Less mystery.
More understanding.
Statistics makes more sense when you can see it, move it and question it.
A lab for curiosity.
Correlation Lab is an educational tool for learning and exploring statistical correlation. It connects the equation, the visualization and the interpretation in one place, for students, teachers and anyone curious about data.
Your data stays yours.
CSV files are read in your browser. Data analysis, example loading and exports run locally. This app does not send input datasets to an analysis server or store them in a database. Your theme preference is saved on this device; input datasets are held in memory and reset on page reload.
What the numbers mean.
- Pearson uses centered, scaled sums; Spearman applies Pearson to average ranks; Kendall uses tau-b with correction for ties.
- Pearson’s two-sided p-value uses the Student t / incomplete-beta relationship under the independent bivariate-normal sampling model. For n = 2 the p-value is 1.
- The Pearson 95% confidence interval uses Fisher’s z approximation for n > 3. Perfect sample correlation yields a degenerate interval, which must not be interpreted as population certainty.
- Rank tests use exhaustive, two-sided permutation for n ≤ 8, including labeled permutations with ties. For Spearman n = 9–500, a reproducible Monte Carlo permutation test uses 9,999 shuffles and plus-one correction; for n > 500, it uses a large-sample t approximation. Kendall n > 8 uses a tie-corrected normal approximation. Rank confidence intervals are not estimated.
- Sample standard deviation uses n − 1. Regression is ordinary least squares with an intercept; its R² is Pearson r squared.
- Potential outliers use the marginal 1.5 × IQR rule with interpolated quartiles. Hiding their markers never changes the sample used for calculation.
Educational use.
All built-in datasets are synthetic teaching examples. Labels such as “weak” and “strong” help build intuition but are not universal scientific thresholds. This site does not support causal conclusions from correlation alone.
Statistical disclaimer.
Results support education and exploration and do not automatically replace professional statistical analysis. Check study design, sampling, dependence, missingness, measurement quality and model assumptions before using a result to make decisions. A significant p-value does not measure effect size or practical importance.
The browser workspace supports up to 2,000 rows, 20 columns and a 2 MB CSV. Values must be finite decimals within the documented magnitude range. Missing or malformed numeric values are reported; rows are never silently dropped.
Read further.
The implementation is independently written and its test fixtures are checked against SciPy’s reference implementation.