Mei, Sheng & Shi (2026)
Nickell bias in panel local projection: Financial crises are worse than you think, Journal of International Economics, 104210.
The four applications use the authors' replication data. They are not redistributed with the package; the authors publish them on GitHub and Mendeley. The test item MSS applications in test/test_mss_apps.jl runs all of them when the data sit under docs/references/code/panel-lp-replication/applications/.
Specifications
Each horizon is a pre-built outcome column, estimated with horizon = 0 as in the authors' notebook. Singletons are kept, as in their code.
| application | outcome | shock | fixed effects | variance |
|---|---|---|---|---|
| Romer–Romer (2017), Fig. 4 | f{h}LNGDP, f{h}UNEMP | CRISIS | country + half-year, OneSweep() | Clustered(:unit; ssc = :none), × 7 |
| Baron–Verner–Xiong (2021), Fig. 5 | 100 × Fd{h}y | R_B, R_N | country | Clustered(:unit, :time; ssc = :mss, psd = :raw) |
| Mian–Sufi–Verner (2017), Fig. 6 | F{h}y | L0HHD_L1GDP, L0NFD_L1GDP | country | Clustered(:unit, :time; ssc = :none, psd = false) |
| Cerra–Saxena (2008), Fig. 7 | cf{h}GRRT_WB | CRISIS | country | Clustered(:unit; ssc = :none) |
r = panellp(@formula(Y ~ CRISIS + l1LNGDP + l2LNGDP + l3LNGDP + l4LNGDP +
l1CRISIS + l2CRISIS + l3CRISIS + l4CRISIS + fe(COUNTDUMS) + fe(halfyear)),
data; id = :COUNTDUMS, time = :halfyear, horizon = 0,
estimator = SplitPanelJackknife(), singletons = :keep, absorb = OneSweep())Results
Table 1 of the paper (SPJ and FE at the horizon where SPJ peaks) is reproduced exactly:
| application | $h$ | SPJ | FE |
|---|---|---|---|
| Romer–Romer, GDP (× 7) | 7 | −6.285 | −5.432 |
| Romer–Romer, unemployment (× 7) | 7 | 2.632 | 2.128 |
| Baron–Verner–Xiong, bank crash | 4 | −4.213 | −3.422 |
| Baron–Verner–Xiong, nonfinancial crash | 6 | −5.876 | −4.561 |
| Mian–Sufi–Verner, household debt | 7 | −0.558 | −0.387 |
| Mian–Sufi–Verner, firm debt | 5 | −0.136 | −0.124 |
| Cerra–Saxena | 10 | −7.893 | −4.507 |
Every coefficient and standard error of the four applications, at every horizon, matches the authors' notebook calls to about 1e-12 (relative; 2e-13 absolute). The per-horizon sample sizes match exactly. The comparison needs the authors' outputs on the local data, written by Rscript crosscheck/R/mss_apps_local.R.
The one-sweep approximation
The authors' code demeans by country and then by half-year once. That is the exact two-way within transformation only when the sample is balanced, which the GDP sample is. There OneSweep() and exact alternating projections agree to rounding (about 1e-14). The unemployment sample is unbalanced, and the published numbers depend on the approximation. Exact two-way fixed effects give (× 7; standard errors clustered by country):
| $h$ | FE, one sweep (published) | FE, exact | SPJ, one sweep (published) | SPJ, exact |
|---|---|---|---|---|
| 4 | 2.049 (0.587) | 2.042 (0.585) | 2.321 (0.815) | 2.321 (0.715) |
| 7 | 2.128 (0.675) | 2.096 (0.672) | 2.632 (1.177) | 2.546 (0.960) |
| 10 | 1.345 (0.592) | 1.286 (0.597) | 1.660 (1.056) | 1.554 (0.928) |
Stata's xtlp …, tfe, whose two-way demeaning iterates, gives the exact columns too: to 1e-13 on the coefficients and 2e-10 on the standard errors, on both Romer–Romer samples and at every horizon (a local test item).
The paper's conclusion is unchanged: SPJ responses exceed FE from $h = 2$ on (at $h = 0, 1$ they are slightly smaller, in both versions). But the peak SPJ response is about 3% smaller, and the $h = 10$ response about 6% smaller. The SPJ standard errors depend on the approximation even more: with exact two-way fixed effects they are 12–18% smaller, while the FE ones move by at most 1%.
The split rule
The authors' code (like xtlp) gives the first $\lceil n_i/2\rceil$ rows of a unit to the first half. The paper defines the halves as $t \le T_h/2$, which is $\lfloor n_i/2\rfloor$. With an odd number of periods the two differ, and the Romer–Romer GDP sample (balanced, $T_h = 65$ at the Table 1 horizon) is affected. SplitPanelJackknife(split = :floor), the paper's rule, gives −6.606 at $h = 7$ instead of the published −6.285, and −5.476 instead of −5.052 at $h = 9$ (× 7).