Sampling, Sampling Distributions, and Statistical Quality Control

Sampling Fundamentals and Sources of Bias

A population encompasses all items of interest in a statistical problem, whereas a sample is a representative subset selected from that population. A population parameter represents a constant, fixed value, while a sample statistic is a random variable whose observed value varies based on the specific random sample drawn. An estimator refers to a sample statistic employed to estimate an unknown population parameter, and the numerical value obtained from a specific sample is an estimate. Statistical bias occurs when a sample statistic systematically overestimates or underestimates a parameter, typically caused by unrepresentative sampling techniques.

Sampling biases compromise inferential accuracy in several distinct forms. Selection bias occurs when portions of the target population are systematically excluded or underrepresented during sample selection, as exemplified by the 1936 Literary Digest presidential poll where telephone directories and automobile registries oversampled affluent households favoring Alf Landon over Franklin Delano Roosevelt. Nonresponse bias arises when individuals who participate in a survey possess preferences that differ systematically from nonrespondents; George Gallup correctly predicted the 1936 election outcome by surveying a representative sample of $50,000$ people with a $10\%$ response rate ($5,000$ respondents). Social-desirability bias emerges when respondents misrepresent their true stances due to perceived social stigma, resulting in systematic discrepancies between socially acceptable survey replies and actual decisions, as observed in polling disparities between live telephone interviews and anonymous online surveys during the 2016 election involving Donald Trump and Hillary Clinton.

Common Sampling Methods

A simple random sample guarantees that every possible sample of size $n$ possesses an identical probability of being chosen from the population. In practical computational environments, simple random samples can be drawn by ordering population items with unique identifiers and utilizing discrete random number generation formulas, such as $=\text{RANDBETWEEN}(\text{lower}, \text{upper})$ in Excel.

Stratified random sampling and cluster sampling offer structured alternatives to simple random sampling. In stratified random sampling, the population is divided into mutually exclusive and collectively exhaustive groups termed strata according to defined characteristics, and independent random samples are collected from each stratum proportionally to its relative population size before being pooled, which ensures subpopulation representation and produces parameter estimates with greater precision. In cluster sampling, the population is separated into mutually exclusive and collectively exhaustive groupings called clusters, frequently based on geographic divisions, and a random sample of whole clusters is selected for analysis. While cluster sampling exhibits less precision for a fixed sample size than simple or stratified random sampling, it substantially reduces operational costs when constructing an exhaustive population frame is difficult or prohibitive.

Sampling Distribution of the Sample Mean

The sampling distribution of the sample mean $\bar{X}$ represents the probability distribution of sample means obtained from all possible samples of size $n$ drawn from a given population. The expected value of the sample mean is equal to the population mean, $E(\bar{X}) = \mu$, confirming that $\bar{X}$ is an unbiased estimator. The dispersion among sample means is measured by the standard error of the sample mean, defined as $se(\bar{X}) = \frac{\sigma}{\sqrt{n}}$, where the variance of the sample mean is $Var(\bar{X}) = \frac{\sigma^2}{n}$. Averaging reduces variability because high and low individual observations offset one another within a sample. When a sample constitutes a substantial share of the population ($n \ge 0.05\,N$), a finite population correction factor is warranted to adjust standard error calculations.

When sampling from a population that follows a normal distribution with mean $\mu$ and standard deviation $\sigma$, the sampling distribution of $\bar{X}$ is strictly normal for any sample size $n$. For non-normal populations, the Central Limit Theorem dictates that the sum or average of a large number of independent observations will possess an approximately normal distribution regardless of the underlying population shape. In practical applications, the normal distribution approximation for $\bar{X}$ is considered valid whenever the sample size satisfies $n \ge 30$.

Sampling Distribution of the Sample Proportion

For categorical variables, statistical inference focuses on the population proportion $p$, with the corresponding point estimator being the sample proportion $\bar{P} = \frac{X}{n}$, where $X$ is the number of successes in $n$ Bernoulli trials. The expected value of the sample proportion matches the population parameter, $E(\bar{P}) = p$, demonstrating that $\bar{P}$ is an unbiased estimator. The variance of the sample proportion is $Var(\bar{P}) = \frac{p(1 - p)}{n}$, and the standard error of the sample proportion is given by $se(\bar{P}) = \sqrt{\frac{p(1 - p)}{n}}$.

The Central Limit Theorem applies directly to categorical data, establishing that the sampling distribution of $\bar{P}$ approaches a normal distribution as the sample size expands. The normal distribution approximation for the sample proportion is justified whenever the expected numbers of successes and failures satisfy the threshold conditions $n p \ge 5$ and $n(1 - p) \ge 5$. Populations whose proportion $p$ departs markedly from $0.50$ require larger sample sizes $n$ to fulfill these conditions and attain normality.

Statistical Quality Control and Process Variation

Statistical quality control applies quantitative methodologies to evaluate, maintain, and enhance an organization's capacity to deliver uniform goods and services. Traditional acceptance sampling inspects finished products after production and discards or repairs nonconforming items, a practice that incurs high costs and risks releasing defective items to consumers. Modern quality management prioritizes the detection approach, which monitors operating conditions along the production line to verify conformity with target specifications before large batches of nonconforming items are manufactured.

Variations in any production system are partitioned into two fundamental categories. Chance variation represents expected, random occurrences inherent to normal machine and human operations that cannot be completely eliminated, and it requires no managerial intervention provided end products conform to acceptable operational tolerance limits. Assignable variation stems from specific, identifiable, and rectifiable causes—such as machine tool wear, incorrect settings, or equipment misalignment—that must be isolated and eliminated to restore operational control.

Control Charts for Numerical and Categorical Variables

Pioneered by Walter A. Shewhart at Bell Telephone Laboratories, control charts track process stability over time by plotting sequential sample statistics against established boundaries. Numerical process monitoring utilizes the $\bar{x}$ chart to supervise process central tendency, alongside $R$ charts or $s$ charts to supervise process variability. Categorical monitoring utilizes the $p$ chart (fraction or percent defective chart) to evaluate the proportion of nonconforming items, as well as the $c$ chart to track counts of defects per unit.

A standard control chart consists of a centerline, an Upper Control Limit ($UCL$), and a Lower Control Limit ($LCL$). The centerline reflects the expected value of the monitored statistic when the system operates in control, while the control limits are set three standard errors away: $UCL = \text{Expected Value} + 3 \times (\text{Standard Error})$ and $LCL = \text{Expected Value} - 3 \times (\text{Standard Error})$. Under the assumption of normality, the interval between $LCL$ and $UCL$ captures a probability of $0.9973$, resulting in a false-alarm probability of $0.0027$ ($1 - 0.9973$). In constructing a $p$ chart, if computational formulas yield $UCL > 1$, the limit is adjusted to $1$, and if $LCL < 0$, it is set to $0$. A process is deemed in statistical control when sample points disperse randomly between the control limits; a process is deemed out of control if any point falls outside the limits, or if non-random behavior—such as sustained runs on one side of the centerline or clear upward or downward trends—is observed.