Two-Stage Causal Estimation for Complex Observational Data

29 concepts25 papers1 proposal

Causal inference aims to answer "what if" questions: how would an outcome change if we intervened on a treatment? In ideal scenarios, randomized controlled trials provide these answers. In the real world, however, we often rely on observational data where "confounders"—hidden factors that influence both the treatment and the outcome—bias our observations.

A growing body of research addresses this by using two-stage estimation. Rather than attempting to map treatments directly to outcomes in a single step, these methods decouple the problem. They first isolate the "clean" variation in the treatment (or a mediator) and then use that variation to estimate the true causal effect.

The Core Mechanism: Decoupling and Integration

The fundamental challenge in causal estimation is that the observed relationship between a treatment $X$ and an outcome $Y$ is "polluted" by unobserved confounders $U$. Two-stage methods break this deadlock by introducing a third variable—such as an instrumental variable (IV), a proxy, or a mediator—that provides a window into the causal mechanism.

In the first stage, the model learns the relationship between the exogenous variable (the instrument or proxy) and the treatment. In the second stage, it maps this learned relationship to the outcome. Hartford et al. popularized this for deep learning with Deep IV, which uses a neural network to model the conditional distribution of treatments given instruments. The second stage then optimizes the outcome function by integrating over this estimated distribution, effectively "averaging out" the influence of unobserved confounders.

Building Intuition with Bridge Functions and Embeddings

Why does this two-step process work better than a direct approach? The intuition lies in the concept of a bridge function. In many confounded settings, the causal effect cannot be identified directly from the data distribution. Instead, we must solve an integral equation where the causal effect is a "bridge" connecting the treatment to the outcome through the instrument or proxy.

To solve this without making rigid linear assumptions, researchers often turn to Reproducing Kernel Hilbert Spaces (RKHS). Singh et al. introduced Kernel Instrumental Variable (KIV) regression, which maps instruments to conditional mean embeddings in the first stage. These embeddings act as high-dimensional features that capture the entire distribution of the treatment. Mastouri et al. and Xu and Gretton extended this to proxy causal learning, showing that even when we cannot observe confounders, we can use "proxies" (noisy measurements of the confounders) to solve for the bridge function via two-stage kernel ridge regression.

Stability and Efficiency through Pseudo-Outcomes

A major advantage of the two-stage framework is its modularity, which avoids the "instability nightmare" of adversarial or minimax optimization. Li et al. argue that by replacing unstable minimax oracles with a two-stage approach involving density estimation and regularized regression, models achieve more reliable convergence and simpler implementation.

This modularity also enables the use of pseudo-outcomes. In this paradigm, the first stage generates synthetic or "hallucinated" outcomes that represent the counterfactual state.

  • Frauen et al. use two-stage pseudo-outcome regression to estimate treatment effects over time, allowing the model to focus on the simple structure of the causal effect rather than the high-dimensional complexity of the entire system.
  • Coston et al. apply this to "runtime confounding," where a complex model uses all available features during training to generate pseudo-outcomes, which a simpler, deployment-ready model then learns to predict using only restricted features.

Generalizing to Complex Treatments and Networks

The two-stage logic is highly adaptable to non-standard data structures:

  1. Continuous Treatments: When treatments are continuous (like drug dosages), traditional methods often struggle with unstable density estimations. Kim and Wang show that a two-stage kernel ridge regression can estimate these effects by first learning a high-dimensional outcome surface and then regressing averaged predictions against treatment levels, bypassing the need for density ratios.
  2. Networked Data: In social or biological networks, confounders often leak through the network structure. Huang et al. propose a two-stage procedure that first disentangles latent instrumental variables from shared environmental confounders using graph neural networks, then freezes these representations to ensure the instrument has no direct, unmediated path to the outcome.
  3. Complex Mediators: When an intervention changes a complex object (like a drug shifting a whole transcriptome), Watson proposes modeling the mapping from treatment to high-dimensional mediator features first. This allows the model to update its understanding of new treatments independently of the stable relationship between the mediator and the outcome.

By decoupling the discovery of causal signals from the prediction of outcomes, two-stage estimation provides a flexible, stable, and theoretically grounded toolkit for navigating the "what ifs" of complex, real-world data.

Key Foundational Work

Deep IV: A Flexible Approach for Counterfactual Prediction (2017) — contributed 1 concept elements to this direction.
Causal Inference via Sparse Additive Models with Application to Online Advertising (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_two_stage_causal_estimation_99f6d6d2,
  title = {two-stage causal estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-stage-causal-estimation/rdr_99f6d6d2049b44f6978c1db0cd14ee0f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_two_stage_causal_estimation_99f6d6d2_v1,
  title = {two-stage causal estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-stage-causal-estimation/rdr_99f6d6d2049b44f6978c1db0cd14ee0f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake