What is a cause?

Before anything else, we need to know what we mean by a “cause.”

What we talk about when we talk about causes?

Imagine an event, process, or state such as a hurricane, going to school, or being a women causing another event, process, or state such as damage, having a high-paying job, or happiness. The first event, process, or state is called the cause and the second event, process, or state is called the effect. The cause is partially (and not necessarily fully) responsible for the effect. To make things more complicated the cause can be the effect of another cause. Similarly, an effect can itself cause other effects.

In the example from the previous lesson, the cause is education and the effect is higher earnings (or simply earnings). In the figure, the direction of the relationship is from the cause (education) to effect (earnings).

In this course, and in empirical science more broadly, we’ll distinguish between three different types of causes:

1. Some causes can be assigned to subjects in practice, i.e., the researcher can choose who receives the cause and who doesn’t. Exercise, taking medication, and enrollment in a training program are all causes that can be assigned.
2. Some causes, however, can only be assigned in theory, and not in practice, mainly due to ethical or practical concerns. Smoking or traffic accidents are some examples of such types of cause. Causes of type 1 and 2 are called events. Hypothetically, we can imagine assigning events to subjects but realistically, due to ethical or practical reasons, we can’t assign them.
3. Lastly, some causes cannot be assigned whatsoever (either in practice or in theory) meaning we can’t even imagine them being under the control of the researcher. Think about blood type, gender, or height. These types of causes are referred to as states or attributes.

Paul Holland, a well-known statistician argues that there is “no causation without manipulation.” He argues that if we can’t manipulate the cause, it makes no sense to investigate a causal effect.

Donald Rubin, (who we’ll hear about a lot in this course), also claims that causes and effects can only be defined in a causal setting if all units can be imagined to receive or not receive the treatment. With this argument, since “femaleness” can only be imagined for, well, females, it can’t be regarded as a cause.

Imagine, the causal effect of a person’s height on their weight. We can’t increase or decrease the person’s height. There is no way to even theoretically observe what the person’s weight would be if she were taller because there is no way to make her taller.

Therefore, to make things a bit easier for us, we’ll mainly focus on causes that can be manipulated such as exercise, nutrition, schooling, peer effect, etc. To understand this, imagine the following statements:

1. Tania didn’t get the job because she is a woman.
2. Tania got the job because she worked hard for it.
3. Tania got the job becuase she was prepared by her advisor for the interview.

The cause is linked to an attribute in statement 1 but to some event/activity that was voluntary in statement 2 and forced on her in statement 3. Holland argues that an attribute can’t be a cause in a causal study. The cause can’t be manipulated in (1), but it can be easily manipulated in (2) and (3). Therefore, statement 1 can’t be framed as a causal study.

This does not mean, however, that we can never study the causal effect of attributes such as gender or race.

For instance, if in statement one example we change the unit of observation from individuals to, let’s say, companies and change the treatment to the share of women who get the job in those companies, we might have a causal study. Note that the share of women in a company is not an attribute that can’t be manipulated. We can easily design or imagine an experiment where we assign different shares of women to each company.

In one of the most famous and experiments in the field of economics, Bertrand and Mullainathan studied the effect of race in the labor market. In the paper, the researchers ran a randomized experiment in which they sent fake resumes to actual companies who were looking for potential job candidates. They found that white-sounding names received 50% more callbacks for interviews that other names.

But didn’t we just say that race is an attribute and, therefore, can’t be manipulated? How is this question causal but not the question in (1)?

Well, Holland would argue that by changing the definition of treatment we can redefine the question. In this paper, Bertrand and Mullainathan did not directly study the effect of race on callbacks but rather the effect of “perceived race” on callbacks. Perceived race (done by manipulating names on a resume) is not an attribute anymore and can be manipulated in an experimental setting.

Practice

Is age an attribute/state or an event/action?

It’s an attribute/state.

It’s an event/action.

Cause and effect in randomized experiments vs. observational studies

A cause is defined in the same way in both randomized experiments and observational studies. The only difference is that the researcher has control over how the cause is distributed among subjects in a randomized experiment and not in an observational study. In the next module, we will learn about how randomization can help us better answer causal questions.

Observational studies usually refer to studies in which assignment to treatment or control isn’t randomized and the researcher has no control over that. In other words, the investigator can’t impose on or withhold from the subject the treatment under investigation. However, in randomized experiments (aka randomized control trials or RCTs), the researcher can control who gets and who doesn’t get the treatment and the assignment to treatment or control is random.

Read more

Want to know more about causes? Read this paper.