Factorial ANOVA Notes

Factorial ANOVA

Factorial Designs

  • Psychologists often investigate the effects of multiple independent variables (IVs) on a dependent variable (DV).
  • Factorial Experimental Design: Research designs involving more than one IV.
    • Main Effect: The impact of one IV on the DV.
    • Interaction: The effect of one IV on the DV depends on the level of another IV.

Factorial Design Example

  • Eskine, Kacinik, & Prinz, 2010:
    • Investigated whether moral judgments vary with the experience of gustatory disgust.
    • Examined if liberals and conservatives differ in making extreme moral judgments based on disgust.
  • Method:
    • Participants identified as liberal or conservative.
    • Participants consumed a bitter beverage, a sweet beverage, or water.
    • They read scenarios depicting moral transgressions.
    • Participants rated the moral wrongness of each transgression.

2 x 3 Design

  • Illustrates a design with two levels of one factor (e.g., Liberal/Conservative) and three levels of another (e.g., Bitter, Neutral/Water, Sweet).

Possible Outcomes

  • Potential study outcomes:
    • Liberals and conservatives differ in moral judgments.
    • Extremity of moral judgments differs across taste conditions.
    • The impact of taste on moral judgments varies based on political ideology.
  • The study could reveal one or more of these possibilities.

Main Effects and Interactions

  • Factorial ANOVA is similar to One-way ANOVA but evaluates multiple effects.
  • With two IVs:
    • Two main effects.
    • One interaction effect.
  • Requires multiple Sum of Squares (SS) estimates and F-ratios.

Sum of Squares Total – SST

  • Aims to explain the total variability in the outcome variable:
    • SST = \sum(X - GMX)^2

Sum of Squares Within (Error) – SSR

  • Error term reflects the pooled variability within each group:
    • SSR = \sum(Xi k - X k)^2

Sum of Squares for the Model – SSM

  • Systematic variation explained by the model represents the improvement over simply guessing the grand mean:
    • SSM = \sum nk (X k - X GM)^2

Sum of Squares for Main Effects – SSA and SSB

  • The sum of squares for the model can be divided into the sum of squares for the main effects:
    • SSA = \sum nk (X k - X GM)^2
    • SSB = \sum nk (X k - X GM)^2

Sum of Squares for the Interaction – SSA x B

  • The sum of squares for the model includes a component for the interaction:
    • SSAxB = SSM - SSA - SSB
    • Degrees of freedom for the interaction term: df AxB = df A \times dfB

Mean Squares and F-Ratios

  • Formulas for Mean Squares (MS) and F-ratios:
    • MSAxB = \frac{SSAxB}{df AxB}
    • MSA = \frac{SSA}{df A}
    • MSB = \frac{SSB}{df B}
    • MSR = \frac{SSR}{df R}
    • F = \frac{MSA}{MSR}
    • F = \frac{MSB}{MSR}
    • F = \frac{MSAxB}{MSR}

Example Results: Descriptive Statistics

  • Presents descriptive statistics for the extremity of moral judgment based on political orientation (Conservative, Liberal) and taste condition (Bitter, Water, Sweet).
  • Includes Mean, Standard Deviation, and N for each group.

Example Results: Tests of Between-Subjects Effects

  • Shows ANOVA results with:
    • Source, Type III Sum of Squares, df, Mean Square, F, Sig., Partial Eta Squared for Corrected Model, Intercept, ConLib (Political Orientation), Condition (Taste), ConLib*Condition (Interaction), and Error.
  • Indicates the significance of main effects and interaction.

Example Results: Marginal Means

  • Reports Estimated Marginal Means for:
    • Political Orientation (Conservative, Liberal).
    • Taste Condition (Bitter, Water, Sweet).
  • Includes Mean, Std. Error, and 95% Confidence Intervals.

Factorial ANOVA – Interpreting and Evaluating Effects

  • F ratios are omnibus tests.
  • Significant Fs require follow-up with planned comparisons or post hoc tests.

Factorial ANOVA – Planned Comparisons for Main Effects

  • Follow-up comparisons for main effects are similar to one-way ANOVA.
    • Logic and rules are identical.
    • Formulas are slightly different.
  • If an IV has only two levels, planned comparisons are unnecessary; means can be directly compared.

Main Effect of A - Pairwise Main Comparisons

  • \psi A = X 1 - X 2
    • \psi is the same as before.
    • SSAcomp is the Sum of Squares A comparison.
    • n is the number of people in each CELL of the design (assumes equal n).
    • b is the number of levels of variable B.
    • The b would change to a and equal the number of levels of variable A if we were evaluating the main effect of B.
  • SSAcomp = \frac{2(bn)(\psi A)^2}

Pairwise Main Comparisons (Continued)

  • MSAcomp = SScomp
    • Note that, here, there is always 1 degree of freedom, so MScomp will always equal SScomp.
  • FAcomp = \frac{MSAcomp}{MSWithin}
    • Note that we use the error term from our omnibus test here.

Main Effect of A – Complex Main Comparisons

  • With coefficients, formulas change slightly.
  • Incorporate the number of levels of B:
    • \psi A = (c1)(X1)+ (c2)(X2 )+ (c3)(X3 )+…
    • SSAcomp = \frac{2(\sum c)^2}{(bn)}

Interaction Effects

  • Analyzing interaction effects is more complex.
  • Requires a plan of attack.
  • Goal: Fully explain the interaction in an interpretable way consistent with theory.
  • Types of comparisons:
    • Simple Effects
    • Pairwise simple comparisons
    • Complex simple comparisons

Simple Effects

  • The effect of one variable at a specific level of another.
    • E.g., the effectiveness of the taste manipulation for conservatives.
  • Compute the sum of squares for the simple comparison:
    • SS A \text{ comp at } bj = \sum nk (X k - X GM)^2

Simple Effects (Continued)

  • MS A \text{ comp. at bj } = \frac{SS}{df}
    • Degrees of freedom equal the number of groups minus 1 (i.e., k-1)
  • F-ratio:
    • F A \text{ comp. at bj } = \frac{MS}{MSWithin}
    • Use the error term from the omnibus test.

Pairwise Simple Comparisons

  • \psi A = X 1 - X 2
    • SSAcomp = \frac{(\psi A)^2}{2n}
    • Where n is the number of people in each cell you are comparing (assumes equal n).
  • FAcomp = \frac{MS}{MSWithin}
    • We once again use the error term from our omnibus test.

Complex Simple Comparisons

  • Need psi again:
    • \psi = (c1)(X1)+ (c2)(X2 )+ (c3)(X3 )+…
  • F formula: F = \frac{MS}{MSWithin}, same omnibus error term.

Computation of Effect Sizes

  • Report a measure of effect size for any significant effect.
  • The most straightforward measure is eta-squared:
    • \eta^2 = \frac{SSM}{SST}