Data Foundations, Measurement and Sampling

Introduction

Business decisions in human resource management (HRM) depend on evidence. Questions about turnover, absenteeism, recruitment effectiveness, employee engagement, pay equity and training outcomes cannot be answered responsibly unless data are properly defined, measured, collected and interpreted. Statistics begins before calculation: it begins with deciding what is to be observed, how it will be represented, and which observations can reasonably support a conclusion about a wider workforce.

This chapter introduces the foundations of business data, measurement and sampling. It distinguishes populations from samples, explains variables and measurement scales, examines reliability and validity, and presents probability and non-probability sampling methods. The emphasis is on HR applications, where data may concern people, organizational processes and sensitive personal characteristics.

Learning Objectives

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

1. distinguish data, information, variables, populations and samples;
2. classify variables and measurement scales appropriately;
3. explain operational definitions and sources of measurement error;
4. assess reliability, validity and data quality;
5. select suitable probability or non-probability sampling methods;
6. calculate and interpret basic sampling quantities, including sampling fractions and standard errors;
7. design a defensible HR data-collection plan; and
8. identify ethical, legal and practical risks in workforce data collection.

1. Data as a Foundation for Business Statistics

Data are recorded observations, measurements, classifications or responses. A collection of data may be numerical, categorical, textual, visual or transactional. Information is data organized and interpreted so that it supports understanding or decision-making. For example, a list of employee absence records is data; a finding that absence is concentrated in one department is information.

A unit of analysis is the entity about which a conclusion is made. In HRM it may be an employee, job application, team, department, work shift or organization. Confusing units of analysis creates serious errors. For example, a survey of employees cannot automatically establish that departments differ unless responses are appropriately grouped and analyzed.

A variable is a characteristic that can take different values across units. Examples include age, monthly pay, employment status, years of service, training hours and engagement score. A constant has the same value for every unit under consideration.

Data may be:

• Primary data, collected for the current investigation, such as responses to a new engagement survey.
• Secondary data, originally collected for another purpose, such as payroll records or a government labour-force database.
• Cross-sectional data, observed at one point or during a short period.
• Longitudinal data, observed repeatedly over time for the same or comparable units.
• Structured data, organized into defined fields, such as a spreadsheet.
• Unstructured or semi-structured data, such as interview transcripts, emails or text responses.

A dataset normally contains rows representing units and columns representing variables. A data dictionary should document each variable, its definition, coding, unit, permitted values and missing-value convention.

2. Populations, Samples and Parameters

The target population is the complete group to which a study intends to generalize. The accessible population is the portion that can actually be identified and contacted. A sampling frame is the operational list or database from which the sample is selected. A frame may be a current employee register, an applicant database or a list of organizational units.

A sample is a subset selected from a population. A numerical characteristic of a population is a parameter; the corresponding characteristic calculated from a sample is a statistic. For example, the population mean monthly absence is μ, whereas the sample mean is x̄.

For a finite population of size N, a simple random sample of size n has sampling fraction:

Here, f is dimensionless, n is the number selected and N is the population size. The fraction indicates how much of the population is observed. It does not, by itself, guarantee representativeness.

If a variable has values x₁, x₂, ..., xₙ for n sampled units, its sample mean is:

The mean is appropriate for quantitative data where addition and division are meaningful. The sample proportion with a specified characteristic is:

where x is the number of sampled units possessing the characteristic and n is the sample size.

Sampling error and standard error

A sample statistic varies from sample to sample. This natural variation is sampling error, not necessarily a mistake. For a simple random sample from a large population, the standard error of a sample mean is approximately:

where s is the sample standard deviation and n is the sample size. This approximation assumes independent observations and a sampling design close to simple random sampling. If sampling without replacement from a finite population that is not large relative to the sample, a finite population correction may be used:

The correction applies when N is known and sampling is without replacement. It reduces the standard error because observing a substantial proportion of the population leaves less uncertainty about the remainder.

For a sample proportion under simple random sampling, an approximate standard error is:

This approximation is most reliable when the sample is sufficiently large and observations are independent. Complex designs, clustering and weighting require design-specific methods.

Worked Example 1: Sampling quantities

An organization has N = 2,000 employees and surveys n = 200 selected employees. The sampling fraction is:

f = 200 ÷ 2,000 = 0.10

Thus, 10% of employees are selected. If 146 respondents report that they would recommend the organization, then:

p̂ = 146 ÷ 200 = 0.73

The sample estimate of the recommendation proportion is 73%. This is an estimate, not proof that exactly 73% of all employees would recommend the organization. Nonresponse and selection procedures must also be examined.

3. Types of Variables and Data

Qualitative and quantitative variables

A qualitative or categorical variable places units into groups. Department, contract type and employment status are examples. A quantitative variable records numerical magnitude, such as salary, age or number of absence days.

Quantitative variables may be:

• Discrete, taking separate countable values, such as number of grievances;
• Continuous, potentially taking any value within an interval, such as time to recruit measured precisely.

A numerical code does not necessarily make a variable quantitative. If permanent employees are coded 1 and temporary employees 2, the codes are labels, not measurements on which averages are substantively meaningful.

Explanatory and response variables

A response variable is the outcome being studied, such as voluntary turnover. An explanatory variable is a factor used to explain or predict it, such as tenure or supervisor support. The distinction is analytical rather than absolute: a variable may be a response in one study and an explanatory variable in another. Association does not establish causation, particularly in observational HR data.

Time-related classifications

A stock measures a level at a specific time, such as the number of employees on 31 December. A flow measures activity over a period, such as hires during a quarter. Rates require a clearly defined numerator, denominator and time period. For example, a turnover rate may be expressed as:

The formula is meaningful only when the separation definition, period and employee denominator are consistent. Organizations may use different conventions; comparisons require equivalent definitions.

4. Measurement Scales

Measurement is the systematic assignment of values or labels to observations according to rules. The scale determines which summaries and calculations are defensible.

Nominal categories should be mutually exclusive and collectively exhaustive, subject to a justified “other” or “prefer not to say” option. Ordinal categories have order, but the difference between adjacent categories may not be equal. A rating of 4 is not necessarily twice as positive as a rating of 2.

An interval scale permits meaningful differences. A ratio scale additionally has a true zero: zero hours worked means no hours worked, and 20 hours is twice 10 hours in the relevant sense. The scale classification guides analysis; it does not automatically determine one universally correct statistical method.

Indexes and composite measures

HR constructs such as engagement, organizational commitment and psychological safety are often not directly observable. They are latent constructs, represented through several measured items. A composite score may be formed by summing or averaging item responses, but this requires a clear scoring rule, evidence that items represent the construct, and appropriate treatment of missing responses.

5. Operationalization and Measurement Quality

A broad concept must be translated into an operational definition: a precise statement of what will be measured, for whom, when, and how. “Employee engagement” might be operationalized as the mean response to eight specified items on a five-point agreement scale among employees employed during the survey period.

Operational definitions improve consistency, but they are not automatically complete or valid. A single question about job satisfaction may not capture the wider construct. A payroll field labelled “absence” may exclude informal leave and therefore differ from an employee’s understanding of absence.

Reliability

Reliability concerns consistency. A reliable instrument produces reasonably stable and consistent results when the underlying condition has not changed. Forms include:

• Test–retest reliability: consistency across time;
• Inter-rater reliability: agreement among assessors;
• Internal consistency: coherence among items intended to measure one construct.

Reliability is necessary but not sufficient for validity. A scale can consistently measure the wrong concept.

Validity

Validity concerns whether evidence supports the intended interpretation and use of a measure. Important forms include:

• Content validity: the measure adequately covers the relevant domain;
• Construct validity: the measure behaves as theory predicts;
• Criterion-related validity: the measure relates appropriately to an external criterion;
• Face validity: the measure appears sensible to relevant users, although appearance alone is weak evidence.

For example, using training attendance as a measure of training effectiveness has limited validity: attendance indicates participation, not necessarily learning or improved performance.

Measurement error

An observed value can be represented conceptually as:

Random error fluctuates unpredictably and can reduce precision. Systematic error, or bias, shifts results in a consistent direction. Sources include ambiguous questions, inaccurate records, recall difficulty, interviewer effects, social desirability and poorly calibrated instruments.

A practical data-quality framework examines:

• Accuracy: closeness to the intended value;
• Completeness: absence of unjustified missing records;
• Consistency: agreement across fields and systems;
• Timeliness: availability when needed;
• Uniqueness: avoidance of duplicate units;
• Validity: conformity with permitted definitions and ranges.

6. Designing Questions and Instruments

Good questions use plain language, specify a reference period and ask one thing at a time. A double-barrelled question such as “How satisfied are you with pay and promotion?” prevents a clear interpretation if respondents hold different views on the two topics.

Avoid leading wording, unexplained technical terms and unnecessary personal questions. Response categories should be balanced and relevant. A “not applicable” option is preferable to forcing an arbitrary answer. Sensitive questions should explain purpose, confidentiality arrangements and voluntary choices where appropriate.

A questionnaire should normally be piloted with people resembling the target respondents. Piloting can reveal ambiguous wording, excessive length, technical problems, missing response categories and unexpected interpretations. Administrative data also require a data audit: definitions, extraction dates, historical changes and access permissions should be documented.

7. Sampling Designs

Probability sampling

In probability sampling, each eligible unit has a known, non-zero probability of selection. This supports defensible inference when implementation is sound.

Simple random sampling selects units so that each sample of a specified size has an equal chance. It is conceptually simple but requires a complete frame.

Systematic sampling selects every k-th unit after a random starting point, where approximately:

N is frame size and n is desired sample size. Systematic sampling is efficient, but periodic patterns in the frame can create bias.

Stratified sampling divides the population into relevant subgroups, or strata, and samples within each. Strata might be region, job family or employment level. It can improve representation and precision when members within strata are relatively similar for the outcome. Disproportionate sampling requires weights when estimating population quantities.

Cluster sampling selects naturally occurring groups, such as branches or teams, and surveys units within selected groups. It can reduce travel and administrative cost, but observations in the same cluster may be similar, reducing effective information.

Multistage sampling combines stages, such as selecting branches, then teams, then employees. Its analysis must reflect unequal selection probabilities and clustering.

Non-probability sampling

Convenience sampling selects readily available participants. Purposive sampling deliberately selects people with relevant experience. Quota sampling fills specified subgroup targets without necessarily randomly selecting within groups. Snowball sampling uses participant referrals and may help reach hidden populations.

These designs can be useful for exploratory work, qualitative research or populations without a reliable frame. However, sampling error formulas for simple random sampling do not automatically justify population-generalization from a non-probability sample. Volunteer surveys are particularly vulnerable to self-selection: people with strong opinions may be more likely to respond.

Nonresponse

Nonresponse occurs when selected units do not provide usable data. It can be unit nonresponse, where an entire case is missing, or item nonresponse, where some questions are unanswered. A high response rate does not guarantee absence of bias, and a lower response rate does not prove bias; the characteristics of responders and nonresponders matter.

Good practice includes clear invitations, reasonable survey length, reminders, accessible modes, confidentiality assurances and analysis of response patterns. Imputation of missing values should be documented and should not conceal substantial data-quality problems.

8. Sample Size and Precision

Sample size should follow the study purpose, population variability, desired precision, confidence level, design and resources. Larger samples generally reduce standard error, but they do not correct systematic bias.

For a large population, an approximate sample-size expression for estimating a population proportion with margin of error E is:

Here, n is required sample size, z is the standard-normal critical value associated with the chosen confidence level, p is the anticipated population proportion, and E is the desired half-width of the margin of error. This expression assumes a probability sample, independent observations and a sufficiently large population. When p is unknown, p = 0.5 produces the largest required sample size for a fixed z and E.

If the population is finite, a finite-population adjustment to the planning value may be written as:

where n₀ is the large-population sample size, nₐ is the adjusted size and N is population size. Nonresponse requires inviting more units than the number of completed responses desired. Cluster designs may require a larger sample because of within-cluster similarity.

9. An Illustrative HR Case

Background

The following is a fictional illustrative case. A 3,600-employee company reports declining engagement and wants to compare employee voice across job levels and regions.

Problem

The HR team proposes emailing a voluntary survey link to all employees and reporting the average score. Senior managers worry that frontline and night-shift employees may be underrepresented. Employees also fear that small-group results could identify individuals.

Analysis

The target population is all employees active on a specified census date. The sampling frame is the verified HR employee register, excluding inactive records. Job level and region are relevant strata because engagement and access to email may differ across them. The construct must be operationalized through a tested set of engagement items, with a defined scoring rule.

A stratified probability sample is selected, with minimum numbers for important subgroups. Survey invitations explain purpose, voluntary participation, confidentiality and reporting thresholds. Results are reported only for groups large enough to protect identity. Response rates are monitored by stratum, and reminders are tailored to work schedules rather than directed only through corporate email.

Outcome

The organization obtains more balanced subgroup representation than it would likely have received from an unrestricted volunteer survey. The HR team finds lower engagement among night-shift employees, but treats this as an association requiring further investigation rather than proof that shift work caused disengagement. Follow-up interviews and operational data are used to explore scheduling, supervision and workload.

Learning points

1. A census invitation is not equivalent to a census result when participation is voluntary.
2. Stratification can address representation, but it does not remove measurement or nonresponse bias.
3. Confidentiality must be designed into sampling, reporting and access controls.
4. Survey results should guide inquiry, not substitute for causal research.

10. Ethics, Governance and Technology

HR data often include sensitive information about health, disability, ethnicity, compensation, performance and protected characteristics. Ethical practice requires a legitimate purpose, data minimization, transparency, secure storage, controlled access and an appropriate retention period. Participation should not be coerced, and employees should understand how data will be used.

Organizations must comply with applicable privacy, employment and equality laws. Legal requirements vary by jurisdiction, so statistical competence does not replace legal advice. Small-cell reporting can create re-identification risk even when names are removed. Linking datasets increases analytical power but also increases privacy risk.

Artificial intelligence and automated analytics can identify patterns in recruitment, absence or turnover, but models inherit defects in historical data. A model trained on biased promotion decisions may reproduce discrimination. Governance should include documented purpose, representative validation data, bias testing, human oversight, explainability appropriate to the use, audit trails and a mechanism for correction or appeal. Data quality and measurement validity remain necessary even when sophisticated software is used.

Best-Practice Checklist

Before collecting or analyzing HR data, ask:

1. What is the precise research question and unit of analysis?
2. What is the target population and what frame represents it?
3. Are variables operationally defined and appropriately scaled?
4. Is the instrument reliable and supported by validity evidence?
5. What sampling design best supports the intended inference?
6. How will nonresponse, missingness and weighting be handled?
7. Are definitions, codes, dates and transformations documented?
8. Are privacy, fairness, security and retention requirements satisfied?
9. Are conclusions limited to what the design can support?

Chapter Summary

Data provide the recorded foundation for statistical reasoning, but useful information depends on appropriate definitions, measurement and context. Variables may be categorical or quantitative, discrete or continuous, and explanatory or response. Measurement scales—nominal, ordinal, interval and ratio—determine which comparisons and calculations are defensible. Operationalization translates abstract HR constructs into observable measures.

Population parameters differ from sample statistics. Probability sampling supports generalization when the sampling frame, selection procedure and implementation are sound. Simple random, systematic, stratified, cluster and multistage designs serve different purposes. Non-probability samples may be practical but commonly limit statistical generalization. Sampling error can be quantified under appropriate assumptions, whereas systematic bias is not corrected simply by increasing sample size.

Reliable and valid measurement, careful questionnaire design, documentation, nonresponse analysis and ethical governance are central to credible HR analytics. Technology and AI can improve efficiency, but they cannot compensate for poor data foundations or justify unsupported causal claims.

Key Terms

Data: Recorded observations, measurements, classifications or responses.

Information: Organized and interpreted data used to support understanding or decisions.

Variable: A characteristic that can take different values across units.

Unit of analysis: The entity about which a conclusion is made.

Population: The complete group relevant to a study.

Sample: A subset selected from a population.

Parameter: A numerical characteristic of a population.

Statistic: A numerical characteristic calculated from a sample.

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

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

Nominal scale: A scale of categories without inherent order.

Ordinal scale: An ordered scale without necessarily equal intervals.

Reliability: Consistency of a measurement process.

Validity: Evidence that a measure supports its intended interpretation or use.

Sampling error: Natural variation between a sample statistic and the population parameter.

Nonresponse: Failure of selected units to provide usable data.

Stratified sampling: Sampling separately within defined population subgroups.

Cluster sampling: Sampling naturally occurring groups rather than, or before, individual units.

Measurement bias: Systematic deviation caused by the measurement process.

Data dictionary: Documentation describing variables, codes, units and permitted values.

Review Questions

Short Answer Questions

1. Distinguish a population parameter from a sample statistic.
2. Why is a numerical code for employment status not necessarily quantitative data?
3. State one difference between reliability and validity.
4. What is the purpose of a sampling frame?
5. Give two possible sources of nonresponse bias.

Long Answer Questions

1. Explain the four measurement scales and illustrate each with an HR example.
2. Compare simple random, stratified and cluster sampling in terms of procedure, strengths and limitations.
3. Explain how an abstract construct such as engagement can be operationalized and evaluated.
4. Discuss why increasing sample size does not necessarily eliminate bias.
5. Describe the ethical and governance requirements for collecting sensitive employee data.

Analytical / Application Questions

1. An organization surveys only employees who attend a voluntary lunchtime webinar. Identify the likely target-population, coverage and self-selection problems, and propose a better design.
2. A company has 1,500 employees and wants to estimate the proportion satisfied with benefits. Explain what information is needed before calculating a defensible sample size.
3. A predictive turnover model uses historical performance ratings and demographic variables. Design an audit plan addressing data quality, validity, fairness, privacy and interpretation of results.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a parameter from a statistic?

A. A parameter describes a sample, whereas a statistic describes a population.

B. A parameter describes a population, whereas a statistic is calculated from a sample.

C. A parameter is always categorical, whereas a statistic is always numerical.

D. A parameter is observed data, whereas a statistic is always missing data.

Correct Answer: B. A parameter describes a population, whereas a statistic is calculated from a sample.

Question 2

Which measurement scale is most appropriate for employee department?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: A. Nominal

Question 3

An organization selects employees separately within each job level to ensure representation. Which sampling method is being used?

A. Convenience sampling

B. Cluster sampling

C. Stratified sampling

D. Snowball sampling

Correct Answer: C. Stratified sampling

Question 4

A sample contains 80 employees from a workforce of 800. What is the sampling fraction?

A. 0.01

B. 0.10

C. 0.20

D. 10.00

Correct Answer: B. 0.10

Question 5

Which statement about reliability is correct?

A. It proves that a measure captures the intended construct.

B. It concerns consistency of measurement.

C. It applies only to ratio-scale variables.

D. It eliminates all systematic bias.

Correct Answer: B. It concerns consistency of measurement.

Question 6

In the formula SE(x̄) = s ÷ √n, what generally happens to the standard error when n increases, assuming s is unchanged?

A. It increases linearly.

B. It remains exactly unchanged.

C. It decreases as the square root of sample size increases.

D. It becomes equal to the population mean.

Correct Answer: C. It decreases as the square root of sample size increases.

Question 7

Which is an example of a flow measure?

A. Employees on the payroll on 31 December

B. Number of employees currently in a department

C. Hires during a quarter

D. Average tenure on a specified date

Correct Answer: C. Hires during a quarter

Question 8

Why can a voluntary employee survey be biased even when every employee receives an invitation?

A. It is automatically a probability sample.

B. Volunteers may differ systematically from nonparticipants.

C. A census invitation prevents measurement error.

D. The sample size must equal zero.

Correct Answer: B. Volunteers may differ systematically from nonparticipants.

Question 9

Which practice best protects confidentiality in subgroup reporting?

A. Publishing individual names with scores

B. Reporting every subgroup regardless of size

C. Using minimum reporting thresholds for small groups

D. Removing the survey reference period

Correct Answer: C. Using minimum reporting thresholds for small groups

Question 10

If no prior estimate of a population proportion is available, which planning value commonly gives the largest required sample size in n = z²p(1 − p) ÷ E²?

A. p = 0.01

B. p = 0.10

C. p = 0.50

D. p = 1.00

Correct Answer: C. p = 0.50