Why Compare Siblings?
A child’s genotype can be associated with an outcome for more than one reason. The child’s own alleles may affect the outcome directly, while relatives’ genotypes may shape the child’s environment. In a sibling study, a sibling’s genotype is one possible source of this family-mediated association. Parental alleles not transmitted to a child can also help characterize parental genetic nurture, as in the earlier material on nontransmitted alleles. These pathways are related, but they are not interchangeable: a sibling-genotype effect and a parental nontransmitted-allele association represent different exposures.
The design’s central advantage is that, conditional on parental genotypes, which alleles each child inherits is partly randomized by Mendelian segregation. Siblings therefore provide genetically informative contrasts within families. The design does not make every comparison automatically causal. Identification depends on which genetic variation is used, how the outcomes are modeled, and whether the assumptions about transmission and family environment are credible.
Representing Sibling Transmission
At one biallelic locus, let \(G_{fi}\) be the allele dosage for child \(i\) in family \(f\), coded as 0, 1, or 2 copies of the specified allele. Let \(P_f\) denote the parental genotypes. The expected dosage for a child, conditional on those parental genotypes, is \(m_f=E(G_{fi}\mid P_f)\). Define the child’s transmission deviation as \(W_{fi}=G_{fi}-m_f\). This separates the part of the child’s genotype predictable from the parents from the random deviation due to which alleles were transmitted.
For example, if both parents are heterozygous, each has one copy of the allele, so \(m_f=1\). A child with dosage 0 has \(W_{fi}=-1\), while a child with dosage 2 has \(W_{fi}=1\). Conditional on the parental genotypes, these deviations capture Mendelian segregation. At multiple loci, the same idea can be applied to allele scores, provided the expected transmissions and dependence among loci are handled appropriately.
A useful model for a focal child \(i\) and sibling \(j\) is
Here \(Y_{fi}\) is the focal child’s outcome; \(\beta\) is the direct effect per allele-dosage unit under the model; \(\gamma\) is the sibling-genotype effect per dosage unit; \(N_{fi}\) is a specified parental nontransmitted-allele measure; \(C_{fi}\) contains measured covariates; \(U_f\) represents shared family factors; and \(\varepsilon_{fi}\) is an individual residual. The coefficients describe conditional associations unless the design assumptions support a causal interpretation. In particular, including \(N_{fi}\) does not ensure that \(\beta\), \(\gamma\), and \(\theta\) can all be separately estimated: their genetic predictors may be dependent or linearly redundant.
Worked Example: From Parental Genotypes to Sibling Deviations
Suppose both parents in a hypothetical family are heterozygous at a locus. Their children have dosages \(G_{f1}=0\) and \(G_{f2}=2\). Find the expected dosage and transmission deviation for each child.
Each parent transmits the specified allele with probability \(1/2\), so the expected dosage is \(m_f=1/2+1/2=1\). Therefore,
The siblings’ deviations are equal in magnitude and opposite in direction in this example. This does not mean that the lower-dosage child causes the higher-dosage child’s outcome. It means that their inherited dosages differ from the same parental expectation. A comparison using these deviations can reduce confounding by differences in parental genotype, but causal interpretation still requires the design assumptions described below.
How the Sibling Model Separates Effects
For a focal child, the model asks whether that child’s outcome varies with the child’s own transmission deviation and with the sibling’s transmission deviation. In the simplified model, replace the genotypes by their deviations after accounting for parental genotypes and other modeled covariates:
Here \(Y_{fi}^{*}\) denotes an outcome adjusted for the specified covariates and parental-genotype expectation; this notation does not imply that every family-specific factor has been removed. If the focal child’s and sibling’s transmission deviations provide independent variation conditional on the parents, the two coefficients can be estimated separately in a suitable sample. The own-deviation coefficient targets the direct effect; the sibling-deviation coefficient targets the sibling-genotype association not explained by the focal child’s own genotype.
Worked Example: Estimating Own and Sibling Effects in a Toy Dataset
Consider four hypothetical focal-child observations after adjustment for parental genotype and measured covariates. The outcome values are on an arbitrary phenotype scale. Suppose the transmission deviations and adjusted outcomes are as follows.
| Family | Focal child \(W_{fi}\) | Sibling \(W_{fj}\) | Adjusted outcome \(Y_{fi}^{*}\) |
|---|---|---|---|
| 1 | -1 | -1 | -0.7 |
| 2 | 1 | -1 | 0.1 |
| 3 | -1 | 1 | -0.1 |
| 4 | 1 | 1 | 0.7 |
Fit the illustrative regression \(Y_{fi}^{*}=\beta W_{fi}+\gamma W_{fj}+\varepsilon_{fi}^{*}\) through the origin. The columns are orthogonal: \(\sum W_{fi}W_{fj}=0\), and each has sum of squares 4. Thus the least-squares coefficient for the focal child is
The sibling coefficient is
In this constructed example, a one-unit increase in focal-child transmission deviation is associated with a 0.4-unit increase in adjusted outcome, conditional on sibling deviation; the corresponding sibling coefficient is 0.3 outcome units per dosage-deviation unit. These calculations illustrate the model, not evidence from a real study. In actual data, uncertainty, covariate specification, and the validity of the assumptions must also be assessed.
Why a Family Fixed-Effect Contrast Can Be Insufficient
A tempting strategy is to compare both siblings’ outcomes and include a family fixed effect. This removes factors shared by siblings, but in a simple two-sibling model it also changes what can be identified. Suppose the two outcomes obey a symmetric model with direct effect \(\beta\), sibling effect \(\gamma\), and shared family component \(U_f\):
Subtracting the second equation from the first cancels the shared family component. It also shows exactly which coefficient combination the within-family contrast contains:
Thus, under this symmetric model, the difference alone identifies information about \(\beta-\gamma\), not \(\beta\) and \(\gamma\) separately. This is an identification limit, not a small-sample problem that disappears merely by collecting more two-sibling families. Separating the effects requires additional identifying variation and a model that uses it—for example, transmission deviations conditional on parental genotypes in a focal-outcome design, or other justified information that distinguishes the pathways.
Worked Example: Interpreting a Within-Family Difference
Suppose a hypothetical two-sibling comparison has \(G_{f1}-G_{f2}=2\) dosage units and \(Y_{f1}-Y_{f2}=0.2\) outcome units. Under the symmetric model above, ignoring the residual difference for this illustration, the contrast implies
outcome units per dosage unit. The contrast does not establish that \(\beta=0.1\) and \(\gamma=0\). For example, \(\beta=0.4\) and \(\gamma=0.3\) produce the same difference, as do \(\beta=0.2\) and \(\gamma=0.1\). The data contrast identifies their difference, not their individual values.
Where Nontransmitted Alleles Fit
For a focal child, parental nontransmitted alleles can provide information about parental genetic influences that are not carried by that child. If a parental score is divided into transmitted and nontransmitted components, an association with the nontransmitted component may be consistent with genetic nurture: parental genotype may influence the child’s environment. But a nontransmitted-allele coefficient is not automatically a sibling effect. It can reflect multiple parental pathways, residual population structure, assortative mating, or other sources of association if the design does not address them.
The relationship between these measures also matters. At a single locus, the parental allele total equals the allele transmitted to the focal child plus the parental allele not transmitted to that child. Consequently, if parental genotype totals are fixed in the analysis, transmitted and nontransmitted dosage at that locus are exact complements. A model cannot estimate separate coefficients for perfectly redundant predictors. At broader scores, dependencies may be less obvious but still affect precision and identification. The analysis should state which genetic contrasts supply independent variation and which pathways the coefficients represent.
Worked Example: A Nontransmitted Allele Is Not Automatically a Nurture Effect
Suppose both parents are heterozygous at one locus, so their combined allele dosage is 2. The focal child inherits dosage 0. The parental alleles not transmitted to that child then have dosage \(N_{fi}=2-0=2\). A sibling in the same family could inherit dosage 2.
The focal child’s nontransmitted dosage is therefore related to the alleles available for transmission to the sibling. An association between \(N_{fi}\) and the focal outcome could arise through parental behavior, through the sibling’s genotype and its effect on the focal child, or through other correlated pathways. To distinguish these interpretations, the analysis needs an explicit model for the sibling genotype and the relevant parental genetic measures; the nontransmitted-allele association alone does not identify which pathway operated.
Assumptions and Limits
A causal reading of sibling-genotype coefficients requires more than a within-family label. The key assumptions should be made explicit and considered in the study context:
- Correct transmission model. Conditional on parental genotypes, the modeled transmission deviations follow the appropriate Mendelian distribution. Genotyping error, unmodeled linkage, or incorrect parental phasing can undermine this representation.
- Conditional independence from family causes. Given parental genotypes and modeled covariates, the random transmissions used for identification are not associated with unmeasured causes of the focal outcome. This is the design’s source of leverage, not a guarantee that every family-level association has vanished.
- Appropriate outcome model. The model must represent the relevant direct and sibling pathways, including nonlinearity or interactions if they are scientifically important. A misspecified linear model may make its coefficients difficult to interpret.
- No unmodeled selection or interference structure. Participation, survival, or sample inclusion should not induce problematic associations between transmissions and outcomes. The meaning of a sibling effect also depends on which sibling is treated as the exposure and which outcome is measured.
- Adequate genetic variation and identification. The own- and sibling-genotype predictors must provide enough distinct variation to estimate their coefficients. A fixed-effect contrast that contains only \(\beta-\gamma\) cannot separate the two, regardless of nominal sample size.
Sibling designs also have practical limits. Shared household environments can be difficult to characterize, and sibling effects may differ by age, birth order, or developmental period. A coefficient averaged across families may not describe every family. Moreover, a genotype can affect more than one family member through multiple mechanisms; the estimated sibling coefficient is a summary for the specified outcome, exposure, and model, not a universal measure of “genetic influence.”
Common Mistake
Wrong reasoning: “A within-family analysis controls for family background, so the coefficient on a sibling’s genotype must be the sibling’s causal effect.”
Correction: A within-family comparison can reduce confounding from shared factors, but it does not automatically isolate each pathway. With two siblings and a symmetric own-plus-sibling model, differencing yields \(\beta-\gamma\), not separate estimates of \(\beta\) and \(\gamma\). A sibling coefficient has a causal interpretation only when the genetic variation used identifies that coefficient and the transmission, modeling, and selection assumptions are credible.
Check Your Understanding
Answer each question by naming the relevant contrast or assumption.
- If both parents are heterozygous at a locus, what are the expected dosage and transmission deviation for a child with dosage 2?
- In the focal-child model, what does the coefficient on the sibling’s genotype represent, conditional on the focal child’s genotype?
- Why does differencing two sibling outcomes in the symmetric model identify \(\beta-\gamma\) rather than both coefficients separately?
- Why is an association with parental nontransmitted alleles not, on its own, proof of a parental environmental mechanism?
- Name one assumption required to interpret transmission-deviation associations causally.