Instrumental Variables in Structural Factor Models

• Instrumental variables in SFM are methods that use exogenous instruments to disentangle causal effects and mitigate endogeneity and latent confounding.
• They rely on strict identification conditions—relevance, exclusion, and exogeneity—to produce unbiased causal estimates.
• Extensions to nonparametric, nonlinear, and multivariate frameworks enhance the robustness and applicability of these approaches in complex empirical analyses.

Instrumental variables (IV) approaches in the context of structural factor models (SFM) constitute a rigorous methodological family for causal inference under latent confounding and endogenous explanatory variables. IV techniques solve identification by leveraging exogenous variation in variables (instruments) that influence the endogenous regressor but do not directly affect the outcome, thus permitting consistent estimation of causal effects even in the presence of unmeasured confounders. Structural equation modeling provides the formal framework, while generalizations to non-linear, non-parametric, and multiple-instrument settings expand the scope and robustness of IV methods for both theoretical investigation and practical data analysis (Wong, 2021, Clarke et al., 2015).

1. Structural Equation Framework and Identification Assumptions

The structural equation model central to IV in SFM is defined on a probability space $(\Omega, \mathcal{A}, P)$ with observed blocks $X$ (endogenous regressor), $Y$ (outcome), $Z$ (instrument) and latent $U$ (confounders), with the canonical system: $X = g(Z, U), \quad Y = f(X, U)$ where $g$, $f$ are (potentially non-linear) mappings. Causal identification with IV hinges on three conditions:

• Relevance: $Z$ exerts a nonzero effect on $X$, i.e., $X$0 is non-constant or full rank.
• Exclusion Restriction: $X$1 affects $X$2 only through $X$3; formally, $X$4 admits no direct dependence on $X$5.
• Exogeneity: $X$6 is independent of the unobserved $X$7 (possibly conditionally).

Graphically, this structure blocks all backdoor paths from $X$8 to $X$9—the only open route is through $Y$0, enabling $Y$1 to serve as a valid source of exogenous variation for causal estimation (Wong, 2021).

2. Causal Effect Definitions and Parametric IV Estimator

The causal effect of $Y$2 on $Y$3, denoted $Y$4, admits the following definitions:

• Linear/constant effect: If $Y$5, with $Y$6, then $Y$7.
• General/nonparametric effect: Let $Y$8 and define the average causal effect at $Y$9 as $Z$0.

In the linear case, two-stage least squares (2SLS) yields the standard IV estimand: $Z$1 where the first stage regresses $Z$2 on $Z$3 to obtain fitted values, and the second stage regresses $Z$4 on these fitted values to estimate $Z$5 (Wong, 2021).

3. Nonparametric Identification and Calculus for IV

Beyond linearity, identification in fully general SEMs proceeds via differentiating conditional means. For scalar $Z$6, under appropriate smoothness and orthogonality conditions,

$Z$7

it follows that $Z$8, so that

$Z$9

If $U$0 depends only on $U$1, and provided the family of conditional distributions $U$2 is “complete,” one recovers $U$3. This calculus approach enables identification in nonparametric and nonlinear SFMs (Wong, 2021).

4. Identification Theorems: Discrete, Mixed, and Multivariate Extensions

Identification extends beyond continuous variables:

• Discrete X, Z: Construction of a contrast matrix $U$4 for $U$5 and a contrast vector $U$6 for $U$7. Under independence, rank, and orthogonality conditions, treatment effect parameters $U$8 solve $U$9.
• Discrete X, Continuous Z: Application of the total derivative formulation leads to a system of differential or integral equations for the effect parameters.
• Vector Cases: Generalizes to multivariate $X = g(Z, U), \quad Y = f(X, U)$0 via Jacobians, with identification through matrix inversion under suitable rank and independence conditions (Wong, 2021).

5. Structural Mean Models and Generalized Method of Moments

Structural mean models (SMMs) provide a semiparametric potential-outcomes framework encompassing additive, multiplicative, and logistic links. SMMs specify shifts in mean outcome under treatment/exposure, characterized by: $X = g(Z, U), \quad Y = f(X, U)$1 With multiple instruments, GMM estimation employs moment conditions of the form $X = g(Z, U), \quad Y = f(X, U)$2, where $X = g(Z, U), \quad Y = f(X, U)$3 is the stacked instrument vector. Weighted quadratic minimization yields semiparametrically efficient estimators for $X = g(Z, U), \quad Y = f(X, U)$4 when instruments are mutually orthogonal binaries. GMM over-identification is tested via the Hansen $X = g(Z, U), \quad Y = f(X, U)$5-test (Clarke et al., 2015).

6. Estimation, Inference, and Empirical Implementation

Estimation procedures differ by model class:

• Linear/Parametric IV (2SLS):

1. Regress $X = g(Z, U), \quad Y = f(X, U)$6 on $X = g(Z, U), \quad Y = f(X, U)$7, obtain fitted values $X = g(Z, U), \quad Y = f(X, U)$8.
2. Regress $X = g(Z, U), \quad Y = f(X, U)$9 on $g$0, extract coefficient $g$1.
3. Estimate asymptotic variance as $g$2.

• Nonparametric/Nonlinear IV:
  • Estimate $g$3 and $g$4 via smoothing.
  • Numerically differentiate to get $g$5.
  • Compute $g$6; invert to recover $g$7 as structure permits (Wong, 2021).
• GMM in SMMs:
  • Specify sample moments $g$8.
  • Employ one-step or two-step weighting; iteratively update $g$9.
  • Use stacked moments when nuisance parameters exist; compute robust standard errors.
  • Confirm instrument validity via Hansen’s $f$0-test (Clarke et al., 2015).

7. Interpretation, Local Effects, and Empirical Findings

If the no effect modification (NEM) assumption fails but monotonicity (no defiers) holds, SMM GMM and 2SLS estimators can be interpreted as weighted averages of local average treatment effects (LATEs) or local risk/odds ratios. Empirical applications using genetic instruments validate the method's efficacy: for example, in a study with two SNPs as instruments for adiposity, causal estimates for hypertension risk via additive, multiplicative, and logistic SMMs were all statistically significant, with Hansen $f$1-test $f$2, indicating validity of instruments and modeling assumptions (Clarke et al., 2015).

Across both linear and highly general nonlinear SFMs, IV techniques under the described frameworks, estimation strategies, and identifying assumptions furnish a unified theoretical and practical arsenal for causal inference in the presence of endogenous predictors and latent confounding, with extensions to multiple-instrument, high-dimensional, and nonparametric regimes (Wong, 2021, Clarke et al., 2015).