Weighted Scan Statistics

• Weighted scan statistics are methods that detect structured anomalies by scanning data regions with weights adjusted for size, position, and other covariates.
• They incorporate variants such as the average likelihood ratio, penalized scans, and weighted U-statistics to balance sensitivity across different scales.
• Their implementation leverages calibrated thresholds and efficient computational techniques to control errors and enhance detection in high-dimensional settings.

Weighted scan statistics are a class of statistical techniques designed for the detection of structured signals or anomalies in data by systematically scanning over regions—intervals, time points, graph substructures, or other index sets—using test statistics that adjust for region size or location. Weighted scan statistics are distinguished from classical scan statistics by the explicit or implicit incorporation of weighting schemes to improve sensitivity, address multiple testing across scales, or control for edge and size effects in the underlying data structure.

1. Fundamental Principles of Weighted Scan Statistics

Weighted scan statistics generalize the maximum likelihood ratio scan by augmenting or modifying the local statistics for each region with either explicit weights (as functions of region size, position, or other covariates) or by aggregating statistics in a way that compensates for the heterogeneity in region contributions. The motivation arises from the observation that unweighted scan statistics often provide greatest sensitivity for the smallest possible regions (intervals, subgraphs, early/late change-points), leading to suboptimal or diluted power for large-extent signals or anomalies.

Core Model Structure

Consider a canonical univariate change-point detection model: $Y_i = f_n(i/n) + Z_i, \quad i=1,\ldots,n,$ where $Z_i \stackrel{\text{i.i.d.}}{\sim} N(0,1)$ and the signal $f_n$ is supported on some unknown interval $I_n \subset [0,1]$. The fundamental problem is hypothesis testing for $H_0: f_n = 0$ versus $H_1: f_n \neq 0$, with the complication that the spatial extent and location of the signal are unknown.

For each candidate region $I$, the local log-likelihood ratio (LLR) is

$$\text{LLR}(I) = \frac{1}{2} Y_n(I)^2, \quad Y_n(I) = \frac{\sum_{i \in I} Y_i}{\sqrt{|I|}},$$

where $|I|$ denotes the size of the region.

Weighted scan approaches apply to this and broader settings by altering aggregation (as in the average likelihood ratio), thresholding, explicit penalty terms, or through structural weights appropriate for graph or high-dimensional data.

2. Weighted Scan Constructions: Classical, Average, and Penalized Variants

Weighted scan test statistics can be grouped according to their operational principles and model context.

Interval Scan and Average Likelihood Ratio (ALR) Statistics

• Scan Statistic (Maximum Likelihood Ratio):

$M_n = \max_{I} Y_n(I)$

This statistic, by maximizing over all intervals $I$, inherently favors short intervals due to the scaling by $1/\sqrt{|I|}$; the maximum is typically attained on the smallest possible region, resulting in a detection threshold $\sqrt{2\log n}$ independent of region size.

• Average Likelihood Ratio Statistic:

$$A_n = \frac{1}{|\mathcal{J}_n|} \sum_{I \in \mathcal{J}_n} \exp\left(\frac{1}{2} Y_n(I)^2\right), \quad |\mathcal{J}_n| \sim n^2/2,$$

Averages the exponentiated LLRs over all intervals, thereby diluting the impact of isolated small-scale anomalies but accumulating evidence for large-scale anomalies spread over many regions.

• Weighted/Blocked/Penalized Scans:

Penalized scan variants subtract a scale-dependent penalty from the local statistic, e.g.,

$$P_n = \max_{I} \left( Y_n(I) - \sqrt{2\log\frac{en}{|I|}} \right).$$

Blocked scans group intervals by size and calibrate critical values by block, aiming for more uniform power across scales.

Weighted U-Statistic Scan for Change-Point Detection

Define for observations $X_1, ..., X_n$ and anti-symmetric kernel $h$,

$$U_n^{\,w}(k) = \sum_{i=1}^k \sum_{j=k+1}^n w\left(\frac{i}{n}, \frac{j}{n}\right) h(X_i, X_j).$$

Weights of the form $w_\gamma(u, v) = [u(1-v)]^{-\gamma}$, with $0 \leq \gamma \leq 1/2$, increase sensitivity near the time series boundaries by inflating the statistic when the candidate change-point is near the beginning or end of the sample.

Weighted Graph Scan Statistics

For multivariate data indexed by nodes in a graph $G = (V, E)$, define a family $\mathcal{R}$ of regions (often ball-subgraphs), fit region-by-region groupwise manifold GLMs, and for each region $R$ compute a standardized score

$\bar L_R = \frac{L_R - |E(R)|}{\sqrt{|E(R)|}},$

where $L_R$ measures the squared difference in groupwise trajectory slopes. The weighted scan corrects for region-size by subtracting

$w(R) = 2\sqrt{\log\frac{|E|}{|E(R)|}}$

to yield the final region score $T(R) = \bar L_R - w(R)$.

3. Detection Boundaries, Power, and Theoretical Guarantees

Detection boundaries describe the minimum signal strength required for consistent identification of the target under alternatives.

Detection Thresholds for Interval Scans

The minimax threshold over all tests is

$$u_n\,\sqrt{|I_n|} \asymp \begin{cases} \sqrt{2\log(1/|I_n|)}, & |I_n| \to 0 \ b_n \to \infty, & |I_n| \not\to 0 \end{cases}$$

where $u_n$ is the amplitude and $|I_n|$ is the extent.

• Scan statistic $M_n$: Detection boundary is $\sqrt{2\log n}$, optimal for $|I_n| \approx 1/n$.
• ALR $A_n$: Achieves $\sqrt{2\log(1/|I_n|)}$ on all large scales; slightly suboptimal on smallest scales but with numerically negligible gap for realistic $n$.
• Condensed ALR: Achieves minimax detection $\sqrt{2\log(1/|I_n|)}$ on all scales with computational efficiency $O(n\log^2 n)$.

Weighted Scan in U-Statistic Change-Point Detection

Under regularity conditions (including Rosenblatt $\alpha$-mixing, bounded and anti-symmetric kernels, moment and variation conditions), the weighted U-statistic scan reaches an asymptotic extreme-value (Gumbel) limiting law under $H_0$: $$\frac{a_n}{\sigma_h}T_n - b_n \xrightarrow{\mathcal{D}} G_2(x) = \exp(-2e^{-x}),$$ with $a_n = \sqrt{2\log\log n}$, $$b_n = 2\log\log n + \frac{1}{2}\log\log\log n - \frac{1}{2}\log\pi$$. The kernel and weight exponent $\gamma$ modulate the boundary sensitivity and power profile.

Graph Scan Statistics: FWER and Power Control

For the maximum weighted region score $T^* = \max_{R\in\mathcal{R}} T(R)$, under the "Avocado" graph structural assumption, the weak family-wise error rate is controlled at level $\alpha$, and the test is consistent for any region $R_0$ if

$$\Delta_{R_0} \gg 2\sqrt{\log\left(\frac{|E|}{|E(R_0)|}\right)}$$

with $\Delta_{R_0}$ the signal-to-noise ratio and $|E(R_0)|$ region size.

4. Principal Advantages and Sensitivity Profiles

Weighted scan statistics address several critical limitations of classical scan tests:

• Sensitivity Across Scales: Unweighted scan statistics (e.g., $M_n$) are most powerful for smallest scales, but diluted for larger anomalies. Weighted/average or penalized statistics (e.g., ALR, condensed ALR) retain or improve power for broader signals by compensating for dilution.
• Power Transfer in Change-Point Detection: In weighted U-statistic scans, raising $\gamma$ (i.e., increasing weights near endpoints) increases sensitivity for early or late changes but at the expense of slightly reduced power at central positions. Choice of kernel (e.g., Wilcoxon for heavy-tailed distributions) further tunes robustness and power.
• Multiple Testing Correction in Structured Data: In graph-based scan statistics, explicit weighting or penalization (such as $w(R)$ for region size) is critical for controlling family-wise error and preventing inflation of type I error due to region multiplicity.

5. Implementation Guidelines and Computational Aspects

Implementation of weighted scan tests depends on model and data structure.

Canonical Univariate and Interval Scans

• For scan $M_n$, simulate the null threshold $K_n$ using Monte Carlo (asymptotically $\sqrt{2 \log n}$). For ALR, the null distribution must be simulated due to complex dependencies.
• Efficient implementation of condensed ALR exploits cumulative sum arrays and dyadic partitioning, yielding complexity $O(n\log^2 n)$.
• Penalty calibration in penalized/blocked scan variants requires Monte Carlo or analytical extreme-value approximations to set critical values per block or scale.

Weighted U-Statistic Change-Point Tests

• Select kernel $h$ (CUSUM for mean shifts, Wilcoxon for robustness).
• Compute $U_n^{w}(k)$ for all candidate change-points with chosen weight exponent $\gamma$.
• Estimate the long-run variance $\sigma_h^2$ using subsampling or block median estimators, even under the alternative.
• Form the normalized statistic and compare to Gumbel-type critical values for explicit level control.

Graph Structured Scans

• Construct candidate regions (e.g., ball-subgraphs), fit regionwise manifold GLMs for each group, and calculate raw and penalized statistics.
• Enumerate all candidate regions efficiently ($O(D \cdot |V|)$) using precomputed subgraphs and prune search based on interim thresholds.
• Apply the greedy peeling algorithm to ensure non-overlapping discoveries.

6. Empirical Evaluation and Application Domains

Simulation studies and real-world applications demonstrate the utility of weighted scan statistics in diverse domains:

• Interval Scans: The ALR or condensed ALR statistics demonstrate optimal detection power for signals spanning larger extents, outperforming classical scan statistics in these regimes (Chan et al., 2011).
• Change-Point Detection: Weighted U-statistic scans with endpoint-favoring weights excel at detecting boundary changes in time series, confirmed by simulation with normal and heavy-tailed data (Dehling et al., 2020).
• Graph-Based Anomaly Detection: In temporally evolving graphical models, local graph scan statistics identify small subgraphs with groupwise-covariate trends missed by global testing—as shown in applications to gender differences in tobacco usage, baby-name trends, and preclinical Alzheimer’s disease multi-modal imaging studies (Mehta et al., 2017).

A plausible implication is that weighted scan statistics, via judicious choice of weighting and penalization strategies, provide a systematic framework for multi-resolution and structured anomaly detection in high-dimensional, structured, or heterogeneous data settings.

7. Limitations, Calibration, and Considerations

Calibration of critical values for weighted scan tests, particularly those using average or penalized statistics, requires more computational effort than classical unweighted scans, especially under correlated or structured null distributions. For high-dimensional applications (e.g., condensed ALR, graph scans), computational tractability is retained via strategic region thinning or heuristic pruning. However, the selection of weighting schemes or penalties is context-dependent; inappropriate weighting may unduly sacrifice power for certain signal classes, so exploratory analysis or domain knowledge should inform their specification. Additionally, in highly dense or irregular graphs, assumptions such as the Avocado condition may be violated, impacting error control guarantees.

Overall, weighted scan statistics constitute a flexible, theoretically grounded, and practically viable methodology for modern statistical detection problems where the scale, structure, or multiplicity of candidate anomalous regions spans a wide, possibly unknown, range.