Correlation Tests

Correlation tests answer a simple question:

“Are these two variables related — and how strongly?”

SciPy provides three main correlation measures:

Pearson Correlation

Measures linear relationship between two continuous variables

Use when:

• Both variables are continuous
• Relationship looks linear
• Normality assumption is reasonable
• No major outliers

Example

Outputs

• corr → correlation coefficient (range: -1 to +1)
• p → p-value for the significance of the correlation

Interpretation

• corr = 1.0 → perfect positive linear relationship
• corr = -1.0 → perfect negative linear relationship
• corr = 0 → no linear relationship

Practical notes

• Sensitive to outliers
• Only captures linear relationships
• If data is skewed → use Spearman instead

pearman Rank Correlation

Monotonic relationship (non-parametric). Spearman correlates the rank order of the values, not the raw values.

Use when:

• Data is not normal
• Relationship is monotonic (always increasing or decreasing, not necessarily linear)
• Data is ordinal (e.g., Likert scales)
• There are outliers

Example

Interpretation

• Same coefficient range (-1 to +1)
• But correlation is based on ranking

Practical notes

• More robust than Pearson
• Does not assume normality
• Great for messy real-world data

Kendall’s Tau

Rank correlation robust to ties (non-parametric)

Use when:

• Sample sizes are small (< 20)
• Data contains many ties (duplicate values)
• You want a robust rank-based measure

Example

Interpretation

• Correlation coefficient usually smaller in magnitude than Pearson or Spearman
• Works well on small or noisy datasets

Practical notes

• Most robust to ties
• Slowest with large datasets
• Rarely used in large-sample applied work, but great for small surveys

When to Use Which Correlation Test

Scatterplots

Always pair correlation with a plot.

Example:

Visual checks reveal:

• Non-linear patterns
• Outliers
• Clusters
• Heteroscedasticity (unequal variance)

Correlation Matrices

Useful for exploring many variables at once.

Using NumPy + SciPy:

Spearman matrix via pandas:

Multiple Correlation Tests (Loop Example)

When testing many variable pairs:

Effect Sizes for Correlation

Correlation coefficient is already an effect size.

Heuristic (Cohen's guidelines):

• 0.10 → small
• 0.30 → medium
• 0.50 → large

This applies to Pearson and Spearman. Kendall uses slightly smaller thresholds.

Partial Correlation (Not in SciPy)

To measure correlation while controlling for another variable, use Pingouin or Statsmodels.

Example (Pingouin):

Practical Examples — When Correlation Tests Matter

Example 1 — Customer behaviour

• Time on site vs purchase amount
• Use Spearman (non-normal)

Example 2 — Medical data

• Age vs blood pressure
• Likely Pearson

Example 3 — Finance

• Returns of two stocks
• Pearson or Spearman depending on distribution

Example 4 — Survey analysis

• Satisfaction rating vs recommendation likelihood
• Spearman or Kendall (ordinal data)

Example 5 — Performance metrics

• CPU usage vs response time
• Often non-linear → Spearman