Causal Inference from a Single Continuous Trajectory

This direction is merged into another direction: unobserved confounding (unmeasured confounding). Its last version stays readable and citable here.

Version 1 of 112 papers23 conceptsfrozen 2026-10-04

In classical statistics, causal inference usually requires a "population"—a large group of independent units where some receive a treatment and others do not. But in many of the most important systems we study, there is only one unit. There is only one Earth’s climate, one global economy, and, for a doctor treating a specific patient, only one individual whose unique biology is evolving over time.

When we cannot look across a population, we must look across time. A growing body of research is developing methods to extract causal relationships from a single continuous trajectory. By leveraging the temporal structure of a sequence, researchers are finding ways to treat a single history as a series of "mini-experiments," effectively turning time into the replicates we lack.

Time as a Sample Generator

The fundamental challenge of "N-of-1" causal inference is the lack of independent and identically distributed (i.i.d.) samples. If you only have one time series, how do you establish statistical significance?

The key intuition, explored by Mark J. van der Laan and Ivana Malenica in Robust Estimation of Data-Dependent Causal Effects based on Observing a Single Time-Series, is to shift from "marginal" effects to "context-specific" averages. Instead of asking what a treatment does to a general population, we ask what it does to this unit given its specific history. By averaging these conditional effects across many time points, the temporal sequence itself provides the necessary data density.

This works because of time-invariance. As Miruna Oprescu, David Keetae Park, and their colleagues demonstrate in their work on GST-UNet, if we assume the underlying "rules" of a system (like the physics of wildfire smoke or the biology of a patient) remain constant, we can treat different segments of a single long-term trajectory as distinct instances of the same causal mechanism. This allows models to learn complex spatiotemporal dynamics from a single "movie" of data rather than thousands of independent "snapshots."

Unmasking the Hidden Confounder

The hardest problem in causal inference is the "unobserved confounder"—a hidden variable that influences both the cause and the effect. In a single time series, this problem is amplified because we cannot use cross-sectional controls to "cancel out" these hidden factors.

However, the sequence of events contains a hidden signal. Stefan Feuerriegel and Tobias Hatt propose a Sequential Deconfounder that exploits this. Imagine a doctor adjusting a patient's medication over months. The doctor’s decisions aren't random; they are based on the patient's state—some of which is recorded (blood pressure) and some of which is "hidden" (the doctor’s intuition or the patient’s subjective fatigue). By analyzing the sequence of treatment assignments, we can mathematically "reconstruct" these hidden factors. The history of actions acts as a proxy for the unobserved variables that drove those actions.

Similarly, Maha Shadaydeh and Joachim Denzler use variational autoencoders in SCEVAE to infer latent confounders from the temporal structure of a sequence. By creating "synthetic twins" of the data, they can estimate what would have happened in a counterfactual past, even when critical variables were never measured.

From Correlation to Excitation

In complex networks like the stock market or the brain, we often want to know "who influences whom." Traditional methods struggle when they only have one continuous stream of events.

Jalal Etesami and Kun Zhang addressed this in Learning Network of Multivariate Hawkes Processes by looking at "internal excitation." In a Hawkes process, an event in one variable increases the probability of future events in others. They proved that these excitation parameters are mathematically equivalent to the causal graph of the system. If you observe a single sequence for long enough, the way the system "rings" after an event reveals its internal wiring, allowing for causal discovery without ever needing a second trial.

Bridging the Short and Long Term

Finally, this research direction addresses the "horizon" problem: how do we predict long-term outcomes when we only have short-term data? Lu Cheng and Huan Liu developed the LTEE framework, which uses sequential modeling to learn "surrogate representations." By treating the intermediate steps of a sequence as informative proxies, they can bridge the gap between an immediate treatment and its eventual effect. This is further refined by Nathan Kallus and Xiaojie Mao, who show in Long-term Causal Inference Under Persistent Confounding that the sequential progression of these intermediate outcomes can actually help unmask "persistent" confounders that bias the entire timeline.

By treating time not just as a coordinate, but as a source of structural information, these researchers are making it possible to perform rigorous causal science on the unique, non-repeatable systems that define our world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_inference_from_a_single_continuou_3b0e6df9,
  title = {Causal Inference from a Single Continuous Trajectory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-from-a-single-continuous-trajectory/rdr_3b0e6df95e554e8fa94f2cac4c970bce},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_inference_from_a_single_continuou_3b0e6df9_v1,
  title = {Causal Inference from a Single Continuous Trajectory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-from-a-single-continuous-trajectory/rdr_3b0e6df95e554e8fa94f2cac4c970bce/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}