Statistical Foundations for Analytics

Introduction

Business analytics depends on the disciplined conversion of observations into evidence for decisions. Statistical foundations provide the language and methods required to describe customers, employees, transactions, operations and markets without overstating what the data can support. A manager who understands measurement, sampling, probability and uncertainty is better equipped to distinguish a genuine pattern from random variation.

Statistics is commonly divided into two connected activities. Descriptive statistics organise and summarise observed data. Inferential statistics use sample data to make cautious statements about a wider population. Both are essential: a sophisticated inferential method cannot repair poorly measured or biased data.

This chapter develops the core statistical ideas needed to interpret business data: variables and measurement scales, populations and samples, probability, distributions, descriptive measures, confidence intervals, correlation and basic hypothesis testing.

Learning Objectives

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

1. Classify variables and measurement scales appropriately.
2. Explain populations, samples and common sampling methods.
3. Interpret basic probability concepts and business probability models.
4. Calculate and interpret measures of centre, spread and position.
5. Explain normal and other useful distributions.
6. Construct and interpret confidence intervals.
7. Distinguish correlation from causation.
8. Conduct and interpret a basic hypothesis test.
9. Identify statistical, ethical and data-governance risks in business analysis.

1. Variables and Measurement

A variable is a characteristic that can take different values across units such as customers, products, stores or transactions. An individual observed value is called an observation. A dataset is often represented as a rectangular table in which rows are cases and columns are variables.

Variables may be classified as follows:

The same variable may be treated differently depending on the analytical purpose. Age recorded in years is numerical, but an organisation may categorise it into age bands for reporting. Categorisation can simplify communication but discards information and may create arbitrary boundaries.

Measurement Scales

The measurement scale determines which comparisons and calculations are meaningful.

• Nominal scale: categories have no inherent order. Product category and region are nominal. Frequencies and proportions are appropriate; calculating an average category code is not.
• Ordinal scale: categories have an order, but gaps between categories are not necessarily equal. Satisfaction ratings such as poor, fair, good and excellent are ordinal. Medians and rank-based methods are often appropriate.
• Interval scale: ordered values have equal intervals, but zero does not represent an absolute absence. Calendar years and temperature in Celsius are examples. Differences are meaningful, but ratios such as “20°C is twice as hot as 10°C” are not.
• Ratio scale: equal intervals and a meaningful zero exist. Revenue, units sold, distance and elapsed time are ratio variables. Both differences and ratios can be interpreted.

A data dictionary should record each variable’s name, definition, unit, allowable values, source, time period and treatment of missing values. This practice prevents errors such as confusing dollars with thousands of dollars or treating a coded category as a numerical quantity.

2. Populations, Samples and Sampling

A population is the complete set of units relevant to a question. A sample is a subset observed or measured from that population. A numerical feature of a population is a parameter; the corresponding value calculated from a sample is a statistic. For example, the population mean is often denoted by μ, while the sample mean is denoted by x̄.

Sampling is necessary when a census is too costly, slow or impossible. The central challenge is obtaining a sample that represents the target population. A large biased sample can be less useful than a smaller, well-designed sample.

Important sampling methods include:

• Simple random sampling: every population unit has a known and equal selection probability.
• Systematic sampling: select every kth unit after a random starting point. It can be efficient but may be biased if the list has periodic patterns.
• Stratified sampling: divide the population into relevant strata, such as regions, then sample within each stratum. This can improve representation of important subgroups.
• Cluster sampling: select groups such as stores or branches and observe units within selected groups. It may reduce cost but often increases sampling variability.
• Convenience sampling: select easily accessible units. It is inexpensive but usually vulnerable to selection bias.
• Voluntary-response sampling: individuals choose whether to participate. People with strong opinions may be overrepresented.

Sampling Error and Bias

Sampling error is random difference between a sample result and the population parameter. It generally decreases as sample size increases, although not necessarily at a linear rate. Bias is systematic error caused by the design, collection or processing of data. Nonresponse, coverage problems, leading questions and survivorship bias are common examples. Increasing the sample size does not reliably remove bias.

A useful sampling plan specifies the target population, sampling frame, selection method, sample size, inclusion criteria, nonresponse procedures and weighting approach. Analysts should report who was excluded and whether the observed sample differs from the intended population.

3. Probability for Business Decisions

Probability quantifies uncertainty. For an event A, its probability P(A) lies between 0 and 1. If outcomes are mutually exclusive and collectively exhaustive, their probabilities sum to 1.

The complement rule is:

where Aᶜ means that event A does not occur.

For two events A and B, the addition rule is:

The intersection A ∩ B means that both events occur. The subtraction avoids counting their overlap twice. If A and B are mutually exclusive, P(A ∩ B) = 0.

Conditional probability is the probability of A given that B has occurred:

provided P(B) > 0. Events are independent when knowing that B occurred does not change the probability of A, so P(A | B) = P(A). In that case:

Independence must not be assumed merely because two events seem unrelated.

Worked Example: Conditional Probability

A retailer finds that 8% of orders are returned. Of returned orders, 30% were placed using a promotional code. Of all orders, 20% used a promotional code. The probability that an order was returned and used a promotional code is 0.08 × 0.30 = 0.024, or 2.4%. The probability of return given promotional-code use is 0.024 ÷ 0.20 = 0.12, or 12%. This is higher than the overall return rate, suggesting an association that merits investigation; it does not by itself establish that the promotion caused returns.

4. Distributions and Random Variables

A random variable assigns a numerical value to each possible outcome of a random process. A discrete random variable has countable outcomes, whereas a continuous random variable can take values across an interval. A probability distribution describes the probabilities associated with possible values.

The expected value of a discrete random variable X is:

The sum is taken across all possible values x. Expected value is a long-run average, not necessarily a value that must occur in an individual case.

The binomial distribution models the number of successes in n independent trials when each trial has two outcomes and a constant success probability p. Its probability function is:

where k is the number of successes and C(n,k) is the number of ways to choose k successes from n trials. The model is appropriate only when trials are independent, n is fixed, outcomes are binary and p is constant.

The normal distribution is continuous, symmetric and bell-shaped. It is described by a mean μ and standard deviation σ. A standardised value, or z-score, is:

It indicates how many standard deviations x lies above or below the mean. The empirical rule states that approximately 68%, 95% and 99.7% of observations lie within 1, 2 and 3 standard deviations of the mean, respectively, when the distribution is approximately normal. These percentages are approximations, not universal guarantees.

The central limit theorem explains why sample means often have an approximately normal sampling distribution when the sample size is sufficiently large, even when individual observations are not normal. Conditions depend on the population distribution, independence and sample size. Strong skewness, outliers or dependence may require a larger sample or alternative methods.

5. Descriptive Statistics

Measures of Centre

The arithmetic mean is:

where xᵢ is observation i and n is the number of observations. The mean uses every value but is sensitive to extreme observations.

The median is the middle ordered value, or the average of the two middle values when n is even. It is resistant to extreme values. The mode is the most frequent value and is especially useful for categorical data.

For positively skewed business data, such as income or transaction value, the mean is often greater than the median. Reporting both can reveal this feature more effectively than reporting either alone.

Measures of Spread

The range is maximum minus minimum. It is easy to interpret but depends heavily on two observations. The sample variance is:

The sample standard deviation is:

The denominator n − 1, known as Bessel’s correction, is used when estimating population variability from a sample. For a complete population, the population variance uses N rather than N − 1, where N is the population size. Standard deviation has the same units as the original variable; variance has squared units.

The interquartile range is IQR = Q₃ − Q₁, where Q₁ and Q₃ are the first and third quartiles. It describes the spread of the middle 50% and is resistant to outliers. A common exploratory rule flags values below Q₁ − 1.5IQR or above Q₃ + 1.5IQR as potential outliers. Such values should be investigated, not automatically deleted.

Worked Example: Descriptive Measures

Suppose five delivery times, in minutes, are 20, 22, 25, 25 and 38. The mean is 130 ÷ 5 = 26 minutes, while the median is 25 minutes. The range is 38 − 20 = 18 minutes. Deviations from the mean are −6, −4, −1, −1 and 12; their squared sum is 198. Thus sample variance is 198 ÷ 4 = 49.5 minutes² and sample standard deviation is √49.5 ≈ 7.04 minutes. The higher mean than median reflects the long delivery time of 38 minutes.

6. Confidence Intervals

A confidence interval estimates an unknown population parameter using a range of plausible values. For a population mean, when the population standard deviation is unknown, a common interval is:

Here x̄ is the sample mean, s is the sample standard deviation, n is sample size and t* is the critical value from the t distribution with n − 1 degrees of freedom. The formula assumes a random or appropriately designed sample, independent observations and a sufficiently well-behaved population distribution. For large samples, the t critical value approaches the corresponding standard-normal value.

The quantity s ÷ √n is the standard error of the sample mean. A 95% confidence interval procedure captures the true population mean in approximately 95% of repeated samples under the model assumptions. It is not technically correct to say that there is a 95% probability that a fixed parameter lies in this already calculated interval.

Worked Example: Mean Order Value

A random sample of 64 orders has mean value $52 and sample standard deviation $16. Using the large-sample critical value 1.96, the standard error is 16 ÷ √64 = $2. The margin of error is 1.96 × $2 = $3.92. The 95% confidence interval is $52 ± $3.92, or ($48.08, $55.92). The practical interpretation is that the method gives intervals containing the population mean in about 95% of repeated samples, assuming the sampling and measurement conditions are appropriate.

Intervals become narrower with larger n and wider with greater variability or higher confidence. A narrow interval can still be misleading if the sample is biased or the variable is poorly measured.

7. Correlation and Association

Correlation measures the strength and direction of a linear relationship between two numerical variables. Pearson’s sample correlation coefficient r ranges from −1 to +1. Values near +1 indicate a strong positive linear association; values near −1 indicate a strong negative linear association; values near 0 indicate little linear association.

Correlation is not causation. A relationship may arise from a third variable, reverse direction of influence, coincidence, common trends or selection effects. Correlation can also miss a strong nonlinear relationship. Analysts should inspect a scatterplot, check outliers, consider time order and use subject-matter knowledge before interpreting r.

The square of the correlation, r², is sometimes called the coefficient of determination in a simple linear model. It describes the proportion of variation in the response linearly associated with the predictor in that sample, not the proportion proven to be caused by it.

For example, advertising expenditure and sales may correlate positively because advertising increases demand, but also because large stores have both larger budgets and higher sales. Randomised experiments or carefully designed quasi-experimental methods are stronger approaches for causal claims.

8. Basic Hypothesis Testing

Hypothesis testing provides a structured way to evaluate evidence against a claim. The null hypothesis, H₀, usually represents no difference, no association or a specified benchmark. The alternative hypothesis, H₁ or Hₐ, represents the effect or relationship of interest.

A test statistic measures how far the sample result is from the null value relative to expected sampling variation. A p-value is the probability, assuming H₀ is true and the model is appropriate, of obtaining a result at least as extreme as the observed result. A small p-value indicates that the data are less compatible with H₀; it does not prove H₁ and is not the probability that H₀ is true.

For a large-sample test of a population mean against μ₀, a common statistic is:

When σ is unknown and the sample is small under suitable assumptions, a t statistic with n − 1 degrees of freedom is generally used instead. Choose a significance level α, commonly 0.05, before examining results. Reject H₀ when the p-value is less than α; otherwise, fail to reject H₀. “Fail to reject” does not mean that H₀ has been proved true.

A Type I error rejects a true H₀; its controlled probability is α under the test procedure. A Type II error fails to reject a false H₀. Statistical significance does not necessarily imply practical importance. Analysts should report effect size, confidence interval, sample size and business consequences.

Worked Example: Testing a Delivery Target

A courier claims that mean delivery time is 30 minutes. A sample of 100 deliveries has x̄ = 31.2 minutes and s = 6 minutes. Test H₀: μ = 30 against Hₐ: μ > 30. The standard error is 6 ÷ √100 = 0.6 minutes, giving z = (31.2 − 30) ÷ 0.6 = 2.00. For a one-sided test at α = 0.05, the critical z value is approximately 1.645. Since 2.00 > 1.645, reject H₀ under this model. The result provides evidence that mean delivery time exceeds 30 minutes, but management should also assess whether the 1.2-minute difference is operationally important and whether the sample represents normal deliveries.

9. Applied Case: Improving an Online Retailer’s Returns Process

Background. Illustrative retailer NovaHome receives customer orders through web and mobile channels. Management observes that promotional campaigns appear to be associated with higher returns and wants to decide whether to change campaign design.

Problem. The initial dashboard compares return rates for promotion and non-promotion orders but includes only customers who completed a post-purchase survey. The analytics team must determine whether the apparent difference is reliable.

Analysis. First, the team creates a data dictionary defining order, return, promotion, product category and customer identifiers. It discovers that survey respondents are disproportionately frequent customers, creating possible nonresponse bias. The team then defines the target population as all completed orders during the quarter and draws a stratified random sample by product category and sales channel. It reports return proportions with confidence intervals and examines order value, delivery time and product type as potential confounders.

A descriptive analysis shows a 12% return rate for promoted orders and 8% for other orders. The difference is an association, not proof that promotion causes returns. Product categories promoted most heavily also have different sizing and fit risks. The team therefore compares similar categories and runs a controlled pilot in which comparable customers are randomly assigned to alternative promotion designs.

Approach and outcome. The pilot finds that a clearer size guide reduces returns without materially reducing conversion. The original apparent promotion effect becomes smaller after product category is considered. Management adopts the size guide, monitors return rate and conversion jointly, and reviews the experiment after four weeks.

Learning points. Sound analytics begins with a defined population and valid measurement. Sampling design matters, descriptive comparisons should precede modelling, confounding must be considered, and business action should be based on effect size and cost rather than statistical significance alone. Customer privacy also requires access controls, data minimisation and reporting that avoids unnecessary identification.

10. Limitations, Ethics and Best Practice

Statistical conclusions are only as credible as the data-generating process. Common limitations include missing data, measurement error, nonrepresentative samples, changing definitions, duplicated records, dependence between observations and multiple testing. Analysts should document assumptions and conduct sensitivity checks.

Ethical practice requires transparency about uncertainty, avoidance of misleading graphs, protection of confidential information and fair treatment of groups. Personal data should be collected for legitimate purposes, retained only as needed and accessed by authorised personnel. Automated decisions based on statistical scores should be assessed for disparate impact and reviewed by accountable people.

Useful professional practices include:

1. Define the decision and population before analysing.
2. Inspect raw data, units, missingness, duplicates and outliers.
3. Use plots and descriptive statistics before formal testing.
4. Match the method to the measurement scale and sampling design.
5. Separate exploratory findings from confirmatory claims.
6. Report uncertainty, effect size and limitations.
7. Avoid causal language unless the design supports causality.
8. Preserve reproducible code, versioned data definitions and an audit trail.
9. Protect privacy and communicate results in plain language.

Contemporary analytics platforms automate calculations and generate forecasts, but automation does not remove the need for statistical judgement. Artificial intelligence can identify patterns, yet it may amplify biased historical data or produce confident but unsupported explanations. Human review, validation on representative data and clear governance remain essential.

Chapter Summary

Statistical foundations enable analysts to move from raw observations to defensible business conclusions. Variables must be correctly defined and measured because the scale determines valid operations. Populations and samples distinguish the target of inference from the observed data, while sampling design affects representativeness and bias. Probability provides a framework for uncertainty, and distributions describe possible outcomes and sampling behaviour. Descriptive statistics summarise centre, spread and position, but must be interpreted with attention to skewness and outliers. Confidence intervals express estimation uncertainty. Correlation describes linear association but cannot establish causation. Hypothesis testing compares evidence with a null claim using p-values, significance levels and test statistics, while practical importance requires effect sizes and business context. Ethical analytics requires transparency, privacy protection, fairness and reproducibility.

Key Terms

Variable: A characteristic that can take different values.

Observation: A recorded value for a case or unit.

Population: The complete set relevant to a research or business question.

Sample: A subset selected from a population.

Parameter: A numerical feature of a population.

Statistic: A numerical feature calculated from a sample.

Nominal scale: A scale consisting of unordered categories.

Ordinal scale: An ordered scale without necessarily equal intervals.

Interval scale: A scale with equal intervals but no meaningful absolute zero.

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

Sampling bias: Systematic distortion caused by how observations are selected or measured.

Random variable: A numerical outcome of a random process.

Standard deviation: A measure of typical distance from the mean.

Standard error: The estimated standard deviation of a sampling statistic.

Confidence interval: An interval produced by a method designed to capture a population parameter at a stated long-run rate.

Correlation: A measure of linear association between two numerical variables.

Null hypothesis: A baseline claim evaluated by a statistical test.

P-value: The probability, under the null model, of a result at least as extreme as that observed.

Type I error: Rejecting a true null hypothesis.

Confounding: A situation in which another variable contributes to an observed relationship.

Review Questions

5 Short Answer Questions

1. What is the difference between a parameter and a statistic?
2. Why is a customer satisfaction rating usually treated as ordinal?
3. State one difference between sampling error and sampling bias.
4. What does a standard deviation measure?
5. Why does correlation not establish causation?

5 Long Answer Questions

1. Explain the four measurement scales and give one business example of each.
2. Compare simple random, stratified, cluster and convenience sampling.
3. Explain the meaning and assumptions of a confidence interval for a population mean.
4. Describe the role of the null hypothesis, p-value and significance level in hypothesis testing.
5. Discuss how outliers and skewness affect the choice between mean, median and standard deviation.

3 Analytical / Application Questions

1. A company surveys only customers who opened its latest email. Identify two possible sources of bias and propose a better sampling approach.
2. A manager observes that stores with larger advertising budgets have higher sales. Explain at least two possible interpretations and identify evidence needed before claiming causation.
3. A sample of delivery times has a mean of 31.2 minutes, standard deviation of 6 minutes and sample size of 100. Using 1.96 as the critical value, construct a 95% large-sample confidence interval and explain its business interpretation.

Multiple Choice Questions with Answers

Question 1

Which measurement scale is most appropriate for product category?

A. Nominal

B. Ordinal

C. Interval

D. Ratio

Correct Answer: A. Nominal

Question 2

Which sampling method divides a population into groups and samples within each group?

A. Cluster sampling

B. Stratified sampling

C. Convenience sampling

D. Voluntary-response sampling

Correct Answer: B. Stratified sampling

Question 3

If P(A) = 0.70, what is P(Aᶜ)?

A. 0.03

B. 0.30

C. 0.70

D. 1.70

Correct Answer: B. 0.30

Question 4

For the values 4, 6 and 8, what is the arithmetic mean?

A. 5

B. 6

C. 7

D. 18

Correct Answer: B. 6

Question 5

Which measure is generally most resistant to an extreme outlier?

A. Mean

B. Range

C. Median

D. Variance

Correct Answer: C. Median

Question 6

In the formula z = (x − μ) ÷ σ, what does z represent?

A. The sample size

B. The number of categories

C. The distance from the mean measured in standard deviations

D. The population variance

Correct Answer: C. The distance from the mean measured in standard deviations

Question 7

A 95% confidence interval procedure means that:

A. 95% of the observations lie inside the calculated interval

B. The parameter has a 95% probability of changing

C. About 95% of intervals from repeated samples would capture the parameter under the model

D. The sample mean is correct with probability 0.95

Correct Answer: C. About 95% of intervals from repeated samples would capture the parameter under the model

Question 8

A correlation close to −1 indicates:

A. A strong negative linear association

B. No possible relationship

C. A strong positive causal effect

D. Equal means

Correct Answer: A. A strong negative linear association

Question 9

If a hypothesis test produces a p-value of 0.02 and α = 0.05, the usual decision is to:

A. Accept that the null hypothesis is certainly true

B. Reject the null hypothesis

C. Increase the p-value to 0.05

D. Conclude that the effect is necessarily important in practice

Correct Answer: B. Reject the null hypothesis

Question 10

For 100 observations with sample standard deviation 6, the standard error of the mean is:

A. 0.06

B. 0.6

C. 6

D. 60

Correct Answer: B. 0.6