An agent-based model of tax compliance with social networks.

BACKGROUND

Tax evasion as a social phenomenon has received increasing amounts of attention in the literature as researchers seek improved approaches to modeling noncompliance. In related research, others have sought approaches that more explicitly account for social phenomena in explaining crime. For example, Glaeser, Sacerdote, and Scheinkman (1996) model the wide variance of crime rates across different geographic areas, taking into account social interactions. Their results indicate that crime (e.g., larceny, burglary, and auto theft) has a significant degree of social interaction that helps explain its variability across different cities, towns and precincts. If in fact most crime does have an element of social interaction as its cause, then it would not be unreasonable to conclude that, since tax evasion is a crime, evasion has social drivers as well. In fact, researchers have tried to incorporate social drivers (e.g., social norms) in explaining noncompliance as conventional economic models over-predict the amount of noncompliant behavior one should observe (Alm, McClelland, and Schulze, 1992). To date, little research has explicitly modeled the role that social networks might play in noncompliant behavior. Though tax authorities tacitly recognize that the sharing of information regarding tax schemes, for example, can promulgate evasive behavior, research and information on understanding the impact of networks has been limited. Part of the explanation for limited understanding of network impacts on compliance may be because of the high degree of complexity involved in modeling taxpayers as part of a linked network of agents.

In this paper, we take advantage of an agent-based computational modeling approach by designing an agent model--the Networked Agent-Based Compliance Model (NACSM)--where taxpayers not only possess heterogeneous characteristics, but also exist within their own respective social networks. In addition, we build on elements of the existing compliance literature by incorporating the traditional aspects of compliance drivers (penalties, audits, and tax rates) as well. After giving a brief overview of agent-based modeling in the first section, we review in the second section the relevant compliance literature as well as research that has applied the computational social science approach to compliance. Then we discuss some of the basics of computational social science and how it relates to compliance. In the third section, we describe our model, and we discuss results in the fourth section. In the fifth section, we discuss some conclusions and point to areas where our research might be extended.

INTRODUCTION

The most pervasive (and well-employed) model of tax compliance, due to Allingham and Sandmo (1972) (hereafter referred to as "A&S"), describes taxpayer expected utility as a function of tax rates, potential penalties, and probability of audit. Essentially, this approach builds on Becker's (1968) rational economic agent approach to crime as a lottery. While this model has been used and modified over the decades to address tax compliance, researchers have long struggled with its major shortcoming--it fails to accurately predict at a macroscopic level the very phenomena it seeks to model. More specifically, the A&S model predicts much lower levels of compliance than actually observed in most industrialized nations. With true audit rates averaging at one percent (or less in some periods) in the past few decades, the real puzzle of compliance, as Alm (1991) notes, is why people pay taxes at all.

That the classic expected utility model of tax compliance yields poor predictions of macro compliance levels is not to say that taxpayers do not consider sanctions and penalties for evading, but rather points to other forces at play in individuals' compliance decisions. Since the advent of this model, however, researchers have widened their lens in looking at the causes of noncompliance. Specifically, the literature has focused on social drivers of compliance such as social norms, individual ethics, and how taxpayers reference themselves to others in their social network. Wenzel (2005) notes that social factors such as "ethics and norms" might not only create deterrents to compliance, but might also mitigate the impact of potential penalties associated with noncompliance. Myles and Naylor (1996) combine the standard model of tax evasion with the idea of social conformity to construct a framework where individuals gain some utility from payIng taxes honestly while also enjoying a payoff from conforming to the established social pattern of behavior.

More current research has sought to modify the A&S model to incorporate social factors. For example, Traxler (2006) modifies the classic approach to include "tax morale" and interprets this modification as an "internalized social norm" for tax compliance. His results yield higher levels of compliance given low audit rates. More recently, and in a somewhat different approach to the compliance issue, Feld and Frey (2007) argue that tax compliance is the result of a "psychological tax contract" where emotional ties and loyalties bond the taxpayer and state together to create a tax morale that reinforces compliance.

The Agent-Based Approach to Social Science

Agent-based computational social science involves analyzing social phenomena from the bottom up--i.e., modeling the individual agents of a system. The agent-based approach represents a rapidly evolving and powerful approach to analyzing complex systems. One of the most important characteristics and advantages of agent-based models (ABM), as compared to traditional comparative static models, for example, is the notion of emergent behavior--global, macroscopic patterns of behavior that originate from individual, microscopic agents following a set of rules and interacting with one another.

As noted by Axtell (2000) there are several distinct advantages of the ABM approach as compared to traditional mathematical modeling. First, agent rationality in such models can be modified to allow for less than completely rational agents. Second, though agents may follow the same methodological rule-set (as they do in our NACSM model) it is uncomplicated to endow agents with heterogeneous tastes and preferences, thus deviating from the standard "one agent represents all" approach common in the economics literature. Third, as agent based models are dynamical in nature, there is a record of each agent's behavior or actions from each period. Thus, agent models create a rich set of synthetic agent data that are the microscopic representatives of the emergent aggregate social structure. Lastly, and importantly for this paper, social networks matter in agent models. Interaction is a key strength of the agent-based approach, as agents can learn and mimic behaviors of other agents. Thus, ABM provides a flexible, natural description of a system that can capture emergent phenomena (Bonabeau, 2002).

Like traditional microsimulation approaches, the ABM approach can also yield insight into the impacts of policy changes on different segments of the population including distributional impacts (e.g., which population segment is most affected by a particular policy change). However, what makes the ABM approach different is that each agent assesses his situation or environment and then subsequently makes decisions according to some individual rule-set. These sets can be distributed across the agents in ways that allow for representation of a heterogeneous population, and can make explicit the results of interactions among the differing agent types. Non-linear, even chaotic, relationships can be observed and evaluated. This is unlike traditional microsimulation, for example, in which microdata do not decide anything--they are simply sifted through a maze of rigid business rules based upon statistical characteristics (e.g., taxable income, gender, head-of-household, etc.). Furthermore, with ABM, agents if so designed can evolve and exhibit unanticipated behaviors and patterns. Moreover, agent models can converge to a stable equilibrium, but as mentioned earlier, can have a great degree of variance in local agent behavior. For example, different agents may journey the gamut of compliance outcomes, with some underreporting income in some periods and reporting all income in other periods. Yet at the macroscopic view, we are in an overall compliant equilibrium setting. In other words, agent models can be characterized by high degrees of local disequilibria, but exhibit a stable macro-level equilibrium.

OVERVIEW OF ABM COMPLIANCE LITERATURE

Though a comprehensive review of the compliance literature is beyond the scope of this paper, it is necessary to briefly review the body of research in compliance that uses computational agent-based models. (1) There are several papers that have used mulfiagent-based simulation (MABS) models to examine the tax compliance issue (Bloomquist, 2006). Each of these models attempts to overcome the conventional A&S type approach to modeling compliance.

Bloomquist (2004) developed the tax compliance simulator (TCS)--an agent-based model designed to analyze taxpayer behavior under different enforcement regimes. The TCS can simulate agent responses to changes in audits and penalties, while also examining agent compliance behavior given the opportunities to evade afforded by "less visible" income. Relevant for our paper, Bloomquist incorporates social networks and finds that as the size of an agent's social network increases, the voluntary compliance rate in the population increases as well, thus pointing to the indirect deterrent effects of audits.