4.5 Fractional Factorial Experiments

Fractional Factorial Experiments Introduction Fractional factorial experiments are designed experiments that use only a fraction of the runs required for a full factorial design, while still allowing insight into the most important factors and interactions. They are especially useful when: - The number of factors is large. - Resources (time, material, budget) are constrained. - The purpose is screening: finding the vital few factors among many. This article explains the essential concepts, notation, planning, analysis, and interpretation of fractional factorial experiments at a level sufficient for rigorous application in process improvement and problem solving. --- Full Factorial vs Fractional Factorial Designs Full Factorial Designs In a 2-level full factorial design with k factors: - Each factor has 2 levels (often coded as -1 and +1 or low and high). - Total runs = (2^k). Examples: - 3 factors → (2^3 = 8) runs. - 6 factors → (2^6 = 64) runs. A full factorial design allows: - Estimation of all main effects. - Estimation of all 2-factor interactions. - Estimation of higher-order interactions. However, as k grows, runs grow exponentially and rapidly become impractical. Motivation for Fractional Designs Many real processes have: - Several potential factors. - Limited budget for experimentation. Experience and theory often suggest that: - Main effects and low-order interactions (especially 2-factor) are usually most important. - Higher-order interactions (3-factor and above) are often negligible. Fractional factorial designs systematically use this assumption to: - Reduce the number of runs. - Retain the ability to estimate main effects and key low-order interactions, accepting some ambiguity (aliasing) with higher-order effects. --- Basic Notation and Terminology 2^(k-p) Designs A 2-level fractional factorial design is commonly described as: - (2^{k-p}) Where: - k = number of factors. - p = degree of fractionation. - Runs = (2^{k-p}). Examples: - (2^{4-1}): 4 factors in a half-fraction = 8 runs. - (2^{5-2}): 5 factors in a quarter-fraction = 8 runs. - (2^{7-3}): 7 factors in an eighth-fraction = 16 runs. Generators and Defining Relations A fractional factorial is built by: - Starting from a smaller design core (e.g., 3 factors). - Defining additional factors as products of existing ones. A generator is an equation that defines a factor in terms of others. Example (half fraction, 4 factors A, B, C, D): - Generator: (D = ABC) This means that the level of D in each run is set as the product (interaction) of A, B, and C. From the generator, a defining relation is constructed: - Multiply both sides by D: (D \cdot D = ABCD) - In coded 2-level algebra: (D^2 = I), so: - (I = ABCD) (I is the identity element) The full set of words derived from generators (including I) is the defining relation. It encodes which effects are aliased. --- Aliasing and Resolution Aliasing (Confounding) Concepts In fractional factorial designs, some effects share the same column of coded levels and cannot be separated; they are aliased or confounded. If: - Effect X and Effect Y are aliased, - Then their combined effect is estimated together, not individually. Aliasing is unavoidable when using fractions; understanding it is crucial for correct interpretation. Deriving Alias Structure Using the defining relation, you can find which effects are aliased. Example: - Defining relation: (I = ABCD) To find aliases of a main effect A: - Multiply both sides by A: (A = A \cdot ABCD = BCD) - So A is aliased with BCD. To find aliases of a 2-factor interaction AB: - Multiply by AB: (AB = AB \cdot ABCD = CD) - So AB is aliased with CD. This process can be used systematically to find all aliases of: - Main effects. - 2-factor interactions. - Higher-order interactions. Design Resolution Design resolution describes the pattern of aliasing between main effects and interactions. Common resolutions: - Resolution III - Main effects are aliased with 2-factor interactions. - 2-factor interactions are aliased with other 2-factor interactions. - Interpretation of main effects is risky if interactions are not truly negligible. - Resolution IV - Main effects are aliased with 3-factor (and higher) interactions. - 2-factor interactions are aliased with other 2-factor interactions. - Main effects can be interpreted assuming 3-factor+ interactions are negligible. - Resolution V - Main effects are aliased with 4-factor (and higher) interactions. - 2-factor interactions are aliased with 3-factor (and higher) interactions. - Both main effects and 2-factor interactions can be interpreted assuming higher-order interactions are negligible. Higher resolution is generally better but requires more runs. --- Constructing Common 2-Level Fractional Designs Example: 2^(4-1) Half-Fraction, Resolution IV Goal: - 4 factors (A, B, C, D). - Use 8 runs instead of 16. Choose generator: - (D = ABC) Defining relation: - (I = ABCD) Resolution: - The shortest word in the defining relation (excluding I) is length 4 (ABCD). - Therefore, this is a Resolution IV design. Key aliasing: - A is aliased with BCD. - B is aliased with ACD. - C is aliased with ABD. - D is aliased with ABC. - AB is aliased with CD. - AC is aliased with BD. - AD is aliased with BC. Main effects are only aliased with 3-factor interactions; 2-factor interactions are aliased with other 2-factor interactions. Example: 2^(5-1) Half-Fraction, Resolution V Goal: - 5 factors (A, B, C, D, E). - Use 16 runs instead of 32. Choose generator: - (E = ABCD) Defining relation: - (I = ABCDE) Shortest word (excluding I) has 5 letters → Resolution V. Implications: - Main effects aliased with 4-factor interactions. - 2-factor interactions aliased with 3-factor interactions. - Both main effects and 2-factor interactions can be interpreted reliably if higher-order interactions are negligible. --- Design Generators and Defining Relations Multiple Generators For higher fractions (e.g., quarter, eighth), more than one generator is used. Example: 2^(5-2) quarter fraction with 5 factors A, B, C, D, E: Choose: - (D = AB) - (E = AC) Determine defining relation: - Start with: (I = ABD) (from D = AB) - And: (I = ACE) (from E = AC) Combine: - Multiply: ((ABD)(ACE) = I \cdot I = I) - Left side: (A^2 B C D E = BCDE) (since (A^2 = I)) - So: (I = BCDE) Defining relation is: - (I = ABD = ACE = BCDE) Shortest word length: - 3 (ABD, ACE) → Resolution III. Choosing Generators Selection of generators aims to: - Achieve a desired resolution (often IV or V). - Minimize aliasing among effects of interest (primarily main and 2-factor effects). Practical guidelines: - Prefer designs where main effects are not aliased with 2-factor interactions. - When 2-factor interactions are critical, aim for Resolution V. - Use standard generator tables or software to select high-quality designs. --- Planning a Fractional Factorial Experiment Clarify Objectives Before designing a fractional factorial: - Define the primary goal: - Screening many factors? - Understanding specific interactions? - Identify the response(s) to be measured. - List all candidate factors and plausible ranges. Choose Number of Factors and Levels - Select only factors that are realistically adjustable during the experiment. - Use two levels per factor for screening: - Low: current or reduced setting. - High: increased or improved setting. - Ensure the levels are safe and practically feasible. Select Fraction and Resolution Consider: - Number of factors k. - Resource constraints (runs available). - Required interpretability of main and interaction effects. Guidelines: - For screening many factors: - Resolution III or IV designs may be acceptable if main interest is identifying a subset of important factors and most interactions are expected small. - For more detailed understanding: - Use Resolution V or a full factorial for critical factors. Randomization, Replication, and Blocking Even in fractional designs, good experimental practice applies. - Randomization - Randomize run order to protect against time-related and other lurking variables. - Replication - Replicate some or all runs to estimate pure error and improve effect estimation. - Especially important when noise is high or significance testing is required. - Blocking - If runs cannot all be done under homogeneous conditions (e.g., different days or machines), consider treating these as blocks. - Blocks can be incorporated into the design using additional generators, but this changes alias structure. --- Analysis of Fractional Factorial Designs Estimating Effects For 2-level designs, effects are estimated using contrasts. For a factor A: - Effect(A) = (Average response at A high) – (Average response at A low) Equivalent in coded form: - Use +1 for high and -1 for low. - Effect is proportional to the contrast between +1 and -1 runs. Similar logic applies to interactions (e.g., AB, AC): - Compute interaction columns as products of corresponding factor columns. - Estimate effects using contrasts on these columns. Normal or Half-Normal Plots A common method for screening effects: - Create a plot (normal or half-normal) of effect magnitudes. - Under the assumption that most effects are negligible: - Negligible effects align roughly along a reference line. - Significant effects stand out away from the line. This helps identify: - Which main effects appear important. - Which 2-factor interactions may be active. ANOVA and Regression Model Fractional factorial data can be analyzed using: - Linear regression models with coded variables. - ANOVA to test significance of factors and interactions. General model form for 2-level k-factor design: - (Y = \beta0 + \sum \betai Xi + \sum \beta{ij} Xi Xj + \cdots + \varepsilon) Where: - (X_i) are coded factor levels (-1, +1). - (\beta_i) are main-effect coefficients. - (\beta_{ij}) are interaction coefficients. - (\varepsilon) is random error. In fractional designs: - Some coefficients are aliased and cannot be uniquely estimated. - Interpretation must respect the alias structure. --- Interpreting Results with Aliasing Using Aliasing to Interpret Effects When an effect is significant in a fractional design: - It actually estimates a combination of aliased effects. Example: - In a 2^(4-1) design with (I = ABCD), AB is aliased with CD. - A large AB effect may mean: - AB is active, CD is negligible, or - CD is active, AB is negligible, or - Both are active. To interpret: - Use process knowledge: - Is AB interaction physically plausible? - Is CD interaction plausible? - Consider patterns in other effects that may support one interpretation. Assumption of Negligible Higher-Order Interactions Fractional designs often rely on: - The assumption that 3-factor and higher interactions are negligible. In a Resolution IV or V design: - Main effects and key 2-factor interactions are aliased with higher-order interactions. - If higher-order interactions are indeed small: - Estimates of main and 2-factor effects can be treated as approximately pure. If evidence suggests higher-order interactions might be large: - Consider augmenting the design to break aliasing (e.g., fold-over). --- Fold-Over Designs Purpose of Fold-Over A fold-over design augments an existing fractional factorial to: - Resolve certain aliases. - Separate aliased effects. Common method: - Run an additional fraction with specific factor level patterns reversed relative to the original. Simple Fold-Over Example Starting with a 2^(4-1) design: - Generator: (D = ABC) - Defining relation: (I = ABCD) A complete fold-over: - Add another 8 runs where one or more factors have their signs reversed relative to the original design (e.g., reverse all factors). - The combined 16 runs form a full 2^4 design. Benefits: - All aliasing among main effects and 2-factor interactions is removed. - Effects can be estimated independently. Partial fold-over strategies: - Can target specific aliases to resolve only certain confounding while limiting additional runs. --- Practical Considerations and Common Pitfalls When to Use Fractional Factorials Fractional factorial designs are appropriate when: - There are many potential factors. - The primary objective is to identify which few factors matter most. - Resources limit the number of experimental runs. They are less appropriate when: - Every interaction must be independently estimated. - There is strong evidence that higher-order interactions are important. Common Mistakes - Ignoring alias structure - Interpreting effects as if they were unaliased can lead to wrong conclusions. - Using too low a resolution - Resolution III designs can mislead if 2-factor interactions are not negligible. - Over-interpreting small effects - Random variation can make some effects appear non-zero in small designs without replication. - Levels chosen too narrow or too wide - Too narrow: effects appear small or insignificant. - Too wide: operating outside safe or realistic process conditions. - No replication - Lack of replication makes it harder to distinguish real effects from noise. --- Summary Fractional factorial experiments provide a structured way to explore many factors using far fewer runs than full factorial designs, by deliberately accepting and managing aliasing among effects. Key ideas: - 2-level fractional designs are denoted (2^{k-p}), where k is factors, p is fractionation. - Generators and defining relations determine the alias structure. - Design resolution (III, IV, V) describes which orders of effects are aliased. - Analysis involves estimating effects, using normal/half-normal plots, and regression/ANOVA, always interpreted through the alias structure. - Practical use depends on clear objectives, careful design choice, appropriate factor levels, randomization, and, when needed, replication and fold-over to resolve critical aliasing. A solid grasp of these concepts allows efficient, reliable experimentation when a full factorial is impractical, while maintaining control over the trade-off between information gained and resources used.