2
$\begingroup$

How can we best explain causality in layman's terms?

There seem to be two main types of causality. One is probabilistic causation, the other is called determinism in philosophic circles or just causality (physics), or if you wish, deterministic causality. The latter appears to have a mechanistic chain of cause and effect. In probabilistic causation, it seems that a mechanism for causation may be unknown. For example, there is a qualitative difference between "a drought can cause a famine," and "smoking can cause cancer." In the former case, we can easily picture that rainfall can cause crops to grow, in the latter, how smoke grows tumors is not clear, and only the former is clearly deterministic. Is the above understanding correct, and in any case, can we define causation more exactly?

$\endgroup$
5
$\begingroup$

Explaining the concept of causality from a philosophical perspective is a big issue, and it's probably best left to philosophers, though it is useful for a statistician to have a working knowledge of this philosophical topic. Personally, I have found the explanation by Peikoff to be the best explanation I have heard (i.e., that causality is the "law of identity" applied to action). Under this view, asking about causality is really just asking about the actions that things take when subjected to various kinds of stimulus/intervention by other things. Asking why something happens is really just asking for a deeper description of what happens when one thing interacts with another thing.

I have always found that the best way to look at probabilistic causality is within the "subjective" Bayesian framework, which renders it compatible with deterministic metaphysical accounts of causation. Under the Bayesian view, probability is considered to be merely an epistemological tool that allows the user to measure his or her uncertainty about unknown quantities, constrained by rationality criteria. Under this viewpoint, probability is interpreted epistemologically, and so probabilistic causality is likewise interpreted epistemologically, not metaphysically. This view recognises that we are often exposed to data that appears to us to be "stochastic" (i.e., well described by probabilistic methods) even if it is formed by an underlying process that is deterministic. One of the advantages of this framework is that it is agnostic on determinism --- i.e., it is compatible with both deterministic and non-deterministic metaphysical accounts of causality.

The point here is that these different "types" of causality need not be seen as competitors that contradict each other. In the "subjective" Bayesian framework, when we talk about "probabilistic causality", we are really just talking about how we can infer causality using the tools of probability theory. "Causality" may still be interpreted as an underlying metaphysical condition that we are seeking to infer with epistemological tools. Moreover, it can be interpreted deterministically. In this framework the probability conditions are not a form of causality, but rather, conditions that allow inference of an underlying metaphysical causality. For this reason I have never really liked the name "probabilistic causation"; it should be called "probabilistic methods of inferring causation" instead.


Example (smoking): Applying this framework to smoking allows you to accept a simple cellular medical explanation of what cancer is, while also using probabilistic tools to try to infer causation. At a deep physical level, cancer is "caused" by smoking in the sense that various cells in the body (e.g., in the lungs) begin to engage in abnormal (harmful) growth when exposed to smoke particles in a particular way. This process might be entirely deterministic, despite the fact that it occurs at a level of detail that precludes our ability to accurately predict when a particular cell will become cancerous under exposure to smoke. Indeed, it is reasonable to conjecture that smoking leads cells to be exposed to smoke particles of different chemical compositions from different angles, frequencies, etc., and that even if the process of cancer growth in cells is deterministic, the exposure conditions vary so wildly that they appear to us to be stochastic.

We can accept a deterministic account of causation here and still use statistical trials and probability theory to infer causality. Ideally this would involve controlled-trials, or if it involves uncontrolled observation, it comes with appropriate caveats on inference. We try to control for confounders and avoid controlling for colliders so that we get an inference about the underlying causality, rather than just a predictive inference.


Explaining this in lay terms: Boy, that's tough! The first step is to explain it to ourselves in technical terms (see above), so that we get our concepts properly integrated and consistent before subjecting those poor laypersons to our views! If I were trying to explain the above in lay terms, I would say something like this:

When we ask about cause, we are really just asking for a more detailed explanation of how a thing acts when it is exposed to a certain type of condition. For example, how do the lungs of a person react to smoking? Do they become cancerous? Statisticians tell us that the best way to figure that out is by running an experiment where you randomly expose lungs to smoke (e.g., randomly make some people smoke and others not smoke) and then see who gets cancer. We then use statistical methods on this data to see if the people who smoked were more likely to get cancer.

Sometimes we can't do a randomised controlled trial like this. In this case, it might be unethical/impractical to instruct people to smoke, who don't want to smoke. (The trial only works if the randomly assigned smokers agree to smoke!) In these cases we still watch who gets cancer and who smokes, but it is harder to figure out if smoking is causing cancer. There is quite a bit of complicated theory for how to do this; it is quite hard and it is easy to make mistakes.

Statisticians have a lot of well-developed theory about how to tell when one thing causes another. They need this because it is hard to tell when things cause each other. Anyway, this theory of using probability to figure out cause is sometimes referred to as "probabilistic causality"; it tells us the conditions we need to be able to infer that some stimulus causes a particular response.

$\endgroup$
  • $\begingroup$ The first part of the answer is informative. The second part does not seem to be as interesting. Controlled trials can be of value when a deterministic model is being tested. When there is no deterministic model, a controlled trial is 'trail and error,' i.e., it offers very limited information. Lung cancer occurs in non-smokers. There are 15 different cancers that smoking can affect. The mechanism for this is not known, and there are competing theories. $\endgroup$ – Carl Jan 15 at 21:17
  • $\begingroup$ Here are some examples of potential chemical carcinogenic etiologies: (1) cancericidal immunosuppression. (2) Suppression of contact inhibition of cell proliferation. (3a) Direct and (3b) indirect chemical caused DNA damage. (4) Virchow: irritation causes cancer. (5) Other and combined effects. $\endgroup$ – Carl Jan 15 at 21:45
  • $\begingroup$ You claim that "the subjective Bayesian framework...renders it compatible with deterministic metaphysical accounts of causation." And yet from this: "...we suggest that the statistical community should accept formal objective Bayesian techniques with confidence, but should be more cautious about casual objective Bayesian techniques." I have difficulty reconciling these opinions, and this is rather more complicated than a persuasive explanation of causality in layman's terms. $\endgroup$ – Carl Jan 17 at 6:37
  • $\begingroup$ Are you reading "casual" as "causal" or is that a typo? $\endgroup$ – Ben Jan 17 at 7:09
  • $\begingroup$ My bad, indeed not causal, the author meant casual. My problem is that I do not understand your comment about the role of probabilistic causality within the "subjective" Bayesian framework. I really do not see how this is connected to the topic, so it is likely somewhat distant from a lay explanation. Can you explain a bit better? $\endgroup$ – Carl Jan 17 at 8:19
0
$\begingroup$

Deterministic causality is like having a smoking gun as evidence in a murder trial. It is the method of means, opportunity, and method.

Probabilistic causality is like having the gun's smoke, a few shell casings and gunpowder residue on a perpetrator's hands and clothing, but no gun.

Perhaps the best illustration of this difference for smoking was given as the probabilistic cause of cancer by Bradford Hill criteria; a set of nine criteria to provide epidemiologic evidence of a causal relationship between a presumed cause and an observed effect as follows:

Strength (effect size): A small association does not mean that there is not a causal effect, though
  the larger the association, the more likely that it is causal.
Consistency (reproducibility): Consistent findings observed by different persons in different places
  with different samples strengthens the likelihood of an effect.
Specificity: Causation is likely if there is a very specific population at a specific site and 
  disease with no other likely explanation. The more specific an association
  between a factor and an effect is, the bigger the probability of a causal 
  relationship.
Temporality: The effect has to occur after the cause (and if there is an expected delay between the
  cause and expected effect, then the effect must occur after that delay).
Biological gradient: Greater exposure should generally lead to greater incidence of the effect. 
  However, in some cases, the mere presence of the factor can trigger the effect. In other cases, an
  inverse proportion is observed: greater exposure leads to lower incidence.
Plausibility: A plausible mechanism between cause and effect is helpful (but Hill noted that
  knowledge of the mechanism is limited by current knowledge).
Coherence: Coherence between epidemiological and laboratory findings increases the likelihood of an
  effect. However, Hill noted that "... lack of such [laboratory] evidence cannot nullify the
  epidemiological effect on associations".
Experiment: "Occasionally it is possible to appeal to experimental evidence".
Analogy: The effect of similar factors may be considered. 

In general, one can state that probabilistic causation should ideally have evidence similar to Bradford Hill criteria for causation, that is, to be a probable cause, there should be be temporal, quantitative (e.g., dose dependent) and other causal properties that causes have. In other words the evidence for causation needs to pass the duck test of abductive reasoning: If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.