Top of page
Skip to main content

Sampling and sample size: from the sampling frame to the number reported in your thesis

Sample size is more than a calculation. It depends on whether a sampling frame exists, the population to which conclusions should apply, expected response and the observations required by the analysis. This article works through those decisions using chapter 7 of Saunders et al.

Choose the sampling technique before calculating the sample size. Population inference requires a probability-based design and an appropriate sampling frame, with sample size informed by confidence, precision, population size and response. Without a frame, justify a non-probability sample against the analytical requirements and research purpose, rather than attaching a proportion-estimation formula to it.

Start reading

1. Population, target population and sample

Saunders et al. (2023) distinguish the population specified in the research question, the target population that can actually be studied, and the sample selected from that target population. Their example narrows all tablet owners to business students who own tablets at the researcher's university. This is practical, but extending findings to all tablet owners may be misleading.

Sampling is reasonable when a census is impractical because of population size, destructive data collection, cost or time. Careful sampling can free resources for instrument design, testing and following up difficult-to-reach cases. The essential criterion is whether the sample helps answer the research question (Saunders et al., 2023).

2. Probability and non-probability sampling

In probability sampling, selection probabilities are known and often equal, supporting statistical estimation for the target population. In non-probability sampling they are unknown. Researchers must therefore justify transferability or analytical generalisation carefully (Saunders et al., 2023). The Vietnamese edition (Saunders et al., 2010) uses the corresponding probability and non-probability terminology.

Table 1. Synthesised from Saunders et al. (2023, pp. 293–295).
CriterionProbability samplingNon-probability sampling
PrerequisiteAppropriate complete sampling frameNo frame, or probability sampling is inappropriate to the question
GeneralisationStatistical generalisation to the target populationAnalytical or theory-oriented generalisation
Common strategiesSurveys and experimentsCase studies, interviews and online-panel surveys
TechniquesSimple random, systematic, stratified and clusterQuota, purposive, volunteer (snowball or self-selection) and convenience

The textbook notes the growing use of non-probability samples with online questionnaires. Many Vietnamese thesis surveys distributed through Google Forms or social media fall in this category and should be described accordingly.

3. The sampling frame

A sampling frame lists cases in the target population, such as club membership or a payroll register. Probability selection depends on an appropriate frame (Saunders et al., 2023). Existing databases can be incomplete, inaccurate or outdated, so review the following questions.

  • Are listed cases relevant to the research topic?
  • Was the frame compiled recently and kept current?
  • Does it cover the target population?
  • Is the information accurate?
  • Are irrelevant cases excluded?
  • For purchased lists, can the provider's selection procedure be controlled?

Adapted from Checklist 7.1 in Saunders et al. (2023, p. 296).

4. Sample size for probability sampling

Saunders et al. (2023) describe a balance between precision and cost, affected by confidence, acceptable error, target-population size and the planned analyses, especially subgroup analysis. The eventual size combines calculation with judgement.

Business and management studies commonly use 95% confidence and errors of 3% to 5%. The following selected entries from the textbook assume responses from every sampled case.

Table 2. Selected entries from Table 7.1 in Saunders et al. (2023, p. 298).
Target population5% error3% error1% error
100799199
500217340475
1,000278516906
5,0003578793,288
10,0003709644,899
1,000,0003841,0669,513

The absolute number matters more than the sampling fraction. At 5% error, the table requires 384 cases for one million people. Citing De Vaus, the textbook uses this to explain why many market-research samples stop at about 2,000.

The calculation

Appendix 2 gives the minimum sample for estimating a proportion.

n = p% × q% × (z / e%)²

Here p% is the proportion of interest, q% its complement, z the confidence value (1.65 for 90%, 1.96 for 95%, 2.57 for 99%) and e% the permissible error. If p is unknown, 50% gives the conservative case. For a population below 10,000, the textbook applies the following adjustment.

n′ = n / (1 + n/N)

Its example uses p = 40%, q = 60%, 95% confidence and 5% error, yielding n = 369.6. With N = 4,000, the adjusted size is 338.46, rounded up to 339 (Saunders et al., 2023, pp. 805–806).

5. Response rates and invitations

The minimum assumes 100% response. Adjust invitations for the anticipated response rate.

nₐ = n × 100 / re%

Here n is the minimum or population-adjusted minimum, and re% the expected response percentage. The textbook example needs 439 responses and anticipates 30% response, so invitations are sent to about 1,463 customers (Saunders et al., 2023, p. 302).

≈ 50%

A reasonable individual-survey response benchmark, attributed to Baruch and Holtom in Saunders et al. (2023).

35–40%

A corresponding organisational-representative benchmark from the same source.

≈ 20%

Family-business survey response, attributed to Pielsticker and Hiebl in Saunders et al. (2023).

Response rates are declining, and recent business surveys by web or post may achieve only 10% to 20%. The textbook cites Neuman's distinction between total and active response rates.

Total response rate = responses / (sample size − ineligible cases)
Active response rate = responses / (sample size − (ineligible + unreachable cases))

For the example of 517 selected cases, nine ineligible cases, 311 contactable email addresses and 147 responses, the total rate is 28.9% and the active rate is 48.7%.

Using the classification attributed to AAPOR, the textbook distinguishes complete refusal (no answered questions), break-off (below 50%), partial response (50%–80%) and complete response (above 80%). Stating such rules makes decisions about usable questionnaires transparent.

6. Sample size for the planned analysis

The analytical technique can determine requirements for each subgroup. Saunders et al. (2023) cite Stutely's heuristic of at least 30 cases per subgroup and recommend collecting the whole subgroup if the population contains fewer than 30. They also note that large samples can make small effects statistically significant, reinforcing the need to report effect sizes.

For SEM, sample size relates to statistical power rather than solely to proportion-estimation error. Kock and Hadaya (2018) find the familiar PLS-SEM ten-times rule often inaccurate and propose inverse-square-root and gamma-exponential approaches evaluated by Monte Carlo simulation. At 5% significance and 80% power, the supplied article gives this inverse-square-root calculation.

nmin > (2.486 / |βmin|)²

βmin is the smallest path expected to be significant. For .20, the stated minimum is 155. For .15, it is 275.

Memon et al. (2020) review common survey-size heuristics and G*Power analysis and warn against using rules without considering their context and assumptions.

7. Four probability-sampling techniques

Table 3. Synthesised from Table 7.2 and section 7.6 in Saunders et al. (2023, pp. 304–311).
TechniqueProcedureFrame requirementConsideration
Simple randomNumber cases and select random numbers until the target is reached.Accurate and accessible listOften suitable for samples of several hundred or more
SystematicRandom starting point followed by a regular fraction, such as one in fiveAccurate list without problematic periodicityRepeated ordering, such as alternating genders, can undermine selection.
StratifiedDivide the frame into relevant strata and sample within each.Cases can be assigned to strata.Supports subgroup comparisons. Unequal response can require weighting.
ClusterSelect groups randomly and collect data within selected groups.List of clusters rather than every individualEfficient for dispersed populations, with reduced precision

Multistage designs combine techniques. The textbook example randomly selects counties in England and Wales, then electoral areas, then households systematically. Detailed final-stage frames are needed only for selected areas, reducing time and cost (Saunders et al., 2023).

8. Non-probability techniques

Quota sampling

Divide the population into categories, set quotas from available data such as a census, and recruit until each quota is met. Recruitment remains non-random despite attempts to reflect population composition. The textbook describes samples often ranging from 2,000 to 5,000 to support subgroup analysis. Probability-based confidence and sampling-error statements do not follow from quota recruitment. Online panels should report participation, usable responses divided by invitations (Saunders et al., 2023).

Purposive sampling

Use judgement to select cases that best address the question. The textbook lists extreme, intensity, maximum-variation, homogeneous, typical, critical, politically important, opportunistic and theoretical sampling, the latter used in grounded theory. Such samples do not establish statistical population representativeness.

Volunteer sampling

Snowball recruitment starts with a few people who introduce others. It can reach hard-to-identify populations but may produce homogeneous samples because people recruit similar contacts. Self-selection invites volunteers publicly and may attract people with strong views (Saunders et al., 2023).

Convenience sampling

Cases are selected because they are accessible, exposing findings to several sources of bias. An apparently convenient sample can sometimes meet a purposive rationale if selection is linked to the research objective. Using part-time MBA students as proxies for all managers, for example, requires cautious interpretation (Saunders et al., 2023, p. 325).

Table 4. Qualitative-interview guidelines synthesised from Saunders et al. (2023, pp. 314–315), drawing on Saunders (2012) and Saunders and Townsend (2016).
Interview contextIndicative sample
Homogeneous group4–12
Heterogeneous group12–30
One organisation or one analytical group30
Several organisations or analytical groups50
Common overall range15–60

In qualitative work, case selection is central. Many sources recommend continuing until saturation. Francis et al., cited in Saunders et al. (2023), suggest another three interviews after apparent saturation to check it.

9. Assessing representativeness

When census or personnel data exist, compare the sample with the population statistically. In the textbook example, the sample reflects tenure but overrepresents professional and managerial pay grades, a limitation that needs discussion. Other approaches include replication with a differently selected sample, surveying nonrespondents, interviewing them about nonresponse, and comparing early with late respondents (Saunders et al., 2023, pp. 312–313).

10. Reporting example, common errors and checklist

Example for non-probability sampling and SEM

The population was [description]. In the absence of a complete sampling frame, [convenience/quota/snowball] sampling was used with [screening criteria]. Findings are not used for statistical inference to the entire population. The minimum analytical sample followed the inverse-square-root method (Kock & Hadaya, 2018). An expected minimum path of [value] required at least [number] observations. Invitations were sent to [number] people through [channel] between [dates]. [Number] questionnaires were returned and [number] retained after excluding [number] with fewer than [threshold] answered questions or failed screening. Composition by [demographic variables] was compared with [population-data source, if available].
Table 5. Common sampling-reporting problems.
Weak practiceProblemImprovement
“Random sampling” for a Facebook linkNo probability selection from a frameDescribe self-selection or convenience sampling accurately.
5% error formula for convenience recruitmentSampling assumptions are unmet.Justify size against analytical requirements.
Only usable responses reportedResponse and exclusions are unclear.Report invitations, returns, usable cases and exclusion reasons.
Generalising to “Vietnamese consumers”Conclusions exceed the actual target population.Limit inference and discuss coverage.
Ten-times rule without justificationFrequently inaccurate estimatesUse inverse-square-root estimation or power analysis.

Before submission

  • Distinguish population, target population and sample.
  • Name the actual sampling method.
  • For probability sampling, describe and assess the frame.
  • Justify analytical size and anticipated response.
  • Report invitations, returns, usable cases, response rates and exclusions.
  • Assess representativeness or nonresponse bias where feasible.
  • Keep conclusions within the target population.

Conclusion

A defensible sampling account connects the target population, frame availability, selection technique, analytical requirement and achieved usable sample. A large number alone cannot replace that argument.

Non-probability sampling is common in theses. Accurate description, a reasoned sample-size justification and appropriately bounded conclusions allow readers to assess what the evidence supports.

References

Textbooks and methodological sources

  • Saunders, M., Lewis, P., & Thornhill, A. (2010). Phương pháp nghiên cứu trong kinh doanh (Nguyễn Văn Dung, dịch). Nhà xuất bản Tài chính.
  • Saunders, M. N. K., Lewis, P., & Thornhill, A. (2023). Research methods for business students (9th ed.). Pearson.

Additional sources

  • Kock, N., & Hadaya, P. (2018). Minimum sample size estimation in PLS-SEM: The inverse square root and gamma-exponential methods. Information Systems Journal, 28(1), 227–261. https://doi.org/10.1111/isj.12131
  • Memon, M. A., Ting, H., Cheah, J.-H., Thurasamy, R., Chuah, F., & Cham, T. H. (2020). Sample size for survey research: Review and recommendations. Journal of Applied Structural Equation Modeling, 4(2), i–xx. https://doi.org/10.47263/JASEM.4(2)01

This article synthesises the listed academic sources for methodological reference. Numerical guidelines should be interpreted in context and checked against institutional and target-journal requirements.