Granular Instrumental Variables (GIV)

• Granular Instrumental Variables (GIV) are a method that leverages cross-sectional heterogeneity to isolate idiosyncratic shocks as valid instruments for endogenous aggregate variables.
• The methodology constructs instruments by removing latent common components via factor analysis, ensuring that only granular idiosyncratic variation influences the estimation.
• Empirical applications demonstrate that GIV techniques effectively address endogeneity across regimes determined by unit size concentration, impacting instrument strength and inference.

Granular Instrumental Variables (GIV) denotes an instrumental-variables methodology that exploits cross-sectional heterogeneity in panels with latent common shocks to construct instruments for endogenous aggregate regressors from idiosyncratic shocks that survive aggregation (Banafti et al., 2022, Hahn et al., 12 Jun 2026, Ramachandran, 2 Jul 2026). In the canonical formulation, an aggregate structural relationship is estimated by using weighted combinations of unit-level observables that are orthogonal to the factor-loading space, so that common components are removed while granular idiosyncratic variation remains relevant for the endogenous aggregate variable (Hahn et al., 12 Jun 2026, Banafti et al., 2022). In adjacent literatures, the acronym also appears in related but distinct senses, including Group Instrumental Variable in data fusion and interaction-based constructions for invalid-IV robustness; these usages share the objective of building valid moment conditions from structured heterogeneity, but they do not coincide exactly with the macro-panel granular-IV framework (Wu et al., 2022, Sun et al., 2021).

1. Classical instrumental-variable foundations and testability

The classical nonparametric instrumental process is written as

$x = g(z,u), \qquad y = h(x,u),$

with the key assumption that $Z$ is independent of $U$ (Pearl, 2013). In this formulation, exclusion is built in by construction because $Z$ does not appear in the equation for $Y$, so $Z$ may influence $Y$ only through $X$ (Pearl, 2013). For discrete $X,Y,Z$, Pearl derives the instrumental inequality

$$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$

a necessary condition for $Z$0 to be generated by such an instrumental process (Pearl, 2013).

This matters for GIV because granularity does not eliminate the classical IV requirements of relevance, exclusion, and exogeneity. Pearl’s contribution is that latent-variable causal models may induce no conditional independences among observables, yet still imply inequality constraints on the observed distribution; a candidate instrument can therefore be falsified even when standard conditional-independence diagnostics are silent (Pearl, 2013). In discrete settings, Sharma extends this line of reasoning by comparing Bayesian generative classes of valid-IV and invalid-IV models and computing a Validity Ratio after a Pearl-Bonet necessary test is passed (Sharma, 2018). The resulting “Necessary and Probably Sufficient” test does not make IV validity universally testable, but it shows that data-driven validation and search over candidate instruments is possible in a meaningful subset of settings, especially with discrete variables and weak-to-moderate instruments (Sharma, 2018).

For GIV designs, this provides a useful boundary condition. A granular construction can deliver relevance through cross-sectional concentration, but validity still depends on whether the shock-based instrument is orthogonal to the structural disturbance after common components are removed. This suggests that GIV should be understood as a structured IV design rather than an exemption from standard IV scrutiny.

2. Canonical GIV setup and construction from granular shocks

A canonical GIV application studies an aggregate demand equation

$Z$1

together with a disaggregated supply panel

$Z$2

where $Z$3 is endogenous because it responds to $Z$4, and aggregate clearing implies

$Z$5

(Ramachandran, 2 Jul 2026). If the idiosyncratic supply shocks $Z$6 can be isolated, then the share-weighted average

$Z$7

is a valid instrument for $Z$8 (Ramachandran, 2 Jul 2026).

The same logic appears in the high-dimensional theory of GIV. In a simultaneous-equations panel with factor error structure, the generalized granular instrument is

$Z$9

so the instrument is a weighted sum of idiosyncratic shocks after purging latent factors (Banafti et al., 2022). In the simplified homogeneous-loading case, the construction reduces to

$U$0

which is the difference between size-weighted and equal-weighted aggregates (Banafti et al., 2022).

A closely related stylized model writes

$U$1

and defines a basic GIV as

$U$2

with

$U$3

(Hahn et al., 12 Jun 2026). More generally,

$U$4

showing that there are many valid GIVs in the simple model and therefore potential overidentification (Hahn et al., 12 Jun 2026).

Across these formulations, the unifying idea is that aggregate endogeneity is addressed by exploiting the fact that idiosyncratic shocks to sufficiently important units do not wash out in the aggregate. The challenge is not generating variation per se, but separating granular idiosyncratic variation from latent common shocks and simultaneity.

3. Identification geometry: the orthogonal complement of the factor-loading space

A central identification result states that valid GIVs are characterized by the orthogonal complement of the factor-loading space (Hahn et al., 12 Jun 2026). In the latent-factor model

$U$5

if

$U$6

and $U$7 denotes an orthonormal basis for the orthogonal complement of $U$8, then valid instruments are generated by

$U$9

(Hahn et al., 12 Jun 2026). This gives a geometric characterization of the “GIV space”: admissible instruments are precisely those cross-sectional contrasts orthogonal to both the constant vector and the factor-loading space.

The same paper shows that this space can be identified directly from second moments. For demeaned observables

$Z$0

the covariance matrix satisfies

$Z$1

where $Z$2 (Hahn et al., 12 Jun 2026). The eigenspace associated with the smallest eigenvalue, subject to orthogonality to $Z$3, spans the desired complement space. Operationally, with

$Z$4

the estimated GIV basis $Z$5 is formed from the eigenvectors of $Z$6 corresponding to the smallest eigenvalues (Hahn et al., 12 Jun 2026).

This feasible construction does not require a large cross-sectional dimension. The paper explicitly states that it is feasible even when $Z$7 is fixed, and no large-$Z$8 asymptotics are required (Hahn et al., 12 Jun 2026). By contrast, the high-dimensional FGIV literature treats both factors and loadings as unknown, estimates them by PCA or iterative OLS-PCA, and constructs

$Z$9

with $Y$0 obtained from estimated loadings (Banafti et al., 2022). In supply applications with cross-sectional dependence in idiosyncratic errors, that literature also uses the precision-weighted aggregation vector

$Y$1

and shows that sampling error in the estimated precision matrix is negligible under its sparsity conditions (Banafti et al., 2022).

The geometric characterization clarifies why GIV is not merely an ad hoc residualization device. The construction is anchored in the orthogonal complement of the loading space, and feasible procedures either estimate that complement directly from the covariance matrix or approximate it through factor estimation.

4. Instrument strength, dominance, and asymptotic regimes

The asymptotic strength of a granular instrument is governed by cross-sectional concentration. Under the power-law model for unit sizes

$Y$2

with shares

$Y$3

the key concentration object is the Herfindahl-like index

$Y$4

(Ramachandran, 2 Jul 2026). In the large-$Y$5, large-$Y$6 theory, the same logic appears through the asymptotic Herfindahl index $Y$7, whose behavior depends on the tail index $Y$8 (Banafti et al., 2022).

Under the power-law size assumption, GIV strength admits three regimes (Ramachandran, 2 Jul 2026):

Regime	Condition	Main implication
Strong GIV	$Y$9	$Z$0; estimator is consistent and asymptotically normal at the standard $Z$1 rate
Nearly weak GIV	$Z$2 with $Z$3	$Z$4 shrinks with $Z$5; estimator remains consistent and asymptotically normal at rate $Z$6
Weak GIV	$Z$7 with $Z$8	$Z$9 decays like $Y$0; estimator is inconsistent and has a non-standard distribution

More specifically,

$Y$1

and in the nearly weak regime the estimation error satisfies

$Y$2

(Ramachandran, 2 Jul 2026). In the earlier large-$Y$3, large-$Y$4 theory, the granular regime corresponds to $Y$5, where $Y$6, whereas thinner tails imply weakening relevance (Banafti et al., 2022).

Inference follows the same trichotomy. Wald inference is valid in the strong regime and remains valid in the nearly weak regime provided standard errors are scaled correctly (Ramachandran, 2 Jul 2026). In the weak regime, the concentration parameter can grow arbitrarily slowly, normal approximations become unreliable, and Anderson–Rubin confidence sets are recommended: $Y$7 (Ramachandran, 2 Jul 2026).

For feasible GIV, first-stage estimation of the factor structure changes the variance. The feasible instrument is

$Y$8

and valid inference must use standard errors that account for the additional first-stage term (Ramachandran, 2 Jul 2026). The paper proposes a HAC variance estimator and requires $Y$9 for feasible asymptotics in the strong and nearly weak regimes (Ramachandran, 2 Jul 2026). This suggests that the econometrics of GIV is inseparable from the cross-sectional size distribution: granularity is both the source of relevance and the determinant of the appropriate inferential regime.

5. Robust and adjacent extensions

One extension replaces homogeneous-spillover GIV with robust granular instrumental variables (RGIV). In the model

$X$0

RGIV allows $X$1 to vary freely across units, allows unknown heteroskedasticity, and does not require skewness in unit sizes (Qian, 2023). Instead of relying on a single constructed granular instrument, it uses the pairwise orthogonality of idiosyncratic shocks in a continuously updating GMM system based on moments of the form

$X$2

where $X$3 (Qian, 2023). The paper proves consistency and asymptotic normality, proposes a Sargan–Hansen $X$4-test for the uncorrelated-shocks restriction, and a homogeneity test for $X$5 (Qian, 2023). It also shows a limitation: with unrestricted spillover heterogeneity and an unobserved factor structure with unknown loadings, global identification can fail (Qian, 2023).

A different extension addresses invalid-IV uncertainty with many candidate instruments. In semiparametric G-estimation, the analyst specifies that at least $X$6 out of $X$7 candidate instruments are valid, and identification is based on higher-order centered interactions of order $X$8 or more (Sun et al., 2021). The identifying moments are

$X$9

and the resulting estimator is consistent, asymptotically normal, and semiparametrically efficient in a union model indexed by the unknown valid subset size $X,Y,Z$0 (Sun et al., 2021). The paper explicitly describes this approach as closely aligned with the spirit of GIV, because the “granular” object is a collection of high-order interaction functions rather than a hand-selected valid subset (Sun et al., 2021).

In data fusion, the acronym GIV is used in a different sense: Group Instrumental Variable. There, the latent source label $X,Y,Z$1 governing distinct treatment-assignment mechanisms is reconstructed from observed $X,Y,Z$2 and used as an instrument (Wu et al., 2022). The Meta-EM framework learns a representation $X,Y,Z$3, clusters the joint space $X,Y,Z$4 with an EM-style Gaussian mixture, and samples latent group labels from posterior probabilities $X,Y,Z$5 (Wu et al., 2022). The central claim is that, under additive-noise IV assumptions and sufficiently rich representations, the learned group assignments converge to the true latent source labels, making the resulting GIV asymptotically valid for treatment-effect estimation (Wu et al., 2022).

Bayesian model averaging provides yet another adjacent development. The gIVBMA procedure treats instrument choice and covariate choice as model uncertainty problems, averages across different sets of instrumental variables and covariates in a structural equation model, and allows instruments and covariates to switch roles (Steiner et al., 18 Apr 2025). By explicitly accounting for model uncertainty, it provides robustness against invalid instruments and proves model-selection consistency under its stated conditions (Steiner et al., 18 Apr 2025).

Taken together, these extensions show that “granular” can refer to different sources of structure: cross-sectional market concentration, interaction bases over many candidate instruments, latent group labels in fused datasets, or Bayesian averaging over large candidate pools. The common theme is the construction of valid or plausibly valid instruments from structured heterogeneity rather than from a single externally supplied exclusion restriction.

6. Empirical applications and substantive findings

The recent empirical GIV literature spans macro-finance, commodity markets, and spillover systems. In the aggregate equity market multiplier application, the orthogonal-complement GIV framework is applied to a 12-sector panel and to a six-sector “granular core” (Hahn et al., 12 Jun 2026). Using 12 sectors, the paper reports OLS multipliers of about $X,Y,Z$6 in 1993Q1–2018Q4 and $X,Y,Z$7 in 1988Q4–2025Q4, FIV multipliers of about $X,Y,Z$8 and $X,Y,Z$9, and GIV multipliers of about $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$0 and $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$1; the overidentification test strongly rejects with $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$2 (Hahn et al., 12 Jun 2026). Restricting to the six largest sectors, which account for more than 97% of holdings, the GIV multiplier rises to $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$3 and $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$4, and the $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$5-test no longer rejects, with $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$6 and $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$7 (Hahn et al., 12 Jun 2026). The paper interprets the rejection of the 12-sector pooled model as evidence of substantial cross-sector heterogeneity in demand elasticities rather than merely bad instruments (Hahn et al., 12 Jun 2026).

In commodity demand estimation, feasible GIV with corrected standard errors is applied to refined copper, crude oil, and natural gas (Ramachandran, 2 Jul 2026). The reported short-run demand elasticities are $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$8 for copper, $$\max_x \sum_y \left[\max_z P(x,y\mid z)\right] \le 1,$$9 for crude oil, and $Z$00 for natural gas (Ramachandran, 2 Jul 2026). The estimated Pareto tail indices suggest that copper and natural gas are clearly in the strong regime $Z$01, while crude oil is borderline and can extend into the nearly weak region (Ramachandran, 2 Jul 2026). This application illustrates the paper’s main econometric point that regime classification is an empirical object tied to the observed size distribution.

RGIV is applied to Euro area sovereign yield spillovers using daily 10-year zero-coupon yields from Bloomberg over Sept. 1, 2009 to May 31, 2018, with countries aggregated into blocks to accommodate within-region correlation (Qian, 2023). In the preferred specification, the RGIV specification test is not rejected $Z$02, spillover homogeneity is strongly rejected $Z$03, the size-weighted spillover coefficient is about $Z$04 with s.e. $Z$05, the western periphery block is about $Z$06, the core block about $Z$07, and Slovenia about $Z$08 (Qian, 2023). The substantive conclusion is that spillovers are clearly heterogeneous across country groups, a pattern that baseline GIV cannot uncover (Qian, 2023).

Adjacent empirical work reinforces the broader relevance of granular IV ideas. The semiparametric invalid-IV G-estimation paper applies its interaction-instrument approach to UK Biobank data on the causal effect of body mass index on systolic and diastolic blood pressure, showing that estimates vary materially with the analyst’s choice of $Z$09 and that small $Z$10 can produce weak first-stage diagnostics (Sun et al., 2021). The data-fusion paper reports that Meta-EM substantially outperforms summary-IV baselines such as UAS, WAS, ModeIV, and AutoIV, and that EM-based reconstructed GIV often approaches the performance of oracle access to the true source labels on synthetic and semi-synthetic data (Wu et al., 2022). The Bayesian model averaging paper reports that gIVBMA outperforms competing methods in simulations with many weak or invalid instruments and yields applications to malaria and institutions, income per capita, and returns to schooling (Steiner et al., 18 Apr 2025).

The empirical record therefore points in two directions at once. First, GIV can deliver informative identification in environments where aggregate endogeneity is driven by latent common shocks and a few large entities preserve idiosyncratic variation. Second, empirical credibility depends on diagnostics for homogeneity, overidentification, instrument strength, and first-stage estimation error. A plausible implication is that the mature use of GIV is not simply instrument construction, but joint design, regime classification, and specification testing within structurally constrained panel environments.