Do you believe in parallel universes? Every decision creates ripples across infinite realities. What if I told you that causal inference, which is a rigorous and flourishing scientific methodology, could help you glimpse into these hidden worlds?
This article will introduce you to causal inference, a fascinating field that goes beyond traditional correlation analysis. We’ll explore why determining causation is so challenging, when A/B testing becomes impossible, and how causal inference methods can unlock insights from observational data. By the end, you’ll understand how data science can construct “parallel worlds” to answer the eternal question: “What would have happened if…?”
The Correlation vs. Causation Dilemma#
Every data science student has heard this warning.
Even when we observe strong correlations between variables, without establishing causation, we’re left wondering what action to take next. Statistical models and machine learning algorithms, no matter how sophisticated, only describe correlational relationships, NOT causal mechanisms.
I always love the hilarious examples from Spurious Correlations: the number of drowning deaths correlates remarkably with Nicolas Cage movie releases. Without establishing causation, suggesting “Nicolas Cage should stop making movies to reduce drowning deaths” would be absurd.
For business decisions, causal relationships are equally critical. We constantly seek to understand whether “doing X” will improve user satisfaction, making causal inference essential for extracting maximum value from our data.
However, analyzing causal relationships presents formidable challenges.
Three Critical Challenges in Causal Analysis#
Causal analysis faces numerous obstacles. Here are three fundamental challenges that make determining causation so difficult:
Confounding Variables#
Confounding bias occurs when a variable influences both who is selected for the treatment and the outcome of the experiment.
— Judea Pearl, The Book of Why
Confounding variables simultaneously influence both variables of interest, leading researchers to misinterpret relationships.
Imagine a beachside ice cream vendor observing a strong positive correlation between “daily ice cream sales” and “daily drowning incidents.” This doesn’t mean drownings cause ice cream purchases. The confounding variable here is “sunny weather”:
- Sunny days → More beachgoers → Both higher ice cream sales AND more drowning incidents
- Rainy days → Fewer beachgoers → Both lower ice cream sales AND fewer drowning incidents
The trickiest aspect? Some confounding variables remain unobserved. Even if we know weather affects both variables, if no one recorded weather data, we cannot control for its influence statistically.
Self-Selection Bias#
Self-selection bias occurs when individuals freely choose whether to participate in a treatment group, creating systematic differences between groups.
Pre-election polling provides a perfect example. When pollsters call random numbers asking about voting intentions, respondents can choose to continue or hang up. Those who engage are typically more politically active, while those who immediately hang up tend to be politically disinterested. If a poll shows 80% plan to vote, this dramatically overestimates actual turnout due to self-selection bias.
In product development, imagine launching a new red theme option for your note-taking app. Users who switched to red spend significantly more time in the app daily. Can we claim that the red theme is a winner feature? Should you change everyone’s theme to red? Not necessarily! Users who actively explore app settings likely were already power users. The act of choosing red didn’t cause increased usage. Instead, it revealed pre-existing engagement patterns.
The Counterfactual Problem#
Counterfactuals represent scenarios contrary to observed reality:
- If COVID-19 hadn’t emerged in 2020, how many concerts would have occurred that winter?
- When Elon Musk tweets about Dogecoin and prices surge, what would the price have been without his tweet?
- During Black Friday, if an e-commerce site didn’t offer free shipping, how many orders would they have received?
Counterfactuals can be described as “parallel worlds”. We can never know what would have happened if we had made different choices under identical circumstances.
Yet from a data science perspective, estimating counterfactuals resembles missing value imputation. If we understand causal relationships, we can predict these “missing values” and glimpse into parallel worlds where different decisions were made.
When A/B Testing Becomes Impossible#
A/B testing, built on randomized controlled trials (RCTs), is often considered the gold standard for causal inference. By randomly assigning subjects to treatment and control groups, it overcomes the challenges I just mentioned—eliminating confounding variables and self-selection bias.
But experienced data scientists know that many scenarios cannot accommodate A/B testing:
Network Effects: Suppose a messaging app introduces emoji reactions, and you want to test whether this feature increases chat frequency. Even if only the treatment group accesses the feature, you cannot control whether treatment users chat exclusively with other treatment users. If emoji reactions genuinely boost engagement, they will simultaneously increase activity in both treatment and control groups, making the A/B test inconclusive.
Broad Impact Interventions: When interventions affect entire systems or involve huge business decisions, researchers cannot limit experiments to small subsets. We know Black Friday sales dramatically boost revenue, but executives cannot A/B test by offering discounts to only half their customers. I would be pretty upset if I found out that my friend was offered 50% discount on the same site and I wasn’t.
Historical Analysis: For past decisions, we cannot travel back in time to run A/B tests. When analyzing observational data, i.e., naturally occurring variations without experimental manipulation, we face all the challenges mentioned earlier.
When A/B testing proves impossible, data scientists turn to causal inference frameworks.
Causal Inference: Building Parallel Worlds with Data#
Causal inference provides scientific methods for analyzing intervention effects. From a data science perspective, it resembles missing value imputation; once we quantify causal effects, we can calculate missing “parallel world” values and construct unobservable counterfactual outcomes.
Have you ever wondered where you would be today if you had chosen something different in your life? Moved to another city 10 years ago? Dated another person?
Answering these types of questions requires us to create alternative worlds, …, you already know intuitively what counterfactuals are.
- Aleksander Molak, Causal Inference and Discovery in Python
Importantly, causal inference isn’t a specific statistical model but a thinking framework. In this framework, the human judgement is as crucial as data itself. As Pearl explains: “Unlike correlation and most of the other tools of mainstream statistics, causal analysis requires the user to make a subjective commitment… Where causation is concerned, a grain of wise subjectivity tells us more about the real world than any amount of objectivity.”
Methods like Directed Acyclic Graphs (DAGs) structure causal relationships (visually!), with statistical models serving as implementation tools.
Practical Applications#
Causal inference methods typically address two scenarios:
Balancing Groups: When treatment and control groups have unbalanced characteristics due to confounding variables or self-selection, methods like Propensity Score Matching create more balanced comparisons.
Constructing Counterfactuals: When everyone receives treatment, methods like Synthetic Control and Difference-in-Differences estimate what would have happened without intervention.
A Real-World Case Study: Google Pixel#
In my experience as a data scientist on Google’s Pixel phone strategy team, I encountered a perfect example of when causal inference becomes essential. Our team needed to prove that Pixel Watch ownership increases Pixel Phone brand loyalty. Specifically, we’d like to study whether Pixel Watch users were more likely to repurchase Pixel phones.
This presented a classic challenge: we couldn’t randomly assign Pixel Watches to some customers while denying them to others. Instead, we used DoubleML, a modern method combining machine learning with econometric theory, to analyze observational data. By carefully controlling for confounding variables like tech enthusiasm, spending patterns, and device preferences, we successfully quantified the causal effect of Pixel Watch ownership on Pixel Phone repurchase rates (i.e., brand loyalty).
This mirrors how major tech companies routinely apply causal inference. For example, Netflix explains how they employ synthetic control and DoubleML methods in this article.
Conclusion#
Causal inference bridges the gap between correlation and causation when A/B testing isn’t feasible. Whether you’re evaluating marketing campaigns, analyzing policy interventions, or optimizing product features, causal inference provides tools to answer the fundamental question: “What happens if we do this instead of that?”
In our data-rich world, the ability to distinguish correlation from causation is not only an academic exercise, but also a competitive advantage. Causal inference gives us the scientific rigor to peek into parallel worlds and make actionable decisions based on evidence rather than intuition.
The next time you encounter a scenario where A/B testing is impossible, remember: the econometrics theory for causal inference might be your key to unlocking those hidden parallel universes.
Essential Reading: Building Your Causal Inference Foundation#
Ready to dive deeper into the art of constructing parallel worlds with data? Two books have fundamentally shaped how I approach causal problems:
The Book of Why by Judea Pearl transformed my understanding of causation itself. Pearl doesn’t just teach methods. He rewires how you think about cause and effect. A great book to build the intuition behind causal frameworks.
For hands-on implementation, Causal Inference and Discovery in Python by Aleksander Molak bridges theory and practice seamlessly. Molak takes Pearl’s revolutionary concepts and shows you exactly how to implement them in Python. Real code, real data, real solutions.
Both books turned abstract statistical concepts into practical tools I use daily. If you’re serious about moving beyond correlation toward true causal understanding, these aren’t just recommendations. They are prerequisites.