Parametric versus non-parametric
Parametric tests (like the t-test) assume something about the population’s distribution, usually normality, and test a parameter such as a mean.
Non-parametric tests make fewer assumptions. Use them when:
- the data are ranks or categories rather than measurements,
- the sample is small and the distribution is unknown or clearly not normal,
- outliers would distort a parametric test,
- the question is not about a parameter at all (for example, whether a series is random).
Testing a correlation coefficient
To test whether a population correlation is zero (H0: ρ = 0), use:
t = r × √(n − 2) / √(1 − r²) with n − 2 degrees of freedom
Example. r = 0.30 from 30 observations.
- t = 0.30 × √28 / √0.91 = 0.30 × 5.292 / 0.954 = 1.66.
- Two-tailed 5% critical value with 28 df ≈ 2.05.
- 1.66 < 2.05, so fail to reject H0: no significant correlation.
Notice how sample size matters. The same r = 0.30 with 100 observations gives t ≈ 3.11, which is significant. Small correlations need large samples to be detected.
Spearman rank correlation
The Spearman rank correlation measures how closely two rankings agree. Rank each variable, then compute:
rs = 1 − 6 Σd² / [n(n² − 1)]
where d is the difference between each pair of ranks. It is less affected by outliers and does not require normality. For large samples (n > 30), test its significance with the same t-formula as above.
Example use: do funds ranked best on past returns also rank best on future returns?
Chi-square test of independence
When both variables are categorical, put the counts in a contingency table. Example: credit rating category (rows) against sector (columns).
- For each cell, the expected count if the variables were independent = (row total × column total) / grand total.
- Calculate:
χ² = Σ (observed − expected)² / expected
- Degrees of freedom = (rows − 1) × (columns − 1).
- Reject independence if χ² exceeds the critical value. The test is one-tailed (right tail).
A large χ² means the observed counts differ from what independence would predict, so the variables are likely related.
Exam focus
Expect the correlation t-test (watch the n − 2 degrees of freedom), a question on when to use a non-parametric test, and the setup of a chi-square contingency test, including expected counts and degrees of freedom.