Audit: xtlp

xtlp (Mei, Sheng & Shi; MIT licence) is the Stata command of the Mei–Sheng–Shi paper: the panel local projection with fixed effects (m(fe)) or the split-panel jackknife (m(spj)), with one-way (fe) or two-way (tfe) effects. The audited version is 1.1.1 (xtlp.ado, 1995 lines; GitHub commit cd3fa24). The cross-checks of this package run it on the synthetic panels (crosscheck/stata/p1_fe.do, p2_spj.do); see the cross-check map.

Statements marked verified are matched by those cross-checks. The others come from reading the installed source, and line numbers refer to xtlp.ado.

What it computes

At each horizon, xtlp builds the transformed outcome (ytransf: level, diff, cmltdiff y{t+h} − y{t−1}, cmltsum), restricts the sample to complete cases, drops singletons iteratively, demeans the outcome and the regressors, and estimates by OLS without an intercept. With m(spj) it splits each unit's rows in two halves of $\lceil T_i/2 \rceil$ and $\lfloor T_i/2 \rfloor$ (by count, after singleton removal), re-demeans each half, and reports $2\hat\beta - (\hat\beta_a + \hat\beta_b)/2$.

Verified: coefficients, standard errors, observation and cluster counts and critical degrees of freedom agree with panellp (WithinFE, SplitPanelJackknife(split = :ceil)) within 2e-10 for every fe/tfe × vce() specification of the cross-checks, including unbalanced panels, gaps, singleton cascades, and SPJ halves with single rows.

Conventions, and how panellp reproduces them

xtlppanellp and vcov
singletons dropped iteratively (units; also periods under tfe)singletons = :drop (the default); keepsingletons is :keep
tfe demeaning: unit step, then time step, stop at an absolute step below 1e-9 or after 1000 sweepsAlternatingProjections() converges to the exact within transform; the difference is within the tolerance on the test panels
vce(cluster id), vce(cluster t), vce(cluster id t)Clustered(:unit), Clustered(:time), Clustered(:unit, :time); ssc = :stata reproduces the factor $(n-1)/(n-a-K-df_a) \cdot G/(G-1)$ with the nested-effects rule
vce(robust)Robust()
vce(dkraay), lag(m)DriscollKraaySE(:nw94 or m; timelags = :position): lags count positions among the sample's periods, not calendar distance
vce(dkraay …, nodfadj)ssc = :xtscc
asessc = :none (df_r is unchanged)
two-way variance not positive semi-definitepsd = :stata: negative eigenvalues clipped in the metric of the regressors' standard deviations (floored at 1e-3)
e(irf) bands: ±1.96 standard errors for every vce()summarize(…; dist = :normal)

The cross-check do-files set type double: otherwise xtlp stores the transformed outcomes as floats (below).

Findings

Precision and stored results

  • The transformed outcomes (level, diff, cmltdiff) are created with generate and stored in Stata's default type, float, so they carry about 7 significant digits (l.405–428). Only cmltsum is double. set type double restores full precision.
  • e(irf) is filled from float variables too, so its values are float-rounded (l.673–676). e(b#) and e(V#) are full double.
  • The e(irf) bands use ±1.96 for every vce(), while e(df_r#) holds t degrees of freedom (l.723–724).
  • Runs over several horizons store no per-horizon observation or cluster count (e(N#), e(N_g#), e(N_clust#)). They also never post or clear e(), so results of an earlier command persist (l.708–715). A horizon's count needs a hor(0) run on a pre-built lead.
  • hor(0), the default, ignores ytransf() silently: the transforms are built only when hor > 0 (l.397).

Samples and clusters

  • Cluster variables are never marked out of the sample: rows with a missing cluster value stay in it and form their own cluster (reghdfe would drop them).
  • The two-way cluster key is C1 · (max(C2) + 1) + C2 (l.1636). It is unique only for non-negative integer C2. Negative codes, such as %tq or %tm dates before 1960, can merge distinct cells: with max(C2) = 100, (C1 = 2, C2 = −40) and (C1 = 1, C2 = 61) both map to 162.
  • The Driscoll–Kraay gap warning assumes a unit spacing of the time variable and ignores the xtset delta (l.1786). With a delta of 5, e.g. data every 5 years, it warns about gaps that do not exist.

Estimation

  • The tfe demeaning stops when no element changes by more than 1e-9 in absolute value (l.1936–1989). The rule depends on the scale of the data: with large values, rounding noise can keep it running for all 1000 sweeps. On slowly mixing unbalanced panels, the remaining error can be well above 1e-9. When the limit is reached, it returns without a warning.
  • No collinear regressor is dropped. A full set of dummies, or a base level of a factor variable, makes $X'X$ singular: error 198 at $h = 0$ and a missing horizon otherwise.
  • When a horizon's variance is undefined (no residual degrees of freedom, a single cluster or period), its coefficients are posted as missing too, although they could be computed. panellp keeps them and reports a NaN variance.

Degrees of freedom

  • Verified: the Driscoll–Kraay default factor under tfe uses $n - K - N_g + 1$ in its denominator, while clustering by period gives $n - K - N_g$: the two differ even with no lags. DriscollKraaySE(ssc = :stata) reproduces both.
  • The default variance (vce(unadjusted)) under tfe counts $N_g + G_t - 1$ absorbed effects, which assumes one connected component of units and periods (reghdfe counts the components). panellp has no homoskedastic variance.
  • Verified: inside the SPJ halves, singletons are kept, and their scores are $2\tilde X_{it}$. A unit with two or three rows has a single row in a half. The halves are not required to have four periods per unit, although the error message for a singular half says so.

Minor

  • Comments on the ase factors (l.1494–1691) describe factors that the code no longer applies (it sets them to 1).
  • In the several-shocks path, the PSD-repair warning reports a stale horizon (XTLP_H).