t-LAHR inference

Almuzara & Sancibrián (2026) study panel local projections of micro outcomes on a macro shock $X_t$ interacted with a unit-level exposure $s_i$:

\[y_{i,t+h} = \beta(h)\, s_i X_t + \psi(h)' W_{it} + \mu_i(h) + \nu_t(h) + \xi_{it}(h),\]

with unit and time effects (time effects only when $s_i$ varies; with $s_i \equiv 1$ they would absorb $X_t$). The estimand is the slope of the best linear predictor of the unit responses on the exposure, $\beta(h) = \operatorname{Cov}(s_i, \beta_{ih}) / \operatorname{Var}(s_i)$. Aggregate shocks make the errors dependent across units, so the effective sample size is $T$, not $NT$. Their recommendation, t-LAHR, is

  • time-clustered: the variance is clustered by period (CR0, eq. 11);
  • lag-augmented: the regression includes $p(h)$ lags of the outcome and of the shock, $W_{it} = (s_i X_{t-1}, y_{i,t-1}, \dots, s_i X_{t-p}, y_{i,t-p}, \tilde W_{it})$. With $p \ge h$ the scores $X_t \xi_t$ are serially uncorrelated (their Lemma A.1), so no kernel or bandwidth is needed;
  • heteroskedasticity-robust: the variance is the Eicker–Huber–White variance of the synthetic time-series regression (Remark 1), refined in small samples with HC2 and Student-t critical values with the Imbens–Kolesár degrees of freedom (supplement B.2).

It is not a HAR estimator: Driscoll–Kraay, equal-weighted cosine (EWC) and clustering by unit or two-way are the paper's comparators.

r = panellp(@formula(leads(y) ~ s & X + fe(id) + fe(t)), data; id = :id, time = :t,
            horizon = 12, estimator = LagAugmented())       # ICLags(): eq. (15)
summarize(r, TLAHR(); level = 0.90)                         # HC2 + Imbens–Kolesár t(ν)
lagorder(r), r.lags                                         # p(h); the criterion
synthetic(r, 4)                                             # the synthetic series at h = 4

Lag augmentation

LagAugmented adds, at each horizon, $p(h)$ lags of the outcome and of every shock column (named L1.y, L1.(s & X), …). A lag already among the regressors is not added again, whatever its name: columns are compared numerically, so with the shocks s & lags(X, 0:2) the lag of s & L0.X is s & L1.X. A lag that the formula's columns span is left out of the regression, without a warning: with lags(y, 1:2) in the formula, the lag $\Delta y_{t-1} = y_{t-1} - y_{t-2}$ that ldiff(y) adds is redundant, and the formula's two coefficients stay. A lag of a shock lags the whole column, so a time-varying exposure enters as $s_{i,t-\ell} X_{t-\ell}$, as in the authors' code (the paper writes $s_i X_{t-\ell}$). The lagged outcome depends on the response:

responselagged outcome
leads(y), anchor(y, z)$y$
ldiff(y), leads(firstdiff(y))$\Delta y$
cumul(g)$g$ (the authors' cumulative = true)

The rule $p(h)$:

rule$p(h)$source
ICLags() (default)$\max\{1, \min(h, \hat p)\}$, $\hat p$ by AICthe paper, eq. (15) and B.7
ICLags(criterion = :bic)the same with BICsupplement B.3
CappedLags(pmax)$\min(h, p_{\max})$: no lag at $h = 0$the authors' panel_LP.m/.R
ConstantLags(p)$p$"plain" t-LAHR ($p = 1$, Table 1)

The paper keeps at least one lag at every horizon, including $h = 0$, as insurance against near unit roots.

The information criterion

The paper chooses $\hat p$ on the synthetic time series (fn 11; supplement B.1): a panel criterion with penalty in $(NT)^{-1}$ over-fits, because redundant lags improve the fit at the rate $T^{-1}$. The paper leaves several details open; ICLags() pins them:

  • the "panel VAR(p)" is the $h = 0$ regression of the outcome on the formula's right-hand side (with the contemporaneous shock) and $p$ of its own lags, with the formula's fixed effects: B.3 adds one parameter per lag, the lag of $\bar Y$ (vars = :outcome_and_shock also adds the shock's lags, with penalty $p(1 + n_{\text{shock}})g$);
  • each candidate $p$ is fitted on its own sample, the rows where its lags are defined (B.4 sums over $t = p+1, \dots, T$; sample = :common uses the $p_{\max}$-lag sample for all);
  • the residuals are averaged per period with the synthetic weights of the first shock (below), $\bar\xi_t(p)$, and $\hat\sigma^2(p) = \sum_t \omega_t \bar\xi_t(p)^2 / \sum_t \omega_t$ over the periods with positive weight $\omega_t$: the mean square residual of the synthetic regression, which fn 24 weights by $\omega_t$. On a balanced panel with a time-invariant exposure $\omega_t$ is constant and $\hat\sigma^2$ is the mean of $\bar\xi_t^2$, as in B.4. Unweighted, the periods with few units, whose $\bar\xi_t$ have variance $\propto 1/\omega_t$, would dominate the criterion (with staggered entry it then often selects no lag, or $p_{\max}$, for a strong AR(2));
  • $\mathrm{IC}(p) = \log\hat\sigma^2(p) + k(p)\,g$, with $k(p)$ the lags candidate $p$ adds to the regression, $g = 2/T$ (AIC) or $\log T / T$ (BIC) and $T$ the periods of the $p = 0$ regression. A lag already in the formula (or spanned by it) adds nothing: candidates that differ only by such lags are the same regression, and $\hat p$ is the last of them, so writing lags(y, 1) in the formula changes neither $\hat p$ nor $p(h)$. Other ties go to the smaller $p$;
  • $p_{\max}$ is not given in the paper. The default is $\max\{1, \lceil (T - H)^{1/3} \rceil\}$, the authors' code's $p_{\max}$ ($T$ the distinct periods in the data, $H$ the largest horizon): with $T = 124$ and $H = 20$ it is 5; the paper's applications select up to 4.

The criterion needs the first shock to be an aggregate series times an exposure; for other shocks use ConstantLags or CappedLags.

Sensitivity

On the cross-checks' panels (fx_tlahr, fx_tlahr_unbal: a first-order design) every option of the criterion (criterion = :bic, sample = :common, vars = :outcome_and_shock, pmax 2 or 6) selects $\hat p = 1$, as ICLags(); on the test suite's AR(2) panel (ar2panel() in test/test_tlahr.jl: $N = 200$, $T = 60$, autoregressive coefficients 1.2 and −0.5) every option selects $\hat p = 2$. What moves the inference is the rule. Standard error of the shock, with $\nu$ in parentheses, at $h = 0, 2, 4$:

rule ($p(h)$ for $h = 0, 1, 2, \dots$)fx_tlahr_unbalAR(2) panel
ICLags(), any option (1, 1, 1, … and 1, 1, 2, 2, …).110 (11.9), .207 (11.4), .255 (10.6).060 (15.9), .167 (13.4), .263 (13.2)
CappedLags() (0, 1, 2, 3, 4, 4, …).167 (12.6), .251 (9.3), .311 (5.7).229 (16.1), .167 (13.4), .278 (12.9)
ConstantLags(1) (1, 1, 1, …)as ICLags().060 (15.9), .178 (15.3), .260 (15.0)

Without a lag at $h = 0$ (the authors' code) the standard error rises by half on fx_tlahr_unbal and almost fourfold on the AR(2) panel: the lagged outcome absorbs the persistence of $y$. At long horizons the code adds $p_{\max} = 4$ lags each of the outcome and of the shock, eight synthetic regressors, and $\nu$ falls to 5.7. One lag where two are needed (ConstantLags(1) on the AR(2) panel) moves the standard error at $h = 2$ by 7% and raises $\nu$.

The variance

CR0 by period

TLAHR(refine = false) is the variance clustered by period with no small-sample factor and normal critical values, eq. (11):

\[\hat\sigma^2(h) = \frac{\sum_t \big(\sum_i \tilde x_{it}\hat\xi_{it}\big)^2} {\big(\sum_{i,t} \tilde x_{it}^2\big)^2},\]

$\tilde x_{it}$ being the shock residualised on the other regressors and the fixed effects. It equals Clustered(:time; ssc = :none). With WithinFE and no lags it is the paper's "t-HR" comparator.

HC2 and Imbens–Kolesár degrees of freedom

TLAHR() transfers the small-sample refinements of heteroskedasticity-robust inference to the synthetic regression (supplement B.2). For a shock with weights $w_i$:

  1. the synthetic regressors are the per-period weighted averages of the within-transformed regressors, $\bar X_t = \sum_{i \in t} w_i \ddot X_{it} / \omega_t$, $\omega_t = \sum_{i \in t} w_i^2$, one row per period of the sample;
  2. the synthetic regression is weighted by $\omega_t$ (fn 24), so its rows are $\tilde X_t = \sqrt{\omega_t}\,\bar X_t$, and $d_t = 1 - h_t$, with $h_t$ the leverage of period $t$ in $\tilde X$;
  3. the variance is $B\,(\sum_t S_t S_t' / d_t)\,B$, with $S_t$ the scores summed in period $t$ and $B$ the bread of the panel regression;
  4. the critical value is $t_{1-\alpha/2}(\nu_k)$, with the Bell–McCaffrey degrees of freedom of coefficient $k$ under a homoskedastic reference (Imbens & Kolesár 2016): $\nu_k = (\sum\lambda)^2 / \sum\lambda^2$, $\lambda$ the eigenvalues of $G'G$, $G_{\cdot t} = \bar M_{\cdot t} a_t / \sqrt{d_t}$, $a = \tilde X (\tilde X'\tilde X)^{-1} e_k$, $\bar M = I - \tilde X (\tilde X'\tilde X)^{-1} \tilde X'$. Since $G'G = D \bar M D$ with $D = \operatorname{diag}(a / \sqrt d)$, $\sum\lambda = \sum_t a_t^2$ and $\sum\lambda^2 = \sum_{t,s} a_t^2 a_s^2 \bar M_{ts}^2 / (d_t d_s)$, computed without an eigendecomposition.

The period weights. B.2 normalises the averages by $\sum_i \hat s_i^2$, which is constant over time on a balanced panel, and does not say what to do when $\omega_t$ varies (unbalanced panels, time-varying exposures). When every regressor is an aggregate series times the exposure, $\ddot X_{it} = w_i \bar X_t'$, the block of period $t$ in the panel's hat matrix has rank one, and its eigenvalue is the leverage of $\sqrt{\omega_t}\,\bar X_t$: the HC2 variance of the weighted synthetic regression is then CR2 clustered by period (Bell & McCaffrey 2002), and $\nu_k$ its Bell–McCaffrey degrees of freedom. A test checks both on an unbalanced panel, to 1e-12. The authors' code uses the unweighted $\bar X_t$, which gives a period with few units the leverage of a full one; TLAHR(weights = :raw) reproduces it. On a balanced panel with a time-invariant exposure $\omega_t$ is constant and the two coincide.

Stata. $\nu_k$ is, by construction, the degrees of freedom of Stata's regress …, vce(hc2, dfadjust) (e(adj_df)) on the weighted synthetic regression. The standard error is the panel's, $B\,(\sum_t S_t S_t'/d_t)\,B$; it equals that regression's HC2 standard error only when the panel regression is itself a synthetic regression (every regressor an aggregate series times the exposure, on a balanced panel), as in the cross-checks below. With the lagged outcome that LagAugmented adds, a unit-level regressor, they differ by a few percent (the paper's Remark 1 says "≈").

Numerics. $\tilde X'\tilde X$ is never formed: $\tilde X$ has a thin SVD, and singular values below $\max(T, K)\,\varepsilon\,\sigma_1$ are dropped (a comment in the authors' code notes that $\bar X'\bar X$ can be near-singular, $\kappa \approx 10^{11}$, and that a looser pseudo-inverse then makes the degrees of freedom depend on the truncation rank). With $U$ its left singular vectors and $b = a^2/d$, $\sum\lambda^2 = \sum_t b_t^2 d_t^2 + \lVert U' \operatorname{diag}(b)\, U \rVert_F^2 - \sum_t b_t^2 h_t^2$: no $T \times T$ matrix is formed, so the memory is linear in the number of periods (daily panels of thousands of periods). A period with leverage one (as many periods as synthetic regressors, without unit effects) makes the variance undefined: NaN, with a warning.

The absorbed intercept. B.2 builds $\bar X$ from the two-way within-transformed regressors, without a constant. On a balanced panel with time-invariant weights and unit effects, the columns of $\bar X$ sum to zero over the periods, so the leverages add up to at most $\operatorname{rank}(\bar X) \le T - 1$ and every $d_t \ge 1/T$: the synthetic regression behaves as one with an intercept whose leverage is not counted.

The weights

The paper's weights are the demeaned exposure, $\hat s_i = s_i - \bar s$, period by period when $s_{it}$ varies (Remark 3). TLAHR() uses, for a regressor $s_k X_t$, its exposure residualised per period

  • on an intercept, when a fixed effect nested in the periods is absorbed (time effects, or group × time effects, within whose groups it is demeaned), and
  • on the exposures of the other regressors that multiply the same aggregate series (up to scale), shocks or not.

The second rule extends the paper, which covers a scalar exposure. With several exposures of one shock (e.g. turnover in two liquidity groups, as in Jeenas & Lagos 2024), the raw exposures make $\bar X$ nearly singular. By Frisch–Waugh–Lovell, the coefficient on $s_k X_t$ is identified by $s_k$ net of the other exposures of $X_t$. The panel's residual is global, with a coefficient weighted by $X_t^2$; the residual period by period equals it on a balanced panel with time-invariant exposures, where the other exposures' synthetic columns then vanish (up to rounding, which the rank tolerance drops). Since the rule follows the regressors, not their labels, the inference on $s_k X_t$ does not depend on which of them shock names. With time-varying exposures or an unbalanced panel the other columns do not vanish: the contemporaneous ones survive through the unit effects, and the lags of the other exposure that LagAugmented adds stay nearly collinear with the shock's own lags. $\bar X$ is then ill-conditioned and $\nu$ low: with two lags and a time-varying second exposure on the cross-checks' balanced panel, $\nu$ of the first exposure is about 4.9 from $h = 3$ on, against 6.2–6.8 with the raw exposures and 8.1–9.6 with that exposure alone (at $h = 0$, without lags: 11.8, 4.5 and 12.7). These intervals err on the wide side. A period whose residualised exposure vanishes (no more units than exposures, or one exposure common to all units under time effects) has weight exactly zero, not a round-off weight.

TLAHR(weights = :raw) weights by the raw exposure $s_{it}$ and leaves the periods unweighted, as the authors' code. The two agree on balanced panels with time effects and a time-invariant exposure (where the within-transformed regressors sum to zero in each period, the weights differ by a constant and $\omega_t$ is constant), and for $s \equiv 1$ on balanced panels. They differ on unbalanced panels, with time-varying exposures, and with several exposures. On the cross-checks' unbalanced panel, the shock's standard errors differ by up to 3% and its degrees of freedom by up to 16%; with a time-varying exposure, by up to 4% and 29% (1% and 5% on the balanced panel).

Several shocks and the other coefficients

Every coefficient takes its $d_t$ and degrees of freedom from one synthetic regression:

  • a shock, its own, as in the authors' code (which reports one variance per shock);
  • a lag of a shock (L1.(s & X), however it was added), and a regressor with the same exposure as a shock (s & L1.X for the shock s & X), the shock's;
  • any other aggregate series times an exposure (e.g. a control s2 & X), its own;
  • every other coefficient (outcome lags, unit-level controls), the first shock's.

Writing $z^{(j)}_t = (B S_t)_j / \sqrt{d^{(c(j))}_t}$ for coefficient $j$ with owner $c(j)$, the variance is $V_{jl} = \sum_t z^{(j)}_t z^{(l)}_t$. Each coefficient's own variance is the paper's HC2 variance, and $V$ is positive semi-definite. The degrees of freedom of coefficient $j$ use its own contrast $e_j$; the authors' code uses $e_1$ for every coefficient (see the audit).

TLAHR() needs every shock to be an aggregate series (equal across units in every period), or an interaction of aggregate series with unit-level variables: s & X, X & s, log(X) & s & w, lags(s & X, 0:1), X alone (exposure 1). For any other shock, TLAHR() is an error and TLAHR(refine = false) gives the CR0 variance. Neither is defined for SplitPanelJackknife results.

Unit-level controls

B.2 averages the whole regressor vector, so each regressor enters the synthetic regression and takes up leverage there, unit-level controls included, although the panel identifies their coefficients from cross-sectional variation. Irrelevant controls therefore widen the refined intervals. On fx_tlahr with one lag, twenty columns of noise raise the shock's refined standard error by about two thirds, at $h = 0$ (0.096 to 0.158) and at $h = 4$ (0.256 to 0.421), and lower $\nu$ at $h = 4$ from 10.4 to 7.5, while the coefficient and the CR0 standard error do not move. Include the controls the specification needs, no more; TLAHR(refine = false) does not depend on them.

The synthetic series

synthetic(r, h) returns, per period of the horizon's sample, the weighted averages of the within-transformed shock (x), response (y) and residuals (resid), and the weight $\omega_t = \sum_i w_i^2$; synthetic(r, h; term) gives the column of any coefficient in the synthetic regression it uses (above). The local projection is approximately the time-series regression of y on x weighted by $\omega_t$ (fn 24); the paper plots the rescaled series $\sqrt{\omega_t}\,\hat Y_{t+h}$ and $\sqrt{\omega_t}\,X_t$ over time (Fig. 3).

Comparators

the paper's labelpanellp and vcov
1W(i): clustered by unitWithinFE(); Clustered(:unit; ssc = :none)
2W: two-way clusteredWithinFE(); Clustered(:unit, :time; ssc = :none, psd = false)
DK98: Driscoll–Kraay, $M = \lceil 4((T-h)/100)^{2/9} \rceil$WithinFE(); DriscollKraaySE(:as2026; ssc = :none)
t-HR: clustered by period, no lagsWithinFE(); TLAHR(refine = false)
t-HAR: equal-weighted cosineWithinFE(); CosineHAR()
t-LAHR, plain: one lag, normalLagAugmented(ConstantLags(1)); TLAHR(refine = false)
t-LAHR, recommendedLagAugmented(); TLAHR()

The paper sets a negative two-way variance's standard error to zero (supplement D); with psd = false it is NaN here, and neither covers. CosineHAR() uses $B = \lfloor 0.41\,T^{2/3} \rfloor$ cosine terms and t($B$) critical values, as Stata's harreg (Lazarus, Lewis, Stock & Watson 2018); the paper says only "LLSW tuning".

Monte Carlo: Table 1

The stylized design of the paper's §2: $y_{it} = 0.7 y_{i,t-1} + \beta_i X_t + Z_t + u_{it}$, $\beta_i \sim N(1, 1)$, $s_i = \beta_i$, so $\beta(h) = 0.7^h$; $T = 30$, $N = 1500$; $\sigma_u^2 = N(1 - \bar R^2)/\bar R^2$ (fn 10). Coverage of 90% intervals at $h = 0, 1, 2, 5$ over 2000 replications (the :mc_full test item; the paper: 5000 replications), with the paper's values below each row:

$\bar R^2 = 0.9$$h = 0$125
1W(i).482.376.306.371
paper.494.372.344.376
2W.812.726.726.804
paper.819.726.725.811
DK98.706.720.689.774
paper.716.725.695.782
t-LAHR plain.832.830.801.780
paper.851.842.811.806
t-LAHR recommended.902.906.870.874
paper.903.899.872.875
$\bar R^2 = 0.1$$h = 0$125
1W(i).884.868.882.870
paper.888.878.866.880
2W.824.812.808.804
paper.841.829.814.820
DK98.814.801.782.800
paper.809.791.789.810
t-LAHR plain.829.828.828.802
paper.848.849.836.835
t-LAHR recommended.900.901.900.898
paper.903.902.891.895

The information criterion selects one lag in 85% ($\bar R^2 = 0.9$) and 80% ($\bar R^2 = 0.1$) of the replications. The fn 10 calibration reproduces the paper's unit-cluster row. By eq. (7), whose exact calibration is $\sigma_u^2 = N(1 - \bar R^2)(1 + \rho^2\bar\beta^2)/\bar R^2$, it corresponds to $\bar R^2 \approx 0.93$; with the exact calibration the unit-cluster coverage rises to .554, .440, .368, .439, further from the paper. A reduced version ($N = 200$, 300 replications) runs in the test suite.

Cross-checks

  • The authors' panel_LP.R, run on our two synthetic panels (balanced, and with 10% of the rows removed): five specifications ($s_i$ with unit and time effects; $s \equiv 1$ with unit effects; a time-varying exposure; a control; two exposures), $p_{\max} \in \{0, 2\}$, level and cumulative responses, with and without the refinement: 672 rows. With LagAugmented(CappedLags(pmax)), singletons = :keep and TLAHR(weights = :raw), they agree on the balanced panel within 2e-12 (estimates; standard errors 1e-13, degrees of freedom 3e-13, relative) and on the unbalanced one within 2e-9 (their absorption stops at a tolerance of 1e-8). The exception is the second exposure's degrees of freedom, where their code uses the first coefficient's contrast. The paper's weights agree with theirs where $\omega_t$ is constant (above).
  • Stata. reghdfe reproduces the lag-augmented two-way regression and its time-clustered variance (CR1, with the observation, cluster and degrees-of-freedom counts), within 2e-14 (the unbalanced panel's standard errors within 5e-12). On the synthetic regression (aggregate regressors, a balanced panel: $s \equiv 1$ with unit effects, and $\hat s_i$ with unit and time effects), regress …, noconstant vce(hc2, dfadjust) reproduces TLAHR()'s standard errors and degrees of freedom, and harreg …, estimator(ewc) reproduces CosineHAR(), within 3e-14.
referencepanellp and vcov
panel_LP(y, X, s, i, t, FE = cbind(id, t), p_max = p, small_sample = FALSE)leads(y) ~ s & X + fe(id) + fe(t), LagAugmented(CappedLags(p)), singletons = :keep; TLAHR(refine = false)
… small_sample = TRUETLAHR(weights = :raw)
… cumulative = TRUEcumul(y) ~ …
reghdfe F#.y sX L(1/p).y L(1/p).sX, absorb(id t) vce(cluster t)LagAugmented(CappedLags(p)); Clustered(:time)
regress ȳ x̄ L.x̄ L2.x̄, noconstant vce(hc2, dfadjust) (synthetic, demeaned)leads(y) ~ X + lags(X, 1:2) + fe(id); TLAHR()
harreg ȳ x̄ L.x̄ L2.x̄, noconstantthe same; CosineHAR()