Everyone Cares About Causality, So What Is the Point of "Correlational" Data Analysis?

I. Introduction: An "Old Friend" Now Out of Favor

If you follow methodological debates in the social sciences in recent years, you will surely feel a powerful trend—causal inference has come to dominate the discourse almost entirely.

Instrumental variables (IV), difference-in-differences (DID), regression discontinuity, randomized controlled trials—these terms, once found only in advanced econometrics textbooks, have now permeated everyday discussions in sociology, political science, management, public health, and even computer science. Open the review comments of any empirical paper, and "How do you identify the causal relationship?" is almost certainly the first question asked.

In this climate, the status of the word "correlation" has become somewhat awkward. Saying that a study finds a "correlational relationship" rather than a "causal relationship" sounds like saying the paper is "not quite up to standard." Some even feel: Since correlation does not imply causation, what is the point of learning all that correlation analysis, regression analysis, and correlation coefficients?

This article aims to give "correlation" a fair assessment—it is not some second-rate concept; it simply performs a different task from causality.

Key message: Correlation tells you "what appears together with what," while causation tells you "what change will bring about what other change." The two are not superior and inferior—they are answers to different questions.

---

II. Correlation and Causation: What Exactly Is the Difference?

2.1 An Overused Example That Still Clears the Mind Every Time

Ice cream sales and drowning deaths are highly positively correlated.

You would not actually believe that "eating ice cream causes drowning," would you? The mechanism behind it is obvious: Hot weather → more people eat ice cream, and hot weather → more people go swimming → more drownings. There is no causal chain between ice cream and drowning; they are simply two outcomes of a common cause.

This example is repeatedly cited not because it is novel—but because it demonstrates in the clearest way the chasm between correlation and causation: Just because you see two variables moving together does not mean that changing one will change the other.

2.2 A Precise Distinction in Mathematical Language

Let $X$ and $Y$ be two variables.

Correlation concerns: When you know the value of $X$, does your expected value of $Y$ change?

$\mathbb{E}[Y|X=x] \neq \mathbb{E}[Y|X=x']$

If the conditional expectation of $Y$ differs for different values of $X$, we say $X$ and $Y$ are correlated. The core feature of correlation is symmetry: if $X$ and $Y$ are correlated, then $Y$ and $X$ are also correlated. $\text{Corr}(X,Y) = \text{Corr}(Y,X)$.

Causation concerns a completely different question: If you artificially, externally change the value of $X$ (rather than waiting for it to change naturally), will the distribution of $Y$ change as a result?

$\mathbb{E}[Y|\text{do}(X=x)] \neq \mathbb{E}[Y|\text{do}(X=x')]$

Here, $\text{do}(\cdot)$ is the notation introduced by Judea Pearl, representing intervention—not "observing that X = x," but "making X become x." The core feature of causation is asymmetry: X causing Y does not equal Y causing X.

2.3 A Table to Clarify

Dimension	Correlation	Causation
Core question	When X changes, how does Y move along?	If I change X, will Y change as a result?
Mathematical expression	$\mathbb{E}[Y\|X]$	$\mathbb{E}[Y\|\text{do}(X)]$
Symmetry	Symmetric: $\text{Corr}(X,Y)=\text{Corr}(Y,X)$	Asymmetric: X→Y does not equal Y→X
Guidance for intervention	Cannot guide intervention	Directly guides intervention
Basis for judgment	Observed patterns of co-movement	Requires identification strategies to rule out alternative explanations
Typical methods	Correlation coefficient, scatter plot, OLS regression (descriptive only)	RCT, IV, DID, RDD
In one sentence	"X and Y move together"	"Moving X moves Y"

---

III. Why Do We Still Need Correlation? Three Irreplaceable Values

In an era when causal inference is ascendant, what use is correlation? Has it not been "surpassed"?

3.1 Value One: Correlation Is the First Gateway to Causal Discovery

Almost any exploration of causality begins with the discovery of a "strange correlation."

In the 19th century, when John Snow investigated the cholera outbreak in London, what he first discovered was not the causal mechanism of "sewage causes cholera"—he found a spatial correlation: cholera deaths clustered around a water pump on Broad Street. This correlation on the map led him to the pump, and before fully understanding the biological mechanism of "bacteria → cholera," he was already able to stop the outbreak by removing the pump handle.

In economics, many causal studies also begin with correlations. Card & Krueger did not study the effect of the minimum wage on employment because theoretical derivation had failed—classical theory was logically airtight. Rather, researchers had observed a correlation that made classical theory uncomfortable: certain regions that raised the minimum wage showed no signs of declining employment. This "anomalous correlation" drove decades of subsequent causal research.

Correlation is not the enemy of causation—it is the guide to causation. Without strong and robust correlations, most causal research would never even begin.

3.2 Value Two: Many Decisions Do Not Need Causation at All—Prediction Is Enough

This is the most central practical value of correlation, and the clearest division of labor between it and causation.

Consider the following scenarios:

• E-commerce recommendation systems: "Customers who bought A also bought B"—you do not need to know whether "buying A causes buying B" or "buying A and buying B are both expressions of some taste." You only need to know that "if a person bought A, recommending B to them makes them more likely to place an order."
• Credit scoring: Banks assess the default risk of loan applicants. The model might use a variable like "whether the applicant has changed phone numbers frequently." There may be no causal chain behind this—people who change numbers frequently have higher default rates, not because changing numbers itself causes default, but because changing numbers correlates with "life instability." But the bank does not care about causation—the bank only cares whether this variable helps predict "whether this person will repay on time."
• Movie recommendations: Netflix does not need to know whether "liking The Godfather causes you to like Goodfellas." It only needs to know that the taste patterns for these two movies overlap highly.
• Search engines: When you type "cold symptoms," the results returned by the search engine do not need to infer "causal relationships between keywords and web pages." Correlation-based ranking is sufficient.

What do these scenarios have in common? You do not need to intervene in the world—you only need to make the best prediction or ranking given the available information. The task of prediction and the task of intervention require entirely different information structures.

3.3 Value Three: Correlation Is the Technical Foundation of Causal Inference

An easily overlooked fact is that all causal inference methods, at the operational level, rely on the calculation and comparison of correlations.

• Difference-in-differences (DID): Essentially compares the difference between the "before-after change in the treatment group" and the "before-after change in the control group"—you are computing a correlation adjusted by differencing.
• Instrumental variables (IV): The core of the first-stage regression is calculating the correlation between the instrument and the endogenous variable—if the first-stage correlation is too weak (the weak instruments problem), the second-stage causal inference is unreliable.
• Randomized controlled trials (RCT): At the operational level, what you do is compare the mean difference between the treatment and control groups—if randomization is successful, this direct mean difference (essentially a simple correlation) can be interpreted as the causal effect.

Causal inference methods have not "replaced" correlation analysis—they add a layer of "design" and "assumptions" on top of correlation analysis, allowing correlation to be given a causal interpretation.

---

IV. When Is Correlation Enough? When Do You Absolutely Need Causation?

This is the most central practical judgment framework in this article. You can use it to quickly determine what level of analysis any given question requires.

4.1 Correlation Is Enough—Prediction-Type Problems

When the problem you face satisfies the following two conditions, correlation analysis is sufficient:

• Your task is prediction only, not intervention. You will not change the system based on the prediction results; you only want to know "what is most likely to happen given the available information."
• The data-generating process is stable. The historical data you use to train the model and the future scenarios come from the same distribution. If the system itself undergoes structural changes (such as major policy shifts or pandemic shocks), predictions based solely on historical correlations may fail.

Typical scenario checklist:

Scenario	Output you need	Why correlation is enough
E-commerce recommendation	"Which product is this user most likely to click"	No need to change the user, only to rank
Credit scoring	"What is this applicant's default probability"	No need to change the applicant's behavior
Weather forecasting	"What is the probability of rain tomorrow"	No need for artificial rain to determine causation
Natural language translation	"What is the most likely target language for this sentence"	Correlation patterns are sufficient to support accurate translation
Medical screening (note the distinction!)	"How likely is this patient to have a certain disease"	Identifies high-risk populations, but treatment decisions require causal evidence

Special note—the "correlation vs. causation" dividing line in medicine: Using CT imaging for early lung cancer screening is a prediction-type problem—your goal is to "find people most likely to have pulmonary nodules," and correlation-based machine learning models can outperform human radiologists. But once you enter treatment decisions—"After finding a nodule, should we surgically remove it or follow up with observation? Can surgery reduce mortality?"—this becomes a causal question requiring causal evidence from RCTs. No matter how strong, a correlational model cannot answer it.

4.2 Causation Is Absolutely Necessary—Intervention-Type Problems

When the problem you face involves "what would happen if we did this"—that is, counterfactual reasoning—correlation analysis is no longer sufficient.

Judgment criterion: Your conclusion will directly guide an action that changes the state of the system.

Typical scenario checklist:

Scenario	Output you need	Why correlation is not enough
Minimum wage policy evaluation	"If we raise the minimum wage, will employment decrease?"	Historically, "regions that raised the minimum wage" were not randomly selected; observed correlations may be confounded by local economic conditions, political leanings, and other factors
Education investment decisions	"Can smaller class sizes improve student performance?"	Schools with smaller classes also tend to be wealthier, and their parents care more about education—correlation ≠ causation
Drug approval	"Can taking this new drug reduce mortality?"	Without an RCT, you cannot rule out alternative explanations such as "people who choose to take this drug are already healthier/more compliant with medical advice"
Marketing	"Did this advertising campaign actually boost sales?"	The timing and regions of advertising campaigns are often carefully chosen (peak season + high purchasing power areas), not random—the observed "sales increase after advertising" may simply be the natural growth that region and period would have experienced anyway

4.3 A Judgment Flowchart

What does the problem you face require?

"I want to know what is most likely to happen under current conditions"
        ↓
    → Prediction-type problem → Correlation analysis is sufficient
        │
        │  Examples: recommendation systems, credit scoring, translation, weather forecasting
        │
"I want to know if Y will change because I do X"
        ↓
    → Intervention-type problem → Causal inference is needed
        │
        │  Examples: policy evaluation, drug approval, investment decisions
        │
"I want to know why Y changes—the complete chain of influence"
        ↓
    → Mechanism-type problem → Causation + mechanism analysis needed
        │
        │  Examples: How does education affect income through skill accumulation?
        │      Through what channels does the minimum wage affect employment?

---

V. In Which Fields Has Correlation Analysis Created Enormous Value?

To further illustrate that correlation is not a "second-rate concept," here are several fields built entirely on correlational thinking—they have created trillions of dollars in economic value without relying on causal inference.

5.1 Recommendation Systems—The Infrastructure of the Internet Economy

Amazon's "customers who bought this item also bought…", YouTube's "up next," TikTok's "For You" feed, Spotify's "Discover Weekly" playlist—the core engine behind all these user experiences is collaborative filtering or other correlation-driven methods.

These systems do not care whether "watching video A causes you to watch video B." They only care: "What content did users with similar behavioral patterns to yours consume next?"

This is pure correlational thinking—if user A and user B have highly similar past behavioral patterns, then content that B has already watched but A has not yet seen is likely suitable to recommend to A. This prediction requires no causal assumptions.

Recommendation-system-driven revenue accounts for more than 30% of Amazon's total revenue, and Netflix claims its recommendation system saves the company over $1 billion annually in subscription churn costs. These are the economic values contributed by "correlational thinking."

5.2 Credit Scoring and Risk Pricing—The Core Engine of the Financial Industry

When a bank decides whether to approve a loan, or an insurance company sets premium levels, they rely on risk models. The core inputs of these models are not causal relationships but predictive variables highly correlated with default or claim probabilities.

For example, a bank finds a robust correlation between "a person changed their phone number within the last 6 months" and "default." This behavior itself may not cause default—but it is a signal of life instability, and people with unstable lives are indeed more likely to default. Would this variable be used by the bank as a basis for approval? Yes. And from a risk-control perspective, doing so is reasonable.

The logic behind this is pure predictive thinking: "I do not need to understand 'why.' I only need to know 'under condition A, what is the average probability of B.'"

5.3 Anomaly Detection—From Credit Card Fraud to Industrial Quality Control

When you receive a text message from your bank saying "Your credit card just had a suspicious transaction; please confirm whether it was you," the engine behind it is an anomaly detection model.

The basis for this model's judgment is correlational patterns: "Your usual spending times are concentrated between 10 a.m. and 10 p.m., locations are concentrated in your city of residence, and amounts are mostly small to medium. Now there is a transaction at 3 a.m., in a different city, for a large amount—the probability of this occurring in your historical behavioral pattern is extremely low. Flag as anomalous."

Again, no causal inference is involved—the model does not say "spending at 3 a.m. causes fraud." It simply identifies a statistical outlier.

Similar applications span industries: manufacturing plants use correlational patterns to detect defective products on production lines, cybersecurity companies use correlational patterns to identify anomalous traffic, and hospitals use correlational patterns to flag potential medical insurance fraud.

5.4 The Currently Popular Large Language Models—Also a Triumph of Correlation

ChatGPT, Claude, ERNIE Bot… What is the core mechanism of these large language models?

Simply put, given the preceding sequence of words, predict the next most likely word. This "predict the next word" task is essentially an extremely large-scale conditional probability estimation problem—that is, a correlation task. The model does not "understand" the text it generates; it is merely making the highest-probability sequence prediction. Yet within this correlational framework, the model can generate surprisingly fluent and accurate text.

From GPT-3 to GPT-4 to the Claude series, the leap in model capability has not changed this underlying logic—they are still making predictions, not doing causal reasoning.

However, there is also a hidden limitation here. Because large language models are built on correlational thinking, when faced with tasks requiring causal reasoning (such as "if I change this policy, what will happen?"), they may give answers that seem plausible but are actually wrong—because they have seen vast amounts of text containing "policy → outcome" correlations, but they do not truly distinguish correlation from causation. This is why requiring AI to perform causal reasoning is a frontier direction in current research.

---

VI. Summary: The Relationship Between Correlation and Causation in Three Sentences

First: Correlation tells you "what goes with what," and causation tells you "what changing what will change what." The former helps you make optimal decisions in an unchanging world; the latter helps you design how to change the world.

Second: Correlation is not the enemy of causation, but its guide. Almost all causal research begins with discovering an interesting correlation. Without correlation, you would never think to ask about causation.

Third: Choosing between correlation and causation is not a matter of sophistication—it is a matter of the task. Prediction, recommendation, screening, ranking—these tasks do not require causation. Intervention, policy evaluation, decision design—these tasks cannot rely on correlation alone.

Next time someone says "your study only found a correlational relationship," you can ask in return—"Wait, what exactly is your task? If you are only predicting, then correlation is everything you need."

---

VII. Presentation Suggestions for Bilibili/WeChat Official Account

• Bilibili video: For the core visuals, consider a "dual-narrative" approach—one thread tells the classic ice cream and drowning example (correlation ≠ causation), the other thread tells the story of recommendation systems (correlation creates value). Interweave the two threads and bring them together at the "judgment flowchart."
• WeChat official account: Turn the two judgment tables into infographics. Turn the judgment flowchart into a vertical decision tree for easy reading and screenshotting on mobile.
• Recommended titles:
  • Main title: 《Everyone Cares About Causality, So What Is the Point of Correlation Analysis?》
  • Alternative title: 《Correlation ≠ Causation, But Correlation Is Far More Than Something "Not Good Enough"》
• Key quote:

  "Choosing between correlation and causation is not a matter of sophistication—it is a matter of the task. Use correlation for prediction, and causation for intervention."