Skip to content

Latest commit

 

History

History
70 lines (57 loc) · 6.23 KB

deviations.md

File metadata and controls

70 lines (57 loc) · 6.23 KB

Deviations from the reference implementations

Where this package's defaults differ from a reference implementation, the table gives the reason and, where a published number depends on it, the switch that reproduces the reference. The audit pages document each reference in detail.

Mei, Sheng & Shi (LP_panel_all.r, xtlp)

topic reference panellp default switch
two-way demeaning one unit-then-time sweep (R code): exact only on balanced samples alternating projections to convergence absorb = OneSweep()
split-panel jackknife halves first \lceil n_i/2 \rceil rows (R code, xtlp); the paper writes \lfloor n_i/2 \rfloor split = :ceil, as the code SplitPanelJackknife(split = :floor)
small-sample factors (R code, robust = T) N/(N-1) \cdot (n-1)/(n-K); two-way G/(G-1) \cdot (n-1)/(n-N-K), G = \min(N, T) reghdfe's factors (ssc = :stata) Clustered(:unit or :unit, :time; ssc = :mss)
repair of a two-way variance that is not positive semi-definite R code: clip the eigenvalues of V (eigens = T); xtlp: clip in the metric of the regressors' standard deviations psd = :stata, as xtlp and reghdfe psd = :raw
Driscoll–Kraay lags (xtlp, xtscc) by position among the sample's periods by calendar period timelags = :position
xtlp's e(irf) bands ±1.96 standard errors whatever the variance the specification's critical values summarize(…; dist = :normal)
transformed outcomes (xtlp) stored as floats unless set type double double —
a horizon whose variance is undefined (xtlp) coefficients posted as missing coefficients kept, variance NaN —

Ugarte-Ruiz (2026)

topic paper panellp switch
specification rules stated in the text checked (CDOLS()) CDOLS(validate = false)
variance "robust", formula not stated (heteroskedasticity-robust, presumably) any; unit clusters are valid, robust ones are not Robust()

Almuzara & Sancibrián (panel_LP.m/.R)

topic reference code panellp default switch
lags \min(h, p_{\max}): no lag at h = 0; no information criterion the paper's \max\{1, \min(h, \hat p)\} with AIC (ICLags()) LagAugmented(CappedLags(p_max))
the HC2/Imbens–Kolesár refinement off by default on (the paper's recommendation) TLAHR(refine = false)
synthetic regression raw exposure s_{it}; periods unweighted the exposure residualised per period (the paper's \hat s_i, and on the exposures of the other regressors of the same aggregate series); periods weighted by \omega_t (fn 24: CR2 by period) TLAHR(weights = :raw)
degrees-of-freedom contrast e_1 for every coefficient (a bug) each coefficient's own —
singletons kept dropped, as reghdfe and xtlp singletons = :keep
no fixed effects no intercept an intercept 0 +
Date time index (R) broken (day counts) calendar periods with frequency —

Dube, Girardi, Jordà & Taylor (lpdid 1.1.0)

topic lpdid panellp (LPDiD) switch
reweighting factor of the weighted regression (rw) the treated share of the period among clean rows with the outcome, including rows whose \Delta D is undefined and that the regression drops the treated share among the rows of the regression (the equally weighted average of the cohorts' comparisons) compat = :lpdid
switches, for notyet, firsttreat and the check of an absorbing treatment summed from the observed first differences: a switch in a gap or behind a missing value is not seen changes between consecutive observed treatment values compat = :lpdid
first treatment with notyet or nevertreated keeps the switch-off rows of a unit treated from its first row, with a regressor of −1 a control has \Delta D = 0: those rows are left out compat = :lpdid
a treatment that switches off, without nonabsorbing() kept, with a regressor of −1 an error nonabsorbing = L
nocomp with a one-sided window also requires the pre-treatment horizon 2 (or horizon 0) the requested horizons (sample = :common) horizon = -2:H
pooled regressions under nocomp the flags of both pooled windows the horizons of the window —
h = -1 under pmd() reported as zero estimated —
standard errors of regression adjustment no small-sample factor, normal critical values the specification's Clustered(:unit; ssc = :none)
wild bootstrap bootstrap() through boottest not implemented —

Ilzetzki (2024, ivreg2 with dummies)

topic replication package panellp default switch
variance classical, RSS/N (no option to ivreg2) none: the variance is chosen at summarize Homoskedastic(ssc = :none)
critical values Student t with N - K, on a standard error without correction normal for ssc = :none the degrees of freedom of Homoskedastic()
singletons of the line and month dummies kept dropped singletons = :keep
first-stage F Cragg–Donald Kleibergen–Paap with the chosen variance firststage(r).cdF

Variance conventions shared with other software

topic reference panellp switch
Driscoll–Kraay bandwidth xtscc, xtlp: \lfloor 4(T/100)^{2/9} \rfloor lags; Almuzara & Sancibrián: kernel bandwidth \lceil 4((T-h)/100)^{2/9} \rceil DriscollKraaySE(:nw94) :as2026 (M - 1 lags), :mss2026, an integer
Driscoll–Kraay factor xtscc and xtlp …, nodfadj: (n-1)/(n-K-1) \cdot T/(T-1) xtlp's default (ssc = :stata) ssc = :xtscc; ssc = :none (ase)
fixest's two-way variance eigenvalues floored at 1e-16 (vcov_fix) psd = :stata psd = :raw (clips at 0)
equal-weighted cosine variance harreg: B = \lfloor 0.41\,T^{2/3} \rfloor, t(B) the same (CosineHAR()) an integer B
absorption stopping rule xtlp: absolute step below 1e-9; the Almuzara–Sancibrián code: 1e-8 (MATLAB: absolute norm of the change; R: fixest::demean) relative to the column after its first sweep, 1e-12 AlternatingProjections(tol = …)