Introduction
Causal inference and discovery from observational data have been extensively studied, but most prior work focuses on independent and identically distributed (i.i.d.) data. This paper extends the gold-standard Rubin causal framework to the complex world of multivariate point processes—continuous streams of recurring events like stock trades, server logs, or medical alerts. By redefining "treatment" and "outcome" for time-varying data and introducing a stabilized inverse propensity weighting method, the authors provide a rigorous way to measure how one type of event truly influences the rate of another over time.
The Challenge of Multivariate Event Streams
Most causal inference tools are built for static snapshots where an individual receives a treatment once and shows a single outcome (e.g., a patient takes a pill, and we measure if they recover).
However, the real world often looks like a "multivariate event stream." Consider a diabetic patient: they eat meals, take insulin, and exercise multiple times a day, while their blood glucose fluctuates constantly. These events are not independent; they are asynchronous, irregularly spaced, and recur over a shared timeline. Traditional causal methods struggle here because the "treatment" (e.g., having eaten recently) and the "outcome" (the rate of glucose spikes) are constantly changing. The gap this paper fills is providing a mathematically sound framework to establish pairwise causal relationships—saying "Event A causes Event B"—within these continuous, messy streams of data.
Methodological Innovations
The authors bridge the gap between Causal Inference and Multivariate Point Processes (MPPs) through several key conceptual shifts:
- The Proximal Assumption: To make the mathematics tractable, they define "treatment" ($Z_t$) based on a look-back window. If a specific causal event $z$ happened at least once in the last $w$ minutes, the system is considered "treated" at time $t$.
- Redefining the Average Treatment Effect (ATE): In this framework, the ATE is defined as the difference in the intensity rate (how often an outcome event is expected to happen per unit of time) between treated and untreated states, averaged over the entire observation period.
- Propensity Scores for Time: In standard statistics, a propensity score is the probability of being treated. In continuous event streams, the authors define it based on duration: the fraction of time a specific set of historical covariates is observed while the treatment is active versus inactive.
- Inverse Probability Weighting (IPW) for Events: The authors propose an IPW scheme adapted for time. If a certain history rarely leads to treatment, those moments are "up-weighted" to create a balanced comparison, similar to balancing a clinical trial. They specifically highlight a Stabilized IPW (IP-Stable) to prevent extreme weights from skewing the estimate.
Experimental Validation
The researchers evaluated their framework on both simulated and real-world data to test its accuracy and robustness.
- Synthetic Datasets: They generated artificial event streams using PGEM and Hawkes Processes where the "ground truth" causal relationships were known. This allowed them to measure exactly how much error their method had compared to the true causal effect.
- Real-World Diabetes Dataset: A dataset tracking 70 patients' meals, exercise, insulin doses, and blood glucose changes. They compared their model's findings against 11 causal pairs identified by medical experts (e.g., "Exercise causes blood glucose to drop").
Key Results
Superior Accuracy and Stability
On synthetic datasets, the proposed IP-Stable method achieved significantly lower Root Mean Squared Error (RMSE) than standard baselines. In some cases, it reduced error by 6 to 10 times compared to simple conditional intensity scores.
Furthermore, unlike simpler models that degrade rapidly if an incorrect look-back window ($w$) is chosen, the IPW-based methods remained relatively stable. This robustness is crucial for real-world applications where researchers rarely know the exact timing or duration of a causal effect.
Real-World Efficacy
On the Diabetes dataset, the method successfully recovered more expert-validated causal pairs than the baselines. Using the "Hits@K" metric (measuring how many true causes appear in the top $K$ results), IP-Stable consistently outperformed other models, proving its ability to find genuine causal signals in noisy, human-generated data.
Limitations and Future Work
While a significant step forward, the authors acknowledge a few hurdles:
- The Window Choice: The results still depend on the choice of the proximal window ($w$). While they suggest data-driven ways to pick it, an inappropriate window can still lead to biased results.
- Unobserved Confounders: Like most observational causal methods, this framework assumes "strong ignorability"—meaning there isn't some hidden, unmeasured event type that is actually causing both the treatment and the outcome.
- Binary Focus: Currently, the model treats events as binary occurrences (it happened or it didn't). Future work could extend this to "marked" processes, where the magnitude of the event (e.g., the specific dose of a drug or the calorie count of a meal) is also factored into the causal effect.
Ultimately, by shifting the focus from the probability of discrete events to the duration of states, this paper provides a powerful new lens for decoding cause and effect in continuous time.
Summary updated