Propensity score methodology in the presence of network entanglement between treatments

Panos Toulis, Alexander Volfovsky, Edoardo M. AiroldiarXiv 2018 · DOI · arXiv

Propensity Score Methodology in the Presence of Network Entanglement

Abstract

In causal inference, treatment assignment is typically assumed to be independent across units. However, many real-world scenarios involve "entangled treatments," where the treatment of one unit is inherently linked to others, such as the formation of edges in a social or biological network. This paper formalizes the concept of entangled treatments and demonstrates that classical propensity score methods, which ignore these dependencies, suffer from significant selection bias. We propose a novel methodology that models the evolution of the underlying network to estimate propensity scores more accurately. Our approach significantly reduces bias in simulated experiments and introduces new tools for quantifying model misspecification in network settings.

Introduction

Standard causal inference frameworks often rely on the Stable Unit Treatment Value Assumption (SUTVA), which posits that one unit's treatment does not affect another's outcome and that treatment assignments are independent. While much research has focused on "interference" (spillover effects in outcomes), the issue of "entanglement" in treatment assignment remains under-explored. Entangled treatments occur when the assignment for an individual unit depends on a common population structure, most notably a network. For instance, if a treatment is defined as "forming a new professional connection," the assignment is dyadic; one person cannot receive the treatment without another also receiving it.

The Challenge of Entangled Treatments

When treatments are entangled, the probability of a unit being treated is not a simple function of its own covariates. Instead, it is a function of the potential for interaction within the broader network. Classical propensity score models, such as logistic or Poisson regressions, treat these assignments as independent Bernoulli trials. This misspecification leads to "apples-to-oranges" comparisons, where units with high network-driven propensity are incorrectly matched with units that have low propensity, introducing substantial bias into the estimated Average Treatment Effect (ATE).

Methods: Modeling Network Evolution

To address this, the authors propose shifting the focus from individual treatment probabilities to the evolution of the network itself. Instead of modeling the marginal probability of treatment, we model the transition from a pre-treatment network ($G^-$) to a post-treatment network ($G^+$).

Figure 1: Illustration of network evolution from a pre-treatment state ($G^-$) to a post-treatment state ($G^+$), where individual treatments are defined by the changes in connectivity.

As shown in the figure above, the treatment is defined by the change in edges between time points. The proposed methodology (Algorithm 1) involves:

  1. Modeling the transition $G^- \to G^+$ using a network evolution model (e.g., Latent Space or Exponential Random Graph Models).
  2. Sampling multiple probable future networks from this model to capture the underlying uncertainty and dependency structure.
  3. Calculating an individual's propensity score by averaging their treatment status across these simulated trajectories.

Theoretical Framework

The paper introduces a novel similarity function to quantify the degree of model misspecification. This function measures how frequently two different propensity score models would assign individuals to the same subclassification "bins." Through a gradient-based approximation, the authors show that standard linear models are often geometrically "orthogonal" to the true network-based propensity scores. This geometric mismatch explains why classical models fail to capture the propensity of "sociable" or central units to form connections, leading to systematic errors in causal estimation.

Results and Discussion

Simulations involving both small-scale (5 units) and large-scale (100 units) networks demonstrate the efficacy of the network-aware approach. In one large-scale scenario, the Root Mean Squared Error (RMSE) for the classical misspecified model was 57.05, whereas the proposed network-aware model achieved an RMSE of only 3.41—a nearly 17-fold reduction in error.

A key insight is the distinction between entanglement and interference. While interference concerns how a neighbor's treatment affects my outcome, entanglement concerns how a neighbor's status dictates my treatment assignment. Even in the absence of interference, entanglement alone is sufficient to invalidate standard propensity score methods. Furthermore, the research suggests that even a slightly misspecified network model is generally superior to a marginal model that ignores the network structure entirely.

Conclusion

As researchers increasingly apply causal inference to networked data, accounting for treatment entanglement becomes critical. By modeling the underlying network mechanisms rather than treating assignments as independent events, we can produce more robust and unbiased causal estimates. Future work will look to integrate these methods with existing frameworks for handling outcome interference, providing a unified approach to causal inference in complex systems.

Summary updated

Abstract

In experimental design and causal inference, it may happen that the treatment is not defined on individual experimental units, but rather on pairs or, more generally, on groups of units. For example, teachers may choose pairs of students who do not know each other to teach a new curriculum; regulators might allow or disallow merging of firms, and biologists may introduce or inhibit interactions between genes or proteins. In this paper, we formalize this experimental setting, and we refer to the individual treatments in such setting as entangled treatments. We then consider the special case where individual treatments depend on a common population quantity, and develop theory and methodology to deal with this case. In our target applications, the common population quantity is a network, and the individual treatments are defined as functions of the change in the network between two specific time points. Our focus is on estimating the causal effect of entangled treatments in observational studies where entangled treatments are endogenous and cannot be directly manipulated. When treatment cannot be manipulated, be it entangled or not, it is necessary to account for the treatment assignment mechanism to avoid selection bias, commonly through a propensity score methodology. In this paper, we quantify the extent to which classical propensity score methodology ignores treatment entanglement, and characterize the bias in the estimated causal effects. To characterize such bias we introduce a novel similarity function between propensity score models, and a practical approximation of it, which we use to quantify model misspecification of propensity scores due to entanglement. One solution to avoid the bias in the presence of entangled treatments is to model the change in the network, directly, and calculate an individual unit's propensity score by averaging treatment assignments over this change.
BibTeX
@misc{toulis2018propensity,
  author = {Panos Toulis and Alexander Volfovsky and Edoardo M. Airoldi},
  title = {{Propensity score methodology in the presence of network entanglement between treatments}},
  year = {2018},
  eprint = {1801.07310},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1801.07310},
}
Figure