Lead paths and negative horizons

Four general options of panellp come from the event-study use of local projections. They are what Ilzetzki (2024) uses on a continuous, instrumented treatment (replication) and what LP-DiD builds on.

Negative horizons

horizon = -6:12 estimates the regression at horizons before the shock as well: the response is dated before the regressors, so a coefficient away from zero is a pre-trend.

  • leads(y) is $y_{i,t+h}$ and ldiff(y) is $y_{i,t+h} - y_{i,t-1}$ for every integer $h$. At $h = -1$ the long difference is identically zero: the coefficient and its standard error are 0 (the horizon is the origin of the response, as in the figures of Ilzetzki 2024 and the pre1 row of lpdid).
  • The right-hand side at $h < 0$ is that of $h = 0$.
  • cumul(y) has no negative horizons, and LagAugmented sets its lags $p(h)$ for $h \ge 0$ only: both are errors. The other estimators accept them.
  • An integer horizon = H still means 0:H.

The path of a regressor: leads(x)

A local projection of $y_{t+h}$ on $x_t$ includes the effect of the later values of $x$ that $x_t$ predicts. Adding $x_{t+1}, \dots, x_{t+h}$ as controls gives the response to $x_t$ holding its later values fixed:

\[y_{i,t+h} - y_{i,t-1} = \beta_h x_{it} + \sum_{j=1}^{h} \phi_{h,j}\, x_{i,t+j} + \dots\]

This is the regression of Ilzetzki (2024, Figure 9), and what conditioning on future treatment does in LP-DiD (DGJT Section 4.3).

panellp(@formula(ldiff(y) ~ x + lags(x, 1:6) + leads(x) + fe(id) + fe(t)), data;
        id = :id, time = :t, horizon = -6:18)
  • leads(x) without a range is the path: the columns F1.x, …, Fh.x at horizon $h$ and none at $h \le 0$. A row needs the leads of its horizon only. A lead beyond the horizon is not in the regression (a NaN coefficient, not an omitted regressor).
  • leads(x, a:b) adds the same leads at every horizon.
  • In an IV fit the path of the endogenous regressor is exogenous: it is included in both stages and is not instrumented (as in Ilzetzki's ivreg2 call). A path term cannot be inside the (endo ~ inst) block nor nested in another term.

Observation weights

weights = :w is weighted least squares, as Stata's [pweight = w]: the fixed effects are absorbed with weighted means, the scores are $w_{it}\,\tilde x_{it}\,\hat e_{it}$, and the degrees-of-freedom factors count rows. Rows with a missing or zero weight leave the sample; a negative weight is an error. With WithinFE, CDOLS and LPDiD (and with an IV block). TLAHR(refine = true) is not defined for weighted fits.

Pooled fits

pooled = true runs one regression of the response averaged over the horizons, on the rows where every horizon's response is defined: the average response over the window. The horizons must be on one side of the shock (0:8, or -5:-2). The result holds a single fit, stored at the far end of the window (r.horizons), and r.pooled is the window. With the same regressors at every horizon, on a common sample without singletons to drop, it equals the mean of the horizons' coefficients.

The classical variance and the Cragg–Donald F

Homoskedastic is $\hat\sigma^2 (\tilde X'\tilde X)^{-1}$, with $(\hat X'\hat X)^{-1}$ and the structural residuals for LP-IV:

$\hat\sigma^2$critical valuesas
Homoskedastic()$RSS/(n - K - df_a)$$t(n - K - df_a)$regress, reghdfe, vce(unadjusted), ivreg2 …, small
Homoskedastic(ssc = :none)$RSS/n$normalivreg2 without options

It is the variance of the figures in Ilzetzki (2024); it is not robust to heteroskedasticity or to correlation within units, so it is a compatibility option rather than a recommendation. firststage reports, next to the Kleibergen–Paap and effective F, the Cragg–Donald Wald F (cdF), the homoskedastic weak-identification statistic that ivreg2 prints without robust or cluster.