📊 statistics

Inverse sampling intensity weighting for preferential sampling adjustment

This paper proposes and evaluates inverse sampling intensity weighting (ISIW) as a computationally efficient, two-stage alternative to complex latent variable models for adjusting preferential sampling in geostatistics, demonstrating its superior predictive performance under design misspecification through applications in moss biomonitoring and air quality data.

Original authors: Thomas W. Hsiao, Lance A. Waller

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thomas W. Hsiao, Lance A. Waller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Biased Reporter"

Imagine you are trying to map the temperature of an entire city. You want to know the average temperature everywhere, not just where you have thermometers.

In a perfect world, you would place your thermometers randomly across the city. But in the real world, people don't do that. They place thermometers where it's convenient or where they expect something interesting to happen.

The Scenario: Imagine you want to measure air pollution. You might place your sensors near busy highways because you expect the air to be bad there. Or, you might place them in a park because it's easy to access.
The Issue: This is called Preferential Sampling (PS). Your data isn't random; it's "biased" toward certain areas. If you just take the average of your sensors, you might think the whole city is very polluted (if you only measured near highways) or very clean (if you only measured in parks). You are getting a distorted picture of reality.

The Old Solution: The "Shared Secret" Model

For years, statisticians have tried to fix this using a complex method called the Shared Latent Process (SLP) model.

The Analogy: Imagine the pollution (the truth) and the placement of your sensors (the bias) are two actors in a play who are secretly reading from the same script. The old method tries to figure out that secret script. It assumes that the reason you put a sensor in a specific spot is directly linked to the pollution level there.
The Problem: This method is like trying to solve a Rubik's cube while wearing blindfolds. It is mathematically heavy, slow, and requires you to guess exactly how the "script" works. If you guess the script wrong (which happens often in real life), the whole solution falls apart. It's also very hard to compute for large maps.

The New Solution: "Inverse Sampling Intensity Weighting" (ISIW)

The authors of this paper propose a smarter, faster, and more flexible way to fix the bias. They call it ISIW.

The Analogy: Instead of trying to guess the secret script, ISIW acts like a smart editor looking at where the reporters (sensors) actually went.
1. Step 1: Map the Bias. First, the method looks at the map of sensors and asks: "Where are the reporters clustering? Where are they sparse?" It creates a map of "sampling intensity." If 50 sensors are packed into one tiny neighborhood and only 1 is in a huge forest, the method knows the neighborhood is "over-reported."
2. Step 2: The Weighted Vote. Next, it applies a "weight" to every data point.
  - If a sensor is in a crowded, over-sampled area, its voice is muted (given a low weight).
  - If a sensor is in a lonely, under-sampled area, its voice is amplified (given a high weight).
3. Step 3: The Result. By listening to the lonely sensors more and the crowded ones less, the final map becomes a fair representation of the whole city, even though the sensors weren't placed randomly.

Why This Paper is Special

The authors didn't just invent the idea of weighting; they made it fast and robust.

Speed (The Vecchia Approximation): The old methods were like trying to count every grain of sand on a beach to find the average. The new method uses a trick called the "Vecchia approximation." Imagine instead of counting every grain, you count a few representative handfuls and use math to estimate the rest. It's incredibly fast and accurate, allowing them to handle huge datasets that would crash the old computers.
Robustness (The "What If" Test): The authors tested their method against the old "Shared Secret" model in many different scenarios.
- The Finding: When the "secret script" (the old model) was guessed correctly, both methods worked well.
- The Twist: But when the "script" was guessed wrong (which happens often in real life), the old method failed miserably. The new ISIW method, however, kept working perfectly. It didn't need to know the secret script; it just needed to know where the sensors were crowded.
The Surprise: The authors found something counter-intuitive. Usually, in statistics, you need to estimate the "rules" of the game perfectly to win. Here, they found that you don't need to perfectly estimate the rules to get a good prediction. You can have a slightly "wrong" understanding of the math, but if your weighting (the editor's voice) is right, your map of the city will still be accurate.

Real-World Tests

The authors tested this on two real-world maps:

Moss in Spain: Measuring lead pollution in moss. The sensors were placed preferentially (biased). ISIW fixed the map better than the old methods.
Air Quality in California: Measuring PM2.5 (fine dust). Again, sensors were clustered in cities. ISIW provided a more accurate picture of the pollution across the whole state.

The Bottom Line

This paper introduces a new tool for geographers and scientists. It says: "Don't waste time trying to guess why your data is biased. Just look at where the data is crowded, and give the lonely data points a louder voice."

It is faster, easier to use, and more reliable than the traditional methods, especially when the real world doesn't follow the perfect mathematical rules we hope for.

Technical Summary: Inverse Sampling Intensity Weighting for Preferential Sampling Adjustment

Problem Statement
Traditional model-based geostatistics assumes that observation locations ( $X$ ) are fixed and independent of the spatial process of interest ( $S$ ). When this assumption is violated—termed preferential sampling (PS)—standard inference and prediction can suffer from substantial bias. The prevailing solution, the Shared Latent Process (SLP) model (Diggle et al., 2010), induces dependence between $X$ and $S$ via a shared latent variable. While effective when the sampling design is correctly specified (e.g., as a Log-Gaussian Cox Process), the SLP framework faces significant practical limitations: it requires complex joint likelihood specification, is computationally intensive for prediction (requiring posterior summarization for every prediction point), and is sensitive to misspecification of the underlying point process. Furthermore, few off-the-shelf software implementations exist, and integrating SLP with modern scalable Gaussian process (GP) approximation techniques remains challenging.

Methodology
The authors propose Inverse Sampling Intensity Weighting (ISIW) as a computationally efficient, two-stage alternative for PS adjustment. Unlike the SLP, which models the full joint likelihood of the response and observation process, ISIW adjusts for dependence by reweighting observations based on their estimated sampling intensity.

Sampling Intensity Estimation: The first stage estimates the sampling intensity $\lambda(x)$ at observation locations. The authors explore nonparametric Kernel Intensity Estimators (KIE) and a variant incorporating a predicted response surface as a spatial covariate ("ISIW KIE COV"). To mitigate numerical instability caused by extreme weights in sparsely sampled regions, a winsorization procedure caps weights at a specified quantile (93% in their analysis).
Weighted Likelihood Estimation: In the second stage, the estimated inverse intensities are used as weights within a composite likelihood framework. The authors specifically implement ISIW using the Vecchia approximation, a composite likelihood method that decomposes the joint density into a product of conditional distributions. This choice allows for the estimation of full mean and covariance parameters while maintaining scalability to large datasets, addressing limitations of previous ISIW studies that relied on univariate or pairwise composite likelihoods.
Prediction: Final predictions are generated by substituting the ISIW-adjusted parameter estimates into standard kriging equations. This contrasts with the SLP, which requires computing the conditional expectation $E[S(X_0)|X, Y]$ from the joint posterior.

Key Contributions and Experimental Design
The paper provides a comprehensive evaluation of ISIW against Maximum Likelihood Estimation (MLE) and the SLP (implemented via INLA-SPDE) across three simulation studies and two real-world applications:

Simulation 1 (Point Process Misspecification): Data generated from Log-Gaussian Cox (LGCP) and Thomas processes. The study tests robustness when the SLP model is misspecified (e.g., fitting an LGCP-based SLP to Thomas process data).
Simulation 2 (Random Field Misspecification): Data generated from a sum of multiple Gaussian processes with heterogeneous smoothness, violating the single Matérn GP assumption.
Simulation 3 (Real Data Calibration): Simulations calibrated to the Galicia moss and California PM2.5 datasets to assess performance in realistic settings.
Real Data Applications: Analysis of lead concentrations in moss from Galicia, Spain (1997 preferential data) and PM2.5 concentrations from the California EPA Air Quality System (2010).

Results

Predictive Performance: ISIW consistently outperformed MLE across all scenarios. Crucially, ISIW demonstrated superior predictive accuracy compared to the SLP when the sampling design was misspecified (e.g., under the Thomas process or complex latent fields). In these cases, the SLP often suffered from numerical instability or severe underestimation of the mean.
Parameter Estimation vs. Prediction: A counter-intuitive finding was the disconnect between parameter estimation accuracy and predictive performance. While MLE often provided the most accurate estimates of covariance parameters (range, nugget) and the SLP performed well under correct specification, ISIW frequently yielded better predictions despite having higher bias in covariance parameter estimation. This suggests that accurate estimation of the mean ( $\mu$ ) and the application of intensity weights are more critical for kriging performance under PS than precise recovery of covariance parameters.
Computational Efficiency: ISIW combined with the Vecchia approximation was significantly faster than the SLP (INLA-SPDE), particularly for prediction tasks, avoiding the need for posterior sampling at every grid point.
Real Data: In the Galicia and California applications, ISIW produced smoother predictive surfaces than MLE and adjusted predictions similarly to the SLP, though the SLP tended to apply more extreme corrections in data-sparse regions.

Significance and Claims
The authors claim that ISIW offers a practical, intuitive, and fast method for adjusting preferential sampling in geostatistics. Its primary significance lies in its robustness to model misspecification; unlike the SLP, which relies on a fully parametric specification of the observation process, ISIW relies on nonparametric intensity estimation, making it more adaptable to complex, real-world sampling designs where the true point process is unknown or difficult to emulate.

The paper highlights that effective spatial prediction under PS does not necessarily require the complex joint modeling of response and observation processes. Instead, weighting observations by their inverse sampling intensity within a scalable likelihood framework (Vecchia) can yield superior predictive performance, particularly when the underlying sampling mechanism deviates from standard assumptions. The authors conclude that while uncertainty quantification for ISIW remains a challenge, the method represents a viable alternative to the SLP for applied geostatistical problems where computational speed and model robustness are prioritized.