Identifying Models via Heteroskedasticity

• Identification through heteroskedasticity is a strategy that exploits variations in variance—from regime shifts, stochastic volatility, or covariate-dependent scaling—to generate extra restrictions for structural parameter recovery.
• It is applied in diverse models such as heteroskedastic SVARs, fully nonparametric transformation models, and endogenous heteroskedastic IV frameworks to achieve point, partial, or testable identification.
• Constructive estimation methods include eigenvalue decompositions, solving derivative-based ODEs, and control-function regressions, while challenges arise when variance distinctions are weak or proportional.

Searching arXiv for recent and foundational papers on identification through heteroskedasticity, including nonparametric transformation models and heteroskedastic SVARs. Identification through heteroskedasticity denotes a class of identification strategies in which systematic variation in conditional variances is used to recover structural parameters that are not pinned down by mean restrictions alone. In the formulations represented in recent arXiv work, the relevant heteroskedasticity may arise from regime shifts in the covariance matrix of structural shocks, stochastic volatility, covariate-dependent scale in fully nonparametric transformation models, or treatment-dependent dispersion in instrumental-variables settings. Across these settings, the common mechanism is that variance variation generates additional algebraic, differential, or moment restrictions, so that otherwise underidentified objects become point-identified, partially identified, or at least testably restricted (Lütkepohl et al., 2018, Kloodt, 2020, Abrevaya et al., 2018).

1. General identification logic

The most compact statement of the idea is that heteroskedasticity creates observable contrasts across states, covariates, or treatments that a structural representation must reproduce. In covariance-based time-series models, the same contemporaneous matrix must diagonalize multiple reduced-form covariance matrices. In fully nonparametric transformation models, covariate-driven scale variation enters derivative ratios of the conditional distribution and yields a first-order ordinary differential equation for the unknown transformation. In endogenous-treatment models, treatment-induced variance heterogeneity allows a normalization of the outcome equation that restores valid conditional moment restrictions (Lütkepohl et al., 2024, Kloodt, 2020, Abrevaya et al., 2018).

This family resemblance does not imply a single canonical estimator. Rather, the identifying restriction takes different mathematical forms depending on the model class. In structural VARs it is a joint diagonalization problem. In transformation models it is an ODE uniqueness problem. In endogenous heteroskedasticity IV models it is a transformed conditional-mean problem. A plausible implication is that “identification through heteroskedasticity” is best understood as a general strategy for converting second-moment variation into structural information, rather than as one fixed econometric procedure.

2. Covariance shifts and structural identification in VAR-type models

In heteroskedastic structural VARs, the standard setup writes reduced-form innovations as $u_t = B e_t$, with regime-specific diagonal shock covariance $\Lambda_r$, so that

$$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$

With two regimes,

$$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$

Hence the eigenvalues of $M$ are the variance ratios $\delta_i$, and the eigenvectors span the columns of $B$. If the diagonal entries of $\Delta$ are pairwise distinct, the eigensystem is unique up to the usual sign and permutation indeterminacies, so the impact matrix is point-identified up to standard normalization (Bacchiocchi et al., 2024).

The same logic appears in Markov-switching heteroskedastic SVARs. In the $A_0$-parameterization, regime-specific reduced-form covariances satisfy

$\Sigma_m = A_0^{-1} D_m A_0^{-\top},$

with diagonal $\Lambda_r$0. Identification is governed by the vectors of relative variances

$\Lambda_r$1

If the $\Lambda_r$2-th row’s variance-ratio vector differs from all others, the $\Lambda_r$3-th row of $\Lambda_r$4 is unique; if all such vectors are distinct, $\Lambda_r$5 is globally identified under the normalization $\Lambda_r$6 (Lütkepohl et al., 2018).

These results formalize the intuition that heteroskedasticity can replace, complement, or overidentify conventional short-run restrictions. In the homoskedastic case, recursive or exclusion restrictions are often just-identifying. Once several covariance states must be rationalized by the same contemporaneous matrix, those same restrictions can become testable overidentifying restrictions (Lütkepohl et al., 2018).

3. Partial identification, multiplicity, and verification

Point identification through covariance shifts depends critically on distinct variance patterns. When some variance ratios coincide, the relevant eigenvalues are repeated and eigenvectors are no longer unique. In that case the structural matrix is only identified up to orthogonal rotations within multiplicity blocks. Formally, if the indices are partitioned into groups $\Lambda_r$7 with common variance ratios inside each group, then any

$\Lambda_r$8

generates the same pair of reduced-form covariance matrices. The identified object is then an invariant subspace rather than a unique column ordering (Bacchiocchi et al., 2024).

Recent work extends this distinction from global to shock-specific identification. In stochastic-volatility SVARs, let $\Lambda_r$9 denote the variance sequence of shock $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$0. The $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$1-th column of the structural matrix is unique up to sign if $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$2 for all $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$3. If only some sequences are distinct, the model is partially identified: only those shocks have uniquely determined impact vectors and impulse responses (Lütkepohl et al., 2024).

The verification problem is therefore central. One strand uses Savage–Dickey density ratios to test homoskedasticity or equality of variance-ratio patterns. In stochastic-volatility SVARs, the key null is $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$4, where $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$5 governs the scale of log-volatility; small posterior-to-prior density ratios at zero provide evidence against homoskedasticity (Lütkepohl et al., 2024). In Markov-switching heteroskedastic SVARs, analogous Savage–Dickey ratios test hypotheses such as $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$6, which directly target the distinctness conditions required for identification (Lütkepohl et al., 2018). A different route is sample-splitting: under the null of identification, independent subsample maximum-likelihood estimators converge to the same parameter, so a Wald-type statistic has standard $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$7 asymptotics; under nonidentification, the subsample estimators need not agree (Maciejowska, 2022).

A recent extension introduces Sparse Heterogeneous Markov-Switching Heteroskedasticity, in which each structural shock has its own Markov process and some regimes may have zero occupancy. In that framework, a shock is identified up to sign if its variance path differs from every other shock’s path, and the model is designed to verify homoskedasticity reliably through Bayesian model comparison (Shang et al., 17 Mar 2026).

4. Fully nonparametric transformation models

A distinct use of heteroskedasticity appears in the fully nonparametric transformation model

$$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$8

where $$\Sigma_r = \operatorname{Var}(u_t\mid r) = B \Lambda_r B'.$$9 is strictly increasing, $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$0 and $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$1 are unknown, and $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$2 with $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$3 and $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$4. The identification problem is to recover $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$5, $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$6, and $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$7 from the joint law of $$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$8 without parametric restrictions on any of these functions (Kloodt, 2020).

The constructive step is to work with the ratio of conditional derivatives

$$M := \Sigma_2 \Sigma_1^{-1} = B \Delta B^{-1}, \qquad \Delta := \Lambda_2 \Lambda_1^{-1}.$$9

After weighting over $M$0 by a nonnegative function $M$1, one obtains an observable scalar function

$M$2

where

$M$3

The condition $M$4 is the formal heteroskedasticity condition. It requires that the scale function varies sufficiently with the chosen covariate direction. Under that condition, identification of $M$5 becomes the uniqueness of the solution to

$M$6

Location and scale normalizations, such as $M$7 and $M$8, select a unique representative from the family of affine rescalings (Kloodt, 2020).

Once $M$9 is known, the remaining functions are recovered by

$\delta_i$0

The paper explicitly contrasts this heteroskedastic route with homoskedastic identification. If $\delta_i$1 is constant, then $\delta_i$2, the derivative ratio no longer contains $\delta_i$3 itself, and the construction degenerates. In the paper’s formulation, heteroskedasticity serves as a “shape” identifier rather than as a source of exogenous variation (Kloodt, 2020).

5. Endogenous heteroskedasticity in treatment and linear IV models

In treatment-effect models, heteroskedasticity can be structural rather than incidental. The endogenous heteroskedasticity IV model writes

$\delta_i$4

with binary treatment $\delta_i$5, binary instrument $\delta_i$6, and normalized disturbance satisfying $\delta_i$7 and $\delta_i$8. Because $\delta_i$9 depends on the endogenous treatment, individual treatment effects remain heterogeneous even among observationally identical units:

$B$0

In this setting, standard IV can be inconsistent because it conflates mean and variance effects (Abrevaya et al., 2018).

Identification proceeds through instrument-induced contrasts in conditional first and second moments. For each treatment state $B$1, the model defines $B$2 and $B$3, with the key lemma

$B$4

This yields the transformation

$B$5

so that dividing the outcome equation by $B$6 produces a transformed model with conditionally exogenous disturbance. Closed-form expressions then identify $B$7, $B$8, $B$9, and $\Delta$0, and the framework also delivers constructive formulas for counterfactual outcomes, the ATT, and the distribution of individual treatment effects (Abrevaya et al., 2018).

A related problem arises in linear models with continuous endogenous regressors:

$\Delta$1

Here the presence of endogenous heteroskedasticity means that even valid instruments generally do not imply $\Delta$2, so two-stage least squares is inconsistent. The proposed remedy is a control-function approach under the condition $\Delta$3, combined with polynomial structure on $\Delta$4 and $\Delta$5. After estimating the first-stage innovation $\Delta$6, one augments the outcome regression with the implied polynomial interactions. The causal coefficient $\Delta$7 is then identified by partialling out on the enlarged regressor set $\Delta$8 (Alejo et al., 2024).

These cross-sectional results show that identification through heteroskedasticity does not require regime switches in time series. It can also arise because treatment or endogenous regressors affect the scale of the disturbance, provided the model supplies the transformations or control functions needed to separate mean and variance channels.

6. Estimation, applications, and limitations

The constructive nature of these identification results often yields equally constructive estimators. In the nonparametric transformation model, estimation follows the identification proof: estimate the conditional CDF $\Delta$9 and its derivatives nonparametrically, form $A_0$0, aggregate to $A_0$1, locate the root $A_0$2, and recover $A_0$3 from the implied ODE under location-scale normalizations. The paper emphasizes that derivative estimation requires regularization, numerical integration, and careful bandwidth selection, and notes that one practical route to estimating the scalar $A_0$4 is to enforce derivative matching at $A_0$5 in the global solution (Kloodt, 2020).

In heteroskedastic SVARs, estimation is typically Bayesian and state-space based. The cited papers use combinations of forward-filtering backward-sampling for latent states, Gibbs steps for transition probabilities and volatility blocks, and Metropolis–Hastings or generalized-normal updates for structural matrices. Markov-switching heteroskedastic SVARs exploit analytical inverse-gamma structures and Savage–Dickey ratios for fast identification testing (Lütkepohl et al., 2018). Stochastic-volatility variants use non-centered parameterizations, auxiliary mixture methods, and ASIS to improve mixing when heteroskedasticity may be weak (Lütkepohl et al., 2024). Sparse heterogeneous Markov-switching models introduce an inverse-gamma-based Dirichlet distribution to sample normalized regime variances by Gibbs rather than Metropolis–Hastings, and are presented as computationally efficient while permitting identification verification (Shang et al., 17 Mar 2026).

Applications reflect the breadth of the framework. A multicountry panel VAR combines external instruments, heteroskedastic latent factors, and sign or zero restrictions to identify monetary policy and central bank information shocks in the euro area and the United States; in that design, stochastic volatility of the latent factors provides the heteroskedastic variation needed to point-identify the contemporaneous impact matrix (Pfarrhofer et al., 2019). A time-varying identification model allows both the structural matrix and the volatility parameter to switch across regimes, verifies within-regime heteroskedasticity for the monetary policy shock, and then selects among alternative exclusion restrictions by a multi-component spike-and-slab prior (Camehl et al., 2023).

The limitations are equally systematic. Identification fails or weakens when variance changes are proportional, when eigenvalues are repeated, when the contemporaneous matrix changes across regimes, when structural shocks are not conditionally diagonal, or when heteroskedasticity is too weak to distinguish shocks. In transformation models, the route collapses in the homoskedastic case $A_0$6, and it also depends on strict positivity of $A_0$7, support conditions, and monotonicity of $A_0$8 (Kloodt, 2020). In IV models with endogenous heteroskedasticity, weak first stages or failures of the control-function restrictions undermine the transformed moments or augmented-regression identification (Abrevaya et al., 2018, Alejo et al., 2024).

Taken together, these results establish identification through heteroskedasticity as a technically heterogeneous but conceptually unified research program. Whether implemented by generalized eigenvalue decompositions, derivative-ratio ODEs, transformed outcome equations, or control-function regressions, the central insight is the same: variance variation can be an independent source of structural information, and in many models it determines not only whether parameters are estimable, but also whether conventional identifying restrictions are genuinely testable.