4.4.1 2k Full Factorial Designs

2k Full Factorial Designs Introduction to 2ᵏ Full Factorial Designs A 2ᵏ full factorial design is an experimental design where: - Each factor has two levels (typically low and high). - All possible combinations of factor levels are run. - The total number of runs is 2ᵏ, where k is the number of factors. This structure makes 2ᵏ designs powerful, efficient, and mathematically simple, which is why they are a core tool for structured experimentation in process improvement and optimization. --- Basic Structure and Notation Factors, Levels, and Runs In a 2ᵏ design: - Factors: Input variables of interest (e.g., temperature, pressure). - Levels: Two settings for each factor, often coded as: - −1 (or “−”) for the low level - +1 (or “+”) for the high level - Runs: Each unique combination of levels across all factors. For example: - 2¹ design (one factor, two levels): 2 runs. - 2² design (two factors, two levels each): 4 runs. - 2³ design (three factors, two levels each): 8 runs. Coded vs Natural Units Working in coded units (−1, +1) simplifies calculations and interpretation: - Natural units: Actual settings (e.g., 150 °C and 200 °C). - Coded units: −1 and +1 mapped to these natural settings. Coding supports: - Simple algebra for effects and interactions. - Easier scaling of regression coefficients. - Direct comparison of factor impact magnitudes. --- Design Matrices and Treatment Combinations Standard Order and Contrast Coding For a 2ᵏ design, the design matrix lists: - Each run as a row. - Columns for: - Main factors (A, B, C, ...). - Interactions (AB, AC, BC, ABC, ...). Example: 2² design (factors A and B): - Levels: - A: −, + - B: −, + - Design matrix (coded): - Run 1: A = −, B = − - Run 2: A = +, B = − - Run 3: A = −, B = + - Run 4: A = +, B = + The interaction column AB is formed by multiplying the signs: - AB = (sign of A) × (sign of B) - Example: - Run 1: A(−) × B(−) = (+) - Run 2: A(+) × B(−) = (−) - Run 3: A(−) × B(+) = (−) - Run 4: A(+) × B(+) = (+) This principle extends to more factors and interactions. Extension to 2³ and Higher For a 2³ design (factors A, B, C): - 8 runs cover all combinations of − and +. - Interaction columns: - AB = A × B - AC = A × C - BC = B × C - ABC = A × B × C In general: - Number of main effects = k. - Number of 2-factor interactions = C(k, 2). - Total effects including all interactions = 2ᵏ − 1. --- Main Effects and Interaction Effects Main Effects A main effect is the change in response when a factor moves from low to high, averaged over all levels of other factors. For factor A in coded units: - Effect(A) = (Average response at A = +1) − (Average response at A = −1). Key ideas: - A large main effect means factor A strongly influences the response. - The sign indicates direction: - Positive effect: Response increases at high level of A. - Negative effect: Response decreases at high level of A. Interaction Effects An interaction effect occurs when the impact of one factor depends on the level of another factor. For a 2² design with factors A and B: - Effect(AB) measures how the effect of A changes between B = −1 and B = +1 (and vice versa). Conceptually: - If there is no interaction: - The lines on an interaction plot (response vs A level, separate lines for B low/high) are roughly parallel. - With a significant interaction: - Lines are non-parallel, possibly crossing. Important points: - Interactions can be: - 2-factor (AB, AC, BC, ...). - 3-factor or higher (ABC, ABCD, etc.), though higher-order interactions are often small in real systems. - When strong interactions are present: - Interpreting main effects alone can be misleading. - Decisions should focus on factor combinations, not single factors in isolation. --- Calculation of Effects and Contrast Coefficients Contrast and Effect Formulas In 2ᵏ designs, effects are computed using contrasts: - A contrast is a weighted sum of responses, where each run is multiplied by +1 or −1 according to the effect being estimated. For a design with N = 2ᵏ runs: - Effect for a factor or interaction: - Effect = (Sum of signed responses) / (N/2) Where: - Signed responses are obtained by multiplying each run’s response by the sign in that factor or interaction column. - For main effect A: - Multiply all responses where A = +1 by +1. - Multiply all responses where A = −1 by −1. - Sum these and divide by N/2. This method applies to any effect column: B, C, AB, ABC, etc. Example: Effect Interpretation For a 2² design: - N = 4. - Effect(A) = [ (y₂ + y₄) − (y₁ + y₃) ] / 2 - Where y₁...y₄ are responses from runs in standard order. - Similar formulas exist for: - Effect(B) - Effect(AB) The contrast (numerator) shows the direction and magnitude; dividing by N/2 scales to a per-level-change effect. --- Statistical Modeling: Regression and ANOVA Linear Model Structure A 2ᵏ full factorial design supports a linear model in coded units: - Y = β₀ + β₁A + β₂B + β₃C + β₁₂AB + β₁₃AC + β₂₃BC + ... + error Where: - Y is the response. - β₀ is the overall mean. - β terms are coefficients for main effects and interactions. - A, B, C, AB, AC, etc. are coded columns (−1, +1). Key points: - In orthogonal 2ᵏ designs (no missing runs, balanced structure): - Estimates of each β are independent of others. - Main and interaction effects can be interpreted separately. ANOVA and Significance Testing Analysis of variance (ANOVA) is used to test which effects are statistically significant. Elements: - Sum of Squares (SS): - Quantifies variability explained by each effect. - Mean Square (MS): - MS = SS / degrees of freedom (df). - F-ratio: - F = MS(effect) / MS(error). For a 2ᵏ design with replicates: - Error variance is estimated from: - Replication within runs (pure error). - Or residuals from the fitted model. Interpretation: - Larger F and smaller p-values suggest an effect is statistically significant. - Non-significant higher-order interactions can be removed to simplify the model. --- Design Resolution, Orthogonality, and Confounding Orthogonality 2ᵏ full factorial designs are orthogonal when: - The sum of the product of any two columns is zero. - This ensures: - Estimates of each effect are uncorrelated with other effects. - Each effect has its own degrees of freedom and SS. Orthogonality is a key strength of full factorial designs. Confounding and Aliasing In a complete 2ᵏ full factorial design: - There is no confounding among main effects and interactions. - Each effect has a unique estimate. Confounding becomes important when: - Runs are blocked. - Designs are fractionated (not all combinations run). In those cases: - Some effects may be aliased (cannot be distinguished from each other). - For a pure full factorial without blocking or fractions: - Main and interaction effects are unaliased. Blocking in 2ᵏ Designs Blocking is used to account for nuisance variability (e.g., day-to-day differences). In 2ᵏ designs: - Blocks are often constructed using block generators (like fractional designs). - This can introduce confounding between: - Block effects and certain interactions. Key points for blocking: - Choose which interactions are acceptable to sacrifice to blocks. - Typically confound blocks with higher-order interactions assumed to be negligible. --- Graphical Tools for 2ᵏ Designs Main Effects Plots A main effects plot shows: - X-axis: Factor level (−1, +1). - Y-axis: Mean response at each level. Interpretation: - Flat line: Little or no main effect. - Steep line: Strong main effect. - Direction: - Upward: Higher level increases response. - Downward: Higher level decreases response. Interaction Plots An interaction plot shows: - X-axis: Levels of one factor. - Lines: Levels of a second factor. - Y-axis: Mean response. Interpretation: - Approximately parallel lines: - Little or no interaction. - Non-parallel or crossing lines: - Evidence of interaction. Graphical examination helps: - Quickly identify dominant effects. - Check for non-additive behavior. --- Replication, Randomization, and Center Points Replication Replication means repeating runs at the same factor settings. Purposes: - Estimate pure error (natural process variation). - Increase precision of effect estimates. - Enhance ability to test significance using ANOVA. In 2ᵏ designs: - Replicates can be: - Complete (repeat all runs). - Partial (repeat selected runs). Randomization Randomization is the random order of experimental runs. Benefits: - Reduces risk of systematic bias from time-related or sequence-related effects. - Distributes nuisance variation randomly across runs. Good practice: - Randomize run order within practical constraints (e.g., warm-up, cleaning). Center Points For 2ᵏ designs with quantitative factors, center points are runs at the mid-level of each factor. Uses: - Detect curvature: - Compare mean at center to average of factorial points. - Improve prediction in the central region. - Assess whether a linear model is adequate. Interpretation: - If center point mean differs significantly from the predicted mean under linearity: - There is evidence of curvature. - Additional design augmentation (e.g., more levels) may be needed. --- Curvature and Model Adequacy Detecting Curvature Curvature arises when the true relationship between factors and response is not well described by a purely linear model. In a 2ᵏ design: - With center points included: - Compute the difference between: - Mean response at center points. - Average response at factorial corner points. - A significant difference suggests: - Nonlinearity in at least one factor. Response Surface Considerations If curvature is detected: - The 2ᵏ design provides: - Good initial understanding of main trends and interactions. - A foundation for selecting next experiments. - Additional design runs (beyond the scope of a pure 2ᵏ design) can then refine: - Optimal settings. - Quadratic effects. The key is recognizing when linear 2-level modeling is insufficient. --- Practical Use and Interpretation Planning a 2ᵏ Full Factorial Experiment Key planning questions: - Which factors: - Select variables believed to substantially affect the response. - Ranges for levels: - Choose low and high levels that are: - Practically feasible. - Safe and within operational limits. - Wide enough to reveal meaningful changes. - Number of replicates: - Balance resource constraints with the need for error estimation. - Use of center points: - Include if curvature is plausible. Analyzing Results Typical analysis sequence: - Estimate main and interaction effects. - Generate main effect and interaction plots. - Fit a linear model: - Include significant effects based on statistics and subject-matter logic. - Perform ANOVA: - Assess significance and overall model fit. - Review residual plots: - Check assumptions of normality, constant variance, and independence. - Use the model to: - Predict responses. - Identify promising factor settings. Using Results for Factor Setting Decisions Based on significant effects: - Identify factors that: - Strongly improve or degrade the response. - Consider interactions: - Optimal settings of one factor may depend on the level of another. - Confirm practical recommendations: - Check feasibility, cost, and safety of proposed settings. - If needed: - Conduct confirmation runs at chosen settings to verify improvements. --- Limitations of 2ᵏ Full Factorial Designs While powerful, 2ᵏ designs have some limitations: - Number of runs grows exponentially: - Higher k leads to large experiments (e.g., 2⁶ = 64 runs). - Two levels may miss nonlinear effects: - Without center points or additional levels, quadratic behavior is not directly estimated. - Not ideal for many factors with limited resources: - Fractional designs may be more practical when many factors must be screened. Understanding these limitations helps decide when a 2ᵏ full factorial is appropriate and when other designs are preferable. --- Summary 2ᵏ full factorial designs provide a structured way to: - Study the effect of multiple factors, each at two levels. - Estimate main effects and interactions using an orthogonal, balanced design. - Build linear models that support prediction and optimization within the studied region. Essential ideas include: - Coding factors at −1 and +1 for simple effect calculation. - Using contrast-based formulas to estimate main and interaction effects. - Applying ANOVA and regression to test significance and build reliable models. - Interpreting main effects and interaction plots to understand factor behavior. - Using replication, randomization, and center points to address error, bias, and curvature. Mastering these concepts enables rigorous design, analysis, and interpretation of 2ᵏ full factorial experiments for effective data-driven process improvement.