Statistical Methods in Quality Management

Importance of Statistics in Quality Management

Quality management (QM) is a systematic approach to ensuring products, services, or processes meet or exceed customer expectations. It integrates planning, monitoring, and improvement of operations to deliver consistent, reliable outcomes. Core principles include customer focus, process optimization, continuous improvement, and evidence-based decision making. Frameworks such as Total Quality Management (TQM) , Six Sigma, and ISO standards provide structured methods to reduce errors, increase efficiency, foster customer satisfaction, and maintain competitive advantage.

Statistics supplies the tools to measure, analyze, and improve quality scientifically. By applying statistical methods, organizations can:

• Quantify variation in processes.
• Identify root causes of defects.
• Implement solutions for consistent performance.

Key Statistical Techniques in QM

Intuition: Why Statistics Matters

A manufacturing process (or any pipeline) produces a stream of items. Inspecting every item is impossible for time and cost reasons. Instead, sample statistics (e.g., sample mean, sample standard deviation) are used to infer whether the entire output meets gold standards. Statistical techniques like estimation and hypothesis testing become critical for drawing these conclusions.

Six Sigma exemplifies the role of statistics in reducing process variability to achieve near-perfect quality levels. By using hypothesis testing, regression, or DOE, businesses detect deviations from the gold standard and pinpoint which part of the pipeline causes the deviation. This enables data-driven decisions that keep errors within acceptable limits.

Worked Examples (from industry)

• Automobile manufacturer uses regression analysis to identify factors correlated with defects in car components. The regression model reveals which process variables are positively correlated with defects, allowing targeted corrective actions.
• Pharmaceutical company uses design of experiments (DOE) to optimise the formulation of a new drug, systematically varying ingredients to find the best combination.

Benefits of Statistical QM

• Facilitates continuous improvement by providing measurable insights into process performance.
• Enables setting benchmarks and tracking progress.
• Supports evidence-based adjustments to meet quality standards.
• Ensures regulatory compliance in manufacturing, healthcare, and service delivery.
• Minimises waste and enhances customer satisfaction.

Key takeaways

• Quality management relies on evidence-based decisions; statistics provides the scientific measurement and analysis framework.
• Control charts and sampling are foundational for monitoring and cost-efficient inspection.
• Estimation and hypothesis testing let managers infer process behaviour from samples.
• Regression analysis and DOE help identify causes of defects and optimise products.
• Statistical methods transform quality management from abstract principle into a measurable, improvable process.

Hypothesis Testing in Quality Management

Hypothesis testing enables managers to decide whether an observed deviation in a process or product is due to random chance or signals a real problem. In quality management, these decisions have direct operational, financial, and safety consequences.

Type I and Type II Errors

Every test risks one of two mistakes. The null hypothesis ($H_0$) typically represents the “status quo” — that the process is acceptable. Rejecting $H_0$ means concluding there is a problem; failing to reject means no action is taken.

Why each error matters in quality

• Type I error ($\alpha$) – “False alarm.”
  Example: A batch that meets standards is rejected, causing unnecessary rework, scrap, or inspection costs.
  Control: Set a low significance level (commonly $\alpha = 0.05$ or 0.01).

• Type II error ($\beta$) – “Miss.”
  Example: A defective batch passes inspection, leading to recalls, legal liability, or safety hazards (e.g., faulty brake pads, wrong drug concentration).
  Control: Increase sample size or test power ($1-\beta$).

Exam tip: In quality, Type I errors waste money; Type II errors can kill people. Which error matters more depends on the consequence — not on a fixed rule.

Consequences of errors in quality management

Key takeaways – Type I & II errors

• Type I = false positive (reject true $H_0$); Type II = false negative (fail to reject false $H_0$).
• $\alpha$ (significance level) controls Type I; $\beta$ controls Type II (power = $1-\beta$).
• In manufacturing, Type I often drives cost; in safety-critical settings, Type II is paramount.
• Balancing errors requires setting appropriate $\alpha$, increasing sample size, and choosing the correct test.

---

Worked Examples

Example 1: Steel rod diameter (two-tailed test)

Situation: A factory produces steel rods with a target mean diameter $\mu = 5$ mm. A random sample of $n=30$ rods yields $\bar{x}=4.98$ mm, $s=0.05$ mm. Is the process misaligned?

• Hypotheses: $H_0: \mu = 5$ vs. $H_1: \mu \neq 5$ (two-tailed — both under‑ and over‑fill are unacceptable).
• Test statistic (t-test, since $\sigma$ unknown):

$t = \frac{\sqrt{n}(\bar{x}-\mu_0)}{s} = \frac{\sqrt{30}(4.98-5)}{0.05} = -2.2$

• Null distribution: $t$ with $df = 29$.
• p-value: For a two-tailed test, $p = 2 \times P(T_{29} < -2.2) \approx 2 \times 0.018 = 0.036$ (3.6%).
• Decision: $p = 0.036 < \alpha = 0.05$ → reject $H_0$.

Conclusion: The observed deviation is unlikely due to chance. The machinery needs recalibration.

Example 2: Brake pad compressive strength (one-tailed test)

Situation: Minimum acceptable mean compressive strength is 100 MPa. A sample of $n=15$ brake pads gives $\bar{x}=99.3$ MPa, $s=1.6$ MPa. Does the supplier meet the standard?

• Hypotheses: $H_0: \mu \geq 100$ vs. $H_1: \mu < 100$ (one-tailed — only under‑strength is a concern).
• Test statistic:

$t = \frac{\sqrt{15}(99.3-100)}{1.6} = -1.7$

• Null distribution: $t$ with $df = 14$.
• p-value: $P(T_{14} < -1.7) \approx 0.056$ (5.6%).
• Decision: $p = 0.056 > \alpha = 0.05$ → do not reject $H_0$.

Conclusion: There is insufficient evidence to claim the supplier is substandard. The sample mean of 99.3 MPa could be due to random variation. (Recommend increasing sample size for a definitive verdict.)

Exercises for practice

1. Sugar packaging: A company fills 1‑kg bags. Suspecting underfilling, they sample $n=25$ bags: $\bar{x}=0.97$ kg, $s=0.05$ kg. Formulate hypotheses and decide (one‑tailed? two‑tailed?) whether the filling machine needs adjustment.
2. Call center wait times: Before a new scheduling system, average wait time was 4.5 min. After, a sample of $n=50$ calls shows $\bar{x}=4.1$ min, $s=0.8$ min. Test whether the system effectively reduced wait time.

Key takeaways – Hypothesis testing in practice

• Choose the right tail: Two-tailed when any deviation is unacceptable; one-tailed when only a specific direction matters (e.g., under‑strength, over‑weight).
• Use $t$‑test when $\sigma$ unknown — the most common case in quality.
• Context drives the decision: A “non‑significant” result (p > 0.05) may still warrant follow‑up with a larger sample if the consequence of a Type II error is severe.
• Hypothesis testing enables early detection, supplier accountability, evidence-based process changes, and cost savings from reduced waste/rework.

Regression Analysis to Model Input and Output Relationships in Quality Control

Regression analysis models how input variables (independent variables) affect an output variable (dependent variable). In quality control, this quantifies the relationship between process factors and product quality, enabling prediction, identification of key drivers, and optimization of operating conditions. By understanding which inputs matter and by how much, teams can reduce variability, minimize defects, and make data-driven improvements.

Importance of Regression in Quality Control

• Identify key drivers – Determines which factors (e.g., machine temperature, raw material properties) significantly influence quality or defect rates.
• Quantify relationships – Produces a mathematical model that predicts output from input values, supporting proactive decision-making and scenario forecasting.
• Find optimal settings – Helps minimize defects, risks, or costs by selecting input levels that keep the output within acceptable bounds.
• Understand sources of variability – Reduces unwanted variation by isolating which inputs contribute to inconsistency.
• Support Six Sigma initiatives – Provides evidence for reducing defects, rework, and costs through targeted process changes.

Linear Regression: Worked Examples

The general linear regression model with multiple inputs is written as:

$Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \cdots + \epsilon$

where $Y$ is the output, $X_i$ are inputs, $\beta_i$ are coefficients, and $\epsilon$ is random error.

Example 1: Controlling Product Weight in Packaging

A food company packages 1‑kg bags. Two inputs are suspected to affect bag weight: machine speed (bags per minute) and hopper pressure (psi). Data collected at various settings yields the following linear regression results:

• Every increase of 10 bags per minute → bag weight decreases by 0.02 kg.
• Every increase of 5 psi → bag weight increases by 0.01 kg.

The company uses scenario forecasting to predict weight at different input combinations, e.g., machine speed = 40 bags/min and hopper pressure = 15 psi gives a weight closest to 1 kg. Multiple acceptable settings may exist; the regression model provides the flexibility to choose a feasible, compliant combination.

Example 2: Defect Rate in an Assembly Line

An electronics manufacturer wants to predict the percentage of defective units based on two inputs: worker training hours and machine maintenance frequency (days since last maintenance). Regression results:

• Every additional 5 hours of training → defect rate decreases by 4 %.
• Every additional 5 days since last maintenance → defect rate increases by 6 %.

The company seeks an optimum level, not necessarily zero defects, because unlimited training or daily maintenance would incur prohibitive costs. For an acceptable threshold of 10 % defects, the model suggests, e.g., maintaining machines every 10 days and providing at least 15 hours of training. This balances quality with cost.

Logistic Regression in Quality Control

Logistic regression is used when the output variable is binary, e.g., pass/fail, in control/out of control. It models the probability of an event (e.g., defect) as a function of predictors:

$P(Y=1) = \frac{1}{1 + e^{-(\beta_0 + \beta_1 X_1 + \beta_2 X_2 + \cdots)}}$

Example: A manufacturer wants to predict the probability that a product fails quality checks based on temperature, pressure, and machine speed. Logistic regression quantifies how each predictor affects the odds of failure. The manager can:

• Identify which factors contribute most to quality issues (high positive coefficients).
• Estimate the likelihood of defects under current process conditions.
• Determine optimal ranges for predictors to minimize defect probability.
• Pinpoint high-risk factors for immediate corrective action.

Exam tip: Logistic regression is ideal for binary quality outcomes (e.g., defective vs. non‑defective, conforming vs. non‑conforming). It provides probabilities, not continuous predictions, and is sensitive to sample size and class balance.

Key takeaways

• Regression (linear or logistic) quantifies how inputs affect outputs in quality management.
• Linear regression predicts continuous outcomes (weight, defect percentage) and aids scenario forecasting.
• Logistic regression predicts binary outcomes (pass/fail) and estimates probabilities of defects.
• Both methods identify key drivers and support optimization, but practical constraints (cost, feasibility) often require a target acceptable level rather than absolute perfection.
• Regression is a cornerstone for data-driven process improvement, variability reduction, and Six Sigma.

Statistical Decision Theory – Introduction

Statistical decision theory provides a structured framework for making choices when outcomes depend on uncertain events. The core idea: combine probabilities, data, and decision rules to minimize risk or maximize expected payoff. Every business manager faces decisions under uncertainty — the quality of those decisions hinges on the quantity and quality of available information.

---

Elements of the Decision Process

Every decision problem under uncertainty contains four components:

1. Alternative courses of action (acts/strategies) – the options the decision maker can choose (e.g., launch a new product vs. expand marketing).
2. States of nature – uncertain, uncontrollable events that affect outcomes (e.g., high market demand vs. low market demand).
3. Payoff – a numerical value (profit, cost, utility) conditional on the combination of an action and a state. The payoff is always conditional because the state is unknown at decision time.
4. Decision rule – the criterion for selecting the best action, based on the decision maker’s goals and available information.

---

Payoff Matrix

A payoff matrix is a table that displays the payoff for every action–state pair. It enables direct comparison of outcomes under different scenarios.

Payoff matrix = each row an action, each column a state, each cell the payoff for that combination.

Example: Product Launch vs. Marketing Expansion

The decision maker cannot simply pick the maximum payoff (₹200,000) because that payoff is contingent on "high demand" occurring. The matrix forces explicit consideration of both possibilities.

---

Opportunity Loss (Regret) Matrix

The opportunity loss matrix (or regret matrix) quantifies the cost of not choosing the best action for each state. It is constructed as:

$$\text{Regret} = \text{Maximum payoff in that state} - \text{Actual payoff from chosen action}$$

Steps to build it:

1. For each state, identify the maximum payoff across all actions.
2. In each column, subtract every payoff from that column’s maximum.

Regret Matrix for the Example

• High demand: max payoff = ₹200,000 (Launch)
• Low demand: max payoff = ₹100,000 (Expand)

Interpretation: If the firm launches the product and demand turns out to be low, it loses ₹50,000 compared to what it could have earned by expanding marketing.

When to use each:

Exam tip: The regret matrix is not an alternative payoff matrix — it is derived from the payoff matrix. A common mistake is to confuse the two. Regret always uses the difference from the best-in-column.

---

Key takeaways

• Four elements: actions, states of nature, conditional payoffs, and a decision rule.
• A payoff matrix organises all possible outcomes for each action–state pair.
• An opportunity loss matrix shows the regret of not picking the best action in each state; each cell = column max − actual payoff.
• The same example can be analysed with either matrix depending on whether the goal is reward maximisation or regret minimisation.

Concept of Decision-Making Under Uncertainty

Decision theory (or decision analysis) formalizes the process of choosing among alternative courses of action when at least two alternatives exist. The decision environment is classified by the amount of information available about the outcomes:

• Certainty – The outcome of each action is known perfectly. The decision is trivial: pick the action with the largest payoff. Statistically uninteresting.
• Uncertainty – More than one possible outcome (state of nature) exists, but the decision-maker cannot assign probabilities to them.
• Risk – States of nature exist and the decision-maker can assign probabilities (objective or subjective) to each.

Decision Process

1. Identify and define the problem.
2. List all possible future events – the states of nature ($S_1, S_2, \dots, S_n$).
3. Identify all possible courses of action (alternatives) – $A_1, A_2, \dots, A_m$.
4. Construct a payoff table showing the payoff (profit, cost, etc.) for every combination of action and state of nature.
5. Choose a decision criterion (rule) to select the best action.

---

Decision Making Under Uncertainty

The decision-maker has knowledge of the possible states of nature but lacks any information about their likelihoods. Four classic criteria are used:

1. Maximax Criterion (Criterion of Optimism)

Select the alternative that yields the maximum of the maximum payoffs.

• For each alternative, find the highest payoff across all states.
• Choose the alternative with the largest of these maxima.

Exam tip: Maximax is unrealistically optimistic; it assumes the best-case scenario will occur.

2. Maximin Criterion (Criterion of Pessimism)

Select the alternative that yields the maximum of the minimum payoffs.

• For each alternative, find the lowest payoff across all states.
• Choose the alternative with the largest of these minima.

This ensures a guaranteed worst-case floor – a conservative, security-oriented approach.

3. Minimax Regret Criterion (Opportunity Loss / Minimum Regret)

Minimises the maximum opportunity cost (regret) of a decision.

Procedure:

1. For each state of nature, identify the best possible payoff.
2. For each action–state combination, compute regret = best payoff in that state − actual payoff.
3. For each alternative, find the maximum regret.
4. Select the alternative with the minimum of those maximum regrets.

This criterion strikes a balance: it limits the worst regret a manager can experience, though it may sacrifice high rewards.

4. Hurwicz Criterion (Criterion of Realism)

A compromise between maximax and maximin. The decision-maker chooses a coefficient of optimism $\alpha$ ($0 \le \alpha \le 1$), where $\alpha = 1$ gives maximax and $\alpha = 0$ gives maximin.

For each alternative $i$, compute:

$$H_i = \alpha \times (\text{maximum payoff for } A_i) + (1-\alpha) \times (\text{minimum payoff for } A_i)$$

Select the alternative with the highest $H_i$.

---

Worked Example: Biscuit Selection

A shopkeeper must choose one biscuit brand: Marigold, Good Day, or Oreo. Three possible states of nature: expected sales next year – Low, Medium, High. Payoffs in thousands of rupees:

Maximax:
Maximum payoffs: Marigold = 45, Good Day = 65, Oreo = 70 → choose Oreo.

Maximin:
Minimum payoffs: Marigold = 15, Good Day = 20, Oreo = 25 → choose Oreo.

Minimax Regret:

First compute the best payoff in each state:

• Low: 25 (Oreo)
• Medium: 50 (Oreo)
• High: 70 (Oreo)

Regret table (best − actual):

Select Oreo (minimax regret = 0). In this example all three criteria select Oreo.

Hurwicz with $\alpha = 0.5$:

• Marigold: $0.5 \times 45 + 0.5 \times 15 = 30$
• Good Day: $0.5 \times 65 + 0.5 \times 20 = 42.5$
• Oreo: $0.5 \times 70 + 0.5 \times 25 = 47.5$ → choose Oreo.

Since both extremes point to Oreo, any weighted average also points to Oreo. This is not always the case, as seen next.

---

Additional Example: Bottling Company (Take-Home Exercise)

A company must choose among three actions to meet rising demand: Expand existing plant, Construct a new plant, or Subcontract production. Four states of nature: High, Medium, Low, or None (demand). Payoffs in thousands of rupees:

Applying the criteria will yield different optimal actions:

• Maximax → Construct (70).
• Maximin → Subcontract (−10 is the highest minimum).
• Minimax regret → Compute regret and select the action with smallest max regret (likely Subcontract or Expand).
• Hurwicz → Varies with $\alpha$; at $\alpha=0.5$:
  • Expand: $0.5(50) + 0.5(-45) = 2.5$
  • Construct: $0.5(70) + 0.5(-80) = -5$
  • Subcontract: $0.5(30) + 0.5(-10) = 10$ → Subcontract.

Exam tip: When maximax and maximin give different answers, Hurwicz offers a flexible middle ground. Always check how changing $\alpha$ alters the decision.

Key Takeaways

• Under uncertainty, no probabilities are known; under risk, probabilities are assigned.
• Maximax – optimistic; Maximin – pessimistic; Minimax regret – minimises worst-case opportunity loss; Hurwicz – weighted average of best and worst.
• The choice of criterion heavily influences the decision; in some problems different criteria recommend different actions.
• The Hurwicz $\alpha$ parameter reflects the manager’s degree of optimism.

Concept of Decision Trees

A decision tree is a graphical tool for evaluating multiple decision alternatives under uncertainty. It breaks down complex decisions into sequential steps, mimicking natural decision-making. Each node is a decision point or a chance event; branches represent possible choices or outcomes with associated probabilities and payoffs. The structured approach allows decision makers to visualize chains of events and systematically weigh risks and rewards.

Decision trees are widely used in business analytics, finance, marketing, operations, and quality management to optimise strategies and reduce uncertainty. The core objective: compute the expected monetary value (EMV) of each branch and choose the path with the highest expected payoff.

Example: Launching a New Smartwatch

A retail company must decide whether to launch a new smartwatch immediately or conduct market research first. Uncertainties include market reception (success or failure), research outcome (favours launch or not), and associated costs/probabilities. The decision tree captures all paths and their EMVs.

EMV calculation

• Launch immediately:
  $\text{EMV}_1 = 0.6 \times 1{,}000{,}000 + 0.4 \times (-500{,}000) = 600{,}000 - 200{,}000 = \text{Rs. }400{,}000$

• Market research first:

  • If research favours launch (70 % chance), the sub‑branch EMV is the same as above: ₹400,000.
  • If research does not favour launch (30 % chance), the loss is only the research cost: ₹200,000.
  • Overall EMV:
    $\text{EMV}_2 = 0.7 \times 400{,}000 + 0.3 \times 0 - 200{,}000 = 280{,}000 - 200{,}000 = \text{Rs. }80{,}000$

Since ₹400,000 > ₹80,000, the optimal decision is launch immediately.

Exam tip: Always subtract the research cost from the total EMV of the research branch — a common oversight. The EMV of the launch branch already accounts for the profit/loss numbers given.

Decision Trees for Quality Control Decisions

Decision trees also apply to batch inspection in manufacturing. A manager must decide: approve, perform further tests, or reject a batch based on sample defect rates. The tree uses pre‑defined thresholds derived from historical data.

Example: Automobile Parts Batch Inspection

• Each batch contains 1,000 parts; a random sample of 50 is inspected.
• Decision thresholds (from historical analysis):
  • Defect rate < 2 % → approve batch.
  • Defect rate 2–5 % → borderline; conduct further testing.
  • Defect rate > 5 % → reject batch.

For a sample of 50 parts:

• 0 or 1 defective → approve (defect rate ≤ 2 %).
• 2, 3, or 4 defective → further testing.
• 5 or more defects → reject (defect rate ≥ 10 %).

Setting Thresholds with Historical Data

Rather than arbitrary cutoffs, thresholds should be data‑driven. Two common approaches:

1. Percentile analysis – Examine the distribution of defect rates from past batches. Identify natural cutoffs where higher defect rates are associated with high rework costs or customer returns. For example, the 90th percentile might mark the point beyond which rejection is economically justified.

2. Cost‑benefit simulation – Define the costs of each decision:

   • Scrap cost (rejecting a good batch)
   • Inspection cost (further testing)
   • Failure cost (defective products reaching customers – returns, warranty claims, legal risks)

   Simulate different threshold pairs (e.g., lower=1 %, upper=3 %; lower=2 %, upper=5 %) and compute total expected cost. The optimal thresholds minimise total cost.

Exam tip: Threshold selection is a trade‑off between strictness (high rejection → scrap cost) and leniency (high defect rate shipped → failure cost). Always consider both sides in an exam problem.

Automation and Continuous Improvement

• Once thresholds are set, the decision tree can be automated – every new batch is sampled, classified, and routed to the appropriate action without manual intervention.
• Continuous improvement: Re‑evaluate thresholds periodically (monthly/quarterly) using updated defect patterns to adapt to process changes.

Key takeaways

• A decision tree maps decisions, chance events, probabilities, and payoffs into a clear visual framework.
• EMV is the core metric: $\text{EMV} = \sum (\text{probability} \times \text{payoff})$ – choose the branch with the highest EMV.
• In quality control, decision trees classify batches based on sample defect rates and predefined thresholds (e.g., approve if ≤2 defects/n).
• Thresholds should be derived from historical data using percentiles or cost‑benefit simulations, not guesswork.
• Decision trees enable automated quality assurance, reducing subjective judgment and improving product quality.

Statistical Process Control (SPC)

Statistical Process Control (SPC) is a data-driven methodology used to monitor and control processes so that they produce consistent, reliable output. The core insight: every process has inherent variation. SPC helps distinguish normal variation from abnormal variation — enabling early corrective action before defects occur.

Types of Process Variation

• Common cause variation – natural, expected fluctuations from minor, uncontrollable factors. Low magnitude, always present.
• Special cause variation – unusual, large disturbances from specific issues (equipment failure, human error, environmental change). Indicates a process out of control.

SPC provides tools to differentiate these two types and maintain process stability.

Key SPC Tools

Control charts are the most essential; also used: histograms, Pareto charts, process capability analysis. This section focuses on control charts (also called Shewhart charts, after Walter A. Shewhart, 1920s).

---

Control Charts: Structure & Purpose

A control chart is a graphical tool that tracks a process metric over time. Every control chart has three components:

A process is stable (in control) when all points fall within the control limits and show no unusual patterns. Any point outside the limits, or a systematic trend, signals a special cause that requires investigation.

Worked Example: Email Response Time

A company monitors daily average response time (minutes) for 15 days.

• Mean response time = 16 minutes → central line at 16.
• Initially, UCL = 22, LCL = 10 (chosen subjectively as ±6 minutes).

Points on day 13 touch the limit; day 15 exceeds the UCL. Also an increasing trend from day 1 to day 15. This suggests a special cause (e.g., increased queries, system downtime, staffing issues) → investigate and correct.

---

Setting Control Limits: The Empirical Rule

Subjective limits are not ideal. A rigorous, data-driven approach uses confidence intervals based on the normal distribution. The standard rule:

$$\text{UCL} = \mu + 3\sigma \qquad \text{LCL} = \mu - 3\sigma$$

where $\mu$ is the process mean and $\sigma$ the process standard deviation.

Why ±3σ? The Empirical Rule

For normally distributed data:

• $\approx 68\%$ of observations lie within $\mu \pm 1\sigma$
• $\approx 95\%$ within $\mu \pm 2\sigma$
• $\approx 99.7\%$ within $\mu \pm 3\sigma$

Thus, only 0.3% of points fall outside ±3σ under common-cause variation. A point beyond ±3σ is strong evidence of a special cause.

Re‑evaluating the Example

The response‑time data had $\mu = 16$, $\sigma = 3.1$. Using ±3σ:

$$\text{UCL} = 16 + 3(3.1) = 25.3 \quad \text{LCL} = 16 - 3(3.1) = 6.7$$

All 15 observations now fall within these wider limits. The process appears in control under the ±3σ rule (no points outside). This demonstrates how changing the limit rule changes the diagnosis.

Exam tip: The ±3σ rule is the most common control‑limit choice. Always check whether limits are given as subjective values or computed from the data.

---

Six Sigma: A Stricter Standard

Six Sigma is a methodology for reducing defects and variability. In SPC, it extends the control‑limit concept to ±6σ from the mean.

For a perfectly normal process:

• $99.99966\%$ of points lie within $\mu \pm 6\sigma$
• Only $0.00034\%$ fall outside → 3.4 defects per million opportunities

This extremely low defect rate is the Six Sigma target. If a point falls outside ±6σ, it almost certainly indicates a special cause.

Six Sigma is widely adopted in manufacturing, healthcare, and finance to maintain high quality through continuous monitoring and improvement.

---

Key Takeaways

• SPC distinguishes common cause (normal) from special cause (assignable) variation.
• Control charts (Shewhart charts) plot process data over time with a central line and two control limits (UCL, LCL).
• The standard data‑driven limit is $\mu \pm 3\sigma$, based on the empirical rule (99.7% coverage).
• Points outside ±3σ signal a special cause → investigate.
• Six Sigma uses $\mu \pm 6\sigma$ limits, aiming for only 3.4 defects per million.
• Control charts enable early detection of process shifts, trends, and outliers, supporting proactive quality management.

Statistical Process Control – Control Charts

Statistical process control (SPC) is a methodology for monitoring, controlling, and improving production processes. The most widely used SPC tool is the control chart – a visual, time‑ordered display of process variation that flags when a process is drifting outside acceptable limits and corrective action is needed.

Four common types are covered: X‑bar chart, R chart, P chart, and NP chart. Each is designed for a specific data type and monitoring purpose.

---

X‑bar and R Charts (Mean and Range Charts)

Intuition: When the process output is a continuous measurement (weight, length, temperature), we care about both the average level and the spread. The X‑bar chart tracks the mean; the R chart tracks the range (max – min) within each sample. Together they detect shifts in location (e.g., machine calibration drift) and changes in variability (e.g., tool wear).

Use case: Continuous variables – e.g., weight of tablets, length of metal rods, fuel efficiency, time to complete a job.

Two charts in one:

• X‑bar chart – monitors the process mean over time. Centre line = overall mean $\bar{\bar{x}}$. Upper and lower control limits (UCL, LCL) are typically set at $\pm 3\sigma_{\bar{x}}$ (or using factor $A_2$ times average range).
• R chart – monitors process variability. Centre line = average range $\bar{R}$. Limits use factors $D_3$ and $D_4$ (or equivalent).

Worked Example – Pharmaceutical Tablet Weight

A company produces tablets of target weight 500 mg. Every hour a sample of 5 tablets is weighed for 10 hours.

Computed overall mean: $\bar{\bar{x}} = 500.2$ mg
Average range: $\bar{R} = 0.91$ mg
Standard deviation of sample means: $\sigma_{\bar{x}} = 0.39$ mg

Using ±3σ limits (stated as “6‑sigma limits”):

• UCL = $500.2 + 0.5 = 500.7$ mg
• LCL = $500.2 - 0.5 = 499.7$ mg

All sample means fall inside these limits → process mean is stable.

Similarly, the R chart with centre line 0.91 mg shows all ranges within control limits → variability is stable.

Interpretation: If an X‑bar point exceeds UCL/LCL → possible calibration issue or ingredient mixing error. If an R point rises → increasing variability (e.g., dull cutting tool, inconsistent mixing).

Key takeaways – X‑bar and R charts

• Used for continuous data; monitor both mean (X‑bar) and spread (R).
• Centre line for X‑bar = $\bar{\bar{x}}$; for R = $\bar{R}$.
• Points outside control limits signal need for investigation.
• Common in manufacturing, pharma, food processing.

---

P Chart (Proportion Chart)

Intuition: When quality is measured as pass/fail (attribute data), we monitor the proportion of defective items. The P chart tracks this proportion over time, with control limits that adjust for changing sample sizes.

Use case: Categorical data – pass/fail, complaint/not complaint, defective/acceptable.

• Centre line = overall proportion $\bar{p}$.
• Upper control limit and lower control limit vary with sample size because the standard error of a proportion depends on $n$: $\text{Standard error} = \sqrt{\frac{\bar{p}(1-\bar{p})}{n}}$
• UCL and LCL are typically $\bar{p} \pm 3 \times \text{SE}$.

Example – Call Centre Complaints

A call centre records daily customer complaints. Data includes:

Because sample size (number of calls) changes each day, the control limits are not constant – they widen when $n$ is small and narrow when $n$ is large.

A point above the UCL (e.g., day 5 or 10) signals an unusually high complaint rate that warrants investigation – possible staffing shortage, training gap, or software issue. A point below the LCL (e.g., day 6) might indicate a truly good day or underreporting; both should be examined.

Exam tip: P‑chart limits are variable when sample sizes differ. Do not compare a point to a fixed limit – always check the local UCL/LCL for that sample.

Key takeaways – P chart

• Used for attribute data (pass/fail, defective/non‑defective).
• Monitors proportion of defects; limits vary if sample size varies.
• Points outside limits indicate special‑cause variation.
• Investigate both high and low outliers (low may be underreporting).

---

NP Chart (Number of Defectives Chart)

Intuition: When sample size is constant, it is simpler to plot the count of defective items rather than the proportion. The NP chart does exactly that – it tracks the raw number of defects.

Use case: Constant sample size across subgroups.

• Centre line = average number of defectives $n\bar{p}$.
• Control limits: $n\bar{p} \pm 3\sqrt{n\bar{p}(1-\bar{p})}$.

Example: A bakery produces 500 loaves daily and inspects each batch for defects (undercooked, burnt crust). The number of defective loaves is plotted each day. An increasing trend might indicate temperature control problems.

Key takeaways – NP chart

• Similar to P chart but tracks count of defectives.
• Requires constant sample size.
• Simpler to interpret when $n$ does not change.

---

Choosing the Right Control Chart

Summary Comparison

Final Key Takeaways – Control Charts

• All control charts detect special‑cause variation – deviations from normal process behaviour.
• X‑bar and R charts handle continuous measurements; P and NP charts handle pass/fail data.
• Process is in control when all points lie within the limits and no non‑random patterns exist.
• Early detection of out‑of‑control signals prevents costly defects and ensures customer satisfaction.

Design of Experiments (DOE)

Design of Experiments (DOE) is a structured, systematic approach to determine the relationship between factors (independent variables) and outcomes (dependent variables). Intuitively: instead of changing one thing at a time and hoping for improvement, DOE lets you test many variables simultaneously in a carefully planned way—saving time, money, and guesswork.

Why Use DOE?

Key Concepts

• Factor – an input variable under investigation (e.g., compression force, drying time).
• Level – a specific setting or category of a factor (e.g., low/medium/high for compression; short/long for drying time).
• Treatment – a unique combination of factor levels.
• Balanced design – all treatments are tested with equal sample sizes; gives each combination equal importance.
• Full factorial experiment – tests every possible combination of all factors.
  • Total treatments = product of levels of each factor:
    $\text{Number of treatments} = a \times b \times c \times \dots$
  • Example: 3 levels × 2 levels × 2 levels = 12 treatments.
• Randomized block design – total number of experimental runs is decided first, then each run is randomly assigned to one of the treatment buckets.

Worked Example 1: Tablet Manufacturing (Pharmaceutical)

A company must ensure tablets dissolve within the required time. Key factors affecting dissolution rate:

Full factorial design: 3 × 2 × 2 = 12 treatment combinations. Each treatment is tested (e.g., 10 replicates each) to estimate the effect of each factor and identify the best settings.

Possible outcome: High compression force + long drying time produce the most consistent dissolution rates.

Business impact:

• Optimized tablet production → fewer defects, better regulatory compliance.
• Reduced material waste and production time.
• Avoided quality failures that could lead to recalls or customer dissatisfaction.

Worked Example 2: Email Marketing (Retail)

A retail company wants to increase customer engagement through email campaigns. Factors:

DOE approach: send different combinations to separate customer subsets, tracking open rates, click-through rates, and conversions.

Possible outcome: short subject line + percentage discount + image-heavy layout yields highest engagement.

Business impact:

• Data-driven campaign optimization (instead of intuition).
• Higher ROI on marketing efforts.
• Improved customer retention through more relevant, engaging content.

Exam tip: Remember the difference between full factorial (all combinations, exhaustive) and randomized block (random assignment to a subset of combinations). Full factorial is thorough but can be expensive; randomized block is more efficient when resources are limited.

Key Takeaways

• DOE is a structured method to test multiple factors simultaneously, replacing inefficient trial-and-error.
• Factors (inputs) and levels (settings) define the experimental space; each unique combination is a treatment.
• Full factorial design tests every combination; the number of treatments equals the product of the factor levels.
• Balanced designs give equal sample size to each treatment, improving statistical reliability.
• DOE is used across manufacturing (process optimization) and services (marketing campaigns) to reduce cost, improve quality, and increase customer satisfaction.