Principal Component Mediation Analysis

• Principal component-based mediation analysis is a method that replaces high-dimensional, correlated variables with orthogonal components to simplify estimation.
• It transforms exposures or mediators into latent components, addressing multicollinearity and stabilizing direct and indirect effect estimation in complex models.
• Implementation techniques include standard PCA, sparse PCA, and functional PCA, with estimation strategies like bootstrapping and the delta method ensuring reliable inference.

Searching arXiv for recent and foundational papers on principal component-based mediation analysis to ground the article. Principal component-based mediation analysis denotes a class of mediation methods that replace high-dimensional or strongly correlated observed variables with a smaller set of orthogonal latent components before estimating direct and indirect effects. In the literature, this strategy appears in several distinct but related forms: PCA on correlated exposure mixtures used as the treatment in mediation models, PCA or sparse PCA on multiple mediators, likelihood-based component models that jointly project exposures and mediators, and functional principal component analysis for sparse and irregular longitudinal mediators and outcomes. Across these variants, the recurring aims are to address multicollinearity, reduce dimensionality, stabilize estimation, and define mediation effects on component scores rather than on the original variables alone (Wang et al., 13 Sep 2025, Zhao et al., 2018, Zeng et al., 2020, Zhao et al., 2021, Zhao, 2022).

1. Scope and problem settings

The principal component-based approach emerges in settings where conventional mediation analysis is destabilized by dependence structure or dimensionality. In exposure-mixture studies, environmental exposures measured in the same population often share sources and metabolic pathways, producing strong correlations, while only a subset of exposures may exert non-zero effects on the mediator and outcome. The same tutorial also notes potential nonlinear and interactive effects, although its PC-based formulation is restricted to linear models without interaction (Wang et al., 13 Sep 2025).

In high-dimensional multiple-mediator settings, the challenge is different but related. With a single treatment or exposure and a large mediator vector, mediator-specific pathway decomposition becomes difficult when mediators are causally dependent, and the number of possible path-specific terms can grow exponentially with the number of mediators. PCA-based reparameterization was introduced to transform mediators into orthogonal components, whereas sparse PCA was proposed to improve interpretability when dense loadings obscure biological meaning (Zhao et al., 2018).

In multimodal integration problems, PCA has been used on high-dimensional exposures while mediator selection is handled downstream by penalized structural equation models. In an Alzheimer’s disease imaging–proteomics setting, this combination was motivated by the fact that both exposures and mediators were potentially larger than $n$ and highly correlated (Zhao et al., 2021). A closely related framework, Principal Component Mediation Analysis (PCMA), assumes that multiple exposures act on the outcome through a small number of latent, orthogonal components representing parallel mediation mechanisms (Zhao, 2022).

For sparse and irregular longitudinal data, the same dimensionality-reduction logic is implemented with functional principal component analysis rather than ordinary PCA. There, the mediator and outcome are treated as latent smooth stochastic processes, and mediation is performed in the space of functional principal component scores (Zeng et al., 2020).

Setting	Principal-component object	Stated purpose
Exposure mixtures	Exposure PCs from standardized exposures	Alleviating multicollinearity and focusing inference on dominant patterns of variation
Multiple mediators	Mediator PCs or sparse mediator PCs	Defining orthogonal transformed mediators and improving interpretability
Multimodal high-dimensional data	Exposure PCs followed by penalized SEM	Reducing dimensionality and producing uncorrelated exposure composites
Sparse longitudinal mediation	Functional PCs of mediator and outcome trajectories	Dimension reduction, denoising trajectories, and borrowing strength across subjects

2. Core model constructions

In the exposure-mixture formulation, one begins with a standardized exposure vector $X \in \mathbb{R}^p$, a continuous mediator $M$, a continuous outcome $Y$, and covariates $C$. PCA is applied to the standardized exposure matrix $X_{\mathrm{std}}$, yielding a loading matrix $\Lambda \in \mathbb{R}^{p \times K}$ and subject-level PC scores

$$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$

The mediation model is then posed directly on $Z$:

$$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$

$X \in \mathbb{R}^p$0

Under linearity and no exposure–mediator interaction, the natural direct, indirect, and total effects for a contrast $X \in \mathbb{R}^p$1 are

$X \in \mathbb{R}^p$2

When a global mixture effect is desired, the tutorial estimates PC-specific effects and sums component-specific effects for the chosen contrast (Wang et al., 13 Sep 2025).

In classical PCA mediation for multiple mediators, the transformation is applied to the mediator vector rather than the exposure vector. With treatment $X \in \mathbb{R}^p$3, mediator vector $X \in \mathbb{R}^p$4, outcome $X \in \mathbb{R}^p$5, and optional covariates $X \in \mathbb{R}^p$6, the linear structural equation model is

$X \in \mathbb{R}^p$7

Let $X \in \mathbb{R}^p$8 be the loading matrix and $X \in \mathbb{R}^p$9. The transformed models use $M$0 and $M$1, so that the component-specific indirect effect is

$M$2

If $M$3 is square orthogonal, the total indirect effect is invariant under rotation:

$M$4

Sparse PCA modifies this construction by replacing dense loadings with sparse ones, commonly through elastic-net SPCA, so that components involve only a small subset of mediators (Zhao et al., 2018).

PCMA goes further by estimating exposure and mediator projections jointly. For component $M$5, exposure and mediator scores are

$M$6

with $M$7 and $M$8. The single-component linear structural equations are

$M$9

The component-specific indirect, direct, and total effects are then

$Y$0

For multiple components, PCMA uses sequential deflation so that effects add across orthogonal mechanisms (Zhao, 2022).

In the longitudinal framework, the mediator and outcome are smooth processes,

$Y$1

and mediation is performed in the space of functional scores $Y$2 and $Y$3. Under the concurrent outcome model, time-resolved effects are reconstructed from score differences and eigenfunctions (Zeng et al., 2020).

3. Causal estimands and identification

The causal estimands in principal component-based mediation remain natural direct and indirect effects, but the intervention target is the latent component score or score vector. In the exposure-mixture setting, with PC scores $Y$4, potential outcomes are written as $Y$5 and $Y$6, and for a prespecified change $Y$7,

$Y$8

$Y$9

The tutorial notes that $C$0 is often chosen as a one-standard-deviation change in each retained PC or as a mixture-relevant joint shift (Wang et al., 13 Sep 2025).

In multiple-mediator PCA under the linear SEM, the natural indirect effect is the familiar product $C$1, and after transformation the decomposition is rewritten in orthogonal directions. PCMA retains the same product interpretation at the component level, but the components are estimated to align with mediation structure rather than solely with variance (Zhao et al., 2018, Zhao, 2022).

The identification assumptions are conventional for mediation analysis but acquire component-specific forms. The exposure-mixture tutorial states consistency, positivity, SUTVA, no unmeasured confounding for PCs–mediator, mediator–outcome, and PCs–outcome relations conditional on $C$2, and cross-world independence $C$3. It also highlights a sufficiency assumption for retained PCs,

$C$4

which can fail if important exposure information lies in omitted PCs (Wang et al., 13 Sep 2025).

The longitudinal FPCA framework uses analogous assumptions adapted to stochastic processes: ignorability of exposure assignment given covariates and observed time-varying covariates, sequential ignorability for increments of mediator and outcome processes, consistency, positivity, and SUTVA. The paper explicitly states that natural-effects interpretation relies on standard mediation identifiability, and that interventional analogues can be considered under weaker assumptions if needed (Zeng et al., 2020).

A common misconception is that orthogonality of principal components alone guarantees causal interpretability. The literature does not support that conclusion. Orthogonality stabilizes regression and clarifies decomposition in transformed space, but causal interpretation still depends on sequential ignorability, correct model specification, and support conditions; in several formulations, cross-world assumptions remain central (Zhao et al., 2018, Wang et al., 13 Sep 2025).

4. Estimation, inference, and implementation

In the exposure-mixture tutorial, implementation is explicit and software-oriented. Exposures are centered and scaled, PCA is computed with prcomp(center=TRUE, scale.=TRUE), $C$5 is selected by cumulative explained variance or scree plot, PC scores are appended to the analysis dataset, and mediator and outcome models are fit linearly. The example code retains PCs achieving $C$6 cumulative variance. Interval estimation is based on the delta method or the bootstrap, and the tutorial gives the delta-method gradient for the indirect effect $C$7 with $C$8. The implementation guidance also names factoextra::fviz_eig for scree plots and CMAverse::cmest with model="rb", mreg/yreg="linear", EMint=FALSE, estimation="paramfunc", and inference="delta" (Wang et al., 13 Sep 2025).

Sparse PCA mediation for high-dimensional mediators estimates a sparse loading matrix by either penalized eigenvector SPCA or regression-based SPCA. The regression-based formulation solves, for each component,

$C$9

then uses $X_{\mathrm{std}}$0 as transformed mediators. Indirect effects are estimated as $X_{\mathrm{std}}$1, with variance approximated by the delta method or obtained by nonparametric bootstrap over subjects (Zhao et al., 2018).

In the multimodal high-dimensional setting, PCA is followed by a penalized least-squares SEM,

$X_{\mathrm{std}}$2

with structured penalties on $X_{\mathrm{std}}$3, $X_{\mathrm{std}}$4, and $X_{\mathrm{std}}$5. The pathway Lasso–type penalty includes $X_{\mathrm{std}}$6 terms, a group Lasso on product terms performs exposure selection, and a Lasso penalty is applied to direct effects. The optimization is carried out through an augmented Lagrangian with blockwise updates and BIC-based hyperparameter selection (Zhao et al., 2021).

PCMA uses a likelihood-based block coordinate descent. Given current parameters, $X_{\mathrm{std}}$7 and $X_{\mathrm{std}}$8 are updated by constrained quadratic optimization under unit-norm constraints, and the regression parameters $X_{\mathrm{std}}$9, $\Lambda \in \mathbb{R}^{p \times K}$0, $\Lambda \in \mathbb{R}^{p \times K}$1, $\Lambda \in \mathbb{R}^{p \times K}$2, and $\Lambda \in \mathbb{R}^{p \times K}$3 are updated by ordinary least squares on component scores. For multiple components, exposures, mediators, and outcome are residualized after each fitted component, and bootstrap inference is then conducted by resampling component-level data with fixed loadings (Zhao, 2022).

The longitudinal FPCA method uses either PACE-type estimation or Bayesian FPCA. In the Bayesian version, eigenfunctions are represented with thin-plate spline bases, subject-specific scores are estimated jointly with measurement-error variances, and MCMC with Gibbs and Metropolis–Hastings updates propagates uncertainty from eigenfunction estimation through mediation effect reconstruction. Pointwise credible bands and integrated-effect credible intervals are then obtained from posterior draws (Zeng et al., 2020).

5. Empirical performance and substantive applications

Simulation evidence is heterogeneous across formulations. For exposure mixtures, the tutorial reports that global indirect-effect bias depended strongly on how many PCs were retained. Using only the first PC produced relative bias of approximately $\Lambda \in \mathbb{R}^{p \times K}$4–$\Lambda \in \mathbb{R}^{p \times K}$5 across scenarios; using the top 3 PCs produced relative bias of approximately $\Lambda \in \mathbb{R}^{p \times K}$6–$\Lambda \in \mathbb{R}^{p \times K}$7; and retaining enough PCs to explain $\Lambda \in \mathbb{R}^{p \times K}$8 variance produced relative bias of approximately $\Lambda \in \mathbb{R}^{p \times K}$9–$$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$0. The same comparison states that ERS-MA often achieved lower bias, with average relative bias $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$1–$$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$2 versus $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$3–$$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$4 for various PC-MA choices (Wang et al., 13 Sep 2025).

For sparse and irregular longitudinal data, the Bayesian FPCA mediation method was evaluated with $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$5 subjects and irregular observation counts $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$6, $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$7. At $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$8, the reported performance for integrated NIE was bias $$Z_i = \Lambda^\top X_{\mathrm{std},i}, \qquad i=1,\ldots,n.$$9, RMSE $Z$0, and coverage $Z$1, compared with $Z$2 for mixed models and $Z$3 for GEE; for integrated TE, the corresponding values were bias $Z$4, RMSE $Z$5, and coverage $Z$6, compared with $Z$7 and $Z$8 for the alternatives. As observations densified, MFPCA and mixed models became comparable and both outperformed GEE (Zeng et al., 2020).

For PCMA, simulations with two true parallel mediation paths reported that the method correctly identified both components essentially $Z$9 of the time across settings, with loading similarity at least $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$0 at $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$1. The PCA-first baseline frequently missed the second component and had loading similarity around $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$2 in low dimensions (Zhao, 2022). In the multimodal penalized-PCA framework, simulations with $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$3 and $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$4 and $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$5 reported that the estimated number of PCs consistently matched truth at approximately 6, while sensitivity reached $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$6 and specificity was approximately $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$7–$$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$8 for $$M = \alpha_0 + \alpha_Z^\top Z + \alpha_C^\top C + \varepsilon_M,$$9 (Zhao et al., 2021).

Applications likewise vary by domain. In the PROTECT birth cohort, the exposure-mixture tutorial analyzed 11 correlated prenatal phthalates, leukotriene E4 as mediator, neonatal head circumference $X \in \mathbb{R}^p$00-score as outcome, and maternal age, education, and pre-pregnancy BMI as covariates, with sample size $X \in \mathbb{R}^p$01. Five PCs explaining more than $X \in \mathbb{R}^p$02 of variance were retained. Among them, only PC2 showed suggestive NIE under the paper’s reporting convention, and the global NIE for the mixture was approximately $X \in \mathbb{R}^p$03 with $X \in \mathbb{R}^p$04 CI $X \in \mathbb{R}^p$05, indicating limited evidence that LTE4 mediated the overall phthalate mixture effect on head circumference (Wang et al., 13 Sep 2025).

In the baboon longitudinal application, $X \in \mathbb{R}^p$06 female baboons were analyzed with early adversity as binary exposure, strength of adult social bonds as mediator process, and log fecal glucocorticoid concentrations as outcome process. $X \in \mathbb{R}^p$07 functional components explained more than $X \in \mathbb{R}^p$08 of variance for both processes. For exposure defined as at least one adversity, the reported effects were $X \in \mathbb{R}^p$09, $X \in \mathbb{R}^p$10, and $X \in \mathbb{R}^p$11, corresponding to approximately $X \in \mathbb{R}^p$12, $X \in \mathbb{R}^p$13, and $X \in \mathbb{R}^p$14 on the percent-change scale (Zeng et al., 2020).

In Alzheimer’s disease applications, the two multivariate frameworks were applied to ADNI data with $X \in \mathbb{R}^p$15 subjects with mild cognitive impairment. In the penalized SEM study, 320 CSF peptide intensities were reduced to $X \in \mathbb{R}^p$16 PCs explaining approximately $X \in \mathbb{R}^p$17 of exposure variance, and the selected mediators included hippocampus, entorhinal cortex, middle temporal gyrus, angular gyrus, precuneus, insula, fusiform gyrus, frontal pole, precentral gyrus, cerebellar vermal lobules, and ventricular structures. In the likelihood-based PCMA study, three orthogonal components had significant negative indirect effects: for C1, $X \in \mathbb{R}^p$18, $X \in \mathbb{R}^p$19, $X \in \mathbb{R}^p$20, $X \in \mathbb{R}^p$21; for C2, $X \in \mathbb{R}^p$22, $X \in \mathbb{R}^p$23, $X \in \mathbb{R}^p$24, $X \in \mathbb{R}^p$25; and for C3, $X \in \mathbb{R}^p$26, $X \in \mathbb{R}^p$27, $X \in \mathbb{R}^p$28, $X \in \mathbb{R}^p$29 (Zhao et al., 2021, Zhao, 2022).

6. Interpretation, strengths, and limitations

The principal strength shared by these methods is that they replace correlated high-dimensional variables with orthogonal or approximately orthogonal components, thereby improving numerical stability and making mediation estimation feasible in settings where direct regression is ill-posed. In exposure mixtures, this directly addresses multicollinearity. In functional data, FPCA additionally denoises trajectories and accommodates sparse, irregular observation times. In multivariate exposure–mediator settings, component-level formulations can reveal parallel mediation mechanisms or selected protein–structure–memory paths that are difficult to isolate in the original variable space (Wang et al., 13 Sep 2025, Zeng et al., 2020, Zhao et al., 2021, Zhao, 2022).

The central limitation is interpretability. Ordinary PCs are unsupervised and outcome-agnostic: the leading PCs capture variance, not necessarily mediation-relevant signal. A one-unit intervention on a PC while holding other PCs fixed may not correspond to a physically realizable change in the original variables. Exposure-wise attribution through loadings is therefore descriptive and depends on the loading matrix. The same issue motivates sparse PCA, whose sparse loadings improve interpretability but weaken exact orthogonality and make invariance of the total indirect effect approximate rather than exact when $X \in \mathbb{R}^p$30 is not orthonormal (Wang et al., 13 Sep 2025, Zhao et al., 2018).

Several methodological boundaries recur across the literature. Most formulations assume linearity and no exposure–mediator interaction in the outcome model; the tutorial explicitly notes that product and difference methods are not algebraically equivalent outside the linear no-interaction setting. PCMA’s asymptotic theory is derived for low-dimensional settings with fixed $X \in \mathbb{R}^p$31 and $X \in \mathbb{R}^p$32. The longitudinal FPCA approach requires strong and generally untestable sequential ignorability for short-interval mediator–outcome increments. In observational settings, all variants remain vulnerable to unmeasured confounding despite orthogonalization (Wang et al., 13 Sep 2025, Zeng et al., 2020, Zhao, 2022).

Reporting recommendations are correspondingly explicit in the exposure-mixture tutorial. Analysts are advised to provide the loading matrix $X \in \mathbb{R}^p$33 or a heatmap, PC-wise variance explained, and the number of PCs retained with its selection criterion; to state the contrast $X \in \mathbb{R}^p$34 used for NDE, NIE, and TE; to report PC-specific effects with uncertainty; to define precisely how any global effect aggregates across PCs; and to summarize exposure-specific contributions through the loading-based expression

$X \in \mathbb{R}^p$35

The same tutorial recommends sensitivity checks over alternative $X \in \mathbb{R}^p$36, alternative contrasts, and triangulation with complementary methods such as ERS-MA and BKMR-CMA (Wang et al., 13 Sep 2025).

Taken together, these studies suggest that principal component-based mediation analysis is best regarded not as a single estimator but as a family of projection-based mediation strategies. The family is most compelling when correlation or dimensionality is the dominant statistical obstacle, when component-level interpretation is scientifically acceptable, and when the assumptions required for causal mediation analysis can be defended with the same rigor demanded in non-component formulations.