15 - Reasoning About Causation
In many situations, causes are correlated to their effects: a given outcome is more probable when a given factor is present and less probable when it is absence. Without correlation, we cannot assume causation.
However, correlation alone does not imply causation - it merely suggests that there is a connection, link, or association, but that it is not causative in nature.
It can neither be said that low correlation dismisses causation: a poor hunter may miss his prey most of the time, but on the few instances his bullet finds its target, it is definitely the cause of the animal's death.
Nor can it be said that a high correlation establishes causation. There may be no connection, the cause-effect assumption may be reversed, there may be a common factor that cases both, or the cause may contribute to the effect without being sufficient or even necessary.
This is not to say that correlation should entirely be dismissed - it is very often a flag that guides us toward causation, but not without further consideration.
Why Correlation Is Not Causation
The author considers a few relationships in which a correlation does not establish causation:
Accidental Correlation
We may observe that two things seemed to happen at the same time, and imagine that there is causality. For example, if a man who customarily wears a gray suit and has never been in an automobile accident gets into an accident on the day he wore a brown suit, he may remark that the color of his suit is a determining factor. While that is clearly absurd, the behavior of sports fans and stock market investors involve many such misconceptions.
The incidence of accidental correlation is significantly increased over time, as the vast amount of information stored in computer systems facilitates playing with numbers by those who are unaware of what the numbers themselves represent.
The Causal Direction is Reversed
In some instances, we can observe that two factors are strongly correlated, but have the cause and the effect mixed up, sometimes intentionally, sometimes because it is difficult to discern the order in which they occur.
Consider the correlation of drug addiction and psychiatric conditions. This is commonly construed to mean that drugs make people go crazy (so let's ban drugs) rather than considering whether a person with a mental condition might seek out drugs (so let's provide treatment).
In some instances, two factors may cause one another, forming a causal loop. Consider the correlation between health and income: a healthy person can work longer and better to earn a higher income, and having a higher income enables a person to afford better living conditions and medical care.
A causal loop that causes conditions to degenerate is a vicious circle, which is generally accepted to lead to a catastrophe. Consider the pattern of a recession in which decreased income makes people spend less, decreasing the demand for goods, decreasing the demand for labor, decreasing the income of people, which makes them spend less, etc. until there is an intervention and a crash ensues.
Common Causes
The previous example (shark attacks and ice cream) illustrates a situation in which two factors that are correlated derive from a common cause.
Causation Due to Side Effect
In some instances, there is a high correlation where the side effect of a factor, rather than the factor itself, has some connection (not necessarily a causal one) with an effect.
A placebo effect is a common use of this - where it is the positive attitude of the patient for having received a sugar pill, rather than the sugar itself, that is believed to alleviate the patient's discomfort. As such, it cannot be concluded that sugar alleviates pain, as it does not matter what the content of the placebo would be. Further, the patient's actual discomfort may not be lessened at all, but merely their motivation to complain to staff.
In biological research, it is often considered that studying animals in captivity does not give a true indication of the behavior of wild animals because putting animals in a confined environment and feeding them regularly is likely to change their behavior, making the results invalid in the wild.
In psychological research, the Hawthorne effect is very well known - people change heir behavior when they know that they are being observed.
There is also the novelty effect - which deals with a temporary change in behavior for encountering something new. Studies that suggest students perform better academically when they switch their drab uniforms to colorful ones presume there to be a psychological effect to color, but it may merely be because of the presence of something unusual. The results would need to be validated after the novelty has worn off.
Good Evidence For Causation
We are well advised to be cautious about accepting correlation as evidence of causation, but should not dismiss that it may be a good indicator to suggest where we might consider whether a causal relationship exists.
Look for Covariation and Manipulability
Covariation requires a closer mathematical analysis of cause and effect. Where the degree to which the effect occurs maps closely to the degree to which the cause is applied, there is greater probability of causation.
For example, if we were to observe the degree to which soft drinks cause diabetes, a higher incidence of diabetes among those who routinely consume four drinks a day to be higher (though not necessarily double) that of those who routinely consume only two.
This becomes even stronger evidence when it can be directly manipulated and not just passively observed, though ethical concerns arise when the outcome of requiring or encouraging a person to do something they would normally not do results in harm.
Chains of Causation
We often speak of causal relationships as if they were direct, but instead there are multiple steps between the two. For example, we consider that pressing a piano key makes a sound, but in reality, pressing they key moves a hammer, the hammer hits a string, the string vibrates, the vibration stirs air molecules, and the vibration in molecules strikes the eardrum, which is interpreted as a sound.
Illustrating the chain of causation grants the association greater credibility, but it also gives us the ability to test each link in the chain, or to test only those links in the chain that seem disputable, and dismiss a lot of silliness (pressing they key does not make a sound to a person wearing earplugs).
Proving causation between a cause and effect when the chain of events between them is unknown is possible, but weaker and less convincing.
Causation Is Complicated
A few tips for sorting things out:
- Necessary and sufficient: The earlier notion of necessary and sufficient conditions should be taken into account when considering causality.
- Multiple factors: we may discover multiple factors that caused an outcome. A student who failed a test may have been distracted by personal issues, suffering from illness, and failed to study sufficiently.
- Triggering events: There may be conditions that make an effect more likely to occur, but do not cause the effect to occur without a triggering event. A gas leak creates conditions that can result in an explosion, but they will not do so without the triggering event of a spark.
- Proximity: A "proximate" cause occurred closer in time or distance to the effect - it is often assumed, but is not necessarily, the primary cause of the outcome
- Randomness and determination: A random event is one that is not causally determined by what happened earlier - which is to say it is not in the chain of causation, but contributes to it.