STATE COERCION, MORAL ATTITUDES, AND TAX COMPLIANCE: EVIDENCE FROM A NATIONAL FACTORIAL SURVEY EXPERIMENT OF INCOME TAX EVASION

Abstract

Why do some people comply with their obligation to pay taxes while others do not? Scholars of tax behavior, particularly economists and political scientists, have relied on models of state coercion and state reciprocity to answer this question. Neither state coercion nor state reciprocity, however, sufficiently account for individuals who voluntarily comply with their tax obligations to the state. We offer a third explanation, derived from the new sociology of morality and moral psychology, suggesting that two types of moral attitudes (moral imperatives and moral alignment) affect tax compliance. Using a factorial survey experiment of income tax evasion and a survey questionnaire administered to a nationally representative random sample of U.S. adults, we provide a systematic test of the three different models of tax compliance. The results yield strong support for moral attitudes (both moral imperatives and moral alignment) and state coercion, but little support for state reciprocity. We review the implications of our findings in the discussion and conclusion.

INTRODUCTION

The fiscal capacity of a state hinges on tax compliance, or the willingness of taxpayers to pay their taxes (Campbell 1993; Kiser and Karceski 2017; Levi 1988; Martin et al. 2009; Steinmo 1993, 2018; Tilly 1990). In the United States—where fiscal capacity is robust—most people comply with their obligation to pay taxes, but some do not. The tax gap, or the difference between the amount of tax imposed by tax authorities and the amount that is reported and paid by taxpayers, has fluctuated between 15 and 18 percent over the past 30 years (IRS 2016). The result of the tax gap is a loss of hundreds of billions of dollars in tax revenue each year (Mazur and Plumley 2007). This loss of revenue is the primary reason why the United States spends several billion dollars annually controlling tax evasion (IRS 2019). Tax evasion is a taxpayer’s willingness to deliberately break the law in order to reduce one’s tax liability (Kirchler 2007).¹ Tax evasion can come in many forms, but in the United States the average taxpayer evades income taxes in two ways: first, by overstating deductions and exemptions to which they are not entitled, and, second, by understating the amount of earned income. Our paper will focus on the latter and investigate taxable income generated from cash payments.

To prevent underreporting of income, modern states have designed tax systems in ways that limit opportunities for individuals to evade paying taxes. In the United States, most taxpayers are simply unable to misreport earned income due to the ubiquity of third-party reporting systems and personal income withholdings. Under this system, employers, rather than employees, report incomes to the tax administration, thereby making earnings more visible to the Internal Revenue Service (IRS). The IRS estimates that misreporting of income subject to third-party reporting is only 1 percent (IRS 2016). In contrast, misreporting of income subject to little or no third-party reporting, such as cash payments (Slemrod et al. 2017), is 63 percent (IRS 2016). Notably, a minority of individuals—37 percent—are willing to correctly state their income to the IRS despite a lack of oversight.

The challenge for social scientists has been to account for why some individuals willingly pay their taxes on incomes subject to little or no information reporting. The most common explanation of individual compliance under these conditions focuses on the deterrent effects of state coercion, or the ability of the state to discover and punish tax evasion. This explanation, which is rooted in the standard economic model (SEM) of taxation and crime (Allingham and Sandmo 1972; Becker 1968; Srinivasan 1973), specifies the economic costs and benefits of tax evasion. The strength of SEM is that it derives clear predictions. All else equal, as the certainty of detection and the severity of punishment rise, tax compliance should increase. But this rarely occurs in practice. In most OECD countries, audit rates (~ 1 to 2 percent) and levels of punishment (~ 10 to 20 percent fine of taxes owed) levied on tax evaders, if detected, are very low (Andreoni et al. 1998). Under these conditions, SEM predicts that tax evasion should be nearly universal. Yet, even on incomes subject to little or no information reporting, most OECD countries observe high rates of voluntary tax compliance (Alm, McClelland, and Schulze 1992; Kleven et al. 2011; Schneider and Enste 2013).

In reaction to SEM’s shortcomings, social scientists have taken two different theoretical routes. In economics, recent models broaden the scope of incentives and integrate psychological aspects of decision-making that diverge from the core assumptions of SEM. Contemporary models of state coercion, for instance, now include subjective beliefs and perceptions, cognitive biases, and learning effects (Andreoni et al. 1998; Kirchler 2007; Sandmo 2005; Slemrod 2007; Torgler 2007). In political science, models of state reciprocity predominate, which underscore noncoercive aspects of the state such as how the state treats and is perceived by taxpayers (Levi 1988, 1997; Tyler 1990, 2006). According to models of state reciprocity (Luttmer and Singhal 2014), if a state provides public goods that taxes are intended to fund, then individuals should voluntarily comply due to obligations to reciprocate with the state (Levi et al. 2009). In spite of these theoretical advances, research shows that contemporary models do not sufficiently improve upon SEM (Kirchler 2007; Slemrod 2007; Torgler 2007), and that tests of state reciprocity yield mixed support (Ariel 2012; Bergman 2009; Blumenthal et al. 2001; Bodea and LeBas 2014; Carrillo et al. 2017; Castañeda et al. 2020; Castro and Scartascini 2015; De Neve et al. 2019; Dwenger et al. 2016; Fellner et al. 2013; Pampel et al. 2019; Robbins and Kiser 2018). Given the empirical results, we aim to reassess the explanatory power of the microfoundational assumptions upon which SEM rests. We argue that SEM is incomplete and that nonpecuniary motivations beyond obligations to reciprocate with the state are necessary to account for tax compliance.

Drawing on the new sociology of morality and moral psychology, we contend that moral attitudes surrounding taxation are central to understanding tax evasion, and identify two morality-based mechanisms of tax compliance. The first consists of moral imperatives (Hitlin 2008; Hitlin and Harkness 2018; Hitlin and Vaisey 2013; Vaisey and Miles 2014), where people feel a moral obligation or duty to pay their taxes—be it reporting income or claiming accurate deductions—regardless of the conditions (Alm and Torgler 2006; Frey 1997). The literature on tax morale suggests that moral imperatives might be important determinants of tax compliance (Alm and Torgler 2006; Frey 1997; Torgler 2007; Wenzel 2004), and that some taxpayers in the United States feel a “fiscal responsibility” (civic duty or moral responsibility) when it comes to paying taxes (Torgler 2007; Pampel et al. 2019; Williamson 2017). We argue that when moral imperatives are accompanied by some degree of social consensus, rates of tax compliance should be higher than when moral imperatives are weak and not socially shared. The second consists of moral alignment (Haidt 2012), which is an individual’s psychological attachment to, sense of belonging with, and loyalty towards a political party (Green et al. 2004; Huddy 2001). In our model, moral alignment occurs when people perceive their values and morals as reflected in the values and morals of political actors in power. With respect to tax behavior, moral alignment should foster tax compliance to the extent that an individual identifies with the political actors who control the state (in-group members), and we expect tax evasion when there is divergence between values shared by taxpayers and the dominant political actors (out-group members) (Haidt and Graham 2007).

With the three competing models of state coercion, state reciprocity, and moral attitudes in mind, we developed a novel factorial survey experiment of income tax evasion and administered it to a nationally representative random sample of U.S. adults (N = 1,295). In our factorial survey experiment, respondents were exposed to ten different hypothetical situations in which portions of their personal income were derived from cash payments subject to no third-party reporting or income withholding. We experimentally manipulated situational treatments associated with state coercion and state reciprocity, and asked respondents the percentage of cash payments they intended to report to the IRS. We also included nonexperimental survey items that measured individual departures from the core assumptions of SEM, mechanisms of state reciprocity, and features of moral attitudes. Our goal with this design was to not only assess the explanatory power of moral attitudes, but to comparatively evaluate moral attitudes in relation to state coercion and state reciprocity.

Our factorial survey experiment and survey questionnaire yielded four principal discoveries. First, individuals intended to declare, on average, roughly 65 percent of their cash payments to the IRS. Most individuals, to varying degrees, were willing to evade income taxes, with a minority of respondents (12.1 percent) almost always willing to evade and a larger subgroup of respondents (41.2 percent) almost always willing to comply. Second, experimental manipulations of SEM as well as nonexperimental survey items showed that intentions to comply were driven by the certainty of detection and the severity of punishment, and were associated with prior histories of tax evasion and subjective probabilities of being audited. Third, intentions to comply were neither caused by nor associated with the experimental manipulations or nonexperimental measures of state reciprocity. Fourth, intentions to comply were strongly associated with an observational measure of moral imperatives, while moral alignment met our theoretical expectations. Overall, our results imply that tax compliance is caused by or associated with measures of moral attitudes and state coercion, with moral imperatives in particular having a greater influence on intentions to comply than state coercion.

THEORY AND LITERATURE REVIEW

State Coercion

Standard economic model.

Most theories of tax compliance begin with the standard economic model (SEM) (Allingham and Sandmo 1972; Srinivasan 1973), which is a deterrence model of crime (Becker 1968) applied to tax compliance and centered around state coercion. This model specifies the economic costs and benefits of tax evasion and assumes that taxpayers decide how much to evade taxes by maximizing expected utility. SEM contends that the expected utility of a taxpayer is a function of four basic parameters: the tax rate, the taxpayer’s earned income, the chance of getting caught (certainty of detection), and the size of the penalty for evasion (severity of punishment). SEM predicts that, all else equal, as tax rates and earned incomes rise, tax compliance should decrease, and that higher penalties and audit probabilities should increase tax compliance.²

Yet, as the originators of SEM attest: “This is a very simple theory, and it may perhaps be criticized for giving too little attention to nonpecuniary factors in the taxpayer’s decision on whether or not to evade taxes (Allingham and Sandmo 1972: 325).” While forty-plus years of subsequent research and analysis has provided some support for SEM (Andreoni et al. 1998; Kirchler 2007; Slemrod 2007; Torgler 2007), social scientists have identified various empirical and practical shortcomings. First, tests of SEM yield smaller than expected effect sizes (Kirchler 2007; Slemrod 2007; Torgler 2007), suggesting that tax compliance is influenced by something other than the maximization of income. Second, SEM predicts rates of tax compliance much lower than what the real-world produces (Andreoni et al. 1998). For incomes not reported by third-parties, which eliminates all but audit-based detection, most Western countries observe reasonable rates of tax compliance (Alm, McClelland, and Schulze 1992; Kleven et al. 2011; Schneider and Enste 2013). Third, state coercion is costly, stripping states of tax revenue (Campbell 1993; Tilly 1990). To minimize costs, states must rely on voluntary tax compliance, which limits the real-world applicability of the deterrence model. Taken together, SEM works in theory, but is insufficient in practice.

Deviations from SEM.

Contemporary models of state coercion have addressed the shortcomings of SEM by broadening the types of incentives capable of motivating tax compliance and by reassessing and modifying the core assumptions of instrumental rationality and utility maximization. The latter—modifications to the core assumptions of SEM—constitute novel contributions made by economists. To do so, economists have developed the concept of tax morale (Alm and Torgler 2006; Frey 1997; Torgler 2007), which is “as an umbrella term capturing nonpecuniary motivations for tax compliance as well as factors that fall outside the standard, expected utility framework (Luttmer and Singhal 2014: 150).” Much of this work considers psychological aspects of decision-making that diverge from SEM, such as incomplete and asymmetric information, subjective beliefs and perceptions, cognitive biases, and learning effects (Andreoni et al. 1998; Kirchler 2007; Sandmo 2005; Slemrod 2007; Torgler 2007). For example, SEM assumes perfect information and narrow choice sets, while contemporary modifications to SEM attempt to model misperceptions of auditing and sanctioning as well as larger choice sets (Alm 2012). According to such models, subjective probabilities of detection have a much greater effect on tax compliance than objective base rates of detection (Kirchler 2007). Furthermore, research shows that tax compliance—and criminal behavior more generally—follows a learning model in which tax compliance and risk perceptions are a function of new information drawn from prior experiences with tax evasion and the tax authorities (Antonides and Robben 1995; Grigoryeva 2018; Maciejovsky et al. 2007; Matsueda et al. 2006; Mittone 2006). This and countless other findings—such as the role of loss aversion or the overweighting of small probabilities—have been shown to impact tax compliance (Slemrod 2007). Yet the literature suggests that even expanded versions of the state coercion model are unable to explain most of the variation in tax compliance (Luttmer and Singhal 2014).

State Reciprocity

Knowing the poor predictive power of SEM, prominent political scientists and political psychologists have developed a family of models known as state reciprocity. These models emphasize noncoercive, or normative, aspects of the state (Levi 1988, 1997; Tyler 1990) and obligations to reciprocate with the state (Levi et al. 2009; Tyler 2006). What undergirds models of state reciprocity is the felt obligation to voluntarily pay taxes (or comply with laws) given the actions of the state and fellow taxpayers. Margaret Levi (1988, 1997) and Tom Tyler (1990, 2006, 2007) have identified three key ways in which states can foster a willingness to comply with the law: trustworthiness of government, procedural fairness and justice, and ethical reciprocity. We review each of the three exogenous factors below.

Trustworthiness of the state.

Taxation is part of a larger social contract in which the state provides goods and services (a government output) to individuals in exchange for tax payments (a taxpayer input). Trustworthy governments are “…motivated to deliver on their promises and do what is right for the people they serve, seeking policies that truly benefit their societies (Levi et al. 2009: 356).” In other words, policies and goods that are deliverable and serve the people indicate a trustworthy government. According to Levi et al. (2009), the perceived trustworthiness of a government hinges on assessments of the quality of government performance and its administrative competence. Government performance refers to the provision of public goods which ensure a minimal level of social welfare, such as national security, roads, drinkable water, sanitation, and education. Those governments that uphold their fiscal contract and deliver outputs that serve the people are more likely to foster voluntary tax compliance (Alm et al. 1993; Levi 1988, 1997; Smith and Stalans 1991; Timmons 2005). Administrative competence—or the government’s ability to control corruption and solve problems—is a necessary condition for the state to uphold its end of the fiscal contract. When administrative competence is low (e.g., when inefficiency and corruption are high), individuals come to believe that governments are incompetent and unable to deliver on their promises (Levi et al. 2009; Rothstein 2011; Seligson 2002), thereby reducing tax compliance (Williamson 2017).

Procedural fairness and justice.

Governments that exercise their authority through procedures that people perceive as fair and just, are more likely to be seen as legitimate (Tyler 1996, 2006; see also Rothstein 1998, 2011). As Levi et al. (2009) write, “This sense of fairness relies upon the beliefs that government officials follow a set of fair procedures and that they do so in a predictable and trustworthy fashion. When governments apply laws unevenly or target certain groups, disobedience is likely to increase (p. 360).” Applied to tax compliance, unfair and unjust governments undermine their own legitimacy and provoke tax evasion. Impartial political authorities that foster representation for all and abide by the rule of law elicit tax obedience and compliance.

Ethical reciprocity.

Ethical reciprocity refers to conditions under which individual compliance is a function of peer compliance, where peer compliance stems from administrative competence and state coercion (Levi 1997). For tax behavior, ethical reciprocity implies that individuals who believe that they should pay their taxes and who perceive others as behaving according to socially accepted rules (i.e., a social norm of tax compliance) will behave appropriately and pay their taxes (Bergman 2009; Levi 1997; Wenzel 2004; Williamson 2017). But if fellow taxpayers signal that tax fraud is tolerated, then compliance will decrease.

Empirical support is mixed for key components of state reciprocity. While observational research generally supports the positive effects of government trustworthiness and procedural fairness on compliance (e.g., Bodea and LeBas 2014; Jackson et al. 2012; Levi et al. 2009; Pampel et al. 2019; Tyler 1990), a growing body of experimental work suggests that the relation between state reciprocity and tax compliance is tenuous. Field experiments and factorial survey experiments manipulating the uses of tax revenue, corruption, and peer compliance across a wide variety of populations and contexts generate weak to null effects (Ariel 2012; Blumenthal et al. 2001; Castro and Scartascini 2015; De Neve et al. 2019; Dwenger et al. 2016; Fellner et al. 2013; Robbins and Kiser 2018; see Carrillo et al. 2017; Castañeda et al. 2020 for experimental research that supports state reciprocity). In short, empirical tests of state reciprocity show little improvement over models of state coercion.

Morality and Moral Attitudes

What accounts for tax compliance if not state coercion or state reciprocity? What other considerations are salient to individuals as they make decisions about tax compliance? To answer these questions, we turn to the new sociology of morality and moral psychology (Haidt 2012; Haidt and Graham 2007; Haidt and Joseph 2004; Hitlin 2003, 2008; Hitlin and Harkness 2018; Hitlin and Vaisey 2010, 2013; Stets and Carter 2012; Vaisey and Miles 2014), and argue that individuals pay taxes to the extent that tax compliance is viewed as moral. Morals and morality have to do with questions about what is right or wrong, acceptable or unacceptable in a society (Vaisey and Miles 2014). When we make moral judgments that something is good for us or that we have reason to act a particular way, we tend to be moved in the direction of those evaluations (Rosati 2016). Moral attitudes, then, are the moral judgments and evaluations—or the lines we draw between good and bad—that serve as conditions for motivation and action. Vaisey and Miles (2014) refer to these lines as “moral prohibitions”, while Hitlin (2008) calls them “bright lines”: moral signposts people use to demarcate acceptable and unacceptable regions of the social world. In terms of their character, moral attitudes vary between individuals and groups, are socially patterned, and a product of “differential exposure to social experience early in life [that] leads to the development of certain patterns of judgment, perspective, taste, and action (Vaisey and Lizardo 2016: 3).” ³ Below, we review two morality-based mechanisms that might account for variation in voluntary tax compliance: moral imperatives and moral alignment.

Moral imperatives.

Moral imperatives are intrinsic principles or duties—derived from moral judgments and attitudes—that motivate action one should or ought to perform independent of the situation (Bandura 1999; Hitlin 2008). That is, moral imperatives are acquired dispositions that remain constant across circumstances and contexts, and contain both inhibitive and proactive elements: the “should nots” and “ought nots” on the one hand, and the “shoulds” and “oughts” on the other (Bandura 1999). Applied to tax behavior, moral imperatives constitute the intrinsic obligations and duties of taxpayers to pay their taxes because it is the morally right thing to do (Alm and Torgler 2006; Frey 1997).⁴ In the U.S. context, research shows that taxpayers have a strong sense of civic duty and fiscal responsibility to pay taxes (Williamson 2017) and view tax fraud as unjustifiable (Alm and Torgler 2006), that between-person variation in fiscal responsibility has been observed (Pampel et al. 2019), and that tax compliance is linked to high levels of social capital, civic engagement, individualism, and political trust (Alm and Torgler 2006; Putnam 2000).

There are two prevailing views for how moral imperatives might affect tax compliance: (1) they could remove some possible actions (such as tax evasion) from the choice set entirely, or (2) they could affect the likelihood of choosing compliance relative to evasion net of other incentives and preferences. Regarding the first view, Elster (1989) and Etzioni (1988) contend that moral imperatives motivate action unconditionally and automatically without reference to ends or goals. Hitlin (2008) treats moral imperatives as “shortcuts that keep us from having to weight costs and benefits each and every time a short-term temptation rubs up against a long-term goal (p. 19).” Similarly, Wikström (2004, 2006) assumes that only a subgroup of actors— those who experience a discrepancy between their personal moral rules and the moral rules of a setting—weigh the costs and benefits of crime (Antonaccio and Tittle 2008; Kroneberg et al. 2010). In line with these arguments, tax evasion experiments show that a significant proportion of individuals behave honestly regardless of the conditions (see Kirchler 2007 for a review). Torgler (2003) refers to these people as honest taxpayers: individuals who always cooperate and do not search for ways to reduce their tax burden. The first view, in short, contends that moral imperatives (a) narrow the choices available to individuals (Kouchaki et al. 2018), (b) may not operate via rational deliberation of costs and benefits (Elster 1989; Etzioni 1988), and (c) act as general scripts for action (Hitlin 2008).

The alternative view suggests that moral imperatives are one of many parameters individuals evaluate in cases like tax compliance. Actors with moral imperatives prefer to act in accordance with their moral obligations and duties (Bowles 2016; Opp 1999; Smith and Wilson 2018), but these moral imperatives work in tandem with other microfoundations of social action and elements of the situation (Opp 1999). From this logic, moral imperatives offer a baseline starting point at which taxpayers are willing to evade taxes, from which taxpayers deviate in light of other dispositions (e.g, risk aversion) and/or characteristics of the situation (e.g., probability of detection). One goal of our study is to adjudicate between these two competing models of moral imperatives: Do moral imperatives trump other incentives and preferences or work in tandem with them to motivate tax compliance? Do individuals with strong moral obligations disregard state coercion and state reciprocity in their willingness to pay, or do individuals also consider state coercion and state reciprocity regardless of the strength of their moral obligations to pay taxes?

Moral alignment.

A growing body of work treats political partisanship as a social identity (Green et al. 2004): a subjective sense of belonging to a group (Tajfel 1970; Tajfel and Turner 1979). From this perspective, political partisanship is the result of people aligning their self-image to the types of people and groups associated with a political party. Once the alignment process takes place and people come to identify with a group, or in this instance a political party, people sort themselves politically on that basis and seek to protect and advance the status and electoral dominance of the party (Huddy 2001). We argue that people align according to the values and morals espoused by a political party, which should manifest as out-group enmity (e.g., desire to deprive opposition party of resources) and in-group amity (e.g., goodwill toward fellow party members). Recent research in sociology and moral psychology suggests that this could be the case. Hitlin (2003, 2008) argues that values and morals are strongly tied to the self and help shape personal and social identities, while Stets and Carter (2012) show that moral identities frame definitions of the situation that, in turn, help explain decision-making. Likewise, Haidt (2012) argues that an individual’s attitudes and beliefs about the nature and direction of the world are linked to membership in particular groups (Haidt and Graham 2007; Haidt and Joseph 2004). Taken together, we argue that moral alignment—a form of political partisanship—occurs when people perceive their values and morals as reflected in the values and morals of a political party.

Moral alignment has important implications for taxation. Being a moral partisan should affect tax compliance to the extent that political parties with whom one identifies control the state. For instance, a Democratic taxpayer should be more likely to voluntarily comply with tax payments when the state is controlled by Democrats than by Republicans. Tax compliance then is a function of common values shared by taxpayers and political actors, which suggests that the effect of political parties on tax compliance is conditional on whether said characteristics are shared by taxpayers. Neither of these factors—political values of taxpayers or political parties— should directly affect tax compliance (Williamson 2017). Only in alignment will they produce tax compliance for in-group members, and—when disjoint—will they produce tax evasion for out-group members.

Research investigating the effects of moral alignment on tax compliance is sparse. In the two studies closest to our own, Robbins and Kiser (2018) found that undergraduate students expected a hypothetical third-person to overstate deductions if and only if the political party of policymakers manipulated in a vignette diverged from the political party with whom student respondents identified. But the effect size was weak and not robust to alternative design specifications. Likewise, Cullen et al. (2018) observed that reported taxable income at the county-level increased as the aggregate voting behavior of counties became more politically aligned with the political party of the U.S. president. Overall, research has only begun to explore whether and how mechanisms of social identity and moral alignment affect tax compliance.

MATERIALS AND METHODS

Sample and Survey Mode

We developed a novel factorial survey experiment (FSE) of income tax evasion, which is a suitable method for investigating the mechanisms of tax compliance and their complex interactions (Auspurg and Hinz 2015), and supplemented it with a survey questionnaire to measure attitudes and socio-demographic characteristics of respondents. The University of Chicago’s National Opinion Research Center (NORC) coded and administered a web-based version of our study (FSE + survey questions) to a random sample of AmeriSpeak panelists using sampling strata based on age, race-ethnicity, education, and gender (48 strata in total). NORC invited 3,597 panelists to participate in the study on August 3^rd, 2017. 1,295 panelists completed the study by August 20^th, 2017, yielding a survey completion rate of 36 percent and a weighted cumulative response rate of 10.4 percent.⁵ While slight differences exist, benchmark comparisons revealed that the sample characteristics were representative of the U.S. population from which the AmeriSpeak panelists were randomly drawn (see Table 1). Panelists who completed the study were compensated with 3,000 AmeriPoints (the equivalent of $3 USD). Eligibility was restricted to U.S. adults age 18 and older who voluntarily consented to participate. The median study length was 13 minutes.

Table 1.

Comparison of Unweighted and Weighted Sample Characteristics to Population Benchmark

	Unweighted	Weighted	Benchmark	Difference
Household Income
Less than $30,000	24.7	25.6	20	5.6
$30,000 to $74,999	39.5	39.3	35.7	3.6
$75,000 to $124,999	23.9	22.9	23.8	0.9
$125,000 Plus	11.8	12.2	20.5	8.3
Age
18 −34	31	31.6	29.8	1.8
35 – 49	23.6	22.8	24.6	1.8
50 – 64	24.9	25.6	25.6	0
65 Plus	20.4	19.9	19.9	0
Race-Ethnicity
Non-Hispanic White	62.9	64.1	64.1	0
Non-Hispanic Black	11.7	11.8	11.8	0
Hispanic	17.8	15.9	15.9	0
Non-Hispanic Asian and Pacific Islander	2.5	2.7	6.1	3.4
Non-Hispanic Others	5.1	5.6	2.1	3.5
Education Status
Less than High School	4.9	10.8	10.8	0
High School Equivalent	17.7	28.8	28.8	0
Some College or Associate Degree	44.9	33.9	28.5	5.4
Bachelor’s Degree	19	15.5	20.3	4.8
Graduate Degree	13.5	11	11.5	0.5
Household Ownership
Owner Occupied	61.1	65	66.4	1.4
Renter Occupied and Other	38.9	35	33.6	1.4
Children in Household
With 1+ Under 18 Years	33.3	35.3	34	1.3
Without Children Under 18	66.7	64.7	66	1.3
Marital Status
Currently Married	50.3	49.6	53.4	3.8
Separated, Divorced, Widowed, or Single	39.7	50.3	46.6	3.8
Sex
Male	48.9	48.3	48.3	0
Female	51.1	51.7	51.7	0
Average Difference				1.9

Note: The “difference” column refers to the numerical difference between weighted and benchmark values. Benchmark values were drawn from the Current Population Survey.

Factorial Survey Experiment Design and Procedures

The typical FSE presents respondents with hypothetical scenarios (vignettes) that reflect real-life situations and stimuli. Within each scenario or vignette, experimenters manipulate attributes of the situation (dimensions) by randomly assigning attribute values (levels) to elicit judgments, decisions, or intentions with one or more dependent variables (evaluation task). For our FSE, panelists were presented with and asked to evaluate ten vignettes describing a hypothetical scenario in which the respondent is preparing to pay federal income taxes. Each vignette contained situational manipulations central to models of state coercion and state reciprocity. This allowed the researchers to exogenously manipulate features of the social context that are, in reality, highly correlated and endogenously self-selected by states, governments, and tax authorities. To guide the construction of our hypothetical scenario and vignette dimensions, we relied on (a) theory, (b) a pilot study (Robbins and Kiser 2018), (c) the taxation literature, and (d) actual tax collection practices observed in the United States. We ultimately settled on a FSE design that featured nine dimensions ranging from two to four levels per dimension, which yielded a factorial object universe of 17,496 (2¹ × 3 ⁷ × 4 ¹) unique vignettes.⁶ Since impossible combinations of dimensions did not exist, we included all possible combinatorial levels in the factorial object universe.⁷

After consenting to participate, respondents were shown a coversheet, which asked them to imagine the following:

“We will show you 10 scenarios. For each scenario, imagine the following. It is tax season and you are preparing to pay federal income taxes. Last year, you worked full-time as a paid employee. 10 percent of your total earnings was paid in cash, the rest was paid in reported wages. The Internal Revenue Service is not aware of your cash payments.

When you read a scenario, use all of the details to decide what you are likely to do. Each scenario has the same structure. But each scenario has different details. Please answer the questions after each scenario as best as you can.”

In our FSE, we focus on cash payments, and situations in which employers neither withhold nor report employee’s cash payments to tax authorities. We ultimately selected reporting cash payments to the IRS as our tax evasion scenario for four reasons. First, with cash payments, we can easily measure the magnitude of willingness to evade taxes by evaluating the percentage of cash payments respondents declare to tax authorities (an operationalization of tax evasion that corresponds to SEM). Second, most taxpayers believe, with some degree of uncertainty, that cash payments are undetectable by tax authorities since most cash payments are not subject to third-party reporting or income withholding (Schneider and Enste 2013). This shared knowledge creates a baseline condition conducive to tax evasion, while granting us freedom to manipulate the context of taxation without respondents being suspicious of unrealistic or implausible situations. Third, many U.S. taxpayers have experience with “under the table” cash payments and have been given cash for their labor at some point in their lives (Schneider and Enste 2013).⁸ Fourth, cash payments require less explication than other hypothetical tax evasion scenarios. A vignette of tax deductions on a U.S. tax return, for instance, would require detailed instructions and greater cognitive effort on behalf of respondents, which would foster systematic measurement error and item nonresponse (Auspurg and Jackle 2017).

After reading the coversheet, respondents were then shown ten vignettes randomly drawn with replacement from the vignette object universe.⁹ While the levels of each dimension were randomized, the order of dimensions was fixed from vignette-to-vignette. Our pilot study revealed that a similar design did not suffer from order effects or complexity effects (comparing 10 dimensions versus 6 dimensions), but that dimensions of SEM were necessary to manipulate in our design because of missing information effects caused by their exclusion (Robbins and Kiser 2018). For each of the ten vignettes, subjects were shown one evaluation task: the percentage of cash payments (ranging from 0 to 100) that the respondent would report to the IRS. All respondents were allowed to move forward during the study but were prohibited from moving back to correct earlier responses. This was done to maintain the assumption of temporal stability (that the timing of exposure to treatment and control is irrelevant). After evaluating the ten vignettes, participants filled out the sociodemographic questionnaire, were shown a debriefing statement, and thanked for their participation. Figures A1 and A2 in the Appendix provide examples of how the coversheet and vignettes were shown to AmeriSpeak panelists.

Vignette Evaluation Task

We operationalized tax evasion in our FSE with a measure of omission that captures the percentage of cash payments respondents intended to report to the IRS.¹⁰ In this way, our dependent variable measures what respondents say they would do (i.e., intention) given a hypothetical situation, and not necessarily what they did (or did not) do (i.e., behavior). To measure compliance intentions, respondents were shown the following question: “Prior to determining the amount of tax you owe, you must decide how much income from cash payments to report to the Internal Revenue Service. Given the conditions above, what percentage of income from cash payments will you report to the Internal Revenue Service?” We structured the evaluation task as a scale ranging from 0 (complete tax evasion) to 100 (complete tax compliance) in the form of a horizontal VAS with a slider bar (M = 65.12, SD = 39.90, N = 12753, min = 0, max = 100). The slider bar for each VAS was invisible. Respondents were instructed to click anywhere on the VAS in order to activate the slider bar. This was done to avoid anchoring effects caused by the placement of a visible slider bar.

We thus measured tax evasion (and tax compliance) as a behavioral intention. While we agree with the virtues of observed behavior as valid indicators of tax compliance, FSEs allow us to move beyond the logistical constraints of lab experiments and field experiments. With FSEs, we are able to experimentally test and comparatively evaluate operationalizations of state coercion, state reciprocity, and moral attitudes all within the same design (Auspurg and Hinz 2015). Note that attitudes and behavioral intentions measured in surveys and FSEs consistently predict and parallel behavior observed in the empirical world (Hainmueller et al. 2015). The predictive validity of FSEs is conditional on the extent to which a hypothetical scenario parallels choices and conditions observed in the empirical world. When these conditions are met, behavioral intentions strongly predict behavior (Hainmueller et al. 2015). This is why our FSE mirrors taxation in the U.S. context. As a result, we feel justified in testing the entirety of our model with behavioral intentions instead of parts of our model with observed behavior. In criminology, this is a common design choice due to the sensitive nature of crime (Antonaccio and Tittle 2008; Klepper and Nagin 1989; Kroneberg et al. 2010; Thurman 1989) and the difficulty in experimentally manipulating causes of crime (e.g., tax evasion; Alm 2012).

A general problem with measuring behavioral intentions, however, is that respondents might refuse to answer questions about willingness to commit tax fraud or lie because of social desirability. Our web-based general population survey minimized these problems in a number of ways. First, web-based surveys increase the social distance between researcher and respondent, which decreases social desirability bias (Dillman 2000). Second, social desirability distortions are less severe in samples of the general population than the university student population (Peterson 2001). Third, confidentiality and anonymity dampen social desirability bias (Lelkes et al. 2010), both of which AmeriSpeak safeguard by providing researchers with deidentified data. Finally, while measurement error (e.g., social desirability bias) and nonresponse error (e.g., refuse to answer) are different, both are driven by the sensitivity of questions. Since our dependent variable exhibited very little item nonresponse (1.52 percent missingness), we argue that the sensitivity of the topic—tax evasion—was minimal and likely did not motivate social desirability distortions in willingness to evade taxes.¹¹

Vignette Dimensions

In our scenarios, we assumed two given actors: the taxpayer and the U.S. government. For both actors, we exogenously varied incentives (e.g., income levels and tax rates) and manipulated dimensions of state reciprocity. The first four dimensions operationalized parameters common to state coercion (standard economic model), the next four dimensions operationalized parameters central to state reciprocity (trustworthiness of government and ethical reciprocity), and the last dimension operationalized a key component of moral attitudes (moral alignment). Because they could not be included in the vignettes, we measured moral imperatives in the survey questionnaire. Below, we review each vignette dimension and level (see Figure A3 in the Appendix for a list of all vignette dimensions and levels).

State coercion.

The amount of earned income referred to the total income during the last fiscal year, including reported wages and cash payments. We distinguished between three levels of earned income (twelve thousand dollars, fifty thousand dollars, and one hundred and forty thousand dollars) that were plausible and corresponded to the 10^th, 50^th, and 90^th percentile earned income levels in the United States, respectively. For each level of total earnings, we stated that 10 percent was derived from cash payments. This was the amount the respondent decided to report to the IRS. The second dimension, tax rate, refers the income tax rate imposed on the respondent by the federal government. We manipulated three levels of tax rates: 10 percent, 20 percent, and 30 percent. These percentages roughly mirrored the 2017 income tax rates for those who earned $12,000 a year, $50,000 a year, and $140,000 a year, respectively, in the United States.

We manipulated deterrence with two dimensions. The first dimension, certainty of detection, operationalized the probability of the respondent getting caught for evading taxes. The levels of this dimension captured low (1 percent), moderate (6 percent), and high (11 percent) probabilities of detection. The values we manipulated mirrored real-world chances of detection in the United States (i.e., 1 percent) with values slightly greater than what was objectively true (i.e., 6 percent and 11 percent).¹² The second dimension, severity of punishment, provided information about the magnitude of punishment given the respondent’s decision to underreport cash payments to the IRS. This dimension manipulated a wide spectrum of punishments for tax evasion that have been observed in the United States legal system: no punishment, a fine equal to 10 percent of tax on the return, or a fine equal to 20 percent of tax on the return.

State reciprocity.

We employed four dimensions to capture state reciprocity. The first two dimensions, state corruption and state inefficiency, operationalized state-level inputs preventing the optimal provision of public goods that income taxes were intended to fund (Levi 1988, 1997). State corruption consisted of three levels: 1 percent, 11 percent, or 21 percent of congress members used public funds for personal purchases or private gains. All three dosages encapsulated key dimensions of corruption with varying magnitudes (Rothstein 2011). For state inefficiency, we again employed three levels: 1 percent, 11 percent, or 21 percent of the federal budget was wasted every year on government programs that took longer and cost more than what the private sector could provide. The goal of this treatment was to illustrate varying levels of government waste that would yield a suboptimal provision of public goods and services.

The third dimension, public goods provisions, manipulated the types of public goods provided by the state to its citizens. With respect to public goods provisions, we operationalized four levels centered on how federal tax revenue would be spent the following fiscal year: neither an increase nor a decrease in funding for government programs, a 5 percent increase in government funding for healthcare subsidies, the military, or education. Preferences for public goods vary within and between populations. As a result, we presumed a greater willingness on the part of respondents to pay taxes for outputs they desired and evade taxes for outputs they opposed (Scott 1985). Given this logic, we expected familywise variation within this dimension.

The fourth dimension, taxpayer compliance, operationalized ethical reciprocity. For the present study, we manipulated the extent to which fellow taxpayers underreported cash payments to the IRS. We use three levels—10 percent, 40 percent, or 70 percent of all possible taxpayers. The upper bound (i.e., 70 percent) was based on rates of tax evasion estimated for shadow economies, particularly state use tax reporting. State use taxes, which are taxes due on sales purchased from out-of-state but consumed within the state of residence, are largely unenforceable and produce high rates of evasion (Slemrod et al. 2017). We used estimates of state use tax evasion as a benchmark since cash payments exhibit similar issues of monitoring and sanctioning of noncompliance. In short, the levels selected for taxpayer compliance parallel rates of tax evasion on personal incomes in the real-world not subject to third-party reporting or income withholdings.

Moral attitudes.

Testing moral alignment required an interaction between political parties in power and the party affiliation of respondents. The final dimension we manipulate, policymaker characteristics, captured the former. This dimension measured the political party in the United States that controlled the executive and legislative branches of government. Given the two-party system in the United States, we manipulated two levels for this dimension: Democrats or Republicans control the White House and both houses of the U.S. congress. While a main effect for the policymaker characteristics dimension may exist, our primary interest was in heterogenous treatments effects, specifically if and how the policymaker characteristics dimension interacted with the party identification of respondents measured in our survey questionnaire.

Individual-Level Variables

Following the ten randomly selected vignettes, respondents were shown a questionnaire that included survey items of attitudes toward the state, deviations from SEM, and moral attitudes. We measure key theoretical parameters with a survey questionnaire for the following reasons. While situational variables lend themselves to experimental manipulation, dispositional variables prove difficult to manipulate. Some attitudes are best treated as dispositions that affect tax compliance regardless of changes to the tax situation in which compliance decisions are made. This line of argumentation converges with our conceptualization of moral imperatives and other attitudes. Dividing the study into experimental and nonexperimental parts allows us to comparatively evaluate the three models of tax compliance but with caveats: treatments manipulated in the vignettes can be safely treated as causal, while observational treatments like moral imperatives should be interpreted as correlational.

Below, we review and summarize our individual-level variables. Table 2 provides descriptions and descriptive statistics of key individual-level variables. For zero-order correlations between all key individual-level variables, please see Table A1 in the Appendix.

Table 2.

Sample Descriptive Statistics of Key Independent Variables

Name	Definition	Mean	SD	Min/Max	N
State reciprocity
Procedural fairness and justice	4-item row-mean scale (0 = strongly disagree, 6 = strongly agree) (α = .87) (1) The US government makes sure that I get what I need, (2) The outcomes of decisions that the US government makes are fair to me, (3) The US government treats me in a fair way, and (4) The US government respects people like me	2.46	1.36	0/6	1274
Trustworthiness of government	4-item row-mean scale (0 = completely distrust, 8 = completely trust) (α = .83) (1) Congress and the Senate, (2) Public authorities, (3) Courts, and (4) The police	3.65	1.89	0/8	1275
Deviations from SEM
Subjective probability of being audited	Percent chance of being audited by the IRS (0 to 100 VAS)	28.79	28.36	0/100	1248
Prior tax fraud	Underreport income, overstate deductions, or claim false deductions (1 = yes, 0 = no)	0.08	---	0/1	1165
Prior audit	Ever been audited by the IRS (1 = yes, 0 = no)	0.17	---	0/1	1240
Prior punishment	Ever been warned, fined, or punished for evading income taxes (1 = yes, 0 = no)	0.04	---	0/1	1260
Moral attitudes
Moral imperatives	3-item row-mean scale (0 = strongly disagree, 6 = strongly agree) (α = .81) (1) I have a moral obligation to pay taxes, (2) I think it is wrong to evade taxes, and (3) If I did not pay taxes, I would have a bad conscience	4.51	1.48	0/6	1286
Party identification	Unfolding party ID scale from ANES (0 = Strong Democrat, 6 = Strong Republican)	2.65	1.94	0/6	1292
Attitudes toward the state
Identification with the state	4-item row-mean scale (0 = strongly disagree, 6 = strongly agree) (α = .91) (1) Being an American is central to who I am, (2) I identify myself as an American, (3) People who know me well would call me patriotic, and (3) Being a patriot is an important reflection of who I am	4.39	1.33	0/6	1239
Roles of the state	6-item row-mean scale (0 = not at all responsible, 4 = entirely responsible) (α = .91) (1) To ensure a job for everyone who wants one?, (2) To ensure adequate health care for the sick?, (3) To ensure a reasonable standard of living for the old?, (4) To ensure a reasonable standard of living for the unemployed?, (5) To ensure sufficient child care services for working parents?, and (6) To provide paid leave from work for people who temporarily have to care for sick family members?	2.20	1.08	0/4	1265
Public goods preferences	On which program should the US government spend the majority of its federal budget?
Defense programs	Defense programs (e.g., military expenditures, national intelligence agencies, etc.)	0.21	---	0/1	1281
Health insurance programs	Health insurance programs (e.g., Affordable Care Act, Medicaid, Medicare, etc.)	0.28	---	0/1	1281
Safety net programs	Safety net programs (e.g., food stamps, unemployment insurance, Child Tax Credit, etc.)	0.07	---	0/1	1281
Social Security programs	Social Security programs (e.g., retirement benefits, disabled worker benefits, etc.)	0.18	---	0/1	1281
Other public services	Other public services (e.g., transportation and infrastructure, science and medical research, etc.)	0.26	---	0/1	1281

State reciprocity.

We also operationalized state reciprocity with preexisting attitudes toward the state. This was done since individual attitudes may matter more for tax compliance than manipulated treatments of state reciprocity. As such, our questionnaire included survey items tapping into the perceived trustworthiness of government and perceptions of procedural fairness and justice (see Table 2).

Deviations from SEM.

Our survey also included items intended to capture psychological aspects of decision-making. These items measured subjective probabilities of detection, prior history of income tax fraud, prior experience with being audited, and prior punishment for evading income taxes (see Table 2).

Moral attitudes.

To capture moral imperatives, we relied on three items in response to a question asking about obligations people have to pay taxes. Testing moral alignment, on the other hand, required an interaction between the values of political authorities and taxpayers. The policymaker characteristics dimension captured the former, while party identification was used to measure the values of taxpayers. We used an unfolding version of the measure drawn from recent iterations of the American National Election Study (see Table 2).

Individual-Level Control Variables

To account for possible confounding at the individual-level, we controlled for attitudes toward the state, classic sociodemographic characteristics, and design-based features of the study. In terms of attitudes toward the state, we controlled for identification with the state, roles of the state, and preferences for particular public goods. These measures were embedded as survey items in the questionnaire and can be found in Table 2.

In terms of sociodemographic characteristics, we controlled for age, gender, race-ethnicity, educational attainment, marital status, citizenship status, logged per capita household income, employment status, metro area, and U.S. region. With respect to design-based variables, we controlled for telephone service in the household, household internet access, date of completion, and estimated length of the survey. The sociodemographic characteristics and design-based variables were included as deliverables by AmeriSpeak and were not embedded in the FSE or questionnaire.

Analytic Strategy

The research design yields panel data in which i vignettes (i = 1,…, 10) are nested within j individuals (j = 1,…, 1295). As a result, we estimated two-level hierarchical linear models with and without higher level moderation of lower level effects (Gelman and Hill 2007).

The level-1 (or within-level) model takes the following form:

$$Y_{ij}={\beta }_{0j}+{\beta }_{1j}{X}_{ij}+e_{ij}$$ (1)

where Y_ij is a continuous measure of intentions to comply for the ith vignette in the jth individual, β_0j refers to the intercept of intentions to comply for individual j, β_1j is a fixed slope for X_ij (which is a vector of vignette dimensions treated as dummy variables), and e_ij is a disturbance term that varies over the population of vignettes (assumed normal, independent, and identically distributed).

We specify a level-2 model of between-individual variation in intentions to comply by modeling the intercept, β_0j, from equation 1:

$${\beta }_{0j}={\gamma }_{00}+{\gamma }_{01}{W}_j+u_{0j}$$ (2)

where γ₀₀ refers to the overall population intercept for intentions to comply (which is the grand mean of compliance intentions across all individuals when all predictors are equal to 0), γ₀₁ is an overall slope for W_j (which is a vector of individual-level level variables that vary across individuals but not vignettes), and u_0j is a random disturbance term for the deviation of the intercept of an individual from the overall intercept γ₀₀ (assumed normal, independent, and identically distributed).

Cross-level interaction effects test whether the impact of X_ij on intentions to comply is a function of W_j (individual-level variables). The cross-level interactions are modeled with β_1j from equation 1:

$${\beta }_{1j}={\gamma }_{10}+{\gamma }_{11}{W}_j+u_{1j}$$ (3)

where β_1j refers to the slope for X_ij in individual j, γ₁₀ is the overall regression coefficient or slope between Y_ij and X_ij, γ₁₁ is a slope for the cross-level interaction between X_ij and W_j, and u_1j is a random disturbance term for the deviation of the slope of an individual from the overall slope γ₁₀ (assumed normal, independent, and identically distributed).

Since our dependent variable was bounded between 0 and 100, exhibited data “clumping” at 0 and 100 (see Figure 1), and was a quantitative measure, three modeling procedures were possible: classic linear model, Tobit model with left- and right-censoring, and fractional logit model. We ultimately present two-level hierarchical linear models since (a) the interpretation of estimates from linear regression models is more intuitive than Tobit and fractional logit models, (b) HLM fitted values produced estimates within the lower and upper bounds of 0 and 100, respectively, and (c) we found little to no difference in estimates from linear, Tobit, and fractional logit models (see Table S1 in the Supplemental Materials Online).

Figure 1.

Histogram of Percentage of Cash Payments Respondents are Willing to Report to the Internal Revenue Service (N = 12,753)

Given the longstanding debate about how to properly address nonresponse error and adjustment error (Arel-Bundock and Pelc 2018; Gelman 2007; Groves et al. 2009; Pepinsky 2018), we present models using listwise deletion without post-stratification survey weights. HLMs estimated using (a) multiple imputation with and without post-stratification survey weights (m = 50), and (b) listwise deletion with post-stratification survey weights can be found in Tables S2 and S3 of the Supplemental Materials Online. The results are largely consistent across all models (see the Results section for a greater discussion of the similarities and differences). The Supplemental Materials online provide a discussion of our modeling assumptions, and an analysis of the assumptions necessary for causal identification.

RESULTS

Descriptive Analysis

Individuals were—on average—willing to declare more than 65 percent of their cash payments to the IRS (M = 65.12). This central tendency was not consistently observed across vignettes as evidenced by the spread of the dependent variable (SD = 39.90). To further illustrate the distribution of our data, Figure 1 presents a histogram of the percentage of cash payments respondents were willing to report to the IRS across the 12,753 evaluated vignettes. As stated previously, Figure 1 reveals floor and ceiling effects at 0 and 100, respectively, with a propensity toward ceiling effects in the data (i.e., values at 100). This suggests that people were more prone to report 100 percent of their income from cash payments than any other percentage. In fact, 16.81 percent of individuals (N = 217) produced individual-specific means of 100, while 41.21 percent of individuals (N = 542) produced individual-specific means of 90 or above. While the findings suggest a tendency toward tax compliance, 62.52 percent of the values of the dependent variable were below 100. Most individuals, to varying degrees, were willing to evade income taxes, with a minority of respondents always willing to evade: 4.6 percent of individuals (N = 61) produced individual-specific means of 0, while 12.1 percent of individuals (N = 155) produced individual-specific means of 10 or below.¹³

The descriptive analysis suggests that a small subgroup of individuals intended to declare all of their cash payments to the IRS (i.e., fully comply), and that a minority of individuals intended to declare none of their cash payments (i.e., fully evade). The majority of individuals, however, were willing to partially evade (78.59 percent), and declare some—but not all—of their cash payments. Interestingly, the distribution of intentions to comply in Figure 1 was similar to patterns observed in other experimental studies of tax compliance (Alm, McClelland, and Schulze 1992; Alm et al. 2015; Zhang et al. 2016), as well as real-world tax declarations. The descriptive statistics, for instance, parallel estimated rates of tax evasion (~ 63 percent) and tax compliance (~ 37 percent) observed on incomes not reported to the IRS or subject to withholdings (IRS 2016), suggesting that our hypothetical vignettes might be good predictors of actual behavior (see also Alm et al. 2015). The descriptive statistics provide preliminary support for (a) SEM (in that the majority of respondents in the FSE are willing to accept some degree of tax evasion under conditions of low probabilities of detection and low levels of punishment), and (b) moral imperatives (in that a subgroup of respondents in the FSE are willing to fully comply). We now turn to the results of our two-level hierarchical linear models.

Hierarchical Linear Models

To start, a null model yielded a relatively large intra-class correlation (ICC) of .78, where the constant term suggested high mean-levels of reporting cash payments to the IRS, u₀ = 65.04, SE = 0.99, p < .001, with variation observed around u₀ as indicted by the level-2 disturbance term, SD(u_0j) = 35.27 (the level-1 disturbance term also exhibited variation, SD(e_ij) = 18.67). The ICC tells us that 22 percent of the variation in reporting hypothetical cash payments to the IRS was accounted for by situational characteristics of the vignettes (i.e., manipulated dimensions). Stated differently, the bulk of the variation—78 percent—was between individuals, while 22 percent of the variation was within individuals; suggesting that rates of tax compliance reported by the same individual tended to be similar.

Vignette dimensions.

Table 3 reports average treatment effects of the nine vignette dimensions we manipulated in our FSE without (model 1) and with (model 2) individual-level variables. The regression coefficients are substantively meaningful, and represent the amount of change in the percentage of cash payments respondents were willing to report to the IRS due to a one-unit change in a vignette dimension or individual-level variable.¹⁴ We report results of familywise and pairwise tests for vignette dimensions.

Table 3.

Two-Level Hierarchical Linear Models of Vignette Dimensions and Individual-Level Variables on Percentage of Cash Payments

	Model 1	Model 2
Level-1: Vignette Dimensions
Earned income (Ref.: $1,200 cash payments)
$5,000 cash payments	3.97*** (0.58)	4.48*** (0.70)
$14,000 cash payments	6.69*** (0.69)	7.36*** (0.81)
Tax rate (Ref.: 10% tax rate)
20% tax rate	−0.21 (0.43)	−0.39 (0.52)
30% tax rate	−0.78 (0.44)	−1.18* (0.52)
Certainty of detection (Ref.: 1% chance of detection)
6% chance of detection	2.08*** (0.48)	2.08*** (0.57)
11% chance of detection	3.26*** (0.51)	3.23*** (0.59)
Severity of punishment (Ref.: Warning)
10% of tax on the return	1.31** (0.42)	1.33* (0.51)
20% of tax on the return	1.93*** (0.45)	2.71*** (0.55)
State corruption (Ref.: 1% of congress misuse public funds)
11% of congress misuse public funds	−0.16 (0.41)	0.09 (0.49)
21% of congress misuse public funds	−0.22 (0.43)	−0.26 (0.51)
State inefficiency (Ref.: 1% of budget wasted)
11% of budget wasted	−0.68 (0.41)	−0.74 (0.48)
21% of budget wasted	−0.88* (0.44)	−1.34* (0.53)
Taxpayer compliance (Ref.: 10% of taxpayers underreport)
40% of taxpayers underreport	−0.38 (0.40)	−0.21 (0.46)
70% of taxpayers underreport	−0.07 (0.43)	−0.16 (0.50)
Public goods provisions (Ref.: No change in funding)
5% increase in funding for healthcare subsidies	0.18 (0.50)	−0.10 (0.59)
5% increase in funding for the military	−0.08 (0.48)	−0.20 (0.59)
5% increase in funding for education	−0.28 (0.48)	−0.36 (0.57)
Policymaker characteristics (Ref.: Democrat control)
Republican control	0.02 (0.36)	−0.04 (0.43)
Level-2: Individual-Level Variables
Reciprocity
Trustworthiness of government		0.29 (0.71)
Procedural fairness and justice		1.86 (0.99)
Deviations from SEM
Subjective probability of being audited		0.21*** (0.03)
Prior tax fraud (1 = yes, 0 = no)		−24.29*** (3.83)
Prior audit (1 = yes, 0 = no)		−0.85 (3.02)
Prior punishment (1 = yes, 0 = no)		7.93 (6.54)
Moral attitudes
Moral imperatives		7.04*** (0.95)
Party identification		0.74 (0.69)
Attitudes toward the State
Identification with the state		−0.75 (0.98)
Roles of the state		1.33 (1.25)
Public goods preferences (Ref: Defense programs)
Health insurance programs		−2.92 (3.61)
Safety net programs		−4.37 (5.60)
Social Security programs		−1.70 (3.53)
Other public services		−1.84 (3.35)
Intercept	59.81*** (1.32)	20.84 (18.12)
SD(intercept)	35.30	31.00
SD(residuals)	18.39	18.18
Individual-level controls	No	Yes
Observations	12753	9104
Respondents	1291	916

^***p < 0.001,

^**p < 0.01,

^*p < 0.05 (two-tailed)

Note: unstandardized coefficients (robust standard errors).

As model 1 in Table 3 shows, we discovered strong support for state coercion (i.e., SEM). In particular, we found a statistically significant effect of earned income on intentions to report cash payments at the p < .05 level, χ² (2) = 93.86, p < .001. Post hoc comparisons indicated that earning $5,000 and $14,000 in cash payments increased the percentage of cash payments respondents were willing to report to the IRS (see Table 3), and that earning $5,000 significantly differed from earning $14,000, χ² (1) = 31.08, p < .001. We did not, however, find a statistically significant effect of the tax rate on intentions to report cash payments to the IRS at the p < .05 level, χ² (2) = 3.44, p = .179.

We also found that the certainty of detection and severity of punishment strongly influenced the percentage of cash payments respondents were willing to report to the IRS. Regarding the former, certainty of detection significantly increased intentions to comply at the p < .05 level, χ² (2) = 40.93, p < .001. Post hoc comparisons revealed that the effect was monotonic: an increase from 1% to 6% in certainty of detection increased intentions to comply by 2.08, while an increase from 1% to 11% increased intentions to comply by 3.26; both of these effects were statistically significant (see Table 3), as well as the difference—Δ1.18—between 6% and 11%, χ² (1) = 7.56, p < .01. Regarding the latter, severity of punishment also significantly increased intentions to comply at the p < .05 level, χ² (2) = 18.73, p < .001. Post hoc comparisons showed similar monotonic effects: an increase from a warning to a fine of 10% of tax on the return increased intentions to comply by 1.31 (see Table 3), an increase from no punishment to a fine of 20% of tax on the return increased intentions to comply by 1.93 (see Table 3), while the difference between a fine of 10% and 20% of tax on the return—Δ 0.62—was statistically nonsignificant, χ² (1) = 2.33, p = .126.

Turning to the four parameters capturing mechanisms of state reciprocity, we observed no statistical support for any of the dimensions: state corruption, χ² (2) = 0.29, p = .866, state inefficiency, χ² (2) = 4.25, p = .119, taxpayer compliance, χ² (2) = 1.00, p = .607, and public goods provisions, χ² (3) = 0.85, p = .836, all produced statistically nonsignificant results at the p < .05 level (see model 1, Table 3). A similar null effect was observed for policymaker characteristics (see Table 3), which was expected.

Finally, despite the inclusion of individual-level variables and the resulting loss of information due to listwise deletion, the vignette dimensions in models 1 and 2 yielded similar results. Regardless of the model, statistically significant regression coefficients ranged between absolute values of 1 and 7 percentage points, which are small effect sizes.

Individual-level variables.

With respect to overall findings, model 2 in Table 3 shows that individual-level measures of state reciprocity—trustworthiness of government and procedural fairness and justice—were statistically nonsignificant. Measures of deviations from SEM, on the other hand, produced mixed results. Subjective probabilities of being audited were statistically significant and positively related to the individual-specific mean percentage of cash payments respondents were willing to report to the IRS. The estimated effect size was also substantial: for every one-unit increase in the subjective probability of being audited (a scale that ranges from 0 to 100), we observed a 0.21-unit increase in the individual-specific mean percentage of cash payments respondents were willing to report to the IRS (a difference of 21 percentage points between subjective probabilities set at their lowest, 0, and highest, 100, values). Among our learning effects, only one variable generated statistically significant findings: committing tax fraud in the past, which was negatively related individual-specific mean intentions to comply. Again, a nonnegligible effect: individuals who committed tax fraud in the past were willing to report, on average, 24 percentage points less of their cash payments to the IRS than individuals who had never committed tax fraud.

With respect to moral attitudes, moral imperatives and party identification both affected compliance in the hypothesized directions. Moral imperatives exerted a strong and statistically significant effect on individual-specific mean levels of intentions to report cash payments to the IRS. In fact, a one-unit increase in moral imperatives generated a 7.04 percentage point increase in the amount of cash payments respondents were willing to report to the IRS (a difference of 42.24 percentage points between moral imperatives set at their lowest and highest values), a substantial effect. As expected, the average treatment effect of party identification was statistically nonsignificant.

Our controls for attitudes toward the state—identification with the state, roles of the state, and preferences for particular public goods—yielded statistically nonsignificant results.

Multilevel moderation.

In Table 4, we explored the role that moral alignment plays in promoting and undermining tax compliance. To do so, the two-level HLM found in model 2, Table 3 was estimated with higher level moderation of lower level effects. The higher-level moderator was the individual-level measure of party identification and the lower-level effect was the policymaker characteristics vignette dimension. Table 4 provided strong support for the hypothesis that the effect of political values of state authorities on tax compliance was moderated by the party identification of respondents. The results indicate that the interaction term between policymaker characteristics and party identification was statistically significant and positive, b = 0.60, SE = 0.20, p < .01.¹⁵

Table 4.

Two-Level Hierarchical Linear Models with Cross-Level Moderation between Party Identification and Republican Control on Percentage of Cash Payments

	Model 1
Vignette dimension
Policymaker characteristics (Ref.: Democrat control)
Republican control	−1.65* (0.71)
Moral attitudes
Party identification	0.44 (0.70)
Cross-level interaction
Party identification X Republican Control	0.60** (0.20)
Intercept	21.48 (18.13)
SD(intercept)	30.95
SD(republican control)	4.83
SD(residuals)	18.00
Vignette dimensions	Yes
Individual-level variables	Yes
Individual-level controls	Yes
Observations	9104
Respondents	916

^***p < 0.001,

^**p < 0.01,

^*p < 0.05 (two-tailed)

Note: unstandardized coefficients (robust standard errors). Model 1 includes vignette dimensions, individual-level variables, and individual-level controls.

To illustrate how the effect of policymaker characteristics on intentions to comply varies across values of party identification, Figure 2 graphically illustrates the marginal effect of Republican control. The solid blue line indicates how the slope of Republican control changes across values of party identification. 95 percent confidence intervals (CI) demonstrate whether—for a given value of party identification—Republican control significantly affects tax compliance. This occurs when the upper and lower bounds of the 95 percent CI are above (or below) the solid red zero line.

Figure 2.

Marginal effect of Republican control by party identification.

Figure 2 demonstrated that the slope of Republican control was statistically significant and negative when respondents identified as strong Democrats. But as party identification moved from strong Democrat to strong Republican, the slope of Republican control became statistically nonsignificant and approached zero. Only once respondents identified as moderate or strong Republicans did the slope of Republican control become statistically significant and positive. Results suggest that intentions to report cash payments to the IRS are driven by in-group (Republican control, Republican party) and out-group (Republican control, Democratic party) biases.

Sensitivity checks.

Next, we investigate how the results varied by modeling decisions (all sensitivity checks can be found in the Supplemental Materials Online). Stated previously, main effects yielded consistent results regardless of statistical adjustments for item nonresponse (i.e., multiple imputation) and/or known differences between sample and population (i.e., post-stratification survey weights) (Tables S2). The main effects were also robust to estimating random slopes for all vignette dimensions (Table S4). Cross-level interaction effects between Republican control and party identification, however, were more sensitive to statistical adjustments (Table S3). While the cross-level interactions were not sensitive to multiple imputation, post-stratification survey weights (with listwise deletion or multiple imputation) reduced the effect of in-group bias but not out-group bias. Moreover, removing the middle “independent” category did not alter the results (Table S5), while treating party identification as a categorical variable lead to slight differences in statistical significance but not the observed in-group/out-group effect (Table S6). Taken together, substantive results were consistent regardless of estimation procedure.

Heterogeneity checks.

We now examine how treatment effects varied by subgroups (heterogeneity checks can be found in the Supplemental Materials Online). First, we test whether key parameters of SEM were interactively related (versus additive). In support of prior research on taxation (Blackwell 2010), we found that the interaction effects between earned income and tax rates, χ² (4) = 3.15, p = .533, as well as certainty of detection and severity of punishment, χ² (4) = 9.58, p = .050, were statistically nonsignificant (Table S7). Second, cross-level interaction effects between per capita household income and earned income dimension were statistically significant (Table S8). The results show that the marginal effect of earned income on intentions to comply was always positive and increased in magnitude as respondents’ household income increased. In other words, positive effects of the earned income dimension were greater for high-income respondents than low-income respondents. Third, cross-level interaction effects between individual-level measures of state reciprocity (as well as attitudes toward the state) and dimensions of state reciprocity yielded statistically nonsignificant results, showing that the null treatment effects of our manipulations of state reciprocity were constant across peoples’ preexisting attitudes toward the state (Tables S9 through S13). Fourth, cross-level interaction effects between moral imperatives and dimensions of SEM (Table S14) and state reciprocity (Table S15) yielded statistically nonsignificant results, demonstrating that the effect of moral imperatives is additive and appears to act as a baseline starting point for intentions. In short, moral principles did not necessarily trump SEM when situational cues dictated the use of rational decision-making (or vise-versa), and the effect of moral imperatives on intentions to comply was independent of how many other taxpayers evaded taxes. Morality and SEM, in other words, operate additively and unconditionally in the determination of intentions to comply.

DISCUSSION AND CONCLUSION

We leveraged evidence from a nationally representative FSE to theoretically adjudicate between multiple mechanisms of tax compliance. Below, we review a number of significant discoveries and compelling implications that can be drawn from our data and findings.

State coercion is a long-standing tradition within economics, criminology, and political science. Our results reveal both the relevance and shortcomings of SEM for explaining tax compliance. Theoretically, SEM predicts that tax compliance should decrease as earned income and tax rates increase (Allingham and Sandmo 1972; Srinivasan 1973). Empirically, we observed the opposite: earned income yielded the largest effect of any manipulation (greater income led to greater tax compliance), and tax rates were statistically unrelated to tax compliance. While our findings parallel prior research (Alm, Jackson, and McKee 1992), recent inquiries into the tax behavior of high-net-worth individuals presents a mixed picture. Young et al. (2016), for instance, investigated whether top income-earners moved to locations with lower taxes as a way to lessen their tax burden, but found this rarely occurred. On the other hand, Zucman (2015) and colleagues (Alstadsæter et al. 2019) found that the world’s richest 0.01 percent conceal about 40 percent of their total personal fortunes in tax havens, and evade about 25 percent of the income taxes they are obligated to pay. Because of these discrepancies in findings, additional research investigating intentions to comply among high-net-worth individuals would be a welcome addition to the tax literature.

The implications of our results for models of deterrence were also compelling. This tradition generally shows that monitoring and sanctioning increase tax compliance (Andreoni et al. 1998; Kirchler 2007) and reduce general crime to a small degree (Pratt et al. 2006). We found strong support for the additive (rather than multiplicative) effects of certainty of detection and severity of punishment on intentions to comply. All else equal, taxpayers who experience a higher penalty rate or a higher probability of detection are more likely to report cash payments than taxpayers who experience weaker penalties or lower probabilities of detection, though the effect sizes are small. Note that we designed our treatments of state coercion to reflect levels of deterrence reported by the IRS in the real world, which yielded rates of intentions to comply anticipated by SEM. Specifically, when state control of tax evasion is weak, SEM predicts that most individuals will engage in tax evasion (Andreoni et al. 1998; Slemrod 2005). This is what we, and the IRS, observe: 83.19 percent of individuals were willing to underreport cash payments to the IRS (to varying degrees), while the IRS claims that misreporting of income subject to little or no information reporting is around 63 percent (IRS 2016). In our experiments—as in real life—just about everyone cheats, if only a little (Ariely 2012). SEM, however, is unable to account for the large subgroup of individuals who were (almost) always willing to comply or for the between-individual variation in tax compliance that we observe.

Moving beyond SEM, we demonstrated that moral attitudes surrounding taxation motivated intentions to comply. More specifically, we found that moral imperatives exerted the greatest influence on willingness to report hypothetical cash payments to the IRS. As expected, the effect of moral imperatives was independent of fellow taxpayer compliance and key parameters of SEM. The implications are twofold.

First, our results suggest that moral imperatives are key determinants of tax compliance. We found that U.S. taxpayers are intrinsically motivated to pay their taxes and will do so for income beyond the purview of the IRS. Taxpayers are willing to comply with their tax obligations because it is the right thing to do. The implication is that the classical sociologists were right (Durkheim 1961, 1965; Goffman 1959; Weber 1930): morality and moral attitudes matter, and they matter more than deterrence. Tax compliance is not merely coercive or quasi-voluntary, it is primarily voluntary, where the roots of voluntary tax compliance stem from the moral attitudes and judgments people attach to taxes and taxation. Paying taxes, at least in the U.S. context, is a fundamentally moral matter that concerns questions of right or wrong, acceptable or unacceptable. Matters that are deeply tied to our sense of self as citizens and as members of society. As Williamson (2017) writes, “Americans understand taxpaying as a responsibility to the community and the country (p. 168).” While most everyone cheats and evades taxes, if only a little, morality and moral motivations might be the spigot that stem the tide of dishonest behavior (Ariely 2012).

Second, an open sociological question concerns the extent to which moral imperatives interact with parameters of SEM. Some social scientists contend that moral imperatives should be treated as one type of incentive among many (Bowles 2016; Opp 1999; Smith and Wilson 2018), whereas others argue that morality limits the set of alternatives available to actors, eliminating tax evasion from the choice set entirely (Etzioni 1988; Elster 1989; Hitlin 2008; Kouchaki et al. 2018). The latter argument suggests that only a subgroup of dishonest taxpayers not bound by principles of fiscal responsibility weigh the costs and benefits of tax evasion (Kroneberg et al. 2010; Wikström 2004, 2006). Our findings provide evidence that moral imperatives complement, rather than sharpen or dull, other incentives (i.e., the effect of moral imperatives is additive rather than multiplicative). Taxpayers consider parameters of SEM regardless of the strength of their moral imperatives, which act as base rates taxpayers deviate from in light of the situation. Taken together, state coercion works but in practice is an inefficient and costly form of state control. Instead, states should invest in campaigns that target the moral foundations of taxpaying. This could be done with signed pledges and moral reminders (Ariely 2012), or with media campaigns underscoring the benefits of paying taxes to community and country (Bott et al. 2017).

Following insights from the sociology of morality (Hitlin 2003, 2008; Stets and Carter 2012) and moral psychology (Haidt and Joseph 2004; Haidt and Graham 2007; Haidt 2012), we proposed that moral alignment was central to tax compliance as well. Our expectation was strong out-group competition and lower tax compliance when the values of taxpayers and political authorities were disjointed, and higher compliance when they were aligned. The findings supported our expectations. We found that the alignment between the values of political authorities and taxpayers increased willingness to report cash payments, and division decreased it. This is a compelling finding that has implications for our understanding of moral identities and the growing partisanship in U.S. politics. It suggests that political polarization not only undermines voter turnout, political trust, and interest in politics (Layman et al. 2006), it may also play a part in one’s decision to evade taxes (Cullen et al. 2018; Robbins and Kiser 2018). As Haidt (2012) argues, the basic need to justify our own actions and defend the groups to which we belong creates a sense of moral righteousness, which can produce moral wrath directed toward out-group members. One way this moral wrath manifests is tax evasion. We find that the magnitude of this effect, however, is small and counterbalanced by an in-group bias of similar magnitude (an absolute value of 1 to 2 percentage points depending on the model). Taken together, our analysis of the moral attitudes surrounding taxation suggests that the average taxpayer is guided by a moral duty to pay some, if not all, of their taxes in spite of the United States becoming more politically divided.

Where do mechanisms of state reciprocity fit into this discussion? Our findings suggest that individuals tend to focus more on control capacities of the state rather than on their own personal attitudes and beliefs about the state, how the state treats taxpayers, or if fellow taxpayers are paying their taxes. By manipulating vignette dimensions, including individual-level measures, and estimating their complex interactions we were able to rule out numerous channels through which state reciprocity influences tax compliance. A set of findings supported by prior research, some of which shows that dimensions of state reciprocity have little influence on people’s willingness to evade or comply with their tax obligations to the state (Ariel 2012; Blumenthal et al. 2001; Castro and Scartascini 2015; De Neve et al. 2019; Dwenger et al. 2016; Fellner et al. 2013; Robbins and Kiser 2018). Despite these results, it is still possible that state reciprocity motivates compliance in certain settings. It could be that mechanisms of state reciprocity are more salient under conditions of institutional upheaval, widespread political malfeasance, or unfair and unjust legal systems. Many countries in Eastern Europe, Africa, and Latin America—where tax evasion is widespread—experience inadequate public goods provision, weak rule of law, and rampant political corruption (Bergman 2009; Bodea and LeBas 2014; Castañeda et al. 2020; Kiser and Sacks 2011; Kogler et al. 2013; Levi et al. 2009). From a design standpoint, we chose to manipulate features of state reciprocity that would appear plausible and realistic to respondents in the U.S. context (Auspurg and Hinz 2015), so we did not want to set levels of corruption or public good provision to those typically seen in Africa or Eastern Europe. But given the null results, it would be beneficial for future research to investigate manipulations of state reciprocity that are found in developing nations or countries in economic and political transition.

Relatedly, it is unclear whether the dramatic short-term changes to state institutions or the long-term effects of living under (un)trustworthy political regimes affects taxpayer compliance. Steinmo (2018) and many others (Alesina and Giuliano 2015; Bergman 2009; Putnam 1993; Steinmo 1993) argue that state institutions and attitudes toward the state co-evolve in their influence on tax compliance. Consider the case of Sweden. For decades, Swedish state policies have generated stronger positive attitudes toward the state than comparable European countries (Svallfors 2011). Swedes, as a result, are highly intrinsically motivated to pay taxes and are equally concerned with the tax payments of others (Steinmo 1993). This would imply that the state and society affect tax compliance, but not via short-term changes to political institutions that motivate reciprocation with the state as some models of state reciprocity would contend (Levi 1988, 1997; Tyler 1990). Instead, it might be long-term state-society interactions and mechanisms of cultural transmission that account for the moral imperatives surrounding tax compliance. While the positive intercorrelations between moral imperatives and attitudes toward the state support this claim (see Table A1), our call is for fiscal sociology to conduct more comparative research—both experimental (Pampel et al. 2019) and historical (Scheve and Stasavage 2016; Steinmo 2018)—that identifies and traces the origins of tax compliance.

Limitations and Directions for Future Research

Like all research, our study has important limitations. First, the moral imperatives scale we used consisted of three closed-ended questions embedded in a survey. Although our outcome variable was elicited in an experimental context, the relation between moral imperatives and intentions to comply is correlational and not causal. Future research should thus prioritize experimentally manipulating moral imperatives. It is also important to determine how moral imperatives affect intentions to comply because of our measurement strategy. As Miles et al. (2019) show, closed-ended survey items—like the ones used in the present manuscript—are strongly tied to both automatic and deliberate cognition. A compelling stream of research might investigate whether and to what extent moral attitudes surrounding taxation operate automatically and/or deliberatively: How do people deliberate over the costs and benefits of tax evasion in relation to their strongly held moral principles? This could be done by placing subjects under cognitive load and by treating them with subliminal primes that operationalize moral imperatives (Miles 2015).

Second, we are unable to empirically identify whether the effect of moral alignment on intentions to comply is due to expressive or instrumental motivations (Green et al. 2004; Huddy 2001). Is the effect a function of proximity to one’s preferred policies (i.e., instrumental) or moral affiliations with a party (i.e., expressive)? The exact nature of the ingroup favoritism and outgroup animus that we observe is a black box. While prior research shows that political polarization has its roots in partisan identity and partisan affect (i.e., expressive motivations; Iyengar and Westwood 2015), future work that probes the expressive and instrumental foundations of moral alignment is necessary for the field to advance.

Conclusion

Why people pay taxes is one of the oldest questions in the social sciences. Mechanisms of state coercion and state reciprocity have been the predominant answers. With this study, we set out to offer a third explanation of tax compliance derived from the new sociology of morality and moral psychology: moral attitudes. To provide insight into the motivations behind tax compliance, we designed a factorial survey experiment of income tax evasion and administered it to a nationally representative random sample of U.S. adults. We discovered strong support for state coercion and moral attitudes. With this work, our hope is to help sociologists better understand the moral foundations of fiscal capacity, and to lay the groundwork for research that focuses not only on the control capacity of a state but also on the moral capacity of a society.

APPENDIX

Table A1.

Correlations among Individual-Level Variables

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
(1) Trustworthiness of government
(2) Procedural fairness and justice	.57**
(3) Subjective probability of being audited	−.04	.01
(4) Prior tax fraud	−.03	−.03	−.05
(5) Prior audit	−.04	−.04	.11**	.07*
(6) Prior punishment	−.09**	−.11**	.08**	.17**	.26**
(7) Moral imperatives	.33**	.32**	.02	−.15**	−.02	−.13**
(8) Party identification	.09**	.08**	−.03	−.02	.06*	.00	.00
(9) Identification with the state	.23**	.21**	−.01	−02	.09**	−.03	.26**	.30**
(10) Roles of the state	.00	.05	.07**	−.02	−09**	−.01	.07*	−.44**	−.23**

^**p < 0.01,

^*p < 0.05 (two-tailed)

Figure A1.

Income tax evasion vignette coversheet

Figure A2.

Example income tax evasion vignette

Figure A3.

Income tax evasion vignette dimensions and levels