Causality and the Causal Inference Revolution, part 1
Unravelling the web of an emerging scientific paradigm.
On The Value of Causal Thinking
Providing causal explanations lies at the foundation of a myriad of disciplines and fields of knowledge. Diverse theories of causation underpin our very conceptions of reality and have critical implications for scientific and philosophical work. In history, the humanities, and social sciences, attributing causation is an essential part of unraveling major historical and social dynamics that help us explain and interpret society. In philosophy, the metaphysics and epistemology of causation have motivated philosophical inquiries for thousands of years. In epidemiology and medicine, understanding the causes and origins of disease has profoundly altered human capacity to intervene in health behaviors and illnesses, extend human health and wellness, and shape the health of populations. In the realm of natural sciences and mathematics, causal relationships are harnessed in scientific models and used to describe and predict phenomena with increasing precision, unlocking new capabilities and inventions, and even opening new questions about the fundamental fabric of reality.
From the lens of the social sciences we recognize that causal claims are not just the domain of scientists and philosophers, but are a routine part of our social and psychological reality and deeply affect our sense of self-awareness and our worldview. In this sense, humans have been thinking causally for as long as we’ve had minds to think. However, clarity of thought about causality is also profoundly challenging for the human mind, despite causal predictions being a fundamental matter of cognition and a regular part of everyday life.
The matter becomes more complex when we attempt to systematically explore reality. Assessing the nature of cause and effect has become a matter of immense importance for all domains of knowledge as we have reached the limits of contemporary paradigms of scientific and philosophical thought. Estimating ‘causal effects’ has become the key aim throughout much of scientific literature, from psychology to economics, to clinical research and public policy interventions.
Yet inquiry into the nature and characteristics of causality and theories of causation are often limited to obscure subsets of curricula, in advanced methodological classes, or buried deep within specialist literature across many fields. Despite how essential understandings of causation are for knowledge claims that are central to most disciplines, few gain a systematic overview of the concept of causation as it has developed historically, or across disciplines. At most, thinkers will become fluent in the topic as it relates to their own field of study while lacking a deeper understanding of how theories of causation have developed and causality has been understood throughout time and across various cultures and disciplines.
To start solidifying a more systematic and reflexive approach to the emerging paradigm of causality, I suggest a comprehensive approach is needed: interdisciplinary, cross-cultural, and transhistorical approaches that survey perspectives on causation across a diversity of fields of knowledge, and throughout time and place. Relatedly, students should be exposed to the fundamentals earlier in their studies - to prepare students for later advanced studies in their own disciplines, but also to foster a collaborative and interdisciplinary approach to understanding causation while students are still at a stage (undergraduate) where their studies are more general in scope.
A Tentative Outline for a Syllabus on the Cultures and Sciences of Causality, Past to Present:
Hindu/Vedic philosophy, Vedanta & Buddhist Cosmology
Causation in East Asian philosophies
Causation in Indigenous thought
Greek philosophy
Medieval (European) approaches to the philosophy of Causation
Causation in Arabic and Islamic philosophies
Early Modern European approaches to the philosophy of causation
The Emerging Contemporary Paradigm: Approaches in Science, Philosophy and Mathematics
Causality in the Humanities and Social Sciences
Applied Causal Thinking: Tools and Approaches for studying causality, knitting together the natural and social sciences
The Causal Inference Revolution
Attributing causation has become an essential aim of the predominant scientific paradigm in the past 40 years. Even still, it has been little recognized a revolution outside certain domains. Only more recently has the scope of these changes begun to be recognized and widely proliferated. It has been prodding along in the background of many disciplines, but, likely due to the lack of robust interdisciplinary interactions, the true scope of the changes across many fields seems to have gone unnoticed by many. A multitude of fields labor in their own lane using similar approaches, sometimes going by different names even when referring to the same concept. This appears to be changing slowly in the past two decades, as conclusions from certain prominent thinkers on causality have consolidated into a clearer vision of what this paradigm offers.
The emergence and growth of “evidence-based” ____ (medicine, psychology, policy, etc.) is one outcome of these changes. Certainly, the permeation of statistical methodology and quantitative research methods throughout the social sciences has been another avenue for the proliferation of causal inference-oriented thinking (though the two are sometimes confused as being one and the same). The largest gaps between disciplines remain between the philosophical and historical approaches to causation and the quantitative wings of the natural and social sciences.
Even within natural and social sciences, where many of the more recent approaches have been formulated into specific methods, techniques and perspectives, there is a bifurcation that feeds into conflicts about how claims are justified by either end of the spectrum of the sciences. At its worst, natural science-oriented thinkers dismiss social science as baseless and un-empirical, or politically or ideologically captured. Broadly, they appear to clash and fail to learn from the implications of either end of the debate; generally, the social sciences are more likely to be dismissed or even declared fraudulent, especially as a certain, narrow type of scientific and technological advancement is sought by the prevailing neoliberal social and economic order of our day. Coincidentally or not, it is these exact circumstances that the social sciences are often critically challenging, whether intellectually or politically, so it is perhaps not a surprise that there would be an attempt to undermine their status as disciplines in response.
In the social sciences, this conflict around the foundations of knowledge is highly productive, despite appearing at first glance to only be attacking its general influence. The tension resulting from the social sciences feeling the need to develop robust and evidence-based models of social reality has, on the whole, driven a reflexive desire to develop more tools to accurately understand social phenomena. Many have proven inadequate, even entire sub-disciplines proving to be lacking. Others have proven highly influential. The use of quasi-experimental designs, for instance, has given new impetus to the study and implementation of large-scale public policy interventions, granting a new arena of practical impact for knowledge derived from the social sciences. The demand to provide clearer evidence in light of the shifting, multifaceted, and highly complex nature of social reality has led to the developing of new methods and models, such as agent-based modeling or network analysis, that grant new insights into certain aspects of social life.
Questions of causality are threaded throughout all these investigations. For instance: that social, economic and environmental determinants of health may be addressed by large-scale public policy interventions (whether fiscal, regulatory, or social safety net policies) is a cornerstone of public health research and recommendations. However, such interventions evoke concerns about the challenge of attributing causality to the intervention vs other confounding effects; naturally, the larger the intervention in scale, cost, etc, the more accountability is demanded for its quality, efficacy, and impact.
As the demand has grown to further support causal claims and justify certain interventions, whether medical, social, economic or otherwise, the need for a more comprehensive paradigm for attributing causation has become clearer. Consolidating this thought across disciplines and reshaping the overall thrust of scientific research productivity in response to it has been the fruit of several decades of exploration and effort. Only more recently has it truly intensified, as several previously distinct fields of study (science and technology studies, metaresearch, and others) have finally begun to cross-pollinate. Additionally, the need to resolve critical failings has become more apparent in the face various knowledge crises, such as the oft-considered replication crisis, or lesser-recognized but no less challenging crises at the heart of modernity. What remains missing is a concerted effort to systematically integrate the more powerful implications from these various disciplines that have all oriented around related questions at the heart of causal thinking. Most clearly the gap lies between humanitarian, historical and philosophical understandings, and social and natural scientific understandings, a theme we will explore more here and in future posts.
‘Correlation is Not Causation’, but Why Association Still Matters
Undoubtedly ‘correlation does not equal causation’ has become one of the more abused cliches, a side effect of a broader lack of clarity on the implications of advancements in research methodology that focus on causation and that have come to characterize an array of disciplines in the past several decades.
Differentiating between correlations (or association) and causation is certainly a core part of causal inference and scientific endeavors, but we sometimes go too far in discounting associations in our search for a purely causal effect. This race towards ‘purer’ and more complex approaches to distinguishing a causal effect leads to some disjointed reasoning. Conventionally, the drive to identify causal relationships leads to associations becoming discounted, essentially viewed as ‘not real' while the desired causal effects become 'real’.
In one frame of mind, there is truth to this. Most of us when receiving basic instruction in our fields are hit over the head with neat and tidy examples where some association tricked us and the causal effect was actually something counterintuitive. However, in a study of a particular group defined by two factors (variables) that are associated, just because we know that they are not causal - or more accurately, cannot prove they are causal - does not necessarily mean the association ‘isn't real’. Instead, we are disproving certain assumptions about the characteristics of the phenomenon under study, assumptions that we brought to the data. That is what we are dismissing, not necessarily the reality of the relationship itself.
Associations do exist in the real world. The underlying causal effect may be biased or obscured, which is usually what we mean when we are discounting a particular observed association as illusory. Nonetheless, the association is 'what it is'. Associations do bear salient information, hence the idea that every association is just an illusion or an obfuscation of the truth, waiting to collapse, is shortsighted.
Indeed, when causal effects are so highly sought after and associations discounted, the temptation becomes ‘how do we massage this association into causation’. Research is replete with examples of such attempts and this hinders the scientific output of humanity at large. The systemic effects on bodies of literature that come from this are something we are only just barely beginning to grasp (especially in metaresearch and science and technology studies). In the present, we also have limited tools to address these systemic hindrances affecting research at large, which is why when it is brought up, many seem to kind of just collectively shrug when asked what can practically be done about it.
One source of this pressure is that, in the face of the demand to attribute causality, the task of description is systematically undervalued, and description, under causal inference frameworks, is essentially elaborating associations. Associations may not tell us exactly the answer we are looking for, in that they may either be disguising the underlying relationship of interest OR in that they are all we may access in the face of limited means to identify a causal relationship.
Let’s address the latter with the example of cross-sectional evidence. Research efforts frequently encounter limits in identifying causal relationships, and cross-sectional studies are a notable example. Is a cross-sectional study somehow then worthless because it only describes where certain attributes of a defined population lie? Only in the frame of mind that pressures us to think solely in analytic and interventionist terms. Taking even further, is it worthless if a prior cross-sectional study already described that same population? Even then, still, no. This is seemingly obvious, yet this kind of thinking permeates research disciplines.
There is a misstep here that actually reveals a critical oversight. If some span of years have past in a living and dynamic emergent system like a population, with complex social and historical forces operating on it… what makes us think that description is no longer relevant? How would we know whether pivotal changes have or haven’t occurred broadly at the population level? Is it even feasible for most causally-oriented investigations to operate at that level without the foundational, complementary value of descriptive studies? These tacit beliefs betray an underdeveloped model of social reality. And yet even among those deliberately focused on studying society, description is still so heavily undervalued as the drive towards quantitative techniques often comes at the expense of the qualitative, especially as causal inference approaches come to rely ever more complex mathematical modeling.
This helps explain the common anxiety of those in the social sciences, especially fieldwork-based and community-based approaches, given that ‘sample size’ is often naively assumed as the final arbiter of the validity of some research conclusions, or that causality can only ever be fruitfully explored in some particular type of interventionist framework, a la randomized controlled trials or other methods that seek to mimic similar conditions. In reality, the impression that causality can only be usefully explored under certain conditions is not the whole truth; a more comprehensive understanding of causation across disciplines helps ameliorate this impression. We’ll explore exactly how this is the case in more depth in future posts.
When it comes to the former misconception, the idea that associations are merely uncollapsed illusions hiding the causal truth, this too merits deeper exploration of the framework of causal inference. The impression that associations are just unclarified causal relationships misses the mark and can lead to absurd oversights. This is especially the case for complex social topics. For example, you can observe this misconception when ideologues attempt to downplay the significance of something like gender inequality by pointing to data that seemingly suggests that the ‘association’ between gender and inequality disappears when adjusting for other factors. This clash is set up by the fact that theories of social and historical domination are mostly elaborated in the social sciences, and a certain type of thinker willing to dismiss the evidence leveraged by the social sciences often gravitates towards strictly analytic approaches. Such an argument often overlooks the nuanced nature of causality through reductionism, and by ignoring or obliterating social context or theoretical approaches.
For instance, consider a hypothetical study examining the gender discrimination in the workplace. Someone who argues against the idea of gender inequality may argue that when factors such as education, occupation, and work experience are taken into account, the gender discrimination appears to diminish or even disappears entirely (attentive readers may already have seen many examples of this argument). They may interpret this as evidence that gender inequality is not a significant issue, ‘look, when you account for these confounders, it explains the disparity!’. The disparity is attributed to some other factor, such as women's career choices, individual preferences (or at its worst, a misogynistic belief in women’s inferiority).
However, a deeper exploration of the framework of causal inference reveals that this interpretation is faulty in several critical ways:
Oversimplifying Causality: the loose approach to adjusting away major factors such as education or occupation can indeed change the observed association between gender and some outcome of interest (in retrospect, how could it not?). However, the mistake comes from an impoverished view of social phenomena that underestimates the complexity at play in these factors, particularly the multifaceted and interconnected nature of social roles, experience, and attaining some marker of status (educational, occupational, or otherwise). Compressing all that social complexity into a single variable is already a major limitation that few researchers seriously appreciate, as most social science or qualitative researchers will often decry. Under the mistaken premise that associations are just illusions, we are more likely to overlook that highly complex social phenomena, like inequality, operate through systemic effects that are not evenly or directly distributed within one or another variable or neatly compressed into standard variables (like “socio-economic status”, an extremely popular, yet limited attempt to capture something far, far more complex). Even further, aside from the loss that comes from that kind of compression, systemic biases, societal expectations, and historic forces shape fundamental causes that feed into the data we collect on such factors in highly complex ways. When ‘adjusted away’, this may leave serious gaps in our understanding of the data before us.
Context matters: It turns out, yes, when you ignore and cut out indicators of the broader social context, it may appear as if some factor, like gender, all of a sudden is no longer significant. But adjusting for this salient factor is essentially like saying the following, 'Yes, it turns out that when you adjust away the pathways and factors that gender inequalities operate through, things no longer appear unequal!’ You might as well say ‘when you adjust away inequalities, things are no longer unequal!’ Wow, who would have thought! This inadvertent (or purposeful) disregard for the complexities of the situation leaves us spinning wheels with totally suspect conclusions about society that directly contradict other forms of research that deliberately explore the contextual and experiential factors at play. The integration of other forms of knowledge brings clarity to our understanding of whether certain harmful forces, like inequality, are present or not and what to do about them. An inability to appreciate/understand how other modes of research reach conclusions is at the heart of the challenge here.
Causation vs Association: None of this is to say that distinguishing causation and association is not important. It is of critical importance. What’s needed instead is a deeper understanding of the complex web of causality that forms social reality. This example illustrates that we don’t come into research efforts agnostic. We can’t pretend that our preformed impressions don’t have an impact on the way we interpret evidence - hence the blindspot of those seeking to dismiss the prevalence of gender inequalities and their willingness to poorly interpret evidence. The ideologue who dismisses gender discrimination and yet claims to be unbiased and only deriving conclusions from the data is a perfect example of the depth of the biases we bring into such scientific research efforts, no matter how technically skilled we are at some method or analytic approach. Resolving that is a much deeper task; what we can start with is dissuading ourselves of the idea that adjusting for variables of importance somehow negates the existence of significance of the observed association, or that associations do not bear meaningful information even when a comprehensive causal explanation is out of reach.
Instead of blindly dismissing association with cliche phrases, or frantically trying to pretend they don’t exist, what we should do is speak of the causal effect when we want to speak of the causal effect, and speak of associations when we want to speak of associations. When we wish to describe and understand associations make it clearly the aim and don’t shy away from it just because they have to connotation of ‘unproven causal evidence’. Certainly this doesn’t mean we won’t attempt to discern more clearly what IS happening causally, even when using evidence such as cross-sectional data. This is still of critical importance. Rather, we need a deeper consciousness of when and how to make clear causal claims when possible and to differentiate the meaning that comes from examining presumed causal relations, and associations. Extreme care is needed, but at the same time we should not forget why things are associated and yet not causal. Associations are instructive, they bear immensely useful information, and reveal understandings about deeper structures of the phenomena we are studying. The greater the complexity, the more we need to avoid haphazard shortcuts. Finally, we need more than just the ability and fluency to work with data. A theory-impoverished approach to studying causality is simply untenable.
Traditional Thinkers in the Philosophy of Causation and the Under-appreciated Intellectual Heritage of Humanity
David Hume, Karl Popper, Ibn Sina, Aristotle. When it comes to assessing the history of thought on causation these are some of the main thinkers that seem to get all the credit. Certainly they made indelible impressions on how we understand these topics. However, surveying the deep landscape of human thought and broadening one’s perspective on what philosophies you draw inspiration from leads to the recognition that diverse thinkers across many regions and cultures have been leveraging causal claims, and thinking deeply about causality to explain and interpret the world and its phenomena throughout human history. The rote framework for the history of ideas that most are exposed to in conventional education is sadly lacking, and stunts our understandings of the true complexity and breadth of human thought on causation. Consider the 2nd c. Buddhist philosopher Nāgārjuna. Consider the cosmological claims that interpreted and provided meaning about the origins of the universe and humanity that underlie the myths and creation stories of countless different cultures.
The many oral and textual traditions on the origins of the world and human beings do not merely encode differing understandings; they actively shape entire worldviews that lay at the foundation of diverse ways of structuring society and social relations. The causal claims embedded in understandings of the universe and depictions of humanity play a pivotal role in conceptions of human nature. Finally, they may make or break relations between cultures and communities of human thought that, without putting in the necessary work to build common understanding, may fall prey to catastrophic ways of relating between cultures. At its worst, this failure perpetuates material and intellectual imperialism and stunts the growth of human thought.
