Local Instrumental Variable Estimator

• Local Instrumental Variable Estimator is a method that defines threshold-specific causal effects using continuous instruments and latent compliance thresholds.
• It relies on derivative ratios of outcome and treatment functions under monotonicity and robust identification assumptions to pinpoint local treatment effects.
• Recent advances employ semiparametric and doubly robust estimation techniques, accommodating high-dimensional settings and complex model structures.

Searching arXiv for recent and foundational papers on local instrumental variables, continuous instruments, and related IV estimands. arXiv search query: "local instrumental variable continuous instrument doubly robust" The local instrumental variable (LIV) estimator is an instrumental-variable estimator for causal effects with a continuous instrument, defined through a latent-threshold or “local complier” interpretation. Under monotonicity, treatment can be written as $A^z = 1(z \ge T)$, where $T$ is an unobserved threshold; the LIV curve then targets the average treatment effect among units whose treatment status switches at a given instrument value. In one formulation, the target is

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$

the effect among individuals who would take treatment right at instrument value $t$, possibly indexed by effect modifiers $V$. Equivalent marginalized formulations define

$$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$

Across this literature, LIV is identified by a ratio of derivatives of outcome and treatment regression functions with respect to the instrument, and modern work develops semiparametric doubly robust, nonparametric, and high-dimensional estimators, alongside several adjacent IV methods that are explicitly not LIV in the usual econometric sense (Kennedy et al., 2016, Zeng et al., 3 Apr 2025).

1. Estimand and causal interpretation

The defining feature of LIV is that it targets a threshold-specific causal effect rather than an average treatment effect for the full population. Under the latent-threshold representation

$A^z = 1(z\ge T),$

the LIV curve

$$\gamma(t,v)=E\!\left(Y^1 - Y^0 \mid T=t,\; V=v\right)$$

is the average treatment effect among those who would be induced to take treatment above, but not below, the instrument threshold $t$. This gives the estimator its local character: the conditioning event is not observed compliance under a binary shift, but infinitesimal responsiveness to the continuous instrument at $t$ (Kennedy et al., 2016).

A related marginalized formulation defines

$T$0

which the literature describes as the continuous-IV analogue of the LATE idea. In the binary-IV case, LATE is the effect among compliers induced by moving from $T$1 to $T$2; in the continuous-IV setting, the LIV curve localizes the effect at each threshold value $T$3. The same work makes explicit that

$T$4

so the LIV curve is an infinitesimal complier effect (Zeng et al., 3 Apr 2025).

The threshold-based interpretation also distinguishes LIV from a continuous-instrument LATE surface. One paper emphasizes that the LIV curve is a one-dimensional function of threshold $T$5, whereas a continuous-instrument LATE analogue would be a higher-dimensional effect surface comparing two instrument levels $T$6. This difference is substantive rather than notational: LIV indexes the treatment effect by the latent switching threshold, not by a pair of observed instrument values (Kennedy et al., 2016).

2. Identification by derivative ratios

Identification of the LIV curve combines standard IV assumptions with a continuous-instrument monotonicity structure. The core assumptions used in the semiparametric continuous-IV literature are consistency, positivity, unconfoundedness of the instrument conditional on covariates, exclusion restriction, and monotonicity

$T$7

Under these conditions, monotonicity implies the latent-threshold representation $T$8. Identification further requires instrumentation, expressed as

$T$9

on the target set, and continuity of $$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$0 and $$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$1 in $$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$2 (Kennedy et al., 2016).

The central identification formula is the derivative ratio

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$3

The denominator is the derivative of the conditional mean treatment curve and identifies the density of the latent threshold,

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$4

so the numerator can be read as the threshold density multiplied by the local effect (Kennedy et al., 2016).

A closely related nonparametric formulation introduces nuisance regressions

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$5

and defines the dose-response functions

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$6

Under the IV assumptions and continuity of $$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$7,

$$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$8

This representation makes explicit that LIV estimation is a derivative-estimation problem for two dose-response functions (Zeng et al., 3 Apr 2025).

The same paper identifies a second continuous-IV estimand when the instrument is bounded, especially $$\gamma(t,v)=E\!\left(Y^1-Y^0 \mid T=t,\; V=v\right),$$9: the maximal complier class $t$0. Its treatment effect is

$t$1

This is not a local derivative estimand; it generalizes binary-IV LATE to continuous instruments through boundary contrasts rather than threshold localization (Zeng et al., 3 Apr 2025).

3. Semiparametric and nonparametric estimation

A major methodological contribution in the recent literature is a semiparametric doubly robust estimator of the LIV curve for continuous instruments. Rather than treating a parametric LIV curve as exactly correct, one paper defines the target parameter as the weighted least-squares projection

$t$2

with nuisance functions

$t$3

The estimator solves the empirical estimating equation

$t$4

where $t$5 is the efficient influence function. A key technical device is integration by parts, which expresses the influence function in terms of derivatives of the known working model and weight, rather than derivatives of unknown regression functions. The estimator is doubly robust in the sense that

$t$6

and root-$t$7 inference is available when

$t$8

The same work allows sample splitting or cross-fitting to remove Donsker requirements and accommodate highly adaptive learners such as random forests (Kennedy et al., 2016).

A separate nonparametric line treats LIV estimation as direct estimation of derivatives of dose-response functions. One approach constructs a doubly robust pseudo-outcome

$t$9

satisfying

$V$0

if either the instrument density or the outcome regression is correct. The derivative is then estimated locally by

$V$1

The same paper develops a smoothing-based doubly robust derivative estimator, bandwidth selection via a doubly robust pseudo-risk, and plug-in variance estimation based on influence functions. It emphasizes that derivative estimation is harder than dose-response estimation because it has variance scale $V$2 rather than $V$3 (Zeng et al., 3 Apr 2025).

An adjacent nonparametric IV paper extends smoothing splines to endogenous regressors and directly estimates both a global structural function $V$4 and its derivative $V$5. The estimator is one-step, uses a single regularization parameter, and yields a natural cubic spline with a closed-form linear-system solution. However, the paper is explicit that it is not the classic LIV derivative-based estimator used in treatment-effect settings; its identification and estimation are based on the conditional moment restriction $V$6, not on a LIV-specific local treatment effect construction (Beyhum et al., 2023).

4. Binary instruments, binary outcomes, and discrete validity structures

Although LIV is most closely associated with continuous instruments, the binary-IV literature provides its canonical discrete analogue through LATE. Under exclusion, independence, relevance, positivity, and monotonicity, the conditional local average treatment effect is

$V$7

For binary outcomes, the same framework also defines a multiplicative local average treatment effect

$V$8

One paper develops a variation-independent parameterization of the binary IV likelihood in terms of the target $V$9, nuisance functions $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$0, and the complier odds product $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$1. The mapping from the observed-data law to

$$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$2

is a smooth bijection, which supports unconstrained maximum likelihood estimation and truly doubly robust estimating equations for both additive and multiplicative LIV estimands (Wang et al., 2020).

High-dimensional extensions preserve the LATE target while changing the nuisance-estimation architecture. A model-assisted approach with binary instrument $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$3, binary treatment $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$4, and covariates $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$5 uses regularized calibrated estimation for the instrument propensity score and regularized weighted likelihood for treatment and outcome regressions. The resulting estimator is a doubly robust augmented inverse probability weighted ratio, and the main theorem establishes valid Wald confidence intervals under suitable sparsity conditions if the instrument propensity score model is correctly specified, even when the treatment and outcome regression models are misspecified (Sun et al., 2020).

For multivalued discrete instruments, validity need not hold for every pair of instrument values. The VSIV framework introduces pairwise valid instruments, meaning validity only for a subset of ordered value pairs $$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$6. For each valid pair, the pair-specific LATE remains the familiar Wald ratio

$$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$7

VSIV uses testable implications of IV validity to screen out invalid pairs, estimates the remaining pair-specific LATEs, and allows aggregation with researcher-specified weights. The paper proves asymptotic normality and shows that the procedure removes or reduces the asymptotic bias relative to standard LATE estimators that do not remove invalid variation (Sun et al., 2022).

5. Extensions to survival outcomes and weighted derivative functionals

LIV ideas extend beyond uncensored mean outcomes. For survival analysis, one paper targets the local average treatment effect on survival probabilities under both nonignorable and ignorable censoring. With a binary instrument, the target is

$$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$8

and with a continuous instrument it defines a local-shift analogue

$$\gamma(z_0)=\mathbb E\!\left[Y^{a=1}-Y^{a=0}\mid T=z_0\right].$$9

Identification is again ratio-based, and the paper develops efficient-influence-function estimators that are doubly robust and compatible with sample splitting, cross-fitting, and machine-learning nuisance estimation (Lee et al., 2020).

A more recent generalization preserves the derivative logic of LIV but changes the target estimand. For continuous instruments, it defines conditional weighted average derivative effects (CWADEs); for categorical instruments, conditional weighted average treatment effects (CWATEs); and it unifies them through a conditional Riesz representer. In the continuous-IV case,

$A^z = 1(z\ge T),$0

a weighted conditional Wald estimand. The marginal target is

$A^z = 1(z\ge T),$1

so the estimand is an ATE rather than a local threshold-specific effect. The same paper derives a semiparametric observed-data model with a nontrivial tangent space and proposes a locally efficient, triply robust, bounded estimator. This is LIV-like in identification logic, but not LIV in target parameter (Dong et al., 16 Oct 2025).

6. Scope, neighboring methods, and recurring misconceptions

Several IV methods are often discussed alongside LIV even though they do not estimate a local threshold-specific treatment effect. One semiparametric paper on the marginal average treatment effect on the treated uses an instrumental variable to identify

$A^z = 1(z\ge T),$2

rather than a complier-specific or local causal effect. It does not require monotonicity or a compliance class interpretation; instead, its identifying structure is based on a model for selection bias through the extended propensity score $A^z = 1(z\ge T),$3 or the generalized odds-ratio function $A^z = 1(z\ge T),$4 (Liu et al., 2015).

Other IV developments are local only in name or in temporal structure. Instrumental processes based on integrated covariances identify aggregate causal effects such as

$A^z = 1(z\ge T),$5

using long-run covariance matrices in VAR models and Hawkes processes. That framework is explicit that it does not develop a local or marginal treatment-effect framework of the Imbens–Angrist/MTE/LIV type (Mogensen, 2022). Quasi-Bayesian local projections with an instrumental variable estimate dynamic impulse responses through LP-IV moment conditions and a Laplace-type quasi-posterior, but this is local projections with an IV, not the derivative-based LIV estimator in treatment-effects analysis (Tanaka, 26 Mar 2025). In confounded linear MDPs, the IV estimator is described as conceptually LIV-like because it uses an IV to recover structural effects under endogeneity, yet the paper emphasizes that it is not a classical LIV estimator: there is no local derivative with respect to the instrument, no marginal treatment effect interpretation, and the method is embedded in a Bellman system rather than a single-equation threshold model (Lu et al., 2022).

A plausible implication is that “local instrumental variable estimator” now names a family of derivative-based and complier-local estimators centered on threshold-specific causal effects, while a broader IV literature reuses some of the same analytical machinery—ratio identification, orthogonal moments, influence functions, cross-fitting, and robustness—without preserving the classical LIV estimand. Within that broader landscape, the distinctive content of LIV remains the combination of continuous-instrument monotonicity, latent thresholds, and derivative-ratio identification of effects for local compliers (Kennedy et al., 2016, Zeng et al., 3 Apr 2025).