Normal Distribution
nothing is called 95% until it has been counted
Theme

The thread: The tail is where it is read

A p-value, a control limit and a risk figure are all tail statements, and the tail is exactly where every approximation in this subject is worst. The central limit theorem converges in the middle long before it converges where anyone looks.
More blocks, and the light tail is called wrong more often. The share of records whose three-way family call is right, over 400 records at each block count, at 100 readings a block. A 95% interval for the shape is formed and the record is recorded as calling Fréchet, Weibull or Gumbel according to whether that interval sits above zero, below it, or straddles it. The two signed parents go from 51.5% and 71.0% at 20 blocks to certainty by 100. The light-tailed one goes the other way — 83.0%, 75.5%, 55.0%, 36.5%, 4.8% — because its estimate sits at about -0.1157 whatever the record length, and a longer record only shrinks the interval onto that number. The tail past the last observation

Three shapes, one limit

A normalised sum has one limit and a normalised maximum has three, indexed by a single number. Twenty blocks put the sign of that number right 97.3% of the time — and naming the family from a light-tailed record gets worse as the record grows, from 83.0% at twenty blocks to 4.8% at five hundred.

 The normal approximation's error on a sum of 100 exponential draws, under its Berry–Esseen bound. The distance between the exact distribution function and the normal one peaks at 0.0133, at z = -0.01. The Berry–Esseen bound is 0.1146, 8.62 times the real worst error, and larger than the whole 2.5% tail a two-sided test reads. Past the first term of the normal approximation

A bound written for a coin

The Berry–Esseen theorem guarantees how far a standardised sum can be from the normal, and the guarantee is true. On an exponential source it is 8.62 times the real worst error at every sample size, the worst error sits at the centre rather than in a tail, and at a hundred draws the bound is larger than the 2.5% tail it would be asked to vouch for.

 A threshold in the tail is a threshold nothing balances. The share of a coin's imbalance in an indicator 1{x > c} that survives a rule which balances the covariate itself. The smooth curve is 1 − ρ² with ρ = φ(c)/√(p(1−p)), a closed form with no trial in it; the points are counted over 500 trials of 200 units at each threshold. At the median the two agree that about a third survives — the removed share is exactly 2/π — and by two standard deviations 86.9% survives. The closed form is exact in the limit and optimistic by a few points at this many units, because the rule balances the sample's mean rather than the population's. The shape the covariate enters by

A threshold in the tail

How much of a threshold's imbalance a balanced covariate removes is a correlation, and the correlation is a closed form. At the median it is exactly 2/π — the same 2/π a median split throws away — and two standard deviations out it is an eighth.

 Where the normal approximation converges, and where it does not. Relative error against the exact binomial. At n = 1280 the error at the median is 0.96% and three sigma out it is 25.7% — a factor of 27. The tail is where the approximation is used. The distribution itself

The tail converges last

The central limit theorem is usually shown as a shape arriving. What the demonstration leaves out is the rate — and the rate is wildly different in the middle and in the tail, which is where every approximation in the subject is actually read.

 One tail arrives; the other is still on its way at a million. The Kolmogorov distance between the exact law of a normalised maximum and its Gumbel limit, at six block sizes, for two parents that both have that same limit. Both are closed form: the exact law of a maximum is F(x)^n and no simulation is involved. The exponential parent's distance falls from 0.0280 to 2.707e-7 — a factor of a hundred thousand, which is exactly one over n. The normal parent's falls from 0.0522 only to 0.0091, a factor of 5.74, because its rate is one over log n. At a million readings a block the two differ by a factor of 33556.3. The tail past the last observation

The maximum converges slowly

The rate at which a normalised maximum reaches its limit law is computable rather than simulable, because the exact law of a maximum is always available. For a normal parent the distance falls like one over the logarithm of the block and is still 0.0091 at a million readings; for an exponential parent, with the same limit, it is 2.707×10⁻⁷.

 The upper tail of 10 exponential draws: normal, one Edgeworth term, two Edgeworth terms, each against the exact tail. Each curve is an approximation divided by the exact gamma tail, so 1 is exact. Six standard deviations out at n = 10: the normal gives ×0.0000670, one Edgeworth term ×0.00159, two ×0.0167 and the saddlepoint ×1.0007. Past the first term of the normal approximation

A correction that goes below zero

One Edgeworth term takes the normal approximation's error at two standard deviations from 38% to 8% on ten exponential draws, and stretches the range within 10% of the truth from 1.66 to 3.09 standard deviations at a hundred. It also turns negative in the short tail at every sample size — past 3.13 standard deviations at a hundred draws and 9.83 at a hundred thousand — because the region recedes only as the sixth root of n.

 A class that is a subspace has no guarantee below its own dimension. Each cell is the worst case over every unit-variance function in a class of dimension m, for a rule reading k functions: the smallest squared principal-angle cosine between the two subspaces. Wherever k is less than m the number is zero to machine precision, and that is not a weak guarantee but the absence of one — some direction of the class is orthogonal to the entire basis, and against an outcome in that direction the rule does exactly what a coin does. An experimenter who declines to name the shapes and asks instead to be protected against everything smooth is asking for the cells above the diagonal. Choosing what the rule reads

Where the guarantee is exactly zero

An experimenter who declines to name the shapes, and asks instead to be protected against anything in a class, is asking for a number that is not small but zero. Bounding the class is unavoidable, and the two ways of doing it choose different bases.

 One cohort of 40, and two intervals around the end of its curve. A single simulated study of 40 subjects with exponential survival at rate 0.35, dropout at rate 0.15 and follow-up to 6 — the first seed from 8811 upward whose plain band reaches below −0.05, chosen to show the failure rather than its frequency. The step curve is Kaplan–Meier and the smooth curve the truth. The plain band, the estimate plus and minus 1.96 Greenwood standard errors, first dips below zero at t = 3.78 and reaches −0.052; early on it also rises to 1.023, above one. At t = 5 the estimate is 0.069 with 1 subject still under observation, the plain interval runs from −0.052 to 0.191 and the log-log interval from 0.006 to 0.251, against a truth of 0.174. The log-log band is built on a scale that cannot leave [0, 1], and it bends away from the edge rather than through it. When the data stops early

The interval at the end of the curve

The interval most software prints around a survival curve covers 89.7% at five years, where 3.3 of forty subjects are still being watched and where the curve is actually read. The same variance carried on a log–log scale covers 94.8% there — and the failure was never the width.

 The upper tail of 10 exponential draws: normal, two Edgeworth terms, saddlepoint, each against the exact tail. Each curve is an approximation divided by the exact gamma tail, so 1 is exact. Six standard deviations out at n = 10: the normal gives ×0.0000670, one Edgeworth term ×0.00159, two ×0.0167 and the saddlepoint ×1.0007. Past the first term of the normal approximation

An approximation built at the threshold

The saddlepoint approximation reads the tail of a sum of five exponential draws to within 0.19% six standard deviations out, where the normal is short by a factor of more than sixty thousand. It is within 2.2% out to ten standard deviations on a single draw, where there is nothing to average, and within 1.1% on a binomial whose expected count is one. It works because it is built where the tail is read rather than at the mean.

 The normal density at sigma = 1.00. The bands hold 68.27%, 95.45%, 99.73% of the mass. Those figures are integrals of the curve drawn, not the memorised 68-95-99.7. The distribution itself

The shape, and where its mass is

68, 95, 99.7 is recited more often than any other set of numbers in the subject. They are integrals of a specific curve, they are worth computing rather than remembering, and the third one is the one people misuse.

 The record stops here, and the curve does not. The level exceeded once in T blocks, against T, for a normal parent at 365 readings a block. The truth is closed form — the block maximum's own distribution function is Φ(x) raised to the 365, so the T-block level is Φ⁻¹ of the 365-th root of (1 − 1/T), with nothing fitted in it — and the fitted mean over 800 records of 50 blocks sits on top of it, 4.0186 against 4.0330 at 100 blocks. What moves is not the level but its error, which grows from 0.0956 at 10 blocks to 0.6383 at a thousand while the level itself moves only from 3.4421 to 4.5454. The rule marks the largest reading an average record contains, 4.0062: everything to the right of where it crosses is read from a fit rather than from data. The tail past the last observation

A level with no data in it

The largest of fifty block maxima is a 51-block event by its own plotting position, so a hundred-block level is read 1.96 times past the longest event the record contains — and it lands above the largest reading on 52.4% of records. The estimate stays nearly unbiased out there; what grows is its error, sixfold from ten blocks to a thousand.

 Four 95% intervals for the mean of 30 exponential observations, split by the side they miss on. Both bars should read 2.5%. The t interval misses below the mean on 6.38% of samples. Widened until its total is exactly 5%, it misses below on 4.69% and above on 0.30%. Hall's transformation misses below on 3.31% and above on 1.96%. Shape, and what it does to a two-sample test

The side a bound is read from

On thirty exponential observations the upper limit of a 95% t interval is exceeded by the true mean 6.38% of the time, against the 2.5% a safety margin set from it assumes. Widen the interval until its total coverage is exactly 95% and the upper limit is still exceeded 4.69% of the time. A symmetric repair fixes the number that is reported and not the one that is used; Hall's transformation, which bends the interval, takes the same rate to 3.31%.

 The factor that holds all of the next m observations with 95% probability, from a sample of 10. From 10 observations, the band for one future value has factor 2.371, for ten 3.716, for a hundred 4.942 and for a thousand 6.008. The 95%-content tolerance factor is 3.382 and is passed by m = 10; the Bonferroni stretch of the prediction factor reaches 7.567 at a thousand. The interval that holds observations, not a mean

All of the next ten

A warranty, a batch release or a monitoring rule promises something about every one of the next ten observations, not about one. From a sample of ten, the band that holds all ten with 95% probability reaches 3.716 sample standard deviations either side of the mean — already wider than the 3.382 of a tolerance interval for 95% of the population — and it keeps widening: 4.942 for a hundred, 6.008 for a thousand, with no ceiling. A 95% prediction interval, read as the answer, holds all ten 67.9% of the time: more than 0.95 to the tenth power, because the ten succeed and fail together.

 Reading the tail of a sum of thirty from one sample of 30, exponential source. Medians over two thousand samples. Six standard deviations out the normal reads ×6.8e-4 of the exact tail, the empirical saddlepoint ×0.061, a gamma fitted by the sample's skewness ×0.185, and a fixed exponential model ×1.000. The bars run from the tenth to the ninetieth percentile of the two readings that estimate the tail's shape. Past the first term of the normal approximation

A tail the sample never saw

Given the source, the saddlepoint reads the tail of a sum of thirty exponential draws to within a fifth of a per cent six standard deviations out. Given only a sample of thirty from that source, the same tilt reads it at a median of 0.061 of the truth, and nine samples in ten read it low. The error has moved from the approximation to the sample: thirty draws put the skewness at a median of 1.30 against a true 2, and hold nothing past their largest value.

 A cut walked from the median into each tail. What a rule balancing a threshold on each of two normal covariates removes of the interaction between the two thresholds, with the cut moved from the 1st percentile to the 99th, on five copulas at a Spearman rank correlation of 0.4. Every copula reads exactly zero at the median. The three radially symmetric copulas are mirror images about it: the Gaussian peaks at 19.49% near the 95th percentile and is down to 13.44% at the 99th; the Frank peaks at 13.22% near the 85th and falls to 2.06%; the t is still near its highest, 37.45%, at the 99th. The Clayton climbs to 56.37% at the 1st percentile, in the tail its dependence is in, and never passes 5.39% in the other; its reflection is the same curve turned round. The other half of the dependence

A cut walked into the tail

Move a balancing rule's threshold off the median and the interaction it leaks rises from exactly zero — and then either turns back to nothing or climbs to a ceiling, depending on one property of the dependence. With both covariates cut, the whole regression collapses to two probabilities, and far out it settles on 2λ/(1 + λ) of the tail-dependence coefficient: 57.25% under a Clayton, 35.57% under a t, and zero under a Gaussian or a Frank however strongly the covariates are related.

 An overlap that depends on where it is read. Each pair's overlap on the doubled log-likelihood scale — the sum of the two searches' own quantiles less the joint search's quantile, as a share of the smaller — at five quantile levels, over 1200 draws under the first-order autoregression. two dictionaries of independent columns: -0.176 at 0.5, -0.000 at 0.75, 0.238 at 0.9, 0.320 at 0.95, 0.538 at 0.99; a break, and an independent column: -1.211 at 0.5, -0.406 at 0.75, -0.029 at 0.9, 0.205 at 0.95, 0.161 at 0.99; two dictionaries of step columns: -0.397 at 0.5, 0.145 at 0.75, 0.438 at 0.9, 0.541 at 0.95, 0.682 at 0.99; a break, and a step column it contains: 1.000 at 0.5, 1.000 at 0.75, 1.000 at 0.9, 1.000 at 0.95, 1.000 at 0.99; a break, and a whitening window: 0.512 at 0.5, 0.640 at 0.75, 0.716 at 0.9, 0.732 at 0.95, 0.754 at 0.99. Every pair but the containment overlaps more at the 95% point than at the median, the control included. Two searches over different features

The overlap at the critical value

Two searches over independent noise columns are additive on average: the joint search finds what the two find separately, to within a fortieth of the smaller. At the point a test actually uses — the joint statistic's 95% quantile — the same pair overlaps by 0.320, and a rule that adds the two searches' own critical values rejects 1.4% of true nulls instead of 5%. Nothing about the searches causes it. Quantiles of a sum do not add, and on every pair in the field the overlap at the critical value is larger than the overlap the field measured.

 The overlap climbs and the interaction does not. The over-charge at each quantile level split into overlap — the two searches' own quantiles less the sequential run's — and interaction — the joint search's quantile less the sequential run's — for three pairs over 1200 draws. two independent dictionaries: overlap -0.40, 0.01, 1.55, 2.44, 5.72; interaction 0.07, 0.01, 0.11, 0.06, 0.00. two step dictionaries: overlap -0.73, 8.19, 22.53, 31.02, 47.75; interaction 4.87, 4.55, 4.90, 4.75, 1.10. a break and a window: overlap 17.88, 29.52, 43.01, 53.79, 71.61; interaction 1.57, 1.29, 1.19, 3.22, 2.75, at the 50%, 75%, 90%, 95% and 99% points. Overlap and complementarity, separated

The split at the critical value

Split the over-charge of two searches at the joint statistic's 95% point rather than at its mean, and the two components behave differently. The interaction — what only the joint search can reach — stays within two units of its average at every level up to the 95% point. The overlap climbs from nothing to most of the over-charge, on two searches over independent columns as much as on searches that share ground. At the critical value the split's overlap has stopped meaning shared ground: it is mostly the arithmetic of adding tails, and the identity that defines it hides that.

 The interval every package reports first does not cover. Counted coverage of two 95% intervals for the 100-block return level of a normal parent, against the length of the record they were fitted from, over 300 records at each length. The level they are about is known in closed form, so this is coverage of a number rather than agreement between two estimates. The delta-method interval covers 80.3% at 25 blocks and reaches only 89.0% at 200; the profile-likelihood interval sits between 94.0% and 94.7% throughout. The gap is not a small-sample effect that lengthening the record removes — it narrows by 8.7 points for an eightfold longer record. The tail past the last observation

Two intervals for one return level

Two 95% intervals read off the same fits of the same records, against a level known in closed form. The symmetric one covers 80.3% at twenty-five blocks and reaches only 89.0% at two hundred — and 99.24% of its misses are the interval sitting entirely below the truth, which is not the endpoint anybody expects to fail.

 Five upper limits for the mean of 15 observations from a exponential source, by margin and by failure rate. t limit: exceeded 8.51% at a mean margin of 0.523; Hall's transformation: exceeded 4.61% at a mean margin of 1.038; gamma family, shape known: exceeded 2.59% at a mean margin of 0.787; gamma family, shape by likelihood: exceeded 4.04% at a mean margin of 0.746; gamma family, shape by skewness: exceeded 17.51% at a mean margin of 0.421. The maximum-likelihood shape has a median of 1.10 and the skewness-based shape 3.20, against a true shape of 1. Shape, and what it does to a two-sample test

A limit from the family it came from

On fifteen exponential observations an upper limit for the mean built from the gamma family's own exact pivot, with the shape estimated by maximum likelihood, is exceeded 4.04% of the time at a mean margin of 0.746 standard deviations. Hall's transformation is exceeded 4.61% at 1.038, and the same family limit with the shape read from the sample skewness is exceeded 17.51%. Knowing the family and estimating its one parameter well recovers most of what knowing the population buys — and the parameter is estimated well only from the logarithms.

 What a corrected critical value delivers: actual size over nominal, upper tail, sum of 30 exponential draws. At a nominal 1% the upper-tail test's actual size is ×1.703 of nominal with the normal critical value, ×0.991 with one Cornish–Fisher term and ×0.998 with two. At 10⁻⁸ they are ×464, ×1.418 and ×0.932. Past the first term of the normal approximation

The same correction, inverted

One skewness term added to a tail probability keeps a sum of a hundred exponential draws within 10% of the truth out to 3.09 standard deviations. The same term used to move the critical value instead keeps the test's actual size within 10% of nominal out to 5.10 — a level of about one in six million. In the short tail it has a flaw of its own: below a level of about one in a million at ten draws the corrected critical value turns back, and a test asked for a level of 10⁻⁷ rejects 1,149 times too often.

 Four rules for a break and a column. How often each of four rules says something is there, for the pair of searches over a break and a column, under four truths, 400 draws each under a first-order autoregression at 0.8 with every threshold simulated on 1200 null draws. nothing there: joint, calibrated 6.3%, summed thresholds 6.3%, Bonferroni 6.3%, either, uncorrected 12.0%. a shift from row 60: joint, calibrated 65.3%, summed thresholds 64.0%, Bonferroni 55.0%, either, uncorrected 69.3%. a column enters: joint, calibrated 11.8%, summed thresholds 10.8%, Bonferroni 73.0%, either, uncorrected 82.5%. both: joint, calibrated 70.5%, summed thresholds 69.0%, Bonferroni 75.8%, either, uncorrected 84.8%. The first block is each rule's size. Two searches over different features

The power a calibration buys

Charging two searches separately makes a joint test conservative, and calibrating the joint statistic on its own null repairs that. Against a real break, on the pair of a break search and a window search, the repair is worth 47 points of power — 62.0% against 14.5%. On the pair of a break search and a column search it is worth nothing, and worse: a real column is found by the calibrated joint test on 11.8% of samples and by a plain Bonferroni split on 73.0%, because a joint statistic takes its threshold from the noisier of its two searches.

 One order's interaction climbs under long memory. The break-and-window pair's interaction — the joint search's quantile less the sequential run's — at five quantile levels, in the field's order (break first, solid) and the other (window first, dashed), under three laws, 1200 draws each. first-order autoregression: break first 1.57, 1.29, 1.19, 3.22, 2.75; window first 0.98, 1.26, 1.53, 2.55, 2.14. five-period moving average: break first 2.00, 1.88, 2.29, 2.63, 0.60; window first 0.16, 0.21, 0.26, 0.29, 0.00. long memory: break first 0.53, 0.80, 1.05, 0.46, 1.35; window first 2.01, 1.62, 4.75, 6.07, 7.85, at the 50%, 75%, 90%, 95% and 99% points. Under two of the laws, and in the field's order under all three, the interaction stays within a few units of its median. Under long memory with the window searched first it climbs from 2.01 to 7.85, against an average of 1.99; the same order's averages under the other two laws are 0.94 and 0.17. Overlap and complementarity, separated

A tail the law gives the pair

Split the over-charge of two searches at their critical value and the interaction — the configurations only the joint search reaches — stays near its average while the overlap climbs. That held under one law. Under a five-period moving average and under long memory it holds for every pair in the order the field reads, to within three units. In the other order, one pair acquires a tail of its own: with the window searched first and the break second, the interaction under long memory climbs from 2.0 at the median to 7.9 at the 99th percentile, four times its average.

 A leak read off a cohort, at two cuts. The threshold rule's leak as a cohort of 1,000 units estimates it from its own two-by-two table: the bar is the middle eighty per cent of 600 cohorts' plug-in estimates, the open circle their median and the stroke the true leak, for five copulas at a Spearman rank correlation of 0.4. At the 90th percentile the estimate is centred on the truth and its middle eighty per cent spans about fourteen points: Gaussian 12.8% to 26.4% around 19.2%, t on four 21.5% to 35.7% around 28.9%, Frank 5.4% to 18.7% around 12.2%, Clayton 0.8% to 9.6% around 4.4%, Clayton turned over 35.9% to 49.5% around 43.1%. At the 99th it is not centred on anything: Gaussian 13.1% to 100.0% around 13.7%, t on four 17.5% to 72.9% around 37.4%, Frank 14.4% to 100.0% around 2.2%, Clayton 16.1% to 100.0% around 0.6%, Clayton turned over 37.0% to 75.7% around 56.3%, because on three of the five copulas between two cohorts in five and four in five have nobody past both cuts. The other half of the dependence

A leak read off a two-by-two table

The share of an interaction a threshold-balancing rule leaves alone needs one number from the data: how often a unit past the cut on one covariate is past it on the other. A cohort of a thousand estimates it to within a dozen points at the 90th percentile. At the 99th most cohorts have an empty cell in their table, and every one of those reads a leak of exactly one hundred per cent — a reading the bootstrap then confirms with an interval of zero width.

 The top of a heavy-tailed population keeps its lead; the top of a light-tailed one gives it back. Select the top share on the first reading and read the group again: the share of its mean lead the second reading keeps, by integration over the true score (lines) and counted on 400,000 draws a parent in 20 batches (points, with two standard errors). The normal keeps exactly 0.6 at every selection. At the top half the Laplace keeps 0.541, the t 0.535 and the uniform 0.648 — the heavy tails keep LESS than the correlation. By the top one per cent the order has reversed: 0.761, 0.784 and 0.443. At one in ten thousand the t keeps 0.977 and the uniform 0.346. Reversals that are not errors

A lead that a heavy tail keeps

Four populations whose readings all correlate at exactly 0.6, and whose least-squares slopes all read 0.6. Select the top one per cent on one reading and measure them again: they keep 60% of their lead if the true scores are normal, 76.1% if they are Laplace, 78.4% if they are a t on four degrees of freedom — and 44.3% if they are uniform. The correlation predicts the regression of the extremes for one shape of population only.

 Estimating P(Z > 5) = 2.8665×10⁻⁷ with plain draws and with four proposals. One seed each. Plain simulation draws nothing past 5 in 100,000 and estimates zero throughout. At 100,000 draws the proposal N(5, 1) reads 1.009 of the truth, N(9, 1) 1.041, N(4.5, 0.25²) 1.039 and N(5, 0.3²) 1.008. Values above 2.2 are drawn at the top edge. What makes it checkable

The draws aimed at the tail

The chance a standard normal exceeds 5 is 2.8665×10⁻⁷, and a plain simulation needs 349 million draws to estimate it to within ten per cent. Draws aimed at the tail and weighted back need 565. Aimed slightly too narrowly, the same method has an infinite variance, an interval that covers 86.0% and gets worse with more draws, and an effective sample size that reads healthier than a proposal that works.

 What a corrected critical value delivers to a t statistic: actual size over nominal, upper tail, 30 exponential draws. At a nominal 1% the upper-tail t test's actual size is ×0.153 with Student's quantile, ×2.94 with one Cornish–Fisher term and ×1.51 with Hall's cubic, both using the source's skewness, and ×0.878 with Hall's cubic using each sample's own. At 10⁻⁶ they are under a hundredth of nominal, ×1.04e+4, ×50.0 and ×1.59. Past the first term of the normal approximation

The correction a t test would use

On thirty exponential draws the one skewness term that held a sum's critical value within 10% out to five standard deviations nearly triples a t test's 1% size when the source's skewness is known, and its upper critical value turns back — at a level of 0.89% on ten draws, where a test asked for one in a thousand rejects 121 times too often. Hall's monotone cubic removes the turn and gains a floor. Built from the sample's own skewness it holds the short side within 20% from 10% down to one in a hundred thousand, and leaves the long side, the one a safety limit reads, at twice its level: the samples that fail are the ones that hide their skewness.

 A cut on each covariate, every pair of them. The share of the product of two threshold indicators that a rule balancing the two indicators removes, in per cent, at a Spearman rank correlation of 0.4, for cuts at the fifth to the ninety-fifth percentile on each covariate. On the Gaussian the far corner, a cut at the 5th percentile on one and the 95th on the other, is 54.8% against 19.5% with both at the 95th. On the t copula it is 1.1% against 34.5%. Both surfaces have their only zero at the double median and leave the diagonal lower in places: moving one Gaussian cut from the 80th to the 70th lowers the share from 14.0% to 12.1%. The same table at seven correlations

A cut on each side

Every table of the threshold's leak put one quantile on both covariates, which is the diagonal of a surface nobody had drawn. Off the diagonal the copulas swap places: a cut at the 5th percentile on one covariate and the 95th on the other removes 54.8% of the interaction under a Gaussian copula and 1.1% under a t copula, the reverse of their order at the near corner. Walked further out, the far corner has two opposite limits, and which one a Gaussian reaches is decided by whether its correlation is above or below exactly one third.

 Which side the misses fall on. Misses of a nominal 95% forecast interval 1 step ahead on 60 observations of a first-order autoregression at 0.7, split by side. normal shocks: the interval as printed 3.1% below and 3.1% above, shocks from the residuals 3.8% below and 4.2% above, the same, coefficient re-estimated 3.3% below and 4.4% above. Student's t on five: the interval as printed 3.3% below and 3.0% above, shocks from the residuals 3.9% below and 3.0% above, the same, coefficient re-estimated 3.6% below and 3.1% above. skewed shocks: the interval as printed 0.0% below and 6.5% above, shocks from the residuals 4.3% below and 4.1% above, the same, coefficient re-estimated 1.8% below and 4.3% above. The observation that has not happened

A tail sixty residuals have barely seen

A forecast interval that reads a normal quantile has every one of its misses on one side when the shocks are skewed, and the natural repair is to draw future shocks from the model's own residuals. It works on the side: 0.0% below and 6.5% above becomes 4.3% and 4.1%. It pays for that with coverage, 93.5% to 91.6%, and on sixty observations it does this under normal shocks too. And a nominal 99% interval has a ceiling nothing in the resampling can lift, because one new shock beats all of sixty old ones in either direction with probability 3.3%.

 The excess over the Gaussian displacement, by window. By how much the expected leverage per parameter exceeds the Gaussian closed form, with the fresh row integrated out, at windows 20, 40, 80, 160, 320, 640. normal: -0.90%, -0.61%, -0.04%, -0.06%, -0.02%, 0.00%. t on eight: 4.14%, 2.78%, 1.79%, 0.73%, 0.28%, 0.18%. t on five: 10.70%, 7.50%, 3.93%, 2.01%, 1.26%, 0.70%. t on three: 44.44%, 32.67%, 22.62%, 16.43%, 11.70%, 7.63%. exponential: 27.91%, 14.01%, 7.21%, 3.53%, 1.93%, 0.85%. On normal predictors the closed form is exact from a window of 80. A search with no fixed point

An excess and its window

The parameter-count displacement is a Gaussian fact, and on heavy-tailed predictors it runs a few per cent over. Swept across windows from 20 to 640 rows, the excess is about the predictors' excess kurtosis divided by the window: a law with a fourth moment of nine on 80 rows is 7.5% over, and on 640 rows under one. At three degrees of freedom it falls only as the root of the window, from 44% to 7.6%. And the instability once blamed on that law was not there: with the fresh row integrated out, the batches agree to a point and a half, and the mean exists.

 Twenty studies of a thousand readings estimate the regression of the extremes: the t, four degrees parent. Each thin line is one study of 1000 readings from the t, four degrees parent: Tweedie's formula with the log-density's slope estimated by a degree-5 log-spline, drawn up to that study's largest reading. The thick line is the share kept by integration over the true score, and the dashed line the correlation, 0.6. The top ten readings of the median study begin at 2.45. Over 400 studies the corrected share kept by the top one per cent averages 0.8190, with a spread of 0.0806, against 0.7842 by integration; a Gaussian kernel averages 0.7914 with a spread of 0.0853. Reversals that are not errors

The slope of a density nobody can see

Tweedie's formula corrects a reading by the slope of the readings' own log-density, and a study has its readings. Estimated from a thousand of them, the correction for the top one per cent beats the correlation's linear rule on 84.0% to 98.0% of studies from heavy-tailed populations and on 75.0% to 81.5% from a bounded one — and costs an error of 0.09 to 0.12 where the population is normal and the rule was already exact. At 250 readings the log-spline loses to the rule it replaces, and at 16,000 the same log-spline gets worse on a power tail.

 A tilt for the ratio, read where it exists. The saddlepoint density of the draw's mean and mean square, integrated over the samples a tilt can reach that put the t statistic past its exact critical value, over the exact tail those same samples carry, at 30 exponential draws. On the short side the relative error runs from +0.88% at 10% to +2.47% at 10⁻⁵; on the long side from +0.83% to +2.05% at 10⁻⁶. Nothing in the construction was fitted to these tails; its error stays within 2.5% of the truth at every level shown. Past the first term of the normal approximation

A tilt for the ratio

A t statistic is a function of two means, so its saddlepoint tilts the joint law of each draw and its square. On an exponential source every such tilt is a truncated normal, and none has a coefficient of variation above one — the exponential is the edge of its own family. Integrated over the samples a tilt can reach, the saddlepoint reads a t statistic's tails on thirty draws within 2.5% from 10% to one in a million on the long side, with nothing fitted. The samples it cannot reach are a third of all samples, they sit at the centre of the statistic rather than in its tails, and a larger sample has more of them.

 The weight each of a hundred outcomes carries in the fitted value at the 97.5th percentile of a lognormal covariate. One design of a hundred lognormal covariate values. The fitted value at a covariate value of 7.10 is a weighted sum of the outcomes, with weights fixed by the design. Under a line and its square it rests on 10.7 effective points, the largest carrying 19.1% of its variance. Under a natural spline with knots at 0.17, 0.57, 1.18, 3.95 it rests on 3.8, the largest carrying 41.1%. What a summary of a scatter is a property of

A value read off a curve

No coefficient of a spline is read; what is read is the fitted curve's value at a covariate value of interest, and that value is a weighted sum of the outcomes with weights the design fixes before any outcome is seen. In the tail of a lognormal covariate a natural spline's value rests on 4.8 of a hundred points, fewer than the squared term's coefficient's 6.9, and the squared term's own value there rests on 10.2 — yet it is three times further from the truth, because its extra points are borrowed from where the curve has a different shape and buy no precision at all. At four hundred points, moving the last knot past the value raises its count threefold and halves its error; doubling the data does neither.

 What it costs to set the t limit at one level without knowing which skewed source the 15 observations came from. For the t limit on 15 observations, charged twenty per standard deviation of shortfall plus its margin, the loss at each nominal level as an excess over the best level for each source. The best levels are 97% for the exponential, 96.25% for the gamma, shape 2, 97.75% for the gamma, shape ½, 98.25% for the lognormal. The minimax level, 97.25%, costs at most 0.8% on any of them: 0.1% on the exponential, 0.6% on the gamma, shape 2, 0.2% on the gamma, shape ½, 0.8% on the lognormal. The conventional 97.5% costs up to 1.0%. Shape, and what it does to a two-sample test

A level set before the source is known

Every best level for an upper limit on skewed data was found knowing the source, and the fifteen observations that would tell the sources apart are the ones being used to set the limit. Chosen instead to protect the worst of four skewed sources, the t limit's level is 97.25% when a failure is charged by the size of its shortfall, and it costs at most 0.8% more than each source's own best — the sources disagree about the best level and agree that missing it a little is cheap. The fitted gamma family, the construction meant to adapt to the source, is the one that needs it: its best levels run from 90% to 96.5% and its minimax level costs 3.4%. Charged a fixed amount per failure, the t limit's minimax level is 99.97% and costs 6.5%; the conventional 97.5% costs up to 119.9%.

 How many people a study has to retest before its estimated correction is worth having, normal errors. Root mean squared error, over 400 studies of a thousand readings at each retest size, of the share of the top one per cent's lead a second reading keeps: Tweedie's correction with a kernel log-slope and the error variance estimated from the retested people (solid), and the linear rule at the reliability those people estimate (dashed). Laplace parent: 0.121, 0.108, 0.084, 0.086, 0.084, 0.078 against 0.343, 0.259, 0.201, 0.183, 0.171, 0.164. T, four degrees parent: 0.133, 0.103, 0.086, 0.084, 0.080, 0.078 against 0.358, 0.278, 0.235, 0.212, 0.205, 0.187. Uniform parent: 0.247, 0.192, 0.156, 0.140, 0.118, 0.117 against 0.260, 0.208, 0.185, 0.166, 0.160, 0.157 — at 10, 20, 50, 100, 200, 1000 retested. Reversals that are not errors

A correction built on a retest

Tweedie's correction for regression to the mean needs the error variance, and a study gets it from a few people measured twice. Twenty of them are enough: with normal errors the estimated correction then beats the correlation's rule on 73% to 77% of studies of a thousand readings from a heavy-tailed population, though its variance is uncertain by a third. What twenty cannot reveal is the error's shape. Laplace errors of the same variance make the correction overstate the lead a top group keeps by about 0.17, the plain rule then wins on most studies, and a study needs about two hundred retest pairs before a check on its own differences notices.

All themes