Factorial experiments

Table of Contents
1. Factorial Experiments: Basic Concepts
2. 2² Factorial Experiment
3. 3² Factorial Experiment
4. Confounding in Factorial Experiments
5. Comparison: 2² vs 3² Factorial Designs
6. Common Mistakes and Exam Tips
View more

Factorial experiments are research designs where we study the effect of two or more factors simultaneously. Each factor is tested at different levels. The notation 2² means 2 factors each at 2 levels. Similarly, 3² means 2 factors each at 3 levels. These designs help examine both main effects (individual factor effects) and interaction effects (combined factor effects). Confounding is a technique used to manage large factorial experiments by deliberately mixing some effects with blocks to reduce experimental size.

1. Factorial Experiments: Basic Concepts

1.1 Definition and Notation

• Factorial Experiment: An experimental design where all possible combinations of factor levels are tested.
• Factor: An independent variable whose effect on the response is being studied (e.g., Temperature, Pressure, Fertilizer type).
• Level: The specific value or category at which a factor is tested (e.g., High/Low temperature, 30°C/40°C/50°C).
• Treatment Combination: A specific combination of factor levels tested in the experiment.
• Notation sⁿ: 's' = number of levels per factor, 'n' = number of factors.
• Total number of treatment combinations = sⁿ

1.2 Main Effects and Interaction Effects

• Main Effect: The average effect of one factor across all levels of other factors. It measures the independent contribution of a single factor.
• Interaction Effect: When the effect of one factor depends on the level of another factor. The combined effect differs from the sum of individual effects.
• Two-factor Interaction: Denoted as AB (for factors A and B). Measures whether A's effect changes at different levels of B.
• Higher-order Interactions: Three-factor (ABC), four-factor interactions, etc. Generally assumed negligible in practice.

1.3 Advantages of Factorial Designs

• Efficiency: Studies multiple factors simultaneously with the same experimental units used for one factor at a time.
• Interaction Study: Only factorial designs can detect and estimate interaction effects between factors.
• Wider Applicability: Conclusions are valid over a broader range of experimental conditions.
• Reduced Experimental Error: Uses all observations to estimate each effect, increasing precision.

2. 2² Factorial Experiment

2.1 Structure and Treatment Combinations

• Design has 2 factors (A and B), each at 2 levels.
• Total treatment combinations = 2² = 4
• Standard Notation:
  • (1) = Both factors at lower level
  • a = Factor A at higher level, B at lower level
  • b = Factor B at higher level, A at lower level
  • ab = Both factors at higher level
• Levels often denoted as: Lower level (-1 or 0), Higher level (+1 or 1)

2.2 Layout and Representation

Treatment combinations can be represented in a 2×2 factorial table:

2.3 Estimation of Effects in 2² Design

Let the response values be denoted by the same notation as treatment combinations.

2.3.1 Main Effect Formulas

• Main Effect of A:

  A = ¹⁄_2n [a + ab - (1) - b]

  This represents the average change in response when A moves from low to high level.

• Main Effect of B:

  B = ¹⁄_2n [b + ab - (1) - a]

  This represents the average change in response when B moves from low to high level.

• Here, n = number of replications of each treatment combination.

2.3.2 Interaction Effect Formula

• Interaction Effect AB:

  AB = ¹⁄_2n [(1) + ab - a - b]

  Measures how much the effect of A differs at different levels of B (or vice versa).

• Alternative Interpretation: AB = (Effect of A at high B) - (Effect of A at low B)
• If AB ≠ 0, interaction is present; factors do not act independently.

2.4 Degrees of Freedom in 2² Design

• Total degrees of freedom (df) = 4n - 1 (where n = replications)
• Factor A: df = 1
• Factor B: df = 1
• Interaction AB: df = (2-1) × (2-1) = 1
• Error: df = 4(n-1)

2.5 Analysis of Variance (ANOVA) for 2² Design

ANOVA table structure:

• Sum of Squares formulas:
  • SS_A = n × (Effect A)²
  • SS_B = n × (Effect B)²
  • SS_AB = n × (Effect AB)²

2.6 Yates' Algorithm for 2² Design

Yates' Method: A systematic computational procedure to calculate effects in factorial experiments without using formulas.

Steps for Yates' Algorithm:

1. Write treatment combinations in standard order: (1), a, b, ab
2. Write corresponding total responses in a column
3. First Cycle: Add successive pairs, then subtract successive pairs
4. Second Cycle: Repeat the add-subtract process on results from first cycle
5. Divide results by appropriate divisor (2ⁿ × r, where n=2, r=replications) to get effects
6. Order of effects obtained: Total, A, B, AB

3. 3² Factorial Experiment

3.1 Structure and Treatment Combinations

• Design has 2 factors (A and B), each at 3 levels.
• Total treatment combinations = 3² = 9
• Notation: Factors at levels 0, 1, 2 (or Low, Medium, High)
• Treatment combinations: (0,0), (1,0), (2,0), (0,1), (1,1), (2,1), (0,2), (1,2), (2,2)
• Alternative notation: aⁱb^j where i,j = 0,1,2

3.2 Layout Representation

Treatment combinations arranged in a 3×3 factorial table:

3.3 Effects in 3² Design

• Main Effects: Each factor has 2 degrees of freedom (3 levels - 1)
• Main effects split into Linear and Quadratic components
• Linear Effect of A: Measures the straight-line trend as A increases
• Quadratic Effect of A: Measures the curvature (non-linear trend) as A increases
• Same decomposition applies to Factor B

3.4 Interaction Components in 3² Design

• Interaction AB: Has (3-1) × (3-1) = 4 degrees of freedom
• Interaction decomposed into four components:
  • A_LB_L: Linear × Linear interaction
  • A_LB_Q: Linear × Quadratic interaction
  • A_QB_L: Quadratic × Linear interaction
  • A_QB_Q: Quadratic × Quadratic interaction
• Each interaction component has 1 df

3.5 Orthogonal Polynomial Contrasts

Orthogonal Polynomials: Mathematical coefficients used to partition factorial effects into linear and quadratic components.

• For 3 equally-spaced levels:
  • Linear coefficients: -1, 0, +1
  • Quadratic coefficients: +1, -2, +1
• These coefficients are orthogonal (sum of products = 0)
• Used to compute linear and quadratic effects through contrast formulas

3.6 Degrees of Freedom in 3² Design

• Total df = 9n - 1 (where n = replications)
• Factor A: df = 2 (Linear: 1, Quadratic: 1)
• Factor B: df = 2 (Linear: 1, Quadratic: 1)
• Interaction AB: df = 4 (4 interaction components, each with 1 df)
• Error: df = 9(n-1)

3.7 ANOVA Structure for 3² Design

4. Confounding in Factorial Experiments

4.1 Concept and Necessity of Confounding

• Confounding: A technique where certain factorial effects are intentionally mixed (confounded) with block effects to reduce block size.
• Purpose: Makes factorial experiments more manageable when complete replication in homogeneous blocks is impractical.
• Requirement: Arises when the number of treatment combinations exceeds the homogeneous block capacity.
• Example: 2⁴ design has 16 treatments; difficult to have 16 homogeneous experimental units in one block.
• Strategy: Sacrifice information on higher-order interactions (usually negligible) to gain precision through blocking.

4.2 Principles of Confounding

• Confound only high-order interactions: Main effects and low-order interactions are preserved.
• Block size reduction: Reduces blocks to 2^n-p units (for 2ⁿ designs with p effects confounded).
• Information loss: Confounded effects cannot be estimated separately from block effects.
• Replications: Multiple replicates can use different confounding schemes to recover some information.
• Complete vs Partial Confounding:
  • Complete Confounding: Same effect confounded in all replicates; completely loses information on that effect.
  • Partial Confounding: Different effects confounded in different replicates; partial information retained on all effects.

4.3 Confounding in 2² Factorial

• 4 treatment combinations: (1), a, b, ab
• Can divide into 2 blocks of 2 units each
• Possible confounding schemes:
  • Confound AB interaction with blocks (most common, preserves main effects)
  • Confound A main effect (loses information on A)
  • Confound B main effect (loses information on B)
• Standard confounding pattern (AB confounded):
  • Block I: (1), ab (treatments with even number of letters)
  • Block II: a, b (treatments with odd number of letters)
• The AB interaction effect cannot be separated from block difference

4.4 Confounding in 2³ and Higher 2ⁿ Designs

• 2³ Design: 8 treatment combinations; can use 2 blocks of 4 or 4 blocks of 2
• Effects available: A, B, C (main effects); AB, AC, BC (two-factor interactions); ABC (three-factor interaction)
• Common scheme: Confound ABC (highest-order interaction) with blocks
• Confounding pattern for ABC:
  • Block I: (1), ab, ac, bc (treatments where sum of exponents is even)
  • Block II: a, b, c, abc (treatments where sum of exponents is odd)
• Generalized Interaction: For any 2ⁿ design, use the defining contrast to assign treatments to blocks

4.5 Confounding in 3² Factorial

• 9 treatment combinations; can be arranged in 3 blocks of 3 units each
• Commonly confound one interaction component (4 df interaction AB)
• Typical approach: Confound the AB component with 2 df, leaving 2 df estimable
• Confounding scheme using modular arithmetic: Use the relation: i × α + j × β ≡ k (mod 3), where i,j are factor levels (0,1,2)
• Different values of α and β give different confounding patterns
• Common pattern confounds AB interaction, preserving both main effects A and B

4.6 Construction of Confounding Design

Steps to construct a confounded factorial design:

1. Identify the effect to confound: Usually highest-order interaction
2. Determine block size: Based on homogeneity constraints
3. Use defining contrast: Mathematical relation that partitions treatments into blocks
4. Assign treatments: Place treatments satisfying the defining equation in one block, others in remaining blocks
5. Replicate if needed: Use different confounding schemes in different replicates for partial confounding

4.7 Analysis with Confounding

• Confounded effects are not tested in ANOVA; their SS is absorbed in block SS
• Modified ANOVA structure:
  • Replicates (if any)
  • Blocks within replicates
  • Treatments (partitioned into non-confounded factorial effects)
  • Error
• Error term computed from non-confounded interactions with treatment × replicate interaction
• Intra-block analysis: Estimates non-confounded effects within blocks
• With partial confounding, inter-block information can be recovered through combined analysis

4.8 Advantages and Limitations of Confounding

Advantages:

• Reduces block size, making experiments practical in heterogeneous environments
• Increases precision for main effects and low-order interactions
• Efficient use of experimental resources
• Flexible design allowing different confounding patterns across replicates

Limitations:

• Loss of information on confounded effects
• Complexity in design construction and analysis
• Requires careful planning; errors in confounding scheme invalidate results
• Not suitable when all interactions are potentially important

5. Comparison: 2² vs 3² Factorial Designs

6. Common Mistakes and Exam Tips

6.1 Trap Alerts: Common Student Errors

• Mistake: Confusing main effects with simple effects. Correction: Main effect averages across all other factor levels; simple effect is at one specific level.
• Mistake: Assuming no interaction means no relationship. Correction: Factors can have strong main effects without interaction; interaction refers only to non-additive effects.
• Mistake: In 3² designs, treating effects as single df instead of decomposing into linear and quadratic. Correction: Always partition 3-level factor effects into 2 components.
• Mistake: Using wrong divisor in Yates' algorithm. Correction: Divisor is always 2ⁿ times number of replications.
• Mistake: Confounding main effects instead of interactions. Correction: Always confound highest-order interactions first; main effects should never be confounded if avoidable.
• Mistake: In confounding, forgetting that confounded effects cannot be tested. Correction: Confounded SS goes to blocks, not treatments.
• Mistake: Incorrectly assigning treatment combinations to blocks in confounded designs. Correction: Always use the defining contrast systematically.

6.2 Key Points for Quick Revision

• Total treatments in sⁿ design = sⁿ
• Interaction df = product of individual factor df
• In 2² design: 3 effects (A, B, AB), each with 1 df
• In 3² design: A and B each have 2 df; AB has 4 df
• Yates' algorithm order: (1), a, b, ab for 2² design
• Confounding reduces block size by mixing interaction with blocks
• Partial confounding > Complete confounding for information recovery
• Always test confounding scheme before conducting experiment

Understanding 2² and 3² factorial designs with confounding is crucial for designing efficient experiments that balance information requirements with practical constraints. The 2² design offers simplicity for two-level screening, while 3² designs capture non-linear trends through quadratic effects. Confounding enables these designs to be implemented in smaller, more homogeneous blocks by strategically sacrificing information on less important (usually higher-order) interactions. Mastery of effect estimation, ANOVA structure, and confounding principles allows you to design, execute, and analyze factorial experiments that extract maximum information from limited experimental resources.