B <- matrix(c(1, -0.5, -0.25, 1), nrow = 2, byrow = TRUE)
Gamma0 <- c(y = 0.2, x = 0.1)
Gamma1 <- matrix(c(0.3, 0, 0.1, 0.4), nrow = 2, byrow = TRUE)
Omega <- diag(2)
B_inverse <- solve(B)
A0 <- solve(B, Gamma0)
A1 <- solve(B, Gamma1)
Sigma <- B_inverse %*% Omega %*% t(B_inverse)
mu <- solve(diag(2) - A1, A0)Module 12: Interpreting VARs with Impulse Responses
Specify the Shock, Trace the System · Econ 6376
How to Use These Notes
The Core sections are the main reading path and the assessable material. Deeper Dive, Technical Note, and Looking Ahead boxes are optional. Follow the worked derivation to understand where an impulse response comes from; you are not expected to reconstruct the full derivation unaided. You should be able to explain the experiment, interpret a supplied calculation, and read a response plot with its assumptions and units attached.
The chapter follows one Concept → Math → Simulation → Application argument. We first compare two paths, derive their difference, and use the familiar Module 11 system to study identification. We deliberately impose two restrictions that are false for that known generator. Finally, we read Timothy Bianco’s monetary-policy application. We do not repeat a separate simulation and application for each supporting notation or software subsection.
Before running the examples
Run the R chunks in order in a fresh session with vars, ggplot2, and knitr installed. The examples need no download or credentials. The interactive displays work in the HTML served through a local HTTP preview; the static figures and all R examples remain available without them. The simulation is the small, single-process Module 11 illustration. The uncertainty example uses 1,000 bootstrap replications with a fixed seed.
Core Learning Objectives
By the end of the module, you should be able to:
- Define an impulse response by stating what changes and what is held fixed.
- Explain how contemporaneous impact and lagged propagation produce its path.
- Read the shock, response, scale, time horizon, and units from an IRF display.
- Distinguish a response at a horizon from a cumulative response.
- Explain why different identifying restrictions can produce different structural responses from the same reduced-form information.
- Choose an appropriate R object and interpret
ortho = TRUEandortho = FALSEcorrectly. - Interpret pointwise uncertainty bands without treating them as forecast intervals or evidence that an identifying assumption is true.
- Connect Bianco’s identifying assumptions to the monetary-policy responses he reports.
12.1 From a Coefficient to a System Response — Core
Module 11 ended with a useful warning. In our structural equation for , the direct coefficient on was zero, yet the corresponding reduced-form coefficient was . The fitted VAR was not making a mistake. That coefficient combined connections within a simultaneous system. Reading it as the primitive structural cross-lag effect was the mistake.
We now ask a different question: if one specified disturbance changes today, how does each variable respond now and later? This is still marginal reasoning, but the object is a sequence of system responses to a defined shock, rather than an isolated regression coefficient. The reduced coefficients remain useful: they describe how an initial difference propagates through time.
Think again of a stone entering a pond. The point of entry, the connections through the water, and the movement at an observation point are different pieces of the story. An impulse response function, or IRF, records the movement attributable to a particular disturbance over subsequent horizons. Calling it a response to a monetary-policy shock requires an argument about which disturbance represents monetary policy. The curve does not supply that argument by itself.
On the course’s mixing board, we have the same connected autoregressive system as Module 11. No new outcome-memory term is being switched on. We are learning how to interpret the propagation already present in that system. This also extends Module 8’s pulse-response idea: in a VAR, the other endogenous variables are allowed to adjust as part of the system rather than following an externally fixed predictor path.
IRFs and forecast error variance decompositions together are often called innovation accounting. This module develops IRFs. Module 13 returns to the complementary question about the importance of shocks for forecast uncertainty.
Someone reports a coefficient on yesterday’s interest rate in today’s output equation. What else would you need before claiming to know how output responds over the next two years to an unexpected policy easing?
12.2 Two Paths, One Changed Shock — Core
Define the experiment
Imagine running two copies of the same system. Give them the same history up to date . At date , add one specified disturbance to the second copy. At every later date, feed the two copies exactly the same innovations. The difference between the copies is the effect of the extra shock, even if both paths are also being moved by other disturbances.
Write that difference as
Here identifies the shock, is its size, and is time since the shock. Horizon zero is impact, the date the shock arrives. The vector contains one response for every variable in the system. Its entry is the difference in , not the future value of itself.
The comparison changes one shock once. It does not hold an observed variable permanently at a new value. Future innovations are the same across the two copies; they do not have to be zero. In a linear model their common contribution cancels when we subtract the paths.
A response is a deviation from baseline
Picturing the system at rest makes the comparison easy to see: a shock moves it away from that baseline. The baseline can also move; the IRF still measures the gap between the two paths.
For a hypothetical policy example, suppose the federal funds rate is modeled in percentage levels and its response to an identified inflation shock is +0.5 percentage point at horizon four. The rate is then 0.5 percentage point above the baseline without the additional identified inflation shock. A baseline rate of 2% gives a shocked rate of 2.5%; a baseline of 4% gives 4.5%. Both comparisons have the same IRF.
The response alone reveals neither the baseline nor the actual rate. It also does not give the change from horizon three to horizon four. This is a deviation in level units, not a first difference. Subtraction removes the common baseline; recovering either actual path requires additional information about the model, initial history and innovations defining that path.
Keep the familiar system
We use the Module 11 generator, with independent, unit-variance structural disturbances and :
The two variables affect one another contemporaneously. For our first experiment, add one unit to and leave unchanged. Because the generating shock variance is one, that is also one generating-shock standard deviation. It is not a command to raise observed by exactly one unit: the simultaneous system determines the impact on and .
The following code records the system we already solved. describes contemporaneous connections, contains structural lag coefficients, and contains reduced-form lag coefficients.
For a clear first picture, start both copies at the system mean and set their common future innovations to zero. The baseline then remains flat. The shocked copy receives the extra structural disturbance at impact.
H <- 20
q_x <- c(0, 1) # select the x structural disturbance
baseline <- shocked <- matrix(0, nrow = H + 2, ncol = 2)
baseline[1, ] <- shocked[1, ] <- mu # common history at horizon -1
common_innovations <- matrix(0, nrow = H + 1, ncol = 2)
for (h in 0:H) {
baseline[h + 2, ] <- A0 + A1 %*% baseline[h + 1, ] +
common_innovations[h + 1, ]
shocked[h + 2, ] <- A0 + A1 %*% shocked[h + 1, ] +
common_innovations[h + 1, ]
if (h == 0) shocked[h + 2, ] <- shocked[h + 2, ] + B_inverse %*% q_x
}
response <- shocked[-1, ] - baseline[-1, ]
colnames(response) <- c("y", "x")Start with the default shock and inspect impact, the next period, and the complete path. Then switch the receiving variable. The two outcome paths and their difference answer different questions. The fixed baseline is a visual convenience; the difference is the object we want.
The static version below shows the same experiment. In the left panels the dashed baseline and the shocked path are levels. The right panels subtract those levels.
paths_long <- rbind(
data.frame(h = rep(0:H, 2), variable = rep(c("y", "x"), each = H + 1),
value = as.vector(baseline[-1, ]), curve = "Baseline", panel = "Outcome paths"),
data.frame(h = rep(0:H, 2), variable = rep(c("y", "x"), each = H + 1),
value = as.vector(shocked[-1, ]), curve = "Shocked", panel = "Outcome paths"),
data.frame(h = rep(0:H, 2), variable = rep(c("y", "x"), each = H + 1),
value = as.vector(response), curve = "Difference", panel = "IRF: shocked minus baseline"))
paths_long$panel <- factor(paths_long$panel,
levels = c("Outcome paths", "IRF: shocked minus baseline"))
paths_long$variable <- factor(paths_long$variable, levels = c("y", "x"))
ggplot(paths_long, aes(h, value, colour = curve, linetype = curve)) +
geom_line(linewidth = 0.9) + facet_wrap(variable ~ panel, scales = "free_y", ncol = 2) +
scale_colour_manual(values = c(Baseline = "#777777", Shocked = "#167d9a", Difference = "#b54832")) +
scale_linetype_manual(values = c(Baseline = "dashed", Shocked = "solid", Difference = "solid")) +
labs(x = "Horizon after the shock (periods)", y = "Simulated-variable units", colour = NULL, linetype = NULL) +
theme(legend.position = "bottom")An IRF can be negative even while the observed series rises. A negative response means that the shocked path is below the baseline path at that horizon. It need not mean that the series fell relative to its own previous value.
12.3 Derive the Response, Then Record It Compactly — Core
Impact: solve the same simultaneous system
Subtract the baseline equations from the shocked equations at date . The intercepts cancel. The lagged values cancel because the histories match. Only the extra unit in the structural disturbance remains:
Substitution gives , hence . Therefore . In vector form,
Both variables move on impact even though we changed only . The disturbance and the observed variable are different objects. These two equations describe a simultaneous solution, not successive rounds in which moves, then , then again within the period.
One period later: today’s response becomes tomorrow’s history
At , the new structural disturbances are the same in both worlds. Subtraction gives
Use the impact values:
Substitute the second expression into the first:
Thus and . No new pulse was added. The system responds because the previous period’s endogenous variables carry the initial difference forward.
The same operation at every later horizon
For , the response pair always satisfies
Solving that pair is exactly the structural-to-reduced transformation from Module 11. Rather than repeat the substitution at every horizon, use its reduced lag matrix:
For example, the response two periods after the shock is
The response is . The first three response vectors now tell the beginning of the story: an immediate simultaneous effect followed by a changing sequence of lagged effects.
r <- matrix(0, nrow = H + 1, ncol = 2)
colnames(r) <- c("y", "x")
r[1, ] <- B_inverse %*% q_x # row 1 is horizon 0
for (h in 1:H) r[h + 1, ] <- A1 %*% r[h, ]
knitr::kable(data.frame(h = 0:2, r[1:3, ]), digits = 4,
caption = "Responses to the one-unit x structural shock")| h | y | x |
|---|---|---|
| 0 | 0.5714 | 1.1429 |
| 1 | 0.4898 | 0.6367 |
| 2 | 0.3415 | 0.3890 |
stopifnot(max(abs(response - r)) < 1e-12)A compact formula
Let be a selector vector: it contains a one in position and zeros elsewhere. A shock of size to structural disturbance has impact . Each further horizon multiplies that impact by once more:
The matrix power means repeated matrix multiplication; gives impact. It is not the entrywise operation A1^h in R. The loop above is a transparent way to compute the response without making that mistake.
The formula separates impact, determined by the shock and the contemporaneous structure, from propagation, determined by reduced-form lag dynamics. A different shock changes the impact vector. The same then propagates it.
The moving-average view
Our stable VAR(1) also has the representation
The matrix describes the effect of a unit reduced-form innovation from periods ago. This is the multivariate version of the AR-to-MA representation encountered earlier in the course. The infinite sum records the overlapping contributions of many shocks; we constructed one such contribution before writing the sum.
For structural responses, remember . If the structural disturbances have diagonal covariance , write , where . The diagonal entries of are the structural shock standard deviations. Then
Column of is the response at horizon to one standard deviation of structural disturbance . Our example has , so the unit and standard-deviation calculations coincide. That coincidence will not hold automatically for another system or another identification.
For a stable VAR(), let and for negative . The response matrices satisfy
The lag matrices propagate past response vectors in the same way they propagate the original outcomes. Structural responses still attach an impact map: . VAR(1) is the especially transparent case in which the recursion becomes a single matrix power.
12.4 Read the Experiment, Then the Curve — Core
A caption is part of the result
Before interpreting a response, identify the source shock, receiving variable, shock size and sign, response units, and horizon frequency. Then describe impact, timing of a peak, sign changes, persistence, and uncertainty. A curve without those labels does not define an economic experiment.
Consider three different statements:
- Add one unit to a reduced-form innovation in the rate equation.
- Add one standard deviation of an orthogonalized shock associated with the rate equation.
- Normalize an identified policy disturbance so its impact on the observed rate is a fall of 100 basis points.
All can generate legitimate response calculations. They are not automatically the same shock or the same scale. One hundred basis points is one percentage point, not a one-percent proportional change in an interest rate.
Point responses and cumulative responses
The point response answers, “How far is from its baseline at horizon ?” The cumulative response
adds the responses from impact through . If is a level measured in dollars, this is an accumulated dollar-period quantity. It is not the level difference at the final horizon.
cum_r <- apply(r, 2, cumsum)
point_cumulative <- data.frame(h = rep(0:H, 2),
value = c(r[, "y"], cum_r[, "y"]),
quantity = rep(c("Point response: y units", "Cumulative response: y-unit periods"), each = H + 1))
ggplot(point_cumulative, aes(h, value)) + geom_hline(yintercept = 0, colour = "grey70") +
geom_line(colour = "#167d9a", linewidth = 0.9) +
facet_wrap(~ quantity, scales = "free_y", ncol = 2) +
labs(x = "Horizon (periods)", y = NULL)The interpretation depends on what entered the model. If a modeled variable is a first difference, accumulating its response can recover the response of the underlying level. With matching initial levels at ,
For example, responses of a change in a rate measured in percentage points accumulate to a level-rate difference in percentage points. Responses of accumulate to a log-level difference, which is approximately a proportional change when small. Multiplying that log difference by 100 gives an approximate percent change. Summing a response already expressed in levels answers a different question. Always check the transformation before choosing cumulative = TRUE.
Module 8 also connected a cumulative pulse response to a step response. In a linear, time-invariant system, adding the same innovation at every date from the shock onward produces, at horizon , the sum of the corresponding one-time responses through . That is a repeated-input experiment. A single shock does not become permanent merely because we plot its cumulative response.
What linearity and stability tell us
A linear VAR makes responses proportional to shock size: doubling the same shock doubles every response; reversing its sign reverses the curve. In our example a shock gives , and a shock gives . Those are model implications, not empirical proof that monetary easing and tightening have symmetric effects.
For a stable VAR, the response to a one-time shock eventually approaches zero. This says the incremental effect disappears; the observed series can continue fluctuating because later shocks still arrive. A cumulative response can approach a nonzero value even when every individual late response is close to zero. Stability is about lagged propagation, distinct from the invertibility of that makes the current simultaneous solution possible.
A study shows a positive cumulative output response at horizon 20, but its point response is nearly zero by horizon 20. Does that establish a permanent positive difference in output at that date? Which object would answer that question?
12.5 Identification Changes the Shock — Core
Put the generating information out of view
So far we have used the structural truth because we built the simulation. An empirical researcher does not observe that truth. The estimated reduced-form coefficients and innovation covariance do not, by themselves, tell us which combination of innovations is an economic shock.
An impact matrix can express the reduced innovations as , where the components of have unit variance and zero covariance. It must satisfy
For the known generator, . But a different factor of the same generally produces different response columns. The data have not chosen an economic interpretation simply because a covariance factorization succeeds.
Module 11 considered a -first and an -first recursive structure. The -first assumption excludes the current connection; the -first assumption excludes the current connection. Both exclusions are false for our generating system, where the current coefficients are 0.5 and 0.25. Yet both structures reproduce its reduced-form information.
C_truth <- B_inverse # Omega = I in the generating model
C_y_first <- t(chol(Sigma))
reverse_order <- c(2, 1)
C_reordered <- t(chol(Sigma[reverse_order, reverse_order]))
C_x_first <- C_reordered[reverse_order, reverse_order]
dimnames(C_truth) <- dimnames(C_y_first) <- dimnames(C_x_first) <-
list(c("response y", "response x"), c("shock y", "shock x"))
knitr::kable(cbind(Truth = C_truth[, 2],
`y-first x shock` = C_y_first[, 2],
`x-first x shock` = C_x_first[, 2]), digits = 4,
caption = "Impact of one standard deviation of each model's x-associated shock")| Truth | y-first x shock | x-first x shock | |
|---|---|---|---|
| response y | 0.5714 | 0.0000 | 0.8316 |
| response x | 1.1429 | 0.8944 | 1.1780 |
stopifnot(max(abs(C_y_first %*% t(C_y_first) - Sigma)) < 1e-12,
max(abs(C_x_first %*% t(C_x_first) - Sigma)) < 1e-12)R’s chol() returns an upper-triangular factor; the transpose supplies the lower factor used for a -first impact map. For the opposite ordering we factor the covariance in order and then map both rows and columns back to the display order . The names must follow the variables and shocks, not their temporary matrix positions.
Compare the recursive assumptions before revealing the simulation truth. The baseline, reduced dynamics, and reduced covariance remain fixed. The changed impact vector defines a different standardized shock. A one-standard-deviation shock within each structure is the default comparison; it does not turn the three shocks into one identical economic intervention.
Break the interpretation, not the data
The static comparison below repeats the experiment with each impact map. We deliberately maintain identifying restrictions that we know are false, while leaving the reduced system untouched.
maps <- list(`Known simultaneous truth` = C_truth,
`Assume y first` = C_y_first, `Assume x first` = C_x_first)
identified_responses <- array(0, dim = c(H + 1, 2, 3))
for (m in seq_along(maps)) {
identified_responses[1, , m] <- maps[[m]][, 2]
for (h in 1:H) identified_responses[h + 1, , m] <-
A1 %*% identified_responses[h, , m]
}
id_plot <- data.frame(h = rep(0:H, 3),
response = as.vector(identified_responses[, 1, ]),
identification = rep(names(maps), each = H + 1))
ggplot(id_plot, aes(h, response, colour = identification, linetype = identification)) +
geom_hline(yintercept = 0, colour = "grey70") + geom_line(linewidth = 0.9) +
scale_colour_manual(values = c("#c15a32", "#4071a8", "#222222")) +
labs(x = "Horizon (periods)", y = "y response (y units)", colour = NULL, linetype = NULL) +
theme(legend.position = "bottom")Under first, the -associated shock has no impact effect on . Under first, it does. Neither recursive curve recovers the original simultaneous structural experiment. These differences persist even if we know and exactly: they are not sampling error.
The exercise does not say that recursive identification is always wrong. It says that its substantive timing restrictions must be defended. In an application, observation frequency, institutions, information availability, and behavior may support a restriction. A successful factorization cannot do that work. Orthogonality means zero covariance; it does not establish independence in general or give an innovation its economic label.
12.6 Implement the Intended Response in R — Core
Recreate the Module 11 observations
Use the same seeded structural disturbances, sample size, and burn-in. The extra retained starting value lets VAR(p = 1) use the same 2,000 dated regression observations that Module 11 used.
set.seed(637611)
n <- 2000
burn <- 250
eta <- matrix(rnorm(2 * (n + burn)), ncol = 2)
Y <- matrix(0, nrow = n + burn + 1, ncol = 2)
colnames(Y) <- c("y", "x")
Y[1, ] <- mu
for (tt in seq_len(nrow(eta))) {
Y[tt + 1, ] <- A0 + A1 %*% Y[tt, ] + B_inverse %*% eta[tt, ]
}
var_data <- Y[(burn + 1):(burn + n + 1), ]
fit <- VAR(var_data, p = 1, type = "const")
A1_hat <- Acoef(fit)[[1]]
residuals_fit <- residuals(fit)
k_equation <- ncol(fit$datamat) - fit$K
Sigma_hat <- crossprod(residuals_fit) / (fit$obs - k_equation)
P_hat <- t(chol(Sigma_hat))
knitr::kable(A1_hat, digits = 4, caption = "Estimated reduced-form lag matrix")| y.l1 | x.l1 | |
|---|---|---|
| y | 0.3756 | 0.2742 |
| x | 0.1852 | 0.4838 |
The covariance calculation uses the residual degrees-of-freedom divisor used by vars for these orthogonalized responses. A different divisor slightly changes shock standard deviations. The population objects have no hats; , , and are estimates.
What ortho changes
For a reduced-form VAR() fit, the two settings implement different response definitions:
| Call on a reduced-form VAR | Response being computed | When it answers the intended question |
|---|---|---|
irf(fit, ortho = FALSE) |
Columns of ; a one-unit reduced innovation in one equation, other innovation coordinates held at zero | Describing propagation in reduced-form coordinates |
irf(fit, ortho = TRUE) |
Columns of ; one-standard-deviation recursive orthogonal shocks | Maintaining and defending that Cholesky ordering as the shock identification |
irf(structural_fit, ...) for an identified SVAR() object |
Responses using that object’s identified impact matrix | Following the identification actually fitted in the structural model |
With ortho = FALSE, the impact matrix is . The zeros off its diagonal do not establish that the economy has no contemporaneous interactions. They describe the chosen reduced-coordinate experiment. With ortho = TRUE, the impact map comes from the covariance factorization in the fitted variable order. This is not a general “allow contemporaneous effects” switch or a “make the result causal” switch.
irf_reduced <- irf(fit, impulse = "x", response = c("y", "x"),
n.ahead = H, ortho = FALSE, cumulative = FALSE, boot = FALSE)
irf_recursive <- irf(fit, impulse = "x", response = c("y", "x"),
n.ahead = H, ortho = TRUE, cumulative = FALSE, boot = FALSE)
knitr::kable(data.frame(
Definition = c("Unit reduced x innovation", "One-SD y-first x shock"),
`Impact on y` = c(irf_reduced$irf$x[1, "y"], irf_recursive$irf$x[1, "y"]),
`Impact on x` = c(irf_reduced$irf$x[1, "x"], irf_recursive$irf$x[1, "x"]),
check.names = FALSE), digits = 4)| Definition | Impact on y | Impact on x |
|---|---|---|
| Unit reduced x innovation | 0 | 1.0000 |
| One-SD y-first x shock | 0 | 0.8967 |
In this bivariate -first case, the second Cholesky column has only one nonzero entry. Its direction is the same as the unit reduced innovation, but its scale differs. That special feature does not make the two definitions generally interchangeable; the first Cholesky column usually mixes both reduced coordinates. Even with diagonal residual covariance, TRUE and FALSE differ in scale unless the relevant variances are one.
Changing impulse = "x" selects a shock to display; it does not change the ordering. To examine the other recursive assumption, fit the same observations in the other variable order:
fit_x_first <- VAR(var_data[, c("x", "y")], p = 1, type = "const")
irf_x_first <- irf(fit_x_first, impulse = "x", response = c("y", "x"),
n.ahead = H, ortho = TRUE, cumulative = FALSE, boot = FALSE)After mapping variable names back, this is the same reduced forecasting system. Its orthogonalized shocks have a different interpretation. Label the ordering and normalization wherever its curves are shown. n.ahead = 20 includes impact plus twenty later periods, so there are 21 rows.
An identified structural object already carries its restrictions
For an object returned by SVAR(), the vars structural-response method uses the fitted structural impact matrix. Changing ortho does not add or remove another Cholesky decomposition. The relevant choice is to pass the structural object whose restrictions you intend to use, rather than its underlying reduced-form fit. This does not identify the unrestricted simultaneous generator for us: its unknown structure still needs sufficient restrictions.
The package’s structural form uses matrices called A and B; its impact map is solve(A) %*% B. These names are not our course’s intercept and contemporaneous matrix. Translate the model equation before borrowing labels.
In vars 1.6-1, the ortho switch also operates on vec2var objects. For svarest and svecest, the identified structural responses are used under either setting. The VECM objects and their construction belong to Module 13. The software behavior described here was checked against this version’s methods, not inferred from the word “orthogonal.”
ortho = FALSE does not compute a generalized IRF. It returns the reduced propagation coefficients. A generalized response instead compares conditional forecasts after learning about one innovation, allowing the other innovation coordinates to move with it according to their dependence.
For a linear VAR with jointly Gaussian, independent-over-time reduced innovations, a one-standard-deviation generalized response associated with coordinate is
The Gaussian assumption gives the linear conditional mean of the other innovations given coordinate . The same expression describes a linear-projection response more generally, but should not be called an arbitrary distribution’s exact conditional-mean response without further assumptions. The generalized response is invariant to merely reordering variables and mapping their names back. That property does not identify a primitive economic disturbance or justify a causal interpretation. See Pesaran and Shin (1998), and Hamilton §11.4 for the distinction between a coordinate change and conditioning on news.
12.7 Read Uncertainty Around the Response — Core
An estimated curve is not the population curve
The sample estimates both the propagation coefficients and the covariance used for the impact map. If another sample had been observed, its estimated response would generally differ. A confidence band summarizes that sampling uncertainty under the selected specification and identification.
A large response and a precisely estimated response are different. We care about the size in economic units as well as the uncertainty. A band that includes zero does not prove there is no response; it indicates that the procedure does not exclude zero at that horizon and confidence level. A narrow band does not validate the economic timing assumption behind the shock.
Bootstrap intuition
The bootstrap creates repeated pseudo-datasets that imitate the fitted data-generating model. For each dataset, it reestimates the model and recomputes the response path. At a given horizon, the distribution of those recomputed responses supplies interval endpoints.
For the residual bootstrap used here, residuals are sampled as whole vectors across equations. If the and residuals from a date were resampled independently, their contemporaneous dependence would be damaged. Resampling pairs preserves the empirical cross-equation dependence that the covariance and Cholesky calculation use.
We use 1,000 replications, seed 637612, and 95% percentile intervals. These are a reproducible classroom illustration conditional on the fitted model; the replication count is not a validation of frequentist coverage. The identification is the deliberately imposed -first structure, not the known simultaneous truth.
irf_band <- irf(fit, impulse = "x", response = c("y", "x"),
n.ahead = H, ortho = TRUE, cumulative = FALSE,
boot = TRUE, ci = 0.95, runs = 1000, seed = 637612)
band_plot <- data.frame(h = 0:H,
estimate = irf_band$irf$x[, "y"],
lower = irf_band$Lower$x[, "y"],
upper = irf_band$Upper$x[, "y"])ggplot(band_plot, aes(h, estimate)) +
geom_ribbon(aes(ymin = lower, ymax = upper), fill = "#167d9a", alpha = 0.2) +
geom_hline(yintercept = 0, colour = "grey50", linetype = "dashed") +
geom_line(colour = "#167d9a", linewidth = 0.9) +
labs(x = "Horizon (periods)", y = "Estimated y response (y units)")At impact the response of is zero under this ordering in every bootstrap draw. That absence of sampling spread is imposed by identification. It is not strong empirical evidence that the true simultaneous response is zero.
These are pointwise intervals: the coverage statement concerns a specified horizon. Connecting them into a shaded region does not make the region a 95% simultaneous confidence band for the entire path. Looking across many horizons for a zero exclusion is not the same as a calibrated joint test.
The band is also not a prediction interval for future observed . An IRF is a response difference. Forecast uncertainty for an observed future outcome includes future innovations, not just uncertainty in the estimated response coefficients.
For a fitted stationary VAR, center the residual vectors, draw whole vectors with replacement, generate a pseudo-series using the fitted coefficients and an initial history, and refit the same specification. Recompute the chosen impact map and all horizons in each replication. Percentile endpoints are empirical quantiles of the replicated responses separately at each horizon.
The vars::irf() defaults are 100 replications and a 95% interval. We override the replication count and seed explicitly. Its basic residual bootstrap should not be assumed appropriate for every form of heteroskedasticity or misspecification.
For cumulative bands, accumulate each bootstrap response path first, then take quantiles. Adding the lower and upper endpoints of pointwise bands is not the same operation. For a fixed negative rescaling, multiply the curves and both endpoints, then reverse which scaled endpoint is lower. For a data-dependent normalization, apply that normalization inside each bootstrap replication if its estimation uncertainty is to be included.
Imagine that the sample becomes enormous and the bands become very narrow. Which uncertainty in our comparison of two recursive structures would still remain?
12.8 Bianco: From Restrictions to Published Responses — Core
The question before the figure
Timothy Bianco’s Monetary policy and credit flows (2021) asks how monetary policy changes credit flows. An aggregate net change can hide simultaneous expansion of some firms’ debt and contraction of others’. His quarterly Compustat sample runs from 1974Q1 to 2017Q1. Credit creation aggregates debt-weighted positive firm debt-growth rates; credit destruction aggregates absolute negative rates. Both are positive flow measures: a rise in destruction means more credit contraction, not a negative value of the measure.
This is a guided reading of the published analysis. Our simulation supplies a fully reproducible calculation; the paper supplies a substantive example of choosing restrictions and interpreting responses.
Translate the identification into a claim
The model orders the variables as
A lower-triangular impact pattern permits a current shock associated with an earlier variable to affect later variables on impact, but excludes the reverse impact responses. In the following pattern, a star denotes a potentially nonzero entry; columns are ordered shocks and rows are ordered responses:
This is an ordering restriction, not a chain containing only adjacent-variable connections. For example, a policy shock may affect both credit-flow variables on impact, while the first three variables have zero impact responses to that shock. The economic interpretation depends on defending those within-quarter restrictions.
Bianco also excludes the first lag of the policy rate from the price equation, using iterated seemingly unrelated regression to estimate the restricted model. That is an additional lag restriction, not a long-run restriction. It is why estimating an unrestricted VAR and selecting ortho = TRUE would not reproduce the full specification. A direct lag exclusion alone should not be read as a general ban on all indirect dynamic paths.
The paper motivates placing destruction before creation by faster access to existing credit and slower adjustment through repayment or maturity; it also checks the reversed order. These are maintained timing assumptions, not conclusions from a Cholesky factorization. The policy-rate measure incorporates the Wu–Xia shadow rate during the lower-bound period. Bianco, §4.1 and §2.2.
Read the experiment before the findings
Figure 4 follows a one-time, unexpected 100-basis-point policy-rate decrease, or a fall of one percentage point. Horizons are quarters, starting at impact. The model uses two lags selected by BIC. Its bands are 68% mean-bias-corrected residual wild-bootstrap bands with 10,000 repetitions, not the 95% percentile bands in our R example. Their interpretation must retain that confidence level and procedure.
Read the direction and persistence: output rises with a lag, and unemployment falls. Keep the source’s axis scaling visible. We use these panels qualitatively because the available description does not establish the precise transformations needed to translate their magnitudes into economic units.
Separate features imposed by the experiment or restrictions from estimated later responses. The impact normalization names the policy shock. The price response should not be assigned percentage or log-point units merely from an axis multiplier.
The paper’s text supplies a useful quantitative reading of credit creation: it falls by 0.11 percentage points on impact, becomes positive by quarter six, peaks near 0.15 percentage points at quarter eight, and remains positive through quarter twelve. An impact-only summary would miss that sign change. These magnitudes are reported in §4.2; any significance statement must use the paper’s procedure, not treat its 68% bands as a 5% test. Bianco, pp. 11–12.
The lesson is a way to read evidence: name the shock, translate its identification, identify the horizon, then state the response and uncertainty. A plotted response is not proof that its maintained identifying restrictions are true.
Choose the credit-creation panel. Explain why “policy easing reduces credit creation” is incomplete. Then state one identifying assumption required for interpreting this as a monetary-policy response, and one uncertainty statement supported by the displayed band.
12.9 Report What the Response Says — Core
A useful report names the identified shock and assumptions; its size and direction; the receiving variable and units; the horizon or path; and uncertainty. For example, “Under the stated recursive ordering, a one-standard-deviation -associated shock produces an estimated positive response at horizon two, with this pointwise interval.” In our simulation we must add that the imposed ordering is false relative to the known generator. In an application we instead explain the economic argument and its limitations.
The full path matters. An effect can be immediate or delayed, reverse sign, peak well after impact, and fade at a different rate from another response. Those are properties of the specified system experiment. They are why innovation accounting adds information that an isolated reduced-form coefficient cannot provide.
Common Misconceptions — Core
- “The IRF is the future observed path.” It is the deviation from the baseline attributable to a specified shock. The response alone reveals neither the baseline nor the actual outcome level, even when its units are level units.
- “A one-unit shock raises the observed variable by one unit.” Simultaneous impact can change that observed movement; distinguish disturbance units from outcome units.
- “
orthoturns contemporaneous effects on or off.” For reduced VARs it chooses a shock transformation and scale. It does not discover the economy’s timing. - “
ortho = FALSEis generalized or causal.” It returns reduced-coordinate propagation; neither label follows. - “Reordering the variables proves a different forecast model.” Reordering the same unrestricted equations changes Cholesky identification while preserving the reduced forecasting system up to relabeling.
- “A cumulative response is the final level response.” Check whether you accumulated levels, differences, or log differences.
- “A narrow band validates the ordering.” Sampling precision and identification credibility are different issues.
- “A band covering zero proves no effect.” It is a non-exclusion under the chosen procedure, not proof of an exact zero.
- “Pointwise 95% bands cover the whole response path with probability 95%.” Whole-path inference requires a different construction.
Core Practice
- Describe the comparison. In the two-path experiment, which quantities are held fixed after the shock date? Why can the response be negative while both observed paths rise?
- Interpret a supplied step. Given and , explain what each multiplication represents. Why is there no new one-unit disturbance in the horizon-one calculation?
- Read the impact. The extra unit enters only , but the response is . Explain the route without describing sequential within-period rounds.
- Diagnose units. Two plots use the same fitted VAR. One follows a unit reduced innovation; the other a one-standard-deviation Cholesky shock. What must be clarified before comparing their magnitudes?
- Compare identifying assumptions. Explain why the two recursive impact matrices reproduce the same covariance but give different responses. Which restrictions are false in our known generator?
- Choose the R object. You have estimated a structural model using restrictions other than the default ordering. Explain why passing its underlying reduced-form fit to
irf(ortho = TRUE)could answer a different question. What does the switch do for an already identifiedSVAR()object? - Interpret cumulation. A model uses the first difference of a level. Under matching initial levels, what does the accumulated response recover? How does the answer differ when the model already uses the level?
- Read uncertainty. Explain why a zero impact with a zero-width band under a recursive exclusion need not be an empirically established absence of contemporaneous response.
- Read Bianco. Write two sentences about credit creation that identify the shock, distinguish impact from quarter eight, and state the relevant units and uncertainty convention. Give one reason an unrestricted VAR with
ortho = TRUEwould not reproduce his full specification. - Read a policy-rate response. A hypothetical funds-rate IRF is +0.5 percentage point at horizon four. What can you say about the rate if its baseline at that horizon is 3%? What can you say if the baseline is not supplied?
Textbook Connections and Notation
Enders, Applied Econometric Time Series, fourth edition, Chapter 5 §7, pp. 294–300, supplies the main response-construction, orthogonalization, and uncertainty connections. The neighboring FEVD discussion is for Module 13. Enders’s contemporaneous sign convention differs from ours: our scalar equations have positive terms and our contemporaneous matrix has negative off-diagonal entries. Keep one convention throughout a calculation.
Hamilton, Time Series Analysis, §11.4, pp. 318–323, distinguishes reduced-coordinate innovations, orthogonalized responses, and conditional updates. Section 11.6 supplies structural interpretation, and §11.7 discusses sampling uncertainty. Our are structural disturbances, reduced innovations, and fitted residuals. Estimates carry hats. remains the intercept vector.
Bold and here are horizon-indexed matrix coefficients, not the scalar lag polynomials and or the seasonal coefficients of earlier modules. The package’s Phi() returns reduced propagation matrices for a reduced VAR, while Psi() supplies the Cholesky responses. Names alone are not a substitute for reading the model and object class.
For the optional generalized-response discussion, see Pesaran and Shin (1998), “Generalized impulse response analysis in linear multivariate models”. Software references are the vars::irf(), Phi(), and Psi() documentation, checked against the installed vars 1.6-1 methods.
Key Takeaways — Core
- An IRF begins with a defined shock experiment: common history, one changed disturbance, and common subsequent innovations. It gives the deviation from baseline; the baseline and actual outcome level cannot be inferred from the response alone.
- The contemporaneous system determines impact; the reduced lag dynamics propagate it.
- Shock definitions, normalization, transformations, and horizon units are part of the result.
- Structural interpretation remains conditional on identifying assumptions. A factorization or an R option does not establish them.
- Response bands describe sampling uncertainty under a specification; they are neither prediction intervals nor measures of the truth of the restrictions.
- Applied interpretation connects the economic question, restrictions, shock, and full response path.
IRFs trace the consequences of a shock. Module 13 asks how important shocks are for horizon-specific forecast uncertainty through FEVD, returns to Granger-causality testing at the system level, and connects cointegration to VECMs. Cointegrated level systems can have responses that do not decay as in our stable baseline. We will preserve long-run information rather than automatically difference every series.