Statistics 2nd ed

statistical-interaction

Factorial ANOVA

factorial layout
factorial means interaction
factorial interaction

Goal. Test the effects of Method (Lecture vs. Online) and Time (Early vs. Late) on exam scores, and whether there is an interaction between Method and Time.

Design & Experiment

  • Factor A (Method): Lecture vs. Online
  • Factor B (Time): Early vs. Late
  • Balanced design: \(n=5\) per cell ⇒ total \(N=20\).

Students are randomly assigned to one of four cells (Method × Time). After a short module, all students take the same 100-point exam.

Figure 1: 2 × 2 layout (Method × Time).


Data

Scores by cell (five students per cell):

MethodTimeScoresCell Mean
LectureEarly686870727270.0
LectureLate767678808078.0
OnlineEarly707072747472.0
OnlineLate717173757573.0

Within each cell the sample variance is 4 (SD = 2), so the within-cell sum of squares is \((n-1)s^2 = 4\times4 = 16\) per cell.

Figure 2: Means with SEM by Time, separate lines for Method.

Figure 3: Interaction plot (Lecture rises sharply; Online nearly flat).


Step 1 — Marginal Means and Grand Mean

Cell means: \[ \bar X_{\text{Lecture,Early}}=70,\; \bar X_{\text{Lecture,Late}}=78,\; \bar X_{\text{Online,Early}}=72,\; \bar X_{\text{Online,Late}}=73. \] Marginal means: \[ \bar X_{\text{Lecture}}=\frac{70+78}{2}=74,\quad \bar X_{\text{Online}}=\frac{72+73}{2}=72.5; \qquad \bar X_{\text{Early}}=\frac{70+72}{2}=71,\quad \bar X_{\text{Late}}=\frac{78+73}{2}=75.5. \] Grand mean: \[ \bar X=\frac{70+78+72+73}{4}=73.25. \]


Step 2 — Sums of Squares (Between)

Balanced design formulas (with \(n\) per cell, \(a=b=2\)):

  • \(SS_A = nb \sum_a(\bar X_{a\cdot}-\bar X)^2\), here \(nb=10\).
  • \(SS_B = na \sum_b(\bar X_{\cdot b}-\bar X)^2\), here \(na=10\).
  • \(SS_{AB} = n \sum_{a,b}\big(\bar X_{ab}-\bar X_{a\cdot}-\bar X_{\cdot b}+\bar X\big)^2\), here \(n=5\).

Compute each term:

Factor A (Method): \[ \begin{aligned} SS_A &= 10\Big[(74-73.25)^2 + (72.5-73.25)^2\Big]\\ &= 10\big[0.75^2 + (-0.75)^2\big] = 10(0.5625+0.5625)=\mathbf{11.25}. \end{aligned} \]

Factor B (Time): \[ \begin{aligned} SS_B &= 10\Big[(71-73.25)^2 + (75.5-73.25)^2\Big]\\ &= 10\big[(-2.25)^2 + (2.25)^2\big] = 10(5.0625+5.0625)=\mathbf{101.25}. \end{aligned} \]

Interaction \(A\times B\): For each cell compute \(d_{ab}=\bar X_{ab}-\bar X_{a\cdot}-\bar X_{\cdot b}+\bar X\). Here each \(d_{ab}=\pm1.75\) so \(d_{ab}^2=3.0625\) and there are four cells: \[ SS_{AB}=5\times(4\times3.0625)=\mathbf{61.25}. \]


Step 3 — Within-Group (Error) and Total SS

Within each cell, \((n-1)s^2=16\). With 4 cells: \[ SS_{\text{within}}=\mathbf{64.00}. \]

Total: \[ SS_{\text{total}}=SS_A+SS_B+SS_{AB}+SS_{\text{within}} =11.25+101.25+61.25+64.00=\mathbf{238.75}. \]


Step 4 — Degrees of Freedom & Mean Squares

\[ \begin{aligned} &df_A=a-1=1,\quad df_B=b-1=1,\quad df_{AB}=(a-1)(b-1)=1,\\ &df_{\text{within}}=N-ab=20-4=\mathbf{16},\quad df_{\text{total}}=N-1=19. \end{aligned} \] \[ MS_A=\frac{11.25}{1}=11.25,\quad MS_B=\frac{101.25}{1}=101.25,\quad MS_{AB}=\frac{61.25}{1}=61.25,\quad MS_{\text{within}}=\frac{64.00}{16}=\mathbf{4.00}. \]


Step 5 — F Tests & p-values

\[ F_A=\frac{MS_A}{MS_{\text{within}}}=\frac{11.25}{4}= \mathbf{2.8125},\qquad F_B=\frac{MS_B}{MS_{\text{within}}}=\frac{101.25}{4}= \mathbf{25.3125},\qquad F_{AB}=\frac{MS_{AB}}{MS_{\text{within}}}=\frac{61.25}{4}= \mathbf{15.3125}. \] With \(df_1=1\), \(df_2=16\): \[ p_A \approx 0.11\;(\text{n.s.}),\quad p_B < 0.001,\quad p_{AB} \approx 0.001. \]


ANOVA Summary Table

SourceSSdfMSFp
Method (A)11.25111.252.8125≈ 0.11
Time (B)101.251101.2525.3125< 0.001
A × B61.25161.2515.3125≈ 0.001
Within (Error)64.00164.00
Total238.7519

Interpretation

Main effect of Time (B) is significant: Late > Early on average. Main effect of Method (A) is not significant at conventional levels. The interaction (A × B) is significant: Lecture improves markedly from Early→Late, while Online changes little—non-parallel lines in the interaction plot.

Figure 4: Interaction plot highlighting non-parallel lines.

Assumptions (checklist)

  • Independence of observations within and across cells.
  • Approximately normal scores within each cell.
  • Homogeneity of variances across cells (here, each cell variance ≈ 4).

Practice self-test quiz

In the space below, please find practice problems and self-test quizzes. For full access, please signup free.

Lecture 7 — Factorial Designs (Two-way ANOVA)

2x2 factorial layout
2x2 interaction
2x2 anova summary table

A factorial design includes two or more factors studied at once.
This allows us to test not only the effect of each factor separately, but also whether the factors interact.


Example: 2 × 2 Design

  • Factor A: Teaching method (Lecture, Online)
  • Factor B: Time of day (Morning, Afternoon)

This design has 4 groups (2 levels of A × 2 levels of B).

We can test:

  1. The main effect of Factor A (method).
  2. The main effect of Factor B (time).
  3. The interaction between method and time.

The ANOVA Partition

For a 2 × 2 design:

  • Main effect A: $$df_A = a - 1$$
  • Main effect B: $$df_B = b - 1$$
  • Interaction A × B: $$df_{A \times B} = (a - 1)(b - 1)$$
  • Error (within): $$df_{\text{within}} = N - ab$$

Where $$a$$ = levels of Factor A, $$b$$ = levels of Factor B, $$N$$ = total number of observations.


Interaction

An interaction occurs when the effect of one factor depends on the level of the other factor.

  • If lines in a plot are parallel, there is no interaction.
  • If lines cross or diverge, there is an interaction.

Example

Suppose means are:

  • Lecture: Morning = 70, Afternoon = 90
  • Online: Morning = 80, Afternoon = 80

Here:

  • Main effect of method: Online > Lecture overall
  • Main effect of time: Afternoon > Morning overall
  • Interaction: Lecture scores rise with time, Online scores stay flat → non-parallel lines.

Visuals

Figure L7.1 — Factorial Layout (2 × 2). A 2 × 2 grid: Method × Time.

Figure L7.2 — Interaction Plot. Lecture line slopes upward, Online line flat. Caption: “Lines not parallel = interaction.”

Figure L7.3 — ANOVA Summary Table for 2 × 2 design. Source | SS | df | MS | F | p.


Why This Matters

Factorial designs let us test more than one factor at a time.
They are efficient and powerful, and the concept of interaction is central in science.
Two-way ANOVA is the foundation for more complex designs, including repeated measures and mixed ANOVA.

Practice self-test quiz

In the space below, please find practice problems and self-test quizzes. For full access, please signup free.