Working Paper

  • Synthetic Counterfactual Networks
    Working paper, 2026; [web link]
    We propose the Synthetic Counterfactual Network method for imputing how a network would respond to a treatment intervention observed only in other networks. Researchers observe a target network and a fixed number of donor networks before the intervention, and only the donors are observed after the intervention, with the estimand being the post-intervention counterfactual link probability of each pair in the target network. However, the challenges arise from the fact that networks differ in size, lack node correspondence, and consist of two entangled sources of heterogeneity: the difference in distributions of individual latent characteristics and the network-specific formation environments. The proposed method resolves these issues in the following steps: with the large network setting, we first denoise each adjacency matrix by kernel neighborhood smoothing; next, align each donor to the target through the Gromov–Wasserstein optimal coupling, which softly matches individuals by comparing within-network differences in their connection profiles; then choose convex weights so that the aligned donors reproduce the target before the intervention; finally, carry the learned soft alignment and network weights forward to impute the target's post-intervention counterfactual network. In the population problem, we provide conditions for the identification of the counterfactual network. We then establish uniform consistency of the finite-sample estimator and the max-norm rate of convergence.
  • Simple Specification-Free Causal Inference in Large Panels
    arXiv working paper, 2026; [web link]
    Ben Deaner, Chen-Wei Hsiang, Andrei Zeleneev
    Previously circulated as Inferring Treatment Effects in Large Panels by Uncovering Latent Similarities (2025)
    We propose a simple and transparent approach to causal inference using panel data with a large number of units and periods. Instead of imposing a particular parametric specification on the untreated potential outcomes, we assume they follow a nonparametric nonseparable factor model. We use long pre-treatment outcome histories to construct a measure of similarity between the latent factors of different units. Using this measure, we impute missing counterfactual means and propensity scores through kernel smoothing. Under weak smoothness conditions, these estimates can attain the optimal nonparametric convergence rate up to logarithmic factors, provided that the pre-treatment history grows sufficiently quickly with the sample size. We construct a doubly robust estimator of the period-specific average treatment effect on the treated (ATT) and provide conditions under which it is root-N-consistent and asymptotically normal, centered at the true ATT. Simulations demonstrate accurate inference for a wide range of linear and nonlinear data-generating processes.

Work in Progress

  • Generalized TWFE Estimation of Average Treatment Effects
    Alberto Abadie, Susan Athey, Ben Deaner, Chen-Wei Hsiang, Guido Imbens, Devavrat Shah, Andrei Zeleneev
    The past decade has seen a surge of methods for estimating causal parameters in panel data settings, including work on synthetic controls, grouped heterogeneity, two-way fixed effects, difference-in-differences, and factor models, a research agenda that has deeply shaped empirical practice. This article examines the potential of nonparametric (nonlinear) factor models as an encompassing framework that subsumes earlier approaches and extends identification to settings where previous methods are invalid. We leverage the features of modern data-rich environments, where both dimensions of the panel, the cross-section dimension N and the time series dimension T, are large. We establish identification results for the nonparametric factor model. On top of that, we propose a new bias-adjusted matching estimator that leverages the similarity distance in terms of the unobserved latent factors.
  • Nonparametric Item Response Theory via Matrix Completion with Cross-Fitting
    Yu-Chang Chen, Hau-Hung Yang, Chen-Wei Hsiang
    This paper develops a fully nonparametric estimator of the item characteristic curve (ICC) in Item Response Theory (IRT) that does not require any parametric link function on the latent ability. Because the latent ability is unobserved and unidentified up to a monotone transformation, the estimator proceeds in two stages: a cross-fitted nonparametric matrix completion step that delivers a uniform sup-norm estimate of the latent ability quantiles, and a local linear regression of the item responses on the generated quantiles. The cross-fitting design makes the second-stage regression error orthogonal to the first-stage estimator by construction, which converts what would otherwise be a high-level identification condition into a property of the procedure. A stochastic expansion characterizes the gap between the feasible estimator and its oracle benchmark by a kernel-smoothed version of the first-stage error, weighted by the ICC slope. Under the matrix-completion first stage, the achievable uniform rate of the generated quantiles is shown to be of order [log(n)/n]1/3, which pins the asymptotic theory in the moderate first-stage regime rather than the fast regime that has dominated the existing literature on generated regressors. A simulation study confirms the predicted convergence behaviour, and an empirical application to the verbal-aggression item bank documents the robustness of the nonparametric ICC to link misspecification under mixture-Gaussian and other non-standard data-generating processes.
  • Event Study with Time-Adjusted Synthetic Control
    This research proposes a time-adjusted synthetic control method for the event study. The proposed method utilizes a two-step approach in constructing time weights based on the factor model with interactive fixed effects and unit weights modified from the conventional synthetic control method. In the simulation study, under the data generating process with heteroscedasticity across time, the proposed method has the advantage of efficiency with large panel data.

Publication

  • Causal Random Forests Model Using Instrumental Variable Quantile Regression
    Econometrics, 2019, 7(4), 1-22; [web link]
    Jau-er Chen, Chen-Wei Hsiang
    We propose an econometric procedure based mainly on the generalized random forests method. Not only does this process estimate the quantile treatment effect nonparametrically, but our procedure yields a measure of variable importance in terms of heterogeneity among control variables. We also apply the proposed procedure to reinvestigate the distributional effect of 401(k) participation on net financial assets, and the quantile earnings effect of participating in a job training program.

Code

  • PanelIFE [GitHub link]
    R package for "Robust Estimation and Inference in Panels with Interactive Fixed Effects (Armstrong, Weidner, and Zeleneev, 2026) [web link]"
    PanelIFE implements the estimation and inference procedure for panel data with interactive fixed effects. This package provides two different method for estimation: one is the commonly used linear panel data model estimation procedure, and another one is the bias-aware estimation procedure that allows weak factors.
  • PanelLatSim [GitHub link]
    R package for "Simple Specification-Free Causal Inference in Large Panels (Deaner, Hsiang, and Zeleneev, 2026) [web link]"
    PanelLatSim implements the doubly-robust estimation and inference procedure of Deaner, Hsiang, and Zeleneev (2026) for the period-specific average treatment effect on the treated (ATT) in large panels with unobserved confounders. The untreated potential outcomes and the treatments are allowed to follow a nonparametric, nonlinear, and non-separable factor model, which is substantially more general than the additive structure imposed by two-way fixed effects and the interactive structure imposed by synthetic control and matrix completion methods. The missing counterfactual outcomes and propensity scores are imputed by kernel smoothing over a pseudo-distance that uncovers latent similarities between units from their long pre-treatment histories, and the imputed nuisances are combined into a doubly-robust estimator with cross-fitting.


(Last updated: Nov 2026)