Why Genotype-Cell Counts Matter
A two-locus interaction model asks whether the joint pattern across two loci is adequately described by the separate effects of each locus. That question depends not only on the model formula but also on which genotype combinations are observed. If a combination is absent, the data cannot directly reveal its outcome mean. If it is rare, its mean may be estimable under a suitable model but imprecisely estimated. These are distinct problems: lack of identifiability and high uncertainty.
For two loci with three genotype categories each, the observations can be arranged in a \(3\times3\) table. The entries are counts of individuals in each joint genotype cell. A cell with count zero provides no outcome observations for that combination. A cell with a small positive count provides some information, but often much less than the other cells. Neither the total sample size nor a balanced-looking marginal count guarantees adequate support for every joint combination.
From Cell Counts to Estimability
Consider first a saturated model for a quantitative outcome. Let \(\mu_{ij}\) denote the mean outcome in genotype cell \((i,j)\), where \(i\) and \(j\) index the genotype categories at the two loci. A saturated cell-mean model assigns a separate mean to every possible joint genotype. If each cell is observed, these means can be estimated directly by the corresponding sample means. If a cell is empty, its sample mean does not exist, and its cell-specific mean cannot be estimated from the data without additional assumptions.
Many interaction questions concern contrasts among cell means rather than a single cell mean. For a \(2\times2\) subset of genotype categories, the difference-in-differences contrast is \(C=\mu_{11}-\mu_{10}-\mu_{01}+\mu_{00}\). It measures how much the change across the first locus differs between the two categories of the second locus, on the outcome-mean scale. If any of these four means is unavailable in a saturated model, this contrast is not directly estimable.
With independent observations and a common within-cell variance \(\sigma^2\), the sample mean in a cell with count \(n_{ij}\) has variance \(\sigma^2/n_{ij}\). Thus, the variance of a contrast is governed by the counts in the cells it uses—not simply by the total sample size. For independent cells, a contrast with weights \(c_{ij}\) has variance \(\sigma^2\sum c_{ij}^2/n_{ij}\). This result explains why a small count can make a contrast imprecise even when all its cell means are estimable.
Worked Example: A Difference-in-Differences With Unequal Cell Counts
Suppose a hypothetical quantitative-trait study observes four genotype cells with counts \(n_{00}=20\), \(n_{10}=35\), \(n_{01}=30\), and \(n_{11}=40\). The respective sample means are \(10\), \(12\), \(11\), and \(16\) trait units. Assume independent individuals and a common within-cell standard deviation of \(2\) trait units. These assumptions are stipulated for this calculation; in an actual analysis they would need scientific and diagnostic support.
The estimated interaction contrast is \(\hat{C}=16-12-11+10=3\) trait units. Each coefficient in the contrast is either \(+1\) or \(-1\), so its square is 1. The estimated standard error is therefore
For clarity, the sum of reciprocals is approximately \(0.136905\); multiplying by the variance \(2^2=4\) gives \(0.547619\), not \(0.4476\). The estimated contrast is about \(3/0.7400=4.054\) standard errors from zero under the stated model. This calculation illustrates precision, not a general claim that the assumptions hold or that the observed association is causal. The cell with the smallest count contributes the largest reciprocal-count term and therefore contributes most to the variance.
Design-Matrix Rank: A General Diagnostic
Cell counts give an immediate view of support, but parameter identifiability is formally a property of the design matrix. In a linear model \(Y=X\beta+\varepsilon\), \(X\) is the matrix of predictors, \(\beta\) is the coefficient vector, and the usual least-squares coefficients are uniquely identified only when \(X\) has full column rank. If columns are linearly dependent, distinct coefficient vectors can produce the same fitted values.
A saturated two-locus cell-mean model can be represented with one indicator column per genotype cell. Its design-matrix rank is the number of occupied cells: each observed cell contributes a distinct indicator pattern, while an empty cell contributes no row pattern at all. Consequently, a saturated model cannot identify the mean of an empty cell. More generally, if a model uses fewer parameters than there are cells, its rank must still be checked; a sparse or empty cell does not by itself determine whether every coefficient in that reduced model is estimable.
Worked Example: One Empty Cell in a Saturated Genotype Table
Suppose a hypothetical study of two loci has the following counts. Rows denote the first locus’s three genotype categories and columns denote the second locus’s three categories.
| Second locus: 0 | Second locus: 1 | Second locus: 2 | |
|---|---|---|---|
| First locus: 0 | 18 | 7 | 0 |
| First locus: 1 | 12 | 4 | 2 |
| First locus: 2 | 5 | 1 | 9 |
Eight of the nine cells are occupied. In a saturated cell-mean model with nine cell indicators, the design matrix has rank 8, not 9. The mean for the \((0,2)\) cell is not directly estimable because that cell contains no individuals. A contrast requiring that mean, such as a difference between it and another cell mean, is also not estimable without further modeling assumptions.
This does not mean that every possible interaction contrast is unidentified. A contrast involving only observed cells can be estimated from their data. Nor does fitting a lower-dimensional model automatically solve the information problem: it may make a prediction for the empty cell by imposing restrictions, such as an additive structure, but that prediction is then supported by those restrictions rather than by observations in that cell. The analyst must state what the model assumes and avoid presenting an assumption-based prediction as a directly observed cell mean.
Why a Reduced Model Can Still Fail
The rank criterion is especially useful when genotype variables are coded as predictors rather than as cell indicators. For two binary indicators \(A\) and \(B\), a model with an intercept, both main effects, and their product has design columns \(1,A,B,AB\). The interaction coefficient is separately identifiable only if the observed rows provide variation sufficient to distinguish all four columns. An empty joint category can make the product column redundant or identically zero.
Worked Example: An Empty Joint Category and a Product Term
Suppose a hypothetical sample uses binary indicators for two genotype categories. The model is \(Y=\beta_0+\beta_AA+\beta_BB+\beta_{AB}AB+\varepsilon\). Assume that only the combinations \((A,B)=(0,0),(1,0),(0,1)\) are observed; the \((1,1)\) combination is empty.
The distinct design rows are \((1,0,0,0)\), \((1,1,0,0)\), and \((1,0,1,0)\), in the column order \(1,A,B,AB\). The fourth column is zero for every observed individual, so the design matrix has rank 3 rather than 4. Changing \(\beta_{AB}\) has no effect on any fitted value. Therefore, the interaction coefficient is not identifiable from these data.
A model that removes the product term can estimate the intercept and the two main-effect coefficients, provided the remaining columns have full rank. But it does so by imposing no interaction; it does not demonstrate that the interaction is zero. This distinction matters whenever an analysis reports a null interaction estimate despite there being no observed joint genotype combination that could inform it.
Rare Cells: Identifiable Does Not Mean Precise
An occupied cell may preserve full rank while still creating an unstable estimate. In the saturated quantitative model, the cell mean’s standard error under the common-variance assumption is \(\sigma/\sqrt{n_{ij}}\). As \(n_{ij}\) decreases, that standard error increases. For a contrast, the contribution from a cell is proportional to \(c_{ij}^2/n_{ij}\), so a rare cell with a nonzero contrast weight can dominate uncertainty.
A positive count can therefore make a coefficient mathematically estimable without making it useful for precise inference. Estimates based on very few observations may be sensitive to individual outcomes, and standard errors may be large. In models for binary outcomes, sparse genotype cells can also contribute to separation or near-separation, where estimated regression coefficients become extremely large or ordinary large-sample approximations perform poorly. The exact consequences depend on the outcome model and should not be inferred from counts alone.
The standard-error calculation in the earlier worked example assumed independent observations and a common within-cell variance. If the outcome variances differ across cells, or observations are correlated, that formula need not apply as written. The analyst must use a variance estimator and model appropriate to the design. Counts remain valuable diagnostics, but they do not replace checking the model assumptions.
A Practical Assessment
A defensible assessment combines the cell table with a rank check for the actual fitted model. First tabulate joint genotypes, including zero-count combinations; marginal genotype frequencies cannot reveal every sparse joint cell. Then identify the interaction contrasts or coefficients being targeted and determine which cells inform them. Finally, inspect the rank of the design matrix and the uncertainty of estimable quantities. A model that is full rank can still be poorly conditioned, so numerical rank and standard errors should be considered alongside the raw counts.
Show all possible combinations relevant to the coding, including empty cells. Do not rely only on the separate allele or genotype frequencies at each locus.
Name the cell means, interaction coefficient, or contrast of interest. A sparse cell matters directly when the target depends on it.
Determine whether the design matrix identifies the target, then assess standard errors and sensitivity. Rank deficiency indicates non-identifiability; large standard errors indicate weak precision, not the same failure.
If a reduced model supplies predictions for unobserved combinations, state the assumptions that permit those predictions and distinguish them from direct cell-specific evidence.
Common Mistake and Key Takeaway
A common mistake is to say, “The interaction is estimable because the model software returned a coefficient,” or, conversely, “One cell is rare, so the interaction is impossible to estimate.” Neither conclusion follows from the cell count alone. A coefficient may be aliased or otherwise non-identifiable even if software prints a value, and a rare but observed cell may permit estimation with high uncertainty. The correct diagnosis uses the specified design matrix and target contrast, then considers the precision and assumptions behind the estimate.
Check Your Understanding
Answer each question by referring to the model and target quantity, not only to the total sample size.
- In a saturated cell-mean model, what happens to the design-matrix rank when one of nine genotype cells is empty?
- Why can an interaction contrast involving an empty cell fail to be estimable even if the other cells have large counts?
- Under independent observations and common within-cell variance, how does a cell’s count affect its contribution to the variance of a contrast?
- In the model with columns \(1,A,B,AB\), what does an all-zero \(AB\) column imply about the interaction coefficient?
- Explain why a reduced additive model can predict an outcome for an empty joint genotype cell without making that prediction directly supported by observations in that cell.