Source-linked AI summary
A quantile-based g-computation approach to addressing the effects of exposure mixtures
Alexander P. Keil, Jessie P. Buckley, Katie M. OBrien, Kelly K. Ferguson, Shanshan Zhao, Alexandra J. White
TL;DR
Exposure-mixture analyses need methods that estimate joint effects without relying on restrictive directional, linearity, and additivity assumptions. The paper introduces quantile g-computation and evaluates it against WQS regression through simulations spanning confounding, exposure correlation, non-causal exposures, and non-linear effects. Quantile g-computation generally produced unbiased, appropriately calibrated, and often more precise estimates, while WQS could magnify confounding bias and show poor confidence-interval coverage.
Problem
Exposure-mixture methods such as WQS regression rely on directional homogeneity, linearity, and additivity assumptions that may not hold when exposure effects differ or interact.
Method
The paper introduces quantile g-computation, combining WQS’s simple joint-effect framework with causal-inference methods that allow non-linear and non-additive exposure effects.
Results
Quantile g-computation was unbiased with valid confidence intervals across examined non-null, confounding, and non-linear or non-additive scenarios, while WQS was biased in these settings.
Takeaways & Limitations
Quantile g-computation offers a simple approach for estimating mixture associations when exposure effects may be beneficial, harmful, harmless, or uncertain in direction.
Takeaways & Limitations
The simulations did not assess sample splitting in WQS, and the effects of treating quantized exposures as continuous regressors remain broadly unknown.
Abstract
from arXiv · showhide
Exposure mixtures frequently occur in data across many domains, particularly in the fields of environmental and nutritional epidemiology. Various strategies have arisen to answer questions about mixtures, including methods such as weighted quantile sum (WQS) regression that estimate a joint effect of the mixture components.We demonstrate a new approach to estimating the joint effects of a mixture: quantile g-computation. This approach combines the inferential simplicity of WQS regression with the flexibility of g-computation, a method of causal effect estimation. We use simulations to examine whether quantile g-computation and WQS regression can accurately and precisely estimate effects of mixtures in common scenarios. We examine the bias, confidence interval coverage, and bias-variance tradeoff of quantile g-computation and WQS regression, and how these quantities are impacted by the presence of non-causal exposures, exposure correlation, unmeasured confounding, and non-linear effects. Quantile g-computation, unlike WQS regression allows inference on mixture effects that is unbiased with appropriate confidence interval coverage at sample sizes typically encountered in epidemiologic studies and when the assumptions of WQS regression are not met. Further, WQS regression can magnify bias from unmeasured confounding that might occur if important components of the mixture are omitted. Unlike inferential approaches that examine effects of individual exposures, methods like quantile g-computation that can estimate the effect of a mixture are essential for understanding effects of potential public health actions that act on exposure sources. Our approach may serve to help bridge gaps between epidemiologic analysis and interventions such as regulations on industrial emissions or mining processes, dietary changes, or consumer behavioral changes that act on multiple exposures simultaneously.
WQSh
Quantile g-computation generally produced less biased and better-calibrated mixture-effect estimates than WQS regression when directional, linearity, additivity, and confounding assumptions were challenged. Its flexibility came with a possible bias-variance tradeoff and unresolved questions about mixture definition and quantized exposure modeling.
- Validity under non-null effects: Quantile g-computation provided unbiased overall-effect estimates with valid confidence intervals in scenarios where WQS regression was biased or poorly calibrated.For a single causal exposure, WQS had power above 90% but 95% confidence-interval coverage of only 57–83%; quantile g-computation had valid intervals but lower power.
- Validity under confounding: Under negative co-pollutant confounding, quantile g-computation was unbiased across examined total effects and causal-exposure correlations, whereas WQS was biased at all studied confounding levels.WQS bias increased with the strength of the negative confounder-outcome association and decreased with exposure correlation.
- Validity under confounding: As exposure correlation increased to 0.9, quantile g-computation’s individual-effect confidence intervals widened while the overall-effect interval narrowed.The reported confidence-interval width was 3.92 times the standard error.
- Validity under confounding: With increasing noise exposures, unmeasured-confounding bias increased for WQS regression but remained stable for quantile g-computation across studied sample sizes.The difference between methods diminished as sample size increased but remained present at all examined sample sizes.
- Validity under non-linearity and non-additivity: Under non-linear and non-additive effects, quantile g-computation produced unbiased and more precise estimates, whereas WQS produced biased quadratic-effect estimates.This remained true even when WQS allowed quadratic effects of the exposure index.
- Method comparison: Quantile g-computation maintains WQS’s simple inferential framework while accommodating beneficial, harmful, or harmless exposure effects.It combines quantized exposures and a joint effect with causal-inference approaches that allow non-linearity and non-additivity.
- Method comparison: WQS regression was less biased than quantile g-computation in no examined simulation scenario, although some scenarios showed lower WQS variance.The authors identify this as a potential bias-variance tradeoff when choosing between methods.
- Limitations: The simulations did not address WQS sample splitting, and the effects of treating quantized exposures as continuous regressors remain broadly unknown.The authors identify skewed exposures and common non-linear effects as areas requiring further research.
APPENDIX
The appendix presents simulation figures comparing quantile g-computation and WQS regression across sample sizes, exposure counts, correlations, and total effect sizes. It also illustrates qgcomp and gWQS exposure-weight outputs under linear, same-direction effects.
- Package output: Figure A1 shows qgcomp exposure-weight estimates when weights are estimable in a linear/additive model.With same-direction exposure effects, weights represent each exposure’s proportion of the total mixture effect.
- Package output: Figure A2 shows gWQS exposure-weight estimates under a linear model with same-direction exposure effects.The weights represent each exposure’s proportion of the total mixture effect.
- Simulation figures: Figures A3–A7 examine co-pollutant-confounding bias for quantile g-computation and WQS regression across sample sizes, exposure counts, exposure correlations, and total effect sizes.The simulations use 1,000 iterations where specified, with boxplots showing medians, interquartile ranges, and outliers.
APPENDIX TABLES
The appendix tables assess validity of WQS regression and quantile g-computation under null, non-null, non-additive, and quadratic-effect scenarios, including a no-sample-splitting analysis.
- Validity tables: Table A1 evaluates null and non-null estimates when directional homogeneity holds using 1,000 simulated samples of N=100.The table reports validity measures for WQS regression and quantile g-computation, including bias, variance, confidence-interval coverage, and power or type 1 error where defined.
- Validity tables: Table A2 evaluates WQS regression and quantile g-computation when individual exposure effects are non-additive and the overall effect includes linear and squared exposure terms.The scenario uses 1,000 simulated samples of N=100.
- Validity tables: Table A3 evaluates WQS regression without sample splitting and quantile g-computation under the null using 1,000 simulated samples of N=500.The table reports the type 1 error rate and variance-estimator validity using the specified package defaults.