Correlation vs Causation in Research

"Correlation does not imply causation" is probably the most quoted sentence in statistics, yet causal language still slips into thousands of theses every year. Phrases like "social media use causes depression" or "tuition improves marks" are claims your data may not be able to support. This article explains exactly what a correlation tells you, why it falls short of proving cause, and how researchers move closer to causal conclusions.

What a Correlation Actually Measures

Pearson's correlation coefficient \( r \) measures the strength and direction of a linear relationship between two continuous variables. It ranges from −1 (perfect negative) through 0 (no linear relationship) to +1 (perfect positive). It is defined as

$$ r = \frac{\sum (x - \bar{x})(y - \bar{y})}{\sqrt{\sum (x - \bar{x})^2 \sum (y - \bar{y})^2}} $$

That is all it measures: how consistently two variables move together. It says nothing about which one, if either, is pushing the other.

Worked Example 1: Computing Pearson's r

A student researcher in Faisalabad records the weekly hours of private tuition taken by five FSc students and their marks in a mock physics exam.

Step 1: Means. \( \bar{x} = 30/5 = 6 \) and \( \bar{y} = 330/5 = 66 \).

Step 2: Deviations. For x: −4, −2, 0, 2, 4. For y: −11, −3, −6, 8, 12.

Step 3: Sum of cross-products.

$$ S_{xy} = (-4)(-11) + (-2)(-3) + (0)(-6) + (2)(8) + (4)(12) = 44 + 6 + 0 + 16 + 48 = 114 $$

Step 4: Sums of squares.

$$ S_{xx} = 16 + 4 + 0 + 4 + 16 = 40, \qquad S_{yy} = 121 + 9 + 36 + 64 + 144 = 374 $$

Step 5: Correlation.

$$ r = \frac{114}{\sqrt{40 \times 374}} = \frac{114}{\sqrt{14960}} = \frac{114}{122.31} = 0.932 $$

With \( df = n - 2 = 3 \), the two-tailed p-value is .021. Squaring gives \( r^2 = 0.869 \), so about 87% of the variation in marks is shared with tuition hours in this sample.

APA report: Tuition hours were strongly and positively correlated with mock exam marks, r(3) = .93, p = .021.

Notice the careful wording: "correlated with", not "improved" or "caused". Here is why.

Three Reasons a Correlation May Not Be Causal

1. Confounding (the Third-Variable Problem)

A confounder is a variable that influences both x and y. In the tuition example, family income is an obvious candidate: wealthier families can pay for more tuition and also provide quieter study spaces, better nutrition, books and internet access. Income alone could produce the correlation even if tuition had no effect at all.

A famous illustration: across summer months, ice-cream sales and drowning incidents rise together. Ice cream does not cause drowning; hot weather increases both.

2. Reverse Causality

Sometimes the arrow points the other way. Researchers have found that people who exercise more report better mood. Exercise may lift mood, but people who already feel well are also more likely to go for a walk. A cross-sectional survey measured at one point in time cannot separate these directions.

3. Chance and Selection

With many variables, some correlations will be significant purely by chance: test 100 unrelated pairs at \( \alpha = .05 \) and about five will appear "significant". Selection bias can also manufacture correlations, for example when only successful students agree to fill in a questionnaire.

Worked Example 2: Controlling for a Confounder With Partial Correlation

Suppose a larger study of FSc students finds these correlations:

• Tuition hours and marks: \( r_{xy} = 0.60 \)
• Tuition hours and family income: \( r_{xz} = 0.70 \)
• Marks and family income: \( r_{yz} = 0.75 \)

The partial correlation between tuition and marks, holding income constant, is

$$ r_{xy \cdot z} = \frac{r_{xy} - r_{xz}r_{yz}}{\sqrt{(1 - r_{xz}^2)(1 - r_{yz}^2)}} $$

Step 1: Numerator. \( 0.60 - (0.70)(0.75) = 0.60 - 0.525 = 0.075 \).

Step 2: Denominator. \( 1 - 0.70^2 = 0.51 \) and \( 1 - 0.75^2 = 0.4375 \).

$$ \sqrt{0.51 \times 0.4375} = \sqrt{0.2231} = 0.4724 $$

Step 3: Divide.

$$ r_{xy \cdot z} = \frac{0.075}{0.4724} = 0.159 $$

Once income is controlled, the tuition–marks correlation shrinks from .60 to about .16. Most of the original relationship was carried by family income. This does not prove tuition is useless, but it shows how a strong raw correlation can be largely confounded.

APA report: After controlling for family income, the partial correlation between tuition hours and marks was weak, r = .16.

Partial correlation only controls the confounders you measured. An unmeasured variable, such as student motivation, could still distort the picture.

How Researchers Get Closer to Causation

1. Randomised controlled trials (RCTs). Randomly assigning participants to conditions balances confounders, measured and unmeasured, across groups. This is the gold standard.
2. Longitudinal designs. Measuring variables over time establishes temporal order: the cause must come before the effect.
3. Quasi-experiments. Difference-in-differences, regression discontinuity and instrumental variables approximate randomisation using natural features of the data.
4. Statistical control. Multiple regression, ANCOVA and SEM adjust for measured confounders.
5. Bradford Hill considerations. In 1965 the epidemiologist Austin Bradford Hill listed points such as strength, consistency, temporality, dose–response and plausibility, which helped establish that smoking causes lung cancer even without human RCTs.

Writing Responsibly in Your Thesis

• Use "associated with", "related to" or "predicted" for correlational designs.
• Reserve "caused", "led to" or "increased" for experimental evidence.
• Name plausible confounders in your limitations section.
• If you used regression or SEM on cross-sectional data, say that the paths are theoretical, not proven causal effects.

Software such as SPSS (Analyze → Correlate → Bivariate or Partial), R (cor.test, ppcor::pcor), Stata (pwcorr, pcorr) and Python (scipy.stats.pearsonr) will compute these quickly, but the interpretation is your responsibility.

Did You Know?

• The phrase "correlation" in its statistical sense was popularised by Francis Galton in the 1880s while studying heredity.
• Countries that consume more chocolate per person have won more Nobel prizes per capita, a widely cited example of a spurious correlation driven by national wealth.
• A correlation of zero does not mean "no relationship": a perfect U-shaped curve can produce \( r = 0 \).

Key Takeaways

• Pearson's r measures linear association, not cause and effect.
• Confounding, reverse causality and chance can all create correlations without causation.
• Partial correlation and regression control measured confounders, but not unmeasured ones.
• Randomised experiments and longitudinal designs provide the strongest causal evidence.
• Match your wording to your design: "associated with" for surveys, "caused" only for experiments.

If you need help interpreting correlations, running partial correlation or writing up your results chapter, Amathify's experts and PhD scholars offer step-by-step support via WhatsApp or the Ask a Question form.