Long-term causal effects in multiagent economies

Panos Toulis, David C. Parkes

Abstract: The effect of a treatment in a multiagent economy, e.g., a price increase, is causal if the treated economy would be different, e.g., in terms of revenue, relative to the control economy. Causal effects measured in an equilibrium of the economy, the long-term causal effects, are more representative of the value of such treatments. However, the statistical estimation of long-term causal effects is difficult because it has to rely, for practical reasons, on experimental data where agents are randomly assigned to the treated or the control economy, and their actions are observed before an equilibrium is reached. We propose a methodology to define and estimate long-term causal effects, which relies on a model of agent behaviors that plays a two-fold role. First, it predicts how agents would behave under different assignments, and, second, it predicts how agents would behave in equilibrium. These two prediction tasks enable the estimation of long-term causal effects under suitable assumptions, which we state explicitly.

1. Introduction

In multiagent economies, such as online advertising auctions or financial markets, the impact of a policy change (treatment) is rarely immediate. Agents adapt their strategies over time, meaning the short-term response observed during an experiment may differ significantly from the long-term equilibrium. Estimating these long-term causal effects is essential for effective policy-making but is complicated by the dynamic and strategic nature of agent interactions.

2. Problem Formulation

The estimation of long-term causal effects faces two primary obstacles. First, the counterfactual problem: in a typical experiment, we observe a mixture of treated and control agents. We must infer what the outcome would be if the entire population were assigned to one group or the other. Second, the equilibrium problem: experimental data is often collected while the system is still in a transient state. We need to predict the "target of inference"—the state of the economy once it has settled into a new equilibrium.

Conceptual framework for estimating long-term causal effects from observed experimental data.

3. The LACE Methodology

The proposed LACE (Long-term Average Causal Effects) framework addresses these challenges by modeling the latent behaviors of agents rather than just their observed actions. This approach relies on two integrated models:

  • Behavioral Model: Based on behavioral game theory (e.g., Quantal $k$-level models), this component captures how agents of varying sophistication levels respond to incentives.
  • Temporal Model: A Vector Autoregressive (VAR(1)) model is used to characterize how the distribution of these behaviors evolves over time.
Temporal evolution of behavioral parameters under different assignment scenarios.

The LACE algorithm proceeds by estimating behavioral and temporal parameters from short-term data, predicting aggregate behavior for counterfactual assignments, and projecting these dynamics forward to the long-term equilibrium.

4. Experimental Results

The methodology was evaluated using a laboratory dataset from a $5 \times 5$ zero-sum game (Rapoport and Boebel, 1992). By using the initial rounds as experimental data and the final rounds as the long-term ground truth, the authors compared LACE against standard benchmarks.

LACE demonstrated superior predictive accuracy for long-term revenue effects:

  • LACE: Mean Squared Error (MSE) of 0.045
  • Naive Projection: MSE of 0.185
  • Difference-in-Differences (DID): MSE of 0.361

5. Discussion and Limitations

LACE succeeds by modeling the process of adaptation. By incorporating the strategic structure of the game (the payoff matrix), it makes more informed counterfactual predictions than generic statistical models. However, limitations remain: the current model assumes a closed economy with a fixed population and relies on assumptions like "behavioral ignorability" which can be difficult to verify. Future research will focus on scaling these methods to large-scale industrial systems and modeling open economies with agent entry and exit.

Summary updated

Abstract

The effect of a treatment in a multiagent economy, e.g., a price increase, is {\em causal} if the treated economy would be different, e.g., in terms of revenue, relative to the {\em control} economy. Causal effects measured in an equilibrium of the economy, the {\em long-term causal effects}, are more representative of the value of such treatments. However, the statistical estimation of long-term causal effects is difficult because it has to rely, for practical reasons, on experimental data where agents are randomly assigned to the treated or the control economy, and their actions are observed before an equilibrium is reached. We propose a methodology to define and estimate long-term causal effects, which relies on a model of agent behaviors that plays a two-fold role. First, it predicts how agents would behave under different assignments, and, second, it predicts how agents would behave in equilibrium. These two prediction tasks enable the estimation of long-term causal effects under suitable assumptions, which we state explicitly.
BibTeX
@inproceedings{toulis2016long,
  author = {Panos Toulis and David C. Parkes},
  title = {{Long-term causal effects in multiagent economies}},
  booktitle = {Digital},
  year = {2016},
}
Figure