Measurement, Sampling, and Data Quality

Introduction

Research conclusions are only as credible as the data on which they are based. In artificial intelligence and machine learning (AI/ML), this principle is especially important: a model may be technically sophisticated yet produce misleading results if its variables are poorly defined, its sample is unrepresentative, its measurements are unreliable, or its data provenance is unclear. Measurement and sampling are therefore not merely administrative stages of research. They are central to construct validity, reproducibility, fairness, and responsible deployment.

This chapter explains how abstract concepts are operationalized into measurable variables; how scales, questionnaires, and instruments are designed; how reliability and validity are assessed; how probability and non-probability samples are selected; how sample size is estimated; and how researchers manage bias, missing data, and provenance. The discussion is applicable to surveys, experiments, observational studies, usability research, and datasets used to train and evaluate AI systems.

Learning Objectives

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

1. Distinguish constructs, variables, indicators, and operational definitions.

2. Select appropriate measurement scales and design usable survey instruments.

3. Explain and assess reliability and validity.

4. Compare probability and non-probability sampling methods.

5. Estimate sample size for common proportion and mean studies.

6. Identify sources of sampling, measurement, and nonresponse bias.

7. Select appropriate strategies for missing data.

8. Document data provenance and evaluate data quality in AI/ML projects.

1. From Constructs to Measurable Variables

A construct is an abstract idea that a researcher wishes to study, such as trust in an AI system, algorithmic literacy, workload, fairness perception, or student engagement. Constructs cannot usually be observed directly. They are represented through variables and indicators.

A variable is a characteristic that can take different values. Examples include age, response time, number of errors, perceived usefulness, or model accuracy. An indicator is an observable measure used to represent a construct. For example, trust in an automated recommendation system might be indicated by confidence in the system, willingness to follow its recommendation, and perceived reliability.

An operational definition specifies exactly how a construct will be measured. A useful operational definition identifies:

• the population or unit being studied;
• the observable indicators;
• the instrument or data source;
• the response or scoring procedure;
• the time period and context;
• rules for coding and interpretation.

For example, “student engagement” is too broad for direct analysis. An operational definition might state: “Engagement is the mean score on eight five-point items measuring attention, participation, persistence, and interest during the previous four weeks, with higher scores indicating greater engagement.”

Operationalization requires a balance between conceptual fidelity and practical feasibility. A narrow operational definition is easier to measure but may omit important dimensions. A broad definition may be conceptually rich but difficult to administer consistently.

1.1 Dimensions and indicators

Many constructs are multidimensional. Service quality, for instance, may include reliability, responsiveness, assurance, and empathy. Researchers should decide whether the dimensions are reflective or formative.

In a reflective model, the construct is assumed to cause the indicators. If a person has high trust, several trust items are expected to receive high responses. Indicators are often correlated, and removing one may not change the meaning substantially.

In a formative model, indicators combine to form the construct. Socioeconomic status may be formed from income, education, occupation, and housing conditions. Indicators need not be highly correlated, and removing one may change the construct’s meaning.

This distinction affects instrument design and analysis. Treating formative indicators as reflective, or vice versa, can produce invalid conclusions.

1.2 Variables and levels of measurement

The classical levels of measurement are nominal, ordinal, interval, and ratio.

• Nominal variables classify cases without numerical order, such as programming language or device type.
• Ordinal variables have an order, but distances between categories are not necessarily equal, such as low, medium, and high satisfaction.
• Interval variables have ordered values with equal intervals but no meaningful absolute zero. Temperature in Celsius is a standard example.
• Ratio variables have equal intervals and a meaningful zero, such as elapsed time, income, or number of errors.

The level of measurement guides permissible summaries and analyses. Frequencies and proportions are appropriate for nominal variables. Medians and rank-based methods are often appropriate for ordinal variables. Means and standard deviations may be appropriate for interval or ratio variables, subject to distributional assumptions. A five-point Likert item is technically ordinal, although a sum or mean of several well-designed Likert items is often treated as approximately continuous when justified by the research design and distribution.

2. Designing Scales, Surveys, and Instruments

An instrument is a structured tool for collecting measurements. It may be a questionnaire, interview schedule, observation protocol, coding scheme, test, sensor, software log, or benchmark procedure.

2.1 Planning an instrument

Instrument development should begin with a construct map or table of specifications. This table links research questions to constructs, dimensions, indicators, item wording, response formats, and analysis plans. It prevents researchers from collecting attractive but irrelevant data.

A typical development sequence is:

1. Define the construct and its dimensions.

2. Review established theories and instruments.

3. Generate more items than will ultimately be needed.

4. Obtain expert review for relevance and clarity.

5. Conduct cognitive interviews or think-aloud testing.

6. Pilot the instrument with participants similar to the target population.

7. Analyze item performance and revise.

8. Document the final instrument, scoring, and administration procedure.

Using an established scale can improve comparability, but permission, cultural adaptation, translation, and context must be considered. Copying an instrument without understanding its construct or validation evidence is not good methodology.

2.2 Writing effective questionnaire items

A good item measures one idea in language suitable for the target respondents. Avoid double-barrelled questions such as “The system is fast and accurate,” because a respondent may agree with one component but not the other. Avoid leading language, unnecessary technical terms, ambiguous time frames, and emotionally loaded wording.

Specify a reference period where memory matters: “During the last seven days…” is more interpretable than “Usually…”. Response options should be mutually exclusive and collectively exhaustive. Include “not applicable” when a genuine non-applicable condition exists, rather than forcing an artificial answer.

Balanced response scales may reduce acquiescence, but negatively worded items can confuse respondents and introduce method effects. If reverse-scored items are used, they should be written clearly and checked carefully during coding.

A common agreement scale is:

1 = strongly disagree, 2 = disagree, 3 = neither agree nor disagree, 4 = agree, 5 = strongly agree.

The numbers are codes, not automatically quantities with equal psychological distance. Researchers must justify how the resulting scores are summarized.

2.3 Measurement error

An observed score can be represented conceptually as:

X = T + E

where X is the observed score, T is the underlying true score under the measurement model, and E is random or systematic error. This classical model is a conceptual approximation, not a claim that a perfectly observable true score exists.

Random error varies unpredictably and tends to reduce precision. Systematic error consistently pushes measurements in a particular direction, such as social desirability bias or a poorly calibrated sensor. Increasing sample size can reduce the impact of random sampling error, but it does not correct systematic measurement bias.

2.4 Pilot testing

A pilot study is a small-scale trial of the instrument and procedures. It examines comprehension, completion time, missing responses, technical failures, interviewer effects, and preliminary item behavior. A pilot should not be treated as a substitute for the main study’s validation evidence. Its purpose is refinement and risk reduction.

3. Reliability and Validity

Reliability concerns consistency or precision of measurement. Validity concerns the degree to which evidence supports the interpretation and use of scores for their intended purpose. A measure can be reliable without being valid: a defective scale may consistently measure the wrong attribute.

3.1 Forms of reliability

Test–retest reliability assesses stability by administering the same instrument to the same participants on two occasions when the construct is expected to remain stable. It is unsuitable when genuine change is expected or when memory effects are strong.

Inter-rater reliability assesses agreement or consistency among observers or coders. For categorical coding, statistics such as Cohen’s kappa may be appropriate; for continuous ratings, intraclass correlation or related methods may be used. The choice depends on the design and estimand.

Internal consistency concerns the extent to which items intended to measure the same construct produce coherent responses. Cronbach’s alpha is frequently reported, but it is not a universal measure of validity and does not prove unidimensionality. Alpha is influenced by the number of items and assumes conditions such as essentially tau-equivalent items for its classical interpretation. Very high alpha may indicate redundant items rather than excellent measurement.

For a set of k items, a commonly used expression is:

α = [k/(k − 1)] [1 − (Σσᵢ²/σₜ²)]

where σᵢ² is the variance of item i and σₜ² is the variance of the total score formed from the items. This statistic is applicable to a set of items intended to measure a common construct and should be interpreted with item content, dimensionality, and the study context. It is not a fixed pass–fail test; values around .70 are sometimes used as a tentative research guideline, but acceptable levels depend on purpose, construct breadth, and consequences of error.

3.2 Evidence of validity

Content validity concerns whether the items adequately represent the construct domain. It is supported through theory, literature review, expert judgment, and systematic content mapping.

Construct validity concerns whether the measure behaves as theory predicts. Convergent evidence is obtained when related measures correlate as expected. Discriminant evidence is obtained when measures of distinct constructs are not excessively related. Factor analysis can help examine dimensional structure, but statistical fit alone does not establish conceptual validity.

Criterion-related validity examines relationships with an external criterion. Concurrent validity uses a criterion measured at approximately the same time; predictive validity concerns future outcomes. A model-performance metric may have predictive value for operational decisions but still fail to represent a broader construct such as fairness or trust.

Face validity is the apparent relevance of an instrument to non-specialists. It may improve acceptability but is weaker evidence than systematic content, construct, and criterion evidence.

Validity is contextual. Evidence supporting an instrument in one population, language, or decision context may not automatically generalize to another. Measurement invariance should be considered when comparing groups, especially across cultures or demographic categories.

4. Sampling Designs

The target population is the complete group to which the researcher intends to generalize. The accessible population is the portion that can actually be reached. A sampling frame is the operational list or process from which the sample is selected. The unit of analysis is the entity about which conclusions are drawn; the sampling unit is the entity selected. These may differ, as when students are sampled through classes but analyzed individually.

4.1 Probability sampling

In probability sampling, each population unit has a known, non-zero probability of selection. This permits principled estimation of sampling error and supports statistical generalization, provided the frame and implementation are adequate.

Simple random sampling gives every frame unit an equal selection probability. It is conceptually straightforward but may require a complete and accurate frame.

Systematic sampling selects every rth unit after a random start, where r is related to the frame size and desired sample size. It is efficient but can be biased if the frame has periodic structure aligned with the interval.

Stratified sampling divides the population into relevant subgroups, or strata, and samples within each stratum. It can improve precision and ensure representation of small but important groups. Proportionate allocation follows population proportions; disproportionate allocation may deliberately oversample a small group, requiring weights during estimation.

Cluster sampling selects naturally occurring groups, such as schools, hospitals, or geographic areas, and then observes units within selected clusters. It reduces field cost but often increases sampling variance because units within a cluster resemble one another.

Multistage sampling combines stages, such as selecting regions, then institutions, then individuals. It is common in large surveys and requires careful attention to unequal probabilities and design effects.

4.2 Non-probability sampling

Convenience sampling recruits readily available participants. It is inexpensive but vulnerable to selection bias. Purposive sampling deliberately selects information-rich cases according to stated criteria and is common in qualitative or specialized research. Quota sampling fills subgroup targets without random selection within groups. Snowball sampling uses participant referrals and can help reach hidden populations, but network effects may limit diversity.

Non-probability samples are not automatically useless. They may be appropriate for exploratory studies, usability testing, rare populations, or qualitative inquiry. However, researchers should avoid claiming population representativeness without strong supporting evidence. Statistical significance tests based on random-sampling assumptions can be misleading when selection probabilities are unknown.

5. Sample Size Estimation

Sample size depends on the research question, design, precision required, expected variability, confidence level, power, effect size, attrition, and resources. There is no universally correct sample size.

For estimating a population proportion with a large or effectively infinite population, a common planning formula is:

n₀ = z²p(1 − p)/e²

where n₀ is the initial required sample size, z is the standard-normal critical value corresponding to the desired confidence level, p is the anticipated population proportion, and e is the desired margin of error expressed as a proportion. If p is unknown, p = 0.50 gives the largest required sample for a fixed z and e.

For a 95% confidence level, z is approximately 1.96. If p = 0.50 and e = 0.05:

n₀ = (1.96² × 0.50 × 0.50)/(0.05²) ≈ 384.16.

Thus, approximately 385 completed responses are required under the stated assumptions. This calculation concerns precision for a proportion under a simple random sample; it does not account for nonresponse, clustering, weighting, or a finite population.

For a finite population of size N, a commonly used correction is:

n = n₀/[1 + (n₀ − 1)/N].

If N = 1,000 and n₀ = 384.16, then n is approximately 277.7, so about 278 completed responses are required. The correction is relevant when the sample is a substantial fraction of the population and selection is without replacement under an appropriate design.

To account for anticipated nonresponse, divide the required number of completed cases by the expected response rate. If 385 completed responses are needed and the expected response rate is 70%, the number invited is 385/0.70 ≈ 550. This is a planning adjustment, not a guarantee that nonresponse bias has been removed.

For hypothesis testing, sample size should generally be based on an a priori power analysis using the expected effect size, significance level, desired power, outcome variability, and analysis model. “Large enough” cannot be determined solely by a population-size rule.

6. Bias, Missing Data, and Data Quality

Bias is systematic deviation from the target of inference. Sampling bias occurs when inclusion probabilities or participation patterns are related to relevant characteristics. Coverage bias occurs when the sampling frame omits parts of the target population. Nonresponse bias occurs when respondents differ meaningfully from nonrespondents. Measurement bias arises from wording, mode, instrument, interviewer, sensor, or coding procedures.

Researchers should distinguish sampling error from nonsampling error. Sampling error is random variation resulting from observing a sample rather than the entire population. Nonsampling errors include coverage, nonresponse, processing, and measurement problems; increasing sample size does not necessarily reduce them.

6.1 Missing data

Missingness should be diagnosed rather than simply deleted. Missing completely at random (MCAR) means missingness is unrelated to observed and unobserved values. Missing at random (MAR) means missingness may depend on observed variables after conditioning on them. Missing not at random (MNAR) means missingness depends on unobserved or missing values even after accounting for observed information.

Complete-case analysis is simple but can reduce power and create bias, especially when MCAR is implausible. Single-value methods such as mean imputation reduce visible missingness but distort variance and relationships. Multiple imputation creates several plausible completed datasets, analyzes each, and combines results while reflecting uncertainty; it requires a defensible imputation model. In longitudinal or administrative data, model-based methods may be appropriate. Sensitivity analyses should examine how conclusions change under plausible missingness assumptions.

For survey instruments, researchers should report item-level and unit-level missingness, reasons where known, exclusions, and the exact handling procedure. In AI datasets, missing labels, rejected records, and unavailable subgroups must be documented because filtering decisions can alter model behavior.

6.2 Data-quality dimensions

Important dimensions include accuracy, completeness, consistency, timeliness, uniqueness, validity, and relevance. A dataset can be complete but inaccurate, or accurate but outdated. Quality checks may include range rules, type checks, duplicate detection, logical consistency checks, outlier review, timestamp validation, and comparison with authoritative sources.

Automated checks should be supplemented with human review. An unusual value may be an error, but it may also represent a genuine rare event. Blindly deleting outliers can damage validity.

7. Data Provenance and Responsible AI

Data provenance is the documented history of data: where it originated, who collected or transformed it, which versions were used, and under what permissions. A provenance record should include source, collection date, population and sampling process, consent or legal basis, preprocessing, labeling instructions, transformations, exclusions, version identifiers, custodians, and known limitations.

For AI/ML research, provenance should extend to train, validation, and test partitions. Researchers must prevent leakage, in which information from the evaluation data or future events enters training or feature construction. Dataset versioning, immutable raw data, reproducible pipelines, checksums, and audit logs improve traceability.

Ethical and regulatory responsibilities include informed consent where applicable, privacy protection, data minimization, secure access, appropriate retention, and fair treatment of groups. De-identification reduces risk but does not guarantee anonymity, particularly when datasets can be linked. Researchers should assess whether data collection imposes unequal burdens or whether measurement instruments function differently across demographic groups.

A model trained on a convenience sample may perform well for the sampled population yet fail for people absent from the data. Fairness evaluation should therefore examine representation, label quality, subgroup performance, and the consequences of errors. Governance documentation should make limitations visible rather than presenting a dataset as neutral or universally representative.

8. Illustrative Case: Evaluating Trust in an AI Academic Adviser

Background

A university develops an AI academic adviser that recommends courses and alerts students to prerequisite problems. The research team wants to determine whether students trust the system and whether trust differs across student groups.

Problem

The initial plan is to post a voluntary link on the university website and measure trust using three newly written items. This approach risks self-selection, inadequate construct coverage, and weak evidence of reliability and validity.

Analysis

The team defines trust as a multidimensional construct involving perceived competence, reliability, and willingness to rely on recommendations. The target population is all enrolled undergraduate students, while the accessible population is students with active university accounts. The sampling frame is the enrollment registry.

The researchers create a construct map and adapt items from prior validated research, obtaining permission where necessary. Experts review content, and cognitive interviews identify confusion about the phrase “AI adviser.” A pilot reveals that students interpret “trust” as either technical accuracy or privacy, so the instrument adds a separate privacy concern measure.

Approach and solution

The main study uses stratified random sampling by year of study and faculty. Oversampling of first-year students is used because they are of particular interest; analysis weights are planned to restore population proportions for overall estimates. The team estimates the required number of completed responses using a 95% confidence level and 5% margin of error for a proportion, then inflates invitations for expected nonresponse.

The survey records response time, item nonresponse, and device mode. A small number of attention checks are used cautiously and are not treated as a complete quality-control strategy. Reliability is examined for the multi-item trust dimensions, while content and construct evidence are assessed through expert review and relationships with actual recommendation use. Group comparisons are interpreted only after checking whether the scale functions comparably across groups.

Data are stored with versioned instruments and documented transformations. Identifiers are separated from survey responses. The team reports coverage limitations, nonresponse patterns, and the fact that students who never accessed the adviser may have different views from active users.

Outcome

The revised design finds that overall trust is moderate, but perceived competence and privacy concern are distinct dimensions. Trust is higher among frequent users, although the cross-sectional design does not establish that use causes trust. First-generation students are underrepresented among voluntary respondents, but stratified recruitment and weighting improve coverage. The university decides to improve explanation of recommendations and privacy notices rather than using the trust score as a standalone deployment justification.

Learning points

The case demonstrates that operationalization precedes analysis; sampling affects generalizability; reliability is not validity; subgroup comparisons require measurement scrutiny; and provenance and privacy are part of data quality, not optional documentation.

9. Best-Practice Workflow

A defensible measurement and sampling workflow includes the following steps:

1. Define the inferential target and unit of analysis.

2. Map constructs to dimensions and observable indicators.

3. Select or develop instruments using theory and prior evidence.

4. Pilot wording, procedures, technology, and coding.

5. Choose a sampling design consistent with the generalization claim.

6. Estimate sample size using precision or power, then account for design and nonresponse.

7. Predefine data-quality rules and missing-data procedures.

8. Monitor recruitment and compare respondents with the target population where possible.

9. Preserve provenance, versions, permissions, and transformation logs.

10. Report limitations, uncertainty, exclusions, and subgroup performance transparently.

Challenges and Limitations

Measurement is constrained by imperfect constructs, respondent fatigue, cultural interpretation, changing contexts, and trade-offs between brevity and coverage. Reliability statistics can be misused as universal thresholds. Probability sampling may be expensive or impossible when no adequate frame exists. Non-probability samples may be the only practical option but require restrained claims. Sample-size formulas rely on assumptions that real studies may violate. Missing-data methods depend on assumptions that cannot always be verified. Finally, data quality is multidimensional: cleaning data can improve consistency while unintentionally removing meaningful variation.

Chapter Summary

Measurement converts theoretical constructs into observable, documented variables. Strong operational definitions identify dimensions, indicators, scoring, context, and time frame. Instrument quality depends on clear items, suitable response options, pilot testing, and attention to measurement error. Reliability concerns consistency, while validity concerns the defensibility of score interpretations; neither should be reduced to a single statistic.

Probability sampling supports principled generalization when selection probabilities and implementation are appropriate. Stratified, cluster, systematic, and multistage designs each involve trade-offs. Non-probability sampling can support exploratory and specialized research but does not automatically justify population claims. Sample size should reflect precision, power, design effects, nonresponse, and the intended analysis.

Bias, missing data, and poor provenance can undermine even large datasets. Researchers should diagnose missingness, use defensible methods, document decisions, protect privacy, and maintain versioned records. In AI/ML, representative sampling, label quality, leakage prevention, subgroup evaluation, and transparent data documentation are essential to trustworthy research.

Key Terms

Construct: An abstract concept that cannot usually be observed directly.

Operational definition: A precise specification of how a construct will be measured.

Indicator: An observable measure representing a construct or dimension.

Variable: A characteristic that can take different values.

Reliability: The consistency or precision of a measurement.

Validity: Evidence supporting the interpretation and use of a measurement.

Content validity: Evidence that items adequately cover a construct domain.

Construct validity: Evidence that a measure behaves as theoretical expectations predict.

Sampling frame: The list or process from which a sample is selected.

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

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

Stratified sampling: Sampling separately within defined population subgroups.

Cluster sampling: Sampling naturally occurring groups rather than individuals directly.

Nonresponse bias: Bias resulting when respondents differ systematically from nonrespondents.

Measurement bias: Systematic distortion caused by an instrument, procedure, respondent, or coder.

Missing completely at random: Missingness unrelated to observed or unobserved values.

Multiple imputation: A method that represents missing-data uncertainty through several plausible datasets.

Data provenance: Documentation of a dataset’s origins, transformations, versions, and stewardship.

Data leakage: Inclusion of information in model development that would not be available at intended prediction time.

Measurement invariance: Comparable measurement meaning across groups or conditions.

Review Questions

Short Answer Questions

1. What is the difference between a construct and an operational definition?

2. Distinguish reliability from validity.

3. What is a sampling frame, and why can it create coverage bias?

4. State two advantages and two limitations of stratified sampling.

5. Why does increasing sample size not necessarily remove measurement bias?

Long Answer Questions

1. Explain the process of developing a questionnaire from a theoretical construct, including pilot testing and validity evidence.

2. Compare simple random, stratified, cluster, and systematic sampling, including situations in which each may be appropriate.

3. Explain Cronbach’s alpha, its assumptions, and why a high value does not establish validity.

4. Discuss MCAR, MAR, and MNAR missingness and compare complete-case analysis with multiple imputation.

5. Describe a provenance and data-quality framework for an AI/ML research dataset.

Analytical / Application Questions

1. A university has 1,000 students and wants a 95% estimate of a proportion with a 5% margin of error. Using p = 0.50 and z = 1.96, calculate the initial sample size and apply the finite population correction. Explain the assumptions.

2. Design a sampling and measurement plan to evaluate whether an AI medical triage tool is trusted equally by nurses in emergency, surgical, and outpatient units. Address representation, instrument validity, nonresponse, and ethics.

3. An ML dataset contains many more records from urban users than rural users, and the label “successful outcome” is assigned by different teams. Diagnose the major data-quality risks and propose a remediation and documentation plan.

Multiple Choice Questions with Answers

Question 1

Which statement best describes an operational definition?

A. A general theory explaining a phenomenon

B. A precise specification of how a construct will be measured

C. A list of all members of a population

D. A statistical test of a hypothesis

Correct Answer: B. A precise specification of how a construct will be measured

Question 2

Which level of measurement categorizes observations without implying order?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: A. Nominal

Question 3

A measure gives nearly identical results each time but measures the wrong construct. It is best described as:

A. Valid but unreliable

B. Reliable but not necessarily valid

C. Neither measurable nor observable

D. Representative by definition

Correct Answer: B. Reliable but not necessarily valid

Question 4

Which sampling method deliberately samples within population subgroups?

A. Convenience sampling

B. Snowball sampling

C. Stratified sampling

D. Simple accidental sampling

Correct Answer: C. Stratified sampling

Question 5

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

A. 96

B. 196

C. 278

D. 385

Correct Answer: D. 385

Question 6

What is a principal limitation of convenience sampling?

A. It always requires a population registry

B. It prevents any data from being collected

C. It may produce selection bias and weak population generalizability

D. It guarantees equal selection probabilities

Correct Answer: C. It may produce selection bias and weak population generalizability

Question 7

Which missingness condition means missingness may depend on observed variables after conditioning on them?

A. MCAR

B. MAR

C. MNAR

D. 完全 random

Correct Answer: B. MAR

Question 8

Which practice most directly supports data provenance?

A. Deleting all raw files after preprocessing

B. Keeping undocumented manual edits

C. Recording sources, transformations, versions, and permissions

D. Reporting only the final model accuracy

Correct Answer: C. Recording sources, transformations, versions, and permissions

Question 9

What is data leakage in machine learning?

A. Loss of a file during storage

B. Use of information during development that would not be available at prediction time

C. Random sampling from a complete frame

D. An increase in response rate

Correct Answer: B. Use of information during development that would not be available at prediction time

Question 10

Why may cluster sampling have greater sampling variance than simple random sampling at the same nominal sample size?

A. Units within clusters may resemble one another

B. Clusters contain no measurable units

C. It always uses a smaller population

D. It eliminates all nonresponse

Correct Answer: A. Units within clusters may resemble one another