Why Aliasing Matters
A factorial model separates variation into main effects and interactions. That separation relies on the design containing enough distinct combinations of factor levels. When a combination is missing, a main-effect pattern may become indistinguishable from an interaction pattern on the combinations that remain. The data can then determine the fitted means without determining each named effect separately.
Aliasing is a property of the design and the model columns, not a conclusion drawn from whether an effect happens to be statistically significant. It is useful to check it before interpreting individual terms: a coefficient may appear in software output even when the observed design cannot uniquely distinguish it from another coefficient.
Model and Meaning of Aliasing
Consider two factors, \(A\) and \(B\), each with two levels. Code the levels of \(A\) by \(x=-1,+1\) and the levels of \(B\) by \(z=-1,+1\). For an observation in a cell, use the full two-factor model
Here \(\mu\) is the intercept, \(\alpha\) and \(\beta\) are the coded main-effect coefficients, and \(\gamma\) is the interaction coefficient. The four possible cell means are obtained by substituting the four combinations of \(x\) and \(z\). This model represents any four cell means when all four combinations are observed.
For a statistical analysis, suppose the design points are fixed and the errors are independent with mean zero and common variance \(\sigma^2\). Normal errors, \(\varepsilon\sim N(0,\sigma^2)\), are additionally assumed when normal-theory inference is used. The structural question of whether columns are aliased depends on which combinations are observed; it does not require normality.
The interaction column has value \(xz\) in a row with codes \(x,z\). The design matrix is formed by placing the columns \(1,x,z,xz\) side by side for the observed rows. If an exact identity holds among these columns on those rows, it may not hold on the full set of four possible combinations. That difference is central: an identity restricted to observed cells can arise solely because a cell is empty.
Complete Design: Main Effects and Interaction Are Separable
Worked Example: Reading Effects From Four Cell Means
Suppose a hypothetical experiment has the following four observed cell means, in outcome units. The values are \(\bar{x}\)-style sample summaries here only in the ordinary sense of observed averages; they are not claimed to be population means.
| \(x\) | \(z\) | Observed cell mean |
|---|---|---|
| -1 | -1 | 10 |
| -1 | +1 | 14 |
| +1 | -1 | 12 |
| +1 | +1 | 20 |
With all four combinations present, the coded coefficients are contrasts of the cell means. In particular, the intercept is their average:
The coefficient for \(x\) is the average of the two means at \(x=+1\) minus the average of the two at \(x=-1\), divided by 2:
Similarly, the coefficient for \(z\) is
Finally, the interaction coefficient is the difference between the two changes across \(A\), divided by 4. At \(z=-1\), the change across \(A\) is \(12-10=2\); at \(z=+1\), it is \(20-14=6\). Therefore,
The model reproduces each cell mean. For example, at \((x,z)=(-1,-1)\), it gives \(14-2-3+1=10\); at \((+1,+1)\), it gives \(14+2+3+1=20\). All four columns are distinguishable because the complete set of combinations supplies all four different rows of the design.
This example also fixes a useful interpretation check: compare the same factor across its levels while holding the other factor fixed. Comparing \(12-10\) and \(20-14\) compares changes across \(A\), each at a fixed level of \(B\). Comparing cell means in a different direction answers a different question.
One Empty Cell Can Create Complete Aliasing
Now suppose the \((+1,+1)\) combination is absent. On the three observed combinations \((-1,-1),(-1,+1),(+1,-1)\), the interaction column satisfies an exact identity:
For example, at \((-1,-1)\) the right side is \(-1+1+1=1\), which equals \(xz\). The identity also holds at each of the other two observed combinations. It does not hold at the missing \((+1,+1)\) combination: there \(xz=1\), whereas \(-1-x-z=-3\).
Substitute the identity into the model on the observed rows:
Thus, on the observed design, the fitted values depend on \(\mu-\gamma\), \(\alpha-\gamma\), and \(\beta-\gamma\), rather than on all four coefficients separately. The interaction column can be represented using the intercept and both main-effect columns. This is complete aliasing: no amount of replication in just those same three combinations breaks that exact column relationship.
Worked Example: An Empty Combination and Indistinguishable Coefficients
Suppose the hypothetical observed means are 10 at \((-1,-1)\), 14 at \((-1,+1)\), and 12 at \((+1,-1)\), with no observations at \((+1,+1)\). Using the substituted model, write \(C_0=\mu-\gamma\), \(C_x=\alpha-\gamma\), and \(C_z=\beta-\gamma\). The three observed means give
Subtracting the first equation from the second gives \(2C_z=4\), so \(C_z=2\). Subtracting the first from the third gives \(2C_x=2\), so \(C_x=1\). Substitution into the first equation gives \(C_0-1-2=10\), hence \(C_0=13\).
These equations determine the three combinations, but not \(\mu,\alpha,\beta,\gamma\) individually. Let \(t\) be any real number and set \(\gamma=t\). Then the coefficients
give the same observed means for every value of \(t\). For example, at \((-1,-1)\) the prediction is \((13+t)-(1+t)-(2+t)+t=10\); at \((-1,+1)\) it is \((13+t)-(1+t)+(2+t)-t=14\); and at \((+1,-1)\) it is \((13+t)+(1+t)-(2+t)-t=12\). The observed data therefore cannot select one value of \(\gamma\) from this family. The missing cell is not just another cell with a large standard error: the full-model coefficient decomposition is structurally ambiguous.
Partial Aliasing: Unequal Replication Without an Empty Cell
A complete set of cells prevents the exact aliasing in the previous example, but unequal replication can make columns nonorthogonal. Nonorthogonality means that the column values do not have zero cross-product; changes associated with one term are consequently linked to changes associated with another in the fitted analysis. It does not, by itself, mean the terms are inseparable.
Worked Example: Correlated Columns in an Unbalanced Complete Design
Consider a hypothetical design with all four combinations observed, but with replication counts \(n_{--}=4\), \(n_{-+}=1\), \(n_{+-}=1\), and \(n_{++}=2\). The subscripts give \((x,z)\), so there are \(N=8\) observations. In the design matrix, the cross-product of the \(x\) and \(xz\) columns is
Each of these columns has squared length 8, since each entry is either \(-1\) or \(+1\). Their uncentered normalized cross-product is therefore \(-2/\sqrt{8\cdot8}=-0.25\), not zero. The columns are not orthogonal.
However, they are not exactly aliased. All four unique \((x,z)\) rows are present. If a linear combination \(c_0+c_1x+c_2z+c_3xz\) equals zero at all four combinations, adding and subtracting those four equations gives \(4c_0=0\), \(4c_1=0\), \(4c_2=0\), and \(4c_3=0\). Thus every coefficient \(c_j\) is zero, so there is no nontrivial exact column dependence. Unequal replication creates nonorthogonality here, not complete aliasing.
This distinction matters in interpretation. With partial aliasing, the terms remain separately represented by the design, although their estimates can be less independent and their precision can be affected. With complete aliasing, the data cannot distinguish the involved coefficients at all under the specified model.
A Short Design-Column Justification
The distinction can be checked directly from the model matrix. In a complete two-by-two design, the columns for \(1,x,z,xz\) take the values shown below. Repeated observations repeat rows, but cannot create a new dependence if all four rows are already present.
| \(x\) | \(z\) | Intercept \(1\) | Main effect \(x\) | Main effect \(z\) | Interaction \(xz\) |
|---|---|---|---|---|---|
| -1 | -1 | 1 | -1 | -1 | 1 |
| -1 | +1 | 1 | -1 | +1 | -1 |
| +1 | -1 | 1 | +1 | -1 | -1 |
| +1 | +1 | 1 | +1 | +1 | 1 |
One way to verify that these four columns are independent is to suppose \(c_0+c_1x+c_2z+c_3xz=0\) for every row. Add the four row equations: each signed column sums to zero, leaving \(4c_0=0\). Multiply each row equation by \(x\) before adding; the intercept, \(z\), and \(xz\) contributions cancel, leaving \(4c_1=0\). Multiplying instead by \(z\) gives \(4c_2=0\), and multiplying by \(xz\) gives \(4c_3=0\). Hence all four coefficients vanish. If one row is removed, this argument no longer supplies four independent equations; the exact identity among columns may then appear.
This check is a design-level justification, not a test of whether an effect is present. Replicating observed cells can improve information about their means and reduce error, but it cannot recover a missing pattern when the columns remain dependent on every replicated row.
Common Mistakes and Practical Checks
- Calling every nonzero column cross-product complete aliasing. A nonzero cross-product indicates nonorthogonality. Complete aliasing requires an exact linear dependence. Check whether a nontrivial linear combination of columns is zero on every observed row.
- Interpreting a coefficient as separately determined because software prints it. In a rank-deficient design, software may use a constraint or a particular parameterization to produce output. That convention does not make the data distinguish coefficients that have identical effects on observed fitted values.
- Assuming more replication repairs an empty-cell alias. Replication repeats existing design rows. If \(xz=-1-x-z\) on every observed row, repeating those rows preserves the identity.
- Comparing the wrong cell means when describing an effect. To describe the change across \(A\), hold \(B\) fixed. For the complete-design example, those changes are \(12-10=2\) at \(z=-1\) and \(20-14=6\) at \(z=+1\); their difference is \(4=4\hat{\gamma}\).
Check Your Understanding
Use the coded two-factor model and distinguish exact dependence from nonorthogonality.
- If the \((+1,+1)\) cell is missing, verify \(xz=-1-x-z\) at each of the three observed combinations.
- For that same missing-cell design, which coefficient combinations appear in the substituted model: \(\mu+\gamma\), \(\alpha+\gamma\), and \(\beta+\gamma\), or \(\mu-\gamma\), \(\alpha-\gamma\), and \(\beta-\gamma\)?
- Explain why adding observations to each of the three nonempty cells does not remove complete aliasing.
- In a complete design with unequal replication, why does a nonzero cross-product between two columns not prove that their coefficients are completely aliased?
- At \(z=-1\), the cell means are 10 at \(x=-1\) and 12 at \(x=+1\); at \(z=+1\), they are 14 at \(x=-1\) and 20 at \(x=+1\). Find each change across \(A\) and their difference.