State-Dependent Local Projections and Vector Autoregressions

arXiv:2601.01622 · econ.EM · Submitted 2026-01-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "State-Dependent Local Projections and Vector Autoregressions".

Warren: Here's the gist of the paper. State-dependent local projections estimate weighted averages of conditional marginal effects This paper shows that state-dependent LPs estimate weighted averages of conditional marginal effects The result holds without making parametric assumptions and the shock of interest is allowed to influence current and future realizations of the state.

Ruth: Characterization of State-Dependent LPs The first part establishes a general characterization that under minimal assumptions, statedependent LPs recover weighted averages of causal effects This holds for essentially all specifications used in practice.

Warren: Second, the paper shows that state-dependent LPs and VARs target different estimands and proposes a simple VAR-based estimator whose probability limit equals the LP estimand.

Ruth: Third, in instrumental variable (LP-IV) settings, state-dependent weighting can generate nonzero interaction terms, even when the effects are not state-dependent.

Title and authors: Ruth: So we're looking at this paper today titled "State-Dependent Local Projections and Vector Autoregressions," and the authors are Valentin Winkler and Serena Ng. Basically, they're looking at how to use local projections when you have a state variable that can change over time.

Warren: It sounds like they’re trying to figure out if this method of using state-dependent local projections actually gives you what you expect when things get complicated in the real world.

Iris: They're focusing on how these projections estimate weighted averages of causal effects, which is a nice starting point because it suggests a general rule might apply to almost any model setup they use.

Ruth: Exactly. The authors establish this characterization first and say it holds under minimal assumptions for essentially all the specifications people actually use in practice, which is important because we often rely on these projections without knowing exactly how the true data generating process looks.

Warren: So they’re saying that if you use these state-dependent LPs, you should be able to recover those weighted averages of causal effects regardless of your specific model setup, even if it's a bit messy.

Iris: They also point out that state-dependent local projections and vector autoregressions are actually targeting different things, which is a key distinction we need to keep in mind when interpreting the results.

Ruth: Right. And they propose a simple vector autoregression estimator whose probability limit equals the same thing as the local projection estimand, which gives us an alternative way to get that effect estimate if the direct LP approach doesn't work out nicely.

The paper's summary: Warren: What they summarize is that state-dependent LPs are useful, but there’s a catch: they target different estimands than vector autoregressions, so you need to be careful about what you think you’re measuring.

Iris: They show that in instrumental variable settings, state-dependent weighting can generate non-zero interaction terms even when the effects themselves aren't actually state-dependent, which is a tricky thing for identification.

Ruth: That’s interesting because we often look for those interaction terms as proof of state dependence, but the paper suggests that in IV settings, those non-zero coefficients might just be caused by how the weights are set up across different states.

Warren: They then compare these state-dependent LPs with their vector autoregression counterparts and find that in a linear case, they asymptotically yield the same effect estimates, which is good for consistency.

Iris: But they use a simple dynamic stochastic general equilibrium model to show that this equivalence breaks down in the state-dependent case, meaning it doesn't hold even when we assume the state follows a known process.

Ruth: So what they do as a remedy is introduce an impulse response estimate constructed from multiple state-dependent vector autoregressions, which they say is easy to build and asymptotically matches the same effect estimate you get from the local projection approach.

The paper's improvements: Iris: One big improvement they suggest is this new impulse response estimate, which I think is really useful because it bypasses the asymptotic breakdown between LPs and VARs in state-dependent settings.

Warren: That sounds like a concrete tool we could actually use to recover the causal effects when we can't rely on the standard LP approach being equivalent to VARs in this complex setup.

Ruth: And they also derive some specific properties for these state-dependent local projection estimands, showing that for common functional forms like binary or continuous states, the interaction term captures a specific difference between average conditional effects with the same weighting function for both states.

Iris: They give a mathematical property for the estimand beta h of the state-dependent local projection, saying it equals one expectation minus another expectation involving lagged terms and a term related to theta h(S t-one; omega X) <ref:2601.01622#pg2>.

Warren: That formula is dense, but essentially it tells us that the effect we are interested in is just a difference between two expected values of certain lagged variables.

Ruth: And when we get to instrumental variables, they show that the weights on the effects depend on the joint distribution of your instrument and your regressor, which I think is where a lot of people misinterpret those coefficients as being purely about state dependence.

Conclusion: Iris: So to wrap up, they say that generally state-dependent local projections are just as valid as linear LPs because if the shock you're interested in is observed, the weights on the causal effects are identical across states and applications.

Warren: But they issue a warning: if you use instrumental variables, those non-zero interaction coefficients can come from differences in your weighting scheme that have nothing to do with whether the effect itself is state dependent.

Ruth: So to summarize the paper "State-Dependent Local Projections and Vector Autoregressions," it says state-dependent LPs are valid, but you have to be careful when using instrumental variables because of those potential misinterpretations in the weighting scheme.

Warren: They also introduce a specific impulse response estimate that matches the LP estimand asymptotically, which is a nice practical fix for when things don't line up perfectly with VARs.

Iris: I just want to add that they flag that you need to take those independence conditions seriously and test them empirically because relying on those assumptions can lead to misinterpretation of the results.

Ruth: That’s the core message: state-dependent LPs are a valid tool, but you have to interpret those interaction terms with a very clear understanding of what they represent in terms of state dependence versus weighting differences.

Warren: It gives us a way forward for people trying to estimate these effects without committing to one structural model, by providing that impulse response estimate as an asymptotic match.

Columbia University

econ.EM

Submitted: 2026-01-04

Updated: 2026-11-09

Comments: 46 pages, 7 figures. Includes the Supplemental Appendix. Previously titled "When and Why State-Dependent Local Projections Work"

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

The gist: State-dependent local projections estimate weighted averages of conditional marginal effects This paper shows that state-dependent LPs estimate weighted averages of conditional marginal effects The…

Key concepts

State-Dependent LPs
These are statistical tools used to estimate weighted averages of causal effects where the weighting depends on the current state of a system. They are shown to recover these averages under minimal assumptions, making them robust for various practical specifications.
Interaction Term
In common functional forms, this term captures the difference between average conditional effects when using different weighting functions for different states. For continuous states, it relates the estimand to covariance terms involving the state variable and future outcomes.
LP-IV Weights
When using instrumental variables in a state-dependent setting, the weights applied to causal effects depend on the joint distribution of both instruments and regressors. The second weight specifically indicates compliance, showing how strongly a treatment reacts to the instrument.

Terminology

Summary

State-dependent local projections estimate weighted averages of conditional marginal effects This paper shows that state-dependent LPs estimate weighted averages of conditional marginal effects The result holds without making parametric assumptions and the shock of interest is allowed to influence current and future realizations of the state. p. 1

Characterization of State-Dependent LPs

The first part establishes a general characterization that under minimal assumptions, statedependent LPs recover weighted averages of causal effects This holds for essentially all specifications used in practice. Second, the paper shows that state-dependent LPs and VARs target different estimands and proposes a simple VAR-based estimator whose probability limit equals the LP estimand. p. 0 Third, in instrumental variable (LP-IV) settings, state-dependent weighting can generate nonzero interaction terms, even when the effects are not state-dependent. p. 0

Relationship to VARs

The paper compares state-dependent LPs with their VAR counterparts and shows that in the linear case, those two methods asymptotically yield the same effect estimates. Using a simple DSGE model, it is shown with simulations and analytically that this equivalence breaks down in the statedependent case. p. 2 As a remedy, an impulse response estimate constructed from multiple statedependent VAR models is introduced, which is easy to construct and asymptotically matches the state-dependent LP estimand. p. 2

State-Dependent LP Estimands

The paper derives that for common functional forms, such as binary states or continuous states, the interaction term captures the difference between average conditional effects with the same weighting function for both states. For a continuous state, the estimand satisfies βh1 = Cov[˜f(St−1), θh(St−1; ωX)] V[˜f(St−1)]. p. 5 Proposition 3 shows that the estimand βh of the state-dependent local projection (10) has the following property: βh = E[(ft−1Xt)(ft−1Xt)′ −1 E [(ft−1Xt)Yt+h]] = E[ft−1f′t−1 − 1E [ft−1θh(St−1; ωX)]. p. 8

State-Dependent LP-IVs and Interpretation

When using instrumental variables, the weights on the effects depend on the joint distribution of instrument and regressor. The result θIV,h(s) is an integral over a product of three components that are plotted in Figure 4. p. 1 The second weight in (32) can be understood as indicating compliance, i. p. 0e., how strong the treatment reacts to the instrument. p. 2

Conclusion and Caveats

The paper concludes that generally state-dependent LPs are just as valid as linear LPs. If the shock of interest is observed, the weights on the causal effects are identical across states and applications. p. 1 A non-zero interaction coefficient implies state dependence of the effect of interest. p. 1 However, when using instrumental variables, non-zero interaction coefficients in state-dependent LP-IVs can be due to differences in the weighting scheme that have nothing to do with the effect of interest. p. 2 The required independence conditions should be taken seriously and tested empirically. p. 2

Appendix Summary

Table 2 lists some properties of the 44 applied studies using state-dependent LPs listed by Gon¸calves et al. (2024a). This shows that the majority of the studies (29/44) uses a lagged state variable. p. 8 Also only 15/44 studies restrict themselves to one binary state variable, while 19/44 use a continuous state variable. p. 9 Lastly, 19/44 of the state-dependent LP papers use 2SLS methods. p. 3 The specification of this paper is broad enough to cover virtually all estimating equations used in the applied studies. p. 1

Illustrative Example

A smooth transition VAR model is studied where the setup allows for an analytical computation of marginal effects while featuring a rich heterogeneity of causal effects and a continuous state variable. The simulation exercise confirms Proposition 1, showing that the LP estimates correspond to the averages of the marginal effect distributions. p. 1

Solving the DSGE Model

The simple DSGE growth model is solved numerically yielding (ϕ˜(0), ϕ˜(1)) ≈ (0.23, 0.14). Therefore, income evolves approximately as Yt = A(St)ϕ(St−1)Yt−1 + ν + B(St)νXt. p. 1 The Euler equation from the social planner solution is C−1/σt = βEt[C−1/σt At+1]. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved causal effect estimation for state-dependent local projections by using a VAR-based estimator that asymptotically matches the LP estimand, providing an easy-to-construct alternative when asymptotic equivalence between LPs and VARs breaks down in state-dependent settings.

  2. Enhanced interpretation of instrumental variable results by clarifying that non-zero interaction coefficients in state-dependent LP instrumental variables can arise from differences in the weighting scheme across states rather than genuine state dependence of the effect of interest.

  3. Development of a novel impulse response estimate, theta b VAR,h(s), which is identical to the state-dependent LP estimand under assumption EXO3, offering a robust way to recover causal effects when conventional VAR-based estimates differ.

  4. Creation of a framework for interpreting state-dependent LP-IV coefficients by decomposing the estimated estimand into components related to the effect of interest and components related to the weighting scheme, highlighting where misinterpretation occurs.

  5. A method for deriving Local Average Treatment Effects (LATE) from state-dependent LP-IVs by showing that in specific scenarios, these estimators correspond to LATE in microeconometric terms, linking macro econometrics to limited compliance concepts.

Related papers