The effect of priming on fraud: Evidence from a natural field experiment

Abstract We present a natural field experiment to examine if priming can influence behavior in a market for credence goods. 40 testers took 600 taxi journeys in Vienna, Austria, and using a between–subject design we vary the script they spoke, each designed to prime either honesty, dishonesty, or a competitor. We find that the honesty prime increases taxi fares by 5.5% relative to a baseline, the result of overcharging rather than overtreatment. Priming dishonesty and a competitor have no impact on fares. We find that the effects of priming on behavior are likely to be small compared to information asymmetries.


A2 Vienna taxi tariffs
The Viennese taxi fares are taken from https://www.wko.at/branchen/w/transport-verkehr/ befoerderungsgewerbe-personenkraftwagen/Wiener_Taxitarif_1997_2.pdf. In addition to the local authority provides the following additional information about the taxi tariff system, in particular where possible additional surcharges may be added to the fare: -1 surcharge if the taxi was ordered via a "taxi-rank-phone" (Standplatztelefon); -2 surcharges if the taxi was ordered via call centre (Taxifunkzentrale); e 2.00 surcharge for the carriage of 4 or more passengers; e 13.00 surcharge if the destination is the airport and the taxi was not pre-ordered; -Further surcharges e.g. for loading the luggage or for going to the station are prohibited.
-The driver has to inform the passenger about any surcharges added.
Charging anything different than mentioned is an administrative offence.  Notes: Note that we clustered certain countries/indications of driver's ethnicity to these categories (e.g. "African" and "Arabian" are clustered to the ethnicity "African"). However, where ever possible (i.e. as controls), we use the full information about the estimated driver's origin.  Table A5 summarizes the nationalities of the drivers in our experiment.

A4 Analysis using additional baselines
In this section we present the results relating to the OSM and Uber baseline, excluded from Section 5.

A4.1 OSM Baselines
This section presents results for Section 5, using OSM as a baseline. Notes: Robust standard errors in parentheses, clustered by quadruple. Models (1)-(5) are Tobit regressions censored at 1. The presented explanatory variables are dummy variables that take a value of 1 if the observation is taken from that treatment (and 0 otherwise). ***, ** and * denote significance at the 1%, 5% and 10% level.

A4.2 Uber Baselines
This section presents results for Section 5, using the Uber price as a baseline.

A4.3 Google Baselines
To examine this, and determine the level of overcharging relative to our perfectly informed baselines, we compare the observe fare obtained using prime p in quadruple i, f p,i to the fare associated with the journey planned by each of our informed baselines, k, in quadruple i, f k,i , where k ∈ {Google M aps, OSM, U ber}. The overcharging difference is defined as Table A9 presents the overcharging differences for each treatment, using each of the three baselines. As can be seen, there is overcharging in all treatments, and this is significantly different to zero (p < 0.01). Regardless of baseline, similar patterns emerge in the data, although there are no significant differences between treatments (p > 0.1 in all cases). Importantly, this implies that our treatments, when compared to a perfectly informed passenger who is never overcharged, all produce significant levels of overcharging.
We examine treatment differences further using Tobit regressions, for brevity presenting only those using the Google Maps baseline, and the same sets of controls as those previously defined. As can be seen in Table A10, the coefficients on the treatment dummies are Notes: Baseline treatment is taken as the baseline. Robust standard errors in parentheses, clustered by quadruple. Models (1)-(5) are Tobit regressions censored at 0. The presented explanatory variables are dummy variables that take a value of 1 if the observation is taken from that treatment (and 0 otherwise). ***, ** and * denote significance at the 1%, 5% and 10% level. all not significant at the 5% level. However, as the constant is positive and significant this supports our assertions that there is a significant level of overcharging relative to the informed passenger baseline, but no differences between treatments. The estimates from Table A10 suggest that, although there is a significant amount of overcharging relative to a perfectly informed passenger in all treatments, there are no treatment differences when comparing the effects of the primes to the fully informed baselines. This observation suggests that the effect of asymmetry in knowledge on overcharging is greater than the effects of the primes; the vast majority of the overcharging we observe relative to the case where there is no overcharging is a result of our testers being 'foreign', and being viewed as having limited knowledge. Notes: Baseline treatment is taken as the baseline. Robust standard errors in parentheses, clustered by quadruple. The number of observations falls as the number of controls are increased due to missing entries. Models (1)-(5) are Tobit regressions censored at 0. The presented explanatory variables are dummy variables that take a value of 1 if the observation is taken from that treatment (and 0 otherwise). ***, ** and * denote significance at the 1%, 5% and 10% level.  Notes: Baseline treatment is taken as the baseline. Robust standard errors in parentheses, clustered by quadruple. Models (1)-(5) are Tobit regressions censored at 0. The presented explanatory variables are dummy variables that take a value of 1 if the observation is taken from that treatment (and 0 otherwise). ***, ** and * denote significance at the 1%, 5% and 10% level.