Foundations of Populations, Samples and Probability Models

Introduction

Sampling distributions provide the conceptual bridge between the data observed in a quality-management study and the conclusions made about a larger process or population. A quality engineer may inspect a sample of manufactured units, record their dimensions, and use the results to judge process capability, estimate defect rates, or decide whether corrective action is necessary. Such decisions are justified only when the population, sampling process, variables, and probability model are clearly defined.

This chapter establishes the vocabulary and mathematical foundations required for later study of sampling distributions, confidence intervals, hypothesis tests, statistical process control, and designed experiments. Particular attention is given to the distinction between a population and a sample, a parameter and a statistic, and an observation and a random variable. The chapter also reviews expectation, variance, covariance, independence, and probability models commonly used in quality management.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish target populations, accessible populations, samples, parameters, statistics, observations, and random variables.
2. Explain how probability models represent uncertainty in quality data.
3. Calculate and interpret expectation, variance, covariance, and correlation.
4. Distinguish independence from lack of correlation and identify when independence assumptions are reasonable.
5. Select appropriate discrete and continuous probability models for quality-management applications.
6. Recognize the assumptions and limitations underlying common models.
7. Explain how sampling and probability models support responsible quality decisions.

1. Populations, Samples and the Logic of Statistical Inference

1.1 The population as a process or set of units

A population is the complete collection of units, events, or measurements about which a study seeks information. The units may be physical products, service transactions, production cycles, suppliers, or future observations from a process.

In quality management, it is useful to distinguish two populations:

• Finite population: a clearly defined set, such as all 8,000 units produced during a particular shift.
• Conceptual or superpopulation: a wider process-generating population, such as all units that could be produced under specified operating conditions in the future.

The population of interest must be defined by content, location, time, and conditions. “All products” is usually inadequate. “All 500-millilitre bottles produced on Line 3 during the night shift in March under the current filling settings” is more precise.

The target population is the population to which the conclusions are intended to apply. The accessible population is the part that can actually be sampled. Differences between them create coverage error. For example, a supplier audit based only on weekday shipments may not represent weekend production.

1.2 Samples and observations

A sample is a subset of units selected from a population. If the sample contains n units, its measured values may be written as x₁, x₂, ..., x_n. Each x_i is an observation, or realized measurement.

A sample is not merely a smaller population. Its usefulness depends on the selection mechanism. A probability sample gives each relevant unit a known, non-zero selection probability. A simple random sample of size n from a finite population of size N gives every subset of size n the same chance of selection. Other designs include stratified, systematic, cluster, and multistage sampling.

Random sampling reduces selection bias but does not automatically eliminate measurement error, nonresponse, process changes, or poor operational definitions. A large biased sample can be less informative than a smaller, well-designed sample.

1.3 Parameters and statistics

A parameter is a numerical characteristic of a population. A statistic is a numerical function of sample observations. Parameters are commonly denoted by Greek letters, while statistics are often denoted by Roman letters.

The sample mean is

where n is the sample size and x_i is observation i. The sample variance usually uses n-1 rather than n:

This denominator produces an unbiased estimator of the population variance when the observations are independent and identically distributed with finite variance. The formula is not automatically appropriate for clustered, weighted, or strongly dependent samples.

A statistic is itself a random variable before the sample is observed because different random samples produce different values. After observation, it has a numerical realization. This distinction is central to sampling distributions.

2. Random Variables and Probability Distributions

2.1 Random variables

A random variable assigns a numerical value to each outcome of a random experiment. A discrete random variable has a finite or countably infinite set of possible values. A continuous random variable can take values over intervals of the real line.

Examples include:

• X = number of defective units in a sample;
• Y = tensile strength of a randomly selected component;
• T = time until equipment failure;
• X̄ = mean diameter of a randomly selected sample.

The capital letter denotes the random variable; lowercase letters commonly denote a possible realized value.

2.2 Probability mass, density and distribution functions

For a discrete random variable, the probability mass function is

and must satisfy p_X(x) ≥ 0 and ∑_x p_X(x) = 1.

For a continuous random variable, probabilities are represented by a probability density function f_X(x). It must satisfy f_X(x) ≥ 0 and

For a continuous variable, P(X=x)=0 at any single point; probabilities are areas over intervals:

The cumulative distribution function applies to both types:

It is nondecreasing, lies between 0 and 1, and approaches 0 and 1 at the lower and upper extremes, respectively.

A probability model is an idealized representation of a data-generating mechanism. It is not a claim that nature follows a formula exactly. Model adequacy must be evaluated using process knowledge, plots, diagnostics, and subject-matter judgment.

3. Expectation, Variance and Standard Deviation

3.1 Expected value

The expected value is the long-run probability-weighted average of a random variable. For a discrete variable,

For a continuous variable,

The expectation need not be an attainable value. If X is the number of defects, its expectation may be 2.4 even though no individual sample contains 2.4 defects.

Expectation is linear. For constants a and b, provided the expectations exist,

More generally, E(∑a_iX_i)=∑a_iE(X_i). Independence is not required for this property.

3.2 Variance

Variance measures the expected squared deviation from the mean:

where μ=E(X). Its units are squared units of X. The standard deviation is

and has the same units as the original measurement, making it easier to interpret.

For constants a and b,

Thus, adding a fixed calibration offset changes the mean but not the variance; multiplying measurements by a conversion factor changes the standard deviation by the absolute value of that factor.

If X₁, ..., X_n are independent and have finite variances,

If they are identically distributed with variance σ², then

and the standard error of the sample mean is σ/√n. This result requires independence, or a justified alternative accounting for dependence.

Worked Example 1: Expected defects and variation

Suppose the number X of surface defects on a randomly selected panel has the following distribution:

The expected number of defects is

E(X)=0(0.50)+1(0.30)+2(0.15)+3(0.05)=0.75.

Also,

E(X²)=0²(0.50)+1²(0.30)+2²(0.15)+3²(0.05)=1.20.

Therefore,

Var(X)=1.20−0.75²=0.6375,

and σ=√0.6375≈0.798 defects. The mean of 0.75 is a long-run average, not a requirement that panels possess fractional defects.

4. Covariance, Correlation and Independence

4.1 Covariance

For random variables X and Y, covariance is

where μ_X and μ_Y are the respective means. Positive covariance indicates that high values tend to occur together; negative covariance indicates an opposing tendency. Covariance depends on measurement units and is therefore difficult to compare across studies.

The variance of a sum is

For a difference, Var(X−Y)=Var(X)+Var(Y)−2Cov(X,Y). These formulas are important when measurements from the same product, machine, or batch are related.

4.2 Correlation

The Pearson correlation coefficient is

provided both standard deviations are positive and finite. It is dimensionless and lies between −1 and 1. Correlation measures linear association, not general dependence. A correlation near zero does not prove independence.

4.3 Independence

X and Y are independent if, for all relevant sets A and B,

For discrete variables this is equivalent to P(X=x,Y=y)=P(X=x)P(Y=y). Independence implies zero covariance when variances exist, but zero covariance does not generally imply independence. The exception includes important special cases such as jointly normal variables.

In production, independence may be questionable when observations are consecutive readings from the same machine, repeated measurements on one unit, or units from the same batch. Autocorrelation, common-cause changes, and operator effects can make the effective sample size smaller than the nominal sample size.

5. Probability Models Used in Quality Management

5.1 Bernoulli model

A Bernoulli random variable represents one trial with two outcomes, often coded as success and failure or conforming and nonconforming. If the probability of a nonconforming outcome is p, then

with E(X)=p and Var(X)=p(1−p). The model assumes a clearly defined binary outcome and a stable probability for the trial under consideration.

5.2 Binomial model

If n independent Bernoulli trials share the same nonconformance probability p, the number X of nonconforming units follows a binomial distribution:

where C(n,x)=n!/[x!(n−x)!]. Its mean and variance are np and np(1−p).

The binomial model is appropriate for sampling with replacement or for sampling without replacement from a very large population relative to the sample. For a small finite population sampled without replacement, the hypergeometric model is more appropriate.

5.3 Hypergeometric model

Suppose a finite lot contains N units, of which K are nonconforming, and n units are selected without replacement. The number X of nonconforming units has probability

for feasible values of x. Unlike the binomial model, successive selections are dependent because the lot composition changes after each draw. The hypergeometric model underlies many acceptance-sampling calculations.

5.4 Poisson model

The Poisson model describes counts of events in a fixed interval of time, area, length, or volume when events occur independently at a constant average rate λ:

Both the mean and variance equal λ. Applications include flaws per square metre, customer complaints per day, or machine stoppages per week. The assumptions of a constant rate and no interaction between events should be examined. Overdispersion—variance greater than the mean—may indicate heterogeneity, clustering, or an unsuitable Poisson model.

5.5 Uniform model

A continuous variable uniformly distributed on [a,b] has constant density

Its mean is (a+b)/2 and its variance is (b−a)²/12. The model may represent a measurement with no reason to favour one value within a bounded interval, but it should not be selected merely because the data have minimum and maximum values.

5.6 Normal model

A normal random variable with mean μ and standard deviation σ>0 has density

The notation is X∼N(μ,σ²). The normal model is continuous, symmetric, and bell-shaped. It is often useful for dimensional measurements arising from many small additive effects, and it is central to statistical inference through the central limit theorem.

A standardized normal variable is

which has mean 0 and variance 1 when X is normal. Normality should not be assumed solely because a variable is measured on a continuous scale. Histograms, probability plots, process knowledge, and robustness of the intended procedure should guide the decision.

5.7 Exponential model

The exponential model describes a nonnegative waiting time with constant hazard rate λ>0:

Its mean is 1/λ and variance is 1/λ². The survival probability is P(T>t)=e^(−λt). Its memoryless property means that the future waiting time does not depend on elapsed waiting time. This may be unrealistic for equipment subject to wear-out, preventive maintenance, or infant mortality.

5.8 Weibull model

The Weibull model is widely used for reliability and time-to-failure analysis. With shape parameter β>0 and scale parameter η>0, its density is

The hazard function increases when β>1, is constant when β=1, and decreases when β<1. Thus, the shape parameter can represent wear-out, random failure, or early-life failure patterns. Parameter estimates and reliability conclusions depend strongly on censoring assumptions and the quality of failure-time records.

6. Sampling Design and Sampling Distributions

A sampling distribution is the probability distribution of a statistic over repeated samples drawn under the same design and population conditions. For example, X̄ has a sampling distribution, even though one observed sample produces only one value of x̄.

If X₁,...,X_n are independent with common mean μ and variance σ², then

If the parent population is normal, X̄ is exactly normal for every sample size. If the parent population is not normal, the central limit theorem states, under suitable regularity conditions, that the standardized sample mean becomes approximately normal as n increases:

The approximation is affected by skewness, heavy tails, dependence, and outliers; there is no universal sample size at which it becomes exact.

For a sample proportion based on independent Bernoulli observations,

where p̂=X/n. A normal approximation is often reasonable when both np and n(1−p) are sufficiently large, although exact or improved interval methods may be preferable for extreme proportions or small samples.

Worked Example 2: A sample proportion

A process has a true nonconformance probability of p=0.04. In a random sample of n=100 independent units, the expected number of nonconforming units is np=4, and

SD(X)=√[np(1−p)]=√(100×0.04×0.96)≈1.96.

For the sample proportion, E(p̂)=0.04 and

SD(p̂)=√[0.04×0.96/100]≈0.0196.

The standard deviation describes sample-to-sample variation in estimated proportions; it is not the process defect probability itself.

7. Worked Quality-Management Application

Illustrative case: Incoming inspection of electronic connectors

Background. An electronics manufacturer receives connectors from a supplier. Historical records suggest that the supplier’s nonconformance rate may be approximately 2%, but recent customer complaints have increased. The quality team must decide how to model a sample of incoming units and how much confidence to place in the observed count.

Problem. From a shipment of N=2,000 connectors, the team selects n=80 units without replacement and finds x=3 nonconforming units. Should the count be modelled as binomial or hypergeometric, and what is the immediate estimated nonconformance proportion?

Analysis. Because sampling is without replacement from a finite lot, the exact design is hypergeometric if the lot’s true number of nonconforming units is treated as fixed. However, because n/N=80/2000=0.04, the sampling fraction is small. A binomial approximation may be operationally adequate if the lot is regarded as representative of a stable process and the probability is approximately unchanged across draws.

The point estimate is

p̂=x/n=3/80=0.0375,

or 3.75%. The estimate is higher than the historical 2%, but a point estimate alone does not establish a process shift. Sampling variation, lot-to-lot variation, inspection error, and the possibility of clustered defects must be considered.

Approach. The team should verify that the sample was randomly selected across the shipment, confirm the inspection definition, record the locations and defect types, and compare the result with an appropriate confidence procedure for a finite population or binomial process. If defects are concentrated in one production batch, independence is doubtful and stratified analysis may be necessary.

Outcome. The immediate management action is not to label the supplier permanently defective based on three observations. Instead, the result should trigger verification, temporary containment if risk is high, and a follow-up sampling plan. Supplier performance should be evaluated using repeated lots and a pre-specified decision rule.

Learning points. The case demonstrates that population definition, sampling design, and model assumptions are inseparable. A numerical method cannot repair an unrepresentative sample or an ambiguous defect definition.

8. Limitations, Governance and Professional Practice

Model risk and data quality

A probability model may be mathematically correct but operationally inappropriate. Common warning signs include autocorrelation, changing process conditions, mixture populations, measurement-system instability, censoring, rounding, and excessive zero counts. Analysts should examine the data-generating process before fitting a distribution.

Measurement systems

Observed variation combines process variation and measurement variation. If the gauge is biased or imprecise, estimates of μ and σ² describe the measurement system as well as the product. Gauge repeatability and reproducibility studies, calibration, traceability, and clear operational definitions are therefore prerequisites for credible sampling analysis.

Ethical and governance considerations

Quality sampling affects release decisions, worker safety, customer protection, and supplier treatment. Sampling plans should not be manipulated to obtain a preferred result. Exclusion rules, stopping rules, missing observations, and model choices should be documented before results are used for high-consequence decisions. Personal or supplier-identifiable data should be protected, and automated tools should preserve auditability and human accountability.

Technology and contemporary practice

Sensors, laboratory information systems, and industrial IoT platforms can produce high-frequency data, but more data do not necessarily mean more independent information. Streaming observations may be serially correlated and may reflect only times when equipment was operating. Statistical software can calculate distributions accurately while still applying an inappropriate model. Modern practice therefore combines computation with process knowledge, graphical diagnostics, version-controlled analysis, and transparent reporting of assumptions.

Best-Practice Checklist

1. Define the target population and the quality characteristic precisely.
2. Identify the sampling frame and assess coverage gaps.
3. Select a probability-based design where inference is required.
4. Stabilize and validate the measurement system.
5. State whether observations are independent, and investigate dependence.
6. Choose a model based on mechanism and evidence, not convenience alone.
7. Report parameters, statistics, units, uncertainty, and assumptions distinctly.
8. Use sensitivity analysis when several plausible models exist.
9. Separate exploratory findings from pre-specified decisions.
10. Preserve an auditable record of data, exclusions, calculations, and decisions.

Chapter Summary

A population is the complete set of units or process outcomes of interest, while a sample is the subset actually observed. A parameter describes a population; a statistic describes a sample and is a random variable before the sample is realized. Probability distributions formalize uncertainty through mass functions, density functions, and cumulative distribution functions.

Expectation describes a probability-weighted centre, variance describes squared dispersion, and covariance describes joint variation. Correlation standardizes covariance but measures only linear association. Independence is a stronger condition than zero correlation and is often threatened by batch effects, time dependence, repeated measurements, and process changes.

Bernoulli and binomial models describe binary outcomes and counts in independent trials; the hypergeometric model applies to finite-lot sampling without replacement; the Poisson model describes event counts; and uniform, normal, exponential, and Weibull models support measurement, waiting-time, and reliability applications. Each model has assumptions that must be checked.

Sampling distributions explain how statistics vary from sample to sample. Under suitable conditions, the sample mean is unbiased for the population mean and its variance decreases as sample size increases. Sound quality decisions require more than formulas: they require a defined population, defensible sampling, valid measurement, appropriate modelling, and ethical reporting.

Key Terms

Population: The complete set of units or outcomes about which conclusions are intended.

Target population: The population to which a study intends to generalize.

Accessible population: The portion of the target population that can actually be sampled.

Sample: A selected subset of population units.

Observation: A realized measured value from a sampled unit.

Parameter: A numerical characteristic of a population.

Statistic: A numerical function of sample observations.

Random variable: A numerical function whose value is determined by a random outcome.

Probability mass function: A function giving probabilities for discrete values.

Probability density function: A function whose interval areas give probabilities for a continuous variable.

Cumulative distribution function: F_X(x)=P(X≤x) for a random variable X.

Expectation: The probability-weighted average or theoretical mean of a random variable.

Variance: The expected squared deviation from the mean.

Covariance: A measure of joint directional variation between two variables.

Independence: A condition in which the joint probability factors into marginal probabilities.

Binomial distribution: The distribution of successes in independent, identically distributed Bernoulli trials.

Hypergeometric distribution: The distribution of successes in finite-population sampling without replacement.

Poisson distribution: A count model for events occurring at a constant average rate.

Normal distribution: A symmetric continuous model defined by a mean and variance.

Sampling distribution: The probability distribution of a statistic over repeated samples.

Review Questions

Short Answer Questions

1. Distinguish a target population from an accessible population using a quality-management example.
2. Explain why a statistic is considered a random variable before data are collected.
3. Define expectation and state whether it must be an attainable observed value.
4. What is the principal difference between a binomial and a hypergeometric sampling situation?
5. Why does zero correlation not generally establish independence?

Long Answer Questions

1. Explain the roles of populations, samples, parameters, and statistics in statistical inference for quality management.
2. Derive the relationship between the variance of a sum and covariance, and discuss its relevance to correlated quality measurements.
3. Compare the Bernoulli, binomial, Poisson, and hypergeometric models, including assumptions and applications.
4. Explain the central limit theorem’s relevance to sampling distributions and identify conditions that may weaken its practical approximation.
5. Discuss how measurement-system error, sampling bias, and model misspecification can affect a quality decision.

Analytical / Application Questions

1. A plant samples 50 units from a lot of 600 without replacement. Explain which probability model is most exact for the number of nonconforming units and when a binomial approximation might be defensible.
2. Daily complaints are recorded as 0, 1, 2, or more, and the variance of daily counts is substantially greater than the mean. Evaluate whether a Poisson model is adequate and propose possible explanations.
3. Consecutive thickness measurements show a high positive autocorrelation. Explain how this affects the usual formula for the variance of the sample mean and recommend an analytical response.

Multiple Choice Questions with Answers

Question 1

Which statement correctly distinguishes a parameter from a statistic?

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

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

C. A parameter is always random, while a statistic is never random.

D. A parameter and a statistic are different names for the same quantity.

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

Question 2

For a discrete random variable, which condition must a probability mass function satisfy?

A. All probabilities must be greater than 1.

B. The probabilities must sum to 0.

C. The probabilities must be nonnegative and sum to 1.

D. Every possible value must have the same probability.

Correct Answer: C. The probabilities must be nonnegative and sum to 1.

Question 3

If X has mean 4 and variance 9, what are the mean and variance of Y = 2X + 5?

A. Mean 9; variance 18

B. Mean 13; variance 18

C. Mean 13; variance 36

D. Mean 9; variance 36

Correct Answer: C. Mean 13; variance 36

Question 4

Which statement about independence is correct?

A. Zero covariance always proves independence.

B. Independence implies zero covariance when the variances exist.

C. Independence can occur only for continuous variables.

D. Correlation must equal 1 for independent variables.

Correct Answer: B. Independence implies zero covariance when the variances exist.

Question 5

A finite lot contains 100 units, including 10 nonconforming units. Five units are selected without replacement. Which model is most exact for the number of nonconforming units?

A. Normal

B. Poisson

C. Binomial

D. Hypergeometric

Correct Answer: D. Hypergeometric

Question 6

If X follows a Poisson distribution with rate λ = 3, what is Var(X)?

A. 1

B. 3

C. 6

D. 9

Correct Answer: B. 3

Question 7

For independent observations with common variance σ², what is Var(X̄) for a sample of size n?

A. σ²n

B. σ² + n

C. σ²/n

D. σ/n²

Correct Answer: C. σ²/n

Question 8

A process produces binary conforming/nonconforming outcomes with constant probability across independent trials. Which model describes the number of nonconforming units in n trials?

A. Binomial

B. Uniform

C. Exponential

D. Weibull

Correct Answer: A. Binomial

Question 9

What does the standard deviation of a sampling distribution of X̄ represent?

A. The population mean

B. The measurement unit conversion factor

C. The sample size

D. The sample-to-sample variability of the sample mean

Correct Answer: D. The sample-to-sample variability of the sample mean

Question 10

In a Weibull reliability model, what does a shape parameter β greater than 1 generally indicate?

A. A decreasing hazard rate

B. A constant hazard rate

C. An increasing hazard rate associated with wear-out

D. A model that cannot describe failure times

Correct Answer: C. An increasing hazard rate associated with wear-out