4.0 Cause and Effect Versus Association in Toxicology

The differences between cause and effect and association are of huge importance to toxicology. In the interest of public health (and often due to societal and political pressures), decisions and actions are frequently demanded in the face of uncertainty. Regulatory action against factors that appear associated (but are not as a result of cause and effect) will be ineffective, expensive and wasteful. Actions which have no effective public or environmental health outcomes are not neutral; they are harmful. Ineffective actions which are undertaken without the proper considerations will waste the limited funding and resources which are available, reducing the future resources to pursue actions with better outcomes.

The Bradford Hill Criteria

The Bradford Hill Criteria are nine principles which were developed in 1965 by Sir Austin Bradford Hill, an English epidemiologist and statistician. Also known as Hill’s Criteria for Causation, these nine principles are used to determine whether epidemiological evidence is indicative of a causal relationship (between a presumed cause and an observed effect). The Bradford Hill Criteria are widely applicable in public health research, although the exact application and limits of the criteria continue to be debated. However, the Bradford Hill Criteria remain critical for effective toxicological risk assessment.

For a cause and effect relationship to be established, the following Bradford Hill criteria must be addressed:

• 4.1 Strength of Association
• 4.2 Consistency
• 4.3 Specificity
• 4.4 Temporality
• 4.5 Biological Gradient
• 4.6 Biological Plausibility
• 4.7 Coherence
• 4.8 Experimental Manipulation
• 4.9 Analogy

There must be a strong statistical association between the independent variables (cause; such as the amount of a drug consumed per kilogram of a person’s weight) and dependent variables (effect; such as the quantity of a drug measured in the person’s bloodstream). Defining what constitutes a ‘strong’ association is critical. Advances in statistics, computational processing power and a greater understanding of the multifactorial nature of diseases have improved the mathematical techniques originally envisioned by Bradford Hill. Determinant risk factors that are small in magnitude but are statistically strong can be important. In modern times, the degree of statistical significance, not just the magnitude of association, is the accepted standard for judging the strength of an observed association.

The findings must be consistent and repeatable (i.e., replicable using different experimental methods, different populations, over time). It is important to stress the repeatability of scientific findings: the results of a single study, no matter how statistically and scientifically sound, cannot be absolutely relied upon to prove causation because of the potential for unknown and uncontrolled confounding factors. In more recent times, data integration has been used to somewhat re-define what Bradford Hill meant by consistency (although Bradford Hill incorporated some of these ideas under concepts of ‘biological plausibility’ and ‘coherence’) e.g., toxicological studies on a specific mode of action or mechanism of action can support an association found in an epidemiologic study. By combining results from different types of studies, consistency (and biological plausibility and coherence) can be demonstrated. This may reduce the need for multiple observational studies.

Associations are more likely to be causal when they are specific, meaning the exposure causes only one disease. Historically, specificity has been regarded as the weakest of Hill’s criteria. Bradford Hill recognized that many diseases were multifactorial but that “if we knew all the answers we might get back to a single factor [responsible for causation].” In his era ‘exposure’ was typically a proxy for true exposure/dose, i.e., being in an occupational or residential setting. In modern times we define exposure as an actual dose of a chemical, physical or biological agent. While there are examples of highly specific agent-outcome associations (i.e., you take this one chemical, you get this one harmful effect), most exposure and health interaction evaluations center around the effects of complex chemical mixtures combined with low-dose environmental exposures, occupational factors and other risk factors. Despite this complexity, modern molecular biology has demonstrated numerous cases of multiple diseases involving different tissues and systems being due to the same pattern of mutations, i.e., Bradford Hill’s concept of “getting back to a single factor” also appears to be correct in some unexpected cases.

Temporality is absolutely essential for causation: the exposure must occur before the harm is seen. Study designs that ensure a temporal progression between two measures are more convincing of a causal relationship.

xkcd. (2011). Cell phones. https://imgs.xkcd.com/comics/cell_phones.png

“If a dose-response is seen, it is more likely that the association is causal.” Typically dose-response relationships are monotonic, that is, they appear similar to the figure below:
(Adapted from Schiefer, H. B., Irvine, D. G. & Buzik, S. C. (1997). Understanding toxicology: Chemicals, their benefits and risks. CRC Press: Florida.)

However, more complex types of dose-response relationships occur and monotonic dose-response curves can be an overly simplistic representation of causal relationships. Dose-responses may also only appear monotonic due to experimental design. Importantly most dose-response curves, particularly in the low dose areas of the curve (i.e., the left side of the curve) that are of greatest interest to toxicologists, can often be non-linear and vary in shape from one study to the next (sometimes even within the same animal being tested, depending on what is actually being measured). This depends on the unique characteristics of a given population, exposure routes, and the molecular endpoints tested. In the past, for reasons of expediency, computing power and risk assessment conservatism, toxicologists have assumed that the low dose area of the dose-response curve is linear or near-linear. This may not be so for many agents. The phenomenon of ‘hormesis’ has been observed for several agents; notable examples include dietary vitamins, dietary essential minerals, and controversially radiation. Complex synergistic and/or antagonistic effects due to components of a mixture can make biological gradients difficult to detect experimentally.

A comparison of monotonic and non-monotonic dose-response curves are shown in the following figure:

Is cause and effect biologically possible? Historically, plausibility has been based on the presence of existing biological models or mechanisms that explain the association of interest, that is, it was limited by the state of knowledge at the time. However, with the availability of high throughput toxicology screening systems, the knowledge supporting ‘biological plausibility’ can be developed by rapid exploration of the mode(s) of action or mechanism(s) of action after the finding of an association. Biological plausibility uses the concept of data integration.

Coherence asks the question: “If you assume that the established theory is correct, would the observed results fit into that theory?” It acknowledges that paradigm shifts do occur in science but that they are uncommon. If an association is non-coherent, it could be something genuinely new to science; but it is much more likely that such a finding will weigh against a cause and effect relationship. Again, explorations of coherence require mechanistic toxicology studies and data integration.

Evidence drawn from experimental manipulation is often the strongest evidence of cause and effect, e.g., an intentional change that reduces exposure results in a reduction of the disease. However, interventions in multifactorial diseases or in multifaceted exposures may not reverse or slow disease progression over a practical period of experimental observation.

When there is strong evidence of a causal relationship between a particular agent and a specific disease, researchers will be more accepting of weaker evidence that a similar agent may cause a similar disease, i.e., when one causal agent is known, the standards of evidence are lowered for a second causal agent that is similar in some way. In toxicology, this approach (analogy) is used as part of ‘read across strategies:’ when there is toxicological information on one chemical structure, closely related chemical structures are assumed to have similar toxicological properties. At best, the read across approach is a last resort in place of actual data. Actual experimental data is always more reliable. In more recent times, analogy can be derived from known mode(s) of action/mechanism(s) of action followed up by testing.

Adapted from: Wikipedia. V-series nerve agents (VX and VE) – these are structurally similar and both act as nerve agents.