The Illusion of Causality in Charts
filwd.substack.com
filwd.substack.com
The underlying issue (which the article discusses to some extent) is how confounding factors can make the data misleading/allow the data to be misinterpreted.
To discuss "The Illusion of Causality in Charts" I'd want to consider how one chart type vs. another is more susceptible to misinterpretation/more misleading than another. I don't know if that's actually true -- I haven't worked up some examples to check -- but that's what I was hoping for here.
Even leaving out the data (which you rightly point out) you are forced to choose what to plot on x and y, which by convention will communicate IV and DV respectively whether you like it or not.
In any case, what is interesting is that different charts have different ways of implying causality. This is what I try to do towards the end of the article.
But seriously, it's a nicely-put-together article in several ways, despite my criticism.
Graphic discovery, visual revelations etc.
I think the issues described in this piece, and by other comments, are going to get much worse with the (dis)information overload AI can provide. "Hey AI, plot thing I don't like A with bad outcome B, and scale the axes so they look heavily correlated". Then it's picked up on social media, a clout-chasing public official sees it, and now it's used to make policy.
Sometimes you are choosing the narrative consciously (I created this chart to tell a story), and sometimes you are choosing it unconsciously (I just want to scatter plot and see what it shows - but you chose the x and y to plot, and you chose the scatter plot vs some other framework), and sometimes it is chosen for you (chart defaults for example, or north is up on a map).
And it’s not just charts. Statistics on the whole exist to organize raw data. The very act of introducing organization means you have a scheme, framework, lens which with to do so. You have to accept that and become conscious of that.
You cannot do anything as simple as report an average without choosing which data to include and which type of average to use. Or a histogram without choosing the bin sizes, and again, the data to include.
This is all to say nothing of the way the data was produced in the first place. (Separate topic)
1. In general, humans are not trained to be skeptical of data visualizations.
2. Humans are hard-wired to find and act on patterns, illusory or not, at great expense.
Incidentally, I've found that avoiding the words "causes," "causality," and "causation" is almost always the right path or at the least should be the rule as opposed to the exception. In my experience, they rarely clarify and are almost always overreach.
It's a fundamental problem of reality.
The nature of reality itself prevents us from determining causality from observation, this includes looking at a chart.
If you observe two variables. Whether those random variables correlate or not... there is NO way to determine if one variable is causative to another through observation alone. Any causation in a conclusion from observation alone is in actuality only assumed. Note the key phrase here is: "through observation alone."
In order to determine if one thing "causes" another thing, you have to insert yourself into the experiment. It needs to go beyond observation.
The experimenter needs to turn off the cause and turn on the cause in a random pattern and see whether that changes the correlation. Only through this can one determine causation. If you don't agree with this, think about it a bit.
Also note that this is how they approve and validate medicine... they have to prove that the medicine/procedure "causes" a better outcome and the only way to do this is to actually make giving and withholding the medicine as part of the trial.
"What does it mean that something is caused by something else?" At the end of it all, what matters is how it's used in the real world. Personally I find the philosophical discussion to be tiresome.
In law, "to cause" is pretty strict: "but for" A, B would not exist or have happened. Therefore A caused B. That's one version. Other people and regimes have theirs.
This is why it's something I try to avoid.
In any case, descriptions of distributions are more comprehensive and avoid conclusions.
It is literally the basis for medicine. We literally have to have a "hand in the experiment" for clinical trials to with-hold medicine and to give medicine in order to establish that medicine "causes" a "cure". Clinical trials are by design not about just observation.
Likely, you just don't understand what I was saying.
The criteria or definition for " A causes B" that you alluded to is a useful one in the context of medicine:
> The experimenter needs to turn off the cause and turn on the cause in a random pattern and see whether that changes the correlation. Only through this can one determine causation. If you don't agree with this, think about it a bit.
It's useful because it establishes a threshold we can use and act on in the real world.
I think there is more nuance and context here though. In clinical trials, minimum cohort sizes are required, possibly related or proportional to power analysis (turning on and off the cause for one person doesn't give us much confidence but for 1000 people gives much more).
So the definition of causes for clinical trials and medicine hinges on more than just turning on and off, it relies on effect size and population size in the experiment.
Going back to TFA, this is the problem when we bring "cause" into the discussion: the definition of it varies depending on the context.
Of course. Because the clinical trial is statistical so the basis of the trial is trying to come to a conclusion about a population of people via a sample. That fact applies to both correlation or causation. Statistics is like an extension from person to people… rather then coming to a conclusion about something for one person you can do it for a sample of the population.
Causality is independent of the extension. You can measure causality against a sample of a population or even a single thing. The property of inserting yourself into an experiment still exists in both cases. This is basic and a simple thought experiment can determine this.
You have two switches two lights and two people. Both people turn each of their respective switches on and off and the light turns on and off in the expected pattern exactly like the state of the switch.
You know one of the switches is hooked to the light and “causes” the light to turn on and off. The other switch is BS and is turning on and off on some predetermined pattern and the person that’s flipping the related switch memorized the pattern and is making it look like the switch is causative to turning on or off the light.
How do you determine which one is the switch that is causative to the light turning on and off and which switch isn’t?
Can you do it through observation alone? Can you just watch the people flip the switch? Or do you have to insert yourself into the experiment and flip both switches randomly yourself to see which one is causal to the light turning on or off?
The answer is obvious. I’m sort of anticipating a pedantic response where you just “observe” the wiring of the switch and the light to that I would say I’m obviously not talking about that. You can assume all the wiring is identical and the actual mechanism is a perfect black box. We never actually try to determine causality or correlation unless we are dealing with a black or semi black box so please do not go down that pedantic road.
You should walk away from this conversation with new intuition on how reality works.
You shouldn’t be getting to involved in mathematical definitions and details of what involves a clinical trial or pedantic details and formal definitions.
Gain deep understanding of why causality must be determined this way and then that helps you see why the specific detailed protocols of clinical trials were designed that way in the first place.
(I say this coming from an engineering context, where e.g. you can pretty confidently say that your sensor isn't affecting the weather but vice-versa is plausible)
In practice it’s hard to determine causality so people make assumptions. Most conclusions are like that. I said this in the original post that conclusions from observation alone must have assumptions made. Which is fine given available resources. If you find people who smoke weed have lower iq you can come to the conclusion that weed causes iq to lower assuming that all smokers of weed had average iq before smoking and this is fine.
I’m sure you’ve seen many causative conclusions redacted because of incorrect assumptions so it is in general a very unreliable method.
And that’s why in medicine they strictly have to do causative based testing because they can’t afford to have a conclusion based off of an incorrect assumption.
The shape of it is that there is a statistic about population and then that statistic is used to describe a member of that population. For example, a news story that starts with “70% of restaurants fail in their first year, so it’s surprising that new restaurant Pete’s Pizza is opening their third location!”
But it’s only surprising if you know absolutely nothing about Pete and his business. Pete’s a smart guy. He’s running without debt and has community and government ties. His aunt ran a pizza business and gave him her recipes.
In a Bayesian way of thinking, the newscasters statement only makes sense if the only prior they have is the average success rate of restaurants. But that is an admittance that they know nothing about the actual specifics of the current situation, or the person they are talking about. Additionally there is zero causal relationship between group statistics and individual outcomes, the causal relationship goes the other way. Pete’s success will slightly change that 70% metric, but the 70% metric never bound Pete to be “likely to fail”.
Other places I see the “bound by statistics” problem is in healthcare, criminal proceedings, racist rhetoric, and identity politics.
It doesn’t mean the statistics are wrong, though. If there is a 70% chance of failure, there’s also a 30% chance of success. But it’s subjective: use a different reference class and you’ll get a different number.
The opposite problem is also common: assuming that “this time it’s different” without considering the reasons why others have failed.
The general issue is overconfidence and failure to consider alternative scenarios. The future isn’t known to us and it’s easy to fool yourself into thinking it is known.
[1] https://en.m.wikipedia.org/wiki/Reference_class_forecasting
Lots of them opened, of them 70% failed, and one who didn’t happened to be named Pete.
No more interesting than “Pete rolled a 4!” even though 83% of people don’t.
I don’t think it’s valid to define “surprising” in such a self-referential way. When something unusual appears in the news, that doesn’t make it common. The probabilities are different before and after applying a filter.
More precisely, it's not surprising that one exists. It may be surprising that this particular one survived, just as it wouldbe surprising that it's my neighbor who wins the lottery next week. But it's likely that somebody will, so if somebody has won, it won't be a surprise that somebody did.
It’s a question of how many other possibilities are considered similar to the one that happened. From a zoomed-out perspective, one win is as good as any another.
> When something unusual appears in the news, that doesn’t make it common.
This in particular isn’t even close to what I said, which was: rare events can be unsurprising in large datasets — as is the case with both dice rolls and restaurants succeeding.
I guess this is just point of view. "Someone won the lottery" and "I won the lottery" describe the same event from very different perspectives.
[1] https://apps.apple.com/us/app/reflect-track-anything/id64638...
we do have a pretty good intuition for it but if you look at the details and ask people what is the difference between correlation and causality and how do you distinguish it things get rabbit holey pretty quick