Model Risk in Machine-Learning Distributional IV Estimation

We study model risk in machine-learning estimation of the Distributional Instrumental Variable Local Average Treatment Effect (D-IV-LATE), the distributional IV effect for the subpopulation induced into treatment by the instrument. The contribution is not a new neural causal estimand. We implement a reduced-form orthogonal level-score DML estimator for the covariate-adjusted D-IV-LATE target and use it to ask how much the nuisance learner matters for distributional IV conclusions. In simulations with explicit monotone principal strata and known complier truth, Kolmogorov-Arnold Networks (KANs) are faster than Random Forests in every scenario examined, but Random Forests usually produce more accurate D-IV-LATE curves. A targeted KAN ablation selects a width-64 KAN as the best KAN variant among those tested, but this is a speed/accuracy tradeoff rather than evidence of KAN dominance. In a 401(k) application, RF and KAN estimates differ materially, with frequent sign reversals along the estimated curve. The KAN instrument-propensity estimates also concentrate near the boundaries, so the KAN empirical curve is best read as sensitivity evidence. In inference validation, KAN pointwise intervals undercover badly under both asymptotic and bootstrap constructions, while RF asymptotic intervals are better calibrated in the validation designs. The main lesson is cautionary and constructive: speed and architectural flexibility are not enough for causal inference. Applied researchers using ML-based distributional IV estimators would be advised to benchmark nuisance learners, report overlap and calibration diagnostics, and validate inference directly.