Partial Copula: Adjusted Dependence in Statistics

• Partial copula is a dependence measure defined by the copula of conditional probability integral transforms that removes the impact of a conditioning variable.
• It aggregates conditional copulas into an unconditional summary capturing nonlinear traits such as asymmetry and tail dependence beyond standard partial correlations.
• It underpins methods for conditional independence testing and vine copula modeling, facilitating sequential estimation with both nonparametric and parametric techniques.

A partial copula is the copula of conditional probability integral transforms and therefore a dependence object for two variables after removing the influence of a conditioning variable through their full conditional distributions. For continuous random variables, if $U=F_{Y|X}(Y\mid X)$ and $V=F_{Z|X}(Z\mid X)$, then the partial copula of $(Y,Z)$ given $X$ is $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$; equivalently, it is the unconditional copula of $(U,V)$ and the average of the conditional copulas over the distribution of the conditioning variable (Spanhel et al., 2015). In this sense it generalizes partial correlation from linear residual association to a rank-based, fully distributional notion of residual dependence, and it is central both to conditional independence testing and to simplified vine copula modeling (Bergsma, 2011).

1. Formal definition and construction

Let $(X,Y,Z)$ be continuous random variables with conditional distribution functions $F_{Y|X}(y|x)$ and $F_{Z|X}(z|x)$. The conditional probability integral transforms are

$U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$

Then $V=F_{Z|X}(Z\mid X)$0 and $V=F_{Z|X}(Z\mid X)$1 are uniform on $V=F_{Z|X}(Z\mid X)$2, and each is independent of $V=F_{Z|X}(Z\mid X)$3. The partial copula of $V=F_{Z|X}(Z\mid X)$4 given $V=F_{Z|X}(Z\mid X)$5 is the joint distribution of $V=F_{Z|X}(Z\mid X)$6:

$V=F_{Z|X}(Z\mid X)$7

Equivalently, if $V=F_{Z|X}(Z\mid X)$8 denotes the conditional copula of $V=F_{Z|X}(Z\mid X)$9 given $(Y,Z)$0, then

$(Y,Z)$1

Thus the partial copula is the expected conditional copula, or an unconditional summary of a $(Y,Z)$2-varying family of conditional copulas (Spanhel et al., 2015).

This construction is the copula-theoretic analogue of residualization. In the Gaussian case, if $(Y,Z)$3 and $(Y,Z)$4 are residuals from linear regression on $(Y,Z)$5, then

$(Y,Z)$6

and the copula of $(Y,Z)$7 is Gaussian with correlation equal to the classical Pearson partial correlation. For a trivariate normal vector, that correlation is

$(Y,Z)$8

Hence, in the Gaussian setting, the partial copula reduces exactly to the Gaussian copula indexed by the familiar partial correlation parameter (Spanhel et al., 2015).

The same definition underlies the “partial copula transform” used in testing conditional independence. If $(Y,Z)$9, then $X$0. This makes conditional independence testable by first transforming the data with conditional distribution functions and then applying an ordinary independence test to the transformed pairs (Bergsma, 2011).

2. Interpretation as residual dependence and core properties

The central interpretation is “net-of-$X$1” dependence. Because $X$2 and $X$3, the partial copula measures dependence between $X$4 and $X$5 after removing the influence of $X$6 at the level of the full conditional distributions, not merely through linear projections. This gives a nonlinear analogue of partial correlation and explains why the partial copula can capture asymmetry, tail behavior, and other non-Gaussian dependence features (Spanhel et al., 2015).

A basic structural property is invariance under strictly monotone transformations of the target variables. If $X$7 with $X$8 strictly increasing, then $X$9, so the partial copula is unaffected. Recent work extends this to strictly increasing transformations of $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$0 and $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$1 that may be $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$2-specific: if $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$3 and $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$4 are strictly increasing in the first argument for each $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$5, then

$C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$6

and the partial copula is unchanged (Justus et al., 11 Mar 2026).

As a copula, the partial copula satisfies the Fréchet–Hoeffding bounds

$C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$7

where $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$8 and $C_{Y,Z;X}(u,v)=P(U\le u,V\le v)$9. Because it is an average of copulas, pointwise order properties can transfer under sign-stable conditional dependence. If for all $(U,V)$0, $(U,V)$1, then $(U,V)$2; similarly, if $(U,V)$3 for all $(U,V)$4, then $(U,V)$5 (Justus et al., 11 Mar 2026).

Several associated dependence measures are defined directly from the partial copula. Partial Spearman’s rho and partial Kendall’s tau are

$(U,V)$6

and

$(U,V)$7

A key distinction is that partial Spearman’s rho equals the expected conditional Spearman’s rho, whereas partial Kendall’s tau does not, in general, equal the expected conditional Kendall’s tau (Spanhel et al., 2015). Tail dependence coefficients also aggregate cleanly: if $(U,V)$8 and $(U,V)$9 denote the upper and lower tail dependence coefficients of $(X,Y,Z)$0, then $(X,Y,Z)$1 and $(X,Y,Z)$2 (Spanhel et al., 2015).

3. Conditional copulas, simplifying assumption, and classes where equality holds

The conditional copula $(X,Y,Z)$3 and the partial copula $(X,Y,Z)$4 coincide under the simplifying assumption, namely when the conditional copula does not depend on the realized conditioning value:

$(X,Y,Z)$5

In that case,

$(X,Y,Z)$6

Thus the partial copula is exact when the conditional dependence structure is conditioning-invariant (Stöber et al., 2012).

This equality holds for important model classes. The multivariate Normal and Student-$(X,Y,Z)$7 copulas give rise to simplified pair copula constructions, so their conditional copulas do not depend on the conditioning values. In particular, for the Student-$(X,Y,Z)$8 copula,

$(X,Y,Z)$9

which is independent of the realized values of the conditioning variables. In the Archimedean class for $F_{Y|X}(y|x)$0, the only simplified pair copula construction is generated by the gamma Laplace transform

$F_{Y|X}(y|x)$1

Conditioning then yields conditional generators

$F_{Y|X}(y|x)$2

so parameters depend on the number of conditioned variables but not on the conditioning values themselves (Stöber et al., 2012).

Outside these special families, the simplifying assumption generally fails. The trivariate Frank copula provides a standard example: the conditional copula of $F_{Y|X}(y|x)$3 is Ali–Mikhail–Haq with parameter $F_{Y|X}(y|x)$4, so the dependence varies with $F_{Y|X}(y|x)$5. The partial copula is then the mixture

$F_{Y|X}(y|x)$6

which is not equal to the conditional copula unless $F_{Y|X}(y|x)$7 is degenerate (Stöber et al., 2012).

This distinction matters conceptually. The conditional copula describes dependence at a fixed conditioning level; the partial copula is an average over such levels. A plausible implication is that the partial copula is informative as a global adjusted summary, but it can mask heterogeneity when the conditional copula varies strongly with the conditioning variable.

4. Higher-order partial copulas and the partial vine copula

In vine copula constructions, the notion extends from a single conditioning variable to general conditioning sets. Consider a D-vine on $F_{Y|X}(y|x)$8 with an edge $F_{Y|X}(y|x)$9 and conditioning set $F_{Z|X}(z|x)$0. Define the higher-order transforms

$F_{Z|X}(z|x)$1

Each is uniform on $F_{Z|X}(z|x)$2. The $F_{Z|X}(z|x)$3-th order partial copula on that edge is

$F_{Z|X}(z|x)$4

For $F_{Z|X}(z|x)$5, this is the ordinary bivariate partial copula; for $F_{Z|X}(z|x)$6, it is a higher-order partial copula (Spanhel et al., 2015).

A crucial nuance is that these higher-order partial probability integral transforms are uniform but need not be independent of the full conditioning vector. The key lemma states:

$F_{Z|X}(z|x)$7

Hence independence from the conditioning vector characterizes coincidence with the true conditional probability integral transform (Spanhel et al., 2015).

The partial vine copula (PVC) is the simplified vine copula that assigns to each edge the corresponding partial copula. In a D-vine, the first tree uses the true bivariate margins, the second tree uses first-order partial copulas, and trees $F_{Z|X}(z|x)$8 use recursively defined higher-order partial copulas. Its density factorizes as

$F_{Z|X}(z|x)$9

with $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$0 and $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$1 the density of the relevant partial copula (Spanhel et al., 2015).

The PVC is treewise optimal but not globally optimal in general. Sequential minimization of the Kullback–Leibler divergence over trees yields the PVC at each tree conditional on the previous ones, which explains why stepwise estimators converge to it. However, when the simplifying assumption fails, the PVC need not minimize the global Kullback–Leibler divergence over all simplified vine copulas; the best global approximation may instead be a vine pseudo-copula, which allows non-uniform pseudo-observations in upper trees (Spanhel et al., 2015).

Later analytic work sharpened the approximation picture. In dimension three, the family of simplified copulas is dense in the uniform metric $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$2, but nowhere dense with respect to $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$3, weak conditional convergence, total variation, and Kullback–Leibler divergence. Within this framework, the partial vine copula is never the optimal simplified copula approximation under $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$4 for a non-simplified copula, and the mapping that assigns a copula its partial vine copula is discontinuous with respect to $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$5 at non-simplified copulas (Mroz et al., 2020). This suggests that the PVC is best understood as a practically important sequential target rather than as a universal projection operator.

5. Estimation, testing, and inference

A standard estimation strategy is two-step. First estimate the conditional distribution functions $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$6 and $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$7; then compute pseudo-observations

$U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$8

and estimate the copula of $U := F_{Y|X}(Y|X), \qquad V := F_{Z|X}(Z|X).$9 either nonparametrically, semiparametrically, or parametrically. Suggested first-stage methods include kernel or local-likelihood regression for conditional distribution functions, as well as parametric or semiparametric models such as GLMs, quantile regression, location–scale models, and ARMA–GARCH in time series (Spanhel et al., 2015).

For conditional independence testing, the core procedure is to transform the sample using estimated conditional distribution functions and then apply an ordinary independence test to the transformed pairs. Kernel estimators of the form

$V=F_{Z|X}(Z\mid X)$00

are one option for Euclidean covariates. The resulting independence statistics can be covariance, Kendall’s tau, Hoeffding’s $V=F_{Z|X}(Z\mid X)$01, Bergsma’s $V=F_{Z|X}(Z\mid X)$02, or $V=F_{Z|X}(Z\mid X)$03. Under mild conditions, the estimation effect vanishes asymptotically:

$V=F_{Z|X}(Z\mid X)$04

Thus the asymptotic null and alternative behavior is the same as if the true transformed variables were observed (Bergsma, 2011).

For vine copulas, estimation practice is typically sequential. Parametric and nonparametric simplified vine estimators are usually fitted tree-by-tree. Under mild regularity conditions, when the simplifying assumption fails, stepwise estimators converge in probability to the PVC, while joint maximum likelihood converges to pseudo-true parameters minimizing the global Kullback–Leibler divergence. In general, these probability limits differ. This discrepancy is both computational and statistical: the combinatorial search for joint optimization grows as $V=F_{Z|X}(Z\mid X)$05 for $V=F_{Z|X}(Z\mid X)$06 candidate families, versus $V=F_{Z|X}(Z\mid X)$07 in a sequential scheme (Spanhel et al., 2015).

A recurrent practical difficulty is the first-stage estimation of conditional distributions. Nonparametric estimation suffers from the curse of dimensionality for high-dimensional conditioning variables, and feasible nonparametric procedures typically require low-dimensional conditioning sets or structural assumptions such as additivity or single-index structure (Spanhel et al., 2015).

6. Limitations, misconceptions, examples, and recent developments

A fundamental limitation is that conditional independence and partial independence are not equivalent. If $V=F_{Z|X}(Z\mid X)$08, then $V=F_{Z|X}(Z\mid X)$09, the independence copula. The converse fails: $V=F_{Z|X}(Z\mid X)$10 does not imply $V=F_{Z|X}(Z\mid X)$11. The standard counterexample is the exchangeable trivariate Farlie–Gumbel–Morgenstern copula, for which the conditional copula varies with $V=F_{Z|X}(Z\mid X)$12 while the partial copula equals $V=F_{Z|X}(Z\mid X)$13 (Spanhel et al., 2015). The same phenomenon appears in recent covariate-adjusted formulations: opposite-signed conditional dependence can cancel in expectation, leaving an independence partial copula although conditional dependence is nonzero for most conditioning values (Justus et al., 11 Mar 2026).

This non-equivalence also separates the partial copula from partial correlation. In the example with $V=F_{Z|X}(Z\mid X)$14 and conditionally independent variables

$V=F_{Z|X}(Z\mid X)$15

with independent $V=F_{Z|X}(Z\mid X)$16 of variance $V=F_{Z|X}(Z\mid X)$17, one has $V=F_{Z|X}(Z\mid X)$18 and therefore $V=F_{Z|X}(Z\mid X)$19, yet the Pearson partial correlation converges to $V=F_{Z|X}(Z\mid X)$20 as $V=F_{Z|X}(Z\mid X)$21 (Spanhel et al., 2015). This is one of the clearest demonstrations that the partial copula is not merely a rank-transformed variant of linear partial correlation.

The partial copula also lacks several closure and optimality properties one might expect. Even if the joint copula of $V=F_{Z|X}(Z\mid X)$22 is Archimedean, the partial copula need not be Archimedean; the Frank example shows associativity can fail. Likewise, if conditional copulas belong to a parametric family $V=F_{Z|X}(Z\mid X)$23, there need not exist a fixed $V=F_{Z|X}(Z\mid X)$24 such that the partial copula remains in the same family (Spanhel et al., 2015). In $V=F_{Z|X}(Z\mid X)$25, the partial copula is generally not the projector onto the conditional copula process; the relevant $V=F_{Z|X}(Z\mid X)$26 projection is $V=F_{Z|X}(Z\mid X)$27, which differs in general from $V=F_{Z|X}(Z\mid X)$28 (Spanhel et al., 2015).

In multivariate vine settings, further interpretive caution is required. Partial independence on an upper tree edge does not imply conditional independence, nor vice versa, for higher-order partial copulas. The paper on the partial vine copula states that conditional independence and higher-order partial independence are generally unrelated for $V=F_{Z|X}(Z\mid X)$29 (Spanhel et al., 2015). Moreover, examples show that replacing true conditional copulas by partial copulas can induce spurious dependence in upper trees; in a five-dimensional FGM construction, true conditional copulas in the third and fourth trees were independence copulas, but the PVC induced non-trivial upper-tree copulas. This can reduce total Kullback–Leibler divergence, illustrating why the PVC may work well empirically even when the simplifying assumption fails (Spanhel et al., 2015).

Recent work has repositioned the partial copula as a covariate-adjusted dependence representation with direct implications for sign recovery. Because partial Spearman’s rho equals the expected conditional Spearman’s rho, if conditional dependence has constant sign across $V=F_{Z|X}(Z\mid X)$30, then the sign of the partial rank measure reflects that adjusted sign. The same work argues that this can clarify Simpson’s paradox settings: marginal association may have a different sign from the adjusted association, whereas partial-copula-based rank measures align with the adjusted sign when the usual adjustment assumptions hold and conditional dependence does not switch sign (Justus et al., 11 Mar 2026). A plausible implication is that partial copulas are especially informative when the aim is global adjusted dependence rather than local, $V=F_{Z|X}(Z\mid X)$31-specific heterogeneity.

Across these developments, the partial copula has acquired a dual status. It is both a rigorous dependence measure—defined by conditional probability integral transforms and endowed with invariance and averaging properties—and a modeling device that targets tractable approximations in simplified vine constructions. Its usefulness is greatest when interpreted as an adjusted, global summary of conditional dependence rather than as a substitute for the full family of conditional copulas.