Skip to contents

Changelog

Source: NEWS.md

streetlamp 0.2.0

  • lamp_map() draws the geography the package has carried since the start and never showed: a choropleth of any panel column, of the per-area slopes of a mean group elasticity, or of the donor weights behind a synthetic control. The rule that governs the rest of the package governs the map too. An area with no usable data is filled in a neutral grey and counted in the caption, never shaded as though it were quiet, and an area the panel never covered is counted separately from one whose files were missing, because they are not the same thing. Fills are sequential, or diverging and centred on zero where the values cross it, from the same palette as every other figure.

  • lamp_elasticity() defaults to clustered Poisson pseudo-likelihood for the fixed-effects method. This estimates the conditional count mean and avoids treating log-one-plus observed counts as log expected counts. Use family = "ols_log" for the former response scale; CCE still uses that scale. Calendar lags now preserve missing-month gaps. Lag-sum intervals use joint covariance and the backend’s reference distribution.

  • lamp_simulate() supports cohort-specific effects, dynamic effect paths, overdispersed counts, selective missing submissions, and deliberate trend violations. An additive count design supports count-level parallel trends. lamp_simulation_truth() exposes exact area-month counterfactual means on either the count-difference or log-mean scale.

  • lamp_effect_summary() uses joint covariance for effect combinations and each staggered backend’s own overall ATT. lamp_trend_sensitivity() reports conditional sensitivity to specified differential linear trends.

  • lamp_design_audit() records data support and identification questions. lamp_elasticity_robustness() retains complete specification grids, failures, and optional leave-one-area-out results.

  • Crimes-prevented conversions account for log-one-plus stops, use the fitted sample, preserve covariance in event averages, distinguish count-level effects, and transform intervals directly. IHS effects require predictions and are rejected by the simple conversion.

  • New research validation and a hash-verified London association study are separate from historical 0.1.0 results. London is explicitly observational; its comparison-outcome and future-exposure diagnostics are retained.

streetlamp 0.1.0

  • One visual system for every figure and table. lamp_theme() and lamp_colours() are exported and every plot() method uses them: a recessive hairline grid, left-aligned titles, forces and outcomes named in words, thousands separators on axes, and a palette checked as a set for colour-vision deficiency. Event studies, staggered fits and pre-trend diagnostics draw an interval ribbon with the reference period hollow; coefficient plots print the estimate and interval beside each mark; the placebo histogram labels the actual estimate and reports its p value; the synthetic control path labels both series at the line ends and shades the pre-period like the event study does.
  • lamp_table() formats any estimate, pre-trend diagnostic, synthetic control or coverage audit as a presentation table: terms in words, the 95 percent interval in brackets, p values as <0.001 where they are that small, ready for knitr::kable(). The print methods and the report use it, so a console print no longer shows p values in scientific notation.
  • lamp_report() HTML output is styled by a bundled stylesheet, embeds each estimate’s plot (new figures argument), and renders tables with numeric columns right-aligned.
  • The plot() methods for estimates and staggered fits wrap the identifying assumption in the subtitle instead of letting it run off the panel, and lamp_elasticity() no longer leaves a stray space before the full stop in its assumption text.
  • First release. Builds an area-by-month panel of recorded crime in England and Wales from the data.police.uk archive, attaches stop and search intensity as a treatment, and estimates its effect with panel econometrics. Every panel carries a contract recording provenance and coverage; every estimator names its identifying assumption and reports the force-months it excluded.
  • The report builder (lamp_report()), the methods note in inst/NOTES/methods.md, validation scripts in inst/scripts/ for the simulation study, placebo battery, benchmark and fixed-specification reproduction, and five precomputed vignettes.
  • Difference-in-differences with staggered adoption over three backends (lamp_did_staggered()); synthetic control with permutation inference and a dependency-free solver (lamp_synth()); joint own-area and neighbour effects (lamp_spillover()) and the weighted displacement quotient (lamp_displacement_quotient()); stop-crime elasticities with common correlated effects and Pesaran’s CD test (lamp_elasticity()); the allocation model that measures how searching follows crime (lamp_allocation()); and the conversion to crimes prevented per thousand searches with its assumption chain (lamp_crimes_prevented()).
  • Treatment definitions for continuous intensity, binary interventions, staggered adoption and dated shocks (lamp_treatment()), with surge detection (lamp_detect_surges()) documented as endogenous; two-way fixed effects (lamp_twfe()) and event studies (lamp_event_study()) over four families, with clustered standard errors, dispersion and Moran’s I of the residuals; pre-trend testing with the power to detect a linear trend and the bias it would leave (lamp_pretrends()); placebo tests in time, space and outcome (lamp_placebo()); and a simulator with known effects, spillovers on a lattice and force-month missingness (lamp_simulate()).
  • ONS boundaries by paged FeatureServer requests (lamp_boundaries()), spatial adjacency (lamp_adjacency()), Census 2021 population from NOMIS (lamp_population()), LSOA re-vintaging (lamp_revintage()), the area hierarchy (lamp_area_lookup()) and deprivation deciles (lamp_deprivation()); the balanced area-by-month panel (lamp_panel()) with NA, never zero, for force-months without a file; the coverage audit (lamp_coverage(), lamp_coverage_compare(), lamp_changelog()) with partial, refreshed and mismatched-file-type flags and a heat map; and the bundled sample panel and boundaries (lamp_sample_panel(), lamp_sample_boundaries()).
  • Selective download from the data.police.uk archive by HTTP byte range (lamp_archive_index(), lamp_archive_download(), lamp_archive_register(), lamp_archive_snapshot()); one archive version per force-month file with alternatives recorded (lamp_list_versions(), lamp_select_version(), lamp_version_diff()); readers with a fixed schema and a panel contract (lamp_read_crime(), lamp_read_outcomes(), lamp_read_stop_counts(), lamp_contract()); LSOA vintage detection (lamp_lsoa_vintage()); the forces table (lamp_forces()); and two bundled miniature archive snapshots for offline examples and tests.
  • Package skeleton, MIT licence, continuous integration, testthat, pkgdown, lintr; bundled crime and outcome classifications, national policy shocks table, bank holidays table, and ONS LSOA 2011 and 2021 code lists with the 2011 to 2021 best-fit lookup.