Interactive pages for this course. Every number on them is computed in your browser — nothing is a recorded animation, and nothing calls out to a server.
The centre and the spread, computed not eyeballed
Identical numbers, different pictures
A positive test, a rare disease
Average a skewed parent, watch it go normal
Read the interval as a multiplier times a standard error
Read power, not just a p-value
Fit a line, then find its influential point
Read a link by its ranks, a shape by its density
Sample a posterior you cannot solve
Split a series into trend, season and noise — then read its memory
A significant p needs a real lift and enough n
Sample, infer, then model
Edit a small dataset (or load symmetric, skewed, bimodal and with-outlier presets) and watch the mean, median, mode, standard deviation, five-number summary and the 1.5×IQR outlier rule recompute, with a live histogram and box plot — the mean chases the tail while the median resists.
Run a seeded coin or die and watch the running frequency converge to the true probability while the raw count difference drifts the other way; then set a medical test’s prevalence, sensitivity and specificity and read the exact P(disease | +) off a confusion table and a probability tree — a 99% test on a 1% disease is right only about half the time it says yes.
Watch a real Metropolis–Hastings and Gibbs sampler converge to a target. Turn the proposal width and see acceptance and mixing trade off — the sample mean and covariance settle onto the true values as the chain runs.
A two-sample A/B experiment: set the control and variant conversion rates and the per-arm size, and read the pooled two-proportion z, the p-value and the unpooled confidence interval for the lift. See the power to detect the lift, the sample size a target power needs, and why peeking at the results early makes a 5% test reject far more than 5% of the time.
Drag points on a scatter plot and watch least squares, the residual and QQ plots, leverage and Cook’s distance all recompute. Two presets share an x value, so their leverage is identical and only their influence differs.
Choose a parent distribution, a sample size and a number of samples, and watch any parent become normal — the course's central claim, drawn rather than asserted.
Effect size, sample size and alpha, with the two error rates moving in opposition and the power curve computed live.
Good charts vs bad charts, made computable: Anscombe’s quartet (four datasets, identical statistics to two decimals, four different shapes), a truncated axis with a live exaggeration factor, and Simpson’s paradox where the pooled slope reverses within groups.
Non-parametric methods, made computable: a kernel density estimate with an adjustable bandwidth (and Silverman’s rule) that finds two humps where one fitted normal sees only one; a LOESS smoother that follows a bend the global line flattens; and Spearman’s rank correlation shown against Pearson’s r — rho stays 1 on y = x³ and shrugs off an outlier that drags r.
Time-series analysis on a seeded trend + season + noise series: an additive decomposition whose trend, seasonal and residual add back to the series exactly; the ACF and PACF with ±1.96/√N bands, where a seasonal spike stands above the band; and an AR(p) forecast fit by Yule–Walker — for AR(1), φ̂ = ρ₁ exactly.