Descriptive Analytics and Statistical Reasoning

Introduction

Supply-chain professionals work with observations: order quantities, supplier lead times, inventory balances, transport costs, forecast errors, fill rates, and customer demand. Descriptive analytics converts these observations into an interpretable account of what has happened. Statistical reasoning goes further by asking how reliable that account is, what variation means, whether apparent relationships are credible, and how a sample can support cautious conclusions about a wider population.

This distinction is essential. A spreadsheet may report that average supplier lead time is 8.4 days, but managers also need to know whether the average conceals extreme delays, whether a few unusually large orders distort it, and whether the data represent all shipments or only one route. Descriptive statistics provide structured evidence; they do not automatically establish causes or predict the future.

Learning objectives

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

1. Select and calculate appropriate measures of center, position, and variability.
2. Interpret frequency distributions, skewness, and empirical percentiles in supply-chain data.
3. Distinguish association from causation and use correlation appropriately.
4. Detect, investigate, and report outliers without deleting valid observations automatically.
5. Explain population, sample, sampling error, and sampling bias.
6. Conduct a disciplined exploratory analysis in a spreadsheet.
7. Communicate statistical findings with appropriate uncertainty, limitations, and ethical care.

1. Data, variables, and analytical purpose

A unit of analysis is the entity represented by one record: a shipment, purchase order, warehouse-day, supplier, or customer. A variable is a measured characteristic, such as transit time or quantity ordered. Variables may be:

• Categorical: supplier region or transport mode.
• Ordinal: service rating from poor to excellent, where order matters but equal spacing is not guaranteed.
• Discrete quantitative: number of stockouts.
• Continuous quantitative: weight, cost, or time, measured on a scale.

The measurement scale constrains valid analysis. Means are meaningful for monetary amounts and physical quantities, but generally not for supplier codes. Counts and rates require clearly defined denominators. A “97% on-time rate” is incomplete unless the reader knows whether it means 97 of 100 shipments, 9,700 of 10,000 shipment lines, or a weighted customer-service measure.

Before calculating statistics, establish the question, observation unit, time period, inclusion rules, and data-quality controls. A descriptive result is only as defensible as the data-generation process behind it.

2. Measures of center and position

Mean, median, and mode

For observations x₁, x₂, …, xₙ, the arithmetic mean is:

Here xᵢ is observation i, n is the number of observations, and x̄ is the sample mean. All observations must be measured in comparable units. The mean is sensitive to extreme values and is especially informative when a distribution is reasonably symmetric.

The median is the middle ordered observation. If n is odd, it is the value at position (n + 1) ÷ 2; if n is even, it is the average of the two central observations. The median is resistant to extreme values and is often preferable for skewed lead-time or cost data.

The mode is the most frequently occurring value or category. It is useful for identifying the most common order size, transport mode, or service classification, but a dataset may have more than one mode or none.

Worked example: delivery time

Consider delivery times in days: 2, 3, 3, 4, 5, 5, 6, 20.

The mean is (2 + 3 + 3 + 4 + 5 + 5 + 6 + 20) ÷ 8 = 48 ÷ 8 = 6 days. The median is (4 + 5) ÷ 2 = 4.5 days, and the modes are 3 and 5. The mean is pulled upward by the 20-day delay; the median better describes a typical shipment, while the extreme value remains operationally important.

Weighted mean

When observations have different importance or volume, a weighted mean may be appropriate:

where wᵢ is a non-negative weight. Weights must have a defensible meaning. For example, average unit cost across products can be weighted by units purchased. However, weighting can obscure performance for smaller customers, so both weighted and unweighted results may be useful.

Percentiles and quartiles

The pth percentile is a value below which approximately p percent of observations fall. The 50th percentile is the median. Quartiles divide ordered data into four sections: Q₁ is the 25th percentile and Q₃ is the 75th percentile. Spreadsheet software may use different percentile interpolation conventions; analysts should document the function and method used.

Percentiles are valuable for service design. If the 90th percentile of supplier lead time is 12 days, then approximately 90% of observed lead times are at or below 12 days under the chosen dataset and definition. This does not mean that the next shipment has a 90% probability of arriving within 12 days unless the sampling and process assumptions justify that inference.

3. Variability and distributional shape

Averages without variability can be misleading. Two suppliers may both average 8 days, while one consistently delivers in 7–9 days and the other alternates between 2 and 14 days.

Range and interquartile range

The range is maximum minus minimum. It is simple but highly sensitive to two extreme observations. The interquartile range (IQR) is:

It describes the spread of the middle 50% and is resistant to extremes.

Variance and standard deviation

For a sample, the sample variance is:

and the sample standard deviation is:

The denominator n − 1, known as Bessel’s correction, is commonly used when estimating population variability from a sample. If the data constitute the entire defined population of interest, a population variance using denominator n may be appropriate. Variance is expressed in squared units, whereas standard deviation has the original units, making it easier to communicate. Both measures are most interpretable when the distribution is not extremely skewed and observations are comparable.

The coefficient of variation compares relative variability:

Often reported as CV × 100%, it is meaningful only when the mean is positive and the zero point is substantive. It should not be used casually when means are near zero or when variables have incompatible measurement structures.

Worked variability calculation

For lead times 2, 3, 3, 4, 5, 5, 6, 20, x̄ = 6. Deviations are −4, −3, −3, −2, −1, −1, 0, and 14. Squared deviations sum to 16 + 9 + 9 + 4 + 1 + 1 + 0 + 196 = 236. Thus s² = 236 ÷ 7 ≈ 33.71 days² and s ≈ 5.81 days. The large standard deviation confirms that the mean alone does not represent stable performance.

Distribution shape

A distribution may be approximately symmetric, right-skewed, left-skewed, heavy-tailed, or multimodal. In a right-skewed distribution, a small number of high observations commonly make the mean exceed the median. Freight costs, delay durations, and emergency-order quantities often show this pattern. A histogram, dot plot, box plot, or empirical cumulative distribution can reveal shape more effectively than a single statistic.

A histogram depends on bin width, so analysts should test reasonable alternatives rather than treating one visual appearance as definitive. Time series should also be plotted in chronological order; combining months can conceal trend, seasonality, or process changes.

4. Correlation and statistical reasoning

The Pearson sample correlation between variables X and Y is:

where x̄ and ȳ are sample means. The value lies between −1 and +1 when both variables have nonzero variation. Positive values indicate that high values tend to occur with high values; negative values indicate an inverse linear association. A value near zero indicates little linear association, not necessarily no relationship.

Correlation is unitless and measures linear association, not causation. It can be distorted by outliers, restricted ranges, nonlinear patterns, time trends, and confounding variables. For example, promotional activity may correlate with warehouse overtime because both increase during peak season; neither correlation nor temporal coincidence alone proves that overtime caused promotions or that promotions caused inefficiency.

A scatterplot should precede interpretation. Examine whether observations are independent, whether a curved pattern exists, whether groups differ, and whether one point drives the result. Spearman’s rank correlation may be useful for monotonic relationships or ordinal data, but it too does not establish causation.

A useful reasoning sequence is: describe the pattern; identify plausible alternative explanations; inspect design and data provenance; consider whether variables were measured before or after one another; and state what the analysis cannot establish. Regression and causal analysis may be appropriate in later work, but descriptive correlation should not be presented as a causal estimate.

5. Outliers: detection, investigation, and judgment

An outlier is an observation unusually distant from the rest of a dataset. It may represent a data-entry error, a different process, a rare but valid event, or a measurement problem. Outlier detection is therefore an investigative aid, not an automatic deletion rule.

A common rule uses the IQR:

Observations outside these fences are flagged as potential outliers. The rule is a screening convention, not a universal law, and its performance depends on distribution shape and sample size. A z-score can also be calculated as:

This standardizes an observation relative to the sample mean and standard deviation. Thresholds such as |z| > 3 are heuristics and are unreliable when the data are skewed or when the outlier inflates s.

A defensible workflow is to retain the raw value, verify source records, check units and timestamps, determine whether the event belongs to the target population, document the reason for any correction, and report sensitivity analyses with and without the observation when appropriate. For valid disruption events, removing the observation can understate operational risk. Robust summaries such as the median and IQR may be preferable.

6. Sampling and uncertainty

A population is the complete set about which a question is asked; a sample is the observed subset. Sampling is useful when a census is costly or impossible, but conclusions depend on how observations were selected.

• Simple random sampling: every eligible unit has a known equal selection probability.
• Systematic sampling: units are selected at a fixed interval after a suitable random start; periodic patterns can create bias.
• Stratified sampling: the population is divided into relevant groups and sampled within each; it can improve representation of regions or product classes.
• Cluster sampling: whole groups, such as facilities or routes, are selected; observations within clusters may be correlated.
• Convenience sampling: readily available units are selected and often do not represent the target population.

Sampling error is random difference between a sample estimate and the population value. Sampling bias is systematic distortion caused by the selection process, nonresponse, coverage gaps, or changing definitions. Increasing sample size reduces random error under suitable conditions but does not repair systematic bias.

For a simple random sample, the standard error of the sample mean is commonly estimated as:

where s is sample standard deviation and n is sample size. This expression assumes an appropriate random sample and independence, or an adjustment for dependence. A 95% confidence interval for a mean is commonly written:

where t* is the relevant critical value from the t distribution with n − 1 degrees of freedom. The interval procedure has long-run coverage near 95% under its assumptions; it is not correct to say that there is a 95% probability that a fixed population mean lies inside the particular interval after it has been calculated.

In supply chains, clustered observations are common: shipments from one supplier may share policies, weather, and route conditions. Treating them as independent can make uncertainty appear smaller than it is. Analysts should consider aggregation, cluster-robust methods, or hierarchical models when the decision requires inferential precision.

7. Spreadsheet-based exploratory analysis

A disciplined exploratory analysis (EDA) is iterative rather than a sequence of arbitrary clicks.

1. Define scope: state the decision, population, unit, period, and metric definitions.
2. Inspect structure: check row counts, data types, missingness, duplicates, impossible values, and units.
3. Create a data dictionary: record field meaning, source, owner, allowed values, and transformations.
4. Summarize: calculate count, mean, median, quartiles, minimum, maximum, and standard deviation by meaningful groups.
5. Visualize: use histograms, box plots, scatterplots, control charts, and time plots as appropriate.
6. Segment: compare supplier, facility, route, product, and period only when group definitions are stable.
7. Investigate: trace outliers, breaks, and missing values to source systems.
8. Stress-test: compare alternative definitions, weighting choices, and inclusion rules.
9. Communicate: state findings, uncertainty, limitations, and recommended action.

Spreadsheet controls include protected formula cells, input validation, named ranges, explicit units, version history, reconciliation checks, and separate raw-data, calculation, and reporting areas. Functions such as MEDIAN, QUARTILE, STDEV.S, CORREL, COUNTIF, and SUMIFS can support analysis, but function names and percentile methods vary by software. A result should be reproducible by another analyst, not dependent on an undocumented manual sort.

8. Illustrative case: supplier lead-time reliability

Background

A distribution company sources a critical component from Supplier A. The procurement team reports an average lead time of 8 days and proposes reducing safety stock.

Problem

An analyst receives 20 shipment records. The records include two emergency shipments and one shipment delayed by a port closure. Management wants to know whether the average adequately represents replenishment risk.

Analysis

The analyst first confirms that each row represents one completed shipment, that lead time is measured from confirmed dispatch to receipt, and that all 20 shipments belong to the same route and product family. A summary shows a mean of 8.0 days, a median of 6.5 days, a standard deviation of 6.2 days, and a 90th percentile of 16 days. A histogram is right-skewed, and the time plot shows that the longest delay occurred during the port closure.

A box-plot rule flags the port-closure shipment as a potential outlier. Source documents confirm it is valid. The analyst therefore does not delete it. Instead, the report presents two operational views: routine performance excluding externally documented disruption, and all-event risk including disruption. The first supports process benchmarking; the second is relevant to resilience and safety-stock decisions.

Approach and solution

Rather than basing policy on the mean alone, the team reviews the median, 90th percentile, disruption frequency, and the cost of stockout versus holding inventory. It also checks whether emergency shipments were caused by prior supplier delays, because excluding them might remove consequences of poor performance. The recommended decision is a conditional safety-stock policy: use routine lead time for normal replenishment, but maintain a disruption buffer or alternate-source trigger for events that cannot be rapidly mitigated.

Outcome and learning points

The company postpones the proposed stock reduction, establishes a monthly percentile dashboard, and records disruption causes using controlled categories. The case demonstrates that a valid extreme observation may be more important for risk management than for describing routine central tendency. It also shows why a statistical summary must be connected to a decision, a population definition, and a documented data-generating process.

9. Limitations, ethics, and contemporary practice

Descriptive analytics can create false confidence through precise-looking dashboards, inconsistent definitions, selective reporting, or unexamined missing data. A KPI may improve because difficult cases were excluded. Group comparisons may disadvantage a supplier serving remote locations if geography is ignored. Analysts should avoid presenting a low average as proof of good performance when variability or tail risk is material.

Governance requires access controls, audit trails, privacy protection, retention rules, and clear ownership of metric definitions. Shipment and customer data may contain commercially sensitive or personally identifiable information. Aggregation and role-based access should be used where detailed records are unnecessary.

AI-assisted spreadsheet tools can suggest formulas, classifications, or anomalies, but generated outputs require human verification. Models may reproduce historical bias, silently change definitions, or provide plausible explanations unsupported by evidence. Versioned data, reproducible calculations, independent checks, and disclosure of automated assistance remain essential.

Best practice is to report the statistic, denominator, period, unit, method, sample frame, notable exclusions, and uncertainty together. Statistical literacy is not merely the ability to calculate; it is the ability to match a method to a question and communicate what the evidence does and does not support.

Chapter Summary

Descriptive analytics organizes supply-chain observations through measures of center, position, variability, and distributional shape. The mean is sensitive to extremes, whereas the median and IQR are more robust for skewed data. Standard deviation measures dispersion in original units, while the coefficient of variation expresses relative variability under suitable conditions. Correlation describes linear association but cannot establish causation. Outliers must be investigated, not automatically removed. Sampling design determines whether results can reasonably represent a wider population; larger samples reduce random error but do not eliminate bias. Effective spreadsheet EDA combines validation, summaries, visualizations, segmentation, sensitivity analysis, documentation, and ethical governance. Sound statistical reasoning connects each result to its measurement process, assumptions, uncertainty, and managerial decision.

Key Terms

Descriptive analytics: Analysis that summarizes observed data without necessarily making causal or predictive claims.

Unit of analysis: Entity represented by one observation or row.

Mean: Arithmetic average of quantitative observations.

Median: Middle value after observations are ordered.

Mode: Most frequently occurring value or category.

Percentile: Value at or below which a specified percentage of observations fall.

Quartile: One of three values dividing ordered data into four sections.

Interquartile range: Q₃ − Q₁, measuring the spread of the middle 50%.

Variance: Average squared deviation, with the sample version using n − 1.

Standard deviation: Square root of variance, expressed in original units.

Coefficient of variation: Standard deviation divided by a positive mean.

Skewness: Asymmetry in a distribution.

Outlier: Observation unusually distant from the remaining data.

Correlation: Standardized measure of linear association between two variables.

Population: Complete set about which an analysis seeks information.

Sample: Observed subset of a population.

Sampling error: Random difference between a sample estimate and population value.

Sampling bias: Systematic distortion caused by the selection or measurement process.

Exploratory data analysis: Iterative examination using summaries, visualizations, and investigation.

Confidence interval: Interval produced by a procedure intended to achieve stated long-run coverage.

Review Questions

Short Answer Questions

1. Why might the median be preferable to the mean for supplier lead times?
2. Define the interquartile range and explain what it measures.
3. What does a correlation of zero fail to prove?
4. Distinguish sampling error from sampling bias.
5. Why should a confirmed disruption outlier not automatically be deleted?

Long Answer Questions

1. Explain how mean, median, standard deviation, and percentile can jointly describe delivery performance.
2. Discuss the assumptions and limitations of Pearson correlation in supply-chain analysis.
3. Compare simple random, stratified, cluster, and convenience sampling.
4. Develop a spreadsheet workflow for analyzing warehouse order-cycle time.
5. Explain how ethical governance affects the design and reporting of operational KPIs.

Analytical / Application Questions

1. A supplier has a low average lead time but a high 95th percentile. What additional analyses should precede a safety-stock decision?
2. A scatterplot shows a strong positive correlation between demand and overtime. Propose at least three alternative explanations and a design for investigating them.
3. An analyst removes all shipments above the upper IQR fence before reporting supplier performance. Critique this approach and propose a more defensible sensitivity analysis.

Multiple Choice Questions with Answers

Question 1

Which measure is generally most resistant to a single extremely long delivery time?

A. Mean

B. Median

C. Variance

D. Range

Correct Answer: B. Median

Question 2

For the observations 2, 3, 3, 4, 5, 5, 6, and 20, what is the arithmetic mean?

A. 4.5 days

B. 5 days

C. 6 days

D. 8 days

Correct Answer: C. 6 days

Question 3

What does the interquartile range measure?

A. The total distance between the minimum and maximum

B. The spread of the middle 50% of observations

C. The average squared deviation from the mean

D. The percentage of observations above the mean

Correct Answer: B. The spread of the middle 50% of observations

Question 4

Which statement about correlation is correct?

A. A strong correlation proves that one variable causes the other

B. Correlation can only be positive

C. Correlation measures linear association and does not by itself establish causation

D. A correlation near zero proves that no relationship exists

Correct Answer: C. Correlation measures linear association and does not by itself establish causation

Question 5

An observation outside the upper IQR fence should initially be treated as:

A. A confirmed data-entry error

B. A value that must be deleted

C. A potential outlier requiring investigation

D. Evidence that the median is incorrect

Correct Answer: C. A potential outlier requiring investigation

Question 6

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

A. Cluster sampling

B. Convenience sampling

C. Stratified sampling

D. Systematic sampling

Correct Answer: C. Stratified sampling

Question 7

For a sample with standard deviation s and size n, the commonly estimated standard error of the mean is:

A. s × n

B. s ÷ √n

C. s ÷ n²

D. √s ÷ n

Correct Answer: B. s ÷ √n

Question 8

Why can increasing sample size fail to correct sampling bias?

A. Larger samples cannot contain quantitative variables

B. Bias is systematic distortion from the selection or measurement process

C. Larger samples always increase standard error

D. Bias occurs only when the mean is calculated

Correct Answer: B. Bias is systematic distortion from the selection or measurement process

Question 9

A right-skewed lead-time distribution commonly has:

A. A mean greater than the median

B. A mean equal to zero

C. A median greater than every observation

D. No variability

Correct Answer: A. A mean greater than the median

Question 10

What is the most defensible response when a flagged outlier is confirmed to be a valid port-closure delay?

A. Delete it without documentation

B. Replace it with the mean

C. Retain it and consider separate routine-performance and all-event analyses

D. Ignore all other observations

Correct Answer: C. Retain it and consider separate routine-performance and all-event analyses