4.5.3 Experimental Resolution

Experimental Resolution Introduction Experimental resolution describes how clearly a designed experiment can separate the effects of factors and their interactions. In practice, resolution tells you which effects are confounded with which other effects. Understanding experimental resolution is essential when selecting, running, and interpreting factorial and fractional factorial designs. The goal is to obtain trustworthy information about main effects and interactions with the least possible experimental effort. --- Basics of Confounding and Aliasing Confounding, Aliasing, and Clear Effects In fractional factorial designs, not all effects can be estimated independently. Some effects are mathematically tied together; this is confounding (or aliasing). - Confounding: Two or more effects are combined so that their individual contributions cannot be separated based on the data. - Alias structure: The full set of equations that show which effects are confounded with which other effects. - Clear effect: An effect that is not confounded with any other effect of practical interest (e.g., a main effect that is not aliased with another main effect). The resolution of a design summarizes how severely key effects (main effects and low-order interactions) are confounded. --- Defining Relation and Design Generators Design Generators A fractional factorial design is created by defining one or more factors as products of other factors. These are design generators. Example for a 2^(4-1) design: - Factors: A, B, C, D - Generator: D = ABC This means: - The level of D in each run is set equal to the product of A, B, and C. Defining Relation The defining relation is obtained by multiplying both sides of each generator by the right-hand side to set up an equation equal to the identity (I). Using D = ABC: - Multiply both sides by ABCD: - D × ABCD = ABC × ABCD - Since A×A = B×B = C×C = D×D = I (identity), this simplifies to: - I = ABCD The defining relation lists one or more words (terms) that equal I. - Word: A product of factor letters, like ABC or ABD. - Word length: Number of letters in the word (e.g., ABCD has length 4). All alias relationships can be derived from the defining relation by multiplying any effect by each word in the defining relation. --- Concept of Design Resolution Formal Definition of Resolution Experimental resolution is defined as the length of the shortest word (excluding I) in the defining relation of a design. - Resolution III: Shortest word length is 3. - Resolution IV: Shortest word length is 4. - Resolution V: Shortest word length is 5. Higher resolution means better separation of important effects. --- Interpretation of Resolution Levels Resolution III Designs In Resolution III designs: - Main effects may be aliased with two-factor interactions. - Two-factor interactions are aliased with other two-factor interactions. Interpretation: - Main effects are not aliased with other main effects. - Main effects can be biased by active two-factor interactions. - These designs are mainly useful for initial screening where it is acceptable to assume interactions are small or negligible. Key aliasing patterns: - Main effect ~ 2FI (two-factor interaction) - 2FI ~ 2FI Implication: - When a main effect appears significant, it may actually be due to a two-factor interaction. Resolution IV Designs In Resolution IV designs: - Main effects are aliased with three-factor (or higher) interactions. - Two-factor interactions are aliased with other two-factor interactions. Interpretation: - Main effects are clear of other main effects and 2FIs. - If three-factor and higher interactions are negligible, main effects can be interpreted with confidence. - Interactions still require caution. Key aliasing patterns: - Main effect ~ 3FI - 2FI ~ 2FI Common use: - More reliable screening when interactions among factors might exist but are considered less important than main effects. Resolution V Designs In Resolution V designs: - Main effects are aliased with four-factor (and higher) interactions. - Two-factor interactions are aliased with three-factor (and higher) interactions. - Main effects are not aliased with 2FIs or other main effects. - 2FIs are not aliased with other 2FIs. Interpretation: - Main effects and two-factor interactions can both be estimated cleanly if higher-order interactions are negligible. - This is suitable when interaction effects are expected to be important. Key aliasing patterns: - Main effect ~ 4FI - 2FI ~ 3FI --- Aliasing Patterns by Resolution Summary by Resolution Type - Resolution III - Main vs main: clear - Main vs 2FI: aliased - 2FI vs 2FI: aliased - Resolution IV - Main vs main: clear - Main vs 2FI: clear - Main vs 3FI: aliased - 2FI vs 2FI: aliased - Resolution V - Main vs main: clear - Main vs 2FI: clear - 2FI vs 2FI: clear - Main vs 4FI: aliased - 2FI vs 3FI: aliased The guiding assumption in practice is that: - Higher-order interactions (3FI and above) are usually small or negligible. - Therefore, in higher-resolution designs, key lower-order effects are effectively clear. --- Determining Resolution from a Generator Step-by-Step Approach To determine resolution: - Step 1: List generators. Example: D = ABC for a 2^(4-1) design. - Step 2: Form the defining relation. Convert generators to equations equal to I: - D = ABC → I = ABCD - Step 3: Identify all unique words. In this simple case, the only nontrivial word is ABCD. - Step 4: Find the shortest word length. Here, length(ABCD) = 4 → Resolution IV. If there are multiple generators, more words appear: Example: 2^(5-2) design: - Generators: D = AB, E = AC - Defining relation: - I = ABD (from D = AB) - I = ACE (from E = AC) - Multiply: (ABD)(ACE) → I = BCDE - Words: ABD, ACE, BCDE - Shortest word length = 3 (ABD, ACE) → Resolution III. --- Trade-offs Between Resolution and Resources Relationship to Fractionation In 2-level factorial designs: - Full factorial: 2^k runs, highest ability to estimate all effects. - Fractional factorial: 2^(k-p) runs, where p is the degree of fractionation. More fractionation (larger p): - Fewer runs. - More severe confounding. - Usually lower resolution. Less fractionation (smaller p): - More runs. - Less confounding. - Usually higher resolution. Choosing Resolution in Practice When planning an experiment: - If resources are tight and many factors are being screened: - Resolution III may be acceptable under strong assumptions that interactions are negligible. - If main effects must be reliable and some interactions may be important: - Resolution IV is generally more suitable. - If both main effects and two-factor interactions are important: - Aim for Resolution V, accepting the larger run size. A common workflow: - Start with a lower-resolution screening design to reduce the number of factors. - Follow with a higher-resolution or more focused design to study main effects and 2FIs precisely. --- Using Alias Structure with Resolution Building and Reading Alias Structures Once you know the defining relation, construct alias relationships: - Step 1: Identify the defining words (e.g., I = ABCD). - Step 2: For any effect of interest, multiply it by each word in the defining relation to find aliases. Example: For I = ABCD: - Consider effect A: - A × I = A - A × ABCD = BCD - Alias set: A = BCD - Consider effect AB: - AB × I = AB - AB × ABCD = CD - Alias set: AB = CD This shows: - A is aliased with the three-factor interaction BCD. - The 2FI AB is aliased with another 2FI, CD. Resolution tells you in advance what types of effects will appear in these alias sets. --- Impact of Resolution on Interpretation Bias in Estimated Effects When effects are aliased, the estimated coefficient for one effect actually reflects the combined impact of all aliased effects. For example: - In a Resolution III design, if a main effect A is aliased with BC: - The estimated effect for A = true A effect + BC effect. If BC is not negligible, the estimate for A is biased. This is why: - Resolution III designs rely heavily on the assumption that interactions are negligible. - Higher resolution is preferred when this assumption is questionable. Strategies to Address Ambiguity When confounding leads to ambiguous interpretation: - Fold-over designs: - Add a second fraction with selected factor signs reversed to de-alias specific effects. - Follow-up experiments: - Run additional targeted comparisons to isolate suspected interactions. - Model checking: - Evaluate residuals and model fit; unexpected patterns may indicate that confounded interactions are active. These strategies are applied after understanding the resolution and alias structure. --- Practical Selection of Experimental Resolution Factors to Consider When choosing resolution: - Objective of the experiment: - Initial screening → Resolution III or IV may be sufficient. - Detailed effect and interaction estimation → Resolution IV or V. - Number of factors: - More factors increase the temptation to use higher fractionation, but this lowers resolution. - Available runs: - Time, cost, and material often limit how high the resolution can be. The guiding principle: - Use the highest resolution that is feasible with available resources, given the importance of main effects and interactions. --- Summary Experimental resolution indicates how clearly a design can distinguish main effects and interactions from each other. It is determined by: - The defining relation of the design. - The length of the shortest word in that relation. Key points: - Resolution III: Main effects confounded with 2FIs; suitable for rough screening when interactions are assumed negligible. - Resolution IV: Main effects clear of 2FIs; 2FIs confounded with other 2FIs; appropriate for more reliable screening. - Resolution V: Main effects and 2FIs both clear of each other; preferred when interactions are important. Understanding and using experimental resolution involves: - Working with design generators and defining relations. - Interpreting alias structures. - Balancing run size against the need for clarity of effects. Mastering these concepts allows you to choose, execute, and interpret designed experiments with appropriate confidence in the conclusions drawn from the data.