Adaptive-mixture-categorization (AMC)-based g-computation and its application to trace element mixtures and bladder cancer risk

Siting Li, Margaret R. Karagas, Brian P. Jackson, Michael N. Passarelli, Jiang GuiScientific 2022 · DOI

Abstract

Identifying the health impacts of complex exposure mixtures remains a significant challenge in environmental epidemiology. While methods like Weighted Quantile Sum (WQS) regression and quantile-based g-computation have made strides, they often rely on rigid, pre-defined exposure categories. This paper introduces Adaptive-Mixture-Categorization (AMC)-based g-computation, which optimizes exposure thresholds using data-driven search criteria. In simulation studies, AMC-based g-computation demonstrated superior performance in reducing bias and increasing true positive rates compared to traditional methods. Application to the New Hampshire Bladder Cancer Study revealed a significant association between medium-level toenail zinc and bladder cancer risk, a non-linear relationship that standard quantile-based models might overlook.

Introduction: Navigating the "Chemical Soup"

Humans are rarely exposed to environmental contaminants in isolation. Instead, we encounter a "chemical soup" of air pollutants, heavy metals, and synthetic chemicals. Modeling these mixtures is difficult because traditional statistical methods often assume linear relationships or require all components to act in the same direction.

Existing tools like Weighted Quantile Sum (WQS) regression assume "directional homogeneity"—that all elements in a mixture are either harmful or protective. Quantile-based g-computation relaxed this assumption but still relied on fixed "buckets" (e.g., quartiles). If a chemical’s toxic effect occurs at a specific biological threshold not aligned with these arbitrary divisions, the signal is often lost to noise. This paper introduces a method to let the data define those thresholds automatically.

Methodology: The AMC Approach

The proposed Adaptive-Mixture-Categorization (AMC)-based g-computation framework allows the data itself to define the most meaningful exposure categories. The core innovation lies in the AMC procedure:

  1. Optimal Search: Rather than splitting data into equal-sized groups, the algorithm employs a linear search using the F-statistic to identify optimal boundaries for each exposure variable.
  2. Variance Optimization: The method seeks to minimize variance within categories while maximizing variance between them, ensuring that the resulting groups represent distinct exposure profiles.
  3. G-computation Integration: These optimized categories are then integrated into a g-computation model to estimate the joint effect of the mixture, allowing for bidirectional associations (both positive and negative) simultaneously.

Simulation Performance

To validate the method, researchers conducted extensive simulations comparing AMC-based g-computation against WQS and standard quantile-based g-computation. Across various scenarios, the AMC approach consistently achieved lower bias and higher True Positive Rates (TPR).

Figure 1: Comparison of estimate bias across four simulation scenarios for AMC-based g-computation, Quantile-based g-computation, and WQS.

The flexibility of AMC proved particularly advantageous for non-monotonic effects, where the relationship between exposure and outcome changes direction or intensity at specific thresholds.

Application: Trace Elements and Bladder Cancer

The method was applied to data from the New Hampshire Bladder Cancer Study, involving 265 cases of non-muscle invasive bladder cancer (NMIBC) and 353 controls. The analysis focused on 12 trace elements measured in toenail samples.

A key finding was the "inverted U-shaped" relationship between zinc and cancer risk. AMC-based g-computation identified that medium-level toenail zinc (116.7–145.5 μg/g) was significantly associated with increased NMIBC risk ($w = 0.172$; 95% CI 0.016 to 0.254), whereas the highest levels did not show the same statistical significance.

Figure 2: Weights and 95% confidence intervals for 12 trace elements at medium and high exposure levels, highlighting the significant association of medium-level zinc.

While specific elements showed associations, the joint effect of all 12 elements as a single mixture did not reach statistical significance in this specific cohort (OR = 1.88; 95% CI 0.25 to 2.83).

Conclusion

AMC-based g-computation offers a powerful alternative for environmental health researchers by replacing arbitrary quantiles with data-driven thresholds. By capturing non-linear and threshold-specific effects, this method provides a more nuanced understanding of how complex mixtures influence disease risk. Future work will explore its application to other health outcomes where non-linear threshold effects are likely common.

Summary updated

Abstract

Several new statistical methods have been developed to identify the overall impact of an exposure mixture on health outcomes. Weighted quantile sum (WQS) regression assigns the joint mixture effect weights to indicate the overall association of multiple exposures, and quantile-based g-computation is a generalized version of WQS without the restriction of directional homogeneity. This paper proposes an adaptive-mixture-categorization (AMC)-based g-computation approach that combines g-computation with an optimal exposure categorization search using the F statistic. AMC-based g-computation reduces variance within each category and retains the variance between categories to build more powerful predictors. In a simulation study, the performance of association analysis was improved using categorizing by AMC compared with quantiles. We applied this method to assess the association between a mixture of 12 trace element concentrations measured from toenails and the risk of non-muscle invasive bladder cancer. Our findings suggested that medium-level (116.7-145.5 μg/g) vs. low-level (39.5-116.2 μg/g) of toenail zinc had a statistically significant positive association with bladder cancer risk.
BibTeX
@inproceedings{li2022adaptive,
  author = {Siting Li and Margaret R. Karagas and Brian P. Jackson and Michael N. Passarelli and Jiang Gui},
  title = {{Adaptive-mixture-categorization (AMC)-based g-computation and its application to trace element mixtures and bladder cancer risk}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-21747-7},
}
Figure