Causal Inference Under Temporal Undersampling

This direction is merged into another direction: temporal resolution mismatch. Its last version stays readable and citable here.

Version 1 of 1101 papers162 conceptsfrozen 2026-10-04

In the ideal world of causal discovery, we observe a system at a frequency that matches its internal dynamics. If a neuron fires and triggers another, or if a stock price shift influences a commodity, we hope our sensors capture the precise "lead-lag" relationship that reveals who is the driver and who is the passenger.

In reality, we are often "blind" to the true causal frequency. Whether it is fMRI scans capturing brain activity every few seconds or economic indicators updated quarterly, our measurements are frequently much slower than the underlying processes. This discrepancy, known as temporal undersampling, creates a fundamental distortion: time-delayed causal links appear to be instantaneous correlations, and genuine interactions can vanish entirely.

The Illusion of Simultaneity

The most immediate consequence of slow sampling is the loss of temporal precedence. If variable $A$ causes $B$ with a lag of 10 milliseconds, but our sensor only records data every 100 milliseconds, $A$ and $B$ will appear to change at the exact same time.

Alex Tank and Ali Shojaie demonstrate in Identifiability and Estimation of Structural Vector Autoregressive Models for Subsampled and Mixed Frequency Time Series that this temporal aggregation obscures the true lead-lag dynamics. When events are compressed into the same sampling window, standard algorithms cannot distinguish between $A \to B$, $B \to A$, or a hidden common cause. This "blurring" often forces researchers to rely on distributional features—like the "skewness" of the signal—rather than time-ordering to orient the causal arrows, as explored by Rubén Sánchez-Romero and Clark Glymour in the context of fMRI feedback networks.

The Fragility of Granger Causality

One of the most popular tools for time-series analysis, Granger Causality (GC), is particularly vulnerable to these sampling artifacts. GC relies on the idea that if $A$ causes $B$, then past values of $A$ should help predict the future of $B$.

However, Douglas Zhou and David Cai show in Analysis of sampling artifacts on the Granger causality analysis that GC values are not absolute; they are functions of the sampling interval. As the sampling frequency increases toward infinity, the raw causality values paradoxically vanish. Conversely, at certain slow sampling rates, the metric can hit "null points" where a physical connection exists but the statistical evidence disappears. This suggests that a "lack of causality" in a dataset might simply be a mismatch between the researcher's clock and the system's clock.

Recovering the High-Frequency Truth

How can we reconstruct a fast causal structure from slow, "mesochronal" data? The research direction has shifted toward algorithms that treat the observed data as a downsampled projection of a hidden, high-frequency process.

  1. Rate-Agnostic Learning: Sergey Plis and David Danks introduced Mesochronal Structure Learning and RASL. These methods search for a fine-grained "system graph" that, when projected forward in time, matches the coarse "measurement graph" observed in the data. This allows for structure recovery even when the exact sampling rate is unknown.
  2. Exploiting Non-Gaussian Noise: A breakthrough in identifiability comes from the nature of noise. Mingming Gong and Bernhard Schölkopf argue in Discovering Temporal Causal Relations from Subsampled Data that if the underlying noise is non-Gaussian, the original high-frequency model leaves a unique statistical "fingerprint" in the subsampled data. This allows researchers to mathematically distinguish genuine causal influence from the artifacts generated by the sampling process itself.
  3. Continuous-Time Modeling: To move away from the constraints of fixed intervals, researchers like Alexis Bellot and Mihaela van der Schaar have proposed using Neural Ordinary Differential Equations (Neural ODEs). By modeling the system in continuous time, these frameworks become "causally consistent"—the recovered structure remains stable regardless of the frequency at which the data was recorded.

Why It Matters

Understanding temporal undersampling is not just a theoretical exercise; it is a prerequisite for valid science in fields with "slow" sensors. In psychology, Jack Kuipers and Giusi Moffa have shown that large gaps in longitudinal data can miss "micro-cascades" of symptoms in schizophrenia, leading to a misunderstanding of how the disorder progresses. By accounting for the sampling rate, we move from merely describing correlations to identifying the true levers of a system.

Papers in this version

Showing 60 of 101 papers.

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_under_temporal_undersam_f88f598e,
  title = {Causal Inference Under Temporal Undersampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-temporal-undersampling/rdr_f88f598e119f47a4a7b44f433fb6ea69},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_inference_under_temporal_undersam_f88f598e_v1,
  title = {Causal Inference Under Temporal Undersampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-under-temporal-undersampling/rdr_f88f598e119f47a4a7b44f433fb6ea69/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}