Tips and Tricks
What Causes What, and How Would We Know?
Facing a barrage of misinformation about COVID-19, climate change, racism, and other controversies, data scientists can use solid causal reasoning to help us figure out what to believe
An innate desire to understand why something important happened or might happen characterizes our hunt for causality. According to the Polish economist and philosopher Mariusz Maziarz (2020) however, more than 2,000 years of consideration have not brought us agreement on what causality is or how to safely assert it. This makes it hard to distinguish truth from hype.
According to James Woodward (2005) philosophers think of causality in two ways:
1. What it is and how we can infer causal relations, and / or
2. How people in different groups understand relationships between cause and effect.
Woodward’s second point is important, and I will come back to it in another post. For example, how we understand (or fail to understand) causality drives how we interpret statements about the origin of COVID-19; the utility of testing for it; how, when, and for whom testing should be done; whether social distancing, masking, and vaccines really work; and for whom.
How we think about causality and how we choose to infer it also drive where and how we look for causal relationships. This is important in science, medicine, and public policy. Our assumptions about and approaches to causal inference often result in public policies that may not work well to combat a perceived threat such as climate change or COVID-19. They also influence the size of federally or privately funded investments in cancer research, efforts to combat domestic or foreign terrorism, racism, health and wellbeing programs, and many other issues.
Should one causal mechanism apply to everyone, or do causes and effects differ by gender identity, age, race, ethnicity, education, region, country, time period, or income levels? We’ll investigate these issues next time. To do that well, we need a grounding in basic notions of what it means that one thing or event causes another, and how we should try to infer something about causality. We must also understand how to deal with misinformation based largely on poor causal reasoning (not to mention blatant political agendas and lust for power). In this post I provide that basic grounding by summarizing the work of leading experts in the field.
What Causal Notions Guide our Thinking?
Controversies about causality permeate almost every social issue, evidenced by our divided politics, the existence of bizarre conspiracy theories, and disagreements about which theories are conspiratorial vs real. Causal inference approaches (e.g., how we decide what the true causes of our happiness or maladies are) differ substantially on a variety of issues.
Examples include how much time (if any) should exist between the observed cause and its purported effects, and disagreement about the types and amount of evidence needed to safely assume that the situation or event was indeed causal. Did event A cause outcome B or did B cause A? Alternatively, was the downstream change in B that we observed due to something else entirely? Did doing more tests result in an increase in COVID cases, as some prominent folks suggested, or did those additional tests just find more cases that already existed due to other, more-likely causes?
In some situations, A and B may work in cycles, causing each other. For example, investing in the stock market may produce higher incomes and wealth, and these gains may then motivate more investment in stocks. These chicken and egg scenarios can be quite hard to parse.
Five approaches to isolate causal influences were investigated by Maziarz and other philosophers, and include the following. We can use these to improve our causal reasoning, with some provisos noted below.
Regularity assumes that a cause always produces an effect, with no exceptions. Laws of nature provide examples of regularity in action. Regularly occurring wave actions and tides due to gravity are examples. Some early philosophers required that a “B always follows A” pattern must exist to infer causality.
Probabilistic inference does not require that a cause produce an effect every time. In this scenario, causes raise the probability of their effects, but the effects do not always occur. The relationships between smoking and heart disease and various forms of cancer are examples of this — not every smoker gets these diseases, but very high percentages of them do.
Counterfactual inference addresses the question of how things would have been different if A had not occurred. Controversies about whether the government should expand unemployment benefits to reduce economic hardships from the pandemic can be viewed as disagreements about counterfactual scenarios. For example, some argue that providing an extra $300 per week in unemployment benefits is causing labor shortages in the US. They say if we did not do that (i.e., under the counterfactual scenario) more unemployed workers would be looking for jobs. So far though, the evidence supporting this counterfactual notion is weak. Labor shortages have improved very little after expanded COVID-19 benefits have expired, so the extra $300 per week probably is not causing worker shortages (Combs et al, 2021).
Mechanistic inference is another question of ‘how,’ often involving a deep dive into exactly how A causes B. For example, the mRNA science of how the Pfizer and Moderna vaccines bring about a high level of immunity to COVID-19 leads to straightforward explanations of cause and effect. This science highly influenced how those vaccines are made, stored, distributed, and applied, thereby reducing hospitalization and death rates from COVID.
Manipulation inference refers to ‘doing A,’ as opposed to letting A happen naturally. The implication here is that human or other intervention on A is what brings about outcome B. Related questions include how much of A it takes to make B observable, and if we vary the level of A then how much B can we expect to see afterward? Studies designed to find the safest and most effective doses of new drugs are examples of manipulations at work. The relationships between minimum wages, employment, and buying habits also illustrate the effects of manipulation causality.
Each approach has nuances or subtypes that differ in their underlying philosophical bases, challenges, and suitability for public policy, and Maziarz devotes most of his text to these issues. He also describes how these causal inference approaches can overlap with each other as they are applied. The following table is loosely based on his work and presents my summary of the five approaches he describes.
With five approaches to choose from as we make inferences about causality, is there one approach that works best in all situations? Unfortunately not, every approach has strengths and weaknesses, and these are situation specific. Only some pros and cons are noted in the table. See the references for many others.
How Can Data Scientists Improve Causal Reasoning?
Considering multiple approaches to causality will enhance the likelihood that our causal interpretations are correct. If several (or even at least two) valid approaches have been applied and lead to consistent results, and if other well-executed approaches to causality inference do not soon negate these consistencies, then we’re more likely to infer correctly whether an intervention is truly causal.
An example of this notion is provided by Russo and Williamson (2007), who say that combining mechanistic and probabilistic approaches yield stronger insights about causality in healthcare applications. When the mechanism of action for a drug or a vaccine is known, along with the probabilities that its intended effects occur, reasonable inferences about causality are possible.
Multiple causal approaches are especially helpful in situations where well-designed and executed RCTs cannot be conducted. Such RCTs usually combine theory with mechanistic, manipulation, and/or counterfactual approaches. Therefore, RCTs may isolate the effect of an intervention on an outcome of interest. Well-executed RCTs do this by adjusting for many measurable and unmeasurable factors that might otherwise confuse causal interpretations.
There are many situations when RCTs cannot be conducted, though, due to logistical, ethical, behavioral, budgetary, or other reasons. When that is the case, good quasi-experimental or other careful studies can also apply multiple causal inference approaches. Taken together, these will lead to stronger causality statements that we can rest our hats on.
For example, replicating scientific studies by various means in different situations and across species, enhanced by journalistic sleuthing of tobacco company behavior, led to wide acceptance that smoking causes cancer. No RCT could be completed to address that hypothesis in humans because it is not possible or ethical to assign smoking status to people to see who eventually gets ill and who does not.
Another example involves the use of DNA research to acquit people convicted of crimes they did not commit, despite earlier evidence suggesting the opposite. For example, if DNA taken from bodily fluids left at the scene of a murder indicate that the suspect was very unlikely to have committed the crime (because the suspect doesn’t have the same DNA profile as the likely perpetrator), he or she probably did not murder the victim. This would be an example of probabilistic causal inference in action.
Well-conducted probabilistic, mechanistic, counterfactual, or manipulation studies in these and other areas help illuminate causal pathways. They can also provide evidence of the cost of being wrong in earlier investigations.
Limitations and Further Explorations
There are already several useful textbooks and articles about causality by renowned thinkers who publish in academic book houses and peer-reviewed journals. Some of their principles and methods have been described in short articles in Toward Data Science or elsewhere on Medium. There is no need for me to go into details here about their methodologic contributions, nor is it possible to give them the space they deserve in such a short post. Rather, my intent here is to present philosophical views about causal inferences described by Maziarz (2020), and then link those philosophies to some of the principles and methods mentioned by Pearl (2009), Morgan and Winship (2015), and others (e.g., Pearl and Mackenzie (2018); Woodward (2005), and Russo and Williamson (2007)).
A thoughtful combination of philosophy and principles that influence study design and methods can be very valuable for data scientists and other researchers to provide. This will help them cogently describe for their employers, consumers, and policy makers what causes what, what doesn’t, and how best to address vexing business or social issues. I’m just scratching the surface in this endeavor.
As a result, this post leaves a lot of uncovered ground. Examples of other important questions I haven’t addressed include:
How can each of the five methods be applied? What are the appropriate steps for each approach? A thorough review of principles and methods about counterfactuals for use in social science research is provided in the text by Morgan and Winship (2015). They address issues such as whether family background or inherent intelligence is more important in determining how well students do in school, and how much money they eventually earn as adults. Morgan and Winship also address whether socioeconomic status causes differences in levels of health and death rates. Maziarz (2020) reviews the literature and several applications of all five causal inference approaches, focusing on economic policy analyses in many areas.
How can adopting solid conceptual or theoretical frameworks guide causal analyses? Every textbook I’ve read about artificial intelligence, machine learning, and deep learning stresses the value of relying on solid theory and human expertise for guidance. So do Pearl (2009) and Morgan and Winship (2015) as they describe how to use directed acyclic graphs to visually show causal relationships. We live in an uncertain world, but uncertainty does not preclude causal inference if solid theory and methods are applied, repeatedly.
How influential and popular has each of the five approaches been? How much have preferences for one method over others, or a particular combination of methods, helped or hurt the hunt for causal understanding? Maziarz (2020) addresses these questions in economic policy studies, and Morgan and Winship (2015) address these questions for other social sciences. For example, propensity score approaches to match intervention and control subjects, or to give these subjects different weights in the analyses, have exploded in popularity in the last few decades in social science research. Like any other method though, propensity scores can be applied well or poorly, leading either to correct or incorrect causal statements. Morgan and Winship provide detailed examples.
How can readers differentiate between good causal methods vs. those derived from non-scientific conspiracy theories? Pearl’s (2009) text focuses on controversies about smoking and cancer, but the causal ladder method he and his co-author Dana Mackenzie (2018) describe is quite useful to ferret out the falsehoods around many conspiracy theories too. Their causal ladder includes several of the big five approaches to causality described by Maziarz (2020) and summarized in the table above. With reference to COVID-19, other researchers are applying a strong reliance on biochemistry principles, along with clear illustrations of the mechanisms describing how the human body works, to help root out a lot of the misinformation about COVID vaccines.
What is the cost of being wrong about causality? We saw those costs in the form of millions of heart attacks, strokes, disabilities, and deaths that occurred as controversies about smoking were being debated over decades; Pearl (2009) addresses this example well.
We also see the costs of being wrong about causality today, in the form of more cases, long-term disability, and needless deaths as misinformation specialists continue to bandy about baseless views about the supposed non-severity of COVID-19 and the limited utility of masks, social distancing, testing, and vaccines. Unfortunately, the costs of misinformation (i.e., the costs of misusing causal reasoning in arguments about COVID) have been magnified by political fiat as politicians have mistakenly banned or limited the use effective methods to address the pandemic.
Regarding climate change, the costs of being wrong may range from insufficient or misapplied funds to address it, all the way to existential threats as the Earth’s temperature continues to arise unabatedly. Arguments about causality (e.g., did humans cause global warming, and can we really control it?) will continue.
Well-founded causal explanations guided by science can provide a basis for useful discussion about these and many other controversies, but only if we are truly open to funding good research and having respectful and thoughtful discussions in safe environments.
Conclusion
While major advances in causal inference have been made over the last 30 years, especially due to the influence of the authors mentioned above and their colleagues, many more issues must be sorted out. Data scientists who apply solid causality theories, principles, and methods will help us understand how the world really works. Taking advantage of sound causal inference approaches, along with other ways to fight misinformation, will lead to better research, much smarter and more effective public policies, and greater health and wellbeing across our planet.
Acknowledgement
I would like to thank Dr. Mariusz Maziarz for reviewing drafts of this paper and providing many useful comments that clarified my understanding and presentation of causal inference. Any remaining errors are mine.
References
K. Coombs, A. Dube, and C. Jahnke, et al., Early Withdrawal of Pandemic Unemployment Insurance: Effects on Earnings, Employment and Consumption, https://files.michaelstepner.com/pandemicUIexpiration-paper.pdf
M. Maziarz, The Philosophy of Causality in Economics (2020), Routledge — Taylor & Francis Group, New York, NY
S. L. Morgan and C. Winship C, Counterfactuals and Causal Inference: 2nd Edition (2015), Cambridge University Press, Cambridge, UK
J. Pearl, Causality: Models, Reasoning and Inference, 2nd Edition (2009), Cambridge University Press, New York, NY
J. Pearl and D. MacKenzie, The Book of Why: The New Science of Cause and Effect (2018), Basic Books, New York, NY
F. Russo and J. Williamson, Interpreting Causality in the Health Sciences (2007), International Studies in the Philosophy of Science 21(2):157–170
J. Woodward, Making Things Happen: A Theory of Causal Explanation (2005), Oxford University Press, Oxford, England
