Sampling design is the bridge between a research question and the evidence used to answer it. Most business researchers cannot collect data from every customer, employee, household, supplier, or firm in a target market. They therefore select a smaller group and use its data to make carefully qualified statements about a larger population. The quality of those statements depends not only on sample size, but also on how the population is defined, how units are selected, how fieldwork is administered, and how error and nonresponse are controlled.

For MBA researchers, sampling is both a technical and managerial decision. A sample that is statistically elegant but operationally infeasible may be useless; a convenient sample that is inexpensive may produce seriously biased decisions. Sampling design must therefore align the research purpose, population, available frame, budget, timing, expected variability, and acceptable uncertainty.

Learning objectives

1. The population-to-sample logic

A target population is the complete set of people, organizations, transactions, or other units to which the researcher intends to generalize findings. For example, “all active customers of a bank in Kenya who used mobile banking during the last six months” is more useful than “bank customers,” because it specifies the substantive scope and eligibility criteria.

A sampling unit is the element selected at a particular stage of sampling. It may be an individual, household, branch, firm, classroom, invoice, or geographical area. The element is the ultimate unit about which information is collected. In a simple customer survey, the sampling unit and element may both be an individual customer. In cluster sampling, a branch may be selected first, while customers within branches are the elements.

A study population is the accessible portion of the target population from which the sample can actually be drawn. Differences between the target and study populations create coverage problems. A sampling frame is the operational list or system that identifies eligible units and permits selection—for example, a customer database, business registry, electoral register, employee list, or compiled list of retail outlets.

A frame is adequate when it is sufficiently complete, current, accurate, nonduplicative, and organized so that every eligible unit has a known or defensible chance of selection. Important frame defects include:

The difference between the target population and the population represented by the frame is called coverage error. It cannot be corrected merely by increasing sample size.

Example: defining a population precisely

Suppose a retailer wants to estimate satisfaction among “online customers.” A defensible specification might be: “individual customers aged 18 or above who completed at least one online purchase from the retailer between 1 January and 30 June 2025 and for whom a valid email address is available.” This definition establishes the unit, time period, eligibility rule, and frame requirement. It also reveals that customers without valid email addresses may be excluded, which must be acknowledged as a limitation.

2. Probability and non-probability sampling

The central distinction is whether selection probabilities are known and greater than zero.

In probability sampling, each eligible unit has a known, non-zero probability of selection. Randomization is used to select units. This permits estimation of sampling error and supports statistically defensible generalization to the frame population.

In non-probability sampling, selection probabilities are unknown. Units are selected through accessibility, judgment, referral, quotas, or other nonrandom mechanisms. Such samples can be valuable for exploratory, qualitative, pilot, or hard-to-reach research, but conventional margins of error cannot be interpreted as if the sample were a random probability sample.

Criterion	Probability sampling	Non-probability sampling
Selection mechanism	Randomized	Nonrandom or partly controlled
Selection probability	Known or calculable	Unknown
Sampling error estimation	Generally possible	Not defensible in the conventional sense
Cost and speed	Often higher/slower	Often lower/faster
Generalization	Stronger when coverage is good	More limited and conditional
Typical use	Population estimation and testing	Exploration, pilots, qualitative work

2.1 Simple random sampling

Under simple random sampling, every possible sample of a specified size has the same probability of selection. A random-number generator, rather than researcher choice, selects units from a complete frame.

Its strengths are conceptual simplicity and transparent inference. Its weaknesses include the need for a complete frame, possible high cost when units are geographically dispersed, and the possibility that small but important subgroups are poorly represented by chance.

2.2 Systematic sampling

In systematic sampling, the researcher selects every kth unit after a random start. If the frame has N units and the desired sample size is n, the approximate interval is:

If N = 10,000 and n = 400, k = 25. After randomly selecting a starting position between 1 and 25, every 25th unit is selected.

Systematic sampling is efficient and spreads the sample across the frame. It becomes dangerous when the frame has periodicity—for example, if every 25th record has a special pattern related to the outcome. Random ordering or inspection for periodic structure is therefore essential.

2.3 Stratified random sampling

Stratified sampling divides the population into mutually exclusive and collectively exhaustive subgroups, or strata, and draws a probability sample within each stratum. Useful strata may include region, customer type, firm size, or product category.

Stratification improves representation and can increase precision when units within each stratum are relatively similar and strata differ from one another. It is especially useful when management requires separate estimates for important segments.

When disproportionate sampling is used, analysis normally requires weights so that estimates represent the population rather than the achieved sample composition.

2.4 Cluster and multistage sampling

Cluster sampling selects naturally occurring groups, such as branches, schools, villages, or stores, and then surveys all or some elements within selected groups. It reduces travel and administrative costs, but units within the same cluster tend to resemble one another. Consequently, a cluster sample of 400 people may contain less information than a simple random sample of 400 independently dispersed people.

Multistage sampling applies sampling at two or more stages—for example, select regions, then branches within regions, then customers within branches. It is practical for large populations but requires careful probability calculations and often weighting.

A useful measure is the design effect, which describes the inflation in variance caused by a complex design relative to a simple random sample of the same size. For equal-sized clusters, an approximate relationship is:

where m is the average number of sampled elements per cluster and ρ is the intracluster correlation coefficient. If m = 20 and ρ = 0.04, deff ≈ 1 + 19(0.04) = 1.76. The effective sample size is therefore approximately n ÷ deff. Cluster designs should generally sample more clusters with fewer units per cluster when precision is important.

2.5 Common non-probability techniques

Convenience sampling selects readily accessible units, such as visitors to a website or students in a classroom. It is inexpensive but highly vulnerable to selection bias.

Purposive or judgmental sampling selects cases because they possess characteristics relevant to the research question. It is appropriate for expert interviews, critical cases, and qualitative inquiry when information richness is more important than statistical representation.

Quota sampling establishes target numbers for categories such as age or region and then fills each quota through nonrandom selection. It improves visible composition but does not remove selection bias within quotas.

Snowball or respondent-driven referral sampling asks participants to identify other eligible participants. It can reach hidden populations, but networks may be socially homogeneous. Referral incentives and recruitment chains must be governed carefully.

Non-probability samples should not automatically be dismissed. They may be the only feasible approach for undocumented workers, senior executives, early-stage exploratory research, or studies seeking mechanisms and meanings rather than population percentages. However, the researcher must state the limits of inference and avoid presenting a convenience sample as representative without evidence.

3. Determining sample size

Sample size is a design decision, not a universal number. It depends on the research objective, estimator, population variability, confidence level, precision requirement, design effect, expected response rate, and available resources.

For estimating a population proportion under simple random sampling, an initial large-population formula is:

where n₀ is the required completed sample, z is the standard-normal critical value for the selected confidence level, p is the anticipated population proportion, and e is the desired maximum margin of error expressed as a proportion. The formula assumes a probability sample, an approximately independent sample under simple random sampling, and a sufficiently large population.

When no credible prior estimate of p exists, p = 0.50 is conservative because it produces the largest required sample for a given z and e. For a 95% confidence level, z ≈ 1.96. With p = 0.50 and e = 0.05:

Thus, approximately 385 completed responses are required under the stated assumptions.

For a finite population of size N, the finite population correction may be applied:

If N = 2,000 and n₀ = 384.16, then n ≈ 322.6, so the researcher would normally target at least 323 completed responses, subject to rounding and design considerations.

where σ is the anticipated population standard deviation in the measurement unit and e is the desired absolute margin of error in that same unit. A prior study, pilot study, or defensible managerial estimate can provide σ.

These formulas do not automatically determine sample size for hypothesis tests, regression, experiments, or structural models. Such studies require power analysis based on the expected effect size, significance level, desired power, number of predictors or groups, and error structure. Researchers should avoid the rule that “larger is always better”: a very large biased sample remains biased, while an unnecessarily large sample can waste resources and increase operational error.

Adjusting for design and nonresponse

If a complex design has design effect deff, an approximate completed sample requirement is:

If the expected response rate is r, expressed as a decimal, the number of units to invite is:

For example, if n₀ = 385, deff = 1.5, and the expected response rate is 60%, then n_design = 577.5, and n_invited ≈ 963. The researcher should round upward and consider subgroup requirements, ineligible cases, and likely variation in response rates.

A margin of error is not a guarantee that every estimate is within that distance of the truth. It describes repeated-sampling uncertainty under the model and design assumptions. It does not include coverage error, nonresponse bias, measurement error, processing error, or interviewer effects.

4. Sampling error and nonsampling error

Sampling error is the difference between a sample estimate and the corresponding population value caused by observing a sample rather than a census. It is a random phenomenon that can generally be quantified under probability sampling. Increasing a well-designed sample usually reduces sampling error, but at a diminishing rate.

Nonsampling error includes all other errors: defective frames, ambiguous questions, interviewer influence, recording mistakes, coding errors, duplicate cases, data-entry problems, and nonresponse. These errors may be systematic and may not decline when sample size increases.

5. Nonresponse and response quality

Unit nonresponse occurs when a selected unit provides no usable data. Item nonresponse occurs when a participant completes the study but omits particular questions. Nonresponse creates bias when respondents and nonrespondents differ in relevant ways and the difference is not corrected by weighting or other design procedures.

A simple response rate can be reported as completed eligible interviews divided by eligible sampled units, but professional reports should state the exact definition used, including how unknown eligibility and partial interviews were treated. Response rates are useful process indicators, yet a high response rate does not guarantee absence of bias, and a lower rate does not prove severe bias.

Weighting can correct known composition differences, but it cannot reliably repair unmeasured differences. Researchers should report the achieved sample, response outcomes, weighting variables, and limitations transparently.

6. Fieldwork design and quality management

Fieldwork is the controlled implementation of the sampling plan. A fieldwork protocol should specify the population, frame, eligibility screen, selection rules, contact schedule, substitution policy, consent procedure, interviewer instructions, callback rules, data-security controls, quality checks, and escalation process.

Selection integrity

Substitution is a major risk. Replacing a difficult-to-contact selected person with an easy-to-contact volunteer changes the selection mechanism and can create bias. If replacement is unavoidable, it should be predetermined, documented, and consistent with the design. Interviewers should never choose replacements based on convenience or perceived cooperativeness.

Interviewer management

Training should cover the study purpose without encouraging leading behavior, eligibility screening, neutral probing, refusal conversion, confidentiality, safeguarding, and technical procedures. Supervisors can use live monitoring, back-checks, audio review where lawful and consented, and consistency checks. Quality control must be supportive rather than coercive; fabricated interviews and pressure to obtain desired answers are serious research misconduct.

Pilot testing

A pilot tests more than question wording. It should examine frame accuracy, recruitment, average completion time, screen-out rates, skip logic, translation, device compatibility, incentive administration, data exports, and respondent burden. Pilot findings may justify revising the frame, changing contact procedures, or recalculating the sample plan.

Digital and AI-enabled fieldwork

Online panels, mobile surveys, geolocation checks, paradata, and automated duplicate detection can improve operational control. AI may assist with routing, translation support, fraud detection, or identifying inconsistent responses. However, automated systems can exclude people with limited digital access, misclassify language or identity, and create opaque decisions. Human review, audit trails, data minimization, access controls, and disclosure of automated processing are essential.

7. Ethics, governance, and inclusion

Sampling decisions have ethical consequences because inclusion determines whose experiences influence managerial decisions. Researchers should justify exclusions, avoid unnecessary collection of sensitive data, obtain informed consent, protect confidentiality, and comply with applicable data-protection and research-ethics requirements. Participation must be voluntary, and incentives should compensate time without becoming undue influence.

Vulnerable groups may require additional safeguards, such as accessible formats, trusted intermediaries, translated materials, or procedures that protect privacy from employers or household members. Researchers should distinguish anonymity—where identity is not collected—from confidentiality—where identity is collected but protected.

8. Illustrative case: a bank’s mobile-app study

Background

A regional bank wants to estimate satisfaction with its mobile application among 120,000 active users and compare urban and rural customers.

Problem

The customer database includes location and recent activity, but 18% of records have missing or outdated contact details. A web link posted on social media would be fast, but it would overrepresent highly engaged digital users and would not support a defensible comparison.

Analysis

The target population is active mobile-app users aged 18 or above during the previous six months. The customer database is the frame, subject to coverage assessment. Because location is strategically important, the bank divides the frame into urban and rural strata. It anticipates p = 0.50, seeks 95% confidence and a 5% margin of error, and calculates n₀ ≈ 385. The design is stratified rather than clustered, so no major design-effect inflation is expected. Assuming a 55% response rate, the bank invites about 700–800 eligible customers, with invitations allocated by stratum and additional rural oversampling if a reliable rural estimate is required.

Approach

The bank removes duplicates, audits inactive accounts, and records missing-contact coverage. A random sample is selected within each stratum. Selected users receive an invitation, two reminders, and a telephone option for customers with limited internet access. The protocol prohibits replacing nonrespondents with volunteers. Response is monitored by stratum; final estimates are weighted back to the population distribution if disproportionate sampling is used.

Outcome

The bank obtains 430 completed responses, but rural response is lower than urban response. It reports weighted satisfaction estimates, confidence intervals appropriate to the design, response outcomes, and the possibility that customers without usable contact information remain undercovered. Managers use the findings to prioritize app reliability and rural support, but do not claim that the results represent all bank customers.

Learning points

The case shows that a defensible study requires more than a large number of responses. Clear eligibility, frame assessment, stratification, random selection, controlled follow-up, documented nonresponse, and cautious interpretation are jointly responsible for evidence quality.

Challenges and best practices

The most persistent challenges are incomplete frames, inaccessible populations, low response, budget constraints, complex designs, and pressure for rapid results. Best practice is to make trade-offs explicit rather than conceal them. Researchers should:

Chapter Summary

Sampling design determines how research evidence connects to a population. A target population states who or what the research concerns; a sampling frame operationalizes that population; and sampling units define what is selected. Probability methods—simple random, systematic, stratified, cluster, and multistage sampling—support known selection probabilities and more defensible statistical generalization. Non-probability methods—convenience, purposive, quota, and snowball sampling—can be appropriate for exploratory or hard-to-reach research but require cautious claims.

Sample size depends on confidence, precision, variability, population size, design effect, and expected response. The formulas for proportions and means are planning tools, not substitutes for design judgment. Sampling error differs from coverage, measurement, processing, and nonresponse errors. Fieldwork quality requires a controlled protocol, trained personnel, ethical recruitment, pilot testing, monitoring, and transparent reporting. In managerial research, a smaller well-designed sample is often more valuable than a larger biased one.

Key Terms

Target population: The complete population to which the researcher intends to generalize.

Study population: The accessible portion of the target population from which the sample is drawn.

Coverage error: Error caused by mismatch between the target population and the frame.

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

Non-probability sampling: Sampling in which selection probabilities are unknown.

Simple random sampling: Random selection in which every sample of a given size is equally likely.

Stratified sampling: Random sampling conducted separately within defined population strata.

Cluster sampling: Sampling groups of naturally occurring units rather than directly sampling all elements.

Design effect: Variance inflation caused by a complex design relative to simple random sampling.

Sampling error: Random difference between a sample estimate and its population value.

Nonsampling error: Error arising from coverage, measurement, nonresponse, processing, or other sources unrelated to sample selection alone.

Item nonresponse: Failure to answer one or more questions in an otherwise completed study.

Weighting: Applying adjustment factors so estimates reflect the population structure.

Fieldwork protocol: The documented procedures governing recruitment, data collection, supervision, and quality control.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

B. An operational list identifying eligible units from which a sample can be selected

Correct Answer: B. An operational list identifying eligible units from which a sample can be selected

Question 2

Correct Answer: C. Probability sampling has known, non-zero selection probabilities

Question 3

A population has N = 10,000 units and a systematic sample requires n = 400 units. What is the approximate sampling interval?

Question 4

Which method is most appropriate when important customer segments must each be represented through random selection?

Question 5

Using z = 1.96, p = 0.50, and e = 0.05, the approximate initial sample size for a population proportion is:

Question 6

If a design effect is 2.0, a completed sample requirement of 300 under simple random sampling becomes approximately:

Question 7

Question 8

Question 9

Correct Answer: C. Replacing hard-to-contact selected units with convenient volunteers

Question 10

What is the main limitation of a conventional margin of error calculated for a probability sample?

Correct Answer: C. It does not by itself account for coverage and nonsampling errors

Chapter 6 Sampling Design and Fieldwork

Sampling Design and Fieldwork

Introduction