Repository navigation
Audit fixes; release R 4.1.1, Python 2.1.1, Stata 11.1.1 - #25
Merged
Merged
Conversation
Closes audit items XL-1 (R half), R-2a/2b, R-3, R-4, R-5, R-8, R-9, R-10, R-12, R-17 (partial), NEW-1, NEW-2, NEW-5 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. - XL-1: restore the missing factorial(deriv) in the sharp+covariates branch, in the tau block *and* in beta_Y_p_l/r, which had the same omission. Every other branch already carried it, so fixing tau alone would have left the package internally inconsistent. At deriv=2 the estimate was half the truth. - R-2: the effective-N block compared STANDARDIZED X_l/X_r/c against a bandwidth already restored to the original scale, so every observation fell inside the window and the full sample was reported (613/587 vs 161/134). - R-3: the restore block put c, X_l/X_r and Y_l/Y_r back on the original scale but not x and y, so the fuzzy re-split compared two different scales and crashed whenever c != 0. - R-4: the N<20 fallback returned a bandwidth still on the standardized scale. - R-5/R-9/R-12: normalize string options before validating, so "NN"/"TRI" and friends no longer silently take a different code path. - R-10/NEW-2: validate nnmatch and reject collinear covs under covs_drop=FALSE instead of surfacing a raw LAPACK message. - NEW-1/NEW-5: guard degenerate bandwidths and empty-side windows rather than failing later with "missing value where TRUE/FALSE needed". - R-8: document that vcov() off-diagonals are placeholder zeros; the design matrices needed for the cross-covariances are not retained on the fit. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Closes PY-1, PY-2a/2b, PY-4, PY-7, R/PY-12, NEW-1 and the Python half of XL-1 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. - PY-1: sharp/fuzzy deriv>=2 always raised UnboundLocalError after all the computation had completed -- the deriv>=2 branch of the rdmodel assignment was dropped in the port from R. Rewritten to mirror rdrobust.R:834-852. With this, the 3-language baseline goes from 72/85 to 85/85 computable specs with zero movement in previously-working cells. - PY-4: the N<20 fallback called np.max(range_l, range_r), which passes range_r as the AXIS argument and raised TypeError every time it fired. - PY-2: same standardized-vs-raw effective-N bug as the R side. - PY-7: rdplot(kernel="TRI") estimated triangular but labelled the output Uniform. - Shared option normalizer/whitelist so casing no longer changes behaviour. Note: the njit fallback in funs.py must stay inside its except ImportError block; an earlier edit orphaned it after a raise, which was latent only because numba happened to be installed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…r codes
Closes ST-1, ST-2, ST-3, ST-4, ST-5, ST-7, ST-8 and the Stata half of XL-1
from AUDIT_FINDINGS_rdrobust_2026-07-20.md.
- ST-7 (user namespace destruction): st_matrix("b"/"V") created GLOBAL
matrices literally named b and V, and they were dropped afterwards, so a
user matrix of either name was gone for good; a user scalar rho was
clobbered too. Now tempname bmat/Vmat plus a local for rho. Note there were
three further bare references (matrix colnames b / rownames V / colnames V)
that only surfaced AFTER the results table printed.
- ST-2: dups/dupsid were declared as tempvars but gen'd and read under their
LITERAL names, so a user variable of either name broke every default
vce(nn) run behind a misleading rc=3499.
- ST-1: M_l/M_r were set only inside the auto-bandwidth branch, so
h() + masspoints(check) died rc=111 -- or, after a previous run, silently
displayed stale counts.
- ST-8: rho() silently overrode an explicit b(), and a positive rho below
5e-5 rounded to zero and was ignored.
- ST-3: stdvars(ON) and other non-lowercase values silently behaved as off,
defeating the scale default (verified 43x bandwidth difference).
- ST-4: Mata exit(1) -> exit(506); rc=1 is Stata's --Break--, so under qui the
user saw only "--Break--" with no hint of the cause.
- ST-5: rdbwselect printed a blank cutoff because `c_orig' was never defined.
- XL-1: the missing factorial(deriv) in the sharp+covariates branch.
Only the first token of vce() is lowercased: tokens 2-3 are a cluster
variable name and an nnmatch count, and Stata variable names are case
sensitive.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The package shipped without tests. This adds 110 testthat tests covering the behaviours the 2026-07 audit found broken, so they cannot silently regress: bandwidth selection, covariates, fuzzy, scale invariance, mass points, numerical contracts, point estimates, print/summary, rdplot, input validation, variance, clustered vce, and weights. test-known-bugs.R is the deliberate home for defect witnesses: each test either asserts the fixed behaviour or is skipped with the open defect ID, so the file doubles as a live checklist. Wires up testthat edition 3 in DESCRIPTION. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ypos Closes ST-6, ST-9 and ST-10 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. 3-language baseline re-run after these edits: 85/85 + 150/150 in all three languages, 0 discrepant cells. ST-6. `touse' came from marksample and marked the [if]/[in] sample only, while every missing-value drop happens inside the work frame -- so e(sample) claimed rows that never entered the estimation (1390 marked vs e(N)=1244). The drop_cond local already gathers every condition in one place and is written over variables that exist in the caller's frame, so it is re-applied to touse just before ereturn post. e(sample) == e(N) in 4/4 configurations: plain (2740), covs (2670), vce(cr1 cl) (2690), and under an if (1840). Same one-line approach as rdhte ST-8. ST-9. rdplot is eclass and therefore clears e(), so after one rdrobustplot the "must follow rdrobust" guard failed and a SECOND call died rc=301. The estimates are now parked with _estimates hold across the delegation and restored afterwards, with rdplot's rc captured and re-raised in between so a genuine failure still surfaces. Three consecutive calls now all return 0, e(cmd) is still rdrobust afterwards, and the guard still fires rc=301 after an intervening regress. Of the seven accepted-but-dead options, five -- scale, xtitle, ytitle, xlabel, ylabel -- are ordinary twoway options and are now folded into the graph_options() handed to rdplot, so they take effect. scale was the most misleading: its only appearance was a no-op self-assignment. col_dots() and col_lines() are removed instead: they are per-plot marker and line colours and rdplot builds its own twoway call with no hook for them, so they could never have worked; col_dots(red) now returns 198 rather than being ignored. ST-10. rdplot's trailing `*' collected unknown options into `options', which the ado never reads (0 references) -- so a typo was accepted and silently ignored while rdrobust and rdbwselect correctly rc=198. Removed; twoway options keep their own graph_options() slot. A genvars rerun now names the collision and the drop command instead of a bare rc=110. Deliberately NOT a blanket `capture drop rdplot_*' -- that is the destructive pattern fixed in nprobust lprobust.ado ST-12. A user variable rdplot_myvar sharing the prefix survives untouched while genvars completes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Closes ST-13 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. 3-language baseline after these edits: 85/85 + 150/150 in all three languages, 0 discrepant cells. rdrobustplot.sthlp is rewritten to match the command after ST-9: xlabel() and ylabel() added to the syntax; col_dots() and col_lines() removed, with a note saying plainly that they never had any effect and are now rejected rather than ignored; and the claim that "any other option is passed through to rdplot's graph_options()" corrected -- extras go straight to rdplot and must be valid rdplot options. A new Stored results section documents the six r() values the command sets (tau, se_rb, ci_l, ci_r, pvalue, subtitle), which the help had never mentioned, and states the repeatability guarantee. masspoints "false": the finding held for rdplot only. rdrobust.ado and rdbwselect.ado already reject it at a strict earlier gate, which made their later lax inlist(..., "false") checks unreachable dead code that merely looked like a hole. Removed from all three so the two gates agree. masspoints(false) is now rejected everywhere, while off/check/adjust are unaffected. rdrobust.sthlp: rho() precedence over b() is now documented (the ST-8 note), verified by h(0.3) b(0.5) rho(2) giving e(b_l) = .15 = h/rho. e(sample) is documented now that ST-6 made it truthful -- it marks the rows surviving both if/in and the listwise deletion, so count if e(sample) equals e(N). The e(N) line still said "original number of observations"; it has been post-deletion since the extended drop scan, and now says so. Also confirms the ST-10 wildcard removal propagates: a bogus option passed through rdrobustplot is now caught by rdplot rather than silently swallowed, while a valid passthrough such as masspoints(off) still works. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Closes PY-5 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. R bins with findInterval(x_r, jumps_r, rightmost.closed=TRUE), whose intervals are LEFT-closed and right-open -- jumps[i] <= x < jumps[i+1] -- numbered 1..J_star_r. searchsorted(side='left') is the opposite convention, (jumps[i-1], jumps[i]], so two things went wrong at once: an interior point sitting exactly on a bin edge was assigned one bin too low, and the numbering ran 0..J_star_r -- J_star_r+1 distinct values, with bin 0 holding only the minimum -- so the old clamp folded J_star_r into J_star_r-1 and MERGED the outermost two bins. side='right' reproduces findInterval exactly, including its 1-based numbering; only the rightmost.closed adjustment is still needed, and it becomes J_star_r+1 -> J_star_r. The left side already used side='right' and was correct, so only the right-side line changes. Verified against R on the shipped senate data at nbins = (10,10), (17,17), (20,20) and (5,30): row counts now identical and every column (mean_bin, mean_x, mean_y, min_bin, max_bin, N) agrees to 2.22e-16. Before the fix Python produced exactly one bin fewer than R in all four specs (19/20, 31/32, 38/39, 33/34), confirmed by stashing the change so the check is not vacuous. Worth recording: bin counts alone are not a correctness criterion, since empty bins legitimately disappear in both languages. The comparison has to be against R rather than against the requested nbins. rdrobust Python 3/3, R suite green, 3-language baseline 85/85 + 150/150 with 0 discrepant cells. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…rashing Fixes the genvars rc=3301 crash found on 2026-08-13 while working ST-10. It was pre-existing, not a side effect of that change. PART 2 (genvars) runs after cwf back to the caller's frame and after the work frame has been dropped, so its `touse' is still the raw [if]/[in] sample: the missing-value drops only ever happened in the work frame. Two consequences. (a) genvars attached bin statistics to observations that never entered the estimation -- on the shipped senate data, rows with a missing outcome but a valid running variable were given a fitted value. (b) the loop's only guard is ZZ[i,1] != ., which skips a missing x but not a missing y. Such a row still has a valid x, so it gets binned; if it lands in a bin holding no RETAINED observation then select() returns 0 rows and the 1x10 assignment aborts the whole command. That combination is why the trigger looked erratic: it needs both a dropped-but-x-valid row and an empty bin, so missing-y-only or missing-x-only data frequently passed. Instrumenting the loop on the repro showed rows(ZZ)=3000 rather than the 2740 analytic rows, 418 bins with rows(rdplot) =417 -- one empty -- and row 50 (missing y, x=.304) mapping to that empty bin. Fixed by narrowing touse with the same drop_cond the work frame used; the local survives the frame switch, which is the technique already used for ST-6 here and ST-8 in rdhte. Added a defensive rows(_sel) == 1 guard as well, since an empty bin is legitimately possible with mass points or an extended support(): the bin columns are left missing rather than aborting. Verified on rdrobust_senate: 1390 rows, 1297 complete cases, rdplot_hat_y filled for exactly 1297, zero rows filled outside the analytic sample, zero analytic rows left empty. The original repro plus the missing-y, missing-x and missing-covs variants all return rc=0; ST-10's genvars rerun and prefix-safety checks still pass; rdplot R tests green; 3-language baseline 85/85 + 150/150 with 0 discrepant cells. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…off outside x Closes R-6 and the R half of PY-11 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. rdplot's vars_bins carried a fossil index. Line 507 had already been fixed to address the left-side bin arrays as rdplot_bin_l + J_star_l + 1, but the min_bin/max_bin assignment 56 lines further down still used the older rev(-rdplot_bin_l) form. The two agree only when every left bin is occupied. Empty bins are legitimately dropped from vars_bins, so with an asymmetric pattern of holes rev() reverses a list that is no longer symmetric and each surviving left row gets the edges of a different bin -- min_bin and max_bin that do not bracket its own mean_x. The right side was never affected. Both languages carried it: Python/rdrobust/rdplot.py:833-834 is a verbatim port of the same expression and is fixed in the companion Python commit. Verified on a fixture whose left support has a gap in [-0.55, -0.25], so 12 requested left bins produce fewer occupied ones: before the fix mean_x fell outside [min_bin, max_bin] for the affected rows and the midpoint disagreed with rdplot's independently computed mean_bin; after, all three agree to 1e-12. The test asserts the gap actually emptied bins, so it cannot pass vacuously; reverting the fix via git stash fails it. rdbwselect: the c-within-range check at what is now line 215 was unreachable in the case it exists for. With c outside the support, one side is empty and the masspoints block above it computes mass_l/mass_r as 1 - M/N with N = 0, so the function dies at `if (mass_l >= 0.2 | ...)` with "missing value where TRUE/FALSE needed" -- conditionCall() confirms the site -- before the informative message can fire. Moved a copy of the test to line 142, ahead of the masspoints work. This is the same gap the Python front-end had, which is why one audit item touched both languages. rdrobust R suite 312 pass / 0 fail / 0 skip. Numbers unchanged: 16 stashed before/after cases bit-identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…indexing Closes PY-11 and the Python half of R/PY-6 from AUDIT_FINDINGS_rdrobust_2026-07-20.md. rdplot.py:833-834 is a verbatim port of the R expression fixed in the previous commit: rdplot_min_bin_l[np.flip(-rdplot_bin_l)-1] reverses a left-bin list that is no longer symmetric once empty bins have been dropped, so surviving left rows get another bin's edges. Replaced with the direct index rdplot_bin_l + J_star_l, matching R. The rest is validation R has always had and Python did not, so the same call errored in one language and returned numbers in the other: - h and b were unchecked. A negative or NaN bandwidth reached the Cholesky factorization as a degenerate design and surfaced as "LinAlgError: Internal potrf return info = [1]"; a length-3 h matched neither the scalar nor the length-2 branch, so estimation continued with h_l never assigned and the user saw "UnboundLocalError: cannot access local variable 'h_l'" -- an internal-logic error presented as their error. - Negative weights were swept into the NA mask and dropped silently. With the negatives concentrated on one side that emptied a whole side of the cutoff and the user was told "c should be set within the range of x": an error about the cutoff, for a weights mistake. - All-zero weights gave a design matrix of zeros and another LinAlgError. - rdbwselect never checked that c lies inside the support of x, though rdrobust does; the reductions below it died as "zero-size array to reduction operation minimum", naming neither c nor x. Messages mirror R's (rdrobust.R:93-99, 312-328). New tests/test_validation.py covers all four with 8 test functions; the bin-edge case joins test_numerical_contracts.py as the Python twin of the R test. rdrobust Python 19 passed (was 3 before the Mac session added these). Behaviour change worth a release note: calls that used to return numbers with a bad h/b or negative weights now raise, matching R. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Closes R-8 from AUDIT_FINDINGS_rdrobust_2026-07-20.md, which was resolved by documenting rather than by filling in the matrix. vcov() returns the squared standard errors of the Conventional, Bias-Corrected and Robust rows with zeros off the diagonal. Those zeros are placeholders, not estimated covariances and not a claim of independence: the three rows are different estimators of the SAME parameter and are strongly correlated, but the design matrices behind tau_cl and tau_bc are not retained on the fitted object, so the cross-covariances cannot be recovered after the fact. The source has said so since 2026-08-13 and the returned matrix carries an "offdiag" attribute, but nothing user-facing did -- there is no man page for the method, so `vcov(m)` handed back a matrix whose most important property was invisible unless you read the attribute or the source. A user combining rows from it, for instance Wald-testing Conventional against Robust, would be treating two highly correlated numbers as independent and would get a wrong answer with no warning. Adds coef.rdrobust and vcov.rdrobust aliases so `?vcov.rdrobust` resolves, and a section stating what the off-diagonals are, why they cannot be filled in, and what to use instead (object$se / $ci / $pv per row, and the Robust row for reporting). Documentation only; no code path changes. checkRd clean; rdrobust R suite 312 pass / 0 fail / 0 skip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
rdrobustplot's trailing wildcard forwarded unrecognized options to rdplot as top-level options. rdplot silently swallowed them before ST-10 removed its own wildcard, and rejects them rc=198 after -- so legend(off) etc. never had a working route. Two explicit routes now: rdplot ANALYSIS options (masspoints, covs_drop, covs_eval, support, genvars, nochecks, precision -- the contract the sthlp already documented) are declared by name and forwarded as such; everything else, plus a new explicit graph_options(), folds into the graph_options() handed to rdplot, which appends it after its own twoway settings so later-wins applies. A typo'd option now errors from -twoway- instead of vanishing. Verified on Stata 16.1: plain/legend(off)/graph_options()/masspoints(off)/legend+xline all rc=0, typo rc=198, repeat call rc=0 (audit/scratch/fc17/rdrp_check.log). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Conflicts in the N < 20 fallback of R rdrobust.R and Python rdrobust.py, which both sides rewrote: kept main's version (it respects supplied h, b and rho and uses original units), which also covers the audit fix there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ELOG Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er warnings All three languages unless noted. - Bandwidth pilots: a pilot window with fewer distinct x than polynomial coefficients (or an undefined pilot from the previous stage) is not identified. It now leads to "Not enough variability" in R, Python and Stata. Before, R returned a bandwidth from ginv(), NaN or an unrelated error; Python raised LinAlgError; Stata printed a message and went on, usually with the full range as bandwidth. The rule counts distinct values, so it gives the same answer on every platform; identified fits keep their inversion (Cholesky, generalized inverse as fallback in all three). - D1: pad the bwcheck floor and the first pilot bandwidth (h_B = range) by (1 + sqrt(eps)), so the observation that sets them keeps a positive kernel weight. Selected bandwidths move by less than 1e-6 relative. - D2 (R): default ginv.tol 1e-15 instead of 1e-20, as numpy's pinv and Stata. A covariate nearly collinear on one side no longer makes the selected bandwidth depend on row order or noise. - D3: warn when either side has fewer than 10 clusters within h, and when a cluster-robust standard error is exactly 0. - R-17: fail early, with the reason, when a side has fewer than q+1 distinct values of x (p+1 in rdbwselect); a single observation used to crash. - ST-16 (Stata): rdplot per-bin CIs use N-1 degrees of freedom, as R/Python. - Stata: lrdrobust.mlib rebuilt with Stata 16.1. The new exits sit where Stata surfaces their return code (2001). Tests: R 116 (new test-post-campaign.R), Python 37 (new test_post_campaign.py); Senate baselines updated for D1. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
R 4.1.0 never reached CRAN but was public on GitHub with the bugs fixed here, so the CRAN version becomes 4.1.1. Python 2.1.0 and Stata 11.1.0 are already on PyPI and SSC. CHANGELOG entry rewritten for the release. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Made with Stata's translate (smcl2pdf), as before. Picks up the August help corrections, mainly rdrobustplot. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The package now suggests testthat for its test suite, and R CMD check stops when a suggested package is missing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Matias's review of 3f9ac7d (ISSUES.md): 1. Stata: b = h/rho uses the supplied rho, not round(rho, .0001) (rho=.00014 gave b=5000 instead of 3571.43; positive rho below 5e-5 was rejected). rho() still takes precedence over b(); a user scalar rho survives. 2. Python: rdbwselect checks every returned selector row, so all=True no longer returns NaN rows; mserd alone still works on the same data. 3. R and Python: the few-cluster warning counts clusters after the user weights, so zero-weight observations do not count. 4. Python: drop the duplicate h/b precheck; the September _normalize_bandwidth already raises ValueError (complex values too). The early insufficient-support stop raises ValueError. Second review of the full diff against main: - Python: msecomb1/cercomb1 used the built-in min(), which hid an undefined msesum pilot behind mserd; np.min passes the NaN to the guard (R and Stata already stopped). - R: the distinct-value stop ran before the cutoff check, so a cutoff outside the data was reported as a support problem; it now runs after. - R: drop a duplicate masspoints check in rdbwselect that rejected "", which main accepts. - R and Python: rdbwselect rejects negative weights and an nnmatch that is not a positive integer, as rdrobust does (R: nnmatch=0 crashed on 0/0; Python accepted 2.5 and gave a misleading error for 0). - R: vcov.rdrobust no longer adds an "offdiag" attribute (documentation of the diagonal-only contract is kept). - Clustered SEs: "exactly zero" never fired (the SE is ~1e-15, not 0). The warning now uses the deterministic condition instead: p+1 or fewer clusters on a side, where the CR variance is not identified (R, Python, Stata). Stata's new warnings respect nowarnings. - Stata: rdrobustplot passes nochecks (stored in `checks'); rdplot and rdrobustplot accept unused options and col_dots()/col_lines() again, as main does, with a note that they are ignored (main ignored them silently; the earlier version of this PR made them errors). - Python __pycache__ files are no longer in the branch. Tests: R 121, Python 56; the new ones fail on the previous version. CHANGELOG states which checks apply to which functions and languages. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Python: negative weights raise ValueError in rdrobust() and rdbwselect(),
as the CHANGELOG states (they raised a bare Exception).
- Python: a single cluster on a side gives NaN standard errors, as in R,
instead of a ZeroDivisionError in the CR1 factor g/(g-1).
- Stata: a non-numeric nnmatch token in vce(nn ...) exits 125 with a message
(vce(nn cl) used to die inside Mata with "cl not found"); rdrobust and
rdbwselect.
- CHANGELOG: scope the row-order statement for ginv.tol to the verified case
(Kuo data, default options); the hc*/masspoints("off") path keeps a ~1e-3
dependence in near-collinear designs, as in 4.0.0. Entries for the two
fixes above.
- Python tests: expect ValueError; new single-cluster test (57 pass).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
mdcattaneo
approved these changes
Oct 7, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes from a three-language audit and follow-up reviews, incorporating the September changes already on main (PR #23). R 4.1.0 never reached CRAN but was public on GitHub with these bugs, so the CRAN version becomes 4.1.1. Python 2.1.0 and Stata 11.1.0 are already on PyPI and SSC, so they become 2.1.1 and 11.1.1.
Changes in results
deriv >= 2: the point estimate now includesfactorial(deriv)(R, Python, Stata). The SE already had it. Withderiv = 2the estimate was half its correct value.deriv >= 2no longer crashes (UnboundLocalError: rdmodel).rdbwselect()withstdvars = TRUE(the default since 4.1.0): reports the effective sample size, not the full sample, and the cutoff on the original scale (R, Python). Senate data:N_h595/702 → 360/323, matchingrdrobust().sharpbw = TRUEand a nonzero cutoff no longer crashes.rdplot(): bin edges stay with their own bins when a side has empty bins (R, Python). In Stata,genvarsdescribe the estimation sample, and per-bin CIs useN - 1degrees of freedom, as in R and Python.bwcheckfloor and the first pilot bandwidth (h_B = range) are padded by1 + sqrt(eps), so the observation that sets them keeps a positive kernel weight (all three). Selected bandwidths move by less than1e-6relative on the Senate test grids (Senate defaulth: 17.75439722 → 17.75439819). A cost-sharing replication has bandwidth and standard-error movements of about2e-6relative; fixed-bandwidth baseline results are unchanged.ginv.tolis1e-15(was1e-20), as in Python and Stata. A covariate nearly collinear on one side of the cutoff made the selectedhdepend on row order (2e-3 on the Kuo data, default options); that case is fixed. A pre-existing dependence of about 1e-3 remains withvce = "hc0"to"hc3"andmasspoints = "off"in near-collinear designs.Errors and warnings
xthan polynomial coefficients, e.g. a running variable with a few values): all three now stop with "Not enough variability". Before, R returned a bandwidth fromginv(), NaN or an unrelated error (svd, "missing value where TRUE/FALSE needed"); Python raisedLinAlgError; Stata printed a message and continued, usually with the full range. The rule counts distinct values, so the answer is the same on every platform. The author reported agreement across R, Python and Stata on a grid of 60 discrete designs (57 errors, 3 identical bandwidths).q + 1distinct values ofx(p + 1inrdbwselect()) stops early with the reason; one observation on a side used to crash.h; warn with fewer than 10 per side, with a stronger caution atp + 1or fewer. The warning allows for unreliable inference or a degenerate variance without asserting that cluster count alone makes the scalar RD variance unidentified. Stata respectsnowarnings.nnmatch, andrdbwselect()accepts the same option casing asrdrobust(). The Python bandwidth normalizer retainsValueError. R gives an informative error for collinear covariates withcovs_drop = FALSE; Python retains its linear-algebra error there. Stata retains its earlier handling of nonpositive weights and rejects nonnumericnnmatchtokens before they reach Mata.e(sample)is set, user matrices and tempvars are no longer overwritten, error codes are fixed, andrdrobustplotaccepts twoway options. Unidentified bandwidth pilots returnr(2001).Compatibility
The September
N < 20behavior is retained: supplied estimation and bias bandwidths are respected,rhotakes precedence, and fallback bandwidths remain in the original running-variable units.Checks
Checks repeated after the final warning correction:
R CMD check --no-manual --no-build-vignettes:Status: OK.twine checkpass.lrdrobust.mlib. The general-cluster example agrees with an independent CR1 calculation, emits the corrected warning, and respectsnowarnings.Numerical comparisons repeated on
fa3fea8, before the warning-only correction: all 146 R/Python baseline fields and 108 R/Stata numeric fields agree at absolute/relative tolerance1e-6; fixed-bandwidth baseline results are unchanged from main in all three languages. Twelve selected Python replication cases succeed on both versions. Independent Stata checks also cover exactrho()arithmetic, user namespace preservation,e(sample), repeated plotting, graph options, andgenvars.The earlier Windows
--as-cran, 13 reverse-dependency checks, 60-design discrete grid, and full 85-estimation/150-selector parity grids were reported by the author and were not repeated in this final review.The Stata library and help PDFs supplied in the PR are retained. The final warning correction changes no estimation formulas. Submission to CRAN, PyPI, and SSC will be handled by Sebastian after merge.