Correlation, Regression & Their Limits - Interview Revision Guide
If two lines rise together on a dashboard, are they connected - or merely walking in the same direction because the city got hotter? Correlation and regression are powerful because they turn messy business data into patterns; they are dangerous because patterns can look like proof when they are only clues.
- Correlation measures direction and strength of linear association between two quantitative variables; it does not say which variable caused the other.
- Regression estimates how an outcome changes with predictors, usually for prediction, explanation or control.
- A high correlation can be useful for forecasting even when it is not causal, but it is weak evidence for strategy unless a mechanism exists.
- Simple linear regression gives an equation: Y = a + bX, where b is the expected change in Y for a one-unit change in X.
- Judge regression with out-of-sample performance, not just in-sample fit: use R², RMSE, MAE, MAPE and residual checks.
- The biggest limits are confounding variables, reverse causality, nonlinearity, outliers, small samples, multicollinearity and data leakage.
- Best answer structure: define, show the equation, explain interpretation, validate the model, then state the causality caveat.
Think of the topic as a ladder: correlation spots association, regression quantifies a relationship, and causal analysis tests whether a business action will actually move the outcome. Many candidates jump from rung one to rung three; strong candidates do not.
Core Explanation: What Each Tool Actually Does
Correlation is the quick scan. It answers: when X is high, is Y usually high, low or unrelated? The most common measure is Pearson’s r, which ranges from -1 to +1.
- r = +1: perfect positive linear relationship.
- r = 0: no linear relationship, though a curved relationship may still exist.
- r = -1: perfect negative linear relationship.
Regression goes one step further. It gives you an equation, typically:
Y = a + bX + error
Here, Y is the outcome, X is the predictor, a is the intercept, b is the slope, and error is the part the model cannot explain. In business language, the slope b says: “for a one-unit increase in X, what is the expected change in Y?”
Definitions You Can Say in One Breath
- Correlation: A unit-free statistic that measures the direction and strength of a linear relationship between two quantitative variables.
- Regression: A model that estimates the average change in an outcome associated with one or more predictors.
- Ordinary Least Squares: A regression method that estimates coefficients by minimizing the sum of squared residuals.
- Residual: The difference between the observed value and the value predicted by the model.
- Causation: A change in X produces a change in Y, holding competing explanations under control.
Correlation vs Regression vs Causation
Use this distinction in business cases. If app installs and ad spend move together, correlation tells you there is a pattern. Regression estimates expected installs for a given spend. Causal evidence tells you whether extra spend truly created incremental installs rather than merely coinciding with festive demand, influencer buzz or a competitor stockout.
A Small Worked Example: From Correlation to Regression
Assume a brand tracks weekly ad spend and sales index for five weeks. This is hypothetical data to show the calculation, not a real company dataset.
Mean of X = 3. Mean of Y = 18. The deviations give:
- Σ(X - X̄)(Y - Ȳ) = 34
- Σ(X - X̄)² = 10
- Σ(Y - Ȳ)² = 120
Correlation: r = 34 / √(10 × 120) = 0.98. That is a very strong positive linear association.
Regression slope: b = 34 / 10 = 3.4. Intercept a = 18 - 3.4 × 3 = 7.8. So the fitted equation is:
Predicted sales index = 7.8 + 3.4 × ad spend index
If ad spend index becomes 6, predicted sales index = 7.8 + 3.4 × 6 = 28.2.
The calculation does not prove ads caused sales. It only shows that, in these five observations, higher ad spend moved with higher sales. Seasonality, price cuts, distribution expansion or competitor activity may be hidden drivers.
How to Judge a Regression Model
Never say “the model is good because R² is high” and stop there. A useful business regression must fit reasonably, predict unseen data, have sensible residuals and avoid leakage.
The Limits: Where Correlation and Regression Mislead
The limits are not technical trivia; they are where business decisions go wrong.
Case Study: Bajaj Finance and Credit Risk Models
Bajaj Finance shows how correlation and regression-like scoring can improve lending decisions, while still needing governance because repayment patterns are not the same as causal truth.
Bajaj Finance operates in Indian consumer lending, where millions of decisions depend on one question: which applicants are likely to repay? In such a setting, historical repayment data, bureau behaviour, product category, customer vintage, income proxies and channel quality can all be statistically associated with default risk.
The strategic move: the company has built a lending model around data-led underwriting, risk segmentation and disciplined collections. The statistical logic is similar to regression and classification: estimate the relationship between borrower characteristics and repayment outcomes, then use that estimate to approve, price, limit or reject credit. The primary driver is risk-based decisioning at scale, supported by bureau data, internal repayment history, distribution partnerships, collection infrastructure and compliance with Indian financial regulation.
The lesson: a variable may be predictive without being causal. For example, a customer’s past repayment behaviour may predict future repayment, but simply changing that variable on paper does not “cause” lower risk. During shocks, such as income disruption or changes in credit cycles, old correlations can weaken. That is why lenders combine models with policy rules, stress testing, monitoring and human governance.
How AI Changes Correlation, Regression & the Limits of Both
AI does not remove the limits of correlation and regression; it makes the pattern-finding faster and the validation more important.
- AutoML finds nonlinear patterns: tools can test regression, tree-based models and interactions quickly. This improves prediction, but it can hide the simple business logic that managers need.
- Causal ML is becoming more common: uplift modelling, heterogeneous treatment effects and causal forests help estimate which customers may respond to an intervention, not just who is likely to buy anyway.
- LLMs speed up interpretation and feature ideation: they can summarize variables, spot possible confounders and draft model documentation, but they can also suggest leaky or unfair variables if unchecked.
Load a small dataset description, variable dictionary and company context into ChatGPT or Claude. Ask: “List possible confounders, leakage risks, nonlinear relationships and validation checks before I trust this regression.” Then verify the suggestions yourself using the scatterplot, residuals and train-test performance.
Interview Relevance
“Our sales and ad spend have a correlation of 0.85. Can we conclude that increasing ad spend will increase sales? How would you test it?”
A strong answer uses both language and caution: “This is predictive evidence, not causal evidence. I would use it to forecast, but I would test incrementality before changing the budget.”
Common Mistake
The single biggest mistake is saying, “Correlation is high, so X causes Y.” It costs candidates because it shows they may make expensive business decisions from a dashboard pattern. The one-line fix: say “associated with” for correlation and regression, and reserve “causes” for experiments or strong causal designs.
What to Revise Next
Now move from pattern recognition to proof. Revise Correlation vs Causation: Confounders & Study Design to understand how to test whether X truly moves Y, then revise Statistical Traps: p-Hacking, Multiple Comparisons & Simpson's Paradox to avoid the traps that make clean-looking analysis misleading.