**The most counterintuitive finding is a Simpson's paradox. Sepal measurements move in opposite directions in the pooled data and within every species.** Across all 150 flowers, sepal length and sepal width are slightly negatively correlated (r = -0.12). Within each species the relationship is clearly positive: - **Setosa:** r = 0.74 - **Versicolor:** r = 0.53 - **Virginica:** r = 0.46 The pooled sign is the reverse of the sign in all three groups. The pattern is even stronger for petal length versus sepal width. Pooled, r = -0.43, which looks like a moderate inverse relationship. Within each species it is positive: versicolor 0.56, virginica 0.40 and setosa 0.18. **Why it happens:** the species differ in average size in different directions. Setosa has small petals and relatively wide sepals. Versicolor and virginica have larger petals and sepals but narrower sepals. Pooling the species mixes these group offsets into the trend, and the between-group differences outweigh the within-group relationship, which flips the sign. **Implications:** - A correlation computed on the whole dataset would suggest wider sepals go with shorter sepals or petals, which is false inside any one species. - Analyses or feature-selection steps that ignore the species label can draw the wrong conclusion about how the measurements relate. - This is a reason to check relationships within groups, especially when the dataset exists to separate groups. The figures come from the correlation query I ran, grouped by species and overall (n = 50 per species, 150 total).