Probability Distributions and Sampling

Introduction

Probability distributions provide a formal language for uncertainty. Managers use them to estimate demand, assess service levels, forecast defects, price risk, plan capacity and decide how much evidence is needed before taking action. A probability model does not predict one certain outcome; it describes the possible outcomes, their likelihoods and the variation that decision makers should expect.

This chapter develops the principal discrete and continuous distributions used in business analytics: the binomial, Poisson, normal and exponential distributions. It then explains sampling distributions, the central limit theorem and the distinction between population parameters and sample statistics. Particular attention is given to assumptions, interpretation and the consequences of using an inappropriate model.

Learning Objectives

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

1. Distinguish discrete from continuous random variables and describe probability distributions.
2. Apply the binomial, Poisson, normal and exponential models.
3. Calculate and interpret expected values, variances and standard deviations.
4. Explain sampling distributions and the central limit theorem.
5. Construct and interpret approximate probabilities for sample means and proportions.
6. Evaluate assumptions, limitations and managerial implications of distributional models.

1. Random Variables and Probability Distributions

A random variable assigns a numerical value to each outcome of a random experiment. A discrete random variable has countable outcomes, such as the number of complaints received in a day. A continuous random variable can take any value within an interval, such as delivery time or transaction value.

A probability mass function (PMF) gives probabilities for discrete values. For a discrete variable X, the probabilities satisfy:

A probability density function (PDF) describes the relative concentration of a continuous variable. Probabilities are areas under the density curve:

For a continuous variable, P(X = x) = 0 for every single point; probability is assigned to intervals. The cumulative distribution function (CDF) is:

The expected value is the probability-weighted long-run average:

For a continuous variable, E(X) = ∫ x f(x) dx. The variance measures dispersion around the mean:

where μ = E(X). The standard deviation is σ = √Var(X), expressed in the same units as X. Variance is measured in squared units.

A useful transformation rule is:

where a and b are constants. Adding a constant changes the mean but not variability; multiplying by b scales the standard deviation by |b|.

2. Discrete Probability Models

2.1 The Binomial Distribution

The binomial model counts successes in n repeated trials. It applies when:

• the number of trials n is fixed;
• each trial has two mutually exclusive outcomes, success and failure;
• the probability of success p is constant;
• trials are independent, or independence is a reasonable approximation.

If X is the number of successes, then X follows a binomial distribution with parameters n and p, written X ~ Binomial(n,p). Its probability function is:

Here, C(n,x) = n!/[x!(n−x)!] counts the possible arrangements. The mean and variance are:

Worked example. A bank’s quality-control review finds that 4% of online applications require manual correction. If 20 independent applications are selected, the probability that exactly two require correction is:

P(X=2) = C(20,2)(0.04)²(0.96)¹⁸

Since C(20,2)=190, the result is approximately 0.1458, or 14.58%. The expected number requiring correction is np=20×0.04=0.8, and the standard deviation is √(20×0.04×0.96)≈0.876.

A binomial model should not be used automatically for sampling without replacement. If the sample is a substantial fraction of a finite population, trial outcomes are dependent and a hypergeometric model may be more appropriate.

2.2 The Poisson Distribution

The Poisson distribution models the number of events occurring in a fixed interval of time, area, length or volume when events occur independently at a stable average rate. If λ is the expected number of events in the interval:

For a Poisson variable:

If the rate is r events per unit and the interval has length t, then λ = rt. This rate conversion is essential: the units of r and t must be compatible.

Worked example. A help desk receives an average of 6 urgent tickets per hour. Assuming a stable independent arrival rate, the probability of exactly 4 tickets in one hour is:

P(X=4)=e⁻⁶6⁴/4!≈0.1339.

The probability of no tickets is P(X=0)=e⁻⁶≈0.00248. For a 30-minute period, λ = 6 × 0.5 = 3, not 6.

The Poisson process has important extensions. Counts in disjoint intervals are independent, and the number in an interval of length t has mean rt. However, changing arrival rates, clustering, capacity constraints and dependence can invalidate the model. In operational settings, overdispersion—variance greater than the mean—is evidence that a simple Poisson model may be inadequate.

3. Continuous Probability Models

3.1 The Normal Distribution

The normal distribution is continuous, symmetric and bell-shaped. It is characterized by mean μ and standard deviation σ, with σ > 0. Its density is:

The standardized normal variable is:

After standardization, Z has mean 0 and standard deviation 1. Normal tables or software provide P(Z≤z). The transformation is valid when X is normally distributed, or when a normal approximation is justified.

Worked example. Suppose daily demand is approximately normal with mean 500 units and standard deviation 80 units. The probability that demand exceeds 620 is:

z=(620−500)/80=1.50. Therefore, P(X>620)=P(Z>1.50)≈0.0668, or 6.68%. A stock level of 620 units would therefore be exceeded on roughly 7 days in 100 under the model.

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 normal. These are approximations, not universal guarantees.

The normal model is often used for measurement errors, aggregate demand and process outcomes. It is inappropriate for strongly skewed variables, bounded proportions, nonnegative waiting times or count data unless a defensible approximation has been established.

3.2 The Exponential Distribution

The exponential distribution models a nonnegative waiting time until the next event in a Poisson process with constant rate λ > 0. Its density and cumulative distribution are:

The mean and variance are:

The rate λ must be expressed in the reciprocal units of t. The exponential distribution has the memoryless property:

Thus, conditional on having already waited s units, the additional waiting time has the same distribution as a new waiting time. This property is mathematically convenient but may be unrealistic when customers abandon queues, service deteriorates or arrival rates change.

Worked example. If service requests arrive at a rate of 3 per hour, the expected time until the next request is 1/3 hour, or 20 minutes. The probability of waiting more than 30 minutes is:

P(T>0.5)=e⁻(3)(0.5)=e⁻¹⋅⁵≈0.2231.

The exponential model concerns waiting time; the corresponding Poisson model concerns the number of events in an interval. They are linked, but they answer different questions.

4. Choosing and Comparing Distributions

Model selection should begin with the data-generating mechanism, not with convenience. Analysts should inspect histograms, empirical means and variances, time plots, outliers and dependence. Goodness-of-fit tests and information criteria can support, but not replace, subject-matter reasoning. A statistically plausible model may still be operationally wrong if the process changes during the decision horizon.

5. Sampling Distributions

A population is the complete set of units of interest. A sample is a subset observed to learn about that population. A population parameter, such as μ or p, is fixed but usually unknown. A sample statistic, such as the sample mean X̄ or sample proportion p̂, varies from sample to sample.

For a random sample X₁, X₂, …, Xₙ from a population with mean μ and variance σ², the sample mean is:

If observations are independent and have common variance σ²:

The standard error is the standard deviation of a statistic’s sampling distribution, not the standard deviation of individual observations. Increasing n reduces the standard error by the square-root rule: quadrupling the sample size halves it.

When sampling without replacement from a finite population of size N, the finite population correction may be appropriate when n/N is not small:

This assumes simple random sampling and a finite population variance convention consistent with the formula. In practice, clustering, stratification and weighting can make the actual standard error differ substantially from the simple random-sample result.

For a population proportion p, the sample proportion is:

Under independent Bernoulli observations:

The value p is usually unknown; an estimated standard error replaces p with p̂.

6. The Central Limit Theorem

The central limit theorem (CLT) states, broadly, that for independent observations with finite mean μ and finite variance σ², the standardized sample mean approaches a standard normal distribution as n becomes large:

The arrow denotes convergence in distribution. The CLT does not say that individual observations become normal. It says that the distribution of sample means becomes approximately normal. The required sample size depends on skewness, tail behavior, dependence and the accuracy demanded. There is no universal rule that n = 30 always suffices.

If the population itself is normal, X̄ is exactly normal for every n, with mean μ and standard error σ/√n. If σ is unknown and the population is normal, inference about μ uses the t distribution:

where s is the sample standard deviation and degrees of freedom are n − 1. The t distribution has heavier tails than the standard normal, especially for small samples, reflecting uncertainty from estimating σ.

For proportions, a normal approximation is often reasonable when np and n(1−p) are sufficiently large. Rules such as both being at least 5 or 10 are heuristics, not laws. Exact binomial methods are preferable for small samples or extreme probabilities.

7. Sampling and Decision Quality

Sampling error is random variation caused by observing a sample rather than the entire population. It is distinct from nonsampling error, which includes nonresponse, coverage problems, measurement error, interviewer effects, data-processing errors and selection bias. A very large biased sample can be less useful than a smaller probability sample.

A confidence interval for a mean with known σ, under normality or a justified CLT approximation, has the form:

Here zₐ/₂ is the standard-normal critical value leaving α/2 in each tail; the confidence level is 1 − α. When σ is unknown, a t critical value and s replace z and σ. The interval describes a method that captures the fixed parameter in repeated sampling at the stated long-run rate; it does not assign a probability to the already fixed parameter in the frequentist interpretation.

Worked example. A retailer samples 64 delivery times. The sample mean is 42 minutes and the process standard deviation is known to be 8 minutes. A 95% confidence interval is:

42 ± 1.96(8/√64) = 42 ± 1.96(1)

Thus the interval is approximately 40.04 to 43.96 minutes. The margin of error is 1.96 minutes under the stated assumptions.

Sample-size planning for estimating a mean with known target margin E uses:

The result should be rounded upward. The formula assumes a simple random sample, a known or defensibly anticipated σ, the selected confidence level and the desired absolute margin in the same units as the outcome. For a proportion, a conservative planning formula is:

When p is unknown, p = 0.5 gives the largest required sample size for a fixed confidence level and margin.

8. Illustrative Case: Service-Centre Capacity Planning

Background. A fictional regional service centre receives technical-support calls. Historical data suggest an average of 12 calls per hour during the afternoon peak. Management is considering whether one additional specialist is needed.

Problem. The centre wants to estimate the probability of receiving more than 15 calls in a one-hour period and the probability that the interval between calls exceeds 10 minutes.

Analysis. Assume, as an initial model, a Poisson process with λ = 12 calls per hour. The required probability is:

P(X>15)=1−P(X≤15)=1−∑ₓ₌₀¹⁵ e⁻¹²12ˣ/x!.

This probability should be calculated with statistical software to avoid rounding errors. For the waiting time T to the next call, λ = 12 per hour and 10 minutes = 1/6 hour:

P(T>1/6)=e⁻¹²⁄⁶=e⁻²≈0.1353.

Thus, the initial model estimates a 13.53% chance of waiting more than ten minutes for the next call. The probability of more than 15 calls is a separate count probability and must not be substituted with the waiting-time calculation.

Approach and managerial interpretation. The operations team should compare the model with hourly arrival data, examine whether peak rates are stable, and test for clustering caused by system outages or product releases. If the observed variance substantially exceeds the mean, a negative-binomial or time-varying-rate model may be more appropriate. Capacity decisions should also incorporate service-time distributions, abandonment, staffing schedules and the cost of overtime—not merely call counts.

Learning points. First, Poisson and exponential models are complementary rather than interchangeable. Second, a probability estimate is conditional on assumptions. Third, the operational decision requires a cost-based capacity analysis after the statistical model has been validated.

9. Limitations, Governance and Contemporary Practice

Probability distributions are models, not facts. Historical stability may fail because of inflation, platform changes, promotions, regulation, seasonality or strategic responses. Analysts should document the population, sampling frame, observation period, units, missing-data treatment, model assumptions and validation results.

Ethically, sampling plans must avoid systematically excluding vulnerable groups. Customer and employee data should be collected lawfully, minimized and protected. Automated decision systems should disclose material uncertainty and should not treat model-based probabilities as objective certainty. If a distribution is used in credit, insurance, hiring or public services, fairness testing and governance review are essential.

Modern analytics platforms can fit distributions, simulate scenarios and monitor drift. AI systems can assist with anomaly detection and probabilistic forecasting, but they do not eliminate the need for independent validation. A responsible workflow is: define the decision; identify the data-generating process; inspect and clean data; select a candidate model; estimate parameters; validate out of sample; quantify uncertainty; perform sensitivity analysis; and monitor performance after deployment.

Best practice includes reporting both point estimates and uncertainty, distinguishing statistical from practical significance, using robust or exact methods when assumptions are doubtful, and communicating results in operational units. For example, “a 6.7% chance of exceeding 620 units” is more useful to a supply manager than an unexplained z-score, provided the time horizon and assumptions are stated.

Chapter Summary

Probability distributions quantify uncertainty in discrete counts and continuous measurements. The binomial distribution applies to a fixed number of independent two-outcome trials; the Poisson distribution models event counts under a stable rate; the normal distribution models symmetric continuous outcomes and supports standardization; and the exponential distribution models waiting times in a constant-rate Poisson process. Each model has specific assumptions that must be checked.

Sampling distributions describe how statistics vary across repeated samples. The sample mean is unbiased for the population mean, and its standard error is σ/√n under independent equal-variance observations. The central limit theorem explains why sample means are often approximately normal, but its adequacy depends on sample size and population shape. Sound managerial decisions require attention to sampling design, nonsampling error, model validation, uncertainty, ethics and changing operational conditions.

Key Terms

Random variable: A numerical function assigning values to outcomes of a random experiment.

Discrete variable: A variable with countable possible values.

Continuous variable: A variable that can take values across an interval.

Probability mass function: A function assigning probabilities to discrete outcomes.

Probability density function: A function whose area over an interval gives continuous probability.

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

Expected value: The probability-weighted long-run average.

Variance: The expected squared deviation from the mean.

Binomial distribution: A model for successes in a fixed number of independent trials.

Poisson distribution: A model for event counts over a fixed interval at a stable rate.

Normal distribution: A symmetric continuous bell-shaped distribution defined by μ and σ.

Exponential distribution: A model for waiting time until the next event in a constant-rate Poisson process.

Population parameter: A fixed numerical characteristic of a population.

Sample statistic: A numerical quantity calculated from a sample.

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

Standard error: The standard deviation of a statistic’s sampling distribution.

Central limit theorem: A result describing the approximate normality of standardized sample means.

Finite population correction: An adjustment reducing sampling variance when sampling a substantial population fraction.

Nonsampling error: Error caused by coverage, nonresponse, measurement or processing rather than random sampling variation.

Review Questions

5 Short Answer Questions

1. What distinguishes a discrete random variable from a continuous random variable?
2. State the assumptions required for a binomial model.
3. Why are the mean and variance equal in a Poisson distribution?
4. What does the standard error of a sample mean measure?
5. Explain the difference between sampling error and nonsampling error.

5 Long Answer Questions

1. Explain the binomial and Poisson distributions, including their assumptions, formulas and business applications.
2. Describe how a normal variable is standardized and explain the managerial meaning of a z-score.
3. Explain the relationship between a Poisson process and the exponential distribution.
4. Present the central limit theorem and discuss when its approximation may be unreliable.
5. Discuss how ethical and governance considerations affect the use of probability models in management decisions.

3 Analytical / Application Questions

1. A factory finds that 3% of units are defective. For a random sample of 25 independent units, calculate the probability of exactly one defective unit and interpret the result.
2. A call centre receives 10 calls per hour. Using a Poisson model, calculate the probability of no calls in 15 minutes and explain the rate conversion.
3. A sample of 100 customers has an average waiting time of 18 minutes. If the known population standard deviation is 6 minutes, construct a 95% confidence interval for the population mean and identify two assumptions that should be evaluated.

Multiple Choice Questions with Answers

Question 1

Which condition is essential for a binomial model?

A. The number of trials is fixed.

B. The variable must be continuous.

C. The event rate must equal the variance.

D. Waiting times must be memoryless.

Correct Answer: A. The number of trials is fixed.

Question 2

A Poisson process has a rate of 4 events per hour. What is the expected number of events in 30 minutes?

A. 0.5

B. 2

C. 4

D. 8

Correct Answer: B. 2

Question 3

If X has mean μ and standard deviation σ, which expression standardizes X?

A. (X+μ)/σ

B. (X−μ)/σ

C. (X−σ)/μ

D. X/(μσ)

Correct Answer: B. (X−μ)/σ

Question 4

For a continuous random variable, what is P(X = x) under an ordinary continuous distribution?

A. 0

B. 0.5

C. 1

D. It must equal the density at x

Correct Answer: A. 0

Question 5

If the standard deviation of individual observations is 20 and n = 100, what is the standard error of the sample mean under independent sampling?

A. 0.2

B. 2

C. 20

D. 200

Correct Answer: B. 2

Question 6

The exponential distribution is most directly used to model:

A. The number of successes in fixed trials

B. The number of events in a fixed interval

C. The waiting time until the next event

D. The sample variance of a population

Correct Answer: C. The waiting time until the next event

Question 7

A population proportion is estimated from n independent Bernoulli observations. Which is the variance of the sample proportion?

A. p(1−p)/n

B. np(1−p)

C. p/n(1−p)

D. p(1+p)/n

Correct Answer: A. p(1−p)/n

Question 8

What does the central limit theorem primarily describe?

A. The normality of every individual observation

B. The approximate normality of standardized sample means for large samples

C. The equality of a population mean and variance

D. The elimination of nonsampling error

Correct Answer: B. The approximate normality of standardized sample means for large samples

Question 9

If calls arrive at 12 per hour, what is the probability of waiting more than 10 minutes for the next call under an exponential model?

A. e⁻¹²

B. e⁻²

C. 1−e⁻²

D. e²

Correct Answer: B. e⁻²

Question 10

When σ is unknown and the population is normal, inference for a mean commonly uses which distribution?

A. Poisson distribution

B. Binomial distribution

C. Student’s t distribution

D. Exponential distribution

Correct Answer: C. Student’s t distribution