Audit: Ugarte-Ruiz (2026)

Panel local projections without fixed-effects: a cumulative-difference OLS estimator, BBVA Research working paper, 29 June 2026. The paper ships no code. Its Monte Carlo runs with the Stata commands xtarsim (Bruno 2005) and locproj (Ugarte-Ruiz 2023), its split-panel jackknife with locproj …, m(xtspj) (Sun & Dhaene 2019; L369), and the applications' jackknife with xtlp (Mei, Sheng & Shi; L764). Line numbers refer to the paper converted to markdown. The findings below come from reading the paper, the installed sources of xtarsim and locproj, and simulations in this package's test suite.

The estimator is as stated; the stated algebra is not

itempaperwhat holdspanellp
$\gamma_h$, the coefficient on $s_{i,t-1}$$(\rho_y^{h} - 1)\beta$ (L261); $(\rho_y^{h+1} - 1)\beta$ in eqs. 29, 34, 35L261 is a typo, and both are the structural moving-average coefficient, not the probability limit, which is $(\rho_y^{h+1} - 1)\beta / (1 - \rho_y \rho_s)$ (tested)$\gamma_h$ is a nuisance coefficient; not reported as structural
Proposition 1"unbiased" (L269)the proof uses population moments: it shows consistency, not finite-sample unbiasednessdocumented as consistent
Assumption 1$E[s_{i,t} \mid v_{i,t}] = 0$ (L110), $E[s_{i,t} \mid v_{i,s}] = 0$ for $s \ne t$ (L112)contradicts eq. (2), since $s_{i,t}$ contains $v_{i,t}$ and its past; what is needed is that $e$ is independent of the path of $s$—
L233$s_t$ uncorrelated with $s_{t-k}$ for $k > 2$false for an AR(1); the argument needs the Markov property (no partial correlation given $s_{t-1}$)—

The fixed-effects and jackknife comparators have another estimand

In the specification without the shock's lag (eq. 24; Table 9), the fixed-effects and split-panel jackknife LPs in levels do not target $\beta_h$: as $T$ grows they converge to $\beta_h + \beta\rho_y^{h+1}\rho_s/(1 - \rho_y\rho_s)$, +177.8% of $\beta$ at $h = 0$ when $\rho_y = \rho_s = 0.8$ (a test item checks this limit on a long panel). Table 9's jackknife "bias" of +177.45% at $h = 0$ is this omitted-lag term, not an over-correction by the jackknife. The paper ties the upward bias to specifications without shock lags (L834), and with the shock's lag (Table 10) the jackknife's bias at $h = 0$ is −5.57%. The term does not explain everything: the +1 y-lag specification (Table 11) still reports an average jackknife bias of +12.6% (+20.4% at $h = 9$).

The Monte Carlo's jackknife (locproj …, m(xtspj)) is not xtlp's, which this package reproduces. xtspj averages two splits when a unit has an odd number of rows (here at every odd $h$, with $T = 50$), and reports the OLS variance of the within regression with normal critical values. So the paper's jackknife coverage comes from another variance.

Unstated choices the reported numbers depend on

  • Variance. The paper calls the standard errors "robust" (L335–342) and recommends a robust or Driscoll–Kraay variance (L281), without a formula. The CD-OLS coverage pattern in its tables (about 0.95 with $\rho_y = 0$, up to 1.00 at long horizons with $\rho_y = 0.8$) is what heteroskedasticity-robust (HC1) standard errors give when the long-difference error is serially correlated. Simulating the paper's designs (a test item; 2000 panels for the figures): with $\rho_s = 0$, $\rho_y = 0.8$, HC1 overstates the sampling standard deviation by 19–51% at $h = 2$–$9$ (coverage 98–99.8%); with $\rho_s = \rho_y = 0.8$ it understates it by 7–12% at $h = 1$–$4$ (coverage about 93%). Unit-clustered standard errors match it within 4%.
  • The DGP. The signal-to-noise ratio (L305), $\sigma_e$ and the start-up are not defined in the paper. They are xtarsim's: $\sigma_e = 1$, a signal-to-noise ratio over $\sigma_e^2$ that sets $\sigma_\xi$ (2.68, 2.56 and 0.72 for $(\rho_s, \rho_y) = (.8, 0), (0, .8), (.8, .8)$), and a stationary start-up without burn-in. The test suite checks the committed xtarsim panels against these formulas, and that the package's own generator, at the same $\sigma_\xi$, reproduces their moments.
  • Time effects. Eq. (22) has a time effect $u_t$ (L297), but the theory excludes them (L118) and the Monte Carlo specifications have none (L344).

Labels and reproducibility

  • Table 11 is labelled "Lags $S_t$ 0, Lags $y_t$ 0"; by its position it is the "+1 y lag" specification.
  • Table A12 is labelled $\rho_y = 0$; by its position it is the $\rho_s = \rho_y = 0.8$ design.
  • Appendix A.5 reads "0.8 instead of 0.8" (L1164); it means 0.5.
  • CMRE with one added shock lag is the same regression as CMRE1 without added lags, yet the tables report different numbers, so the specifications do not share simulated data (or the implementations differ). Exact reproduction of the tables is not possible.

Limits of the method the paper does not test

  • Heterogeneous shock means. The design's shock has the same mean for every unit. With unit-specific means the pooled regression of $s_{i,t}$ on $s_{i,t-1}$ keeps the mean in the residual, and CD-OLS is inconsistent. The test suite shows a bias of tens of standard errors on a large simulated panel.
  • Unit roots and trends. With $\rho_y = 1$ or unit-specific trend growth, the long difference keeps a unit term.
  • Feedback from $y$ to $s$ is excluded by assumption.

xtarsim

Read from the installed xtarsim.ado (v1.0.0):

  • negative parameters are mis-handled: Stata evaluates ^ before unary minus, and the parameters are pasted into expressions as text, so $-0.5^2$ is $-0.25$; with $\beta < 0$ the command stops with error 411. The paper uses non-negative parameters only;
  • with $\gamma\rho \ne 0$ the start-up of $x$ and of its contribution to $y$ use the same first innovation, so the first periods carry a transient term decaying like $\rho^{t-2}$;
  • a two-way model's time effect enters the dynamic equation as $(\tau_t - \gamma\tau_{t-1})/(1-\gamma)$, not $\tau_t$;
  • seed() seeds the kiss32 generator (the command runs under version 8.2), so the draws are reproducible across Stata versions but not from Julia; the panels are committed.