Have You Got A Causal Effect?

Causation is a tricky thing. I remember discussing different conceptions of causality in undergrad and can’t say I’ve grasped it yet. In some ways, it seems like an easy question — we have lay theories of what causes something else everyday. Still, push harder and things get murky pretty quickly. Giving up isn’t an option though. In scientific study, you really must have an idea of what causes something else. You can’t just say stuff happens. Well, you can do that, but it is unlikely to be published. (Unless you have some embarrassing dirt on the editor which, thinking about it, probably explains some papers I have read). So have you got a causal effect? How would you know?

Casual Inference

Paul Rosenbaum has a simple guide to causal inference. It starts with the basics of looking at a single intervention. You really don’t know what effect it had. Did the doctors kill George Washington with their interventions? How will we even know? Causal inference is a challenge because we don’t see Washington in two different worlds one where he got the doctors’ treatment and one where he didn’t. An approach with more promise is to repeat an intervention (or lack of it) over a number of similar circumstances and to note the outcomes. If you treat lots of people and they die, while you don’t treat lots of others and they survive, the doctors are looking pretty guilty.

Bad Attempts To Capture Causation

Rosenbaum notes a number of truly terrible ways to try and understand causation. A lot of them have to do with violations of random assignment. You want two groups to be as similar as possible before a treatment/intervention to be able to claim that any change might have been caused by the treatment/intervention. If you only use leaches on really sick patients it shouldn’t be too much of a surprise that those who get leaches are more likely to die. Don’t blame the leaches, it was your experimental design that was a problem.

Observational Studies

In life, we often don’t get to run experiments. You can’t randomly assign people to, for example, smoke or not smoke. Instead, you have to use clever thinking to be able to argue that what you see is a result of their smoking and not some third factor that causes both death and smoking. Perhaps people who are ill smoke to relieve their symptoms and tend to die because they were sick before they smoked.

When you observe what happened, rather than randomly assign interventions yourself, you have an observational study. You still want to know, ‘have I got a causal effect?’, but you need to be much more careful and inventive. Can you find some randomness that happens in the world to hitch your study to? A lottery run for other purposes? A somewhat arbitrary cut-off point? A distinction imposed upon groups that are similar to start with but one group accesses a different treatment. You need to be more careful in your claims but if you can accumulate evidence that all points the same way you start to be somewhat convincing.

Have You Got A Causal Effect?

One thing to avoid is the simplistic argument that we can’t really know anything. In many ways it is true, but it isn’t helpful. There are some things that seem so certain given the preponderance of evidence that a reasonable person can act with the assumption that they understand the causal effects. Enough people have died early after smoking, that makes it reasonable to conclude that it causes disease and death. Indeed, it is unreasonable not to conclude that given the evidence. If you want to dispute this you need some pretty convincing counter-evidence.

While getting evidence of causation is hard, the very act of searching for evidence helps us progress our knowledge. The simplistic idea that we can reject well-accepted scientific ideas because your friend knows a smoker who survived to 100, or that a cold day means that climate change is a hoax, is not likely to help further human understanding.

For more on experimentation see here, here, and here.

Read; Paul R. Rosenbaum (2023) Causal Inference, MIT Press Essential Knowledge series