suplrn() stacks the learners of the package (least squares or logistic regression, the lasso, ridge and elastic net, the random forest) into a super learner: K-fold cross-fitted predictions, weights by non-negative least squares that sum to one, or the discrete super learner (the learner of least cross-validated risk), for numeric and 0/1 outcomes, the folds drawn from a seed and returned. With one learner the stack is that learner. dmlx() takes it as a nuisance learner (lrn = "stack", with stk, nstk and disc).
Ctrl-C (Esc in the R GUIs) now stops every long computation: the resampling, permutation, bootstrap and simulation loops, the iterative fits (elastic net, logistic and count regressions, EM, quasi-Newton, Weber point, k-means, power iterations, Louvain, synthetic control weights, fixed-effects sweeps) and the scans over pairs check for an interrupt every few milliseconds of work, free what they hold and end with R’s interrupt, never with a half-made result. In ‘Python’ the same loops raise KeyboardInterrupt.
etasf() fits the temporal ETAS model (marks and the modified Omori law), mvhwk() the multivariate Hawkes process of several types that excite each other, with its excitation matrix and spectral radius, and sthwk() the space-time Hawkes process of crime, with intensities at query points; hwksim() simulates from a space-time fit. proof("etasf") names the theorem that ties the Omori kernel to the Lomax kernel of hwkfit().
sdbsim() simulates the Short-D’Orsogna-Brantingham model of burglary hot spots, as agents on a lattice or as its reaction-diffusion limit with suppressed sites, and sdbthr() gives the linear stability of its uniform state: the threshold beyond which hot spots form and the fastest-growing wavelength. proof("sdbthr") and proof("sdbsim") name the Lean theorems on the steady state, the trace and the decay.
hypq() solves Larson’s hypercube queue of patrol units exactly (workloads, dispatch shares, lost or waiting calls, travel times) and sqrtlw() gives Kolesar and Blum’s square-root law of response distance; proof("sqrtlw") names its scaling theorems.
kmrec() gives Kaplan-Meier curves by group with Greenwood’s standard errors, survfit()’s log-scale intervals and medians, and the log-rank test; with unit it takes event times and studies the gaps between a unit’s events (repeat incidents at a place), the last one censored.
trndrr() measures the trend in a series of counts as a Poisson rate ratio per period (quasi-Poisson when overdispersed), by Mann-Kendall’s test (exact up to 8 periods) and the Theil-Sen slope, with the change on the period before and a test of equal seasonal totals; stpcp() finds a step change with its permutation p-value and Pettitt’s test; cusrv() watches counts by CUSUM; xcser() correlates two series, matched by period name, at lags as ccf() does.
hwkfit() fits a Hawkes (self-exciting) process to event times by maximum likelihood with the analytic gradient: exponential, Weibull or Lomax kernels, a constant baseline or a trend with a cycle, several of them compared by AIC, and the branching arithmetic (cluster size, stationary rate, share triggered) that proof("hwkfit") names. hwkres() gives the time-rescaling residuals with the Kolmogorov-Smirnov test of ks.test(), and hwksim() simulates the process, in time or in space and time, by its branching construction.
crsev() builds a severity-weighted crime index from counts by offence (the Crime Severity Index of Statistics Canada: weighted counts per population against a base, the weights given or made from incarceration rates and mean sentences); zcomp() ranks units on a composite of standardised counts across categories.
cconc() measures how concentrated a total is across units: the Gini coefficient and its exact split into the share of units at zero and the Gini of the rest, the Lorenz curve, top shares, dispersion against the Poisson, the bound chance alone allows and a simulated Poisson null, the Gini within groups, the urban-scaling exponent, and for the places of events in order the growth of distinct places. hilltl() estimates a power-law tail index: the discrete maximum likelihood estimate through the Hurwitz zeta function, or the Hill estimators, with the Kolmogorov-Smirnov distance of the fit. proof("cconc") names the Lean theorems on the split at zero.
crcap() estimates a hidden population from two lists (Lincoln-Petersen with its bounds over a box of list dependence, Chapman’s estimate with Seber’s standard error) or three (the never-listed cell under no three-way interaction, its sensitivity to that interaction, and the pairwise estimates); dkfig() bounds the dark figure of crime: the true victimisation rate under limits on a survey’s false positives and negatives with the breakdown point of a conclusion, the total when a survey covers only some offences, and the offences behind incidents counted under the hierarchy rule. proof("crcap") and proof("dkfig") name the Lean theorems.
bandx() reads band labels (“1-5”, “Over 60”, “Under 10”) into bounds, places a value in each band, expands a banded frequency table and shows how much a statistic depends on the cap assumed for an open band; pbbnd() gives the bounds a rounded or suppressed published cell allows and the envelope of the change between two published columns; yoypx() gives year-over-year (or any period) change within groups with the exact interval of the percent change, exact p-values adjusted for multiplicity, and the envelope that rounding allows. proof("pbbnd") names the Lean theorems on the envelope.
jdgiv() gives the judge-leniency instrument’s first stage, intention-to- treat effect and Wald ratio with the compliers’ effect over a grid of defier shares and effects, or the exact decomposition by type on a known population; jmono() gives each judge’s decision rate with a chi-square test of equal rates, the many-judge slope test of nested judges, and the comparison of judges within groups of cases with adjusted p-values and slopes; mmshr() gives the shares of sentences at, below and above a mandatory minimum. proof("jdgiv") and proof("jmono") name the Lean theorems.
agemix() gives the aggregate age-crime curve of a mixture of latent types and the covariance across types between ages that a single type forces to 0, with its permutation test on panel data; cumrsk() gives the cumulative risk of a first arrest, conviction or admission by age from a life table of first events, with Greenwood’s interval or with deaths competing; jpipe() gives rates between stages of the justice pipeline from window sums lagged by the time between the stages. proof("agemix") names the Lean theorem.
mndla() classifies segregation placements under the Mandela Rules (more than 15 days, or Rules 43 and 44 with hours out of cell) and gives the qualifying share by group over placements, people with any qualifying placement and people by their cumulative days, with exact intervals.
stkflw() relates person-days, the average daily population, admissions and the length of stay (any two give the rest) with changes, rates and the stock an earlier release removes; lostay() summarises lengths of stay, their bands and thresholds and the lognormal, exponential and Pareto fits; bklog() applies Little’s law to a docket; dktflw() gives clearance, rolling clearance, the pending caseload, Little’s law on published figures and the trend of the outcome mix; lftab() builds the actuarial life table; trnmat() gives the transition matrix between states with its stationary distribution. proof("stkflw") and proof("bklog") name the Lean theorems.
cntreg() fits a count on a group by Poisson and negative binomial regression (the iterations of glm() and MASS::glm.nb()), with adjustment columns, an exposure offset and clusters, and reports the rate ratios with Wald intervals: model-based, cluster-robust (the sandwich of sandwich::vcovCL()), and from an exchangeable GEE (as ‘geepack’ estimates it). proof("cntreg") names the Lean theorems on swapping the groups.
grprr() gives rates or proportions by group and their ratios and differences against a reference group: Wilson, Clopper-Pearson and exact Poisson intervals, log-normal and exact conditional intervals of the ratio, Newcombe’s interval of a difference of proportions, and each group’s over-representation among the events. dastd() gives directly standardised rates with Fay and Feuer’s interval, and their ratios and differences; sirex() gives standardised incidence ratios with exact intervals, funnel limits and Marshall’s empirical Bayes shrinkage.
ctabx() cross-tabulates labels (counting rows, summing a count column, taking the largest or smallest value, or counting distinct ids per cell) and tests the table: Pearson’s chi-square (with Yates’ correction on request), the likelihood-ratio test, Fisher’s exact test with the conditional odds ratio of a 2 x 2 table, Cramer’s V with its noncentral chi-square interval and Bergsma’s bias correction, the contingency coefficient, phi and the diagonal share. proof("ctabx") names the Lean theorems on the 2 x 2 statistic.
clprep() maps a published dataset’s columns onto a generic function’s arguments from inst/presets/presets.tsv (one row per preset and argument: column, recode, filter, units, note, source), with recipes for derived columns, so that generic(clprep(data, preset)) gives the published analysis. The presets cover the Ontario segregation and deaths-in-custody tables, the race-based use-of-force files, the Special Investigations Unit cases, the Structured Intervention Unit tables and the Toronto incident layers. Presets also feed the risk-tool audits: fairdf(), tmpaud(), calaud() and scrdsp() take a clprep() result as their first argument, for a generic file of area, risk, outcome and group, for fairsim() records, and for the layouts of two published studies of risk classification.
Audits of risk tools and predictive policing. fairdf() gives the parity gap and disparate impact ratio between groups with Newcombe and Katz intervals, the gaps in true and false positive rates and the average odds difference with their intervals, and the bias amplification score with a bootstrap interval; tmpaud() the same by city and period with each city’s range. calaud() rolls records up to areas and asks whether the order of predicted risk matches the order of recorded outcomes (Spearman’s rho with the AS 89 and large-sample p-values) and whether the mean rank gap of each group’s areas departs from chance (a permutation test); scrdsp() a score’s spread across groups with the one-way analysis of variance and Welch’s test. lipaud() is the Lipschitz audit of individual fairness. nsbnd() bounds a true base rate behind a noisily recorded one for noise rates in boxes, and says which groups’ order survives every admissible noise; impfpr() is Chouldechova’s identity; rnkres() the pairs a noisy ranking orders. oaxbd() is the Oaxaca-Blinder decomposition of a gap at either group’s coefficients and the threefold one, with delta-method standard errors. fairsim() simulates records with a known disparity and nordet() the noisy-OR chance of detection by patrols. proof() names the Lean theorems behind nsbnd(), impfpr(), rnkres() and oaxbd().
relchk() recomputes a reported odds ratio, risk ratio or risk difference from its table under every labelling (each reference group, each outcome as the event, groups in rows or in columns) and says which readings reproduce the value as rounded: the labelling the report states, another one, or none. Zero cells give Inf or NaN, never a correction. proof("relchk") names the Lean theorems on how the contrasts change under relabelling.
catchk() audits categorical inputs before an analysis: storage type, levels and counts, missing values, unused and rare levels, labels that differ only by case or white space, 0/1 and logical codings, numbers that look like category codes, value labels that disagree with the values, and the reference level a default would take and why. It can check a stated reference level, and changes nothing.
clrsmp() splits a change in a clearance rate into rate and mix effects and flags Simpson’s paradox.
expbnd() bounds rates and disparities when exposure is known only through a proxy.
misrd() gives the risk difference when group membership is misrecorded.
rtmhs() measures regression to the mean at selected hot spots.
incap() gives the crime rate and share prevented under incapacitation.
incapr() gives crimes prevented under desistance and replacement.
sellab() bounds the failure rate of a release rule from released defendants alone.
losbnd() bounds the mean time in custody from completed and open stays.
sntbnd() bounds the effect of a sentence on a binary outcome without assumptions.
hzdsel() shows how period-two hazards differ after selective depletion.
Networks: coofg() builds the co-offending graph of an incident table; ntdeg(), ntbtw(), ntcls() and ntevc() give degree, betweenness, closeness and eigenvector centrality; ntcore() cores, ntcmp() components, ntlvn() and ntmod() Louvain communities and modularity, ntbrk() Gould-Fernandez brokerage and ntchg() the Cheeger bounds. Results are by node label in order of first appearance.
Spatial weights and autocorrelation: spwt() (k nearest, distance band, contiguity; binary or inverse distance), gearc() (Geary’s C), gorg() (the Getis-Ord G), lmorn() (local Moran with conditional permutations), locgi() (local G and G*) and dirmn() (directional mean).
hdbang() (head-bang smoothing, weighted or not), crwalk() (a correlated random walk of a series: fit, expected displacement and simulation) and rnnhcl() (nearest-neighbour clustering against a population at risk). nnidx(), ripk(), fzmode() and mantst() take metric = "manhattan"; fzmode(), kmpart(), nnhclu() and stacel() take zones as counts at centroids (count).
proof() names the Lean theorem behind each function; the proofs ship in inst/lean.
A vector longer than the largest integer is refused with an error, by the exported functions and by the compiled entry points alike; the compiled code used to truncate its length.
eceplt() gives the calibrated curve its limits at extreme prior odds: a calibrated ratio of 0 or infinity used to give NaN, or an error.
incapr() refuses a t0 + s beyond the length of lam however large the two are; past the largest integer it used to warn of an integer overflow and stop with an unrelated error.
sntbnd() orders the two values of z by Unicode code point, so the default treat and the order of the terms are the same on every machine and in every locale ("B" sorts before "a").
Every error message starts with the name of the function that raised it ("outtst: `q` must be at most 1."); the words after it are unchanged.
A matrix with more than one row and column is refused where a vector is expected, naming the argument; it used to be read as one long vector.
Arguments that hold one value per group are matched by name. Named inputs whose names come in another order were read by position, a wrong answer: vodtst(c(white = 600, black = 400), c(black = 300, white = 700)) read 300 as the white night count. Now every later argument is put in the order of the first by name; names that differ, or a name given twice, are an error that lists them. Unnamed inputs are still matched by position. This holds in vodtst(), dspbm(), dyadrt() (the row and column names of count too), clrsmp(), expbnd(), incap(), gtmp(), hzdsel(), and for loci in kinidx() and patidx(); per-unit arguments of outtst(), rtmhs(), lrpost(), sellab() and sntbnd() are matched by name when every one has unique names.
Per-group arguments also accept a list of single numbers and a one-row data frame; a one-row or one-column matrix gives its names.
vodtst() takes one table of groups by day and night (a matrix, 2-d table or data frame with columns named day and night) with night left out.
vodtst(), dspbm() and expbnd() return a ref column marking the reference group; a factor ref is read by its label.
sntbnd() no longer picks the treatment by sorting the values of z: treat is required unless z is logical (TRUE is the treatment) or 0 and 1 (1 is).
A labelled vector (with a labels attribute, as haven makes) is refused with a request to decode it first: its numbers are category codes.
Learning: enpath(), the elastic net path by coordinate descent with cross-validation over given folds (the objective and scale of glmnet); lgtreg(), logistic regression; rfor(), random forests for regression and probabilities.
Causal estimates: dmlx(), double machine learning by cross-fitting for the partially linear, interactive, IV, difference-in-differences and sample-selection models, with AIPW and TMLE from the same nuisances.
enpath() screens binomial paths by the sequential strong rule (scr); rfor() grows honest trees (hon), limits split imbalance (imb) and takes mtry by the rule of grf by default; dmlx() adds two-way clusters, the experimental DiD scores (score = "exp") and passes tol, scr and hon to its learners.
Policy evaluation on unit-by-period panels. didfe() is the two-way fixed-effects difference in differences by least squares (alternating projections, weights, one-way or two-way cluster-robust errors with the small-sample factor of ‘fixest’), or by Poisson or negative binomial regression for counts, with the estimate leaving out each treated unit, the Goodman-Bacon decomposition, and matching on the outcomes before treatment. evtst() is the event study with the joint test of the leads and the raw trends of each adoption cohort. stgdid() gives the group-time effects of Callaway and Sant’Anna, with the doubly robust, IPW and regression estimators for covariates, Wooldridge’s regression on cohort-period cells, and de Chaisemartin and D’Haultfoeuille’s estimator for treatments that switch on and off, with cluster-bootstrap errors. synthc() is the synthetic control of one treated unit with the placebo test, and sdidx() the synthetic difference in differences with placebo or jackknife errors; both solve their weights exactly (Wolfe’s minimum-norm point). evtst() bounds the effects under departures from parallel trends (Rambachan and Roth’s relative magnitudes) with a breakdown value. wcbx() gives the restricted wild cluster bootstrap (every sign pattern when few), permutation of adoption dates across clusters, and simulated power with the minimum detectable effect; wystep() the Westfall-Young stepdown p-values.
Spatial crime analysis. knxnr() is the Knox test of near repeats: pairs by band of distance and time against their exact expectation under exchangeable times, with permutation and Poisson p-values. kdsurf() gives kernel intensity surfaces (the bandwidth and values of MASS::kde2d(), the edge corrections of spatstat.explore::density.ppp()) with the smallest areas holding each share of the incidents; dbscn() DBSCAN clusters as dbscan::dbscan() numbers them; hsper() hot grid cells per period, their overlap and how long each stays hot; ebshr() the empirical-Bayes shrinkage of hot-spot counts and the fall to expect without any programme; kscan() Kulldorff’s space-time permutation scan with secondary clusters. gmoran() and bvmor() give global and bivariate Moran’s I with their inference (spdep::moran.test(), spdep::moran_bv()), one variable or a sweep of columns; sarml() fits the spatial lag, error and combined models by maximum likelihood, and the lag model by two-stage least squares, with Durbin terms and the OLS fit beside (spatialreg::lagsarlm(), errorsarlm(), sacsarlm(), stsls()); spill() gives direct, indirect and total effects with delta-method standard errors (spatialreg::impacts()). crossk() gives the cross-type K function and the kernel segregation test, qdrtst() the quadrat test, morsta() Morisita’s index, dissix() the dissimilarity, entropy, Atkinson and Gini indices with Morrill’s adjustment. pinpl() puts points in polygons and reports the population their regions cover, arwt() moves counts between two sets of areal units, and bufas() assigns treated, comparison and excluded places around sites. proof("ebshr"), proof("kscan"), proof("dissix"), proof("spill") and proof("bvmor") name the Lean theorems in inst/lean/Hotspot.lean.
Design-based causal effects of a 0/1 treatment. cbatt() gives Hajek-normalised inverse probability weighting, g-computation, subclassification and caliper matching on the propensity (ATE, ATT or ATC), and entropy balancing for the ATT, with bootstrap errors; smdbal() the standardised mean differences and variance ratios before and after weighting with the overlap of the propensity; evlx() the E-value of a risk, odds or hazard ratio or a standardised difference.
svymx() gives design-based means, totals, proportions, ratios and variances from stratified, clustered and weighted samples, by domain and with differences between domains, with linearisation errors, the finite-population correction and a description of the design.
