7. Sampling, Access and Data Management

Introduction and Context

Research conclusions are never stronger than the relationship between the phenomenon being studied and the evidence actually obtained. Sampling determines whose experiences, behaviours, records, or opinions enter the study. Access determines whether those cases can be reached in an ethically and practically acceptable manner. Data management determines whether the resulting evidence remains accurate, secure, interpretable, and reproducible. These issues are therefore not administrative details added after research design; they are central components of validity.

In business and management research, the population of interest is often organisational rather than purely individual. A study may concern firms, business units, employees, managers, customers, suppliers, projects, transactions, or online communities. The researcher must consequently make several linked decisions: what constitutes a case, who can provide valid information, how cases will be identified, how many are needed, how access will be negotiated, and how data will be preserved and controlled. Sampling choices also depend on the research paradigm. Probability sampling supports statistical generalisation under defined assumptions, whereas purposive and other non-probability approaches may be more appropriate for explanatory, exploratory, interpretive, or hard-to-reach research.

This chapter treats sampling, access, and data management as an integrated design problem. A defensible sample that cannot be accessed is useless; accessible data that are poorly documented cannot support credible analysis; and a large dataset obtained through systematic non-response may produce less valid inference than a smaller, carefully designed sample.

Learning Objectives

By the end of this chapter, readers should be able to:

1. Distinguish populations, sampling frames, units of analysis, cases, and informants.

2. Select and justify probability and non-probability sampling strategies.

3. Estimate sample requirements using precision, power, design effect, and anticipated response.

4. Develop procedures for organisational access, recruitment, consent, and retention.

5. Diagnose non-response, attrition, coverage error, and selection bias.

6. Construct a secure, documented, and reproducible data-management plan.

7. Evaluate sampling and access decisions ethically and critically in doctoral research.

From Research Population to Analysable Case

The target population is the complete set to which the researcher intends to relate the findings. The accessible population is the portion that can realistically be contacted under the study’s resources and permissions. The distinction matters because conclusions may silently shift from “all medium-sized manufacturing firms” to “firms listed in one commercial directory.”

A sampling unit is the element selected at a particular stage of sampling. It may be an individual employee, a firm, a branch, a team, or a customer account. The unit of analysis is the entity about which a claim is made. These are not always identical. For example, a researcher may sample individual managers but analyse firms by aggregating their responses. Such aggregation requires theoretical and statistical justification, including sufficient within-firm agreement. A case is the bounded empirical entity examined in the study; an informant is a person who supplies information about that case. One firm may therefore be one case with several informants.

The sampling frame is the operational list or system from which selection occurs. It may be a membership register, employee database, customer file, business directory, voter-like register, transaction archive, or constructed list of organisations. A frame should be assessed for coverage, duplication, eligibility, currency, accessibility, and the possibility that inclusion itself is correlated with the research outcome. Coverage error occurs when the frame excludes eligible units, includes ineligible units, or represents some units more than once.

Defining eligibility before selection is essential. Inclusion criteria specify who or what qualifies; exclusion criteria prevent ambiguous or ethically inappropriate cases from entering the study. For longitudinal research, the researcher must also define the observation window, panel membership, and rules for replacing or retaining cases that cease to qualify.

Probability Sampling and Statistical Generalisation

Probability sampling gives each eligible unit a known, non-zero probability of selection. The probability need not be equal, but it must be known or estimable. This feature permits design-based estimation and, when assumptions are satisfied, supports generalisation from sample to population.

Simple random sampling selects units so that every possible sample of a specified size has the same probability. It is conceptually transparent but requires a complete frame and may be inefficient when the population is geographically dispersed or heterogeneous.

Systematic sampling selects every kth unit after a random starting point, where k is approximately the population size divided by the desired sample size. It is operationally convenient, but periodic patterns in the frame can create bias. A customer list ordered by purchase frequency, for instance, may produce a distorted sample if the interval aligns with a recurring pattern.

Stratified sampling divides the population into mutually exclusive strata and samples within each. Strata may be based on region, sector, firm size, seniority, or another substantively important characteristic. Proportionate allocation reflects population proportions; disproportionate allocation deliberately oversamples small or strategically important groups and requires weighting during analysis if population estimates are desired. Stratification can improve precision when strata are internally similar and meaningfully different from one another.

Cluster sampling selects naturally occurring groups, such as branches, schools, firms, or geographic areas, and then studies all or some units within selected clusters. It reduces travel and access costs, but units within a cluster tend to resemble one another. This intracluster correlation reduces the effective sample size and creates a design effect. A rough expression is:

Design effect = 1 + (average cluster size − 1) × intracluster correlation.

Multistage sampling combines stages, such as selecting regions, then firms, then employees. It is often the most realistic design for large organisational populations, but each stage introduces possible coverage and selection errors. Selection probabilities should be recorded so that weights can be calculated.

Probability sampling does not automatically guarantee representativeness. A low response rate, inaccurate frame, unequal attrition, or flawed weighting can still produce biased estimates. Its strength is transparent selection, not immunity from all forms of error.

Non-Probability Sampling and Deliberate Case Selection

Non-probability sampling is appropriate when a complete frame is unavailable, the population is hidden or specialised, the study seeks theoretical rather than statistical generalisation, or cases must possess particular experience. Its use should be justified by the research question rather than by convenience alone.

Convenience sampling selects accessible participants. It may be acceptable for pilot testing, instrument refinement, or early exploratory work, but claims should be limited because inclusion is likely related to availability and motivation.

Purposive sampling selects information-rich cases according to explicit criteria. Maximum-variation sampling seeks contrasting cases; homogeneous sampling examines a narrowly defined group; typical-case sampling explores ordinary instances; extreme or deviant-case sampling investigates unusual outcomes; and critical-case sampling selects cases whose logic is especially revealing. The sampling logic should be stated before recruitment and revised transparently if field knowledge changes it.

Quota sampling fills specified categories without random selection within categories. It can improve compositional diversity but does not remove selection bias. Snowball or respondent-driven approaches use participant networks to reach hidden populations. They may provide access where formal frames do not exist, although network dependence and homophily can narrow the resulting sample. Theoretical sampling, associated with grounded theory, proceeds iteratively: emerging concepts guide the selection of subsequent cases until additional data no longer substantially develop the theoretical categories.

Qualitative adequacy is not determined by a universal numerical rule. Researchers should consider information power, heterogeneity, case complexity, quality of interviews, analytic strategy, and the point of theoretical saturation or sufficiency. Saturation must be demonstrated analytically rather than asserted merely because a planned number of interviews has been completed.

Sample Size, Precision and Power

Sample size is a design decision linking the research question, estimator, acceptable uncertainty, population structure, analysis, and resources. Larger samples generally reduce random sampling error, but they cannot correct systematic bias. A large, poorly covered sample may be less credible than a smaller, well-designed one.

For estimating a proportion under simple random sampling, a commonly used initial expression is:

n₀ = z²p(1 − p) / e²,

where z is the critical value for the confidence level, p is the anticipated proportion, and e is the desired margin of error. When p is unknown, 0.50 produces the most conservative requirement. For a finite population, a correction may be applied:

n = n₀ / [1 + (n₀ − 1)/N].

These calculations assume a particular design and do not account for non-response. Cluster designs require inflation by the design effect; stratified designs may improve precision; and unequal selection probabilities require appropriate weighting.

In hypothesis-testing research, statistical power is the probability of detecting an effect of a specified size when that effect exists. Power depends on sample size, significance level, effect size, variance, and the analysis model. A power analysis should be conducted before data collection, using a substantively meaningful minimum effect rather than an optimistic expectation. Researchers should distinguish an a priori power analysis from a sensitivity analysis, which asks what effect size the available sample can detect. “Observed power” calculated after a non-significant result is generally uninformative and should not substitute for confidence intervals and effect estimates.

Sample-size planning must also account for clustered observations, multiple testing, covariates, subgroup comparisons, missing data, and anticipated attrition. In longitudinal studies, recruitment targets may need substantial inflation because the final analytic sample is determined by retention at the final wave, not enrolment alone. Researchers should report the basis of the calculation, assumptions, software or procedure used, and departures from the original plan.

Access as a Negotiated Research Relationship

Organisational access is both a practical and a political process. A gatekeeper may control entry but may not be the appropriate informant. Senior sponsorship can facilitate access while also creating perceived pressure on employees to participate. Access negotiations should therefore separate organisational permission from individual consent.

An effective access proposal explains the research purpose, eligibility requirements, time commitment, data requested, confidentiality arrangements, risks, expected outputs, and responsibilities of both parties. It should identify what the organisation will and will not receive. Promising individual-level findings or unrestricted access to raw data can create later ethical and analytical problems. Where findings may be commercially sensitive, the agreement should specify aggregation, anonymisation, publication review boundaries, and intellectual property without allowing the sponsor to suppress legitimate results.

A staged approach is often useful: establish a sponsor, conduct a feasibility conversation, obtain formal organisational permission, pilot recruitment, and then begin full collection. Researchers should maintain an access log recording contacts, permissions, amendments, refusals, and conditions. This record supports reflexivity and explains how the field setting shaped the evidence.

Recruitment communications should be neutral and avoid implying that participation affects employment, promotion, service access, or academic evaluation. Employees should be able to decline without explanation. Where the researcher has a managerial or supervisory relationship with participants, an independent recruiter or confidential platform is preferable.

Non-Response, Attrition and Fieldwork Adaptation

Unit non-response occurs when a selected unit provides no usable response; item non-response occurs when a respondent omits particular questions. Attrition is loss of participants across waves. These problems are not merely numerical. They may alter the composition of the sample if non-respondents differ systematically from respondents.

Researchers should distinguish refusal, ineligibility, inability to contact, partial completion, and technical failure. A contact protocol may include an advance notice, accessible invitation, reminders at planned intervals, alternative response modes, and a clear closing date. Incentives should be proportionate and not coercive. Burden can be reduced through concise instruments, mobile compatibility, flexible interview scheduling, and clear explanations of value and confidentiality.

Response rates should be defined using an established denominator and reported alongside the recruitment flow. Non-response analysis may compare early and late respondents, known frame characteristics, wave-specific retention, or auxiliary administrative variables. Weighting adjustments, imputation, sensitivity analysis, and pattern-mixture or selection models may be appropriate, but no statistical technique can fully repair unmeasured differences. In qualitative research, researchers should analyse whose voices are absent and whether recruitment channels systematically exclude particular perspectives.

Retention is strengthened by collecting only necessary contact information, confirming preferred communication channels, maintaining trust, providing reminders, and documenting changes in participants’ circumstances. Contact details must be stored separately from research responses whenever possible.

Data Management as Research Infrastructure

A data-management plan should be designed before collection and updated throughout the project. It identifies the data to be created, formats, ownership, storage, access rights, security controls, documentation, retention period, sharing conditions, and destruction procedures.

A reproducible workflow begins with a stable project structure. Raw data should be preserved as read-only; cleaned data should be generated through documented scripts or transformation logs; and analysis datasets should be distinguished from intermediate files and final outputs. File names should be systematic, versioned, and free from ambiguous labels such as “final_final.” A data dictionary should define variable names, labels, units, coding, missing-value conventions, derived measures, and permitted ranges. Qualitative projects require equivalent documentation for interview protocols, transcripts, coding frameworks, analytic memos, and case identifiers.

Data quality controls include range checks, duplicate detection, consistency checks, validation of skip patterns, audit samples, and comparison between source and entered data. Researchers should preserve a decision log explaining exclusions, recoding, merging, deviations from protocol, and changes to instruments. These records are essential for auditability and doctoral examination.

Security should be proportionate to sensitivity. Access should follow least-privilege principles, with strong authentication, encryption in transit and at rest, controlled sharing, secure backups, and tested recovery procedures. Identifiers should be separated from substantive data and replaced with pseudonyms or study IDs. Pseudonymisation reduces direct identifiability but is not equivalent to anonymisation if the linkage key remains available. Small-cell disclosure, distinctive quotations, organisational names, and combinations of seemingly harmless variables can re-identify participants.

Cloud platforms and automated transcription can improve efficiency but require scrutiny of jurisdiction, provider terms, retention, encryption, human review, and secondary use. Generative AI tools should not receive confidential raw data unless institutional policy, contractual safeguards, and informed consent permit it. If used for transcription, coding assistance, or synthetic-data testing, the researcher remains responsible for verification, provenance, and disclosure of the tool’s role.

Open science does not mean indiscriminate release. Data sharing should balance reproducibility with consent, confidentiality, commercial sensitivity, legal obligations, and participant safety. Where raw data cannot be shared, researchers can provide code, metadata, a de-identified subset, synthetic data, or controlled-access arrangements.

An Integrated Applied Case

Background

The following is an illustrative case. A doctoral researcher is studying whether digital-service capability improves resilience among medium-sized logistics firms. The proposed design combines a survey of firms with interviews involving operations managers and information-technology leaders.

Situation and Problem

The researcher initially proposes emailing a questionnaire to every firm listed in a commercial directory. The directory contains duplicates, excludes informal firms, and provides only general contact addresses. The proposed analysis requires firm-level responses, yet the instrument can be completed by any employee. The researcher also plans to interview “senior staff” without specifying the knowledge required.

Analysis

The target population is medium-sized logistics firms operating nationally during the study period. The case is the firm, while the informant is a knowledgeable manager. The directory is an imperfect frame, creating coverage error. A single respondent may provide useful information, but claims about firm capability require a role-based eligibility rule and, ideally, multiple informants or corroborating records. A power calculation based on independent individual responses would be inappropriate if the unit of analysis is the firm.

Approach

The researcher combines several business-register sources, removes duplicates, verifies eligibility through screening questions, and stratifies firms by region and turnover band. Within each stratum, firms are randomly selected. The recruitment letter is addressed to an operations or technology leader and explains that the organisation, not the individual, is the analytical unit. Selected firms may nominate a knowledgeable respondent, while a subsample is invited to provide a second informant.

The researcher conducts an a priori power analysis for the planned regression, inflates the recruitment target for expected non-response, and records selection probabilities. For interviews, purposive maximum-variation sampling selects firms with high, medium, and low reported capability, subject to sufficient information power. Organisational permission is obtained separately from employee consent. Survey identifiers are stored apart from responses, interview transcripts are pseudonymised, and all transformations are scripted.

A dashboard records invitations, eligibility, completion, refusal, and attrition. A comparison with frame variables shows lower response among smaller firms. The researcher applies transparent non-response adjustments, conducts sensitivity analyses, and qualifies generalisations accordingly.

Outcome

The final study cannot claim that every logistics firm was represented equally, but it can explain coverage, selection, response, and weighting decisions. The mixed evidence is stronger because the sampling unit, unit of analysis, informant competence, and data-management procedures are aligned.

Learning Points

The case demonstrates that sampling is inseparable from construct measurement and analysis. Access arrangements influence who responds; informant selection influences whether firm-level claims are credible; and documentation determines whether decisions can be audited. A defensible study is not defined by sample size alone but by coherence between population, frame, selection, measurement, analysis, and governance.

Ethical and Governance Considerations

Sampling may reproduce exclusion when databases omit precarious workers, small enterprises, minority groups, or people with limited digital access. Researchers should consider whether inclusion criteria are substantively necessary or merely convenient. Vulnerable participants require additional safeguards, accessible information, and careful assessment of coercion and confidentiality.

In organisational research, confidentiality can be difficult because a distinctive role, event, or quotation may identify a participant even after names are removed. Researchers should explain these limits honestly. Data protection obligations may govern lawful basis, purpose limitation, minimisation, retention, international transfer, breach response, and participant rights. Institutional review should cover recruitment, incentives, gatekeeper influence, recording, secondary use, and destruction.

Researchers also have a duty to report deviations. Replacing inaccessible firms with convenient alternatives, stopping recruitment after favourable responses, or deleting inconvenient cases without a pre-specified rationale threatens research integrity. Deviations may be unavoidable, but they must be documented, justified, and reflected in the interpretation.

Current Developments and Best Practice

Digital populations increasingly require hybrid frames combining administrative records, platform data, professional networks, and verified web sources. Adaptive and responsive designs can redirect effort toward under-represented groups during fieldwork, although changes to selection probabilities must be documented. Passive data and sensor-based research create new opportunities for longitudinal analysis but intensify consent, ownership, and re-identification concerns.

Contemporary best practice is characterised by alignment and transparency. Researchers should define the population and case before selecting participants; preserve a record of the frame and eligibility rules; justify probability or purposive selection; calculate sample requirements from the planned analysis; separate organisational permission from personal consent; monitor recruitment in real time; analyse missingness and attrition; protect identifiers; version data and code; and report limitations without overstating generalisability. These practices transform sampling and data management from procedural afterthoughts into sources of epistemic credibility.

Chapter Summary

Sampling specifies whose evidence enters a study, access determines whether eligible cases can be reached, and data management preserves the integrity and reproducibility of what is collected. Probability sampling enables design-based generalisation when frames, selection probabilities, response, and weighting are properly handled. Non-probability sampling can be rigorous when purposive, theoretical, or network-based selection is explicitly connected to the research question and claims are appropriately bounded. Sample size should reflect precision, statistical power, clustering, subgroup analysis, missingness, and attrition rather than a fixed convention. Organisational access requires negotiated permissions, independent consent, careful gatekeeper management, and transparent records. Finally, secure storage, documentation, version control, pseudonymisation, quality checks, and responsible sharing are essential components of doctoral research integrity.

Key Terms

Target population: The full set of entities to which the researcher intends to generalise.

Accessible population: The portion of the target population that can realistically be reached.

Sampling frame: The operational list or system from which cases are selected.

Sampling unit: The element selected at a particular stage of sampling.

Unit of analysis: The entity about which the research claim is made.

Case: A bounded empirical entity examined in a study.

Informant: A person who supplies information about a case or phenomenon.

Coverage error: Bias caused by omissions, duplications, or inappropriate inclusions in a sampling frame.

Probability sampling: Selection in which each eligible unit has a known, non-zero selection probability.

Stratified sampling: Sampling separately within defined population subgroups.

Cluster sampling: Selecting naturally occurring groups rather than isolated individuals.

Design effect: Inflation of variance caused especially by clustering or unequal selection.

Purposive sampling: Deliberate selection of cases according to research-relevant criteria.

Theoretical sampling: Iterative case selection guided by emerging theoretical categories.

Non-response: Failure of a selected unit to provide usable data.

Attrition: Loss of participants or cases across research waves.

Statistical power: Probability of detecting a specified effect when it exists.

Pseudonymisation: Replacing direct identifiers while retaining a separately protected linkage mechanism.

Data dictionary: Documentation defining variables, codes, units, and missing-value rules.

Reproducibility: The capacity for another researcher to understand and recreate an analytical workflow from documented materials.

Review Questions

Short Answer Questions

1. Distinguish the target population, accessible population, and sampling frame.

2. Why might the sampling unit differ from the unit of analysis?

3. What is the principal advantage of stratified sampling?

4. Define unit non-response and item non-response.

5. Why should raw data be preserved separately from cleaned data?

Long Answer Questions

1. Compare probability and non-probability sampling, explaining when each may be appropriate in doctoral business research.

2. Explain how clustering, non-response, and attrition affect sample-size planning.

3. Discuss the principal stages in negotiating ethical access to an organisation.

4. Explain how a data dictionary, decision log, version control, and access controls contribute to reproducibility.

5. Critically evaluate the claim that a large sample is necessarily a representative sample.

Analytical / Application Questions

1. A study samples employees from ten firms but makes claims about firm-level culture. Identify the design problems and propose a revised sampling and informant strategy.

2. A survey has strong overall participation but very low participation from small firms. What analyses and corrective actions should the researcher consider?

3. Design an integrated sampling, access, and data-management plan for a longitudinal study of digital transformation in medium-sized organisations.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a sampling unit from a unit of analysis?

A. The sampling unit is always an individual, whereas the unit of analysis is always an organisation.

B. The sampling unit is selected during sampling, whereas the unit of analysis is the entity about which claims are made.

C. The sampling unit is used only in qualitative research, whereas the unit of analysis is used only in quantitative research.

D. There is no meaningful distinction between the two concepts.

Correct Answer: B. The sampling unit is selected during sampling, whereas the unit of analysis is the entity about which claims are made.

Question 2

A researcher wants estimates for small regional subgroups that would be poorly represented in a proportionate sample. Which design is most appropriate?

A. Convenience sampling

B. Disproportionate stratified sampling with appropriate weighting

C. Unstructured snowball sampling

D. Systematic sampling without a frame

Correct Answer: B. Disproportionate stratified sampling with appropriate weighting

Question 3

Why can cluster sampling produce less statistical precision than simple random sampling of the same number of individuals?

A. Clusters eliminate all sampling error.

B. Individuals within clusters are often similar, reducing the effective independent sample size.

C. Cluster samples cannot be weighted.

D. Clusters always contain ineligible cases.

Correct Answer: B. Individuals within clusters are often similar, reducing the effective independent sample size.

Question 4

A directory excludes informal firms and contains duplicate entries. This primarily creates which problem?

A. Coverage error

B. Item non-response

C. Statistical power

D. Theoretical saturation

Correct Answer: A. Coverage error

Question 5

Which statement about statistical power is most accurate?

A. It is the probability that the null hypothesis is true.

B. It is determined only by the response rate.

C. It is the probability of detecting a specified effect if that effect exists.

D. It guarantees that a statistically significant result is practically important.

Correct Answer: C. It is the probability of detecting a specified effect if that effect exists.

Question 6

An employee is invited to participate by their direct supervisor. What is the principal ethical concern?

A. The sample automatically becomes probability-based.

B. The employee may perceive participation as compulsory or linked to employment consequences.

C. The response cannot be stored digitally.

D. The study must use cluster sampling.

Correct Answer: B. The employee may perceive participation as compulsory or linked to employment consequences.

Question 7

Which procedure best supports reproducibility?

A. Overwriting the original dataset after cleaning

B. Keeping raw data, transformation scripts, metadata, and decision logs in controlled versions

C. Relying on the researcher’s memory of recoding decisions

D. Storing all files under the name 'final'

Correct Answer: B. Keeping raw data, transformation scripts, metadata, and decision logs in controlled versions

Question 8

A participant leaves a three-wave study before the final wave. This is an example of:

A. Coverage error

B. Attrition

C. Stratification

D. Frame duplication

Correct Answer: B. Attrition

Question 9

Which approach is most suitable when cases are selected because they possess particular experience relevant to an emerging theory?

A. Theoretical sampling

B. Simple random sampling

C. Quota sampling without eligibility criteria

D. Systematic sampling

Correct Answer: A. Theoretical sampling

Question 10

Why is pseudonymisation not identical to anonymisation?

A. Pseudonymised data cannot be encrypted.

B. Pseudonymisation always makes data public.

C. A protected linkage key may still permit re-identification.

D. Anonymisation requires a larger sample.

Correct Answer: C. A protected linkage key may still permit re-identification.