High-Dimensional Multiple Testing & False Discovery Rate (FDR)

Consider the outcome table of $m$ simultaneous hypothesis tests: The total number of hypotheses $m$ and total rejected hypotheses $R$ are observed, while $U, V, T, S$ and the number of true nulls $m_0$ are unobservable latent quantities. ### The Family-Wise Error Rate (FWER) Classical multiple comparison theory controls the **Family-Wise Error Rate (FWER)**: the probability of committing at least one Type I error: $$\mathrm{FWER} = P(V \ge 1)$$ 1. **The Bonferroni Correction**: By Boole's inequality (the union bound): $$P(V \ge 1) = P\left( \bigcup_{i \in I_0} \{ p_i \le \alpha_{\text{ind}} \} \right) \le \sum_{i \in I_0} P(p_i \le \alpha_{\text{ind}}) = m_0 \alpha_{\text{ind}} \le m \alpha_{\text{ind}}$$ To guarantee $\mathrm{FWER} \le \alpha$, Bonferroni rejects $H_{0, i}$ if: $$p_i \le \frac{\alpha}{m}$$ *Fatal Limitation of Bonferroni in Big Data*: If $m = 1,000,000$, Bonferroni requires $p \le 0.05 / 10^6 = 5 \times 10^{-8}$. This criterion is paralyzingly conservative: to prevent a single false positive, it sacrifices virtually all statistical power, discarding dozens of genuine scientific discoveries.

Yoav Benjamini and Yosef Hochberg (1995) revolutionized high-dimensional statistics by proposing the False Discovery Rate (FDR):

where $Q = V / R$ is the False Discovery Proportion (FDP) if $R > 0$, and $0$ if $R = 0$.

Rather than requiring that zero false positives occur across the entire study ($P(V \ge 1) \le \alpha$), FDR controls the expected proportion of false discoveries among all rejected hypotheses. Tolerating a small, controlled fraction of false discoveries (e.g., $\mathrm{FDR} \le 0.05$, meaning that among 1,000 declared discoveries, roughly 50 are false positives) yields immense statistical power in big data analysis.

The Benjamini-Hochberg (BH) Step-Up Procedure

To control FDR at level $q^* \in (0, 1)$:

1. Sort the $m$ observed $p$-values in ascending order:

1. Find the largest index $k$ such that:

1. If such a $k$ exists, reject all null hypotheses corresponding to the $k$ smallest $p$-values:

If no such $k$ exists, reject nothing.

Theorem (Benjamini-Hochberg 1995): If the $p$-values corresponding to the $m_0$ true null hypotheses are independent (or satisfy Positive Regression Dependency on a Subset - PRDS), the BH procedure guarantees:

for any configuration of false null hypotheses.

John D. Storey (2002) noted that when $m_0 < m$, the BH procedure is conservative by a factor of $\pi_0 = m_0 / m$. Storey formulated the positive False Discovery Rate (pFDR):

Estimating $\pi_0$ (The Proportion of True Nulls)

Because true null $p$-values are uniformly distributed on $[0, 1]$, while alternative $p$-values concentrate near $0$, $p$-values in the upper interval $(\lambda, 1]$ (typically $\lambda = 0.50$) consist almost entirely of true nulls. Storey's estimator of $\pi_0$ is:

The $q$-value of an observed hypothesis is defined as the minimum FDR at which that hypothesis would be declared significant:

$q$-values provide an intuitive Bayesian-like posterior probability metric for high-dimensional discovery reporting.

Bradley Efron (2004, 2012) developed the Local False Discovery Rate (lfdr) using an empirical Bayes two-groups mixture model.

Let test statistics be converted to $z$-scores $z_i = \Phi^{-1}(1 - p_i) \sim f(z)$. The density decomposes as a mixture:

where $f_0(z)$ is the theoretical or empirical null density (often modeled as $\mathcal{N}(\delta, \sigma_0^2)$ to absorb unobserved confounding) and $f_1(z)$ is the non-null alternative density.

By Bayes' rule, the posterior probability that a specific observation with $z$-score $z$ is a true null is:

Unlike tail-area FDR (which averages over all points more extreme than $z$), the local FDR quantifies the exact posterior probability of false discovery at the specific point $z$, providing the modern gold standard for large-scale genetic and proteomic screening.

In high-dimensional sparse regression $\mathbf{y} = \mathbf{X} \boldsymbol{\beta} + \boldsymbol{\epsilon}$, the Benjamini-Hochberg procedure cannot be applied directly to individual regression $p$-values: the columns of $\mathbf{X}$ are mutually correlated, which causes the single-feature $p$-values to be intensely dependent, violating independence assumptions.

Rina Foygel Barber and Emmanuel Candès (2015), along with Candès et al. (2018), revolutionized high-dimensional variable selection by inventing the Knockoffs Framework. Knockoffs control the False Discovery Rate exactly in finite samples without relying on asymptotic approximations or making any assumptions about the sparsity of the true coefficients.

1. Model-X Knockoffs Construction

Let $\mathbf{x} = (X_1, \dots, X_p)^T$ be the random vector of covariates. A knockoff vector $\tilde{\mathbf{x}} = (\tilde{X}_1, \dots, \tilde{X}_p)^T$ is an artificially generated synthetic feature vector satisfying two conditions:

1. Swap Invariance: For any subset of indices $S \subseteq \{1, \dots, p\}$, swapping the true features $X_j$ with their knockoffs $\tilde{X}_j$ for all $j \in S$ leaves the joint distribution unchanged:

2. Conditional Independence from Response: Knockoffs are generated without ever looking at the response variable $\mathbf{y}$:

For Gaussian covariates $\mathbf{x} \sim \mathcal{N}(\mathbf{0}, \mathbf{\Sigma})$, knockoffs are generated as:

where $\mathbf{D}$ is a positive diagonal matrix chosen such that $2 \mathbf{\Sigma} - \mathbf{D} \succeq \mathbf{0}$.

2. Feature Statistics and the Knockoff Filter

Fit a sparse learning algorithm (such as cross-validated Lasso) on the augmented feature matrix $[\mathbf{X}, \tilde{\mathbf{X}}]$ of dimension $n \times 2p$. For each variable $j \in \{1, \dots, p\}$, define a Feature Statistic $W_j$ comparing the strength of original variable $j$ against its synthetic knockoff $\tilde{j}$:

where $W_j > 0$ indicates that the real feature entered the model before its knockoff, while $W_j < 0$ indicates that the knockoff appeared stronger. A canonical example is the Lasso signed maximum lambda:

By the swap invariance property, if variable $j$ is a true null ($\beta_j^* = 0$), the sign of $W_j$ is an independent, fair coin flip:

3. The Knockoff+ Threshold

Sort the absolute values $|W_j|$. The data-dependent selection threshold is:

The selected set of discoveries is $\hat{S} = \{ j : W_j \ge T \}$.

Theorem (Barber and Candès 2015): For any linear or non-linear regression model with arbitrary unknown signal magnitude, the Knockoff+ procedure guarantees exact finite-sample FDR control:

Knockoffs eliminate the reliance on heuristic variable selection thresholds, providing a mathematically airtight shield against false scientific discoveries in high dimensions.

In high-dimensional scientific research, analysts frequently perform Data Dredging or "p-hacking": they run the Lasso on thousands of variables, identify the 5 variables with non-zero coefficients $\hat{S} = \{j : \hat{\beta}_j \ne 0\}$, and then fit an unpenalized OLS model on those 5 selected variables, reporting standard Student's $t$-test $p$-values.

Richard Berk et al. (2013), followed by Jonathan Taylor and Robert Tibshirani (2015), proved that naive post-selection $p$-values are catastrophically invalid: because the variables were selected specifically because they had the largest empirical correlations with $\mathbf{y}$, standard $p$-values are biased downward by orders of magnitude, generating massive rates of false scientific publication.

The Polyhedral Lemma

Taylor, Tibshirani, and collaborators developed Exact Selective Inference, conditioning hypothesis tests on the mathematical event that the specific model was selected:

The Polyhedral Lemma: For the Lasso (and many forward-stepwise selection algorithms), the selection event $\{\hat{S} = S, \mathrm{sign}(\hat{\boldsymbol{\beta}}_S) = \mathbf{s}\}$ can be expressed as a set of linear inequality constraints on the response vector $\mathbf{y}$:

where the matrix $\mathbf{A}$ and vector $\mathbf{b}$ depend on the design matrix $\mathbf{X}$ and active signs $\mathbf{s}$.

Conditioning a multivariate Gaussian observation $\mathbf{y} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 \mathbf{I})$ on lying inside a polyhedron $\{\mathbf{A} \mathbf{y} \le \mathbf{b}\}$ restricts the projected test statistic to a Truncated Gaussian Distribution:

where $\mathcal{V}^-$ and $\mathcal{V}^+$ are computable coordinate truncation limits. Evaluating tail probabilities under this truncated Gaussian distribution produces exact, uniformly distributed post-selection $p$-values, restoring rigorous statistical integrity to exploratory model search.

In large-scale biological and epidemiological testing (e.g., DNA methylation arrays, transcriptomics), test statistics are severely contaminated by unmeasured technical noise and latent confounding factors (e.g., batch effects, laboratory temperature, sample collection time, cell type heterogeneity). If ignored, these latent factors induce widespread spurious correlations across thousands of genes.

Jeffrey T. Leek and John D. Storey (2007) introduced Surrogate Variable Analysis (SVA) to estimate and remove latent confounders directly from high-dimensional testing matrices without requiring explicit metadata.

The SVA Decomposition

Let $\mathbf{Y} \in \mathbb{R}^{m \times n}$ be the gene expression matrix ($m$ genes, $n$ samples), modeled as:

where $\mathbf{x}_i$ are the observed primary clinical variables, and $\mathbf{z}_i \in \mathbb{R}^K$ are unobserved latent surrogate variables with coefficient loadings $\boldsymbol{\gamma}_j$.

1. Fit the primary model $\mathbf{Y} = \mathbf{X} \mathbf{B} + \mathbf{R}$ and compute the residual matrix:

1. Compute the SVD of the residual matrix: $\mathbf{R} = \mathbf{U} \mathbf{D} \mathbf{V}^T$.
2. Identify the number of statistically significant surrogate factors $K$ using permutation tests or the Marchenko-Pastur spectral cutoff.
3. Construct the surrogate variables $\hat{\mathbf{z}}_k$ from the right singular vectors of $\mathbf{R}$.
4. Re-run multiple hypothesis testing including the estimated surrogate variables as nuisance covariates:

SVA purges unmeasured technical variation and batch effects, eliminating thousands of false positive correlations and rescuing reproducibility in high-dimensional science.

In high-dimensional genomics and proteomics, hypotheses $H_{0, 1}, \dots, H_{0, m}$ are not exchangeable: extensive prior biological metadata exists. For instance, in a GWAS studying coronary artery disease, single nucleotide polymorphisms (SNPs) located in known cardiac lipid metabolic pathways or protein-coding exons are biologically far more likely to be true non-null signals than intergenic non-coding variants.

Standard Benjamini-Hochberg treats all $m$ hypotheses with equal egalitarian weight, sacrificing statistical power on high-prior variants. Christopher Genovese, Kathryn Roeder, and Larry Wasserman (2006) introduced the Weighted False Discovery Rate (wFDR).

Mathematical Formulation of Weighted FDR

Assign each hypothesis $i \in \{1, \dots, m\}$ a non-negative prior weight $w_i \ge 0$, normalized such that:

Features with strong biological prior evidence receive weights $w_i > 1$, while low-prior background features receive weights $w_i < 1$.

Define the weighted $p$-value as:

Apply the standard Benjamini-Hochberg procedure to the weighted $p$-values:

1. Sort weighted $p$-values: $p_{(1)}^w \le p_{(2)}^w \le \dots \le p_{(m)}^w$.
2. Find the largest index $k$ such that:

1. Reject all corresponding hypotheses $H_{(1)}, \dots, H_{(k)}$.

Theorem (Genovese, Roeder, Wasserman 2006): If the weights $w_i$ are determined independently of the current test statistics (e.g., from prior historical literature, external databases, or sample splitting), the weighted BH procedure strictly controls the False Discovery Rate at level $q^*$:

Weighting boosts discovery power by $40\%$ to $200\%$ on biologically informative candidates without introducing false positive bias.

In modern big data analytics, black-box machine learning algorithms (such as deep neural networks, Random Forests, and XGBoost) generate point predictions $\hat{y} = f(\mathbf{x})$. However, point predictions convey zero statistical confidence: in medical diagnostics or autonomous vehicle steering, knowing whether a prediction has $\pm 2\%$ or $\pm 80\%$ uncertainty is a matter of life and death.

Vladimir Vovk, Alexander Gammerman, and Glenn Shafer (2005), followed by Emmanuel Candès, Yaniv Romano, and Rina Foygel Barber (2019), developed Conformal Prediction, generating exact, distribution-free prediction intervals for any arbitrary black-box predictive model.

1. Split Conformal Prediction

Let $(X_1, Y_1), \dots, (X_n, Y_n)$ be training data, and $(X_{n+1}, Y_{n+1})$ be a new unseen test observation. The only assumption required is Exchangeability (satisfied by i.i.d. sampling).

1. Data Splitting: Split the historical data into a training set $\mathcal{D}_{\text{train}}$ of size $n_1$ and a calibration set $\mathcal{D}_{\text{cal}}$ of size $n_2$.

2. Model Training: Fit an arbitrary model $\hat{\mu}(\mathbf{x})$ on $\mathcal{D}_{\text{train}}$.

3. Non-Conformity Scoring: For each observation $i \in \mathcal{D}_{\text{cal}}$, evaluate the prediction residual:

4. Quantile Selection: Let $q_{\alpha}$ be the $\lceil (n_2 + 1)(1 - \alpha) \rceil / n_2$ empirical quantile of the calibration scores $\{E_1, \dots, E_{n_2}\}$.

5. Prediction Band Construction: For a new query point $X_{n+1}$, the conformal prediction interval is:

Theorem (Exact Finite-Sample Conformal Coverage): For any black-box predictive algorithm, any feature dimension $p$, and any underlying probability distribution:

The validity of the confidence band is mathematically exact in finite samples. It requires zero Gaussian assumptions, zero asymptotic expansions, and zero regularity conditions, establishing conformal prediction as the premier uncertainty quantification engine for modern big data.

In Section 8.2, we proved that the standard Benjamini-Hochberg (BH) procedure controls the False Discovery Rate at level $\frac{m_0}{m} q^* \le q^*$ when $p$-values are independent or satisfy Positive Regression Dependency on a Subset (PRDS). However, in dense time series, complex spatial models, or high-dimensional brain networks with negative feedback loops, test statistics can exhibit arbitrary negative correlations.

Yoav Benjamini and Daniel Yekutieli (2001) established the universal BY Procedure for controlling FDR under arbitrary, unknown dependence structures.

The Benjamini-Yekutieli (BY) Step-Up Algorithm

To control FDR at level $q^*$ under arbitrary dependence:

1. Sort the $m$ observed $p$-values: $p_{(1)} \le p_{(2)} \le \dots \le p_{(m)}$.
2. Compute the harmonic correction factor:

where $\gamma \approx 0.577215$ is the Euler-Mascheroni constant.

1. Find the largest index $k$ such that:

1. Reject all hypotheses $H_{(1)}, \dots, H_{(k)}$.

Theorem (Benjamini-Yekutieli 2001): For any dependency structure whatsoever among the $m$ test statistics (including severe negative correlations):

While the factor $c(m) \approx \ln m$ reduces statistical power relative to standard BH (e.g., for $m = 10,000$, $c(m) \approx \ln(10,000) \approx 9.21$, reducing the threshold by roughly a factor of 9), the BY procedure provides an unshakeable mathematical guarantee when dependence structures cannot be verified.

In Unit 8, we synthesized the statistical theory of high-dimensional multiple testing and False Discovery Rate (FDR) control. We formulated the multiple testing crisis, demonstrating that testing $m = 100,000$ hypotheses at $\alpha = 0.05$ produces 5,000 spurious discoveries under the global null. We analyzed the Family-Wise Error Rate (FWER) and the union-bound Bonferroni correction $\alpha/m$, showing that Bonferroni destroys statistical power in big data science. We formulated the False Discovery Rate of Benjamini and Hochberg ($\mathrm{FDR} = \mathbb{E}[V / \max(R, 1)]$) and proved the validity of the step-up thresholding rule $p_{(k)} \le \frac{k}{m} q^*$. We detailed Storey's positive FDR (pFDR) and $q$-value methodology via estimation of $\pi_0$. We examined Bradley Efron's empirical Bayes Local False Discovery Rate ($\mathrm{lfdr}(z) = \pi_0 f_0(z) / f(z)$) and two-groups mixture model. Finally, we studied the Knockoffs framework of Barber and Candès for exact finite-sample FDR control in high-dimensional regression, post-selection inference via the Polyhedral Lemma, Surrogate Variable Analysis (SVA) for latent confounders, weighted FDR with functional biological priors, and split Conformal Prediction for exact distribution-free predictive intervals.

In Unit 8.10, we examined Split Conformal Prediction, which constructs prediction intervals by adding and subtracting a constant residual scalar: $\hat{\mu}(\mathbf{x}) \pm q_\alpha$. While mathematically exact in marginal coverage, this constant-width interval fails under Heteroskedasticity: if the data variance $\sigma^2(\mathbf{x})$ is small when $x \approx 0$ but large when $x \gg 0$, a constant-width interval will be unnecessarily wide for small $x$ and dangerously narrow for large $x$, failing conditional coverage.

Yaniv Romano, Matteo Sesia, and Emmanuel Candès (2019) introduced Conformalized Quantile Regression (CQR), marrying quantile regression with conformal prediction to deliver variable-width adaptive intervals with exact finite-sample guarantees.

1. Dual Quantile Training

1. Train two non-parametric quantile regressors on the training set:

• $\hat{q}_{\alpha/2}(\mathbf{x})$: estimates the lower conditional quantile $\tau_1 = \alpha/2$ (e.g., $2.5\%$).
• $\hat{q}_{1 - \alpha/2}(\mathbf{x})$: estimates the upper conditional quantile $\tau_2 = 1 - \alpha/2$ (e.g., $97.5\%$).

These models can be fitted via Pinball loss minimization using Gradient Boosting (LightGBM/XGBoost) or deep neural networks.

2. Conformal Non-Conformity Scoring

For each observation $i$ in the calibration set $\mathcal{D}_{\text{cal}}$, evaluate the signed conformity score measuring how far the observation falls outside the quantile band:

• If $Y_i$ falls inside the predicted band $[\hat{q}_{\alpha/2}, \hat{q}_{1-\alpha/2}]$, $E_i \le 0$.
• If $Y_i$ falls outside the band, $E_i > 0$ measures the distance to the nearest boundary.

3. Conformalized Adaptive Prediction Interval

Let $Q_{1-\alpha}(E, \mathcal{D}_{\text{cal}})$ be the $(1-\alpha)(1 + 1/|\mathcal{D}_{\text{cal}}|)$ empirical quantile of the calibration scores. For any new query point $\mathbf{x}_{n+1}$, the CQR prediction interval is:

Theoretical Guarantee: The CQR interval guarantees exact finite-sample coverage $P(Y_{n+1} \in \mathcal{C}(\mathbf{x}_{n+1})) \ge 1 - \alpha$, while dynamically expanding in volatile regions and contracting in quiet regions, providing the state-of-the-art framework for uncertainty quantification in modern machine learning.

In this laboratory case study, we simulate a Genome-Wide Association Study (GWAS) testing $m = 10,000$ genetic variants on $n = 1,000$ individuals. Among the 10,000 variants, exactly $m_1 = 50$ variants have true causal biological effects on the clinical phenotype, while $9,950$ variants are pure Gaussian noise ($m_0 = 9,950$). We evaluate four multiple testing procedures at nominal significance level $\alpha = q^* = 0.05$: The simulation vividly illustrates the core thesis of modern high-dimensional inference: uncorrected testing generates over $90\%$ false discoveries, Bonferroni eliminates false discoveries at the cost of discarding $72\%$ of true signals, while the Benjamini-Hochberg FDR procedure triples statistical power while strictly bounding false discovery rates below $5\%$.

In Section 8.2, we stated that the Benjamini-Hochberg (BH) procedure controls the False Discovery Rate under the Positive Regression Dependency on a Subset (PRDS) property. Here, we present the mathematical definition and theoretical proof sketch.

Definition of PRDS

A vector of test statistics $\mathbf{X} = (X_1, \dots, X_m)$ is said to satisfy Positive Regression Dependency on a Subset $I_0$ (where $I_0$ is the set of true null hypotheses) if for any increasing set $D \subseteq \mathbb{R}^m$ and for any index $i \in I_0$:

PRDS is satisfied by:

• Multivariate normal test statistics with non-negative correlation matrix ($\Sigma_{jk} \ge 0$).
• Absolute Student-$t$ statistics sharing a common denominator variance estimate.
• One-sided tests comparing treatment groups against a common control group (Dunnett's test).

The Benjamini-Yekutieli Proof of FDR Control under PRDS

Let $V$ be the number of false discoveries and $R$ the total discoveries.

Condition on the $p$-value of true null hypothesis $i$: $P_i = p$. Under the BH step-up rule, $H_i$ is rejected if and only if $P_i \le \frac{R}{m} q^*$, which is equivalent to $R \ge \frac{m P_i}{q^*}$.

Summing over $p$ and using the PRDS property to bound the conditional probabilities yields:

This rigorous proof guarantees that in genetic association studies with non-negative linkage disequilibrium (LD) blocks, the Benjamini-Hochberg procedure strictly controls FDR.

In functional Magnetic Resonance Imaging (fMRI) brain research, scanners record Blood-Oxygen-Level-Dependent (BOLD) signals across a three-dimensional grid of $p \approx 200,000$ voxels over time. Because neighboring brain voxels share extensive biological connectivity and hemodynamic vascular flow, voxel-level test statistics exhibit intense local spatial correlation.

Karl Friston, Keith Worsley, and colleagues (1994, 1996) formulated Random Field Theory (RFT) for statistical parametric mapping.

The Euler Characteristic Heuristic

Instead of treating voxels as independent hypotheses, RFT models the 3D test statistic map as a continuous Gaussian random field $X(\mathbf{r})$ on a spatial manifold $S \subset \mathbb{R}^3$. When the random field is thresholded at height $u$, the excursion set $A_u = \{ \mathbf{r} \in S : X(\mathbf{r}) \ge u \}$ consists of isolated islands or clusters.

The topological Euler Characteristic $\chi(A_u)$ counts:

At high thresholds $u \gg 0$, holes and voids disappear, and the Euler characteristic equals the exact count of activated brain clusters!

Theorem (Worsley's Tube Formula): The Family-Wise Error Rate of detecting at least one spurious brain cluster is given by the expected Euler characteristic:

where $R_d(S)$ are the geometric Resel (resolution element) counts of the brain search volume, and $\rho_d(u)$ are universal Euler characteristic density polynomials. RFT controls family-wise error across continuous spatial dimensions, providing the regulatory foundation for cognitive neuroscience publications.

When reporting statistical discoveries in large-scale genomic, proteomic, and neuroimaging studies:

1. Never Report Uncorrected $p$-Values: In studies testing $m \ge 1,000$ hypotheses, reporting raw $p \le 0.05$ as evidence of discovery is statistical malpractice that generates massive false discovery proportions.

2. Default to the Benjamini-Hochberg Procedure at 5% FDR: The BH procedure provides the optimal balance of rigorous false discovery control and high statistical discovery power. If positive regression dependency cannot be assumed, deploy the Benjamini-Yekutieli correction.

3. Report Both FDR Adjusted $p$-Values and Storey $q$-Values: Provide complete volcano plots displaying both effect sizes (log fold change) and $q$-values ($-\log_{10} q$) to provide transparent, reproducible evidence for regulatory and peer-reviewed scientific evaluation.

Standard high-dimensional multiple testing procedures assume exchangeability among hypotheses or treat hypothesis indices as an unordered set. In biological imaging, geospatial surveillance, and neuroimaging (fMRI), test statistics exhibit prominent spatial autocorrelation: neighboring voxels or coordinates share latent physiological noise and true signals form contiguous clusters rather than isolated spikes.

Applying standard BH under spatial clustering can lead to high variability in empirical false discoveries. Topological approaches based on Gaussian Random Fields establish cluster-extent thresholding:

1. Define a primary voxel-level intensity threshold $u$ (such as $p < 0.001$).
2. Identify contiguous connected clusters of suprathreshold voxels exceeding spatial volume threshold $k$.
3. Compute cluster-level significance using the Euler characteristic heuristic of Gaussian excursion sets:

where $\mathcal{L}_d(\Omega)$ represents the Lipschitz-Killing curvature measures of the spatial search domain $\Omega$, and $\rho_d(u)$ represents the Euler characteristic density of dimension $d$. This spatial topological correction achieves rigorous Family-Wise Error Rate or Cluster-FDR control while multiplying statistical power relative to unadjusted voxelwise testing.