Causal humility is the attitude that arises from the following two premises. First, the causal structure of the world is rich and layered, ranging from microscopic causal processes to group level causation and beyond. Second, humans are only good at reasoning at the causal level that has contributed to our survival during our evolutionary history; when we attempt to reason at other causal levels, we are ham-fisted and error prone.

As an example, consider the PM2.5 particles emitted by the cars currently passing me by on a nearby road as I’m sitting here on my deck. Some of those particles will plausibly make their way into my lungs, cross the blood-brain barrier, and increase my risk of developing some brain disease or other by an unmeasurably small probability.

Good luck “proving” the above causal claim. To do so, you have two options. One option is to conduct a population level study that examines the relationship between, say, living close to a major highway and developing averse long-term health outcomes. Studies like these are notoriously difficult to get right given the pre-existing differences between those who live next to major highways and those who don’t.

The second option is to develop a mechanistic account of the various molecular-level goings on that could plausibly take place. The particle lands somewhere in the lung, makes its way to blood, passes through the cells in the brain, and so on. The problem with these stories, of course, is that we can never observe any given PM2.5 particle doing all these things.

It’s taken years of consistent effort by researchers to start establishing the causal effects of PM2.5 exposure. Contrast that to our ability to infer causal relationships at a level that is natural to us. For example, suppose there’s a diesel truck idling on that same nearby road and its fumes make their way into my office through an open window. Not only am I able to immediately pinpoint the cause of the bad smell; I’m also able to generate countless counterfactuals to understand what I could’ve done differently to prevent the event, such as closing the window that’s open.

My claim is that the only difference between the diesel truck and PM2.5 scenarios is that one of them happens at a causal level that is natural to us. If figuring out microscopic causal relationship carried a similar survival value to us than figuring out the sources of bad smells does, we would be have an equal ability to reason about the effects of individual PM2.5 particles as we have to reason about the effects of nearby idling diesel trucks.

Randomized controlled experiments are a delightful demonstration of our limited causal reasoning ability. When toddlers start learning about causality, they sometimes walk around with toy hammers to verify that, when you hit things with hammers, those things tend to break. However, because toddlers are conducting these experiments at a causal level that is natural to humans, they are able to generalise their findings in a remarkable way. After hitting things with hammers for a few dozen times, toddlers have an excellent idea of which kinds of items make for good hammers, which kinds of things tend to break when you hit them, and which kinds of conditions are favourable for breaking things with hammers.

Moreover, toddlers are able to reason about the downstream causal effects in a sophisticated manner. For example, it may only take a handful of experiments for a toddler to learn that their parent is equally unhappy about getting hit by a plastic hammer than he or she is about getting hit by a wooden hammer.

Randomized controlled trials are intuitive and appealing because they invoke the notion of causation-as-agency, which is intensely familiar to all of us. Since we all have learned a lot about causality by hitting things with hammers, we have a natural tendency to try to reason about all causal domains by invoking the “intervention” metaphor. (In our natural causal domain, the equivalent of randomisation is free will, which is a topic for another post.) Unfortunately, this type of reasoning tends not to work particularly well when extended outside our everyday causal experience, or what I’ve above called the “natural” causal level.

As an example, consider a typical tech industry experiment that’s randomized among millions of customers across the world. Here, we don’t have a good idea of what the hammer is, what exactly it’s breaking, and how any of it relates to almost anything we’ve learned in the past. Consequently, experiments in tech are generally used as accounting rather than learning devices (e.g. this team improved that KPI by 0.4%, and so on).

Causal humility is the attitude that the causal structure of the world is rich but that our tools for learning about that causal structure are almost comically limited. One of the implications from this is that we certainly should not pretend that any particular methodology reveals or even approximates the truth about causality. Randomized experiments, for example, may be a suitable tool when our goal is to answer certain types of accounting questions. But such experiments can hardly represent the “gold standard” for causal inference considering that the vast majority of our causal knowledge has nothing to do with randomization or experiments.

The Google Scholar search “PM2.5 long term health” yields over 300,000 results. Not all of those papers were strictly necessary, of course. But none of them alone could’ve served as the knock-down study to establish, once and for all, the negative health consequence of PM2.5 particles. Instead, these studies reflect the messy, arduous and frustrating process of causal reasoning spanning from the microscopic to the macroscopic level. What’s more, it’ll probably take another 300,000 papers to figure out the exact conditions under which PM2.5 exposure has its effects, what are the factors that protect against those effects, and so on.

If there’s any “method” to causal inference, it’s the process of articulating, criticising and refining assumptions. Sometimes, we need to call it and make a decision based on our necessarily limited understanding. So it’s not a coincidence that you won’t see me hanging out on my deck during the rush hour.