Marketing research rarely permits observation of every member of a target market. A retailer may want to understand all current customers, a bank may study all small-business clients, and a public agency may assess every household in a region. A census can be expensive, slow, and administratively difficult. Sampling provides a disciplined way to draw conclusions about a population from a subset of its members.

Sampling is not simply the selection of a convenient number of respondents. It is a chain of decisions involving the definition of the population, identification of a sampling frame, selection of sampling units, choice of a sampling design, determination of sample size, implementation, weighting, and evaluation of coverage and nonresponse. Errors introduced at any stage can threaten validity even when the final sample is large.

For graduate-level marketing research, the central principle is that sample quality depends more on design and execution than on size alone. A small probability sample with excellent coverage may be more defensible than a large opt-in sample with unknown selection mechanisms.

Learning Objectives

1. The Logic of Sampling

The target population is the complete set of units about which the researcher wishes to make claims. The survey population is the portion of the target population that the research procedures can actually reach. A sampling frame is the operational list, database, map, panel, or other mechanism from which the sample is selected.

A population may be defined by people, households, organizations, transactions, stores, advertisements, or time periods. Definitions should specify content, geographic scope, time reference, and eligibility. “Our customers” is insufficient unless the researcher specifies, for example, customers who made at least one purchase online in the last six months and reside in the country of interest.

An element is the basic entity about which information is collected. A sampling unit is the entity selected at a particular stage. In a household survey, a country may be divided into areas, areas into households, and households into adults. The household is a primary sampling unit at one stage, while the adult is the ultimate sampling unit. The observation unit is the entity represented in the data record; it may differ from the sampling unit.

A useful distinction is between the parameter, a population characteristic such as the true proportion preferring a brand, and the statistic, its estimate calculated from sample data. Sampling error is the random difference between a statistic and its population parameter under repeated samples. Nonsampling errors include coverage error, nonresponse, measurement error, processing error, and specification error. Increasing sample size primarily reduces random sampling error; it does not automatically correct nonsampling error.

Sampling frames and frame quality

A high-quality frame is current, sufficiently complete, nonduplicative, and contains information needed for selection and contact. Common frame defects include:

Frame construction is often a substantive research task. A customer relationship management database may omit cash purchasers; a business directory may exclude informal firms; and a social-media platform may overrepresent highly connected users.

2. Probability and Non-Probability Sampling

In probability sampling, every eligible frame unit has a known, nonzero probability of selection. This permits design-based estimates of sampling error and supports defensible population inference, provided the frame and implementation are adequate. In non-probability sampling, selection probabilities are unknown or zero for some units. Such samples can be useful for exploratory research, rare-access populations, qualitative work, rapid testing, and situations without a usable frame, but conventional margins of error should not be presented as if they were valid probability-sample estimates.

Method	Main procedure	Strengths	Main limitations
Simple random sampling	Select units randomly from a complete frame	Transparent; straightforward theory	Requires a complete frame; may be costly geographically
Systematic sampling	Select every kth unit after a random start	Efficient and easy to administer	Periodicity in the frame can bias results
Stratified sampling	Divide frame into strata and sample within each	Ensures subgroup representation; improves precision when strata are homogeneous	Requires suitable classification information
Cluster sampling	Select groups, then units within groups	Reduces travel and contact cost	Intracluster similarity increases variance
Multistage sampling	Sample successively at several levels	Practical for dispersed populations	More complex weighting and variance estimation
Convenience sampling	Recruit readily available units	Fast and inexpensive	Unknown selection bias; weak generalizability
Quota sampling	Fill subgroup quotas using nonrandom recruitment	Controls visible composition; operationally fast	Within-quota selection remains uncontrolled
Purposive sampling	Select cases judged information-rich	Useful for expert, critical, or rare cases	Depends on researcher judgment; not design-unbiased
Snowball or respondent-driven recruitment	Participants recruit others	Reaches hidden or networked populations	Network dependence and recruitment bias
Opt-in online panel	Volunteers join and answer invitations	Efficient digital fieldwork	Self-selection, panel conditioning, uncertain coverage

Probability designs in greater depth

Simple random sampling gives each unit the same selection probability. It is conceptually clean but can be inefficient when the population is geographically dispersed or contains important small subgroups.

Systematic sampling selects every kth frame unit after a random start. If population size is N and desired sample size is n, the approximate interval is k = N/n. The interval should be interpreted carefully when the frame has periodic ordering, such as alternating high- and low-value customers.

Stratified sampling divides the population into mutually exclusive and collectively exhaustive strata, such as region, customer tenure, or firm size. Proportional allocation samples in proportion to stratum size. Disproportionate allocation deliberately oversamples small or strategically important groups, followed by weighting. Stratification improves precision when units within strata are relatively similar on the study variable and strata differ from one another.

Cluster sampling selects naturally occurring groups, such as stores, postal areas, or corporate branches. It is economical when travel or contact costs dominate, but members of the same cluster often resemble one another. Consequently, the effective information is less than the raw respondent count suggests.

Multistage sampling combines these approaches—for example, select regions, then stores, then customers. Each stage requires explicit selection probabilities and appropriate variance estimation.

3. Determining Sample Size

Sample size is a design decision, not a universal constant. It depends on the study objective, estimator, desired precision, confidence level, population variability, design, subgroup requirements, expected response, cost, and analysis plan.

Estimating a proportion

For a simple random sample estimating a population proportion, the approximate sample size for a specified margin of error is:

Here, n₀ is the initial sample size, z is the standard-normal critical value corresponding to the confidence level, p is the anticipated population proportion, and e is the desired absolute margin of error. The formula assumes a probability sample, a sufficiently large population, an approximately unbiased estimator, and a simple random design. The margin of error and p are proportions, not percentages.

If no credible prior estimate of p exists, p = 0.50 is conservative because p(1 − p) is then maximized. For 95% confidence, z ≈ 1.96. With e = 0.05:

The researcher rounds upward to 385 completed cases. This is not a guarantee that every estimate in the study will have a five-point margin of error; subgroup estimates, complex designs, and nonresponse can require more.

For a finite population of size N, the finite-population correction can be applied as:

This correction is material when the sample is a substantial fraction of the population. If N = 2,000 and n₀ = 385, then n ≈ 323. The correction should not be used to excuse a poor frame or nonprobability recruitment.

Estimating a mean

For a continuous measure such as monthly expenditure, the corresponding approximate formula is:

where σ is the anticipated population standard deviation in the measurement units and e is the desired absolute precision in those same units. A pilot study, prior research, or administrative data may provide σ. If the measure is highly skewed, researchers should consider transformation, robust estimators, quantile precision, or simulation rather than relying mechanically on a normal approximation.

Design effect and effective sample size

Complex designs often produce greater variance than simple random sampling. The design effect is:

where b is the average number of sampled units per cluster and ρ is the intracluster correlation coefficient. The formula assumes a reasonably homogeneous cluster size and is an approximation. If b = 20 and ρ = 0.03, deff ≈ 1.57. The required completed sample becomes approximately 385 × 1.57 ≈ 605, before adjustment for nonresponse.

The effective sample size summarizes the information equivalent to a simple random sample:

Weight variation can also inflate variance. A commonly used approximation for unequal weights is:

where CV(w) is the coefficient of variation of the weights, defined as their standard deviation divided by their mean. This approximation is useful for planning but should not replace a survey-specific variance estimator.

Nonresponse inflation

If the expected response rate is r, expressed as a proportion, the number of cases to contact is approximately:

For 605 completed interviews and an expected response rate of 60%, n_contact = 605 / 0.60 ≈ 1,009 invitations. This calculation assumes that respondents and nonrespondents do not differ in ways that affect the estimate after available adjustments. Response-rate inflation increases completed cases; it does not eliminate nonresponse bias.

Power for hypothesis tests

When the primary objective is testing a difference rather than estimating a proportion to a chosen margin of error, sample size should be based on statistical power. Power is the probability of rejecting a null hypothesis when a specified alternative is true. A two-group comparison generally requires the anticipated effect size, significance level α, desired power 1 − β, outcome variability, allocation ratio, and design effect.

For two independent means with equal group sizes and common standard deviation σ, a planning approximation for the number per group is:

where Δ is the smallest practically important difference. This formula assumes independent observations, approximately normal means, equal variance, equal allocation, and a two-sided test. For proportions, binary regression, repeated measures, interaction effects, and complex surveys, software or simulation is generally preferable. Researchers must distinguish statistical significance from managerial importance: a trivial effect can be statistically significant in a very large sample.

Cost and optimization

Sample-size planning should consider the cost of each stage. If fixed cost is C₀ and variable cost per completed case is c, total cost is approximately:

In cluster designs, the relevant decision is often how many clusters and how many units per cluster to sample. More units within the same cluster may be cheaper but less informative than adding clusters. A cost-efficient design balances precision, travel or contact costs, subgroup needs, and operational risk. In practice, researchers should prepare scenarios rather than claim a single mathematically optimal sample size based on uncertain cost and variance assumptions.

4. Weighting and Estimation

A survey weight represents how much population information a sampled unit stands for. The base weight is the inverse of the unit’s selection probability:

where πᵢ is the inclusion probability for unit i. In multistage designs, πᵢ is the product of conditional selection probabilities across stages.

Weights may then be adjusted for nonresponse and calibrated or post-stratified to known population totals. A weighted estimate of a mean is commonly expressed as:

where yᵢ is the observed value and wᵢ is the final analysis weight. Weighted estimates can reduce coverage and composition bias, but weighting may increase variance, especially when weights are highly unequal. Analysts should document weight construction, trimming rules, calibration variables, and variance estimation procedures.

Standard errors must reflect stratification, clustering, and weights. Ordinary formulas that treat all observations as independent simple-random observations can substantially understate uncertainty. Survey-capable statistical software, replicate weights, Taylor linearization, or appropriate resampling methods should be used.

5. Coverage Error, Nonresponse, and Data Quality

Coverage error occurs when the frame differs from the target population. It has two components: omission and erroneous inclusion. Researchers should compare frame counts with external benchmarks, deduplicate records, identify unreachable units, and consider dual-frame or area-based solutions where no single frame is complete.

Unit nonresponse occurs when a selected unit does not participate. Item nonresponse occurs when a participating respondent omits particular questions. Nonresponse bias arises when responding and nonresponding units differ on variables relevant to the estimate and the difference is not corrected by design or adjustment.

A response rate is operationally useful but not itself a complete measure of bias. A high response rate can coexist with coverage problems, while a lower rate may produce little bias if respondents and nonrespondents are similar on relevant variables. Diagnostics include comparisons with frame data, paradata such as call attempts and contact mode, response propensity models, early–late comparisons, and sensitivity analyses under alternative assumptions.

Effective practices include multiple contact attempts, varied contact times, clear invitations, mobile-friendly instruments, reasonable incentives, language accessibility, interviewer training, and refusal conversion where ethically appropriate. Incentives should compensate time without becoming coercive, particularly for financially vulnerable participants.

6. Choosing Among Designs: A Decision Framework

The design should follow the inference required. If the client needs defensible population estimates and subgroup comparisons, a probability design is generally preferred. If the objective is concept generation, usability testing, expert insight, or exploration of a difficult-to-frame population, purposive or other non-probability methods may be appropriate. Mixed designs can combine a probability survey with purposive qualitative interviews, but findings must not be merged as though they have identical inferential status.

7. Illustrative Case: National Customer Experience Survey

Background and problem

A fictional national telecommunications company wants to estimate the proportion of active customers who would recommend its service. Management requires a national estimate and reliable comparisons across four regions and two customer segments: prepaid and contract customers. The customer database contains 4 million records, but it includes duplicate accounts, recently disconnected customers, and customers without valid contact permissions.

Analysis and approach

The research team first defines an eligible customer as an individual account with an active service at the frame-extraction date. Records are deduplicated and checked for eligibility. Because regional and segment comparisons are required, the team creates eight strata: four regions × two segments. A stratified probability sample is selected from the cleaned frame, with disproportionate selection of smaller segments to ensure analytical precision.

A pilot suggests p = 0.50, a 95% confidence level, and a five-percentage-point national margin of error. The simple-random requirement is approximately 385 completes. Anticipated clustering and weighting imply a planning design effect of 1.5, producing approximately 578 completes. The team increases this to 640 to protect subgroup estimates and expects a 64% response rate, requiring approximately 1,000 invitations.

Selection probabilities and nonresponse adjustments are recorded. Final weights are calibrated to known regional and segment totals. The analysis uses survey procedures that account for stratification and weights. The report provides national estimates, regional confidence intervals, response outcomes, frame exclusions, and sensitivity analyses.

Outcome and learning points

The survey produces a precise national estimate, but one region has substantially lower response among prepaid customers. The team does not simply report the raw overall proportion. It examines response patterns, applies documented adjustments, compares weighted estimates with administrative indicators, and presents uncertainty and limitations. The case demonstrates that sample-size arithmetic cannot compensate for a deficient frame, and that subgroup objectives must be incorporated before fieldwork begins.

8. Ethics, Governance, and Contemporary Practice

Sampling decisions affect whose experiences become visible in managerial decisions. Excluding people without internet access, omitting minority-language participants, or using a customer list without appropriate consent can create both methodological and ethical harm. Researchers should minimize personal-data collection, establish lawful processing and retention procedures, protect contact information, and separate survey responses from identifying data where possible.

Online panels, passive behavioral data, mobile location data, and AI-assisted recruitment can improve speed but introduce opaque selection mechanisms, algorithmic exclusion, bot activity, duplicate respondents, and privacy risks. AI may assist frame deduplication, fraud detection, or response-propensity modeling, but model outputs require validation and should not obscure who was excluded or how weights were formed. Transparency about recruitment source, eligibility, incentives, exclusions, response outcomes, weighting, and uncertainty is essential.

Best practice is to preregister or document the sampling protocol, maintain an auditable sample-selection log, monitor deviations during fieldwork, use appropriate survey estimators, and distinguish exploratory evidence from population-representative evidence. Claims should never exceed the design’s inferential support.

Challenges and Limitations

Sampling theory cannot fully resolve ambiguous population definitions, poor frames, inaccessible respondents, deceptive answers, changing eligibility, or low-quality measurement. Probability sampling can be expensive and may still suffer from nonresponse or coverage error. Non-probability samples can be valuable but depend heavily on assumptions about selection and adjustment; statistical generalization requires caution. Sample-size formulas also rely on planning inputs that may be uncertain, and subgroup, interaction, and longitudinal analyses often require substantially larger samples than a single overall estimate.

Chapter Summary

Sampling design connects a research question to a defensible set of observations. Researchers must define the target population, frame, elements, sampling units, and estimands before selecting cases. Probability methods provide known selection probabilities and a basis for design-based inference; non-probability methods can be appropriate for exploration, specialized access, and qualitative purposes but do not automatically support conventional margins of error.

Sample size depends on precision, confidence, variability, power, finite-population size, design effect, response rate, subgroup requirements, and cost. Stratification can improve precision and representation, whereas clustering reduces cost but often increases variance. Weights correct unequal selection and adjustment requirements but may increase uncertainty. Coverage error and nonresponse bias are distinct from sampling error and cannot be solved merely by adding respondents. Ethical, transparent, and technology-aware implementation is necessary for credible marketing evidence.

Key Terms

Target population: Complete group about which the researcher intends to make claims.

Survey population: Portion of the target population reachable under the study procedures.

Probability sampling: Sampling in which each eligible unit has a known, nonzero selection probability.

Non-probability sampling: Sampling in which selection probabilities are unknown or not positive for all units.

Cluster sampling: Sampling groups of population elements rather than individual elements initially.

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

Nonresponse bias: Bias arising when participation is related to the study outcome or its determinants.

Design effect: Ratio of variance under a complex design to variance under simple random sampling.

Effective sample size: Simple-random-sample equivalent amount of statistical information.

Sampling weight: Value representing the population contribution of a sampled unit.

Margin of error: Planned maximum distance between an estimate and parameter under stated assumptions.

Statistical power: Probability of detecting a specified effect when that effect is present.

Calibration: Adjustment of weights so weighted sample totals agree with known population benchmarks.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

Correct Answer: B. The operational list or mechanism from which units are selected

Question 2

Under simple random sampling, what is the approximate sample size for a proportion with 95% confidence, p = 0.50, and a margin of error of 0.05?

Question 3

Which method deliberately divides a population into subgroups and samples within each subgroup?

Question 4

If an estimated simple-random sample requirement is 400 and the expected response rate is 50%, approximately how many units should be contacted?

Question 5

A design effect greater than 1 generally indicates that the complex design has what consequence relative to simple random sampling?

Question 6

Question 7

Question 8

What is the primary limitation of conventional margins of error for a convenience sample?

Question 9

If the average cluster size is b = 20 and the intracluster correlation is ρ = 0.03, the approximate cluster design effect is:

Question 10

B. Using multiple contact attempts and comparing respondents with frame information

Correct Answer: B. Using multiple contact attempts and comparing respondents with frame information

Chapter 8 Sampling Design and Sample Size

Sampling Design and Sample Size

Introduction