POI-Anchored Residuals in Model Diagnostics

• POI-Anchored Residuals are specialized diagnostics that focus on specific points-of-interest to evaluate localized model fit and robustness under distributional shifts.
• They leverage anchor regression and spatial analysis techniques by penalizing deviations at key locations to enforce residual invariance across environments.
• This approach is applied to enhance regression models and spatial point process assessments, ensuring reliable diagnostics and improved out-of-distribution performance.

Point-Of-Interest (POI)-anchored residuals are specialized residual diagnostics that evaluate the fit of statistical or machine learning models at specific locations or under specific source conditions, with applications spanning generalized regression, spatial point process modeling, and out-of-distribution (OOD) robustness contexts. In both causal regression and spatial statistics, POI-anchoring focuses residual analysis or invariance specifically on subsets—such as anchor variables in causal models or fixed spatial locations in point processes—to obtain localized or distributionally robust diagnostic power.

1. Fundamental Concepts of POI-Anchored Residuals

POI-anchored residuals extend classical residual-based diagnostics by focusing on specific, user-selected indices—referred to as “points of interest” (POIs)—rather than aggregating or averaging over the full sample. In causal OOD generalization, the POIs are typically values or functions of exogenous “anchor” variables whose distribution may change across environments. In spatial point process analysis, POIs correspond to fixed spatial locations where the adequacy of the model is assessed by direct comparison of observed events to model-predicted local intensity.

This anchoring enables diagnostics or regularization that target robustness or local fit at user-specified, operationally-relevant loci. In anchor regression, POI-anchored residuals are key to enforcing residual invariance across anchor environments, directly linking model calibration to OOD generalization (Kook et al., 2021). In spatial analysis, POI-anchored residual fields sharpen model checks at informative or high-value spatial sites (Baddeley et al., 2012).

2. POI-Anchored Residuals in Distributional Anchor Regression

Distributional Anchor Regression generalizes the anchor regression framework by combining a fully parametric transformation model for the conditional distribution with a causal penalty on POI-anchored (score) residuals (Kook et al., 2021). The approach is summarized as follows:

• The conditional distribution of the response $Y$ given covariates $X$ is modeled as:

$$F_{Y\mid X}(y\mid x;\,\theta) = F_0(h(y) - \eta(x;\beta))$$

where $F_0$ is a predetermined CDF (e.g., Gaussian, logistic), $h(y)$ is a baseline transformation (basis expansion, step function for ordinal data), and $\eta(x;\beta)$ is typically linear.

• The score residual at a data point $(y_i, x_i)$ is defined by introducing a fictitious intercept $-\alpha$ on the transformation scale and computing the derivative of the log-likelihood with respect to $\alpha$ at zero:

$$r_i(\theta) = \left.\frac{\partial}{\partial\alpha}\ell(\theta, \alpha; y_i, x_i)\right|_{\alpha=0}$$

• Letting $A_i$ denote anchor variables (exogenous source nodes), POI-anchoring targets invariance of residuals with respect to $A$. The anchor penalty penalizes the $L_2$ norm of the projection of residuals onto the anchor space:

$$\lambda \|P_A r(\theta)\|_2^2 \quad \text{where} \quad P_A = A (A^\top A)^{-1} A^\top$$

• The complete objective is

$$\theta^*(\lambda) = \arg\min_\theta \left\{ -\frac{1}{n} \sum_{i=1}^n \ell(\theta; y_i, x_i) + \lambda\frac{1}{n}\sum_{i=1}^n r_i(\theta)^2 \right\}$$

This structure guarantees, in the population limit, $$\mathbb{E}[r(\theta^*) | \mathrm{do}(A=a)] \approx 0$$ for all possible anchor interventions, enforcing distributional robustness for shifts in $A$ and thus across OOD environments.

3. Computation and Interpretation in Spatial Point Process Models

POI-anchored residuals in spatial point process models are constructed by comparing observed points to the fitted Papangelou conditional intensity at pre-specified spatial POIs (Baddeley et al., 2012). The key diagnostics are:

• Raw (martingale) residual:

$r(u_j) = I(u_j \in X) - \hat\lambda(u_j; X)$

where $I(u_j \in X)$ indicates if an event occurred at $u_j$ and $\hat\lambda(u_j; X)$ is the fitted conditional intensity at $u_j$.

• Pearson and deviance residuals are computed as pointwise functions of the same innovation.
• For pseudo-residuals corresponding to a summary $S(X) = \sum_j I(u_j \in X)$, one computes:

$R_S = \sum_j [I(u_j) - \hat\lambda(u_j; X)]$

These residuals can be aggregated, smoothed, or visualized at the POIs, and statistical tests can be constructed by normalizing their sum relative to their model-implied variance.

Residual Type	Formula at $u_j$	Purpose
Raw (martingale)	$I(u_j \in X) - \hat\lambda(u_j; X)$	Direct innovation, event vs intensity
Pearson	$$\frac{I(u_j \in X) - \hat\lambda(u_j; X)}{\sqrt{\hat\lambda(u_j; X)}}$$	Variance normalization
Deviance	$$\operatorname{sign}[I(u_j) - \hat\lambda(u_j)] \sqrt{2[I(u_j)\log(I(u_j)/\hat\lambda(u_j)) - (I(u_j)-\hat\lambda(u_j))]}$$	Likelihood-based distance

A theoretical justification is provided by the spatial martingale property: under the true model, $r(u)$ has mean zero, and integrating $r(u)$ against test functions provides sensitivity to local and global lack-of-fit.

4. Causal and Structural Invariance via POI-Anchored Residual Regularization

Imposing POI-anchored residual invariance underpins the causal guarantee for OOD generalization in anchor regression and its distributional analogues (Kook et al., 2021). By minimizing the squared anchor-projected residuals, the learning algorithm ensures that model-prediction errors are uninformative about exogenous anchor values in any environment, i.e.,

$$\mathbb{E}[r(\theta^*) | \mathrm{do}(A=a)] \approx 0 \quad \forall a$$

This residual invariance holds not only in expectation but also, for common score residuals, with covariance bounded by the anchor penalty. In the linear-Gaussian limit, this approach interpolates between ordinary least squares (no anchor penalty) and two-stage least squares (infinite penalty), yielding models that are robust to anchor-induced distributional shifts.

5. Implementation, Applications, and Inference

For distributional anchor regression, score residuals are computed using either closed-form expressions (for probit, logistic, or martingale types) or via automatic differentiation over transformation models, and the anchor penalty is optimized via convex (or bi-convex in the min-max form) programming.

In spatial point process models, the procedure involves:

• Fitting a model (e.g., Gibbs/Markov or maximum/pseudo-likelihood) to the observed point pattern.
• Selecting POIs and calculating the fitted conditional intensity at those locations.
• Computing raw, Pearson, or deviance residuals at each POI.
• Aggregating or smoothing residuals to construct formal or informal fit diagnostics.

POI-anchored residuals are implemented in statistical software, such as the spatstat package in R, which provides direct functionality for computing, visualizing, and aggregating POI-anchored residuals (Baddeley et al., 2012).

6. Theoretical Guarantees and Limitations

Theoretical support relies on martingale difference properties (for spatial models) and causal invariance results (for anchor regression). Under the true model, POI-anchored residuals are mean zero and independent of anchor/environment, yielding reliable OOD generalization under structural equation model assumptions.

A plausible implication is that POI-anchored residuals provide principled tools for diagnosing or correcting localized model misspecification where the distribution of input variables (anchors or spatial location) changes between training and test scenarios. However, the effectiveness of POI-anchored regularization is contingent on accurate specification of the data-generating structure—specifically, the correct identification of anchors as exogenous sources.

7. Connections to Broader Residual and Model Assessment Theory

POI-anchored residuals are conceptually related to, but distinct from, general residual-based methods in model assessment. Unlike aggregate diagnostics—such as those based on cumulative distribution of residuals—POI-anchored approaches provide targeted, interpretable assessment at prespecified loci. In high-dimensional or causal settings, this aligns with recent advances in robust regression that exploit exogenous leverages for fine-grained control of model sensitivity to environment changes (Kook et al., 2021). In spatial and functional data analysis, POI-anchoring extends diagnostics from global fit to spatially inhomogeneous or structure-guided checks (Baddeley et al., 2012).

In summary, POI-anchored residuals represent a key methodological tool for localized, robust model assessment and causal OOD generalization, grounded in rigorous statistical theory and supported by software implementations in both regression and spatial statistics.