Association Metrics: r, η, and Cramér’s V

What the three measures in the pairwise association matrix mean, how to interpret their values, and why they cannot be compared with each other directly.

Overview

The Stat Analysis matrix shows how strongly every pair of variables in your dataset is associated. Statulator automatically picks the most appropriate measure of association based on whether each variable in the pair is numeric (continuous) or categorical (discrete). There are three possible pair types, each with its own measure:

Association measure shown for each type of variable pair
Variable pairMeasure shownSymbolRange
Num × Num Pearson correlation r −1 to +1
Num × Cat Eta (correlation ratio) η 0 to 1
Cat × Cat Cramér’s V V 0 to 1

The colour and shading of each cell in the matrix encode the strength of the association. Keep in mind, however, that the three measures are not directly comparable: an r of 0.17, an η of 0.17, and a V of 0.17 mean different things, and the thresholds for “small / medium / large” differ across them. See Can I compare the three measures directly? below.

Pearson Correlation (r) — Numeric × Numeric

What it measures: the strength and direction of the linear relationship between two numeric variables.

Why it is used here: when both variables are numeric and roughly continuous, the natural question is “do high values of one go with high values of the other?” Pearson’s r answers this directly, on a scale that is symmetric, signed, and well understood.

\[ r = \frac{\sum (x_i - \bar{x})(y_i - \bar{y})}{\sqrt{\sum (x_i - \bar{x})^2 \cdot \sum (y_i - \bar{y})^2}} \]
Interpretation
Interpreting the magnitude of Pearson's r
|r|Typical labelMeaning
0.0–0.1TrivialEssentially no linear relationship.
0.1–0.3Small / weakHint of a relationship; large samples are needed to detect it reliably.
0.3–0.5ModerateA noticeable trend, visible in a scatter plot.
0.5–0.7StrongClear linear pattern.
0.7–1.0Very strongTight linear relationship, near-deterministic at 1.0.

The sign matters: r = −0.22 means a weak negative trend (high values of one variable tend to go with low values of the other); r = +0.22 is the same strength in the positive direction.

Caveat: Pearson’s r only measures linear association. Two variables can have a strong curved or non-monotonic relationship and still show r close to 0 — always look at the scatter plot before concluding “no relationship”.

Next step: click the cell to open the full Numeric × Numeric analysis — scatter plot, regression line, confidence interval, Spearman’s ρ, and a hypothesis test of whether r differs from zero.

Eta (η) — Numeric × Categorical

What it measures: how much of the variation in a numeric variable is explained by group membership in a categorical variable — equivalently, the strength of association between a numeric response and a categorical predictor.

Why it is used here: a scatter plot does not make sense when one variable has no natural ordering. The question becomes “do the mean values of the numeric variable differ across the categories?” Eta captures that on a 0-to-1 scale and is the square root of the η² statistic from one-way ANOVA.

\[ \eta^2 = \frac{SS_{\text{between}}}{SS_{\text{total}}} = \frac{\sum_g n_g (\bar{y}_g - \bar{y})^2}{\sum_i (y_i - \bar{y})^2}, \qquad \eta = \sqrt{\eta^2} \]
  • The numerator sums over each category \( g \) with size \( n_g \) and mean \( \bar{y}_g \); the denominator is the total variation in \( y \).
  • \( \eta^2 \) is identical to the R² of a one-way ANOVA model with the categorical variable as the only predictor.
Interpretation

Cohen’s conventions, treating η² like R²:

Interpreting the magnitude of eta squared and eta
η²ηEffect size
0.010.10Small
0.060.24Medium
0.140.37Large

Worked example: η = 0.17 between Age (numeric) and Treatment Group (categorical) means η² ≈ 0.029 — about 2.9% of the variation in Age is explained by group membership. This is a small effect, close to the border with medium: the group means of Age differ, but the spread within each group is much larger than the gap between them.

Direction: eta has no sign — it is always positive, because categorical variables have no “positive direction”. Open the full analysis to see which groups have higher or lower mean values.

Caveat: η assumes a one-way layout (one categorical predictor, one numeric response) and treats the categories as unordered. If your categories are actually ordinal (e.g., low / medium / high), a Spearman or Kendall rank correlation may capture more of the relationship.

Next step: click the cell to open the full Numeric × Categorical analysis — box plots, group means with confidence intervals, t-test or one-way ANOVA (with Mann-Whitney U or Kruskal-Wallis as non-parametric alternatives), and effect sizes.

Cramér’s V — Categorical × Categorical

What it measures: the strength of association between two categorical variables, based on the chi-square statistic of their contingency table and normalised so that V always lies between 0 and 1 regardless of table size.

Why it is used here: when both variables are categorical, the natural question is “does the distribution of one variable depend on the other?” The raw chi-square statistic answers this, but its scale depends on the sample size and the number of cells, making it impossible to compare across tables. Cramér’s V removes both dependencies.

\[ V = \sqrt{\frac{\chi^2}{n \cdot (\min(r, c) - 1)}} \]
  • \( n \) = sample size;   \( r \) = number of row categories;   \( c \) = number of column categories.
  • \( \chi^2 \) = Pearson chi-square statistic of the contingency table.
Interpretation

The “small / medium / large” thresholds shift with the table size, because larger tables can hold more partial associations (Cohen 1988, adapting his w guidelines for chi-square tests):

Cramer's V effect-size thresholds by table size
df* = min(r, c) − 1SmallMediumLarge
1 (2 × 2 table)0.100.300.50
2 (3-level)0.070.210.35
3 (4-level)0.060.170.29
4 (5-level)0.050.150.25

Worked example: V = 0.13 between Outcome and Treatment Group (both categorical) is small to small-medium by every row of the table above. The two variables share modest information — knowing one shifts the distribution of the other a little, but not dramatically.

Direction: Cramér’s V has no sign — categorical variables have no direction. To see which cells contribute most to the dependence, examine the standardised residuals in the contingency table.

Caveat: Cramér’s V can be inflated when one variable has many sparsely populated levels. If cells have expected counts < 5, the chi-square statistic (and therefore V) is unreliable — switch to Fisher’s exact test.

Next step: click the cell to open the full Categorical × Categorical analysis — contingency table with observed and expected frequencies, chi-square test of independence (or Fisher’s exact test for sparse tables), and odds ratio and relative risk when both variables have two levels.

Can I Compare the Three Measures Directly?

No, not directly. A V of 0.17, an η of 0.17, and an r of 0.17 do not mean the same thing. Each measure has its own scale, its own thresholds, and its own underlying assumptions. Use the matrix to rank the strongest associations within each variable-type group, not to compare values across groups.

A rough rule of thumb:

Why These Three Metrics, and Not Others?

Statulator uses r, η, and V because they are the standard, most widely taught measures for each pairing of variable types, with consistent scales and well-established statistical tests behind them. Alternatives exist and may be added in future updates:

References

  1. Pearson, K. (1895). Note on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58, 240–242.
  2. Pearson, K. (1905). On the general theory of skew correlation and non-linear regression. Drapers’ Company Research Memoirs, Biometric Series II.
  3. Cramér, H. (1946). Mathematical Methods of Statistics. Princeton University Press, §21.9.
  4. Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum., Chapters 7–8: effect sizes for chi-square tests and ANOVA.
  5. Snedecor, G. W., & Cochran, W. G. (1989). Statistical Methods (8th ed.). Iowa State University Press., Chapter 10: correlation.