7. Quality Engineering, Experimentation and Robust Design

Introduction and Context

Total Quality Management is not limited to inspecting finished products or correcting defects after they occur. At an advanced level, quality is designed into products, processes and operating systems. Quality engineering provides the analytical methods needed to understand how inputs influence outputs, reduce unwanted variation and achieve performance that remains stable under changing conditions. Design of experiments (DOE), response analysis, Taguchi methods, parameter design, tolerance design and the quality loss function form an integrated body of knowledge for this purpose.

Manufacturing and service processes are affected by many variables: material properties, machine settings, environmental conditions, operator practices, supplier variation and usage conditions. Studying one factor at a time often fails because factors may interact. A process may perform well under controlled laboratory conditions but fail when temperature, vibration, load or customer behaviour changes. Robust design addresses this problem by seeking settings that deliver consistent performance despite noise factors.

The relevance of these methods extends beyond manufacturing. They are used in chemical processing, pharmaceutical development, product engineering, logistics, healthcare, software performance and service operations. For production managers, the central challenge is to make informed trade-offs among quality, cost, speed, flexibility and manufacturability. Statistical experimentation supports these decisions with evidence rather than intuition.

Learning Objectives

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

1. Explain the role of quality engineering in Total Quality Management.

2. Distinguish response, control, noise and signal factors in an experiment.

3. Design and interpret factorial, fractional factorial and response-surface experiments.

4. Apply Taguchi concepts of robust design, parameter design and tolerance design.

5. Use signal-to-noise ratios and quality loss functions appropriately.

6. Evaluate interactions, statistical significance, practical significance and confirmation results.

7. Make balanced decisions involving quality, cost, speed and manufacturability.

8. Recognize limitations, ethical issues and implementation risks in experimentation.

Quality Engineering as a Design Discipline

Quality engineering may be understood as the systematic use of engineering, statistical and managerial methods to improve product and process performance. Its emphasis is preventive: rather than relying primarily on inspection, it identifies sources of variation and designs conditions under which acceptable performance is naturally produced.

A useful distinction is between common-cause variation and special-cause variation. Common causes are inherent in the existing system, such as normal material variation or machine capability. Special causes arise from unusual events, such as a damaged tool, incorrect setup or contaminated batch. Statistical process control is particularly useful for distinguishing these forms of variation. DOE goes further by deliberately changing selected inputs to learn how they affect an output.

The output being studied is called the response. Inputs deliberately changed by the investigator are control factors or experimental factors. Conditions that cannot be economically or practically controlled during normal operation are noise factors. A factor may be controllable during an experiment but noisy in production. For example, curing temperature may be controlled in a development laboratory but fluctuate in a production oven. The objective of robust design is to make the response insensitive to important noise factors.

Quality engineering is closely connected with concurrent engineering. Product designers, process engineers, quality specialists, procurement personnel and production managers should collaborate because design decisions determine much of the eventual cost and variation. A dimension that is unnecessarily tight may increase machining and inspection costs, whereas a loose tolerance may cause poor assembly or field failure. Optimization therefore requires a systems perspective.

The Logic of Designed Experiments

An experiment is a planned and structured investigation in which one or more factors are varied so that their effects on a response can be estimated. Good experimentation differs from informal trial and error in four respects: the objective is defined in advance, factors and levels are selected deliberately, observations are collected according to a design, and conclusions are supported by analysis.

A typical DOE begins with a practical problem. The team then defines the response measure, identifies potentially important factors, selects feasible levels, chooses an experimental design, conducts randomized trials, analyzes the results, confirms the proposed settings and establishes control conditions for routine operation.

Responses and measurement systems

The response must represent the customer or process requirement. Common responses include strength, cycle time, defect rate, energy consumption, dimensional error and customer waiting time. A response may be:

• Nominal-the-best, where a target value is desired, such as a shaft diameter.
• Smaller-the-better, where reduction is desirable, such as defects, leakage or contamination.
• Larger-the-better, where increase is desirable, such as strength or battery life.

Before experimentation, the measurement system must be capable. If gauge error is large relative to the expected factor effects, the experiment may identify measurement noise rather than process behaviour. Measurement-system analysis, calibration and operational definitions are therefore essential. A statistically sophisticated design cannot compensate for an invalid response measure.

Factors, levels and interactions

A factor is an input that may influence the response. Its levels are the settings or categories tested. For example, cutting speed may be tested at 80 and 120 metres per minute, while coolant type may have three categories.

The effect of a factor is not always constant. An interaction exists when the effect of one factor depends on the level of another. Suppose higher temperature improves curing when catalyst concentration is low but damages the product when catalyst concentration is high. The temperature effect depends on catalyst concentration; examining either factor alone would be misleading. Interactions are especially important in production because attempts to optimize one parameter independently may worsen performance elsewhere.

Replication, randomization and blocking

Replication means conducting repeated experimental runs under the same treatment combination. It provides an estimate of experimental error and improves precision. Replication is different from repeated measurements of one specimen; the latter may reveal measurement repeatability but not process-to-process variation.

Randomization assigns runs or treatments in a random order. It protects the experiment against time trends, operator learning, warm-up effects and other uncontrolled influences. When randomization is impractical, the limitation should be recognized and the run order managed carefully.

Blocking groups experimental runs according to a nuisance source, such as shift, raw-material lot, machine or day. Differences among blocks are separated from treatment effects. Blocking is valuable when a nuisance factor is known to influence the response but is not itself the principal subject of study.

Major Experimental Designs

Full factorial designs

In a two-level full factorial design, all combinations of factor levels are tested. With k two-level factors, the number of treatment combinations is 2^k. A full factorial design estimates main effects and interactions represented by the design. For instance, a 2^3 design has eight combinations for three factors.

The major strength of a full factorial design is information. It shows not merely whether factors matter, but how they work together. Its weakness is that the number of runs increases rapidly as factors are added. At the screening stage, many factors may be considered, making a full design expensive.

Fractional factorial designs

A fractional factorial design uses a carefully selected fraction of the full set of combinations. It reduces the number of runs but introduces aliasing, meaning that certain effects are confounded with one another. A main effect may be aliased with an interaction, or two interactions may be aliased. The design must therefore be selected with a clear understanding of which effects are likely to be important.

Fractional factorial designs are particularly useful for screening a large number of factors. They should not be treated as automatically superior to full designs. If strong interactions are expected, excessive fractionation may produce ambiguous conclusions. Resolution, defining relations and the practical knowledge of the process should guide the choice.

Response surface methodology

Once important factors have been screened, the objective often shifts from identifying factors to finding the best operating region. Response surface methodology (RSM) models the response as a function of quantitative factors, commonly using a first- or second-order polynomial. Central composite and Box–Behnken designs are widely used for fitting quadratic response surfaces.

A second-order model may be written as:

Y = β0 + ΣβiXi + ΣβiiXi² + ΣβijXiXj + ε

where Y is the response, Xi are coded factors, β terms are estimated coefficients and ε represents unexplained error. The squared terms allow the model to represent curvature, while cross-product terms represent interactions.

RSM can identify a stationary point, but statistical optimization is not automatically operational optimization. A mathematically ideal setting may be difficult to maintain, unsafe, expensive or incompatible with equipment constraints. The recommended region should therefore be evaluated using confirmation trials and practical engineering judgment.

Mixture and constrained experiments

When the inputs are proportions that must sum to a fixed total, ordinary factorial designs are inappropriate. Mixture designs are used for formulations such as chemical blends, food recipes and pharmaceutical compositions. Other experiments may involve constraints, such as a total cycle time, limited energy or a required production rate. Constrained optimization makes the feasible operating region explicit rather than recommending impossible settings.

Taguchi’s Approach to Robust Design

Genichi Taguchi made a major contribution by framing quality as performance on target with minimum sensitivity to noise. His approach emphasizes designing products and processes so that variation is reduced before production begins.

Taguchi separates factors into three broad groups. Signal factors are inputs selected by the user or operator to obtain a desired output. Control factors are design or process variables chosen by the engineer. Noise factors cause variation and may be environmental, manufacturing-related or usage-related. A robust design uses control factors to reduce the response’s sensitivity to noise factors.

Parameter design

Parameter design seeks appropriate levels of control factors. It generally uses an inner array for control factors and an outer array for noise factors. Each control-factor combination is evaluated across selected noise conditions. The analysis identifies settings that provide a desirable average response and low variation.

This approach is valuable when noise factors cannot be eliminated economically. For example, a sealing process may need to perform across different material lots and ambient temperatures. Rather than attempting to hold every condition constant, the experiment deliberately exposes the process to representative noise conditions and selects settings that remain stable.

Taguchi orthogonal arrays provide balanced arrangements of factor levels and can reduce the number of experimental runs. They are efficient, but they do not eliminate the need for engineering interpretation. Important interactions must be represented; otherwise, a sparse design may miss the mechanism causing variation.

Signal-to-noise ratios

The Taguchi signal-to-noise ratio combines location and variation into a measure that can be optimized. Common forms include:

For smaller-the-better characteristics:

S/N = −10 log10[(1/n)Σyi²]

For larger-the-better characteristics:

S/N = −10 log10[(1/n)Σ(1/yi²)]

For nominal-the-best characteristics, a form based on the squared mean relative to variance is commonly used. The exact interpretation depends on the objective and measurement scale. A higher S/N ratio generally indicates a more desirable combination of performance and stability for the selected characteristic.

S/N analysis should not replace examination of the actual data. Two settings may have similar S/N ratios but very different means, and an improvement in variation may be accompanied by an unacceptable shift from the target. Managers should inspect means, standard deviations, confidence intervals, engineering limits and economic consequences together.

Tolerance Design and the Quality Loss Function

Parameter design chooses process or product settings; tolerance design determines the permissible range around nominal values. Tightening a tolerance usually reduces variation but increases manufacturing, inspection and sometimes material costs. The decision should be based on total value, not on the assumption that tighter is always better.

Taguchi’s quality loss function challenges the traditional view that all units within specification are equally good. For a nominal-the-best characteristic, the quadratic loss function is commonly represented as:

L(y) = k(y − T)²

where y is the observed value, T is the target and k converts deviation into monetary or consequence-based loss. The function implies that loss increases continuously as performance moves away from target, even before the specification limit is crossed.

For a smaller-the-better characteristic, loss may be represented as proportional to y²; for a larger-the-better characteristic, an inverse relationship may be appropriate. The model is not a universal law of economics. It is a decision framework requiring a credible estimate of k, based on rework, warranty, energy, customer dissatisfaction, safety or other consequences.

The loss-function perspective supports a better discussion between design and production. A product close to the specification limit may pass inspection but have a higher expected field cost than one centred on target. Conversely, an extremely narrow tolerance may impose production costs that exceed the reduction in customer loss. The appropriate choice balances both sides.

Statistical and Engineering Interpretation

DOE analysis commonly uses analysis of variance (ANOVA), regression models, effect plots and residual analysis. ANOVA partitions observed variation into contributions associated with factors, interactions and error. A p-value can indicate whether an effect is statistically distinguishable from random experimental error under the model assumptions. It does not, by itself, establish practical importance.

A small effect may be statistically significant with many observations but irrelevant to production. A large effect may fail to reach conventional significance in a small experiment because the study has low power. Confidence intervals, effect size, process capability, cost and engineering consequences should therefore accompany significance testing.

Residuals are examined to assess model adequacy. Patterns in residuals may indicate nonlinearity, unequal variance, dependence, outliers or omitted variables. Transformations may sometimes improve the model, but transformation should not be used merely to conceal inconvenient results. The experiment should be repeated or redesigned when assumptions are seriously violated.

After an optimum is identified, a confirmation experiment is essential. The proposed settings are run independently, preferably using normal production materials, equipment and operators. The observed result is compared with the predicted response and the required target. A successful confirmation supports implementation; a failure may indicate model error, scale-up effects, uncontrolled noise or an unrealistic experimental region.

An Integrated Improvement Process

A practical quality-engineering project can follow this sequence:

1. Translate the customer or business need into measurable critical-to-quality responses.

2. Map the process and establish a credible cause-and-effect structure.

3. Verify the measurement system and define operating procedures.

4. Identify control, noise and nuisance factors through technical knowledge and prior data.

5. Select a design that reflects the objective: screening, characterization, optimization or robustness.

6. Plan replication, randomization, blocking, safety and data recording.

7. Execute the experiment under controlled and documented conditions.

8. Analyze main effects, interactions, variation, model adequacy and practical consequences.

9. Conduct confirmation trials at the proposed settings.

10. Standardize the process, update control plans and monitor performance after implementation.

This sequence should not be treated as a rigid statistical ritual. The experimental design must evolve as knowledge improves. Screening may reveal that a factor is unimportant, while a previously ignored interaction may require a second experiment. Learning is iterative.

Illustrative Case: Robust Curing of an Industrial Adhesive

Background

The following is an illustrative case. A manufacturer of electrical assemblies uses an adhesive to bond a protective cover to a housing. Field complaints indicate occasional bond separation. Laboratory testing suggests that bond strength is affected by curing temperature, cure time, adhesive mix ratio and surface preparation. Production also experiences variation in material lot, ambient humidity and operator application pressure.

Situation and problem

The process currently uses a nominal temperature and time selected through historical trial and error. Average bond strength meets the requirement, but the lower tail of the distribution is unsatisfactory. Increasing temperature appears to improve strength in some trials but causes brittleness in others. Management wants improvement without adding a lengthy inspection operation or substantially increasing cycle time.

Analysis and experimental strategy

The response is defined as bond strength, with a secondary response measuring failure mode because cohesive failure is preferred to adhesive separation. A measurement-system study confirms that the test method is sufficiently repeatable. Control factors are curing temperature, cure time, mix ratio and surface-preparation method. Noise factors are adhesive lot, humidity and application pressure.

A two-level screening experiment is first selected to identify the important control factors and interactions. Because material lot and production day may affect results, runs are randomized within blocks. The team anticipates that temperature and mix ratio may interact, so the selected design must preserve the ability to estimate that interaction.

The analysis shows that surface preparation has the largest effect on average strength, while temperature and mix ratio interact strongly. A high temperature is beneficial at one mix ratio but harmful at another. Cure time has a modest effect within the tested region. Residual analysis indicates that a simple linear model is adequate for screening but that curvature may exist around the promising region.

A second response-surface experiment is conducted with temperature and mix ratio as quantitative factors. Noise conditions are deliberately varied across material lots and humidity conditions. The selected region produces a slightly lower maximum strength than the original high-temperature setting but substantially reduces variation and eliminates the brittle failure mode.

Approach and implementation

The proposed settings are tested in a confirmation run using routine operators, normal equipment and a new adhesive lot. The process is then updated: surface preparation is standardized, mix-ratio controls are improved, the curing window is centred in the robust region and humidity is recorded as a process condition rather than ignored. A control chart is introduced for bond strength, while periodic destructive testing is retained as a verification activity.

Outcome and learning points

The case demonstrates that the strongest average response is not necessarily the best operating point. The robust setting provides more consistent performance under realistic noise conditions. It also shows why interaction analysis matters: changing temperature without considering mix ratio could have produced an apparently contradictory result. Finally, the improvement depends on implementation discipline. A statistically identified setting has little value if surface preparation, measurement and material identification remain uncontrolled.

Practical and Managerial Applications

Quality engineering supports decisions at several levels. At the product-design level, it helps select materials, dimensions and architectures that are tolerant of manufacturing and usage variation. At the process-design level, it identifies stable operating windows and reduces dependence on inspection. At the plant level, it supports decisions about equipment capability, preventive maintenance, supplier specifications and automation.

For production managers, the most important managerial implication is that optimization is multi-objective. A setting that minimizes defects may increase energy consumption; a shorter cycle may increase variation; a tighter tolerance may reduce warranty cost but increase machining cost. Multi-response optimization can be used to identify feasible compromises, but the final decision requires explicit priorities and stakeholder agreement.

Economic analysis should include total cost: experimentation, implementation, scrap, rework, inspection, energy, downtime, warranty and customer impact. A low-cost experiment that produces an unmanufacturable recommendation is not a successful project. Similarly, a modest process change that reduces chronic variation may generate greater value than a dramatic but fragile improvement.

Challenges, Limitations and Common Mistakes

Several failures recur in industrial experimentation. The first is choosing a response that is convenient to measure but weakly connected to customer value. The second is changing too many conditions at once without a design, making causation impossible to establish. The third is ignoring interactions because the team expects each factor to act independently.

Other common problems include inadequate replication, non-random run order, uncontrolled measurement drift, excessive reliance on p-values, fitting a model outside the experimental region and stopping after the first apparently successful trial. Confounding in fractional designs may be misunderstood, leading to overconfident conclusions. Taguchi arrays may also be applied mechanically without representing critical interactions or verifying the predicted settings.

Noise factors require special care. Deliberately introducing a noise condition may create safety, regulatory or equipment risks. The team must define safe operating boundaries and distinguish realistic variation from abnormal misuse. Robustness does not mean making a process immune to every conceivable condition; it means reducing sensitivity to relevant and foreseeable variation.

Governance, Ethics and Professional Responsibility

Experimentation affects people, assets and customers. Production experiments should have documented authority, risk assessment and clear stop criteria. In regulated industries, changes may require validation, traceability and approval before release. If experiments involve human subjects, patients, employees or customer service outcomes, applicable ethical and privacy requirements must be followed.

Data integrity is a professional obligation. Runs should not be deleted because they produce inconvenient results, and exclusions must be justified before or during analysis. Analysts should disclose deviations from the design, failed measurements and uncertainty. A statistically significant improvement that compromises worker safety, environmental compliance or product reliability is not a quality improvement.

Digital Tools, Automation and Emerging Developments

Modern manufacturing execution systems, laboratory information systems and industrial Internet of Things platforms can provide synchronized data on machine settings, environmental conditions, material lots and responses. Automated data capture reduces transcription error and makes experimentation more scalable. Digital twins and simulation can narrow the experimental region, although simulation results must be validated against physical behaviour.

Machine learning can assist in identifying nonlinear relationships and prioritizing experiments, especially when historical data are extensive. However, predictive accuracy does not automatically establish causal effects. A model may learn correlations created by a particular supplier, machine or time period. Designed experiments remain valuable because deliberate intervention supports causal inference.

Emerging approaches include sequential experimentation, Bayesian optimization, active learning and adaptive process control. These methods can reduce the number of physical trials by selecting the next experiment using information from earlier runs. Their effectiveness depends on sound measurement, well-defined constraints and human oversight. Explainability, cybersecurity, data ownership and model drift also require governance.

Best Practices for Robust Quality Engineering

Effective practice rests on a few consistent principles:

• Begin with a customer-relevant response and a clear decision problem.
• Involve process experts, operators, maintenance personnel and statisticians early.
• Verify the measurement system before interpreting process effects.
• Use engineering knowledge to select meaningful factor ranges and realistic noise conditions.
• Plan for interactions rather than assuming independence.
• Randomize and block where appropriate, and document deviations.
• Separate statistical significance from practical and economic significance.
• Use confirmation trials under normal operating conditions.
• Convert experimental findings into standard work, control plans and training.
• Monitor the process after implementation and revisit assumptions when conditions change.

Chapter Summary

Quality engineering applies statistical and engineering reasoning to design products and processes that deliver target performance with low variation. DOE provides a structured way to investigate factors, interactions and operating regions. Full factorial designs offer rich information, fractional factorial designs support efficient screening, and response surface methods help optimize quantitative factors. Replication, randomization, blocking and measurement-system capability protect the validity of conclusions.

Taguchi’s robust-design approach separates control and noise factors and seeks settings that are insensitive to realistic disturbances. Parameter design selects robust factor levels, while tolerance design balances variation reduction against manufacturing cost. The quality loss function emphasizes that departure from target can create loss even within specification. Sound practice combines statistical analysis with engineering judgment, confirmation trials, economic reasoning, ethical governance and continuing process control.

Key Terms

Quality engineering: The systematic use of engineering and statistical methods to design and improve quality.

Design of experiments: A planned method for changing factors and estimating their effects on responses.

Response: The measured output or performance characteristic studied in an experiment.

Control factor: An input deliberately selected or adjusted by the investigator.

Noise factor: A source of variation that is difficult or uneconomic to control in routine operation.

Signal factor: An input selected to obtain a desired output level.

Interaction: A condition in which the effect of one factor depends on the level of another.

Replication: Repetition of experimental treatment combinations to estimate experimental error.

Randomization: Random assignment or ordering of experimental runs to reduce bias.

Blocking: Grouping runs to separate the influence of a known nuisance source.

Full factorial design: A design containing every combination of selected factor levels.

Fractional factorial design: A strategically selected portion of a full factorial design.

Aliasing: Confounding of effects caused by using a reduced experimental design.

Response surface methodology: Methods for modelling and optimizing responses using quantitative factors.

Parameter design: Selection of control-factor settings to improve performance and robustness.

Tolerance design: Determination of acceptable ranges around nominal product or process values.

Signal-to-noise ratio: A Taguchi measure combining response level and variation.

Quality loss function: A model relating departure from target to economic or consequential loss.

Confirmation experiment: An independent trial used to verify an experimental recommendation.

Robust design: Design that maintains desired performance despite relevant noise factors.

Review Questions

Short Answer Questions

1. Distinguish between a control factor and a noise factor, using one production example.

2. Why are replication and randomization important in DOE?

3. What is meant by an interaction between two factors?

4. Differentiate parameter design from tolerance design.

5. What does the quality loss function imply about products within specification limits?

Long Answer Questions

1. Explain the logic and major stages of a quality-engineering experiment from problem definition through confirmation.

2. Compare full factorial, fractional factorial and response-surface designs, including their appropriate uses.

3. Explain Taguchi robust design, including control factors, noise factors, inner and outer arrays, and signal-to-noise ratios.

4. Discuss how statistical significance, practical significance and economic significance should be integrated in process decisions.

5. Explain how tolerance decisions affect quality, cost, manufacturability and customer value.

Analytical / Application Questions

1. A process has four potentially important factors, but only twelve experimental runs are affordable. Propose an initial experimental strategy and explain how you would manage aliasing, interactions and confirmation.

2. A curing process has high average strength but large variation across humidity conditions. Design a robust-design study identifying the response, control factors, noise factors and implementation measures.

3. A manager proposes tightening a product tolerance because it will reduce deviation from target. Analyze the decision using the quality loss function and identify the additional cost and process information required before approval.

22. Multiple Choice Questions with Answers

Question 1

A process engineer deliberately varies temperature, pressure and feed rate while measuring surface roughness. Surface roughness is the:

A. Noise factor

B. Response

C. Blocking variable

D. Signal factor

Correct Answer: B. Response

Question 2

Which feature most directly enables a designed experiment to reveal that the effect of temperature depends on catalyst concentration?

A. Replication only

B. Measurement calibration only

C. Inclusion of the relevant interaction structure

D. Use of specification limits

Correct Answer: C. Inclusion of the relevant interaction structure

Question 3

A fractional factorial design is selected primarily because the team wants to:

A. Estimate every possible high-order interaction without ambiguity

B. Screen many factors with fewer runs

C. Avoid the need for randomization

D. Replace confirmation experiments

Correct Answer: B. Screen many factors with fewer runs

Question 4

In robust design, a noise factor is best described as a factor that:

A. Is always statistically insignificant

B. Cannot be measured reliably

C. Causes relevant variation but is difficult or uneconomic to control routinely

D. Must be removed from the experiment

Correct Answer: C. Causes relevant variation but is difficult or uneconomic to control routinely

Question 5

A higher Taguchi signal-to-noise ratio generally indicates:

A. A more desirable combination of performance and stability for the selected objective

B. A larger sample size regardless of response quality

C. That the process mean is always exactly on target

D. That all interactions are absent

Correct Answer: A. A more desirable combination of performance and stability for the selected objective

Question 6

A confirmation experiment should normally be conducted:

A. Only with ideal laboratory materials

B. After selecting proposed settings and preferably under normal operating conditions

C. Before identifying any important factors

D. Instead of measuring the response

Correct Answer: B. After selecting proposed settings and preferably under normal operating conditions

Question 7

According to the nominal-the-best quality loss function, loss is minimized when the characteristic:

A. Is anywhere inside the specification limits

B. Is at the target value

C. Is as close as possible to the upper specification limit

D. Has the largest permissible tolerance

Correct Answer: B. Is at the target value

Question 8

A statistically significant factor has a very small effect, but changing it would require expensive equipment modification. The best managerial response is to:

A. Implement it automatically because significance proves economic value

B. Ignore all statistical evidence

C. Compare effect size and total economic consequences before deciding

D. Remove the factor from the model

Correct Answer: C. Compare effect size and total economic consequences before deciding

Question 9

Why is measurement-system evaluation needed before interpreting a DOE?

A. It guarantees that all factors will be significant

B. It determines whether observed differences can be distinguished from measurement error

C. It eliminates the need for replication

D. It converts a fractional design into a full design

Correct Answer: B. It determines whether observed differences can be distinguished from measurement error

Question 10

In the adhesive case, changing temperature without considering mix ratio could lead to a wrong recommendation because temperature and mix ratio:

A. Are both responses

B. Form an interaction

C. Are both blocking variables

D. Cannot be measured simultaneously

Correct Answer: B. Form an interaction