Business Data, Measurement and Quality

Introduction

Business analytics begins before a statistical model is selected. It begins with deciding what is to be measured, how observations will be obtained, and whether the resulting data are sufficiently accurate and complete for the decision at hand. A sophisticated model cannot compensate for a poorly defined variable, an incomplete sampling frame, systematic non-response, or duplicated records.

This chapter develops the foundations required to work responsibly with business data. It distinguishes qualitative and quantitative data, explains the four principal measurement scales, examines populations, samples and sampling frames, and compares common data-collection methods. It then introduces practical methods for identifying bias, missing values, outliers and other data-quality problems. The emphasis is not merely on classification, but on deciding what a dataset can legitimately support.

Learning Objectives

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

1. Classify variables as qualitative or quantitative and distinguish discrete from continuous quantitative data.
2. Select an appropriate measurement scale and explain its implications for analysis.
3. Define a population, sample, sampling frame and sampling unit.
4. Compare probability and non-probability sampling methods.
5. Evaluate surveys, experiments, observations, interviews and secondary data sources.
6. Detect and document bias, missing values, duplicates, inconsistent formats and outliers.
7. Apply a structured data-quality workflow to a business problem.
8. Recognise ethical, privacy and governance responsibilities in data collection and preparation.

1. What Is Business Data?

Data are recorded representations of observations, transactions, events or judgments. Information is data interpreted in a useful context. For example, the value 125 is data; “125 units sold in Store A during March” is more informative because its meaning, unit, location and time are specified.

A unit of analysis is the entity about which a conclusion is made. It may be a customer, invoice, store, employee, product, website visit or day. A variable is a characteristic recorded for each unit, such as customer age, order value or payment method. An individual recorded value is an observation.

A data dictionary should normally specify:

The distinction between a variable’s meaning and its storage format is important. A postal code may be stored as digits but is usually qualitative: arithmetic on postal codes is meaningless. Conversely, a quantity may be stored as text because of a faulty import even though it is quantitative in meaning.

1.1 Qualitative and quantitative data

Qualitative data describe categories, attributes or labels. Examples include product category, region, customer type and payment method. Categories may be:

• Nominal, where there is no meaningful order, such as browser type.
• Ordinal, where categories have an order, such as low, medium and high satisfaction.

Quantitative data represent numerical amounts for which arithmetic operations have substantive meaning. They may be:

• Discrete, arising from counts, such as number of complaints or units sold. Discrete values are often whole numbers, although the relevant issue is whether intermediate values are meaningful.
• Continuous, arising from measurement, such as delivery time, weight or temperature. In principle, continuous variables can take any value within an interval, subject to measurement precision.

A categorical variable can be coded numerically without becoming quantitative. If “online” is coded 1 and “store” is coded 2, the average of these codes has no business interpretation.

1.2 Structured, semi-structured and unstructured data

Structured data fit a predefined schema, such as rows in a sales table. Semi-structured data contain some organisation but not a fixed relational format, such as JSON transaction logs or XML documents. Unstructured data include free-text reviews, images, audio and video. A business analytics project may combine all three, but each requires different methods of validation and interpretation.

2. Measurement and Scales of Measurement

Measurement is the assignment of values or labels to attributes according to defined rules. A measurement scale determines which comparisons and calculations are justified. The traditional hierarchy consists of nominal, ordinal, interval and ratio scales.

2.1 Nominal scale

Nominal values identify categories. The numbers used to represent categories are labels rather than quantities. A frequency table and percentage distribution are usually appropriate. The mode is the most frequently observed category. A mean is not meaningful for nominal codes.

2.2 Ordinal scale

Ordinal values indicate relative position, but the distances between adjacent categories are not necessarily equal. The difference between “poor” and “fair” may not be equivalent to the difference between “good” and “excellent.” Medians, percentiles and rank-based comparisons are generally defensible. Treating a five-point satisfaction scale as numerical may sometimes be a practical modelling choice, but it is an assumption that should be stated and tested rather than silently made.

2.3 Interval scale

An interval scale has equal intervals, but zero is arbitrary rather than an absence of the measured attribute. For example, 20°C is not twice as warm as 10°C. Differences are meaningful, while ratio statements are not. Calendar dates can support differences such as elapsed days, but “year 2000 is twice year 1000” is meaningless.

2.4 Ratio scale

A ratio scale has equal intervals and a meaningful zero. Revenue of £200 is twice revenue of £100, and a duration of 20 minutes is twice 10 minutes. Ratio-scale variables support addition, subtraction, multiplication and division, provided the units and context are appropriate.

2.5 Operational definitions

An operational definition states exactly how a concept will be measured. “Customer loyalty” is a concept, not yet a variable. Possible operationalisations include repeat-purchase rate, number of purchases in 12 months, or an agreed survey scale. Different operationalisations can produce different conclusions. A sound definition specifies the population, time period, unit, calculation rule and treatment of exceptional cases.

For example, a monthly conversion rate may be defined as:

Here, “eligible session” must be defined. If automated bots and internal employee visits are included in the denominator, the rate may be misleading. The formula is appropriate only when the numerator is a subset of the denominator and both refer to the same period and population.

3. Populations, Samples and Sampling Frames

A population is the complete set of units relevant to a question. A retailer studying all customers who purchased in 2025 has a finite target population. A sample is a subset observed in order to learn about that population. A census attempts to observe every unit in the defined population.

The sampling unit is the basic unit selected, such as a customer, household, store or transaction. The sampling frame is the operational list or mechanism from which the sample is drawn. The frame may be a customer database, employee register, list of branches or set of website visitors.

Three populations should be distinguished:

1. Target population: the population to which the organisation wants to generalise.
2. Frame population: units actually represented by the sampling frame.
3. Study population: units that meet the practical inclusion criteria for the study.

A gap between these populations creates coverage error. For example, a survey sent only to loyalty-program members cannot automatically represent all customers. A frame may also contain duplicate, outdated or ineligible entries.

3.1 Sampling error and non-sampling error

Sampling error is the natural difference between a sample result and the corresponding population value caused by observing only part of the population. It tends to decrease as a well-designed random sample becomes larger, although it is not eliminated completely.

Non-sampling errors include coverage error, non-response, measurement error, interviewer effects, processing errors and careless coding. Increasing the sample size does not necessarily correct these problems and can sometimes make a biased result appear more precise.

For a simple random sample estimating a population proportion, an approximate standard error is:

where p̂ is the sample proportion and n is the sample size. This approximation assumes a probability sample, independent observations and a sufficiently large sample for the normal approximation. It describes sampling variability, not bias. When sampling without replacement from a substantial fraction of a finite population, a finite-population correction may be relevant:

where N is population size and n is sample size. The correction is commonly considered when the sampling fraction is large. It does not repair a defective frame or non-response bias.

3.2 Probability sampling

In probability sampling, each eligible unit has a known, non-zero probability of selection.

• Simple random sampling gives each unit an equal selection opportunity. It is conceptually strong but requires a complete and accurate frame.
• Systematic sampling selects every kth unit after a random starting point. If N units are in the frame and n are required, a basic interval is k ≈ N ÷ n. Periodic patterns in the list can make this method biased.
• Stratified sampling divides the population into relevant subgroups, or strata, and samples within each. It can ensure representation of small but important groups and improve precision when strata are internally similar.
• Cluster sampling selects groups such as branches or schools and then observes units within selected groups. It can reduce travel and administration costs, but units in the same cluster may be similar, reducing statistical efficiency.

3.3 Non-probability sampling

Convenience sampling selects easily accessible units. Voluntary-response sampling relies on people choosing to participate. Quota sampling fills category targets without necessarily selecting randomly within categories. Purposive sampling deliberately selects cases judged relevant by the researcher.

These methods may be useful for exploratory work, qualitative research or hard-to-reach populations. However, selection probabilities are not known, so sampling error and generalisation are difficult to quantify. Claims should be limited to the observed participants or clearly qualified.

4. Data-Collection Methods

4.1 Surveys and questionnaires

Surveys efficiently collect standardised responses from many people. Good questions are specific, neutral and relevant. A question should not be double-barrelled, such as “How satisfied are you with delivery speed and packaging?” because a respondent may hold different views about the two issues. Response options should be exhaustive and mutually exclusive where appropriate, and “not applicable” should be offered when necessary.

Leading wording, confusing scales, recall demands and excessive length can create measurement error. A pilot test can reveal ambiguous wording, missing response options and technical problems. Response rates should be reported, but a high response rate alone does not prove representativeness.

4.2 Interviews and focus groups

Interviews permit clarification and detailed exploration. They are valuable when the subject is complex, but interviewer behaviour and social desirability can influence answers. Focus groups generate interaction and insight into language and perceptions; they are not normally designed to estimate population percentages.

4.3 Observation and transaction data

Observation records behaviour rather than relying entirely on self-report. Point-of-sale systems, web logs and sensors can provide high-volume behavioural data. Nevertheless, logs reflect system design: events may be missing, duplicated, filtered or recorded differently after a software change. Observational data reveal association and behaviour but do not automatically establish causation.

4.4 Experiments and A/B tests

An experiment deliberately assigns conditions to units and compares outcomes. In an A/B test, eligible users may be randomly assigned to version A or B. Random assignment helps balance confounding factors on average, while a predefined outcome and analysis plan reduce selective interpretation. Ethical and operational constraints may limit experimentation, and results may not generalise beyond the tested population or period.

4.5 Secondary data

Secondary data were collected for another purpose. Examples include government statistics, industry reports, accounting systems and public datasets. Before use, evaluate provenance, definitions, collection date, coverage, revisions, licensing restrictions and comparability. A dataset can be accurate for its original purpose yet unsuitable for a new question because its unit, time period or definitions differ.

5. Data Quality: Dimensions and Workflow

Data quality is fitness for a defined use, not an absolute property. Important dimensions include:

A practical quality workflow is:

1. Define the analytical purpose and acceptance criteria.
2. Inspect the schema, metadata and data dictionary.
3. Profile fields using counts, distinct values, ranges and missingness.
4. Validate formats, domains, relationships and business rules.
5. Investigate anomalies with source owners.
6. Correct, exclude, impute or retain values using documented rules.
7. Re-profile after cleaning and preserve an audit trail.
8. Monitor quality over time rather than treating cleaning as a one-time task.

5.1 Missing values

A missing value may mean “not collected,” “not applicable,” “refused,” “unknown” or “not yet available.” These meanings must not be collapsed automatically. Common mechanisms are:

• MCAR: missing completely at random; missingness is unrelated to observed and unobserved values.
• MAR: missing at random conditional on observed variables.
• MNAR: missingness depends on the unobserved value itself or related unobserved factors.

The labels describe assumptions about the process, not convenient descriptions of the dataset. Missingness should be quantified by field and subgroup. Possible treatments include analysis of complete cases, an explicit “unknown” category, single-value replacement, model-based imputation or collecting the data again. Each has consequences. Replacing missing income with the mean reduces apparent variation and may distort relationships. Any imputation should be flagged and sensitivity-tested where decisions are material.

5.2 Invalid values and inconsistent formats

Validation rules should reflect business meaning. Examples include:

• quantity must be a non-negative integer;
• discount must lie between 0% and 100%;
• dispatch date cannot precede order date;
• a country code must follow the agreed standard;
• an identifier should not be silently converted to a number if leading zeros matter.

Dates such as 03/04/2025 are ambiguous across locales. Use an unambiguous standard such as YYYY-MM-DD internally, while displaying dates according to user requirements.

5.3 Duplicates and entity resolution

Duplicate rows may represent repeated ingestion, or they may be legitimate repeated transactions. Deduplication requires a business definition of uniqueness. A transaction identifier may be unique, whereas a customer may legitimately place many orders. Near-duplicates arise from spelling variation, changed addresses or multiple identifiers. Matching rules should be conservative, documented and reviewed because false matches can merge different people and create serious privacy and financial errors.

5.4 Outliers

An outlier is an observation unusually distant from much of the data. It may be a recording error, a valid rare event, a different population or evidence of a process change. It is not automatically an error.

For a sample with mean x̄ and standard deviation s, a standardised score is:

where x is an observation. Large absolute values may flag unusual observations, but no universal cutoff proves that a value is wrong. The mean and standard deviation themselves can be strongly affected by extreme observations.

For a boxplot-based rule, define the interquartile range as:

where Q₁ and Q₃ are the first and third quartiles. A common screening rule flags values below Q₁ − 1.5 × IQR or above Q₃ + 1.5 × IQR. This is a flagging convention, not a deletion rule. Investigate the source, compare with business context, and consider robust summaries, transformations or separate segment analysis.

5.5 Bias

Bias is systematic error that causes a method or result to depart from the target quantity. Important forms include:

• Selection bias: included units differ systematically from excluded units.
• Non-response bias: respondents differ from non-respondents in relevant ways.
• Coverage bias: the frame omits parts of the target population.
• Measurement bias: wording, instrument or process systematically distorts values.
• Survivorship bias: analysis includes only cases that remain observable.
• Confirmation bias: analysts favour evidence supporting an existing belief.
• Automation or algorithmic bias: historical or design choices produce systematically unequal outcomes.

Bias should be addressed through design where possible: improve the frame, randomise selection, pilot instruments, monitor subgroup response, document exclusions and compare alternative specifications. Weighting can sometimes adjust known selection differences, but it cannot reliably recover information about completely unobserved groups.

6. Applied Case: Improving a Retail Customer Survey

The following is an illustrative case.

Background

A national retailer wants to estimate customer satisfaction after a new delivery service is introduced. Management proposes emailing a survey to all loyalty-program members who made an online purchase during the previous month.

Problem

The target population is all customers who received an online delivery, but the proposed frame excludes non-members and customers whose email addresses are missing or invalid. Customers with strong positive or negative experiences may be more likely to respond. The proposed survey also asks, “How satisfied were you with our fast delivery and helpful support?” This is double-barrelled and may not apply to customers who did not contact support.

Analysis and approach

The analytics team first defines the unit of analysis as a completed delivery and creates a frame from delivery records rather than loyalty membership. It removes cancelled orders, identifies duplicate delivery IDs and records frame exclusions. Customers are stratified by region and delivery outcome, with random selection within strata. A short pilot tests the instrument.

The revised questionnaire asks separate questions about delivery time, condition of goods and support contact, each with a “not applicable” option. It records response status and sends limited reminders. Response rates and satisfaction distributions are compared across region, customer type and delivery outcome. The team reports the frame coverage, selection method, response rate and limitations rather than presenting the respondent average as automatically representative.

Outcome

The pilot reveals that customers with delayed deliveries are under-represented among respondents and that one region has a high proportion of invalid email addresses. The retailer therefore supplements email with SMS invitations where consent exists, investigates the regional data problem and reports subgroup results separately. A modest weighting adjustment is considered using known delivery-status distributions, but the report explains that weighting cannot correct all non-response bias.

Learning points

The main improvement did not come from a more complicated statistical model. It came from aligning the target population and sampling frame, clarifying measurement, monitoring missingness and documenting limitations. Data quality is therefore part of research design, not merely a technical cleaning stage after collection.

7. Ethics, Governance and Responsible Use

Data collection should have a legitimate purpose, a lawful basis where required, and a level of collection proportionate to that purpose. Organisations should communicate what is collected, why it is collected, how long it is retained and who may access it. Personal data should be minimised, protected and de-identified where feasible. Access controls, encryption, retention schedules and audit logs support responsible governance.

Consent is not a universal solution to poor design. It should be informed, specific and freely given where consent is the applicable basis. Analysts should avoid re-identification, discriminatory targeting and the use of sensitive attributes without strong justification and safeguards. Automated decisions require attention to explainability, human oversight, fairness and monitoring for unequal error rates.

Data lineage records where data originated, how it changed and which version supported a report. Reproducible preparation scripts and a data-quality log make errors easier to detect and decisions easier to defend.

8. Contemporary Practice and Best Practices

Modern organisations increasingly combine relational databases, cloud platforms, application programming interfaces, sensors, text and machine-generated data. Automated validation can detect schema changes, impossible values and unusual volumes. Machine-learning systems can assist entity matching and anomaly detection, but they should support—not replace—human review. Models may reproduce historical bias and can be confidently wrong when source data change.

Recommended practice is to:

• begin with a precise decision and operational definitions;
• maintain a data dictionary and ownership responsibilities;
• distinguish identifiers, categories, measurements and derived fields;
• use probability sampling when population generalisation is required;
• pilot collection instruments and monitor subgroup coverage;
• profile missingness and anomalies before modelling;
• never delete outliers without a documented reason;
• preserve raw data separately from cleaned analytical data;
• record transformations, assumptions and exclusions;
• assess privacy, fairness and security throughout the data lifecycle.

Chapter Summary

Business data consist of observations about defined units, recorded through variables with particular meanings and measurement scales. Qualitative variables represent categories, while quantitative variables represent amounts; numerical coding does not by itself make a variable quantitative. Nominal, ordinal, interval and ratio scales determine which comparisons and calculations are defensible.

A population is the complete set of relevant units, whereas a sample is the observed subset. The sampling frame operationalises selection, and differences between the target population and frame can create coverage error. Probability sampling supports stronger generalisation than convenience or voluntary-response sampling, although all designs remain vulnerable to non-sampling error.

Surveys, interviews, observations, experiments and secondary sources serve different purposes and have different risks. Data quality includes accuracy, completeness, validity, consistency, uniqueness, timeliness and integrity. Missing values, invalid formats, duplicates and outliers require investigation rather than automatic deletion or replacement. Bias is systematic error and cannot generally be solved simply by increasing sample size.

Sound analytics therefore depends on clear operational definitions, appropriate sampling, careful measurement, documented quality checks, ethical governance and transparent communication of limitations.

Key Terms

Accuracy: Degree to which a recorded value represents the real-world attribute.

Bias: Systematic error that shifts results away from the target quantity.

Census: Attempt to observe every unit in a defined population.

Cluster sampling: Sampling groups of units, such as branches, rather than selecting all units individually.

Continuous data: Quantitative data that can, in principle, take any value within an interval.

Data dictionary: Documentation defining fields, meanings, types, units and permitted values.

Frame population: Units represented by the sampling frame.

Interquartile range: Difference between the third and first quartiles.

Measurement error: Difference between a recorded value and the intended attribute caused by the measurement process.

Missing value: An absent or unavailable recorded value whose meaning should be investigated.

Nominal scale: Measurement scale containing categories without meaningful order.

Ordinal scale: Scale with meaningful order but not necessarily equal intervals.

Outlier: An observation unusually distant from much of the data.

Population: Complete set of units relevant to a research or business question.

Quantitative data: Numerical data representing meaningful amounts or counts.

Ratio scale: Scale with equal intervals and a meaningful zero.

Sampling frame: Operational list or mechanism from which a sample is selected.

Sampling unit: Basic entity selected during sampling.

Stratified sampling: Sampling separately within defined subgroups or strata.

Validity: Extent to which data conform to specified rules and represent the intended concept.

Review Questions

Short Answer Questions

1. Distinguish qualitative data from quantitative data and give one example of each.
2. Why is a postal code usually treated as nominal data even when stored numerically?
3. Define population, sample and sampling frame.
4. State two causes of missing values and explain why their meanings may differ.
5. Why should an outlier not be deleted automatically?

Long Answer Questions

1. Explain the four measurement scales and discuss the statistical summaries appropriate for each.
2. Compare simple random, stratified, cluster and convenience sampling in terms of design and limitations.
3. Discuss how questionnaire wording can create measurement bias. Include examples of poor and improved questions.
4. Explain the dimensions of data quality and propose a quality-assurance workflow for a sales dataset.
5. Discuss the ethical and governance responsibilities associated with collecting and analysing customer data.

Analytical / Application Questions

1. A company surveys only customers who opened a promotional email. Identify likely sources of bias and design a better sampling approach for estimating satisfaction among all customers.
2. A transaction file contains negative quantities, duplicated order IDs, dates in multiple formats and unusually large orders. Propose an investigation and treatment plan, explaining which values should be corrected, retained or escalated.
3. A manager claims that a satisfaction score coded from 1 to 5 is ratio-scale data because 4 is twice 2. Evaluate the claim and recommend an appropriate analysis strategy.

Multiple Choice Questions with Answers

Question 1

Which variable is quantitative and discrete?

A. Payment method

B. Number of returned items

C. Customer region

D. Satisfaction category

Correct Answer: B. Number of returned items

Question 2

Which measurement scale has ordered categories but does not guarantee equal distances between categories?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: B. Ordinal

Question 3

A sampling frame is best described as:

A. The complete set of conclusions from a study

B. The operational list or mechanism from which a sample is selected

C. The number of observations required in a sample

D. A summary of the population mean

Correct Answer: B. The operational list or mechanism from which a sample is selected

Question 4

Which method ensures that specified subgroups are represented by sampling within each subgroup?

A. Cluster sampling

B. Convenience sampling

C. Stratified sampling

D. Voluntary-response sampling

Correct Answer: C. Stratified sampling

Question 5

If 80 of 400 eligible website sessions result in purchases, what is the conversion rate?

A. 5%

B. 20%

C. 32%

D. 80%

Correct Answer: B. 20%

Question 6

Which issue occurs when a customer database excludes people who are not loyalty-program members, although the target population includes all customers?

A. Coverage error

B. Random sampling error

C. Unit conversion error

D. Rounding error

Correct Answer: A. Coverage error

Question 7

Which statement about an outlier is correct?

A. It is always a data-entry mistake

B. It must always be removed before analysis

C. It should be investigated in relation to its source and context

D. It proves that the sample is biased

Correct Answer: C. It should be investigated in relation to its source and context

Question 8

Replacing every missing income value with the sample mean will generally:

A. Increase the observed variation

B. Reduce the observed variation and potentially distort relationships

C. Eliminate all non-response bias

D. Convert qualitative data into ratio data

Correct Answer: B. Reduce the observed variation and potentially distort relationships

Question 9

Which is an example of a double-barrelled survey question?

A. How many orders did you place last month?

B. Which delivery channel did you use?

C. How satisfied were you with delivery speed and packaging?

D. Was the order applicable to you?

Correct Answer: C. How satisfied were you with delivery speed and packaging?

Question 10

A numerical code assigned to payment methods should generally be treated as:

A. Ratio-scale data

B. Interval-scale data

C. Continuous data

D. Categorical data

Correct Answer: D. Categorical data