Exposure mixtures and the near future of improving our understanding of worker health

Exposure mixtures are sets of exposures that co-occur in place and time. Traditionally, epidemiologists have focused work within mixtures on isolating effects of single exposures by statistical adjustment for other elements of the mixture as confounders. Increasingly, however, mixtures methodology focuses on pushing epidemiology beyond experimental paradigms that control co-exposures into a research model that embraces their variability. For example, the National Institutes of Environmental Health Sciences (part of the US National Institutes of Health) recently funded a set of research teams to develop statistical methodologies for mixtures aimed to improve “the scientific basis for regulatory frameworks that fully account for real world exposures.”¹ This paradigm shift is needed because effects of harmful exposure may be exacerbated by the presence of other exposures within a mixture, and efforts to reduce levels of a single harmful exposure can often result in reducing levels of many other exposures when that exposure occurs within a mixture. This critical change in focus in service of public health and scientific knowledge is evidenced by recent working groups framed around mixtures within the US Environmental Protection Agency (https://www.epa.gov/healthresearch/cumulative-impacts-research) and the US National Academies of Science (https://nap.nationalacademies.org/read/29058/chapter/1).

Methodology for mixtures has seen a weaker uptake in occupational studies, however. In the current issue of Occupational and Environmental Medicine, two studies address problems of occupational exposures that occur within a mixture. In one of these, Van Buren et al. address associations between adverse pregnancy and birth outcomes and occupational exposure to noise during pre-conception and pregnancy among a sample of live births without congenital anomalies in the United States.² In the other, Stjernbrandt et al. estimate associations between incidence of surgery for carpel tunnel syndrome (CTS) and multiple, occupational biomechanical risk factors among male construction workers in Sweden.³ In this commentary, I use these studies as touch-points to describe how recent developments in methods for environmental exposure mixtures might be better utilized in occupational settings.

One common parameter of many of the recently developed methods for exposure mixtures is an “overall effect” of a mixture. For example, single index methods such as weighted quantile sum regression (WQSR) quantify a mixture in term of a univariate “index” exposure and estimate associations between the index exposure and health outcomes, yielding an exposure response for the mixture itself.⁴ WQSR yields estimates similar to standard regression models, which can provide appealingly simple interpretation, but also requires strong modeling and causal assumptions. Subsequently, Bayesian Kernel Machine Regression (BKMR) was developed specifically to relax the strong assumptions inherent in other methods, at the cost of some interpretability.⁵ BKMR yields a posterior exposure-response surface distribution, which can be used to summarize mixture-health associations graphically in terms of a smooth regression line, similar to how spline regression has been used to assess non-linear associations between silica exposure and lung cancer among exposed workers. Generally, BKMR requires continuously measured exposures. More recently, quantile g-computation (QGC) was developed to estimate low-dimensional exposure-response parameters similar to WQSR but under relaxed assumptions and reduced computational costs.⁶ This approach, like WQSR, estimates the association between a health outcome and a simultaneous one-quantile increase in all exposures within the mixture (or other schema such as per doubling of all exposures).

Generally speaking, each of these methods was developed for single-time exposure data but each can accommodate longitudinally measured exposures and time-to-event outcomes, which are common in occupational cohort studies. QGC has been extended to estimate hazard ratios from Cox regression,⁷ while WQSR can implement a pooled-logistic approach to survival analysis, and BKMR permits binary health outcomes via a probit regression framework. Notably, however, none of these methods can address healthy worker survivor bias (HWSB), a common concern with longitudinally measured occupational exposures which results in bias that cannot be addressed by any standard regression approach.⁸ Few studies have addressed a mixture while at the same time addressing HWSB. A notable example is that of Neophytou et al. who used the parametric g-formula to assess the combined effect of reducing respirable elemental carbon and respirable dust on mortality from chronic obstructive pulmonary disease.⁹

For the study of Van Buren et al., with a relatively short window of exposure, methods like WQSR and QGC would have been wholly appropriate to use to study associations between adverse birth outcomes and noise while also accounting for the fact that one of their highest exposed occupational groups (Transportation and Material Moving) would likely also be simultaneously exposed to other hazards like diesel exhaust and vibration. In contrast, Stjernbrandt et al. could have utilized methods like the parametric g-formula to assess the simultaneous effect on CTS of reducing multiple biomechanical factors at once. Such methods could be used to assess the impact of reducing (for example) intensity of upper extremity load and frequency of repetitive wrist flexion and extension to the lowest levels observed. Newer methods for HWSB like inverse probability weighting could be used to estimate a joint exposure-response curve for multiple factors.¹⁰ In addition to these methods better addressing confounding by concomitant occupational exposures, they would estimate quantities that more closely resemble impacts of reducing exposures in the workplace that would result from industrial hygienic controls.

Workers are often simultaneously exposed to multiple hazards at work. Similarly, industrial hygienic measures such as requiring respirators, improving ventilation, and badging requirements to limit time within highly exposed work areas can impact a worker’s exposure to multiple hazards at once. To reduce the potential for bias from multiple, potentially confounding exposures, occupational study design has often focused on identifying workplaces free of other exposures to mimic experimental design. In contrast, environmental epidemiologists have recently begun to embrace the complexity of real-world exposures scenarios, and environmental epidemiology has consequently seen a recent surge of new methodologies for assessing the simultaneous effects of multiple exposures at once. To improve the scientific basis underlying worker protections, and for the sake of better understanding the effect of work on health, occupational epidemiologists would do well to adopt and adapt some of these methodologies.