Correlation and causation help us understand why something happens. Correlation is simply something that happens while something else is happening. A correlation could indicate causation, but it does not always. Causation is the something that causes something else to happen. The "why" may seem straightforward. It is actually more challenging than you may think. Our brains are naturally wired to seek causation. It is understandable that we may confuse the easier-to-see correlations with causation. Together, we will dig into the nuances creating causality challenges. In the concluding section, we explore cognitive bias challenges and tools to make great decisions in the service of causality. Choice architecture makes causally-consistent decisions easier to achieve. Definitive Choice is an easy-to-use app to help you make the best decisions.
About the author: Jeff Hulett is a career banker, data scientist, and choice architect. Jeff has held banking and consulting leadership roles at Wells Fargo, Citibank, KPMG, and IBM. Today, Jeff is an executive with the Definitive Companies. He teaches personal finance at James Madison University and provides personal finance seminars. Chec out his book -- Making Choices, Making Money: Your Guide to Making Confident Financial Decisions -- at jeffhulett.com.
The earlier pushing the pickup example seems obvious. If you are in the bed of the pickup, no matter how hard you push, clearly, you are not causing the pickup to move. The power causing the pickup to move is coming from the person on the ground behind the truck. But what if there was something you could not easily see about the pickup's environment? What if the picture of the pickup was taken parallel to the road? Then, as you gain more perspective, you realize the road is at a steep grade.
A new take on causality - a better perspective
Now, given our improved perspective, even the person behind the truck is not doing much to cause the truck to move. Gravity is causing the truck to move. Causality is not always obvious -- it takes perspective and investigation. But if it is challenging for people to parse causality, what about Artificial Intelligence? A.I. are the machines that mimic some forms of human intelligence.
In general, A.I. is particularly well-suited for determining data correlations. However, artificial intelligence struggles with telling the difference between correlation and causation. In fact, in many cases, A.I. leaves causality determination to its human programmers. A good example of the A.I. correlation-causation challenge is from meteorology. An A.I. may consider a simple two-column dataset, with observed weather (rainy or sunny) and barometric pressure. The A.I. could easily and accurately regress one data column on the other with the outcome suggesting a significant correlation.
The A.I. will say "Yup, there is a correlation between barometric pressure and observed weather. Sunny days occur when the barometric pressure is higher. Cloudy or rainy days happen when the barometric pressure is lower." But what is the cause? The A.I. will suggest "Hey, if you want sunny weather, all you have to do is increase the barometric pressure!" Barometric pressure clearly is not a dial we can turn to cause better weather! Naturally, we all know that barometric pressure is only a correlated indication of the weather. But the A.I. will not be able to judge the difference between causal and correlated data relationships.
Causality is not as important for marketing some products - A Netflix example.
A.I. is used on big consumer platforms like Amazon or Netflix. For example, on Netflix, the A.I. uses next-show models. These are models that consider your past viewing behavior and the viewing behavior of people like you to predict what movie or show you are most likely to watch next. Now, did the fact that you watched "Stranger Things" in the past cause you to watch "Bridgerton" next? No, of course not. But there are correlations between your and others' past show behavior that the A.I. is using to nudge you for your next Netflix show. No one likes a show hole! Next-show models, while sophisticated in their use of big data, are only good at first-step correlations. Next-show models are generally not able to reach a second-step casualty. We discuss correlation and causality steps below.
Causality is very important for marketing other products - A consumer credit example.
In today's more automated consumer credit industry, lenders use statistically modeled credit assessment scores to recommend credit decisions. The more famous model is the FICO score, but there are others. Consumer product lenders, like mortgage, auto, credit card, and others, have regulatory requirements to clearly disclose why a loan is declined. Under the Fair Credit Reporting Act and the Equal Credit Opportunity Act, a lender is required to give a declined loan applicant an adverse action notice as to why a loan was declined. As such, it is difficult to use opaque, causally challenged A.I., like unsupervised Machine Learning techniques, because of the difficulty in providing a decline reason. Highly transparent traditional models, like the FICO score, are often used because of decline reason explainability needs. Decline reasons are lifted directly from the traditional model scorecard.
But even in the case of lending, the adverse action regulation only resolves part of the causality puzzle. What caused a loan applicant to have poor payment behavior is a bigger societal question. Systemically biased data - also known as structurally biased data - is resident in the credit modeling data provided by the Credit Reporting Agencies. [i] In terms of the cause of poor payment behavior, systemic bias enables the causality arrow to point in both directions:
The reasoning for 2-way causality enabled by the credit data is straightforward:
"...Since the data used for credit algorithms is from our recent past and our recent past contains racism, then it follows the data itself may be systemically biased by race or other protected classes."
However, the law only requires the lender to disclose why they declined the loan [P] --> [C], not why the loan applicant was having payment difficulty [P] <-- [C]. There is a big difference... and society is starting to wake up to that difference! [ii]
Generally, the best way to establish the degree to which characteristics may be considered causal is via Randomized Control Trial ("RCT") testing. This is the gold standard for determining causality. The RCT is a test in which subjects are randomly assigned to one of two groups: one (the experimental group) receiving the intervention that is being tested, and the other (the comparison group or control) receiving an alternative (conventional) treatment. In the social media world, A/B testing can be used as RCTs. Increasingly, natural experiments are being utilized as valid RCTs. These are experiments where social science test and control groups are available "in the wild" because of naturally occurring group conditions creating observable differences.
A Natural Experiment Example: In Helena, Montana a smoking ban was in effect in all public spaces, including bars and restaurants, during the six-month period from June 2002 to December 2002. Helena is geographically isolated and served by only one hospital. The investigators observed that the rate of heart attacks dropped by 40% while the smoking ban was in effect. Opponents of the law prevailed in getting the enforcement of the law suspended after six months, after which the rate of heart attacks went back up. [iii] This study was an example of a natural experiment, called a case-crossover experiment, where the exposure is removed for a time and then returned.
Judea Pearl, in The Book Of Why does a nice job of describing the importance of correlation and causation in terms of a ladder or stairs. Causality needs to get to at least the second step of the causation stairs, whereas correlation is only at the first step. Notice, we have to climb the “see” and “do” steps before we get to the “imagine” step on the stairs.
Finally, the statistical truism is that "correlation does not imply causation." The opposite may be true as well. Causation could exist when correlation does not. [iv] More often than not, this is NOT the case -- but it does happen.
Causation can occur without correlation when a lack of change in the variables is present.
Lack of change in variables occurs most often with insufficient samples. In the most basic example, if we have a sample of 1, we have no correlation, because there’s no other data point to compare against. There’s no correlation. If I hit a glass with a hammer once, we have a clear, obvious causative effect, but because I did it once, there’s no correlation because there’s no other variable to compare it against.
Causation can occur without correlation in mined data subsets.
For example, we know there’s a causative effect between alcohol consumption and automotive fatalities. Drinking and driving – or operating a vehicle under the impairing influence of any substance – leads to fatalities. In a normal dataset, if we compared the number of drinks consumed per day and vehicular fatality outcome, we’d see a clear correlation.
However, what if we restricted that dataset to people who consumed 10 or more daily drinks? Even though we have a known causative relationship, we might not see a correlation because the chances of dying from all kinds of outcomes due to alcoholism interfere with the correlation. That much drinking per day will kill you for any number of reasons.
Conclusion
We need to look inside our own brains to understand why causality and correlation are easy to confuse. The answer, at its core, comes down to our neurobiology. Our brain’s natural decision default setting is called the fast brain. Fast-brained decisions are an evolutionary-tuned, fight-or-flight sort of decision. The good news is they are fast and they help us make accurate lifesaving decisions like - "Should I run from this lion!?" The bad news is, unless there is a jailbreak from the local zoo, we rarely make these evolutionary-tuned decisions anymore. In today’s world, the vast, vast majority of decisions we make are complex, slow-brain decisions. Our brains are NOT yet tuned to make more complex decisions quickly. Perhaps, someday evolution will catch up. The psychological difference between the fast-brain and the slow-brain type of decision presents as one or many cognitive biases. The foundational cognitive bias is confirmation bias. [v]
In the causality and correlation context, our natural response is to rely upon fast, easy-to-see correlations and associate them with causality. This is also known as representativeness bias. This is often just wrong. Confirmation bias is all around us, it can be seen in many other decision contexts. Earlier, several examples of using RCT to help make cause-related decisions were provided. Additionally, great decisions can occur by using choice architecture to help you make those decisions. Choice architecture, like Definitive Choice, separates your cognitive biases from your valuable judgment. Choice architecture helps you make the best decisions in the service of separating causality from correlation.
Resources
To achieve fast, confident, and transparent decisions, choice architecture is an essential companion. Choice architecture is decision science and behavioral science-informed tools. Choice architecture will help you curate data. Choice architecture will help you overcome your naturally occurring cognitive biases leading to not fully understanding complex and dynamic self-interests, preferences, and utility.
Definitive Choice is an app decision solution to help you understand your self-interests and action on almost all life decisions. It provides a straightforward user experience. The number-crunching occurs in the background by time-tested decision science algorithms. It uses a proprietary "Decision 6(tm)" approach that organizes the preference criteria (what is important to you?) and alternatives (what are the choices?) in a series of bite-size ranking decisions. Since it is on your smartphone, you can use it while you are curating data to support the decision. It is like having a decision expert in your pocket. The results dashboard provides a rank-ordered list of recommended "best choices," tailored to your preferences.
Also, Definitive Choice comes pre-loaded with many decision templates. You will want to customize your own preferences (aka criteria) and alternatives, but the preloaded templates provide a nice starting point.
Using decision process solutions enables DECISION A-C-T:
Accelerated: faster, less costly decisions. It enables a nimble decision environment.
Confidence-inspired: process causes people to be more confident in the decision, increasing buy-in, and decision up-take.
Transparency-enabled: reporting, documentation, and charts to help communicate the decision.
Notes
[i] To be clear, it is not that credit bureau data is wrong, the CB data is incomplete. It underrepresents certain groups of people.
Andrews, How Flawed Data Aggravates Inequality in Credit, Stanford University Human Centered Artificial Intelligence, 2021
[ii] Hulett, Resolving Lending Bias - a proposal to improve credit decisions with more accurate credit data, The Curiosity Vine, 2021
[iii] Sargent, Shepard, Glantz, Reduced incidence of admissions for myocardial infarction associated with public smoking ban: before and after study, British Medical Journal, 2004
[iv] Thanks to Chris Penn for the causation without causality examples.
Penn, Can Causation Exist Without Correlation? Yes!, Awaken Your Superhero, 2018
[v] Hulett, Great decision-making and how confidence changes the game, The Curiosity Vine, 2022