation that represents the businesses in an industry, and many publish sales estimates, members’ surveys and market reports. For a niche, such as chilled coffee drinks in one city, an association or a trade magazine may be the only source. Read three things before using one: whose sales it covers (usually its members’, not the whole market), how it got the figure (a census of members, a sample, or an estimate), and why it published it (associations exist to promote their industry, so they have a reason to report growth).

Syndicated and panel data

OpenStax describes syndicated data as a type of external secondary data available through subscription services: a research firm collects the data once and sells it to many clients. Two kinds matter most for consumer goods. Retail measurement records sales at the checkouts of the stores that supply their scanner data, by brand, store type and week. A consumer panel is a sample of households or people who record what they buy over months or years, so it shows who buys, how often, what else they buy and when they switch brands.

Each has its own blind spot. Retail measurement misses stores that don’t share their data, and channels such as Lantern’s own online shop; panels depend on the households willing to record their shopping, which may differ from those who aren’t. A panel produces panel data in B01’s sense, the same units observed over time, which is why it can show switching that a single survey can’t. Syndicated data is often expensive, and its licence usually limits who may see it and what may be published.

Company filings

Companies that report to a securities regulator publish detailed filings, a free source on competitors and customers. In the US, the SEC’s EDGAR database gives free public access to them. The annual report on Form 10-K contains audited annual financial statements, a discussion of the material risk factors facing the company and management’s discussion and analysis of its results; Form 10-Q gives unaudited quarterly statements, and Form 8-K reports material events between them.

F00’s fictional Harbourline Roasters Inc., a listed roaster with $640 million of revenue, would file a 10-K whose risk factors and management’s discussion might say which product lines are growing; its revenue, though, would cover every product and every region together. In the UK, the Companies House register can be searched free of charge, without registering, for a company’s filing history, its officers and the dates of its accounts. Companies House checks that filings are complete and valid, but it says it makes only basic checks on accounts and cannot take responsibility for the consequences of errors or omissions in the register’s data.

Big data and digital traces

OpenStax’s Principles of Marketing describes big data as “the countless number of records that continues in an increasing capacity and at a faster rate”, and characterises it by three Vs: volume (how much there is), velocity (how fast it grows) and variety (how many kinds of data and contributors). For a marketer the main sources are digital traces of behaviour: web and app analytics (sessions, pages viewed, abandoned baskets, the web data of B01), transaction and loyalty-card records (transactional data linked to a customer), search trends (how interest in a term rises and falls) and social listening, which tracks what people say in public posts, reviews and forums. A business’s own analytics and loyalty-card records are internal data; search trends, public posts and data bought from platforms are external.

OpenStax places such data in a marketing information system, which a business uses to collect, analyse and report findings from its internal and external data; it calls the system an ever-changing database that needs marketers to act on what it holds, and among the names for that information it lists marketing intelligence, M01’s market intelligence. It also notes a difficulty: because the data keep changing, it is hard to settle on an approach to analysing them. Nearly all of these records are secondary data, created by running a website, a till or a social network rather than to answer a research question, so the six questions in the next section apply to them in full.

Volume isn’t representativeness. A million posts come from the people who chose to post, often the keenest and the angriest, plus bots, duplicate accounts and paid reviewers. However large it is, a social-listening count is a self-selected sample of opinion: it can show what is being said and how the conversation changes, not what share of customers think it. Loyalty-card data covers only card holders, and web analytics only the visits it records.

Definitions and platforms change without notice. A change in an analytics tool’s session settings, in how a platform ranks or filters posts, or in which reviews it shows can move a series overnight, the digital version of a series that jumps when its method changes. Record the settings, the query and the date with every extract, and check for changes before reading a trend.

Privacy limits use. Records that can be linked to a person are personal data, so using them for research needs a legal basis under the rules B01 outlined, and customers may not expect their purchases or posts to be studied. A public post is not consent to be profiled: analyse themes and counts, and never quote or report anyone in a way that identifies them.

Evaluating a source

Six questions cover almost every problem with a secondary source. The first two are about the source, the other four about how well its numbers fit your question.

The questions echo the data-quality dimensions of B01 and B02, timeliness among them, applied to a whole source rather than to values in a table. Write the answers down: a short source log with the source, its date, its definitions, its coverage and its link saves the next person, or you next year, from repeating the work.

Combining sources

Triangulation is checking a finding or an estimate against evidence from other sources or methods. Agreement raises confidence; disagreement is a clue, usually to a difference in definitions, coverage or timing that has to be explained before either figure is used. The method has four moves: put each figure on the same unit and period; adjust for what each covers, for example grossing up a members-only figure by the members’ share; compare; and explain each gap.

The result is usually a range, not a point. Two sources that bracket a figure from below and above don’t prove it, but they show it isn’t contradicted, and they tell you which assumption would have to change for it to be wrong. Report the range with what each end covers, and carry it into the next step of the research: market sizing (Module 3) starts from it.

The limits of secondary data

Secondary data can’t tell you how people will react to something that doesn’t exist yet: no statistics office has measured interest in Lantern Cold Brew. Its definitions may never match yours, its areas may be wrong for a business that serves one district, and the latest figures may describe a market that has since moved. And publishing someone else’s data has its own rules: check the licence, and cite the source.