[Paper Review] Upstream Causes of Downstream Effects

Bradley Saul, Michael G. Hudgens+1 more|arXiv (Cornell University)|May 22, 2017

Advanced Causal Inference Techniques38 references3 citations

TL;DR

This paper extends causal g-methods to estimate the effects of spatiotemporally varying nutrient pollution on downstream chlorophyll a levels in the Cape Fear River, using observational water quality data. It demonstrates that nitrate concentrations 86–109 km upstream significantly influence chlorophyll a at Lock and Dam 1, offering a framework for ecological causal inference in non-experimental stream monitoring data.

ABSTRACT

The United States Environmental Protection Agency considers nutrient pollution in stream ecosystems one of the U.S. most pressing environmental challenges. But limited independent replicates, lack of experimental randomization, and space- and time-varying confounding handicap causal inference on effects of nutrient pollution. In this paper the causal g-methods developed by Robins and colleagues are extended to allow for exposures to vary in time and space in order to assess the effects of nutrient pollution on chlorophyll a, a proxy for algal production. Publicly available data from the North Carolina Cape Fear River and a simulation study are used to show how causal effects of upstream nutrient concentrations on downstream chlorophyll a levels may be estimated from typical water quality monitoring data. Estimates obtained from the parametric g-formula, a marginal structural model, and a structural nested model indicate that chlorophyll a concentrations at Lock and Dam 1 were influenced by nitrate concentrations measured 86 to 109 km upstream, an area where four major industrial and municipal point sources discharge wastewater.

Motivation & Objective

• To address the challenge of inferring causal effects of nutrient pollution on algal blooms in river systems where experimental manipulation is infeasible.
• To extend causal g-methods to handle time- and space-varying exposures and outcomes, particularly in the presence of spatial interference.
• To apply marginal structural models, parametric g-formula, and structural nested models to observational water quality data from the Cape Fear River.
• To evaluate the statistical performance and robustness of these methods under limited independent replicates and confounding.
• To provide a replicable, open-source framework for ecological causal inference applicable to environmental policy and watershed management.

Proposed method

• Adapts the potential outcomes framework to model causal effects of upstream nutrient concentrations on downstream chlorophyll a levels.
• Applies parametric g-formula, marginal structural models (MSMs), and structural nested models (SNMs) to estimate average causal effects in a spatiotemporal setting.
• Incorporates space- and time-varying confounding through modeling of time-dependent covariates and exposure histories.
• Uses estimating equation theory with small-sample corrections (e.g., Fay and Graubard, 2001) to improve inference in data with limited independent replicates.
• Derives closed-form estimators for structural nested models and implements them via custom R packages (updown, geex, capefear).
• Validates methods through a simulation study under known data-generating mechanisms, comparing bias, variance, and coverage across methods.

Experimental results

Research questions

• RQ1 What is the causal effect of upstream nitrate concentrations on downstream chlorophyll a levels in the Cape Fear River?
• RQ2 How do different g-methods (g-formula, MSM, SNM) compare in estimating causal effects when exposure and confounding vary over time and space?
• RQ3 Which upstream reach of the river contributes most significantly to chlorophyll a levels at Lock and Dam 1?
• RQ4 How can causal inference be reliably performed in observational stream monitoring data with limited replication and confounding?
• RQ5 Can structural nested models be practically implemented and validated in ecological monitoring contexts?

Key findings

• Chlorophyll a concentrations at Lock and Dam 1 were significantly influenced by nitrate levels measured 86–109 km upstream.
• The parametric g-formula, marginal structural model, and structural nested model all yielded consistent estimates of causal effects, supporting robustness of findings.
• Four major point sources—Tarheel Plant, Dupont Fayetteville Works, Cedar Creek Site, and Rockville Creek Wastewater Treatment Plant—were located in the 86–109 km upstream zone, suggesting a likely origin of the causal exposure.
• The simulation study confirmed that the g-methods produced low-bias estimates with appropriate coverage under various data-generating mechanisms.
• The study demonstrates that causal inference is feasible in observational stream data when proper modeling of time- and space-varying confounding is applied.
• The R package updown and associated tools enable replication and extension of the analysis, supporting broader application in environmental causal inference.