Dynamic Local Average Treatment Effects

• Dynamic local average treatment effects are local causal parameters defined over treatment paths, cohorts, event times, or horizons, capturing evolving compliance and instrument exposure.
• They are identified using frameworks such as dynamic treatment regimes, staggered DID-IV, and time-series IV models, each relying on assumptions like one-sided noncompliance and sequential ignorability.
• Advanced semiparametric estimation and aggregation methods are essential for accurately estimating these heterogeneous, time-varying effects where traditional IV techniques may fall short.

Dynamic local average treatment effects are local causal parameters for environments in which treatment assignment, instrument exposure, or both evolve over time. In place of the single cross-sectional complier effect of the canonical Imbens–Angrist setup, the relevant object can be indexed by treatment path, exposure cohort and calendar time, event time, or horizon after a time-$t$ perturbation, and the corresponding complier population can also vary with those indices. Recent formulations include dynamic treatment regimes with adaptive encouragements, instrumented difference-in-differences with staggered instrument exposure, and time-series IV models with horizon-specific local effects (Sojitra et al., 2024, Raaschou-Pedersen, 5 May 2026, Miyaji, 2024, Casini et al., 16 Sep 2025).

1. Definitions and sources of dynamic locality

The main unifying idea is that “local” no longer refers only to one instrument-induced treatment margin. In dynamic settings, locality can be path-specific, cohort-specific, event-time-specific, or horizon-specific. Accordingly, dynamic LATEs are not a single estimand but a family of estimands defined on different complier subpopulations.

Framework	Canonical local effect	Defining local population
Dynamic treatment regimes	$\tau_d=\theta(d,d)$, with $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$	Units whose treatment path under encouragement $z$ equals $d$
Staggered IDiD / DID-IV	$CLATT_{e,e+l}$ or $LATT(e,t)$	Exposure cohort $e$, period $t$ or event time $l$, and instrument-induced treated compliers
Time-series IV	$\tau_d=\theta(d,d)$0	Dates in the subset where the first stage exists, at horizon $\tau_d=\theta(d,d)$1, conditional on $\tau_d=\theta(d,d)$2

This indexing matters because papers are explicit that the local effect generally changes as the instrument changes over time, as the exposure cohort changes, or as the horizon changes. In the staggered DID-IV formulation, the dynamic local effect is “not one stable complier effect traced over time” but a sequence of cohort- and event-time-specific effects for potentially different complier populations. In the time-series formulation, the parameter is local both in treatment type and in time, because it is defined only on the subpopulation where the first stage is operative (Sojitra et al., 2024, Miyaji, 2024, Raaschou-Pedersen, 5 May 2026, Casini et al., 16 Sep 2025).

A second unifying theme is that dynamic locality is inseparable from dynamic compliance. In sequential environments, a unit may comply in one period but not another; in staggered-exposure settings, compliance is indexed by exposure cohort and period; in time series, the “complier” can be a date or episode for which the instrument shifts the treatment. This is why dynamic LATEs require more structure than static LATEs even when all variables remain binary.

2. Sequential encouragement and dynamic treatment regimes

One direct formulation of dynamic LATEs studies dynamic treatment regimes with adaptive encouragements and one-sided noncompliance. In the two-period setup the observed data are

$\tau_d=\theta(d,d)$3

where $\tau_d=\theta(d,d)$4 is the period-$\tau_d=\theta(d,d)$5 encouragement, $\tau_d=\theta(d,d)$6 is the realized treatment, $\tau_d=\theta(d,d)$7 is baseline state, $\tau_d=\theta(d,d)$8 is intermediate state, and $\tau_d=\theta(d,d)$9 is the final outcome. The core dynamic estimand is

$\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$0

Here $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$1 is a treatment path and $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$2 is an encouragement path.

The key baseline assumptions are dynamic exclusion through treatment, sequential ignorability of instruments conditional on observed history, sequential overlap, sequential relevance, and one-sided noncompliance. In two periods, one-sided noncompliance is

$\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$3

and in $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$4 periods,

$\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$5

Under these assumptions, the paper shows that dynamic LATEs corresponding to treating at single time steps are nonparametrically identified. For $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$6,

$\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$7

The same logic extends to $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$8 periods for any treatment vector with a single one: $\theta(z,d)=E[Y(d)-Y(0)\mid D(z)=d]$9

The paper also identifies a dynamic mixture parameter,

$z$0

and proves that it is a convex combination of path-specific dynamic LATEs over the non-never-takers induced by $z$1. This matters because baseline assumptions identify some mixtures even when they do not identify every path-specific local effect.

A central negative result is equally important. Sequential monotonicity alone does not identify the dynamic when-to-treat LATE; one-sided noncompliance does essential identifying work. Moreover, one-sided noncompliance is still insufficient to identify the always-treat effect

$z$2

in general. The paper constructs observationally equivalent data-generating processes with different values of $z$3. Thus the cleanly identified objects under baseline assumptions are the single-period dynamic LATEs, not arbitrary multi-period treatment-path effects.

Always-treat dynamic LATEs become identifiable only under additional structure. One route is a cross-period effect-compliance mean-independence condition linking the period-1 treatment effect to later compliance behavior. Another is “staggered compliance,” under which first-period compliers deterministically comply later when encouraged. In the two-period staggered-compliance case, the always-treat effect simplifies to a local Wald-type ratio with an adaptive period-2 encouragement rule, and analogous many-period results are derived for the all-ones treatment path (Sojitra et al., 2024).

3. Staggered instrument exposure and instrumented difference-in-differences

A second major use of dynamic local effects arises in panel or repeated-cross-section settings with staggered instrument exposure. In the basic two-period DID-IV design, the local parameter is the local average treatment effect on the treated,

$z$4

identified by the Wald-DID ratio

$z$5

The multi-period extension defines the cohort-specific local average treatment effect on the treated,

$z$6

so the dynamic local effect is cohort-specific, relative-time-specific, local, and exposed-group-specific.

The identification assumptions combine IV restrictions with DID restrictions: no carryover, exclusion, monotonicity, no anticipation in the first stage, parallel trends in treatment, parallel trends in outcomes, and relevance. A distinctive feature is that the outcome parallel-trends condition is not the usual untreated-outcomes restriction. It is parallel trends in the unexposed outcome path, allowing treatment to evolve endogenously over time even without instrument exposure. This reinterpretation is central to the DID-IV design (Miyaji, 2024).

A closely related staggered-adoption framework defines the cohort-time local parameter

$z$7

Here $z$8 is the first period of instrument exposure, so $z$9 indexes the exposure cohort and $d$0 indexes calendar time. The instrument is absorbing: $d$1 The framework supports two control-group constructions based on instrument exposure,

$d$2

corresponding to never-exposed and not-yet-exposed controls.

Identification requires exclusion/no direct effect/no carryover in outcomes,

$d$3

period-by-period monotonicity,

$d$4

no anticipation in the first stage, overlap, and conditional parallel trends both for the unexposed treatment path and for unexposed outcomes. In panel data the identified parameter is

$d$5

with an analogous repeated-cross-section expression.

A major conceptual result is the Bloom-type relation. Under absorbing treatment and one-sided compliance,

$d$6

Thus, under those additional conditions, the dynamic local effect becomes a convex combination of exposure-cohort-specific dynamic ATT terms. Without one-sided compliance, however, $d$7 remains a genuine complier-type local effect rather than an ATT (Raaschou-Pedersen, 5 May 2026).

4. Time-series formulations

A third formulation treats dynamic LATEs directly in time series. The observed process is

$d$8

with histories summarized by

$d$9

The framework defines horizon-$CLATT_{e,e+l}$0 potential outcomes after perturbing the time-$CLATT_{e,e+l}$1 treatment or instrument, so the target can be interpreted as an IV impulse response at horizon $CLATT_{e,e+l}$2.

The paper’s time- and horizon-specific local parameter is the generalized Wald estimand

$CLATT_{e,e+l}$3

Identification requires conditional independence of the instrument given $CLATT_{e,e+l}$4, exclusion,

$CLATT_{e,e+l}$5

a partial first-stage condition on an unknown subset $CLATT_{e,e+l}$6 of dates, monotonicity on that subset, and local absolute continuity. Under these conditions, the generalized Wald estimand equals a weighted average of marginal treatment effects for compliers in the relevant subpopulation.

The most distinctive result is that, with a binary instrument and smoothness assumptions, individual compliers can be identified in time series data. Under a deterministic complier-status assumption and local LLN/smoothness conditions, a date $CLATT_{e,e+l}$7 is a complier if and only if

$CLATT_{e,e+l}$8

and a date $CLATT_{e,e+l}$9 is a complier if and only if

$LATT(e,t)$0

This makes the relevant local subpopulation observable in a way unavailable in the standard cross-sectional LATE framework.

Once non-compliers can be identified, exclusion becomes directly testable on subsets of non-compliers. Under additional comparability and LLN conditions, the difference in average outcomes for identified non-compliers across instrument values converges to zero under the null. The paper therefore proposes a $LATT(e,t)$1-test comparing average outcomes among identified non-compliers across policy and control observations.

The same framework treats weak identification as a time-varying first-stage problem rather than a purely global one. It defines a strongest-subsample statistic,

$LATT(e,t)$2

and uses dynamic programming to search for the most strongly identified subsample. In the application to heteroskedasticity-based identification of monetary policy effects, the paper finds that about $LATT(e,t)$3 of observations are compliers and fails to reject the exclusion restriction (Casini et al., 16 Sep 2025).

5. Estimation and inference

The recent dynamic-LATE literature is also a literature on semiparametric estimation. In dynamic treatment regimes with one-sided noncompliance, the proposed estimators are automatic debiased machine learning procedures based on nuisance regressions, Riesz representers, Neyman-orthogonal moments, and cross-fitting. For two-period when-to-treat effects,

$LATT(e,t)$4

and the paper establishes consistency and asymptotic normality under strong overlap, nuisance consistency, and product-rate conditions. For the staggered-compliance always-treat estimand, it constructs an analogous orthogonalized ratio estimator (Sojitra et al., 2024).

In staggered IDiD, the semiparametric contribution is explicit. The paper derives efficient influence functions for $LATT(e,t)$5 in both panel and repeated-cross-section settings, constructs doubly robust estimands, and then builds DML estimators via cross-fitting. In panel data the doubly robust estimand is

$LATT(e,t)$6

and the paper proves double robustness in the sense that consistency holds if either the propensity model is correct or the relevant regression models are correct. It also shows that the DML estimators are equivalent to the cross-fitted estimating-equation estimators based on the EIF (Raaschou-Pedersen, 5 May 2026).

Dynamic local effects create nonstandard inference problems when different instruments, periods, or timing margins identify different complier effects. In a static but directly relevant result, overidentified 2SLS with multiple instruments is shown to converge to a weighted average of distinct instrument-specific LATEs, and the postulated overidentified moment condition does not generally hold at that pseudo-true value. Consequently, the conventional heteroskedasticity-robust 2SLS variance estimator is inconsistent, and a Hall–Inoue-type misspecification-robust variance estimator is required. The same logic transfers immediately to dynamic applications with multiple period-specific or timing-specific instruments, where pooled IV coefficients can average distinct dynamic LATEs (Lee, 2018).

Weak identification remains a separate issue. A static LATE paper with high-dimensional covariates develops a conditional QLR-style procedure with Neyman-orthogonal scores, lasso-based nuisance estimation, and cross-fitting, delivering uniformly correct asymptotic size irrespective of the strength of identification. That paper is not dynamic, but it is methodologically relevant because dynamic LATE designs often feature even weaker period-specific first stages and richer state vectors. This suggests that dynamic extensions will likely require dynamic orthogonal scores together with identification-robust test inversion rather than standard Wald approximations (Ma, 2023).

6. External validity, aggregation, and recurring misconceptions

Dynamic LATE inherits the classical external-validity problem of LATE and typically intensifies it. In the static latent-index framework, the standard critique is that the LATE is intrinsically local, the treated subgroup is policy-specific, and the complier group depends on the instrument. With a binary instrument, observed variation identifies treatment responses only on the complier interval of the latent resistance variable and does not directly identify always-takers, never-takers, or other policy-relevant weighted averages. The paper therefore frames standard LATE as lacking external validity “in counterfactual environments” (Han et al., 2020).

The same critique applies a fortiori in dynamic settings. A period-$LATT(e,t)$7 local effect is tied to a specific source of exogenous variation at a specific time and history. In staggered DID-IV, the relevant margin is cohort- and event-time-specific. In time series, the meaningful local parameter can be restricted to the subset of dates where the first stage exists. In sequential dynamic treatment regimes, different encouragement paths induce different path-specific complier groups. A dynamic LATE is therefore not a policy-invariant effect.

This directly motivates aggregation. In staggered IDiD, general summaries are written as

$LATT(e,t)$8

and event-study summaries such as

$LATT(e,t)$9

aggregate cohort-time local effects at a fixed horizon $e$0. In the earlier DID-IV formulation, the recommended aggregation weights are based on the size of the first stage in each cohort-period cell, so the summary parameter is explicitly weighted by the share of compliers in that cell. These are transparent weighting schemes, not implicit regression averages (Miyaji, 2024, Raaschou-Pedersen, 5 May 2026).

Several recurrent misunderstandings follow from ignoring this structure. One is that a dynamic local effect is “the” treatment effect traced over time. The staggered DID-IV papers reject that interpretation: each $e$1 or $e$2 may vary because treatment effects vary over time, because treatment effects vary across cohorts, and because the set of compliers changes across $e$3. A second misunderstanding is to conflate dynamic ATT and dynamic LATE. Under one-sided compliance and absorbing treatment, $e$4 becomes a convex combination of $e$5 terms, but without those conditions it remains a complier-specific local effect rather than an ATT. A third misunderstanding is that pooled TWFE-IV coefficients provide a meaningful dynamic summary. The DID-IV paper is explicit that conventional TWFE-IV can use already-exposed cohorts as controls, inducing bad comparisons and negative weights under heterogeneity (Miyaji, 2024, Raaschou-Pedersen, 5 May 2026).

The broader methodological implication is that dynamic LATE research sits between IV logic and dynamic policy evaluation. The static policy-evaluation literature shows that many counterfactual parameters can be written as weighted averages of marginal treatment effects, with policy-specific weights over latent resistance. This suggests a general lesson for dynamic work: local causal effects are often only the first step, and moving from local sequential IV effects to counterfactual policy effects will generally require explicit aggregation, stronger structure, or partial identification rather than simple transport of a single local estimate (Han et al., 2020).