--- title: Local Distributional Treatment Effect url: https://www.emergentmind.com/topics/local-distributional-treatment-effect type: topic --- # Local Distributional Treatment Effect Local Distributional Treatment Effect (LDTE) denotes a family of causal parameters that compare distributions rather than expectations on a localized margin. The locality can refer to the complier subpopulation under imperfect compliance, a fixed covariate value \(x\), a small interval around an outcome threshold \(y_0\), a kernel neighborhood in outcome space, or the limit at a regression-discontinuity cutoff. The term is therefore not tied to a single universal estimand. In instrumental-variables formulations, it is the difference in the treated and untreated potential-outcome cumulative distribution functions (CDFs) among compliers; in conditional distributional regression, it is often a covariate-specific difference of conditional CDFs; and in related work on the distribution of treatment effects, locality is attached instead to the conditional law of the individual treatment effect itself [2506.12765] [1806.09386] [2407.16037] [2407.14635]. ## 1. Terminological scope and principal estimands The main usages of LDTE in the recent literature differ along two axes: the object being localized and the design used for identification. Some papers localize in the population, as with compliers under imperfect compliance. Others localize in covariate space, outcome space, or design space, as in regression discontinuity. A related but distinct strand studies the distribution of the treatment effect \(\Delta_i=Y_i(1)-Y_i(0)\) rather than the distribution of outcomes under treatment and control [2509.15594] [2411.08778] [2504.03992] [2407.14635]. | Strand | Estimand | Locality notion | |---|---|---| | IV with imperfect compliance | \(\beta(y)=F_{Y(1)}(y\mid C)-F_{Y(0)}(y\mid C)\) [2509.15594] | complier subpopulation | | D-IV-LATE | \(\Delta(y)=F_{Y(1)\mid C}(y)-F_{Y(0)\mid C}(y)\) [2506.12765] | complier subpopulation | | GAMLSS conditional LDTE | \(\Delta(y\mid x)=F_Y(y\mid T=1,X=x)-F_Y(y\mid T=0,X=x)\) [1806.09386] | fixed \(x\) | | Randomized local effect | \(\tau_{\mathrm{loc}}(y_0;h)\) from CDF differences over \([y_0,y_0+h]\) [2407.16037] | local interval in outcome space | | Causal-DRF / CKTE | \(\tau_h(x;y)=E[K_h(Y-y)\mid D=1,X=x]-E[K_h(Y-y)\mid D=0,X=x]\) [2411.08778] | fixed \(x\), kernel neighborhood around \(y\) | | Distribution of treatment effects | \(F_{\Delta\mid X}(\tau\mid x)=P[\Delta_i\le \tau\mid X_i=x]\) [2407.14635] | fixed \(x\) on treatment-effect distribution | A recurrent source of confusion is the presumption that “local” has a stable meaning across these formulations. The literature does not support that presumption. This suggests that any statement about an LDTE must specify the locality mechanism, the identifying assumptions, and whether the target is a difference of outcome distributions or the distribution of the treatment effect itself. ## 2. Complier-specific LDTE under imperfect compliance In the instrumental-variables literature with binary instrument \(Z\) and binary treatment uptake, the LDTE is defined on the complier subpopulation. Let \(Y_i(w)\) be the potential outcome under treatment \(W=w\in\{0,1\}\), let \(W_i(z)\) be the potential treatment uptake under instrument \(Z_i=z\in\{0,1\}\), and impose monotonicity \(W_i(1)\ge W_i(0)\). The complier population is \[ C=\{i:W_i(1)>W_i(0)\}. \] The Distributional IV-LATE at threshold \(y\) is then \[ \Delta(y)=\mathbb{P}(Y_i(1)\le y\mid C)-\mathbb{P}(Y_i(0)\le y\mid C) =F_{Y(1)\mid C}(y)-F_{Y(0)\mid C}(y), \] with the same object denoted \(\beta(y)\) in the regression-adjusted imperfect-compliance framework [2506.12765] [2509.15594]. Identification proceeds through a Wald-like ratio of intent-to-treat differences. In the D-IV-LATE exposition, \(\Delta(y)\) is written as \(\alpha(y)/\beta\), where \[ \alpha(y)=\mathbb{E}_X\!\left[\mathbb{E}[\mathbf{1}\{Y\le y\}\mid Z=1,X]-\mathbb{E}[\mathbf{1}\{Y\le y\}\mid Z=0,X]\right], \] and \[ \beta=\mathbb{E}_X\!\left[\mathbb{E}[W\mid Z=1,X]-\mathbb{E}[W\mid Z=0,X]\right]. \] In the covariate-adaptive randomization formulation, the same logic is written with strata \(S\) as \[ \beta(y)= \frac{ \sum_s p(s)\{E[1\{Y\le y\}\mid Z=1,S=s]-E[1\{Y\le y\}\mid Z=0,S=s]\} }{ \sum_s p(s)\{E[D\mid Z=1,S=s]-E[D\mid Z=0,S=s]\} }. \] The familiar simplified expression is \[ \tau_{LDTE}(y)=\frac{F_{Y\mid Z=1}(y)-F_{Y\mid Z=0}(y)}{\Pr(D(1)>D(0))}. \] The underlying IV assumptions are random assignment conditional on strata, exclusion, and monotonicity [2509.15594]. Estimation is organized through Neyman-orthogonal moments and cross-fitting. In D-IV-LATE, the nuisance functions are \[ \mu(y,w,x)=\mathbb{E}[\mathbf{1}\{Y\le y\}\mid W=w,X=x],\quad p(z,x)=\mathbb{P}(W=1\mid Z=z,X=x),\quad \pi(x)=\mathbb{P}(Z=1\mid X=x), \] with orthogonal score functions \(\psi_{\alpha,i}(y;\mu,p,\pi)\) and \(\psi_{\beta,i}(p,\pi)\). Cross-fitting trains nuisance estimators on \(K-1\) folds and evaluates scores on the held-out fold, and the final estimator is the ratio of average score terms. The regression-adjusted imperfect-compliance estimator uses augmented residuals \(\Xi^Y_{z,i}(y)\) and \(\Xi^D_{z,i}\), again with \(L\)-fold cross-fitting and flexible first-stage estimators such as logistic, random forest, and boosting [2506.12765] [2509.15594]. The asymptotic theory is correspondingly strong. Under regularity and \(o(n^{-1/4})\) nuisance rates, the D-IV-LATE estimator is asymptotically normal for each fixed \(y\), with variance \[ V(y)=\frac{1}{\beta^2}\, \mathbb{E}\!\left[ \bigl(\psi_{\alpha,i}(y)-\Delta(y)\psi_{\beta,i}\bigr)^2 \right]. \] The imperfect-compliance paper further derives weak convergence in \(\ell^\infty(\mathcal{Y})\), states that the estimator achieves the semiparametric efficiency bound, and emphasizes applicability to continuous, discrete, and mixed discrete-continuous outcomes under stratified block designs and simple random sampling [2506.12765] [2509.15594]. A distinctive contribution of the D-IV-LATE paper is its emphasis on nuisance-model choice. It contrasts a Random-Forest-based DML estimator with a KAN-based D-IV-LATE, where “KAN” is defined as a Kernel-Augmented Nuisance estimator that fits each nuisance function via a combined parametric+nonparametric (RKHS) loss with a kernel penalty \(\lambda\|f\|_{\mathcal H}^2\) chosen by CV. The paper argues that the selection of nuisance-function estimators is not a mere implementation detail but a pivotal choice that can profoundly impact research outcomes, a claim supported by a 401(k) application in SIPP 1991. There, the RF-DML estimate is negative for very low wealth thresholds, peaks near \(0.9\) at \(y\approx \$18.5k\) for moderate wealth, and remains positive but declines to around \(0.4\) above \(y>\$120{,}000\); the abstract reports that the KAN-based estimator suggests more complex treatment effect heterogeneity [2506.12765]. ## 3. Conditional LDTEs from distributional regression A second major usage of LDTE is covariate-specific. In the GAMLSS formulation, one observes \(\{(Y_i,X_i,T_i)\}_{i=1}^n\) and models \[ Y_i\mid X_i,T_i\sim D(\theta_i),\qquad \theta_i=(\mu_i,\sigma_i,\nu_i,\tau_i), \] with each parameter linked to an additive predictor of the form \[ g_k(\theta_{ik})=\beta_{k0}+\beta_{k1}T_i+s_k(X_i). \] The Local Distributional Treatment Effect at \(y\) and \(x\) is \[ \Delta(y\mid x)=F_Y(y\mid T=1,X=x)-F_Y(y\mid T=0,X=x) =F_D(y;\hat\theta(1;x))-F_D(y;\hat\theta(0;x)). \] The same framework defines a local density treatment effect \[ \delta(y\mid x)=f_Y(y\mid 1,x)-f_Y(y\mid 0,x), \] and a local quantile treatment effect \[ QTE(\tau\mid x)=Q_Y(\tau\mid 1,x)-Q_Y(\tau\mid 0,x). \] Because the response family can be nonnormal and the covariate effects nonlinear, the approach explicitly targets location, scale, skewness, kurtosis, and zero-mass type features rather than only the conditional mean [1806.09386]. The estimation workflow is fully specified. One chooses a distribution family \(D\), chooses link functions \(g_k\), specifies additive predictors using terms such as P-splines, spatial terms, or random effects, estimates by maximum-likelihood or Bayesian methods, predicts \(\hat\theta(1;x)\) and \(\hat\theta(0;x)\), and then evaluates \(\Delta(y\mid x)\), \(\delta(y\mid x)\), and \(QTE(\tau\mid x)\) either by built-in cdf/pdf/quantile functions or by numerical inversion and integration. Inference is based on a parametric bootstrap with \(B\approx 500\)–\(1{,}000\), and model validation uses normalized quantile residuals (Dunn–Smyth residuals), Q–Q plots, and moments near \(N(0,1)\) [1806.09386]. Randomized experiments generate a related conditional distributional-regression formulation. There, for each treatment arm \(w\in\{0,1\}\) and each threshold \(y\), the nuisance function is \[ \gamma_y^{(w)}(x)=E[1\{Y\le y\}\mid W=w,X=x], \] and the Neyman-orthogonal score is \[ \psi_y^{(w)}(Z;\theta_y,\gamma_y) = \frac{\mathbf{1}\{W=w\}}{\pi_w}\,[\mathbf{1}\{Y\le y\}-\gamma_y^{(w)}(X)] +\gamma_y^{(w)}(X)-\theta_y^{(w)}. \] Cross-fitted regression adjustment yields \(\hat F_w(y)\), and the local effect on a narrow interval is the “probability treatment effect” \[ \hat\tau_{\mathrm{loc}}(y_0;h) = [\hat F_1(y_0+h)-\hat F_1(y_0)]-[\hat F_0(y_0+h)-\hat F_0(y_0)]. \] Uniform inference is based on a multiplier bootstrap, and the paper reports substantial finite-sample variance reduction from machine learning adjustment, with Monte Carlo RMSE reductions and narrower confidence intervals in the Ferraro & Price water-nudge experiment [2407.16037]. ## 4. Kernel, forest, conformal, and finite-location localizations Some LDTE formulations localize directly in outcome space through kernels. In Causal-DRF, the Local Distributional Treatment Effect is also called the Conditional Kernel Treatment Effect (CKTE): \[ \tau_h(x;y)=E[K_h(Y-y)\mid D=1,X=x]-E[K_h(Y-y)\mid D=0,X=x]. \] Here \(K_h(u)=K(u/h)/h\), and \(\tau_h(x;y)\) is estimated by honest distributional random forests. Splits maximize an MMD-type statistic based on kernel means over a grid \(y_1<\cdots