Applying the Bradford Hill criteria in the 21st century: how data integration has changed causal inference in molecular epidemiology

Defining what constitutes a “strong” association is critical to the assessment of potentially causal relationships. Advances in statistical theory and the computational processing power have allowed scientists to delineate strong versus weak associations using more defensible mathematical criteria than Hill had in mind. Strength is no longer interpreted as simply the magnitude of an association. Furthermore, researchers have gained a greater appreciation for multi-factoral diseases and the existence of determinant risk factors that are small in magnitude yet statistically strong. Today, statistical significance—not the magnitude of association—is the accepted benchmark for judging the strength of an observed association, and thus its potential causality.

An example can be seen in the analysis and subsequent re-analysis of pulmonary function in a cohort of 106 workers at a flavorings manufacturing facility that used a variety of chemicals, including acetaldehyde, acetoin, benzaldehyde, butyric acid, and diacetyl [12, 13]. In the original study conducted by the National Institute for Occupational Safety and Health (NIOSH), researchers retrospectively analyzed spirometry reports and job title records collected by the cohort’s employer [13]. The authors presented statistically significant effect estimates showing that employees in jobs with higher potential for flavoring chemical exposures had 2.8 times greater annual declines in forced expiratory volume (FEV) than employees in lower-exposure jobs. This led authors to conclude that there was a statistically strong association between occupational exposure to flavorings and restrictive pulmonary disease [13]. However, as Ronk et al. [12] pointed out, the NIOSH researchers did not account for the inherently correlated nature of the longitudinal spirometry test data in their choice of regression analysis, which would affect the data variability and therefore standard error estimates and subsequent statistical inference [14, 15]. Ronk and colleagues re-analyzed the same data set using generalized estimating equations (GEE) that account for these correlations and did not find any statistically significant associations [12]. The varied outcomes and author interpretations associated with these two studies underscores how the use of different statistical methods can lead to statistically different results, thus impacting the application of strength of association.

Traditionally, Hill’s consistency criterion is upheld when multiple epidemiologic studies using a variety of locations, populations, and methods show a consistent association between two variables with respect to the null hypothesis. Hill stressed the importance of repetitive findings because a single study, no matter how statistically sound, cannot be relied upon to prove causation due to ever-present threats to internal validity. This criterion is still very appropriate for determining causal relationships; however, data integration practices have led to an evolution in thought on what constitutes consistency. The concept of data integration is inherently influential in the interpretation of the consistency criterion as it speaks to understanding a consistent story across multiple disciplines or practices. For example, through the lens of data integration, molecular experimentation can bolster epidemiologic findings by providing supportive evidence for a mechanistic hypothesis, thereby lessening the need for repetition among numerous observational studies. In vitro toxicology studies that suggest a mode of action such as genotoxicity or altered gene expression can support an association found in an epidemiologic study. By integrating results from multiple types of studies, researchers can show consistency in the causal story by illuminating various mechanistic points along the exposure-to-effects paradigm. This is a much broader interpretation of consistency than Hill’s original concept of repetitive epidemiologic findings.

The story of benzene-associated Acute Myeloid Leukemia (AML) illustrates the application of the consistency criterion in light of modern data integration. Both animal models and in vitro human cell cultures demonstrated that hydroquinone and para-benzoquinone are the active metabolites of benzene [15, 16]. Additionally, it was shown that hydroquinone induces cell changes that are consistent with various cellular changes known to mark the early progression of AML in humans [16, 17]. These molecular-level studies supported available human in vivo data (i.e., standard epidemiological studies), thereby lessening the need for additional observational studies to support a causal relationship.

Similarly, data integration played a role in the demonstration of consistency to support a causal relationship between polychlorinated biphenyl (PCB) exposure and melanoma. Consistency among epidemiologic studies of PCB exposure and melanoma, and in vitro mechanistic studies with human melanocytes support a plausible mechanism by which PCBs disrupt melanogenesis [18, 19]. Collectively, these data contributed to the decision by the International Agency for Research on Cancer Monograph Working Group to upgrade PCBs to a Group 1 carcinogen [18, 20]. Consistency between rodent and human bioassays also demonstrates support for a mechanism of carcinogenicity via initial binding to the aryl-hydrocarbon receptor (AhR) by PCB 126 and 2,3,7,8-tetrachlorodibenzo-para-dioxin, (TCDD) in other cancers [18, 20]. These examples illustrate how advanced molecular analyses can be integrated with the results of observational studies to demonstrate consistent research findings supporting a potentially causal relationship.

Hill suggested that associations are more likely to be causal when they are specific, meaning the exposure causes only one disease. While Hill understood that some diseases had multiple causes or risk factors, he suggested that “if we knew all the answers we might get back to a single factor” responsible for causation. This view is indicative of the fact that, in Hill’s era, exposure was often defined in terms of proxies for true exposures, such as an occupational setting or a residential location. Today, we attempt to specifically define exposures not in terms of a person’s surroundings or conditions, but rather as an actual dose of a chemical, physical, or biological agent. While some examples of highly specific agent-outcome associations exist, most exposure and health concerns at the forefront of research today center around complex chemical mixtures and low-dose environmental and occupational exposures complicated by a variety of risk factors.

The original criterion of specificity is widely considered weak or irrelevant from an epidemiologic standpoint. However, specificity may have new and interesting implications in the broader context of data integration. For example, researchers can demonstrate a molecular mechanism of action with precisely defined (i.e., specific) relationships between the agent and the effects using a variety of research methodologies. Asbestos exposure and the development of asbestosis is one example. In addition to the common use of occupational history as a surrogate for asbestos exposure in an epidemiological framework, advances such as refined standardized criteria for clinical diagnosis of asbestosis, microscopic lung fiber burden analyses and identification of asbestos bodies, as well as increased understanding of the relative potency of different fiber types have further clarified how asbestosis may be specifically caused by asbestos exposures [21‐24]. With data integration, specificity evolves into a more powerful criterion, and the lack of specificity can help to narrow down specific agents associated with disease. For example, complex mixtures of chemicals (e.g., tobacco smoke) typically lack specificity when studied using classic epidemiology designs, since multiple diseases can result from the exposures. However, it is possible that data integration may elucidate some mechanistic specificity among the varied disease endpoints associated with these complex carcinogenic mixtures.

Temporality is perhaps the only criterion which epidemiologists universally agree is essential to causal inference. Consider that Rothman and Greenland, despite finding a lack of utility or practicality in any of the other criteria, referred to temporality as “inarguable” [25]. Hill explained that for an exposure-disease relationship to be causal, exposure must precede the onset of disease. Thus, epidemiologic study designs which ensure a temporal progression between the two measures are more persuasive in causal inference.

A modern example of expanded temporal analysis using data integration is illustrated by studies of low-dose exposures to arsenic through drinking water and food. Arsenic levels in hair and nails serves as a biomarker of past exposure [26, 27], and drinking water analytical records from an individual’s past and present residences can be used to create an estimate of historic environmental exposure [28]. Limited windows of exposure can be evaluated to determine effects of exposure during sensitive stages [29, 30]. By integrating new data and knowledge from these tools, temporal relationships can be considered even within cross-sectional or ecological studies that do not implicitly establish temporality within the study design.

Today, our understanding of temporality now includes a wider range of precisely defined wider exposure windows, some of which are more relevant to disease outcomes than previously thought. Through epigenetic mechanisms (i.e., DNA methylation, histone modifications), exposures that occur during specific periods of development or even in previous generations can result in phenotypic differences in offspring [31]. Such changes could be responsible for generational effects of synthetic estrogen diethylstilbestrol (DES) exposure which can lead to increased risk of breast cancer multiple generations removed from the initial exposure [32]. Analytical techniques are improving to detect these changes and to determine which epigenetic alterations may serve as indicators of disease potential and persistent biomarkers of a previous exposure [33]. Understanding the molecular-level changes that precede an observable outcome can help establish the temporal progression in a multigenerational causal story [34].

Hill wrote that “if a dose response is seen, it is more likely that the association is causal.” According to the traditional interpretation of biological gradient, the presence of a dose–response relationship supports the causal association between an exposure and an effect [25, 35]. In traditional epidemiology, a monotonic biological gradient, wherein increased exposure resulted in increased incidence of disease, provides the clearest evidence of a causal relationship. However, Hill acknowledged that more complex dose–response relationships may exist, and modern studies have confirmed that a monotonic dose–response curve is an overly simplistic representation of most causal relationships. In fact, most dose–response curves are non-linear and can even vary in shape from one study to the next depending on unique characteristics of the given population, exposure routes, and molecular endpoints assessed [36]. Furthermore, individual susceptibility and synergistic or antagonistic effects of cumulative exposures can make some biological gradients even more difficult to characterize. An example of this effect can be seen in aryl hydrocarbon receptor (AhR)-based mechanisms: many exogenous and endogenous agents can act as partial agonists/antagonists of AhR, and thus modulate the dose–response effect of 2,3,7,8-tetrachlorodibenzodioxin (TCDD) which affects gene expression via AhR [9]. Integration of advanced statistical capabilities, data modeling techniques, and knowledge from increased understanding of biomolecular interactions have resulted in the descriptions of more defined dose–response curves, capable of showing molecular effects at very low levels of exposure. Additionally, growing knowledge of genetic polymorphisms has illuminated the reasons behind individual variations in biological response to toxic insult and the dose–response relationships [8].

It is now possible to observe threshold responses in the low-dose range, rather than assuming linearity for all substances. Furthermore, experimental support for a dose–response phenomenon referred to as hormesis has increased with improved molecular techniques. Hormesis is characterized by low dose stimulation and a high dose inhibition [37]. The dose–response curve associated with this phenomenon is biphasic and, depending on the endpoint measured, is either J or U shaped [38]. Hormesis has been observed in both toxicology and pharmacology, and the features of the observed dose–response are consistent and independent of the biological model, endpoint measured, chemical or physical stressor, and mechanism [37]. The most distinctive feature of hormesis is that it is repeatedly observed below the typical threshold dose [37].

Biological gradient is an example of how data integration can complicate causal inference. New tools and technical capabilities have allowed researchers to characterize a variety of low-level molecular endpoints that may not lead to disease or observable adverse outcomes on a larger scale. For example, innate responses can repair, eliminate, or reverse molecular changes caused by low levels of exposure. Thus, molecular changes within the no-observable-adverse-effect level (NOAEL) may not contribute to disease and are more indicative of a threshold dose response. Understanding the mechanisms at low level exposures allows us to elucidate a dose–response curve. For example, the in vitro endpoints for asbestos toxicity include generation of oxidative stress which results in genotoxicity and chromosome damage via DNA adduct formation [39]. However, damage at low levels, while measurable in vitro, is removed via cellular apoptosis which represents adaptive response and a threshold effect. Thus, responses at these low levels may not be indicative of disease, but rather adaptive responses that indicate a threshold must be overcome prior to disease initiation.

Additionally, modern analytics have shown that epigenetic endpoints can occur in the low-dose range of environmental chemical exposures, though these measured changes may not lead to observable disease. For example, Kim et al. [40] observed non-monotonic dose-dependent alterations in DNA methylation among mouse liver samples from offspring exposed perinatally to multiple doses of BPA through the maternal diet. These changes may provide insight regarding a mechanism of action for BPA during developmental exposure; however, further information regarding phenotypic changes is necessary to determine whether epigenetic changes at low level exposures are significant indicators of a dose-disease response relationship. Thus, biological gradient can be broadened to include molecular dose–response relationships, if the actual response occurs at a dose that is also associated with disease onset or progression.

Even at the time it was introduced, biological plausibility represented fundamental concepts of data integration—the criterion implies that epidemiology and biology must interact [5]. Plausibility has historically been judged based on the presence of existing biological or social models that explain the association of interest. Hill’s criterion of plausibility is satisfied if the relationship is consistent with the current body of knowledge regarding the etiology and mechanism of disease; though, Hill admitted that this interpretation of biological plausibility was dependent on the current state of knowledge. Today, tools such as high-throughput screening assays can be used to study a specific biologically plausible pathway and identify toxic agents that interfere with that pathway in defined ways. Indeed, opening the ‘black box’ through integrating molecular epidemiological advancements has allowed researchers to illuminate more steps in the exposure-to-effect paradigm, contributing to an understanding of biological plausibility for suggested causal relationships.

The elucidation of biological pathways leading to liver toxicity have played a large role in advancing the interpretation of biological plausibility, and the integration of knowledge from various evidence streams has aided in those interpretations. The liver is typically the first organ with appreciable capacity for oxidative metabolism that an agent encounters after ingestion, and is therefore a key organ for studying potential toxicity [16]. Liver effects demonstrated using techniques such as high-throughput in vitro and in silico cell manipulation, can be seen as a harbinger for further toxic endpoints that might occur with more refined, realistic exposures [41, 42]. However, as demonstrated by the newly-developed “virtual liver” [43], the future of testing biological plausibility likely lies with in silico experimentation. Researchers can now predict plausible relationships using in vitro and in silico screening tools targeting defined disease mechanisms, which represents a potential paradigm shift in how scientists frame causal research questions and design studies.

Historically, causal inference was approached with the assumption of a single-factor direct relationship (i.e. A causes B). However, researchers now understand that many disease outcomes are a result of the interplay and balance between multiple contributing and intermediary factors. As such, demonstrating the biological plausibility of a causal relationship can be complex. However, improved statistical techniques can help researchers to understand complex disease progression from a molecular standpoint, where multiple risk factors, confounders, adaptive responses, and mediating mechanisms intersect [44‐46]. For example, the biostatistical approach of mediation analysis allows for the disentanglement and decomposition of the various biological pathways of direct and indirect effects that play a role in filling the “black box” between exposures and observable outcomes [47].

Coherence has been viewed as being similar to biological plausibility, in that the cause-and-effect story should make sense with all knowledge available to the researcher, and this criterion has not changed greatly since its inception. Indeed, Hill identified histopathological evidence of bronchial epithelium changes and animal-based toxicity tests for the carcinogenicity of cigarette smoke as an example of a coherent story among several avenues of study design. Today, coherence is another area in which molecular-based studies have been used to demonstrate a comprehensible story regarding various aspects of the exposure-to-disease paradigm. For example, lung tissue fiber analysis by scanning transmission electron microscopy (STEM) has expanded our knowledge of internal biologically effective amphibole dose relating to altered structure and function of lung tissue, supporting the conclusion that amphibole asbestos fibers induce mesothelioma [48].

Alternatively, advanced mechanistic studies can elucidate an incoherent body of epidemiologic literature, thereby strengthening the causal inference in one direction or another. Consider for example the carcinogenicity of hexavalent chromium [Cr(VI)]. The body of epidemiologic literature regarding the carcinogenicity of Cr(VI) is limited and conflicting, particularly regarding ingestion exposures (e.g., drinking water) and cancers outside the respiratory system (e.g., cancers of the GI tract). However, a recent array of genomic, pharmacokinetic, and mechanistic research—including metabolism, bioavailability and kinetic studies, mutagenic mode of action studies, and gene expression profiling—demonstrate that ingested Cr(VI) does indeed have a carcinogenic profile [49, 50].

Hill explained that evidence drawn from experimental manipulation—particularly epidemiologic studies in disease risk declines following an intervention or cessation of exposure—may lead to the strongest support for causal inference. Yet in modern contexts, experimentation must consider that many diseases result from multifaceted exposures and follow complex progression pathways. Cessation of exposure as Hill described may not reverse or appreciably slow the progression of disease. In some cases, multiple risk factors, including diet, exercise, smoking, chemical exposures, and genetic predisposition can contribute to disease onset and progression. Thus, while the combination of these factors may culminate in disease, experimental manipulation of a single contributory factor may or may not result in observable decreases in disease incidence.

Researchers using a data integration framework can now draw from toxicological findings for experimental insight into causality. In vitro studies that test mechanistic pathways and demonstrate the biological role of an agent in disease progression may result in knowledge that can be used to predict potential human health outcomes in a much more time-efficient manner than human studies, particularly for adverse outcomes with a long latency period.

The expanded understanding of temporality in light of data from varied evidence streams can also affect interpretation of the experiment criterion. Individual exposures can cause epigenetic modifications to parental DNA that result in an observed effect in future offspring, even though there is no direct exposure to the offspring. Experimental studies in animal models are often necessary to provide mechanistic support for an epidemiologic observation that involves complex temporality. For example, multiple animal studies provide support for the hypothesis that epigenetic changes induced by DES exposure in utero may be causative of transgenerational effects of DES exposure in females [32, 51‐54]. Because epigenetic analyses in transgenerational human studies take decades and are riddled with potential confounders, reliance on animal models and advanced analytical techniques can help to support determination of a causal relationship.

Hill implied that when there is strong evidence of a causal relationship between a particular agent and a specific disease, researchers should be more accepting of weaker evidence that a similar agent may cause a similar disease. Analogy has been interpreted to mean that when one causal agent is known, the standards of evidence are lowered for a second causal agent that is similar in some way [55]. Some modern epidemiologists have argued that a lack of analogy does not preclude causation, but simply implies a lack of creativity on the researcher’s part [56]. Indeed, some might argue that enough knowledge exists and is accessible today to identify an analogy for every situation, especially if the researcher pulls that knowledge from multiple disciplines and across evidence streams. Today, researchers have a wider range of tools by which to seek an analogy, including disease progression pattern, common risk factors and confounders, and biological mechanisms of action. Therefore, the modern value of analogy is not gained from confirming a causal inference, but rather from proposing and testing mechanistic hypotheses.

As an example, analogous mechanistic hypothesis testing has been conducted on carbon nanotubes (CNTs) using the extensive literature on the mechanistic toxicity of asbestos fibers. Models based on molecular structure and physical–chemical characteristics such as aspect ratio predict a mechanism of action similar to that of asbestos [57]. The physical morphology of CNTs appears similar to that of asbestos fibers; thus, respirable-sized fibers are expected to behave similarly in occupational settings and lead to similar lung translocation and deposition. Additionally, asbestos fibers are known to cause inflammation and fibrosis of the lung pleura as a precursor to mesothelioma; these same outcomes have been demonstrated following CNT exposure [58, 59]. Further, CNTs have been found to stimulate the release of acute phase cytokines from human macrophages and mesothelial cells exposed to CNTs of varying lengths, demonstrating that CNT exposure results in a length-dependent pro-inflammatory response, similar to that of asbestos [60]. These findings enhance the asbestos analogy by confirming that CNTs may be capable of causing disease that begins with pleural inflammation—the same mechanism responsible for asbestos-related mesothelioma. However, the results also demonstrate that not all CNTs have the same potential for carcinogenicity, implying that proactive design of engineered CNTs can limit the risks and allow for safe use of the compounds in a variety of applications—and that the analogy to asbestos should not be viewed in a way that limits continued research.