No figure on this site is a drawing that was made once and saved. Each one is a function:
it takes parameters and returns SVG, so the same generator produces the p4 plate and the
p6m plate without either being redrawn.
That is the reason the collection can keep growing without the illustrations drifting apart.
A generator is written once, checked once, and every essay that calls it inherits the same
line weights, the same colour roles, and the same behaviour in dark mode. There are
22 of them so far.
base-rate0 0.250 0.500 0.750 1 how common the condition is chance a positive result is true 1 in 10,000 1 in 1,000 1 in 100 1 in 10 1 in 1 1.8% 66.7% one test, every prevalence the base rate outweighs the test
bootstrap-failsthe sample mean 93.6% nominal 95% the sample maximum 0.0% nominal 95% 3,000 samples of 40 the same procedure, two statistics resampling cannot see past the data
coverage-curve0.400 0.600 0.800 1 0.200 0.400 0.600 0.800 the true proportion actual coverage of a nominal 95% interval nominal 95% Wald — the textbook one Wilson Agresti–Coull Clopper–Pearson summed over all 31 outcomes no interval is 95% until it is counted
coverage-oscillates0.800 0.850 0.900 0.950 25 50 75 100 sample size coverage at a true proportion of 0.15 n = 19 n = 20 Wilson Wald exact coverage at every n from 10 to 120 more data is not automatically better here
forking-paths0 0.200 0.400 0.600 0.800 5 10 15 20 analyses available to the researcher chance of finding something significant the nominal 5% if the analyses were independent correlated, as real ones are no effect present anywhere every individual analysis is correct
middle-and-tailsample size relative error 10 20 40 80 160 320 640 1280 10% 1% at the median two sigma out three sigma out exact binomial against its normal approximation the tail converges last
p-under-the-null0 500 1e+3 0.200 0.400 0.600 0.800 p-value count flat KS distance 0.0165 a p-value that is not uniform is not a p-value
power-curve0 0.250 0.500 0.750 1 0 0.500 1 true effect, in standard deviations probability of rejecting 80%, the usual target 56% at d = 0.5 4,000 simulated studies per point the second number a p-value needs
regression-to-meanfirst measurement second +0.55 -0.59 predicted 0.64 from the correlation alone nobody was treated
same-p-different-meaningten observations 0.758 sd p = 0.04 a hundred 0.208 sd p = 0.04 two thousand 0.046 sd p = 0.04 the effect that gives the same p-value all three reach p = 0.04 the number does not say how big the effect is
simpson-regionfraction of group A given the treatment group B 0.0 0.5 1.0 32% reverse shaded: the overall comparison reverses the rates are identical everywhere here
simpson-tabletreated control verdict group A 93.0% 87.0% treatment group B 73.0% 69.0% treatment both together 78.1% 82.6% control group A: 88 treated of 351 · group B: 263 of 351 wins in both groups, loses overall the rates never change only the allocation does
sums-converge0 0.100 0.200 0.300 0.400 -2 0 2 4 standardised sum density skew 0.747 2/√n = 0.707 40,000 sums, one seed each the rate is predicted, not just the shape
t-against-normal0 0.100 0.200 0.300 0.400 -4 -2 0 2 4 standard errors from the mean density t 2.57 z 1.96 solid: t · dashed: normal 31% wider at 5 df
t-or-zt interval, 7 df 95.0% ± 0.3 at 2 s.e. z interval 91.3% ± 0.4 at 2 s.e. 20,000 samples, one seed each nominal 95% counted, not assumed
tail-ratio1 sigma out 0.982 exact 1.3e-1 2 sigma out 0.691 exact 2.8e-2 3 sigma out 0.335 exact 4.3e-3 4 sigma out 0.104 exact 4.6e-4 1.0 would be exact approximate ÷ exact understated where it matters most
the-density0 0.100 0.200 0.300 0.400 -4 -2 0 2 standard deviations from the mean density 68.27% 95.45% 99.73% the bands are integrated, not recalled sigma = 1.00
thousand-people9 have it, test positive 50 do not have it, test positive 1 have it, test negative 15% of positives are true each dot is one person the arithmetic is not in dispute
twenty-intervalsthe truth, 0.35 0.0 0.2 0.4 0.6 0.8 0 of 20 missed the 95% belongs to the procedure, not to one interval
what-normal-looks-like20 independent samples, 40 points each all twenty are normal this is what the noise looks like
what-width-costsWald 0.241 covers 94.2% Wilson 0.247 covers 96.5% Agresti–Coull 0.259 covers 96.5% Clopper–Pearson 0.273 covers 98.3% expected width, and what it buys orange: fails its nominal level the shortest interval is the one that misses
winners-curse0 250 500 750 1e+3 -0.500 0 0.500 1 effect the study reports studies the truth, 0.3 what gets published, 0.61 shaded: reached p below 0.05 power 20%, inflation 2.04×