Reference Genotypes and Effect Contrasts

Why the Reference Genotype Matters

A genotype model needs a way to compare the trait across genotype groups. With indicator coding, one genotype is chosen as the reference genotype; the model represents the other genotypes by their differences from it. This choice determines how the coefficients are written and interpreted. It does not, by itself, change the data, the genotype-specific fitted values, or the underlying comparisons that can be made.

The reference genotype is a model baseline, not necessarily the most common, normal, or biologically preferred genotype. Choosing a different baseline can make a particular contrast easier to read, but the reference choice should not be confused with a claim about which genotype is biologically standard.

Reference-Coded Models for Genotype-Specific Means

Consider a locus with three possible genotypes, denoted \(AA\), \(Aa\), and \(aa\). Let \(Y\) be a quantitative trait and define the genotype-specific mean as \(\mu_g=E(Y\mid G=g)\), where \(G\) denotes genotype. For now, assume a model with an intercept and one indicator for each non-reference genotype. This is a saturated model for the three genotype means: it gives each genotype its own fitted mean.

If \(AA\) is the reference genotype, let \(I(G=Aa)\) equal 1 for individuals with genotype \(Aa\) and 0 otherwise, and define \(I(G=aa)\) in the same way. The model is:

For an \(AA\) individual, both indicators are 0, so the model gives \(E(Y\mid G=AA)=\beta_0\). For an \(Aa\) individual, the first indicator is 1 and the second is 0, so \(E(Y\mid G=Aa)=\beta_0+\beta_{Aa}\). For an \(aa\) individual, the fitted mean is \(\beta_0+\beta_{aa}\). Therefore:

The intercept is the mean for the reference genotype. Each other coefficient is a contrast: the mean for one genotype minus the mean for the reference. A positive coefficient means a larger expected trait value than the reference; a negative coefficient means a smaller expected value. These are differences in the trait’s units.

Changing the Reference Changes the Coefficients

The model can use \(Aa\) or \(aa\) as its reference instead. To see the effect, keep the same three genotype-specific means, \(\mu_{AA}\), \(\mu_{Aa}\), and \(\mu_{aa}\). If genotype \(r\) is the reference, then the intercept is \(\mu_r\), and the coefficient for any other genotype \(g\) is \(\mu_g-\mu_r\). The omitted reference category has no indicator coefficient; its contrast with itself is zero.

In the previous example, changing the reference from \(AA\) to \(Aa\) makes the intercept 14 mm. The coefficient for \(AA\) becomes \(11-14=-3\) mm, and the coefficient for \(aa\) becomes \(9-14=-5\) mm. The fitted means remain 11, 14, and 9 mm: only their representation as an intercept plus indicator coefficients has changed.

Genotype-Specific Risks as Contrasts

The same reference principle applies when the outcome is binary. Let \(Y=1\) indicate that an individual has a specified phenotype, and let \(Y=0\) indicate that the phenotype is absent. The genotype-specific risk is \(p_g=P(Y=1\mid G=g)\). Because the conditional mean of a binary outcome equals its probability of being 1, a model that represents genotype-specific risks directly has the same indicator form, with \(p_g\) in place of \(\mu_g\). In this response-scale model, the intercept is the reference-genotype risk and each indicator coefficient is a risk difference from that reference.

This interpretation depends on the response scale. A coefficient in a model whose response is the risk is a difference in risks. A coefficient in a model using a transformed response scale need not be a risk difference. The reference-genotype contrast idea remains useful, but the scale must be stated before interpreting a coefficient.

Why Re-Referencing Preserves Fitted Values

The invariance can be verified directly. Let the original reference be genotype \(r\), and write its fitted genotype means as \(\mu_g\). Under reference coding, the intercept is \(\mu_r\), and each non-reference coefficient is \(\mu_g-\mu_r\). Now choose a different reference \(s\). The new intercept is \(\mu_s\), and the coefficient for genotype \(g\ne s\) is \(\mu_g-\mu_s\). For any genotype \(g\ne s\), the new model adds its coefficient to the new intercept:

For genotype \(s\), all indicators are zero, so its fitted value is the new intercept \(\mu_s\). Thus every genotype retains its original fitted value. This algebra shows that a reference change is a re-expression of the same set of genotype-specific means, not a change to those means.

For example, if \(r=AA\) and \(s=Aa\), a coefficient that originally compared \(aa\) with \(AA\) becomes a coefficient comparing \(aa\) with \(Aa\). The new \(aa\) contrast is the old \(aa\)-versus-\(AA\) contrast minus the old \(Aa\)-versus-\(AA\) contrast:

The arithmetic changes because the comparison changes. The fitted genotype means do not. Inference about a particular contrast should likewise be framed as a test or interval for that contrast; a coefficient’s numerical value alone is meaningful only together with its reference and scale.

Common Mistakes and Practical Interpretation

• Calling the reference genotype the control group. The reference is a coding choice. It need not be an experimental control, the most frequent genotype, or the genotype without a phenotype.
• Interpreting an indicator coefficient as the genotype’s mean. The coefficient is the difference between that genotype’s mean and the reference mean. The mean itself is the intercept plus the coefficient.
• Comparing coefficient values from different references as if they were the same contrast. A coefficient always compares its genotype with the model’s current reference. Changing the reference changes what that coefficient compares.
• Confusing a risk difference with a percentage-point statement. A difference of 0.08 in risks is 8 percentage points, not an 8% relative increase. Relative comparisons require a different calculation.
• Treating an observed contrast as causal. A difference in genotype-specific sample means or risks is an association in the observed data. A causal interpretation needs additional assumptions and study design support.

When reporting a genotype model, state the reference genotype, the outcome scale, and the units or probability scale. For instance, “Using \(AA\) as the reference, the fitted mean for \(Aa\) was 3 mm higher” identifies the comparison and gives it a clear interpretation. If the baseline is changed, report the new reference rather than presenting coefficients without their coding context.