Why Causal Knowledge Is Pragmatically Possible: Bridging Philosophical Skepticism and Scientific…

When we observe the world, we note that events consistently follow one another. As David Hume famously argued in An Enquiry Concerning…

When we observe the world, we note that events consistently follow one another. As David Hume famously argued in An Enquiry Concerning Human Understanding (1748), we do not perceive an inherent “force” or necessary connection that binds a cause to its effect — only a regular succession of events. This observation has led some to claim that a fully justified account of causation is epistemologically elusive (Hume, 1748; Cartwright, 1983). I find it frustrating when some scientists, in my view, are too quick to dismiss causal inference in favor of vague correlational claims. To be clear though, I, personally, am not after some metaphysical grounding which I believe to be unachievable, but after a pragmatic way of thinking.

Some of my neuroscience colleagues believe that everything is correlational and that hedge words like suggest are the right way of dealing with the problem. This article argues that randomized experiments are the way.

The Challenge of Observational Causality

In any complex system, consider two variables, X and Y. While we can observe their joint probability distribution p(x,y), determining whether X causes Y, Y causes X, or if an unobserved variable Z influences both, remains challenging (Salmon, 1984). This difficulty is central to the philosophical debate on causation. Yet, this very challenge underscores the pragmatic methods we use to decipher how the world operates. If we want to ask if a neuron influences another, a paper influences its readers, or a transistor influences another transistor.

The Interventionist Framework

Judea Pearl’s work has been instrumental in formalizing causal inference. In standard Bayesian reasoning, we consider p(y∣x) — the probability of Y given X. However, Pearl’s introduction of the “do” operator allows us to express the outcome of an intervention, written as p(y∣do(x)), which represents the probability of Y when X is actively set to a value (Pearl, 2009). This framework underpins modern methods for establishing causal relationships by focusing on interventions rather than mere observation. Key is that we ask if X causally influences Y. Not what the nature of the path is through which X influences Y (which may be unknowable, so what).

Three Sources of Pragmatic Causal Knowledge

In the middle is us experimenting the world. Everything we know emanates from there.

1. Randomized Experiments:
   Randomized controlled trials (RCTs) are the gold standard for estimating causal effects. By randomizing treatment assignments, RCTs mitigate confounding influences, providing unbiased estimates of the effect of an intervention. For instance, experimental studies consistently demonstrate that activating a switch leads to illumination:
   p(light on|do(switch on))≈1,p(light off|do(switch off))≈1 (e.g. Rubin, 1974). Under pretty harmless assumptions we obtain a proof of at least estimating the average effects in an unbiased way.

2. Domain-Specific Understanding:
   In fields like neuroscience or electronics, domain expertise allows practitioners to draw on established principles. While a simple hand movement may not directly cause illumination, our knowledge of circuitry informs us that closing a switch completes an electrical circuit, thereby producing light. Such insights refine our causal models and improve experimental design. They usually ultimately derive from prior randomized experiments. A certain amount of domain understanding is also what enables approaches like quasiexperiments to produce rather convincing causal results.

3. Scientific Testimony and Cumulative Evidence:
   The advancement of scientific knowledge relies on the collective validation of results. When multiple, independent randomized studies yield similar causal conclusions, the scientific community gains a pragmatic, if not philosophically definitive, understanding of the underlying causal mechanisms (see application in neuroscience). Again, this approach justifies itself from past success at describing causality.

Reconciling Pragmatism with Philosophical Skepticism

It is important to recognize that while experimental methods, such as randomized experiments and conceptual ideas, such as the do-operator have proven effective, they do not entirely resolve the deep epistemological challenges posed by Hume and others. Instead, they offer a pragmatic means to navigate complex systems and make reliable predictions — an approach that is invaluable in both neuroscience and computer science. As Woodward (2003) argues, causal explanation in science is best understood as a tool for manipulation and prediction rather than a direct window into metaphysical truths.

Conclusion

The philosophical debate over the nature of causation remains a rich and ongoing discussion. However, the practical methods described above illustrate that, for all intents and purposes, we can acquire and use causal knowledge to effectively interact with and understand the world. This pragmatic perspective is essential for driving progress in scientific inquiry, even if the ultimate metaphysical nature of causation continues to be debated. And let’s be real, we are willing to trust our lives to our causal understanding of the world: we are willing to sit in tin boxes that move at speeds incompatible with human survival if they were to crash.

P.s. for the neuroscientists

In neuroscience, causality is often hard and in many domains not currently possible. This in no way undoes the problem that knowledge obtained from randomized experiments (or derived from insights obtaind from other randomized experiments) can be made meaningfully reliable in a pragmatic way while correlational knowledge simply can’t. Correlational knowledge simply does not provide meaningful causal insights.