3. Causal reasoning

Many of the most significant problems we face involve reasoning about

causal connections. But thinking clearly about causal claims is a very tricky

business. For one thing, in colloquial speech the term ‘cause’ is used in a

number of different ways (e.g., to signify a probabilistic causal connection

or a necessary or sufficient causal condition). Further, there are various

ways in which controlled experiments can go awry—and these are often

quite difficult to detect. We don’t intend to give an easy-to-use strategy for

reasoning about any and all causal claims. In fact, we doubt such a strategy

exists. But there is a fairly simple strategy for avoiding a certain kind of

pitfall when it comes to causal reasoning. (This strategy is implicit in much

of Robyn Dawes’s work, e.g., 2001.) The neglected risk involves accepting as

a causal explanation a narrative that lacks a control. To see this, consider

two points: the importance of controls in causal reasoning and our tendency

to accept plausible sounding narratives (Trout 1998, esp. chapter 8).

Suppose we want to know whether Snake Oil Hooch is an effective

cure for the common cold. We give the Snake Oil Hooch to 100 people

with a cold, and all of them get better within a week. Is that good evidence

for thinking we have a cure for the common cold? Of course not. We

would expect (near enough) all 100 people who get over a cold to get over

it without a medical cure. To know whether the Snake Oil Hooch had an

effect, we would want to run a controlled, double-blind experiment. We

would want to study a control group—a group of people with colds who

aren’t given the Snake Oil Hooch. And we would want the experiment to be double-blind (neither the subjects nor the people who diagnose the

subjects know whether subjects are in the control or experimental group).

If there is no significant difference in outcomes, then it is reasonable to

suppose that the Snake Oil has no effect.

Most people recognize the importance of controls in causal reasoning.

And yet, we have a tendency to accept compelling narratives as evidence

for causal claims. Let’s consider a somewhat amusing example. Some people

believe that shaving hair causes it to grow back thicker. This is supported

by observation: When people start shaving, their hair does tend to

start coming in thicker. And people also often support this causal claim

with some sort of ‘‘intuitive’’ explanation. We’ve heard two. One likens

hair to certain plants that when pruned come back thicker; another holds

that the razor ‘‘stimulates’’ the hair follicles. But this is an old wives’ tale.

It is an example of correlation but not causation. Those who propose this

hypothesis don’t have a control group. While they know that shaving and

thicker hair are correlated, they don’t know what would have happened to

the hair without the shaving. And in fact, merely raising the question leads

to an obvious alternative hypothesis: Most people start shaving in early

puberty, when their facial and body hair begin to grow in. But puberty

itself explains the increased hair growth.

This problem suggests a relatively simple debiasing strategy. When

faced with a causal hypothesis of the form X causes Y that is supported by a

narrative but no control, it is often useful to ask a very simple question of

the form: What would have happened to Y without X? It is important to

recognize that this debiasing query, by itself, doesn’t immediately lead us

to the right conclusion about what’s causing what. But if one has no idea

what would have happened to Y without X, and if one recognizes the importance

of controls in causal reasoning, then at the very least this should

give one pause. Perhaps one doesn’t have particularly powerful reasons to

believe the causal claim. So while the consider-the-control strategy is only

a first step in thinking about causal claims, it is often an effective first step

in opening up fruitful lines of investigation that can perhaps help us to

avoid the temptation of falling for an intuitive but mistaken story. Let’s see

how this might work with three high-stakes examples.

Many of the most significant problems we face involve reasoning about

causal connections. But thinking clearly about causal claims is a very tricky

business. For one thing, in colloquial speech the term ‘cause’ is used in a

number of different ways (e.g., to signify a probabilistic causal connection

or a necessary or sufficient causal condition). Further, there are various

ways in which controlled experiments can go awry—and these are often

quite difficult to detect. We don’t intend to give an easy-to-use strategy for

reasoning about any and all causal claims. In fact, we doubt such a strategy

exists. But there is a fairly simple strategy for avoiding a certain kind of

pitfall when it comes to causal reasoning. (This strategy is implicit in much

of Robyn Dawes’s work, e.g., 2001.) The neglected risk involves accepting as

a causal explanation a narrative that lacks a control. To see this, consider

two points: the importance of controls in causal reasoning and our tendency

to accept plausible sounding narratives (Trout 1998, esp. chapter 8).

Suppose we want to know whether Snake Oil Hooch is an effective

cure for the common cold. We give the Snake Oil Hooch to 100 people

with a cold, and all of them get better within a week. Is that good evidence

for thinking we have a cure for the common cold? Of course not. We

would expect (near enough) all 100 people who get over a cold to get over

it without a medical cure. To know whether the Snake Oil Hooch had an

effect, we would want to run a controlled, double-blind experiment. We

would want to study a control group—a group of people with colds who

aren’t given the Snake Oil Hooch. And we would want the experiment to be double-blind (neither the subjects nor the people who diagnose the

subjects know whether subjects are in the control or experimental group).

If there is no significant difference in outcomes, then it is reasonable to

suppose that the Snake Oil has no effect.

Most people recognize the importance of controls in causal reasoning.

And yet, we have a tendency to accept compelling narratives as evidence

for causal claims. Let’s consider a somewhat amusing example. Some people

believe that shaving hair causes it to grow back thicker. This is supported

by observation: When people start shaving, their hair does tend to

start coming in thicker. And people also often support this causal claim

with some sort of ‘‘intuitive’’ explanation. We’ve heard two. One likens

hair to certain plants that when pruned come back thicker; another holds

that the razor ‘‘stimulates’’ the hair follicles. But this is an old wives’ tale.

It is an example of correlation but not causation. Those who propose this

hypothesis don’t have a control group. While they know that shaving and

thicker hair are correlated, they don’t know what would have happened to

the hair without the shaving. And in fact, merely raising the question leads

to an obvious alternative hypothesis: Most people start shaving in early

puberty, when their facial and body hair begin to grow in. But puberty

itself explains the increased hair growth.

This problem suggests a relatively simple debiasing strategy. When

faced with a causal hypothesis of the form X causes Y that is supported by a

narrative but no control, it is often useful to ask a very simple question of

the form: What would have happened to Y without X? It is important to

recognize that this debiasing query, by itself, doesn’t immediately lead us

to the right conclusion about what’s causing what. But if one has no idea

what would have happened to Y without X, and if one recognizes the importance

of controls in causal reasoning, then at the very least this should

give one pause. Perhaps one doesn’t have particularly powerful reasons to

believe the causal claim. So while the consider-the-control strategy is only

a first step in thinking about causal claims, it is often an effective first step

in opening up fruitful lines of investigation that can perhaps help us to

avoid the temptation of falling for an intuitive but mistaken story. Let’s see

how this might work with three high-stakes examples.