Why Pairwise Checks Have a Limit
When several variables are incomplete, chained equations specify a conditional model for each variable given the others. It is natural to review these models one at a time: does each include relevant predictors, represent the intended nonlinear terms, and preserve the analysis relationship? Such checks are useful, but they do not answer a further question: could all the specified conditionals be conditional distributions of one common joint distribution?
That distinction matters when an analysis contains several predictors, interactions, or nonlinear terms. A set of conditional models can look reasonable in separate pairwise comparisons while imposing requirements that cannot all hold together. Alternatively, the conditionals can agree on pairwise relationships but fail to reveal a higher-order dependence that affects a joint interaction. Compatibility is a property of the collection, not a conclusion obtained by adding up separate pairwise endorsements.
A full conditional for \(X\), for example, is the distribution of \(X\) given all the other variables in the system, such as \(Y\) and \(Z\). A pairwise conditional or association examines fewer variables, for example the relationship between \(X\) and \(Y\) without conditioning on \(Z\). Pairwise summaries do not generally determine full conditionals. Even if every individual conditional regression has sensible fitted values and includes familiar terms, the collection may fail a joint consistency requirement.
A Gaussian Compatibility Constraint
A tractable setting makes the issue precise. Suppose \(X_1,\ldots,X_p\) are intended to have a nonsingular multivariate normal joint distribution with mean vector \(\boldsymbol{\mu}\) and positive-definite precision matrix \(\Omega\), the inverse of the covariance matrix. The full conditional distribution of \(X_j\) given the other variables is normal. Write its mean as an intercept plus slopes \(b_{jk}\) on each \(X_k\), \(k\ne j\), and its variance as \(\tau_j^2\).
Here \(X_{-j}\) denotes all variables other than \(X_j\). The formula follows by collecting the terms involving \(X_j\) in the multivariate normal log density. The quadratic term in \(X_j\) has coefficient \(-\Omega_{jj}/2\), and the linear terms involving another variable have coefficient \(-\Omega_{jk}X_k\). Completing the square gives conditional variance \(1/\Omega_{jj}\) and slope \(-\Omega_{jk}/\Omega_{jj}\).
Because a precision matrix is symmetric, the slope for \(X_k\) in the conditional model for \(X_j\) cannot be chosen independently of the slope for \(X_j\) in the conditional model for \(X_k\). Substitution gives the necessary constraint below. In this Gaussian setting, the slopes and variances must satisfy these symmetry constraints, the implied precision matrix must be positive definite, and the intercepts must be consistent with a common mean. Together these conditions give a proper multivariate normal joint distribution.
This condition is stronger than saying that both conditional regressions are individually reasonable. It links their coefficients and residual variances. It is also specific to the stated Gaussian setup: other distributions have their own compatibility requirements, so this equation should not be used as a universal test for every chained-equations model.
Worked Example: Two Conditional Slopes Do Not Agree
Suppose three hypothetical continuous variables \(X,Y,Z\) have proposed Gaussian full conditionals, each with conditional variance \(1\):
For the pair \(X,Y\), the first conditional gives \(b_{XY}=0.4\) and \(\tau_X^2=1\). The second gives \(b_{YX}=0.1\) and \(\tau_Y^2=1\). The Gaussian compatibility requirement would be
Since \(0.4\ne0.1\), the required equality fails. Each displayed conditional is a valid normal model by itself, but the three cannot all be the full conditionals of the same multivariate normal joint distribution. This is a collection-level failure that would not be caught by checking only that each regression can be fitted.
Pairwise Independence Can Hide a Three-Way Interaction
A different limitation is that pairwise relationships can leave higher-order structure unidentified. Consider three hypothetical binary variables \(X,Y,Z\), each coded as either \(-1\) or \(1\). For a constant \(d\) with \(|d|<1\), define a joint distribution by
This is a valid probability distribution: each of its eight probabilities is positive because \(1+dxyz>0\), and the probabilities sum to \(1\), since the four \(xyz=1\) outcomes and four \(xyz=-1\) outcomes contribute equally to the total. The term \(dxyz\) is a three-way interaction. It changes the distribution of the triple without changing any of the pairwise margins.
Worked Example: Pairwise Independence With a Three-Way Dependence
Take \(d=0.6\). Each outcome with \(xyz=1\) has probability \((1+0.6)/8=0.2\), and each outcome with \(xyz=-1\) has probability \((1-0.6)/8=0.05\). There are four outcomes of each type, so the total probability is \(4(0.2)+4(0.05)=1\).
To obtain the marginal probability of \(X=x,Y=y\), sum over the two values of \(z\):
Thus \(X\) and \(Y\) are pairwise independent and each combination has probability \(1/4\). The same calculation applies to the other two pairs. Yet, conditional on \(Y=y,Z=z\), the denominator is \(P(Y=y,Z=z)=1/4\), and therefore
For \(Y=1,Z=1\), this probability is \((1+0.6)/2=0.8\); for \(Y=1,Z=-1\), it is \((1-0.6)/2=0.2\). The pairwise conditional probability \(P(X=1\mid Y=1)\), by contrast, is \(0.5\). Pairwise independence therefore does not rule out a substantial dependence of \(X\) on the combination of \(Y\) and \(Z\). In an imputation problem, a model that only reflects pairwise relationships may miss a dependence relevant to an interaction in the analysis model.
The example does not say that every pairwise check is useless. It shows what such a check cannot establish: pairwise behavior alone does not specify the joint distribution. A model can reproduce pairwise margins yet differ in its conditional behavior given a combination of predictors.
Symmetry Is Not Enough
Even when Gaussian slope symmetry holds, a proper joint distribution still requires a positive-definite precision matrix. This condition ensures that the implied quadratic form defines a normalizable density with finite variances. A collection of individually proper normal conditional distributions can fail this requirement.
Worked Example: Symmetric Slopes but No Proper Gaussian Joint
Suppose \(X,Y,Z\) each have conditional variance \(1\), and each conditional mean is \(0.8\) times each of the other two variables. All the opposite-direction slopes agree, so the Gaussian slope symmetry equations hold. The implied precision matrix has diagonal entries \(1\) and off-diagonal entries \(-0.8\):
For the vector \((1,1,1)^\mathsf{T}\), multiplication by \(\Omega\) gives \((-0.6,-0.6,-0.6)^\mathsf{T}\). Thus \(-0.6\) is an eigenvalue, so \(\Omega\) is not positive definite. No proper multivariate normal joint distribution has this precision matrix. This example shows why checking only pairwise coefficient symmetry is not a complete Gaussian compatibility audit.
Using Pairwise Audits Carefully
A pairwise review is best understood as a screening step, not a certificate. It can find omissions or inconsistencies worth investigating, but the absence of an obvious pairwise problem does not prove joint compatibility. In particular, when the analysis uses an interaction, a nonlinear transformation, or several incomplete predictors, ask whether the relevant conditional models account for the joint combinations of variables, rather than only their separate relationships.
For a proposed MICE specification, a practical review can proceed as follows:
For each incomplete variable, record which other variables and which functions of them enter its imputation model. Distinguish a full conditional from a pairwise summary.
Where a tractable joint model is intended, derive the restrictions its conditionals impose. For Gaussian linear conditionals, check slope-variance symmetry and positive definiteness.
Check whether the conditional behavior changes across combinations of predictors when the analysis model requires such changes. Pairwise plots or regressions cannot, on their own, verify this.
Review whether the imputation models preserve the relationships needed for the analysis, then examine imputed and observed relationships across relevant predictor combinations. These checks provide evidence about behavior, not a general proof of compatibility.
The mathematical audit is easiest when a joint distribution is explicitly specified and the conditionals can be derived from it. When separate conditional models are specified directly, a joint distribution may be difficult to identify or may not exist. In such cases, diagnostics can reveal problematic behavior, but passing a collection of empirical checks does not establish a theorem of compatibility.
What Compatibility Does—and Does Not—Guarantee
Compatibility is a coherence property: it asks whether the conditional specifications can be understood as pieces of one joint distribution. It does not by itself ensure that the chosen joint distribution is scientifically appropriate, that its assumptions match the data, or that an analysis is free of bias. Conversely, incompatible conditionals do not mean that chained equations cannot be run. They mean that the specified conditionals cannot all be interpreted as full conditionals of one common joint model.
A chained-equations algorithm updates variables using the specified conditional models. Under suitable conditions, an algorithm may have a stationary distribution, but that fact alone does not show that its stationary distribution has the specified models as its full conditionals. In an incompatible system, the behavior can depend on details such as update order and model specification. Therefore, do not infer joint coherence merely from a chain that runs, stable-looking summaries, or plausible pairwise plots.
The main alternatives are complementary rather than interchangeable. A joint model specifies a common distribution directly and derives its conditionals, making coherence explicit but requiring a defensible joint model. A conditional specification can be more flexible for modeling each variable, but then compatibility must be considered rather than presumed. Pairwise audits are comparatively simple and useful for screening; their limitation is that they cannot establish all the joint restrictions or rule out higher-order dependence.
Common Mistake
- Wrong reasoning: “Every predictor has a plausible pairwise relationship with the incomplete variable, so the collection of imputation models must be compatible.”
- Correction: Plausible pairwise relationships do not verify the restrictions linking full conditionals, and they do not reveal all higher-order interactions. Check the joint requirements where they can be derived, and treat empirical diagnostics as evidence rather than proof.
- Another warning: Matching opposite-direction slopes is not enough in the Gaussian case; the implied precision matrix must also be positive definite.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For Gaussian full conditionals, \(X\mid Y\) has slope \(0.3\) and variance \(2\), while \(Y\mid X\) has slope \(0.6\) and variance \(4\). Does this pair satisfy the slope-variance symmetry requirement?
- Why does pairwise independence among three variables fail to rule out an interaction involving all three?
- In a Gaussian compatibility audit, the scaled slopes satisfy symmetry, but the implied precision matrix has a negative eigenvalue. What conclusion follows?
- Does a chain that runs and appears stable prove that its specified conditional models are compatible? Explain.
Solutions. (1) Yes: \(0.3/2=0.15\) and \(0.6/4=0.15\), so this pair satisfies the symmetry equation. This alone does not establish compatibility of a larger system or positive definiteness. (2) Pairwise margins average over the third variable and can erase dependence that appears only for combinations of all three; the binary example exhibits exactly this pattern. (3) The precision matrix is not positive definite, so it cannot define a proper multivariate normal joint distribution, even though the symmetry condition holds. (4) No. Running and apparent stability do not show that a stationary distribution exists with the specified conditionals as its full conditionals; algorithmic behavior is not a proof of compatibility.