Tag: Statistics

Counterfactuals

Counterfactuals

Posted on by jabittan
I have written a lot lately about the incredible human ability to imagine worlds that don’t exist. An important way that we understand the world is by imagining what would happen if we did something that we have not yet done or if we imagine what would have happened had we done something different in the past. We are able to use our experiences about the world and our intuition on causality to imagine a different state of affairs from what currently exists. Innovation, scientific advancements, and social cooperation all depend on our ability to imagine different worlds and intuit causal chains between our current world and the imagined reality we desire.
In The Book of Why Jude Pearl writes, “counterfactuals are an essential part of how humans learn about the world and how our actions affect it. While we can never walk down both the paths that diverge in a wood, in a great many cases we can know, with some degree of confidence, what lies down each.”
A criticism of modern science and statistics is the reliance on randomized controlled trials and the fact that we cannot run an RCT on many of the things we study. We cannot run RCTs on our planet to determine the role of meteor impacts or lightning strikes on the emergence of life. We cannot run RCTs on the toxicity of snake venoms in human subjects. We cannot run RCTs on giving stimulus checks  to Americans during the COVID-19 Pandemic. Due to physical limitations and ethical considerations, RCTs are not always possible. Nevertheless, we can still study the world and use counterfactuals to think about the role of specific interventions.
If we forced ourselves to only accept knowledge based on RCTs then we would not be able to study the areas I mentioned above. We cannot go down both paths in randomized experiments with those choices. We either ethically cannot administer an RCT or we are stuck with the way history played out. We can, however, employ counterfactuals, imagining different worlds in our heads to think about what would have happened had we gone down another path. In this process we might make errors, but we can continually learn and improve our mental models. We can study what did happen, think about what we can observe based on causal structures, and better understand what would have happened had we done something different. This is how much of human progress has moved forward, without RCTs and with counterfactuals, imagining how the world could be different, how people, places, societies, and molecules could have reacted differently with different actions and conditions.
Dose-Response Curves

Dose-Response Curves

Posted on by jabittan
One limitation of linear regression models, explains Judea Pearl in his book The Book of Why is that they are unable to accurately model interactions or relationships that don’t follow linear relationships. This lesson was hammered into my head by a statistics professor at the University of Nevada, Reno when discussing binomial variables. For variables where there are only two possible options, such as yes or no, a linear regression model doesn’t work. When the Challenger Shuttle’s O-ring failed, it was because the team had run a linear regression model to determine a binomial variable, the O-ring fails or it’s integrity holds. However, there are other situations where a linear regression becomes problematic.
 
 
In the book, Pearl writes, “linear models cannot represent dose-response curves that are not straight lines. They cannot represent threshold effects, such as a drug that has increasing effects up to a certain dosage and then no further effect.”
 
 
Linear relationship models become problematic when the effect of a variable is not constant over dosage. In the field of study that I was trained in, political science, this isn’t a big deal. In my field, simply demonstrating that there is a mostly consistent connection between ratings of trust in public institutions and receipt of GI benefits, for example, is usually sufficient. However, in fields like medicine or nuclear physics, it is important to recognize that a linear regression model might be ill suited to the actual reality of the variable.
 
 
A drug that is ineffective at small doses, becomes effective at moderate doses, but quickly becomes deadly at high doses shouldn’t be modeled with a linear regression model. This type of drug is one that the general public needs to be especially careful with, since so many individuals approach medicine with a “if some is good then more is better” mindset. Within physics, as was seen in the Challenger example, the outcomes can also be a matter of life. If a particular rubber for tires holds its strength but fails at a given threshold, if a rubber seal fails at a low temperature, or if a nuclear cooling pool will flash boil at a certain heat, then linear regression models will be inadequate for making predictions about the true nature of variables.
 
 
This is an important thing for us to think about when we consider the way that science is used in general discussion. We should recognize that people assume a linear relationship based on an experimental study, and we should look for binomial variables or potential non-linear relationships when thinking about a study and its conclusions. Improving our thinking about linear regression and dose-response curves can help us be smarter when it comes to things that matter like global pandemics and even more general discussions about what we think the government should or should not do.

Ignorability

Ignorability

Posted on by jabittan
The idea of ignorability helps us in science by playing a role in randomized trials. In the real world, there are too many potential variables to be able to comprehensively predict exactly how a given intervention will play out in every case. We almost always have outliers that have wildly different outcomes compared to what we would have predicted. Quite often some strange factor that could not be controlled or predicted caused the individual case to differ dramatically from the norm.
Thanks to concepts of ignorability, we don’t have to spend too much time worrying about the causal structures that created a single outlier. In The Book of Why Judea Pearl tries his best to provide a definition of ingorability for those who need to assess whether ignorability holds in a given outlier decision. He writes, “the assignment of patients to either treatment or control is ignorable if patients who would have one potential outcome are just as likely to be in the treatment or control group as the patients who would have a different potential outcome.”
What Pearl means is that ignorability applies when there is not a determining factor that makes people with any given outcome more likely to be in a control or treatment group. When people are randomized into control versus treatment, then there is not likely to be a commonality among people in either group that makes them more or less likely to have a given reaction. So a random outlier in one group can be expected to be offset by a random outlier in the other group (not literally a direct opposite, but we shouldn’t see a trend of specific outliers all in either treatment or control).
Ignroability does not apply in situations where there is a self-selection effect for control or treatment. In the world of the COVID-19 Pandemic, this applies in situations like human challenge trials. It is unlikely that people who know they are at risk of bad reactions to a vaccine would self-select into a human challenge trial. This same sort of thing happens with corporate health benefits initiatives, smart phone beta-testers, and general inadvertent errors in scientific studies. Outliers may not be outliers we can ignore if there is a self-selection effect, and the outcomes that we observe may reflect something other than what we are studying, meaning that we can’t apply ignorability in a way that allows us to draw a conclusion specifically on our intervention.
Alternative, Nonexistent Worlds - Judea Pearl - The Book of Why - Joe Abittan

Alternative, Nonexistent Worlds

Posted on by jabittan
Judea Pearl’s The Book of Why hinges on a unique ability that human animals have. Our ability to imagine alternative, nonexistent worlds is what has set us on new pathways and allowed us to dominate the planet. We can think of what would happen if we acted in a certain manner, used a tool in a new way, or if two objects collided together. We can visualize future outcomes of our actions and of the actions of other bodies and predict what can be done to create desired future outcomes.
In the book he writes, “our ability to conceive of alternative, nonexistent worlds separated us from our protohuman ancestors and indeed from any other creature on the planet. Every other creature can see what is. Our gift, which may sometimes be a curse, is that we can see what might have been.”
Pearl argues that our ability to see different possibilities, to imagine new worlds, and to be able to predict actions and behaviors that would realize that imagined world is not something we should ignore. He argues that this ability allows us to move beyond correlations, beyond statistical regressions, and into a world where our causal thinking helps drive our advancement toward the worlds we want.
It is important to note that he is not advocating for holding a belief and setting out to prove it with data and science, but rather than we use data and science combined with our ability to think causally to better understand the world. We do not have to be stuck in a state where we understand statistical techniques but deny plausible causal pathways. We can identify and define causal pathways, even if we cannot fully define causal mechanisms. Our ability to reason through alternative, nonexistent worlds is what allows us to think causally and apply this causal reasoning to statistical relationships. Doing so, Pearl argues, will save lives, help propel technological innovation, and will push science to new frontiers to improve life on our planet.
Regression Coefficients

Regression Coefficients

Posted on by jabittan
Statistical regression is a great thing. We can generate a scatter plot, generate a line of best fit, and measure how well that line describes the relationship between the individual points within the data. The better the line fits (the more that individual points stick close to the line) the better the line describes the relationships and trends in our data. However, this doesn’t mean that the regression coefficients tell us anything about causality. It is tempting to say that a causal relationship exists when we see a trend line with lots of tight fitting dots around and two different variables on an X and Y axis, but this can be misleading.
In The Book of Why Judea Pearl writes, “Regression coefficients, whether adjusted or not, are only statistical trends, conveying no causal information in themselves.” It is easy to forget this, even if you have had a statistics class and know that correlation does not imply causation. Humans are pattern recognition machines, but we go a step beyond simply recognizing a pattern, we instantly set about trying to understand what is causing the pattern. However, our regression coefficients and scatter plots don’t always hold clear causal information. Quite often there is a third hidden variable that cannot be measured directly that is influencing the relationship we discover in our regression coefficients.
Pearl continues, “sometimes a regression coefficient represents a causal effect, and sometimes it does not – and you can’t rely on the data alone to tell you the difference.” Imagine a graph with a regression line running through a plot of force applied by a hydraulic press and fracture rates for ceramic mugs. One axis may be pressure, and the other axis may be thickness of the ceramic mug. The individual points represent the point at which individual mugs fractured We would be able to generate a regression line by testing the fracture strength of mugs of different thickness, and from this line we would be able to develop pretty solid causal inferences about thickness and fracture rates. A clear causal link could be identified by the regression coefficients in this scenario.
However, we could also imagine a graph that plotted murder rates in European cities and the spread of Christianity. With one axis being the number of years a city has had a Catholic bishop and the other axis being the number of murders, we may find that murders decrease the longer a city has had a bishop.  From this, we might be tempted to say that Christianity (particularly the location of a Bishop in a town) reduces murder. But what would we point to as the causal mechanism? Would it be religious beliefs adopted by people interacting with the church? Would it be that marriage rules that limited polygamy ensured more men found wives and became less murderous as a result? Would it be that some divinity smiled upon the praying people and made them to be less murderous? A regression like the one I described above wouldn’t tell us anything about the causal mechanism in effect in this instance. Our causal-thinking minds, however, would still generate causal hypothesis, some of which would be reasonable but others less so (this example comes from the wonderful The WEIRDest People in the World by Joseph Henrich).
Regression coefficients can be helpful, but they are less helpful when we cannot understand the causal mechanisms at play. Understanding the causal mechanisms can help us better understand the relationship represented by the regression coefficients, but the coefficient itself only represents a relationship, not a causal structure. Approaching data and looking for trends doesn’t help us generate useful information. We must first have a sense of a potential causal mechanism, then examine the data to see if our proposed causal mechanism has support or not. This is how we can use data and find support for causal hypothesis within regression coefficients.
Laboratory Proof

Laboratory Proof

Posted on by jabittan
“If the standard of laboratory proof had been applied to scurvy,” writes Judea Pearl in The Book of Why, “then sailors would have continued dying right up until the 1930’s, because until the discovery of vitamin C, there was no laboratory proof that citrus fruits prevented scurvy.” Pearl’s quote shows that high scientific standards for definitive and exact causality are not always for the greater good. Sometimes modern science will spurn clear statistical relationships and evidence because statistical relationships alone cannot be counted on as concrete causal evidence. A clear answer will not be given because some marginal unknowns may still exist, and this can have its own costs.
Sailors did not know why or how citrus fruits prevented scurvy, but observations demonstrated that citrus fruits managed to prevent scurvy. There was no clear understanding of what scurvy was or why citrus fruits were helpful, but it was commonly understood that a causal relationship existed. People acted on these observations and lives were saved.
On two episodes, the Don’t Panic Geocast has talked about journal articles in the British Medical Journal that make the same point as Pearl. As a critique of the need for randomized controlled trials, the two journal articles highlight the troubling reality that there have not been any randomized controlled trials on the effectiveness of parachute usage when jumping from airplanes. The articles are hilarious and clearly satirical, but ultimately come to the same point that Pearl does with the quote above – laboratory proof is not always necessary, practical, or reasonable when lives are on the line.
Pearl argues that we can rely on our abilities to identify causality even without laboratory proof when we have sufficient statistical analysis and understanding of relationships. Statisticians always tell us that correlation is not causation and that observational studies are not sufficient to determine causality, yet the citrus fruit and parachute examples highlight that this mindset is not always appropriate. Sometimes more realistic and common sense understanding of causation – even if supported with just correlational relationships and statistics – are more important than laboratory proof.
Slope is Agnostic to Cause and Effect

Slope is Agnostic to Cause and Effect

Posted on by jabittan
I like statistics. I like to think statistically, to recognize that there is a percent chance of one outcome that can be influenced by other factors. I enjoy looking at best fit lines, seeing that there are correlations between different variables, and seeing how trend-lines change if you control for different variables. However, statistics and trend lines don’t actually tell us anything about causality.
In The Book of Why Judea Pearl writes, “the slope (after scaling) is the same no matter whether you plot X against Y or Y against X. In other words, the slope is completely agnostic as to cause and effect. One variable could cause the other, or they could both be effects of a third cause; for the purpose of prediction, it does not matter.”
In statistics we all know that correlation is not causation, but this quote helps us remember important information when we see a statistical analysis and a plot with linear regression line running through it. The regression line is like the owl that Pearl had described earlier in the book. The owl is able to predict where a mouse is likely to be and able to predict which direction it will run, but the owl does not seem to know why a mouse is likely to be in a given location or why it is likely to run in one direction over another. It simply knows from experience and observation what a mouse is likely to do.
The regression line is a best fit for numerous observations, but it doesn’t tell us whether one variable causes another or whether both are influenced in a similar manner by another variable. The regression line knows where the mouse might be and where it might run, but it doesn’t know why.
In statistics courses we end at this point of correlation. We might look for other variables that are correlated or try to control for third variables to see if the relationship remains, but we never answer the question of causality, we never get to the why. Pearl thinks this is a limitation we do not need to put on ourselves. Humans, unlike owls, can understand causality, we can recognize the various reasons why a mouse might be hiding under a bush, and why it may chose to run in one direction rather than another. Correlations can help us start to see where relationships exist, but it is the ability of our mind to understand causal pathways that helps us determine causation.
Pearl argues that statisticians avoid these causal arguments out of caution, but that it only ends up creating more problems down the line. Important statistical research in areas of high interest or concern to law-makers, business people, or the general public are carried beyond the cautious bounds that causality-averse statisticians place on their work. Showing correlations without making an effort to understand the causality behind it makes scientific work vulnerable to the epistemically malevolent who would like to use correlations to their own ends. While statisticians rigorously train themselves to understand that correlation is not causation, the general public and those struck with motivated reasoning don’t hold themselves to the same standard. Leaving statistical analysis at the level of correlation means that others can attribute the cause and effect of their choice to the data, and the proposed causal pathways can be wildly inaccurate and even dangerous. Pearl suggests that statisticians and researchers are thus obligated to do more with causal structures, to round off  their work and better develop ideas of causation that can be defended once their work is beyond the world of academic journals.
The Fundamental Nature of Cause and Effect

The Fundamental Nature of Cause and Effect

Posted on by jabittan
In my undergraduate and graduate studies I had a few statistics classes and I remember the challenge of learning probability. Probability, odds, and statistics are not always easy to understand and interpret. There are some concepts that are pretty straightforward, and others that seem to contradict what we would expect if we had not gone through the math and if we had not studied the concepts in depth. To contrast the difficult and sometimes counter-intuitive nature of statistics, we can think about causality, which is a challenging concept, but unlike statistics, is something we are able to intuit from very young age.
In The Book of Why Judea Pearl writes, “In both a cognitive and a philosophical sense, the idea of cause and effect is much more fundamental than probability. We begin learning causes and effects before we understand language and before we understand mathematics.”
As Pearl explains, we see causality naturally and experience causality as we move through our lives. From a young child who learns that if they cry they receive attention to a nuclear physicist who learns what happens when two atoms collide at high energy levels, our minds are constantly looking at the world and looking for causes. It begins by making observations of phenomena around us and continues as we predict what outcomes would happen based on certain system inputs. Eventually, our minds reach a point where we can understand why our predictions are accurate or inaccurate, and we can imagine new ways to bring about certain outcomes. Even if we cannot explain all of this, we can still understand causation at a fundamental and intuitive level.
However, many of us deny that we can see and understand the world in a causal way. I am personally guilty of thinking in a purely statistical way and ignoring the causal. The classes I took in college helped me understand statistics and probability, but also told me not to trust my intuitive causal thinking. Books like Kahneman’s Thinking Fast and Slow cemented this mindset for me. Rationality, we believe, requires that we think statistically and discount our intuitions for fear of bias. Modern science says we can only trust evidence when it is backed by randomized controlled trials and directs us to think of the world through correlations and statistical relationships, not through a lens of causality.
Pearl pushes back against this notion. By arguing that causality is fundamental to the human mind, he implies that our causal reasoning can and should be trusted. Throughout the book he demonstrates that a purely statistical way of thinking leaves us falling short of the knowledge we really need to improve the world. He demonstrates that complex tactics to remove variables from equations in statistical methods are often unnecessary, and that we can accept the results of experiments and interventions even when they are not fully randomized controlled trials.  For much of human history our causal thinking nature has lead us astray, but I think that Pearl argues that we have overcorrected in modern statistics and science, and that we need to return to our causal roots to move forward and solve problems that statistics tells us are impossible to solve.
Predictions & Explanations

Predictions & Explanations

Posted on by jabittan
The human mind has incredible predictive abilities, but our explanatory abilities do not always turn out to be as equally incredible. Prediction is relatively easy when compared to explanation. Animals can predict where a food source will be without being able to explain how it got there. For most of human history our ancestors were able to predict that the sun would rise the next day without having any way of explaining why it would rise. Computer programs today can predict our next move in chess but few can explain their prediction or why we would make the choice that was predicted.
As Judea Pearl writes in The Book of Why, “Good predictions need not have good explanations. The owl can be a good hunter without understanding why the rat always goes from point A to point B.” Prediction is possible with statistics and good observations. With a large enough database, we can make a prediction about what percentage of individuals will have negative reactions to medications, we can predict when a traffic jam will occur, and we can predict how an animal will behave. What is harder, according to Pearl, is moving to the stage where we describe why we observe the relationships that statistics reveal.
Statistics alone cannot tell us why particular patterns emerge. Statistics cannot identify causal structures. As a result, we continually tell ourselves that correlation is not causation and that we can only determine what relationships are truly causal through randomized controlled trials. Pearl would argue that this is incorrect, and he would argue that this idea results from the fact that statistics is trying to answer a completely different question than causation. Approaching statistical questions from a causal lens may lead to inaccurate interpretations of data or “p-hacking” an academic term used to describe efforts to get the statistical results you wanted to see. The key is not hunting for causation within statistics, but understanding causation and supporting it through evidence uncovered via statistics.
Seeing the difference between causation and statistics is helpful when thinking about the world. Being stuck without a way to see and determine causation leads to situations like tobacco companies claiming that cigarettes don’t cause cancer or oil and gas companies claiming that humans don’t contribute to global warming. Causal thinking, however, utilizes our ability to develop explanations and applies those explanations to the world. Our ability to predict different outcomes based on different interventions helps us interpret and understand the data that the world produces. We may not see the exact picture in the data, but we can understand it and use it to help us make better decisions that will lead to more accurate causal understandings over time.
Hope in Big Data

Hope in Big Data

Posted on by jabittan
Most of us probably don’t work with huge data sets, but all of us contribute to huge data sets. We know the world of big data is out there, and we know people are working with big data, but there are not many of us who truly know what it means and how we should think about any of it. In The Book of Why, Judea Pearl argues that even many of those doing research and running companies based on big data don’t fully understand what it all means.
Pearl is critical of researchers and entrepreneurs who lack causal understandings but pursue new knowledge and information by pulling correlations and statistics out of large data sets. There are some companies that are taking advantage of the fact that huge amounts of computing power can give us insights into data sets that we never before could have generated, however, these insights are not always as meaningful as we are lead to believe.
Pearl writes, “The hope – and at present, it is usually a silent one – is that the data themselves will guide us to the right answers whenever causal questions come up.”
My last post was about the overuse of the phrase: correlation is not causation. Finding correlations and relationships in data is meaningless if we don’t also have causal understandings in mind. This is the critique that Pearl makes with the quote above. If we don’t have a way of understanding basic causal structures, then the phrase is right, correlations don’t mean anything. Many companies and researchers are in a stage where they are finding correlations and unexpected statistical results in big data, but they lack causal understandings to do anything meaningful with the data. In the world of public policy this feels like the saying, a solution in search of a problem or in the world of healthcare like a pay and chase scenario.
Pearl argues throughout the book that we are better at identifying causal structures than we are lead to believe in our statistics courses. He also argues that understanding causality is key to unlocking the potential of big data and actually getting something useful out of massive datasets. Without a grounding in causality, we are wasting our time with the statistical research we do. We are running around with solutions in the forms of big data correlations that don’t have a causal underpinning. It is as if we are paying fraudulent claims, then chasing down some of the money we spent and congratulating ourselves on preventing fraud. The end result is a poor use of data that we prop up as a magnanimous solution.