Audit: Ramey's infrastructure replication code

Replication files of Ramey, "The Macroeconomic Consequences of Infrastructure Investment" (README dated June 5, 2021): Dynare models with Stata programs that plot their impulse responses and compute multipliers (Models/), and Stata programs for the empirical figures (Empirical/). They are a local reference here (not committed). The findings come from reading the code; line numbers refer to the files named.

The local projection is a cross-section LP-IV

The only local projection, Empirical/arra_highway_irfs.do (Figs. 6–7), is not a pooled panel local projection. Each regression is a cross-section of states in 2009 (Alaska and Nebraska dropped, l.110). The panel only serves to build long differences:

\[y_{i,2009+h} - y_{i,2008} = a_h + \beta_h\,\mathrm{ARRA}_{i,2009} + \gamma_h'\Delta c_{i,2009} + \delta_h \Delta \mathrm{HE}_{i,2008} + u_{ih}, \qquad h = 0, \dots, 4,\]

where ARRA is the highway apportionment per capita and HE highway employment. The apportionment is instrumented by two road factors ($road_factors, l.32), and the controls are political changes (l.132) and lagged highway employment (l.159). It is estimated by two-step efficient GMM (ivregress gmm …, robust, l.168–175) with 90% bands of ±1.65 standard errors (l.193–194). The outcomes are ARRA obligations and outlays, highway spending and highway employment.

panellp estimates panel LP-IV by two-stage least squares (since v0.2), not by two-step efficient GMM, so this specification is not reproduced.

Findings

  • The instruments. A comment says the standard errors use "the 3 instruments" (l.134), but the regressions use the two of $road_factors. The three-factor list L1_all_factors (l.36) is defined and never used.
  • Missing values become zeros. replace FHWA_<name> = 0 if FHWA_<name> == . (l.41) applies to every FHWA flow. A missing later year then enters a long difference as a spurious 0.
  • Flows, not cumulative multipliers. The responses are long differences of annual flows. The cumulative variables (cumadd_0910_*, cumadd_0912_*, l.42–45, 68) are built and never used.
  • A control left out. The file notes that the control S12ces_sa_pc1208_n "is significant and affects results" (l.125), and the baseline excludes it. The population control of Fig. 7 is a manual edit of l.159/162.
  • Weak instruments. Two-step GMM on 48 states, with normal critical values. The first-stage regression is printed (l.151), but no weak-instrument or overidentification test accompanies the estimates.
  • Unused results. The population-weighted regressions (l.174–178) are computed and not plotted (the figure combines the unweighted ones, l.207).
  • Saved settings. The do-file has no xtset. Its leads and lags rely on the panel settings stored in the .dta file.
  • Empirical/bouakez_reg.do (§3.3) runs time-series regressions of TFP on capital stocks with leads of their differences and a trend, in the style of dynamic OLS, with heteroskedasticity-robust (HC1) standard errors (vce(robust), l.30–34). Inference on a cointegrating regression needs long-run-variance standard errors. The time index is built from the row order (ydate = y(1960) + _n − 1, l.18).
  • Models/graphirfs_nk.do, l.55. replace grossrproduct = grossr*grossrproduct[_n-1] refers to grossr, which no variable is named. Stata abbreviates it to grossrproduct, so the running product is squared and the present-value multiplier of that file is wrong. The figures plot impulse responses only. The Figure 4 code that the README assigns to this file is commented out (l.166–220).
  • The README names Simulations_irf.xlsx (the file is Simulation_irfs.xlsx), neo_iadj_util.mod (neo_iadj_utiliz.mod), and Multipliers.do, Graphs_bea_govk.do and Bouakez_reg.do (lower case on disk).
  • Timing in the model files. The $h = 0$ row of the Dynare output is the steady state, and the shock hits at $h = 1$ (sheet neo_gc of Simulation_irfs.xlsx). So a cumulative multiplier "at $h = 4$" covers the impact quarter and three more.