Dubious versus trustworthy faces: what difference does it make for tax compliance?

ABSTRACT

We find experimental evidence that the decision problem of tax compliance changes if subjects’ declarations are not randomly assessed, but is based on their appearance as captured by pictures of their faces, even if the aggregate audit probability does not change. Some subjects may fear that their picture looks rather dubious, whereas others may believe that their picture looks more trustworthy than average. Depending on these beliefs, they may adjust their compliance decisions. Our experimental design allows us to disentangle these potentially countervailing effects.

KEYWORDS:

• Compliance
• face picture
• audit probability
• tax evasion

JEL CLASSIFICATION:

• H26
• C91
• H31

Acknowledgements

We thank Kai A. Konrad for valuable suggestions, some contributions in writing a former draft and a careful discussion of the experimental results. We also thank Miguel A. Fonseca, Volker Meier and Hans-Werner Sinn as well as participants of the APET meeting 2013 in Lisbon, of the 2014 conference on ‘Taxation, Social Norms and Compliance’ held by the Center for Economic Behavior and Institution Design in Nuremberg/Germany and of the CESifo Group Seminar in Munich in 2014 for helpful comments. For providing laboratory resources, we kindly thank MELESSA of the University of Munich. Thanks also to Hans Müller for developing and programming the web-based environment. Tim Lohse is grateful for hospitality at the Center for Economic Studies (CES) when finalizing this paper. The usual caveat applies.

Notes

¹ A recent exception is by Van Leeuwen et al. (2014). In an ultimatum game, they analyse whether facial cues provide a credible signal of destructive behaviour.

² Coricelli et al. (2010) use images of faces in a compliance experiment, as well, but with a rather different focus: a subject’s cheating behaviour is revealed by publicly displaying his picture in the laboratory. The risk of public exposure of deception deters evasion significantly.

³ This economic problem borrows from the compliance set-up in Konrad, Lohse, and Qari (2012). The tax declaration decision was framed as a customs decision. Subjects played travellers whose endowments were either above or below a tax-exempt amount.

⁴ In case of an uneven number of under-reporters where one under-reporter represents a ‘tie’, the tie is randomly selected for an audit with a 50% chance by the computer. Besides the audit mechanism used here – audit of 50% of all high-income individuals who declared low income – an alternative mechanism could be an audit of 50% of all individuals declaring a low income, regardless of their true income. However, from the point of view of a potential cheater, both mechanisms work in exactly the same way. The selection of a particular mechanism is therefore solely a framing issue. We believe the audit mechanism used here is more useful for an experimental study as it is easiest to understand.

⁵ Already when registering for the experiment, the subjects were informed about them being photographed. After having read the instructions and before starting the experiment, subjects were asked to sign a declaration of consent about the use of their face picture for scientific purposes. Subjects withholding approval would have been paid the show-up fee and asked to leave. Of course, subjects had the possibility to leave at any time in the course of the experiment and revoke their declaration of consent but then no show-up fee would have been paid. However, all subjects signed the declarations and no one quit.

⁶ We assume all individuals to be risk neutral. Rabin (2000) shows that, within the expected-utility framework, anything but risk neutrality over modest stakes would imply rather unrealistic risk aversion over large stakes. In our empirical analysis we generate a risk measure by using data from a standard risk elicitation game in the style of Holt and Laury (2002) which participating subjects had to play. This risk measure has only little explanatory power in our data (see Section III for details).

⁷ Indifference condition: 800 = p(1000 – 200 – 100) + (1 – p)1000 <=> p = 2/3.

⁸ Indifference condition: 800 = p(1000 – 200 – 300) + (1 – p)1000 <=> p = 2/5.

⁹ The participants were recruited using the software ORSEE (Greiner 2004).

¹⁰ A further incentivation of the officers is not necessary since the research focus is entirely on the declaring individuals. Besides, recall that each officer potentially sees the face picture of a subject just once.

¹¹ The predictions are calculated by inserting the coefficient of the constant for Base and the coefficient of the constant plus the coefficient of Pic into the logistic cdf, respectively.

¹² versus : ,

¹³ versus : ,

¹⁴ Sample size may also be an issue. As the expected effect size is unknown a priori, it is unclear which sample size is needed to carry out formal hypothesis tests. For the discussion on the (ab)use of applying uniformly a 5% level and related problems, see, e.g. McCloskey and Ziliak (1996) and Krämer (2011).

¹⁵ This weak effect is in line with an analysis of the observed correlation between the subjective probability assessment and individuals’ reporting decisions, which is small and insignificant.