Skip to content

About

Panel local projections for Julia, with fixed effects and IV estimation, jackknife estimation, robust inference and t-LAHR inference following Almuzara and Sancibrián (2026)

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

PanelLocalProjections.jl

CI

Panel local projections in Julia. For each horizon h, the package estimates a panel regression of y_{i,t+h} on a shock, and gathers the estimators and inference procedures of the recent panel-LP literature in one place:

Method Reference
Fixed-effects (within/LSDV) panel LP, with unit-, time- and two-way-clustered, heteroskedasticity-robust, Driscoll–Kraay and equal-weighted cosine (EWC) standard errors Jordà (2005); Driscoll & Kraay (1998); Cameron, Gelbach & Miller (2011); Lazarus, Lewis, Stock & Watson (2018)
Split-panel jackknife (SPJ) correction of the Nickell bias of the FE estimator Mei, Sheng & Shi (2026, JIE)
Cumulative-difference OLS (CD-OLS), which avoids the within transformation Ugarte-Ruiz (2026, BBVA Research)
t-LAHR inference for aggregate (macro) shocks: time-clustered, lag-augmented, heteroskedasticity-robust, with an HC2 + Imbens–Kolesár refinement Almuzara & Sancibrián (2026)
Panel LP-IV (2SLS, with lag augmentation and t-LAHR), with Kleibergen–Paap, Montiel Olea–Pflueger and Cragg–Donald first-stage diagnostics Almuzara & Sancibrián (2026, §3.4); Kleibergen & Paap (2006); Montiel Olea & Pflueger (2013)
LP-DiD: difference-in-differences by local projections on clean-control samples, for absorbing, non-absorbing and one-off treatments, with reweighting, regression adjustment, pre-mean differencing, pooled estimates and a pre-trend test Dube, Girardi, Jordà & Taylor (2025, JAE)
Pre-trends (negative horizons) and the response holding the later path of the treatment fixed, for a continuous, instrumented treatment Ilzetzki (2024, AER)

Version 0.2.1, for Julia 1.10 (LTS) and later. What comes after it (joint bands across horizons, weak-IV-robust inference) is listed in TODO.md; the changes of each version are in CHANGELOG.md. The documentation covers the methods, the conventions of each reference implementation and how panellp reproduces them, the cross-checks, and audits of the reference code.

Installation

The package is not in the General registry. Install it from GitHub:

using Pkg
Pkg.add(url = "https://github.com/andrerecio/PanelLocalProjections.jl")                   # main
Pkg.add(url = "https://github.com/andrerecio/PanelLocalProjections.jl", rev = "v0.2.1")   # or a tagged release

It depends on StatsModels, Tables, StatsAPI, StatsFuns, RecipesBase and Julia's standard library. Makie is optional (a package extension, for irfplot).

Quick start

panellp(formula, data; id, time, horizon) returns one fit per horizon. data is any Tables.jl table in long format (a DataFrame, a NamedTuple of vectors), one row per unit and period, in any order; id and time name the unit and period columns. Periods are integers (consecutive integers are consecutive periods), or dates with frequency = :year, :halfyear, :quarter or :month. Gaps and missing values are allowed: an undefined lead or lag drops the row from that horizon's sample.

This block runs as pasted:

using PanelLocalProjections, Random

# A simulated panel: 50 units, 40 periods, y_it = μ_i + 0.5 y_{i,t-1} + x_it + e_it
rng = Xoshiro(1)
N, T = 50, 40
id = repeat(1:N; inner = T)
t = repeat(1:T; outer = N)
x = randn(rng, N * T)
y = zeros(N * T)
for i in 1:N
    μ = randn(rng)
    for p in 1:T
        k = (i - 1) * T + p
        y[k] = μ + (p == 1 ? 0.0 : 0.5 * y[k - 1]) + x[k] + randn(rng)
    end
end
data = (; id, t, y, x)

# y_{i,t+h} on x_it for h = 0, …, 6, with unit fixed effects
r = panellp(@formula(leads(y) ~ x + lags(x, 1) + lags(y, 1) + fe(id)), data;
            id = :id, time = :t, horizon = 6)
s = summarize(r, Clustered(:unit))
Impulse response of leads(y) to x
Variance: Clustered(:unit; ssc = :stata); 95% bands
h        coef         se        lower      upper  df  nobs
0     0.98566  0.0219284     0.941593    1.02973  49  1950
1    0.511577  0.0310447     0.449191   0.573964  49  1900
2    0.232064  0.0432445     0.145161   0.318967  49  1850
3    0.104635  0.0403817    0.0234849   0.185785  49  1800
4   0.0660779  0.0376194  -0.00952113   0.141677  49  1750
5  -0.0173723  0.0394707   -0.0966917   0.061947  49  1700
6  -0.0250721  0.0444163     -0.11433  0.0641857  49  1650

The true response is 0.5^h. The output is from Julia 1.13; the random stream, and so the digits, change with the Julia version.

The formula helpers:

Side Terms
response leads(y) (y_{i,t+h}), ldiff(y) (y_{i,t+h} − y_{i,t−1}), anchor(y, z) (y_{i,t+h} − z_{i,t}), cumul(y) (Σ_{k≤h} y_{i,t+k}), or a pre-built column with horizon = 0
regressors columns and functions of columns, lags(x, 1:p), leads(x, 1:p), firstdiff(x), interactions s & X, fixed effects fe(a) and fe(a) & fe(b) (absorbed), an IV block (endo ~ inst), and leads(x) for the path x_{i,t+1}, …, x_{i,t+h} at horizon h

horizon = 6 means the horizons 0:6; a range may start below zero (horizon = -4:8) for pre-trends. weights = :w gives weighted least squares, and pooled = true one regression of the response averaged over the horizons.

Results

Variances are computed after estimation, from score sums stored at every horizon, so several can be compared without re-estimating:

summarize(r, DriscollKraaySE(); level = 0.90)   # another variance, other bands
summarize(r, Clustered(:unit); term = "L1.y")   # another coefficient

coefnames(r)                           # "x", "L1.x", "L1.y"
coefpath(r)                            # the shock's coefficient at every horizon
coef(r, 2), nobs(r, 2)                 # all coefficients, and the observations, at h = 2
horizons(r), nobspath(r), esample(r, 2)

cv = vcov(Clustered(:unit, :time), r)  # variances at every horizon
vcov(cv, 2)                            # the matrix at h = 2
stderror(cv; term = "x")               # standard errors across horizons

s.horizon, s.coef, s.se, s.lower, s.upper, s.df, s.nobs

A summary is a Tables.jl table, and it plots with Makie or Plots.jl:

using DataFrames
DataFrame(s)

using CairoMakie      # loads the Makie extension
irfplot(s)            # a Figure; irfplot!(ax, s; color, bandalpha) draws on an axis

import Plots
Plots.plot(s)         # the same with Plots.jl (Makie and Plots both export `plot`)

Estimators and variances

estimator = What it fits What it needs
WithinFE() (default) OLS after absorbing the fe() terms; pooled OLS with an intercept when there are none —
SplitPanelJackknife() the fixed-effects estimator with the split-panel jackknife correction fe(id) in the formula; no IV block
CDOLS() pooled OLS of the long difference on the shock and its first lag the response ldiff(y) (or cumul(g) of a growth rate), lags(shock, 1) among the regressors, no unit fixed effects (time effects are allowed); no IV block. CDOLS(validate = false) skips the checks
LagAugmented(rule) fixed effects, with p(h) lags of the outcome and of every shock added at horizon h unit fixed effects. The default rule ICLags() needs the first shock to be an aggregate series, alone or times an exposure (X, s & X); ConstantLags(p) and CappedLags(pmax) work with any shock
an IV block (endo ~ inst) two-stage least squares at every horizon; firststage(r) gives the diagnostics WithinFE() or LagAugmented(rule)
LPDiD(…) the long difference on the treatment switch, on the newly treated rows and the clean controls of each horizon the formula ldiff(y) ~ firstdiff(D) + fe(time) with a 0/1 treatment D; no IV block. nonabsorbing = L for a treatment that can switch off
Variance What it is Works with
Robust() heteroskedasticity-robust (HC1) every estimator
Clustered(:unit), Clustered(:time), Clustered(:unit, :time), Clustered(:region) one- or two-way cluster-robust every estimator. A column that is not the unit, the period or a fixed effect is declared at estimation: panellp(…; clusters = (:region,))
DriscollKraaySE(bw) Driscoll–Kraay, Bartlett kernel every estimator
CosineHAR(B) equal-weighted cosine (EWC) every estimator
TLAHR(refine = false) clustered by period, no small-sample factor, normal critical values; the t-LAHR variance of an LP-IV every estimator but SplitPanelJackknife; meant for LagAugmented
TLAHR() t-LAHR with the HC2 + Imbens–Kolesár refinement unweighted OLS fits (no IV block, not SplitPanelJackknife) whose shocks are an aggregate series, alone or times an exposure; meant for LagAugmented
Homoskedastic() classical, σ̂²(X̃'X̃)⁻¹; a compatibility option (regress, ivreg2 without robust) least-squares and 2SLS fits (not SplitPanelJackknife, not regression-adjustment LP-DiD)

Small-sample factors, critical values, the Driscoll–Kraay bandwidth and the repair of a two-way variance are options of each specification (ssc, psd, timelags); see Inference.

Usage by method

data and firms stand for your panel.

# Mei–Sheng–Shi: fixed effects against the split-panel jackknife, unit-clustered standard errors
f     = @formula(leads(y) ~ x + lags(x, 1:4) + lags(y, 1:4) + fe(id))
r_fe  = panellp(f, data; id = :id, time = :year, horizon = 10)
r_spj = panellp(f, data; id = :id, time = :year, horizon = 10,
                estimator = SplitPanelJackknife())
summarize(r_spj, Clustered(:unit))

# Ugarte-Ruiz: cumulative-difference OLS
r_cd  = panellp(@formula(ldiff(y) ~ s + lags(s, 1)), data; id = :id, time = :year,
                horizon = 9, estimator = CDOLS())
summarize(r_cd, Clustered(:unit))

# Almuzara–Sancibrián: an aggregate shock mp interacted with an exposure lev, t-LAHR
# inference. `shock` is required when the formula has an interaction and other regressors
r_la  = panellp(@formula(ldiff(logk) ~ lev & mp + lev + fe(firm) + fe(sector) & fe(date)),
                firms; id = :firm, time = :date, frequency = :quarter, horizon = 12,
                shock = "lev & mp", estimator = LagAugmented(ICLags()))
summarize(r_la, TLAHR(); level = 0.90)
lagorder(r_la)                             # the lags p(h) added at each horizon

# Panel LP-IV: s × ffr instrumented by s × mp, lag-augmented
r_iv  = panellp(@formula(leads(y) ~ (s & ffr ~ s & mp) + fe(id) + fe(t)), data;
                id = :id, time = :t, horizon = 12, estimator = LagAugmented())
summarize(r_iv, TLAHR(refine = false))     # TLAHR() is not defined for IV
firststage(r_iv)                           # Kleibergen–Paap F at every horizon

In the third call the periods are dates, so fe(date) and clusters on date group by quarter. The Montiel Olea–Pflueger effective F is reported for one endogenous regressor; with LagAugmented the lags of the endogenous shock are endogenous too, and it is NaN.

LP-DiD

A binary treatment D that units enter at different dates (Dube, Girardi, Jordà & Taylor 2025). The formula is the regression; the estimator picks, at each horizon, the rows entering treatment and the clean controls (not yet treated at t + h):

f  = @formula(ldiff(y) ~ firstdiff(D) + fe(t))
es = panellp(f, data; id = :id, time = :t, horizon = -5:8, estimator = LPDiD())
summarize(es, Clustered(:unit))            # event study; h = −1 is zero by construction
pretrendtest(es)                           # joint test of the coefficients at h ≤ −2
averageeffect(es)                          # average over the post- and pre-treatment windows

# the equally weighted effect on the treated, with a covariate and lagged outcome growth
# (regression adjustment: its scores are the influence function, so no small-sample factor)
rw = panellp(@formula(ldiff(y) ~ firstdiff(D) + x + lags(firstdiff(y), 1) + fe(t)), data;
             id = :id, time = :t, horizon = -5:8, estimator = LPDiD(reweight = true))
summarize(rw, Clustered(:unit; ssc = :none))

# one pooled estimate over the post-treatment window; the same rows at every horizon
panellp(f, data; id = :id, time = :t, horizon = 0:8, pooled = true, estimator = LPDiD())
panellp(f, data; id = :id, time = :t, horizon = -5:8, sample = :common, estimator = LPDiD())

# a treatment that switches on and off, with effects assumed to stabilise within 5 periods
panellp(@formula(ldiff(y2) ~ firstdiff(D2) + fe(t)), data; id = :id, time = :t,
        horizon = -5:8, estimator = LPDiD(nonabsorbing = 5))

Other options: control = :nevertreated or :notyet, firsttreat, oneoff (a treatment that lasts one period), untreated_before, pmd = k or :max (pre-mean differencing), weights. It is written from the paper and reproduces the Stata command lpdid 1.1.0 option by option; the LP-DiD page has the map between the two, and the audit the places where they differ (the reweighted estimate on unbalanced panels, among them).

Pre-trends and the path of the treatment

For a continuous, instrumented treatment, as in Ilzetzki (2024): horizons from −6, and at h > 0 the leads x_{t+1}, …, x_{t+h} as controls, so that the coefficient of x_t is the response holding later treatment fixed.

r = panellp(@formula(ldiff(y) ~ (x ~ z) + lags(x, 1:6) + leads(x) + fe(id) + fe(t)), data;
            id = :id, time = :t, horizon = -6:18)
summarize(r, Clustered(:unit))
firststage(r; vcov = Clustered(:unit))     # Kleibergen–Paap, effective and Cragg–Donald F

The replication page reproduces his Figure 9 on the replication data (singletons = :keep and Homoskedastic(ssc = :none) for the published numbers).

Example: monetary policy and US states

How much more does personal income fall in states more exposed to manufacturing after a rise in the 1-year Treasury rate? The rate is instrumented by monetary policy surprises, a panel LP-IV in Almuzara & Sancibrián's setting. The data are public: BEA state personal income, FRED's 1-year rate (GS1), and the Jarociński–Karadi surprises. examples/us_states/build.jl downloads them and builds the panel (julia --project examples/us_states/build.jl, from the repository root); the plot comes from examples/us_states/figure.jl.

using PanelLocalProjections
include("examples/us_states/csv.jl")       # readtable
d = readtable("docs/data/us_states/processed/state_panel.csv")

# ldiff(lpi): 100 × (log income_{t+h} − log income_{t−1}); s: 1998 manufacturing share
# (standardised); gs1: 1-year rate; mp: surprise; state and quarter effects
iv = panellp(@formula(ldiff(lpi) ~ (s & firstdiff(gs1) ~ s & mp) + fe(fips) + fe(t)), d;
             id = :fips, time = :t, horizon = 12, estimator = LagAugmented())
summarize(iv, TLAHR(refine = false); level = 0.90)    # t-LAHR for LP-IV (A&S A.4)
summarize(iv, Clustered(:unit); level = 0.90)         # clustered by state, for comparison
firststage(iv).kpF                                    # Kleibergen–Paap F, clustered by period

using CairoMakie
irfplot(summarize(iv, TLAHR(refine = false); level = 0.90))

LP-IV and reduced-form responses of US state personal income, with t-LAHR and state-clustered 90% bands

Clustered by state, the responses look significant at several horizons. The t-LAHR standard errors are 1.5 to 2.7 times those clustered by state, and none of the t-LAHR bands excludes zero. The first stage shows the same thing: its Kleibergen–Paap F is about 2 clustered by period, but 35 robust and 200,000 clustered by state. The shock is aggregate, so the effective sample size is the 84 quarters, not the 4,284 state-quarters. See the example page for the details.

Data: BEA Regional Economic Accounts; Board of Governors of the Federal Reserve System (H.15), via FRED; Jarociński & Karadi (2020, AEJ: Macroeconomics), updated surprises (CC BY 4.0). The data are not distributed with the package.

Conventions

Leads, lags and differences are always computed on the (unit, calendar period) grid, so gaps in a unit's time series are respected and nothing is shifted by row position. Every convention on which reference implementations disagree is an explicit option, and each is cross-checked against the reference that uses it: degrees-of-freedom factors, singleton handling, two-way demeaning, the Driscoll–Kraay bandwidth and lag ordering, the split-panel rule and critical values. The references are the Stata commands xtlp, reghdfe, xtreg, regress, xtscc, locproj, xtspj, ivreghdfe, ivreg2, xtivreg2, weakivtest, harreg and lpdid, R fixest, and the authors' own R code (Mei–Sheng–Shi's LP_panel, Almuzara–Sancibrián's panel_LP.R). crosscheck/README.md maps each reference command to the panellp options that reproduce it, and Deviations lists the defaults that differ from a reference.

Documentation

The pages are in docs/src:

To build the HTML pages locally (Julia 1.11 or later), into docs/build:

julia --project=docs -e 'using Pkg; Pkg.instantiate()'
julia --project=docs docs/make.jl

Acknowledgements and inspiration

The fixed-effect absorption, the OLS solver and every variance estimator are implemented in this package. Their design takes inspiration from:

None of them is a dependency. Source files that follow one of these projects say so in a comment.

Citing

If you use the package, please cite it together with the paper whose method you apply; see CITATION.bib.

  • Mei, Z., Sheng, L. & Shi, Z. (2026). Nickell bias in panel local projection: Financial crises are worse than you think. Journal of International Economics, 104210.
  • Ugarte-Ruiz, A. (2026). Panel local projections without fixed-effects: A cumulative-difference OLS estimator. BBVA Research.
  • Almuzara, M. & Sancibrián, V. (2026). Micro responses to macro shocks. Working paper, July 2026; earlier version: Federal Reserve Bank of New York Staff Report No. 1090.
  • Dube, A., Girardi, D., Jordà, Ò. & Taylor, A. M. (2025). A local projections approach to difference-in-differences. Journal of Applied Econometrics 40(7), 741–758.
  • Ilzetzki, E. (2024). Learning by necessity: Government demand, capacity constraints, and productivity growth. American Economic Review.

License

MIT; see LICENSE.

About

Panel local projections for Julia, with fixed effects and IV estimation, jackknife estimation, robust inference and t-LAHR inference following Almuzara and Sancibrián (2026)

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages