Skip to contents

Changelog

Source: NEWS.md

drlate 0.3.1

First CRAN release. (The 0.3.0 submission was revised at CRAN’s request to document the return value of the confint() method; the feature set below is unchanged.)

New outcome and treatment model families, completing parity with the Stata lateffects omodel/tmodel options:

  • omodel gains "probit" (binary outcome, probit link) and "flogit" / "fprobit" for fractional outcomes in [0, 1] (e.g. proportions or rates). tmodel gains "probit". The fractional families share all estimation with their binary counterparts and only relax the response to the unit interval. They reuse the probit/logit quasi-likelihood scoring already validated for the instrument propensity score, and the test suite checks them against first principles: the fractional families coincide with their binary counterparts on a 0/1 response, and every fit reproduces the corresponding weighted glm estimate.

Postestimation diagnostics mirroring the Stata lateffects suite (StataNow):

  • complier_means() reports population versus complier covariate means, the complier averages computed with the normalized Abadie-kappa weights (Stata’s estat compliers). kappa_weights() returns those weights (the genkappa object) for use in other complier summaries.
  • balance_test() implements the Imai and Ratkovic (2014) overidentification test for whether the instrument propensity score balances the covariates (Stata’s latebalance overid); cluster-robust when the fit is.
  • balance(detail = TRUE) adds IPW-weighted arm means and unweighted and weighted variance ratios alongside the standardized mean differences (Stata’s latebalance summarize).
  • plot() gains type = "balance_density" (covariate kernel densities by instrument arm, raw versus weighted; Stata’s latebalance density) and a geom = "density" option for type = "overlap" (Stata’s lateoverlap).

These diagnostics are verified against their standard references: the Abadie-kappa identity for the complier means, the Imai and Ratkovic (2014) statistic for the balance test, and the bootstrap for the standard errors.

drlate 0.2.0

  • New kappa-weighting estimators of the LATE from Słoczyński, Uysal, and Wooldridge (2025, JBES 43(1), 164-177): method = "kappa" (tau_a), "kappa0" (tau_a,0), and "kappa10" (tau_a,10), validated against the Stata kappalate command. Cluster-robust SEs, sampling weights, the bootstrap, and (for "kappa"/"kappa0") Fieller confidence sets carry over from the existing machinery.
  • Printed output now shows the kappalate names for the IPW estimators: normalized IPW is tau_u, unnormalized IPW is tau_a,1.
  • New ivmodel = "probit" (kappalate’s zmodel(probit)) for the weighting estimators ("ipw" and the kappa methods), completing coverage of the kappalate command’s options.
  • drlate_compare() now reports each kappa estimator’s own normalization in the normalized column, and ?drlate documents that the kappa denominators are kappa-weight means — estimating the same complier share as the IPW first-stage contrast, but as a different sample statistic.

drlate 0.1.0 (patch)

Changes from an internal econometric audit (Monte Carlo evidence in data-raw/mc-review.R and data-raw/mc-weak2.R):

  • Correctness (deliberate divergence from Stata 1.0.0): the unnormalized LATT-AIPW estimator now computes the treated share with sampling weights. Stata uses an unweighted mean there, which leaves its w1 moment condition nonzero under pweights and invalidates the joint variance; with uniform weights the two coincide exactly (all validated configurations are unaffected).
  • Fieller sets handle the degenerate-quadratic regimes explicitly (half-line when denom^2 = q * V_dd, single-point tangency) instead of collapsing them into “whole line”, and the complement-set print states that the set is unbounded.
  • The weak-instrument advisory now triggers at first-stage F < 10 (|z| < 3.16) instead of |z| < 2, and the printout reports z^2 ~ F.
  • Cluster bootstrap warns when there are fewer than 30 clusters; documentation notes that failed bootstrap draws concentrate where identification is weak (prefer the Fieller set in that regime) and that seeded results are reproducible per fixed number of cores.
  • drlate_compare() documents that IPW/RA rows use reduced adjustment sets (estimator and specification change together) and de-duplicates rows after normalization auto-switching.
  • dr_hausman() documents that the LATT and ATT halves adjust on the instrument- and treatment-equation covariates respectively.

drlate 0.1.0

Extensions beyond the Stata original:

  • Diagnostics: plot() methods for instrument propensity score overlap, covariate balance (love plot), and implied-weight distributions; balance() returns standardized mean differences; print()/summary() report first-stage strength and flag weak instruments.
  • Bootstrap inference: drlate(vcov = "bootstrap") provides nonparametric bootstrap standard errors and percentile confidence intervals (cluster bootstrap when cluster is supplied), with optional parallelism.
  • Weak-instrument-robust inference: confint(method = "fieller") inverts the joint test of the numerator and denominator, returning bounded, complement, or whole-line confidence sets as appropriate.
  • DR Hausman test: dr_hausman() implements the doubly robust test of unconfoundedness from Słoczyński, Uysal & Wooldridge (2022, Section 5) — proposed in the paper but not available in the Stata package — with an analytic standard error from one jointly stacked moment system.
  • Estimator comparison: drlate_compare() runs IPWRA/IPW/AIPW/RA in one call with a comparison table and dot-whisker plot.
  • Documentation site: pkgdown website with the primer vignette featured, deployed via GitHub Actions.

drlate 0.0.0.9000

  • Initial R port of the Stata package drlate v1.0.0 (SSC S459708).
  • LATE and LATT estimation via IPWRA, IPW, AIPW, and RA.
  • Linear, logit, and Poisson outcome/treatment models.
  • Logit (MLE), CBPS, and IPT instrument propensity score models.
  • Joint sandwich inference across all estimation stages, with robust and cluster-robust variants and sampling weights.
  • Overlap diagnostics (pstolerance, osample).
  • Golden-fixture test harness for numerical equivalence with Stata: estimates match to ~1e-9 and standard errors to ~1e-6 across all 33 validation scenarios.