Top of page
Skip to main content

Secondary data: where to find it and how to assess it for your dissertation

An existing dataset can save months of collection, but it was created for somebody else's question. Your task is to show that its measures, coverage and quality are suitable for yours.

Secondary data were collected for another purpose and are reanalysed for your study. They can save money and time, offer strong coverage and enable longitudinal work, but you cannot control the original measures, sample or changes in definitions. Assess overall suitability, detailed suitability, and costs against benefits before committing to a dataset.

Start reading

1. What are secondary data?

Saunders et al. (2023, p. 342) include both raw data and published summaries originally collected for another purpose. Once assessed as suitable, they can support new interpretations and conclusions. Statistical agencies, industry associations, news archives and organisational records offer alternatives to collecting a new survey.

The primary–secondary distinction depends on perspective. Schindler (2022, p. 57) uses annual reports to illustrate this: a report may be a primary documentary source for a company's official position, while its figures are secondary evidence assembled from underlying records. Internal KPI dashboards can similarly be reused.

A practical distinction

When you did not design the original collection, someone else selected the questions, sample and coding. Those decisions become constraints on your analysis.

2. Three groups of secondary data

Secondary evidence can be numerical or qualitative, structured in tables and databases or unstructured as texts, audio and images. Saunders et al. (2023, pp. 345–357) distinguish survey, documentary and multiple-source data. Their discussion, drawing on Taylor, also highlights the large volume of unstructured information.

Table 1. Secondary-data groups summarised from Saunders et al. (2023, pp. 345–357).
GroupSubgroupsExamplesMain consideration
SurveyCensuses; continuous or regular surveys; one-off surveysPopulation census, household living-standards surveys, business surveysUsually structured and quantitative; definitions, coverage and response must be examined
DocumentaryText; sound; images and videoMinutes, email, websites, reports, customer reviews, podcasts and videoInterpret within the creator's perspective and purpose
Multiple sourceCross-sectional, longitudinal or continuously updatedCompany financial databases, regional yearbooks, transaction data and big dataSomeone has already combined underlying sources

Censuses can offer broad coverage and detailed documentation; one-off surveys may be more topic-specific but harder to obtain at microdata level (Saunders et al., 2023, p. 346). Multiple-source evidence is increasingly important. The book describes big data through volume, velocity and variety, while cautioning that specialist software and computational requirements may make a resource impractical for a student project (pp. 357–358).

A single source rarely contains every required variable. Combining sources into an analytical dataset is therefore common and can take considerable time (Saunders et al., 2023, p. 350).

3. Why consider existing data?

Table 2. Benefits synthesised from Saunders et al. (2023, pp. 358–360) and Bryman and Bell (2019, pp. 123–126).
BenefitValue for a dissertationSource
Time and cost savingsMore time for theory, analysis and interpretation; sometimes the only feasible routeSaunders et al., p. 358
Potentially strong qualityEstablished sampling and quality-control procedures, sometimes national coverageBryman and Bell, p. 123
Subgroup analysisLarge samples can support groups a small new survey would missBryman and Bell, pp. 124–125
Longitudinal and international comparisonsAccess to multiple years or countries within a limited projectSaunders et al., p. 359; Bryman and Bell, p. 125
Non-reactive evidenceCollection has already occurred, often with de-identification; review requirements must still be checkedSaunders et al., p. 359
VerificationAccessible data allow others to examine the analysisSaunders et al., p. 360

A dataset is not exhausted by its first analysis: different relationships or analytical techniques can produce new understanding. Reusing evidence also respects participants' investment of time. Some funders require applicants to check whether suitable evidence already exists before financing new collection (Bryman and Bell, 2019, pp. 123, 126). Saunders et al. (2023, pp. 359–360), drawing on Dale and colleagues, use reanalysis of medical records to illustrate how evidence can answer questions beyond its original purpose.

4. Risks you must evaluate

The common difficulty is that the data were designed for another question (Saunders et al., 2023, pp. 360–363; Bryman and Bell, 2019, p. 127).

Table 3. Risks synthesised from the two textbooks.
RiskManifestationSource
Purpose mismatchOutdated information, inconsistent collection or omitted topicsSaunders et al., pp. 360–361
Access and costCommercial reports or custom tables require payment or permissionSaunders et al., p. 361
Aggregation and definitionsOnly grouped values are available, or questions change between wavesSaunders et al., p. 361
Limited quality controlNonresponse, misreporting and incomplete coverage affect even official figuresSaunders et al., p. 362; Bryman and Bell, p. 127
Selective presentationOrganisations highlight favourable information; online reviews may be manipulatedSaunders et al., pp. 362–363
ComplexityHundreds of variables, unfamiliar coding and household–individual structuresBryman and Bell, p. 127
Missing key variablesThe required concept is absent or measured inconsistentlyBryman and Bell, p. 127

Ecological fallacy

Do not use regional aggregates to infer individual behaviour. The unit represented by the data must match the inference being made. Robinson (1950) showed that aggregate correlations can differ markedly, even in direction, from individual-level correlations; Bryman and Bell (2019, p. 127) discuss the implications for secondary analysis.

A Vietnamese dissertation example

If your observations are provinces, your conclusions concern provinces. A relationship between provincial governance scores and provincial investment differs from a claim that firms rating government more positively invest more. The latter requires suitable firm-level evidence.

Official statistics may measure something different

Distinguish survey statistics from administrative records. Recorded violations reflect enforcement and reporting as well as underlying behaviour; an inspection campaign can increase recorded offences without an equivalent behavioural change. Unemployment indicators may exclude discouraged workers depending on the definition. Compare suitable representative-survey evidence where possible (Bryman and Bell, 2019, pp. 127–128).

5. Vietnam: changing definitions and administrative boundaries

The book's example of marital-status categories changing between UK censuses shows why apparently identical variables may not be comparable over time (Saunders et al., 2023, pp. 361–362). Boundary changes and revised indicators create similar issues in Vietnamese data.

34

Provinces and centrally governed cities covered by the PCI 2.0 framework described by VCCI in 2026.

137

Indicators: 111 from business surveys and 26 from official statistical data.

103

Inherited indicators, alongside 34 new indicators.

VCCI's announcement of 9 May 2026 described PCI 2.0 as scheduled for release on 15 May, with three pillars and ten subindices. A multi-year provincial analysis must therefore decide how earlier provinces map onto the new boundaries and which indicators remain comparable. A new composite index should not simply be appended to the old series without checking definitions.

The National Statistics Office's account of the Vietnam Household Living Standards Survey also illustrates changing content. It describes biennial even-year collection during 2002–2010 and annual collection during 2011–2024, with a narrower set of topics in odd years. A project requiring detailed housing or healthcare-expenditure variables must check the questionnaire for each wave rather than assume those variables exist in every year (National Statistics Office, 2026; Saunders et al., 2023, p. 373).

Before merging years

Obtain every wave's questionnaire and methodology. Build a crosswalk showing question wording, response options and geographic units for each analytical variable. Resolve and report mismatches in the methodology chapter.

6. Finding and locating data

Saunders et al. (2023, pp. 364–368) describe an iterative detective process.

Identify what you need

Derive variables or qualitative themes from the question, objectives and literature. For deductive quantitative work, list requirements as carefully as you would design a questionnaire.

Estimate whether it exists

Check data sources in published studies, then consult libraries, supervisors, statistical agencies and professional associations.

Retrieve the actual resource

Repositories may offer a direct download; organisational access may require lengthy negotiation.

Look outside your own industry. Environmental, legal, social and government sources can be especially valuable for business research (Schindler, 2022, p. 57).

7. Assessing suitability at three levels

Apply the same care as you would to primary evidence. Ask for a sample and documentation before committing, and stop if an essential level fails (Saunders et al., 2023, pp. 369–370).

Level 1: overall suitability

Measurement validity: does the information represent the required concept? Production output is not sales. Minutes may omit material a chair did not want recorded. Review how earlier studies have handled the same source and acknowledge unresolved limitations (Saunders et al., 2023, p. 371).

Coverage: does the resource cover the intended population, period and location? Will enough observations remain after excluding irrelevant cases? An unmeasured variable may be tolerable for some descriptive questions but seriously weaken an explanatory model (pp. 371–373).

Level 2: detailed suitability

Reliability: identify who is responsible, related publications, contact information and provenance. An academic-sounding institutional name is not independent evidence of quality (Saunders et al., 2023, p. 373).

Validity and credibility: examine who collected the evidence, sampling, response rates, sampling error, circumstances and instruments. Blogs or minutes may be evidence of the author's perspective rather than an objective record. Always check denominators: a 50% increase from two to three differs greatly from a 20% increase from 1,000 to 1,200 (pp. 374–376).

Measurement bias: look for deliberate distortion, changing collection practices and techniques that do not measure the intended topic. Independent sources can help assess consistency, though agreement alone does not remove every bias (pp. 376–377).

Level 3: costs and benefits

Consider money, search time, access negotiations and conversion or entry work. A spreadsheet may save considerable effort. Imperfect data may still answer part of the question if their limitations are compatible with the proposed inference; availability alone is not sufficient justification (Saunders et al., 2023, p. 377).

Schindler's five online-source criteria

Table 4. Paraphrased from Schindler (2022, pp. 62–63).
CriterionQuestions
PurposeWhy does the source exist, and how do its interests shape the information?
ScopeHow current, comprehensive and geographically relevant is it?
AuthorityWho created or sponsored it, with which expertise and underlying sources?
AudienceFor whom was it produced, and how does that affect presentation?
FormatCan the relevant material be located and downloaded in a usable form?

8. Ethics when reusing evidence

Public access does not automatically remove ethical review. Consider harm, expectations about research use and data management. Obtain permission when required for organisational or members-only material even if you are an employee or member (Saunders et al., 2023, pp. 359, 370–371).

The book discusses a student analysing a vehicle manufacturer's public Facebook page and deciding on anonymisation and face blurring. That context-specific decision is not general permission to reuse identifiable posts. In Vietnam, consider Law No. 91/2025/QH15 and institutional requirements. The research-ethics article develops this issue further.

9. Reporting template, common mistakes and checklist

Illustrative methodology paragraph

The study uses secondary data from [dataset], collected by [organisation] in [year] for [original purpose]. The original sample contained [n] [units], selected using [sampling method], with a response rate of [x%]. Applying [exclusion criteria] left [n′] analytical observations. Variable [A] is measured by [question and options]. Because no direct measure of [concept] is available, [proxy] is used following [author/year]. Between [waves], [wording/definition/boundary] changed; this was handled through [procedure]. Access was granted under [conditions], and the analytical file was de-identified.

Table 5. Common mistakes synthesised from Saunders et al. (2023) and Bryman and Bell (2019).
Weak approachProblemBetter approach
“Official data are reliable”Reputation does not establish measurement suitabilityExplain collection, sampling, measures and fit
Merging years without reviewing questionnairesDefinitions or boundaries may have changedCreate and report a wave-by-wave crosswalk
Drawing individual conclusions from provincial dataEcological fallacyMatch inference to the unit of analysis
Using an undisclosed proxyImplies direct measurementName and justify the proxy
Scraping posts because they are publicIgnores ethics and personal-data requirementsObtain review, protect identities and check applicable rules

Checklist

Condensed from Saunders et al. (2023, pp. 377–378).

  • Required concepts are measured directly or through justified proxies.
  • Population, geography, period and remaining sample size are appropriate.
  • Source, original purpose, audience and interests are understood.
  • Collection, sampling, errors and response are documented.
  • Definitions and measures have been compared across waves and against independent evidence.
  • Permission, ethics and de-identification requirements are addressed.
  • Access time, cost and format are proportionate to the benefit.

Conclusion

Secondary analysis has a different burden of justification: demonstrate that someone else's instrument, created for another purpose, measures what your question requires. Most weaknesses can be found before modelling by reading each wave's documentation. Revised Vietnamese administrative boundaries and indicators make that preparation particularly important.

References

Textbooks

  • Bryman, A., & Bell, E. (2019). Social research methods (5th Canadian ed.). Oxford University Press.
  • Saunders, M. N. K., Lewis, P., & Thornhill, A. (2023). Research methods for business students (9th ed.). Pearson. Pearson
  • Schindler, P. S. (2022). Business research methods (14th ed.). McGraw-Hill.

Papers, legislation and data sources

  • Cục Thống kê. (2026). Kết quả Khảo sát mức sống dân cư Việt Nam năm 2024. Bộ Tài chính. https://www.nso.gov.vn/?p=60506
  • Quốc hội. (2025). Luật Bảo vệ dữ liệu cá nhân số 91/2025/QH15. https://thuvienphapluat.vn
  • Robinson, W. S. (1950). Ecological correlations and the behavior of individuals. American Sociological Review, 15(3), 351–357. https://doi.org/10.2307/2087176
  • VCCI. (2026, 9 tháng 5). Lần đầu công bố xếp hạng năng lực cạnh tranh 34 tỉnh, thành phố. Liên đoàn Thương mại và Công nghiệp Việt Nam. https://www.vcci.com.vn

Textbooks are paraphrased with page references; secondary citations are identified. PCI and household-survey details refer to the linked VCCI and statistical-office publications. The reporting template and Vietnamese applications were written for this article.