Back to List

Geometry and Effects in a Full Factorial Design: What All Combinations Separate

Explains how all combinations in 2² and 2³ full factorial designs, ±1 coding, balance, and orthogonality separate main effects from interactions.

Intermediate
|
30min
|
Verified (2026-08-14)
full factorial designbalanceorthogonalitymain effectinteractioncoded levelJMP
Progress0/28 (0%)

When two factors, A and B, are studied at low and high levels, four combinations result. These four points are not merely a list: they are the four vertices of a square. That symmetric structure separates main effects from interactions.

This unit has one question.


When every combination of factor levels is run, which effects can be viewed without mixing them together?
Full factorial designA design that includes all combinations of levels
balanceEach level and combination appears with equal frequency
orthogonalitySeparate structure where the product of effect columns is 0
main effectHigh-low differences averaged across different factor levels

A 2² full factorial design includes all four vertices

When low is coded −1 and high +1, the design table has these columns.

ABAB
−1−1+1
+1−1−1
−1+1−1
+1+1+1

The AB column is A×B. Each of A, B, and AB contains equal numbers of +1 and −1 values, so the columns are balanced. The sum of elementwise products for any two different columns is zero. That is orthogonality in this setting.

U17 · Figure 01
Orthogonal designs estimate main effects and interactions separately
A effect80B effect50AB effect60
The two-level effect is high-level mean−low-level mean. The regression coefficient is half the effect at ±1 coding.

Because of orthogonality, the +/− values of B and AB cancel in a balanced way when the A effect is calculated. An estimate of one effect is not pulled by the true value of another. Missing runs, unequal weights, or execution failures can break this perfect orthogonal structure.

A main effect is the difference between the high-level and low-level means

A effect = mean(Y | A=+1) − mean(Y | A=−1)

Because this averages over both B levels, it is the average change associated with A. The B effect is defined the same way. The AB effect is the mean at the two vertices where AB is +1 minus the mean at the two vertices where AB is −1.

If the ±1-coded regression equation is ŷ=b₀+bAA+bBB+bABAB, each regression coefficient is half of its factorial effect. Moving A from −1 to +1 spans two coded units.

Do not erase an interaction after looking only at main effects

When AB is large, the average main effect of A merely averages effects under two different B conditions. Interpret an interaction plot and condition-specific simple effects together. Removing related main effects from a model with an interaction breaks hierarchy.

Nonparallel lines show dependence on conditions

Connect the mean at A low and A high separately for B=low and B=high.

  • parallel lines: AB effect is 0
  • different slopes: interaction
  • crossing lines: the direction of the A effect can reverse by B level

Slightly nonparallel lines do not by themselves establish a practically important interaction. Estimate pure error through independent replication, then inspect effect CIs, ANOVA, and residuals. In a saturated 2² run once at every combination, fitting all effects leaves no residual degrees of freedom.

A 2³ design has eight vertices of a cube

Three two-level factors produce 8 runs and effects A, B, C, AB, AC, BC, and ABC. When all combinations are observed, these seven effect columns are orthogonal. With k factors, the number of runs grows quickly as 2^k. That growth motivates fractional factorial designs in U20, but we must first understand what the full design separates in order to understand the cost of saving runs.

In-Silico Lab: match hand calculations to regression coefficients

The Lab creates all four vertices of Y=70+8A+5B+γAB.

  1. Verify that the A effect is 16 and the B effect is 10.
  2. With γ=6, see why the AB effect is 12 and the regression coefficient is 6.
  3. Calculate the +/− sums in each column and the cross-products between columns to confirm balance and orthogonality.
  4. Consider what problem occurs if one of the four responses is removed.
In-Silico Lab · U17

2² Restore effects directly from the design table

Compare the A·B·AB effects and ±1 coded regression coefficients for the four combinations of responses.

If this is your first time: What should I press?
  1. 1. Read the question firstIn the Lab title, check the one thing you will compare this time.
  2. 2. Change just one conditionInitially, change only one of the inputs: n, effect, or spread.
  3. 3. new composite specimen pressureNew synthetic data is created. The same conditions may vary depending on the sample.
  4. 4. Pictures and calculation results CompareWrite in one sentence what moves and what stays the same before and after the change.

If it gets stuckresetGo back to see the default results and change just one condition. This Lab is not a correct answer tester but a pattern observation tool.

Synthetic observations of the same settings

runABABY
1-1-1163.0
2-11-161.0
31-1-167.0
411189.0

calculation result

A effect16.0
B effect10.0
AB effect12.0
coded expression70 + 8A + 5B + 6AB

The sign of the effect can be interpreted by writing down high−low together with the good direction of the response.

educational synthetic modelbjs-factorial-sequence-v1. Actual research judgments must separately reflect experimental units, missingness, distribution, multiplicity, pre-planning, and domain criteria.

JMP output connects the design table and effect plots

Design Table

−1/+1 combination, check balance and number of runs.

Effect Estimates

The difference between the high-level average and the low-level average is read as an effect.

Interaction / Cube Plot

Check vertex responses and non-parallelities in space.

Do not confuse coded levels in a Design Table with physical units. Effect Estimates show direction according to the coding convention. Cube Plots and Interaction Plots are mean summaries of raw data, so also inspect the spread of replicate points and residuals.

Example result statement

A 2² full factorial design of temperature (A: 30/37°C) and pH (B: 6.8/7.4) was independently replicated in two batches. Using high−low contrasts, the temperature effect was +16.2, the pH effect +9.8, and the AB interaction effect +11.5 units. The positive AB effect indicated that the pH effect was larger at high temperature. The coded and natural-unit equations, run order, and residual diagnostics were retained together.

At the end of this unit

  • A full factorial design includes every combination of factor levels.
  • ±1 coding makes balance and orthogonality explicit.
  • A main effect is the high−low difference averaged over the other factors.
  • A coded regression coefficient is half of a two-level factorial effect.
  • When an interaction is large, interpret effects by condition.
  • A saturated design without replication has no residual degrees of freedom.
Full factorial designs separate main effects and interactions thanks to all combinations and orthogonal structures. The effect code is meaningful only when it is written together with the coding and response definition.

The next unit explains why the quality of evidence can differ with run order, replication, and blocks even when the same four combinations are run.

Official supplementary resources

The design and Lab in this article use synthetic educational data.

💬 Questions & Comments

0 comments

You can post without signing in. Guest comments cannot be edited or deleted by their author.

0/2000

Loading...