Two variables move together. The lazy conclusion: one causes the other. The usual reality: a hidden confounder drives both, and the direct link is an illusion. The classic structure:
X ⇠Z⇢ Y
Zthe confounder. It causes both X and Y, manufacturing a correlation between them even though X does not cause Y at all.
Tap each famous pair to reveal its hidden Z. Then use the second tool to watch a confounder fabricate a correlation — and watch it vanish when you account for Z.
Interactive 1a — tap a pair to expose the confounder.
Interactive 1b — turn the confounder up. The raw correlation appears from nothing.
Raw correlation (X,Y)
—
True direct effect
0.00
What rescues you. Three options: (1) a true experiment (randomise X, breaking Z's influence); (2) control for Z by adding it to a regression ("all else equal"); (3) honesty — write "is associated with", report the confounders you couldn't rule out, and don't claim cause. On observational DBA data, option 3 is often the right one.
TRAP 2 · OVERFITTING
"MY MODEL IS 98% ACCURATE"
Memorising is not learning
A model flexible enough can fit your data perfectly — including its random noise. It looks brilliant on the data it trained on and fails on anything new. The defence is the train/test split: hide some data, train on the rest, and judge the model only on the hidden part it never saw.
This is the bias–variance tradeoff. Slide model complexity from a stiff straight line to a wild wiggly curve. Watch training error fall forever while test error falls, bottoms out, then climbs — the U-shape that defines the sweet spot.
Interactive 2 — find the complexity that minimises TEST error, not training error.
Training error
—
Test error (unseen)
—
Diagnosis
—
The tells of overfitting: a big gap between training and test performance; a model with almost as many parameters as data points; "perfect" accuracy. The fixes: more data, a simpler model, cross-validation, and regularisation. Always report performance on held-out data, never on the training set.
TRAP 3 · DATA QUALITY
GARBAGE IN, GARBAGE OUT
No method rescues bad data
The most sophisticated model in the world, fed flawed data, produces confident nonsense. Before any analysis, audit the inputs. Toggle the problems present in a dataset and watch a defensibility score collapse — each issue maps to a real threat to validity.
Interactive 3a — toggle the issues your data has.
100
Confidence in your conclusions
Interactive 3b — one outlier, one big lie. Drag the rogue point and watch the regression line and mean swing.
Slope with outlier
—
Slope without it
—
The discipline: plot your raw data before modelling; investigate outliers (a data-entry error? or your most important case?); report missingness and how you handled it; and be explicit about who is not in your sample. A limitations section that names these honestly is a sign of rigour, not weakness.
Scroll progress saves on this device. Part of the DBA Data-Analysis series by Jan Erik Meidell.