UNIVERSITY OF AMSTERDAM
FACULTY OF ECONOMICS AND BUSINESS
Kristína Sacherová
Student number: 11376139How do people lie and can the other person spot
deception: Experimental evidence from a
buyer-seller game.
Master thesis
ECTS: 15
Author: Kristína Sacherová
Supervisor: dr. Jeroen van de Ven
Study programme: Business Economics: Managerial Economics and Strategy
Academic Year: 2016/2017
Statement of Originality
I hereby declare that I, Kristína Sacherová with UvA student number 11376139, wrote this thesis independently under the leadership of my supervisor and that the references include all resources and literature I have used.
I declare to take full responsibility for the contents of this document. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
I hereby proclaim that the thesis has not been used to obtain a different or the same degree. I grant a permission to reproduce and to distribute copies of this thesis document in whole or in part.
Amsterdam, August 15, 2017
Abstract
The detection of misleading information belongs to one of the most critical challenges that online shoppers face. In this paper, I present experimental evidence on the existence of linguistic cues to deception in informal written conversations as well as on the ability of people to spot when being lied to. Subjects in pairs of two play a variation of buyer-seller game. Sellers have informational advantage and are incentivized to lie in some cases. The height of these incentives is manipulated. Buyers decide whether to follow seller’s recommendation. Communication takes place via chat session. The analysis of recorded messages identifies systematic distinctions in the language of deceivers related mostly to the word count, pronouns, the emphasis of trustworthiness and the use of context-specific phrases. Buyers show the above-chance ability to recognize false advices perceiving informal language, lengthier texts and the increased referring to money or promises as cues to honesty.
Keywords
Deception Lie detection Informal written communication Linguistic analysis Actual cues to deception Perceived cues to deception Experiment Buyer-seller game Asymmetric information
- 1 -
1. Introduction
„On the internet, nobody knows you’re a dog”, says a cartoon drawn by Steiner (1993)
printed in The New Yorker magazine. It has already been twenty-four years since its publication, but concerns about veracity of online content are becoming even more pressing in recent years. This is due to the extensive digitalization of current era. According to World Economic Forum reports, the number of online users has already exceeded 3 billion1 and there will be more than 20 billion gadgets connected to the Internet in 20202. The continued innovation has been shaping and changing everyday aspects of people’s lives, their manners and habits. One of the most affected areas is shopping. Nowadays, more than 50% of shoppers buy products online on weekly or monthly basis, as a survey conducted by PwC in 2016 revealed3. Hand-in-hand with transition of businesses from retail to online world, Internet advertisement is booming and has already outpaced more traditional ways of promoting such as TV commercials or campaigns in press4. Considering all these facts, Internet may be justly deemed as the most important mass communication channel. However, the anonymity, which this easily accessible medium provides, brings with it numerous challenges. The detection of misleading and false information definitely belongs to the most critical ones that shoppers making online purchases need to deal with.
Indeed, many studies reported people´s vulnerability to be easily deceived via Internet (see e.g. Burgoon et al., 2004). Due to decreased possibility to validate truthfulness of online information, people cannot solely use advertised messages as guidelines for decision-making when shopping online. As de Haan et al. (2011) discusses in their paper, not all features of products can be realistically assessed prior to the trade realisation and actual delivery whilst the most discussable and distorted aspect is quality. This provides scope for abusive practices of online sellers, particularly when there is a possibility to earn more from transaction if lie or disinformation is not detected (Chen et al., 2013). Therefore, trust is an important feature of interactions via Internet. But can buyers determine which information should be considered credible and whom to trust? And is it possible at all? Are there some systematic cues in the behaviour of sellers associated with deceptive practices?
1 The approximate figure for 2015 was taken from WEF Digital Media and Society Report (2016). Online available at http://www3.weforum.org/docs/WEFUSA_DigitalMediaAndSociety_Report2016.pdf.
2 The estimate of 20 billion online devices based on WEF Global IT Report (2015). Online available at
http://www3.weforum.org/docs/WEF_Global_IT_Report_2015.pdf.
3 The survey took place in 25 countries on the sample of around 23 thousand shoppers. For more results and information about study, see https://www.pwc.com/gx/en/industries/retail-consumer/global-total-retail.html. 4
The information based on findings of Global Entertainment & Media Outlook 2017-2021 issued by PwC in 2017. Online available at http://www.pwc.com/gx/en/industries/entertainment-media/outlook.html.
- 2 -
In spite of the importance and potential applicability of the ability to spot deceivers, economic literature does not provide sufficient answers to these questions. To look for some new insights on this issue, a laboratory experiment was designed to simulate trading via Internet. Participants in the role of sellers were incentivized to sell a product as the high quality one sometimes even at the price of giving false advice to buyers. On the other side, it was in the best interest of subjects representing buyers to find out the actual state of the product on sale. The only available information, on which they were possibly able to base their decision, was the seller’s recommendation and two-minute informal free-format chat. Although simplified, this design mimics very economically meaningful and up-to-date form of communication in online environment. Buy-sell groups on social sites where second-hand items are traded are a relevant example of place where similar interactions take place. To my knowledge, there are no deception-related papers in economics in this context. Besides, the study has additional strengths and advantages over existing literature, such as no instructions affecting the decision whether and when to lie or the possibility to distinguish between different groups of liars. The investigation focused on the content of recorded conversations. Texts were analysed from two perspectives. Firstly, I looked for linguistic particularities in messages of honest and deceiving sellers. Secondly, I examined the impact of various texting styles and language features on the buyer’s choice to follow corresponding advice. The analysis revealed systematic distinctions in the language of deceivers related mostly to the word count, pronouns, the emphasis of trustworthiness and the use of context-specific phrases. Furthermore, I identified some differences in the vocabulary of occasional and more frequent liars. In the experiment, buyers showed the ability to recognize cases when they are being deceived. Subjects perceived informal language, lengthier messages, the increased referring to money or mentioning of the word “promise” as significant cues to honesty.
The rest of the thesis is organized in the following way. Section 2 contains a critical review of related literature. In Section 3, I talk about the way how the experiment was designed and conducted; then, I define the research question and hypothesize about potential answers. In Section 4, I discuss my approach to data analysis and present obtained results. Section 5 summarizes outcomes of the analysis and concludes. At the end of the paper, a set of Appendices with additional information, figures, statistical and regression tables is enclosed.
- 3 -
2. Literature Review
Deception is an important topic from both, social and economic, perspectives (Mazar and Ariely, 2006). For a person to become a good “lie detector”, one would need to appropriately combine knowledge from various disciplines, such as linguistics, psychology, criminology, anthropology, neurology or economics, to know which verbal and non-verbal cues to seek for (Navarro, 2012). This section provides an overview of relevant available literature presenting methods, findings, strengths as well as shortcomings of these studies.To conclude, I highlight strong points, novelties and overall contribution of this experimental investigation.
2.1. Psychology research on deception
In psychology, lying and its detection have been frequently studied topics for many decades (for an overview of earlier studies, see Hyman, 1989). On the contrary, economics at that time was not concerned about deception, since standard economic models rely on an assumption that lying occurs whenever potential benefits from deceiving exceed risks that arise if it fails (Lewicki, 1984). Nevertheless, this belief has proved to be unrealistic and deceptive behaviour has become a relevant subject of interest for economists, as well.
Although both economic and psychology studies investigate the same phenomenon, there are several notable differences in their practices resulting in diverse outcomes. Compared to economics, psychology approaches research qualitatively focusing more on cues like body and facial movements, postures or non-verbal expressions (see, for instance, Ekman, 2009). Furthermore, settings in which psychology studies take place are oftentimes of low economic importance. Also, since deceptive behaviour is extensively studied in lab, a distinct opinion of psychology and economics on whether and how to motivate participants of experiments makes comparability and applicability of results difficult (Hertwig et al., 2001). Based on a review of literature by DePaulo et al. (2003), incentives make lie more detectable. Cues to deception are more evident if some monetary, material or identity-related reward is provided, whilst the authors suggest the last one to be the most influential stimulus. As proposed by Zuckerman et al. (1981), the reason why these forms of exogenous motivation have such an effect is that liars’ rhetoric is less spontaneous because of excessive controlling of behaviour. In general, most psychology-based research coincide in a finding that people cannot systematically distinguish truths from lies and just sparsely perform significantly better than chance (Vrij, 2000).
- 4 -
2.2. Methods used for examination of deception and its detection
2.2.1. Monologue vs. dialogue and observational vs. experimental data
As a study by Feldman et al. (2002) revealed during a ten-minute dialogue people lie two to three times. However, much of published deception-related work follows from analysis of individual statements, testimonies or letters (e.g. Chen and Houser, 2013). Although this fraction of research brings valuable insights into debate, due to a lack of interaction in monologue-based samples, outcomes from such settings have limited relevance in everyday situations. Hence, conversational format is preferred by many researchers (e.g. Belot and van de Ven, 2016). Unfortunately, due to privacy concerns and scarcity of observational data most researchers have to conduct experiments to collect communication data for their analyses5. In a majority of dialogue-based experimental studies, subjects interact simply in pairs of two. Prior to start of an experimental session, each person is usually assigned a fixed role. The most common is a sender-receiver or seller-buyer paradigm where one party is better informed and typically instructed or incentivized to lie.
On-going communication in these games is commonly referred to as cheap talk. In spite of inability to verify the transmitted information, it has been shown that cheap talk is actually convincing (e.g. Loewenstein et al., 2011 or Coffman and Niehaus, 2015). Also, de Haan et al. (2011) observed in their experimental investigation of advertising that cheap talk carries more information than expected by standard economic theory. Cheap talk-related articles present important findings relevant also for examining of deception but some crucial features are not included there. For instance, most such studies do not deal with accurate measuring of buyer’s transaction valuation, thus lack data about the truthfulness of information.
2.2.2. Types of communication and underlying credibility
Burgoon et al. (2003) explored a relationship between the level of involvement and interactivity of communication and its reliability and subsequent decision regarding trust. Texting is assumed to be the least interactive followed by audio and audio-visual settings, whilst face-to-face communication ranks the highest with regards to the degree of interactivity and involvement of participants. For treatment without deception, examined aspects moved in
5
There exist some other ways how to get observational data. For example, some studies use transcripts of existing dialogues from TV shows or interviews (see e.g. Belot et al., 2010).
- 5 -
the same direction. When the possibility to deceive was introduced, people were no longer good at detecting deceits.
On the contrary, Belot and van de Ven (2016) found above-chance ability of receivers to predict credibility of seller’s advice in a face-to-face buyer-seller setting. The authors also studied whether contextual richness and opportunity to ask questions influence cooperation and deception detection rate. Their data showed no support for the latter. With more complex contexts, the probability that buyer followed seller’s recommendation was slightly higher; however, the effect was not statistically significant. It was argued that even though there is more space to cover a disinformation when selling a more sophisticated product, complex situations also require more cognitive thinking and concentration which may result in less convincing and inconsistent recommendations.
Despite the fact that content of messages is easier to control than body language, voice tone or facial expressions, there are still numerous indications of deceit present in written conversations (Meyer, 2010). Actually, there are several papers providing evidence that people are to a certain degree able to determine credibility of written statements. As, for instance, Utikal (2013) showed that participants could distinguish whether harm caused by their companion was intended or happened just by accident based on a letter of apology.
2.3. Actual and perceived linguistic cues associated with deception 2.3.1. The language of liars
Words can reveal a lot about the author and his or her state of mind (Pennebaker et al., 2003). As a consequence, specific phrases and language patterns may signal dishonesty of a person since deceiving is naturally accompanied with emotional and psychological changes. A growing body of contemporary research concerned with lying have turned its focus on particularities in content and form of truthful and deceptive communication. Most of these linguistic analyses pinpoint the following dimensions of communication as being indicative whether somebody is lying: i, quantity in terms of number of syllables, words and sentences; ii, self-referencing and use of other-person pronouns; iii, expressiveness and emotiveness; and iv, overall cognitive complexity.
Hancock et al. (2004) analysed more than 200 instant-message conversations where participants in pairs of two were given a certain topic (e.g. “Talk about a mistake you made recently”) to chat about without any time constraints. Senders were told to occasionally lie
- 6 -
and given few minutes to prepare. Receivers did not know about these instructions and were informed only about the topic of upcoming chat. The results showed that deceivers wrote more than truth-tellers having their stories also richer from perspective of sense words and other-person pronouns. Furthermore, deceptive messages contained less cognitive words. Burgoon et al. (2003) reported evidence from a cooperation game and a mock theft experiment using various communication channels (email, audio- and videoconferencing, face-to-face interaction). Their results indicate that there were systematic distinctions in language style of authentic and fake testimonies. In contrast to Hancock et al. (2004), the researchers found support for less talkative and less emotive liars. Inconsistent findings regarding these two indicators across studies may be explained by strong context-dependency of these dimensions and differences in experimental design (e.g. monologue versus dialogue formats). On the contrary, the conclusions regarding cognitive complexity are similar.
Newmann et al. (2003) were interested in automated lie detection. Data, which they analysed, comes from a mock theft experiment and four monologue-based laboratory experiments where students were asked to present their opinions either in written or spoken format. Based on these samples, multivariate profiles of deceit were developed using Linguistic Inquiry and Word Count (LIWC) software. These were then used to determine truthfulness of the same statements. The test resulted in accurate prediction, on average, in two out of three times. Considering individual outcomes of these five studies, the authors proposed the following three linguistic features to be decent predictors of lies: i, low incidence of first-person and third-person pronouns; ii, more frequent usage of phrases expressing negative emotions; iii, simpler syntax and fewer cognitive words. Negative association of self-referencing and deception, the finding consistent with the vast majority of literature, is usually reasoned by the willingness of deceivers to isolate themselves from lies (for a review, see e.g. Vrij, 2000). Similar argument is commonly used when hypothesising about increased use of other-focused pronouns (Ickes, 1986 or DePaulo et al., 2003), though this study showed the opposite evidence. Yet, the findings concerning referencing to others are less straightforward.
A set-up of written-format experiment by Chen and Houser (2013) consisted of two parts. The first one was based on a mistress game with allowed communication in form of secret letters, in the second stage third-party evaluators assessed whether the content of these letters can be viewed as a promise or just empty words. This classification was then compared with real actions of mistresses during the game. Unlike previous mentioned studies, the authors did not specifically instruct participants to lie; rather they designed the setting to naturally motivate
- 7 -
people to deceive under some circumstances. The analysis revealed that longer letters are more likely to be perceived as trustworthy promises, which is in line with suggestions of earlier research (e.g. Wood et al., 1985). The occurrence of words related to monetary rewards was another thing viewed by evaluators as a cue to commitment. However, at the same time this perceived cue to honesty was the most prevalent marker of senders’ disloyalty.
2.3.2. Responses from receivers
While it may seem sufficient to look only at messages by potential liar, Niederhoffer and Pennebaker (2002) suggested that analysing reactions of other involved parties may make illustration of interesting features of truthful and deceptive communication clearer and more complete. In their paper, they pointed out that people mirror some communication patterns used by their counterpart. This phenomenon is also known as linguistic style matching. For instance, the study by Hancock et al. (2004), where also messages of receivers were analysed, presents convincing evidence in favour of linguistic style matching. In addition, the following patterns in communication by receivers when being confronted by deceiver were identified: i. messages were longer but sentences, on average, consisted of fewer words; ii. language contained more sense-based descriptions; and iii. conversations were all-in-all longer also because of more questions raised by receiver. By definition, monologue-based methods lack this part of the story specific for more interactive frameworks.
2.4. Contribution of this study
To conclude the section, I turn attention back to this experimental investigation and discuss how it complements existing literature. The main novelty is that lying behaviour and the ability to detect lies are examined in an economically-relevant situation where communication goes via chat session. To my knowledge, there are no such economic studies on deception combining buyer-seller paradigm, free-format written chat method and informal dialogue. Another strong point is an incentive scheme which provides participants with stark motivation to act in their own interest and, hence, the decision whether and when to lie is voluntary rather than artificially instructed. Moreover, the variability in height and order of incentives allows for identification of more frequent and just occasional liars and linguistic dimensions of their texting style can be evaluated separately and compared. Last but not least, the sample contains 1269 transcripts of casual two-minute conversations which I consider as a reasonable evidence for answering the questions of interest.
- 8 -
3. Methodology
To shed more light on how people lie, whether the ones that receive information are able to distinguish true from false messages and how they respond when being deceived, a “Lemon Game” experiment was conducted. During 14 lab sessions, online buyer-seller relationship with asymmetric information on sellers’ side was studied. Altogether, 1269 chat-based conversations in English6 were recorded and processed for further analysis.
3.1. Experimental design
Subjects were divided into buyers and sellers. Prior to the start of every round, each seller was matched with a buyer. On computer screens of all participants, a self-taken photo of their counterpart was displayed. Afterwards, the game started while its timing looked as follows. At first, seller privately and randomly drew a card which could be either green or red, both options with the probability of 0.5. The colour of the card determined the condition of a product which was on sale; green represented flawless product whereas the red card implied that the item was in bad condition. The draw was followed by seller’s claim about the condition of the product. Buyer saw only this recommendation, never the card itself. Buyer and seller were then given two minutes to communicate without any further restrictions on topic or the length of messages7. The communication went on in a chat window where conversation history, the snapshot and remaining time were visible. When the time was up, there was a turn for buyer to decide and submit in private which product to “buy”. At that moment, the game ended and there was no further feedback to any party about the outcome of this particular interaction.
Each lab session consisted of two parts during which the above-described game was played. The game was repeated 6 to 8 times per part depending on the number of subjects who participated in given session. In every round, participants were reassigned with a new partner. The game was designed to simulate actual business interactions in which a trade is being realised. In such economically meaningful situations, sellers are motivated to convince buyers that the product is good even if it has defects and sell it for the highest price possible. On the other hand, it is in the buyers’ best interest to find out the actual state of the product. In this
6
The original dataset had a total size of 1680 observations; however, due to language limitations of the author and LIWC sofware 411 conversations in Dutch and Chinese were excluded from the sample.
7
- 9 -
set-up, three different levels of incentives for sellers were used; particularly high, low and very low. For buyers, the right guess of the product’s condition always meant the payoff of 30€ while the purchase of the product with defects was worth nothing. All three possible payoff matrices are presented in Table 1 below (note that payoffs are in euros). Regardless their role, participants were informed about relevant incentive structure in given part and potential payoffs of their counterpart.
For all payoff structures, the game has multiple equilibria (for deeper analysis, explanation and proofs see Belot and van de Ven, 2016, Section 3.1 and Appendix 2).
Table 1 - The payoff matrices by incentive levels for seller
In addition to the variation in the height of incentives for sellers, their order was manipulated, too. The following 4 treatments were introduced to study the effect of changes of incentives on lying propensity8: i, low – high; ii, high – low; iii, low – low; and iv, very low9 – high.
3.2. Procedures
The fully-computerized experimental sessions were organized at the Centre for Research in Experimental Economics and Political Decision Making (CREED) laboratory at University of Amsterdam in 2016, with a total of 216 participants recruited from CREED student database. Preceding the game, all subjects had to authorize the use of their snapshot taken in the lab by signing a consent form and confirm that they have read and understood the instructions (for the full text of the instructions, see Ledeboer, 2013). In addition after completing both parts, several survey questions asking about participant’s game plan and his or her opinions of
8 The first word indicates the payoff structure used in part 1, the post-dash one represents the height of incentives in part 2.
9 The last treatment, i.e. very low – high, was added to check whether low incentives are sufficiently weak compared to high ones. According to Ledeboer (2016), the incentive structures were designed properly.
- 10 -
actions of others had to be answered. Afterwards, each participant collected in private his or her money earned. The payment consisted of 5-euro show-up fee and earnings determined by the performance in randomly chosen round of the game.
3.3. Research questions and hypotheses
Devices used to detect deceit are still far from being infallible and unbeatable. Many mechanisms and contexts in which lying occurs, as for instance texting - a field of interest in this study, have been almost untouched and further examination is needed. Thus, I have defined the following research questions to be answered by the experimental evidence:
Research questions: How do people deceive in the informal chat-based setting? Is receiver of
messages able to spot deception and distinguish truthful information from lies?
Based on already-published findings of economic research (for more detailed information, see Section 2), I hypothesise as follow:
Hypothesis #1: The language of deceivers differs from the one used by honest people. Hypothesis #2: Buyers can accurately perceive cues of trustworthiness and identify lies.
My analysis deals with the occurrence of actual and perceived cues and their relation to deception, but also with their aggregate potential to uncover lies in this specific framework.
For purposes of investigation of deceptive behaviour, written-format communication has several advantages over other more interactive settings such as audio calls or face-to-face interactions. Most non-verbal cues like mimics or body postures do not play any role in this environment, so researchers are left with fewer indicators to explore. Moreover, words, which are believed to have the highest importance out of the remaining cues, are more easily recordable, traceable and quantifiable than other signals, which simplifies context even more. Indeed, as Schaffer (2016) writes, “words can, and do, reveal deception”. In coming paragraphs, I develop a set of auxiliary propositions breaking down the problem into more specific linguistic issues to be eventually able to answer the research questions raised.
3.3.1. Word count
According to Hancock et al. (2004) who found evidence in favour of the increased word count in messages of liars, I expect sellers who lie to be more talkative. I reckon it based on the
- 11 -
similarity of setups of these studies (i.e. transcripts of online chats). However, the evidence for this feature is inconsistent (recall the conclusions of Burgoon et al., 2003).
Hancock et al. (2004) proposed various reasons for longer false messages by liars. Besides the senders’ willingness to make up more credible story, the authors observed that receivers in deceptive conversations ask for more details and additional information which may lead not only to continuing in the conversation for a longer time but also to the increased word count of individual responses. One could argue that due to two-minute time restriction this proposition is not very relevant for this case. However, considering that the communication is free-format and there is no topic explicitly given to talk about, I believe that the number of questions can positively affect the length of dialogue and alike to previous studies I suppose that buyers use more questions when their counterpart lies.
As reported by Wood et al. (1985) or Chen et al. (2013), senders who devoted more time and effort are more likely to be viewed as being honest. In line with these findings, I also suspect the higher number of words used by seller to be considered as a cue to trustworthiness.
3.3.2. Pronoun usage
Another important linguistic dimension of interest is the pronoun usage. On the basis of published studies, I expect that sellers who lie refer to themselves less frequently than honest sellers and, in reverse, use other-referencing to a greater extent. The first pronoun-related proposition is compatible with the outcomes of the majority of literature. Generally, psychologists and economics argue that people attempt to dissociate themselves from their lies (e.g. Vrij, 2000; Newmann et al., 2003; Hancock et al., 2004). This is often accompanied by aiming attention on others which is in line with the latter proposition. Although there is no consensus yet on the regularity in use of second and third-person and impersonal pronouns by deceivers, I again hypothesize similarly to Hancock et al. (2004).
3.3.3. Structure and vocabulary
The following group of propositions is oriented on the structure, complexity, content, vocabulary of messages and the occurrence of specific context-related words and phrases. Regarding the association between deceiving and neutrality (or if examining it from the opposite perspective - expressiveness and emotiveness), there is a disagreement among
- 12 -
researchers. Considering this set-up, I favour the conclusions of the review by DePaulo et al. (2003) or the experimental evidence of Hancock et al. (2004) and suppose more emotions and sense-based words to appear in false communication as a potential result of mannerism. To invent a convincing tale is a cognitively demanding task (e.g. DePaulo et al., 2003 or Vrij et al., 2008). The fact that lies are less cognitively complex is broadly documented in the literature which suggests that, for instance, the lower rate of exclusive words (Hancock et al., 2004) or the higher occurrence of shorter and simpler sentences (Burgoon et al., 2003) tend to be characteristic for liars. I anticipate the same patterns to exist in this experimental sample. Research in psychology and criminology has been also interested in how people who tell actual and false information differ when referring to truthfulness in their messages. Similarly to Meyer (2010), I assume that liars have a tendency to overuse words such as “sure”, “true”, “honesty”, “promise” or “trust”.
Chen and Houser (2013) analysed the occurrence and the effects of mentioning of money in their study of promises. Considering parallels of theirs and this setting (mainly the form of incentives provided), I propose that the use of money-related phrases is higher in deceptive communication, but at the same time is mistakenly perceived by buyers as a cue to honesty. Also, the next premise is devoted to the buyers’ perception on the content and style of messages. According to Cialdini (1993), if you are likeable, your odds of being perceived trustworthy are considerably higher. Analogously, I presume more friendly and personal-oriented messages, which for instance include interjections like “haha”, to be more trusted. Last but not least with regard to vocabulary, I analyse the presence of words and phrases specific to this experimental design and context, namely “green (card)”, “red (card)”, “good condition” or “bad condition”. Since there has been no such linguistic analysis conducted in a similar setting, I do not list any propositions backed by previous research.
3.3.4. Complementary investigation
In addition to above-mentioned propositions, I also test all the other combinations of specified linguistic dimensions and their relation to truthfulness and perceived trustworthiness to see whether there are any further important patterns which are, however, not yet reported in the literature.
- 13 -
4. Data analysis and results
Herein I present outcomes of statistical and econometric analyses conducted on the sample consisting of 17 255 messages. Specifically, sellers sent all together 9480 messages; the other 7775 messages are replies from buyers. My approach looks as follows. Firstly, I specify what is considered to be lying in this framework. Then, I briefly describe the behaviour of participants in the experiment with regards to their role and particular incentive structures and examine whether buyers were good at detecting lies. Also, I summarize conclusions of previous investigation on the same dataset by Ledeboer (2016). For the most part, I am concerned with the analysis of linguistic dimensions of conversations and report evidence to either support or reject hypotheses stated in the previous section.
4.1. Behaviour of participants in the experiment
The Oxford English Dictionary defines “lie” as an intentionally false statement10
. This definition is applicable in this context, too. Sellers see the actual colour of the card, thus claiming the other option should be considered as deceiving11. Indeed, I use the same logic but distinguish between two possible scenarios. More than 95% of lies represent cases where sellers indicate green after drawing red card. This is a rational move taking into account expected payoffs and information asymmetry. However, nearly 13% of sellers tried at least once to deceive in a more complicated manner; they recommended choosing red in spite of green card being drawn expecting buyers not to follow their advice. 4 out of 108 sellers did it two or more times. I denote such behaviour as a “sophisticated” lie. The idea is based on a concept of sophisticated deception introduced by Sutter (2009) who proposed that truth telling can also be a way to deceive if person anticipates not being trusted. In this setting, sellers went one step ahead and pretended to be honest hoping their match would not believe them. The most important patterns in behaviour of sellers and buyers in the experiment, such as the percentage of lies or the rate of guessing green, are summarized in Table 2. The table also illustrates the effect of incentive structure on decision-making of participants. To test whether the height of incentives provided mattered, I conducted several statistical tests12 and discovered the following patterns. Predictably enough, the occurrence of lying was
10
Retrieved July 15, 2017, from https://en.oxforddictionaries.com/definition/lie#lie_Noun_200. 11
Some studies differentiate and define lying and deceiving as two different phenomena (e.g. Bok, 1978). In this paper, I do not make any distiction between them and use the terms as synonyms similarly to DePaulo (2003). 12
I assumed that the variables of interest are not normally distributed (binary) and groups of observations are independent. Thence, I used a non-parametric alternative to two sample t-test, Wilcoxon rank sum test.
- 14 -
significantly higher for conversations where sellers were given high expected payoffs whereas there was not found any significant difference in deception rates between two low incentive structures (Wilcoxon rank sum test, two-tailed, Z = - 4,400, p < 0,001 for high vs. both low incentive schemes; Z = 0,190, p = 0,849 for low vs. very low incentive scheme). As Figure 1 confirms, buyers seem to have foreseen these practices acting in accordance with advices more frequently in lower incentive cases (Wilcoxon rank sum test, two-tailed, Z = 5,282, p < 0,001 for high vs. low incentive schemes; Z = 0,416, p = 0,678 for low vs. very low incentive scheme). Regarding the reaction on different recommendations, buyers realised that a claim indicating green is less trustworthy. Such awareness exists under all payoff structures and is significant at least at 5% level (Wilcoxon rank sum test, two-tailed, Z = 7,577, p < 0,001 overall; Z = 3,707, p < 0,001 for high incentives; Z = 5,524, p < 0.001 for low incentives; Z = 2,040, p = 0,041 for very low incentives). Nonetheless, buyers listened to “green claims” above chance levels, to be more precise in almost 57% of the cases.
Table 2 – Behaviour of sellers and buyers: Summary statistics by incentive scheme
- 15 -
4.2. Buyers’ ability to detect deception
Were buyers able to recognize a liar? I proceeded similarly to Belot and van de Ven (2016) when determining buyers’ accuracy to predict the truthfulness of sellers’ claims and measured it as the difference between the fraction of times when buyers chose green and the green card had been drawn and the fraction of times when buyers chose green but the seller had actually faced a red card. The formula looks as follows:
bi ity to detect deceit ( guess green | green ) ( guess green | red )
Table 3 shows how buyers succeeded in the game. The results indicate that buyers could identify false recommendations no matter the incentive structure or specification (corresponding p-values resulting from two-tailed Wilcoxon rank sum tests are reported below lie detection ability parameters). The only exception is the insignificant difference for claims where sellers stated only green in the very low incentives condition which I, however, attribute to considerably smaller sample size of given group.
Table 3 - Accuracy of buyers’ deception detection
4.3. Earlier findings from analysis of the dataset
Using the same experimental evidence, Ledeboer (2016) studied in her thesis how peoples’ decisions to lie were affected by expected payoffs and gender and nationality of participants in matched pairs. She revealed that the choice to deceive depended on previous choices and circumstances. Regardless the order of incentive structures in two parts there was found no considerable difference in lying rates under high payoff schemes (a comparison of the part 2 of Treatment i, and the part 1 of Treatment ii,). Nevertheless, when participants began with higher monetary payoffs (Treatment ii,) they were significantly less inclined to lie in latter
- 16 -
“less-paid” rounds. Regarding gender, the author concluded that men and women behaved similarly when deceiving with respect to the height of incentives. Likewise, no support for any effect of homophily on the decision-making of sellers or the perception of buyers was found in the data. Further interesting finding was that buyers were more accurate in distinguishing lies in part 2. The author interpreted it as a consequence of getting used to the setting since there was no feedback provided suggesting no possibility for learning.
4.4. Linguistic analysis
In Section 3, I defined my research question and hypothesised about potential answer(s). This subsection is devoted to the examination of language features of recorded messages with respect to their truthfulness and trustworthiness.
Preceding any analyses, LIWC 2015 software was used to process texts of individual conversations into a set of output variables assessing linguistic dimensions (for a programme specification, see Pennebaker et al., 2015). The majority of these parameters indicate the share of specific words in overall passages (for more detailed information on each category relevant for this study, see Table 4 in Appendix A). The only exceptions are two word-count and four standardized summary variables capturing the level of analytical thinking, clout, authenticity and emotional tone. Moreover, I created several additional variables to be able to investigate all features of texts discussed in Section 3.3. Firstly, a measure of the rate of question-asking by buyers was assigned to each conversation. Then, I generated necessary binary variables capturing the occurrence of words in messages of sellers13. I was primarily interested in words and phrases related to trustworthiness, monetary aspect, texting style and this context, namely
“trust”, “promise”, “honest”, “money”, “haha”, “green”, “red” and “good condition”.
Furthermore, I looked at the vocabulary of subjects participating in the experiment. In Tables 5 and 6 in Appendix B, there are listed twenty most frequently used words alongside with four lists of the most common phrases. On average, one in ten words was either “I” or “you” - two words which ranked the highest with respect to the occurrence. The examples of typical expressions were “trust me” or “I have a green card”. Considering the format and context of conversations, it is not too surprising that subjects used mostly short words. Only 7,4% of their communication corresponds to so-called big words containing more than 6 letters. For comparison, in formal texts almost one fifth of words belong to this category.
- 17 -
To study the presence of linguistic cues referred to in Section 3.3, I performed multiple tests on the whole dataset as well as separately for cases in which sellers indicated to have drawn green. This was done to check for robustness of outcomes by excluding sophisticated lies from the analysis since I reckon this type of conversations could be quite specific by their definition and nature. In addition, I controlled for the height of incentives provided. Naturally, the percentage of lies was higher for high incentive cases. This suggests that also subjects, who did not lie under low incentive schemes, tried to deceive when payoffs increased. Assuming that they were not so accustomed to such situations, I suspect their style of texting may have had particular features which were not evident in messages of frequent liars potentially leading to several distinct findings of the analysis.
Since neither of word-count and LIWC variables do follow a normal distribution (all p-values resulting from Shapiro-Wilk test for normality are smaller than 0,001), I used Wilcoxon rank-sum test, a non-parametric alternative to two-tailed t-test, to compare means of opposing independent samples - truths vs. lies to look for distinctions in language of sellers and claims
followed vs. claims not followed to identify how buyers react to various texting styles.
Afterwards, I analysed whether some combinations of studied cues have a joint potential to some extent determine truthfulness of claims and perception of buyers. This econometric investigation is desirable to further examine the presence and significance of linguistic features since their correlation may possibly lead to misinterpretation of outcomes resulting from statistical comparison. To do so, I worked with number of subsets and specifications. I started with linear probability model treating each conversation as an independent observation14. Then, I estimated logit transformation of the same model to check for approximate validity of p-values resulting from the previous regression15. Last but not least, I adjusted the independence assumption to bring it closer to reality16 and conducted linear fixed effects estimation controlling for subject-specific characteristics in texting style and perceptions. In the end, all three models yielded similar results. Noteworthy findings are discussed individually in following subsections. The summary of all results from this analysis can be found in Appendices D and E (see Tables 7 to 12).
14 Besides, I assumed random assignment of treatments and linear relationship between predictors and
corresponding dependent variables. To deal with heteroskedasticity which is present in linear probability model by its construction (binary character of dependent variable), I used robust standard errors.
15 There were several reasons why I perfomed logit estimation (the presence of outliers; predicted probabilities outside the range of <0;1>; non-normally distributed errors). However, since in almost all of the cases results of both regressions were similar with regard to significance and direction of estimated effects, I consider LPM estimates to be valid for hypothesis testing and subsequent inference.
16
Instead of considering each conversation as independent, I clustered dialogues by relevant subject ID and assumed independence of participants.
- 18 -
4.4.1. Actual cues to deception
Truth-telling subjects sent on average almost 8 messages during a two-minute conversation whereas liars tended to write one text less. The same holds for the number of words. Messages of deceivers advising green were by more than four words shorter compared to ones by sincere sellers. Means are significantly different at 5% level (p-values 0,013 and 0,0001 for
all sellers and sellers claiming green respectively). After controlling for other linguistic
dimensions, sellers claiming colour truthfully still appeared significantly more talkative, however, notably strong differences were found only for high incentive cases (all significant at 1% level). Based on these findings, it seems that devoting effort to write longer message is used by honest highly-motivated subjects as something like a way of showing their sincerity. The above-stated result seems even more robust after closer inspection of buyers´ side of communication. I found on average 24% fewer questions raised by participants who were advised honestly with p-value smaller than 0,001. Although excluding conversations following red claims increases p-value to 0,047, the significance of the mean difference is still preserved. Correspondingly, almost all econometric model specifications confirm the pattern of less curious buyers in truthful situations. The only exceptions were the cases in which sellers facing high incentives recommended green. For these, buyers’ rate of asking was similar regardless the veracity of claims. Considering negative relation between the height of incentives and the number of questions, I suppose buyers expected that almost every advice could be false and decided to choose colour at random instead of querying their match.
Consistently with previously published research, this experiment also provides evidence that the pronoun usage is an important aspect for determination of trustworthiness. For instance, the first-person singular pronouns corresponded to 8,46% of words in messages of sellers claiming green truthfully compared to 7,31% share of the vocabulary of liars who also indicated to have drawn green (significantly different at 1% level). The econometric investigation, however, identifies such distinctions exclusively for low incentive cases. This suggests that the lower degree of self-referencing was common only for more accustomed liars. On the contrary, I found significant support for excessive occurrence of second-person pronouns in messages of deceivers facing high incentives (all relevant p-values smaller than 0,05). This indicates that occasional and frequent liars possibly had different style of texting. Regarding the presence of other-person or impersonal pronouns and all kinds of their plural forms, this data presents no convincing support for increased usage by any group of sellers.
- 19 -
Looking at emotional aspect, participants expressed themselves in a quite similar tone irrespective of whether they were sincere or not. The only substantial variance was found for the presence of negative emotions in communication following claims in which subjects recommended choosing red product. If such claims were false, sellers used on average over two times less negatively-toned expressions. Although this difference is just marginally significant with p-value equal to 0,123, I attribute it to the small sample size (only 20 cases of sophisticated lies) and consider it as partial evidence against the proposition that liars´ style of texting is more emotional. In addition, I examined how participants used negative forms of words. Although statistical tests indicate the increased occurrence of negations in honest conversations (p-values 0,044 and 0,065 for all sellers and sellers claiming green respectively), after controlling for other linguistic categories there was found just weak positive association between truth-telling and the use of negations for high incentive cases. The words indicating certainty, such as sure, never or always, occurred more frequently in truthful messages. Likewise, testimonies referring to trust, promises or honesty were more likely associated with truths. Figure 2 shows the percentage of conversations in which sellers mentioned at least once the word related to trustworthiness by the veracity of corresponding recommendation. While all mean differences showed in the graph are statistically significant (p-values equal at most 0,076), econometric models points out only words “promise” and “trust” to be differently used by liars. In low incentive cases, messages including “promise” were 20% more likely to be linked to truths (significant at least at 5% level). When high incentives applied, emphasising the word “trust” was a considerable cue to honesty (p-values smaller than 0,01 for green claims). This implies that there is another distinct pattern in the vocabulary of frequent and occasional liars. The first group seems to have refrained from promising, the latter apparently avoided mentioning phrases related to trust.
- 20 -
Messages of sellers who advised the authentic colour of their card contained significantly more context-specific words and phrases. For instance if looking at claims that the green was drawn, up to 81% of honest sellers used the word ”green” opposing to only 62% of deceiving subjects. Statistical tests and econometric tests for all relevant expressions17 recognize substantial differences at least at 10 % significance level.
The analysis revealed no other important distinct characteristics of the language of honest and lying sellers. With regard to the lower cognitive complexity and the increased referring to money which have been pinpointed by earlier studies as likely indicators that some information is misleading, I found no sufficient support in the data. Contrariwise, the firstly stated linguistic dimension appears to be highly insignificant in relation to truthfulness (corresponding p-values higher than 0,65). On the other hand, trustworthy subject seems to have talked about money to a greater extent (6,7% vs. 4,6% for all truths vs. all lies respectively). The difference is, however, not statistically significant for any specification.
4.4.2. Perceived cues to deception
Claims followed by lengthier messages were significantly more likely trusted. Statistical tests comparing means yield p-values 0,03 and 0,0001 for all sellers and sellers claiming green respectively. Buyers seem to have been able to identify this cue to honesty in high incentive conditions (significant at 1% level). When low incentives applied, the same was not possible because of the similarity in length of messages of sellers irrespective of the claim truthfulness. Buyers oftentimes distrusted advices of sellers who excessively reached out and referred to them as their match. Specifically, outcomes of the statistical analysis show that claims followed by messages with higher occurrence of second-person pronouns were significantly less frequently followed (p-values 0,027 and 0,048 for all sellers and sellers claiming green respectively). All econometric specifications coincide in a finding that this holds only for high incentive cases. Considering findings reported Section 4.4.1, I suspect that these detected deceivers are probably sellers who lied just occasionally when payoffs were high enough. Buyers placed more confidence in advices of sellers who lead conversations in more casual and relaxed way. The experimental evidence provides acceptable support that sellers writing
17 To be more precise, it does not holds for the presence of word “red” for the subset claims green and the word “green” for sellers facing high incentives. This insignificance can be, however, anticipated due to the design of experiment and the nature of these specific cases.
- 21 -
in informal language18 were more likely considered credible (p-values from Wilcoxon rank-sum test equal to 0,011 and 0,104 for all sellers and sellers claiming green respectively). Similar findings are obtained after controlling for other linguistic dimensions, buyers seem to have been inclined to trust less formal messages slightly more (significant at 10% level for all three model specifications when analysing the whole dataset). Moreover, 42% of conversations after which buyers acted in accordance with claims contained the word “haha” in contrast to only one third of dialogues by sellers whose suggested option was not chosen. The difference is significant at 5% level.
Also, buyers listened to recommendations considerably more times when sellers mentioned the word “promise” or “money” (corresponding p-values for all mean comparisons equal at most 0,02).19 The fact that claims including a promise were viewed as trustworthy is
confirmed by all statistical and econometric specifications. Claims of sellers referring to monetary aspects were significantly more frequently followed only under low incentive structures (significant at least at 5% level).
4.4.3. Limitations of applied statistical and econometric methods
To conclude, I shortly discuss some limitations and drawbacks of applied analytic tools. The statistical comparison of sample means was conducted to outline elementary relations in the data and represented only initial step before more complex investigation. Though its outcome is easily understandable, it misses broader context due to separate evaluation of each linguistic dimension. Regression estimates present more realistic view on actual patterns since it allows for control over more aspects at the same time. However, it is still not possible to treat for everything which possibly affect studied phenomenon. Omitting such feature can lead to so-called omitted variable bias. For example, in this analysis I assigned every subject some unspecified fixed characteristics. Probably, it would be more appropriate to use variables representing more specific attributes of subjects instead, such as their level of English or personality type. Also, each subject could treat each match differently, for instance, because of personal preferences about person’ s appearance on the snapshot. I did not control for the possibility of such discrimination. In addition, for inference to be valid several assumptions have to be met. These were discussed in the beginning of Section 4.
18
Informal language measured by LIWC as one of its linguistic dimensions (see Appendix A for more details). 19
Subjects were also inclined to listen to advices of sellers who expressed negative emotions or mentioned the word “red”. These findings are very instinctive but at the same time quite irrelevant for examination of lie detection given the fact that both features prevalently appeared in dialogues following claims that red was drawn.
- 22 -
5. Conclusions and discussion
The detection of false information is one of the most critical challenges that people making online purchases have to deal with. In this thesis, I presented experimental evidence from a buyer-seller game simulating behaviour of people trading a product over Internet. Interests of involved subjects in this relationship were not perfectly aligned. Sellers had informational advantage and were incentivized to lie in some cases; buyers had to decide whether and when to trust their match. The only available information, on which they were able to base their decision, was the seller’s recommendation and two-minute informal free-format chat conversation. Recently popular buy-sell groups on social sites where second-hand items are traded represent relevant real-life examples of similar interactions. The outcome of this investigation is important not only because of the high economic relevance of studied relationship, but especially due to paper’ s contribution and novelties which complement the few applicable findings of previous investigation which was, however, conducted in relatively different setups.
At first, I discuss the behaviour of subjects who were assigned the role of sellers. It is not very surprising that the manipulation of payoff structure resulted in the positive association of the height of incentives and the expected probability to lie. More interesting and less obvious are finding of the analysis of linguistic dimensions. This identified the following patterns in the language of liars to be significantly distinctive from honest communication in this specific context: (i) the lower number of words in messages; (ii) the higher incidence of buyer’s questioning during conversation; (iii) the reduced usage of context-related phrases and words emphasizing credibility and certainty. Most of these findings are in line with existing economic research, only the lower incidence of words related to trustworthiness does not correspond with expected pattern proposed by psychology studies. I attribute this inconsistency to dissimilarities in settings and procedures. Furthermore, the variation in the height of incentives allowed for comparison in texting style of frequent and occasional liars. These differences potentially represent new insights added to the debate about deceptive communication. The first group assuming to be deceiving more frequently regardless the incentive levels used in their messages less first-person singular pronouns. These sellers also refrained from making promises to their match. In the texts of latter group consisting of deceivers which lied mainly under higher incentive conditions, there was found considerably more second-person pronouns whilst these subjects avoided mentioning phrases with the word
- 23 -
“trust”. It needs to be, however, noted that the distinction in these two categories of liars depended purely on the share of deceits in respective subsets, not on individual behaviour of sellers in the experiment.
Next, I examine how buyers succeeded in their lie detection efforts. According to applied definition of deception detection rate based on the paper by Belot et al. (2016), buyers proved to have been capable of identifying deceptive claims. In addition, I noticed their increased caution when evaluating recommendations made by sellers facing high incentive cases and the fact that they realised the lower credibility of advices indicating green. The subsequent linguistic investigation suggests that subjects in the role of buyers correctly perceived lengthier messages and sellers referring to promises as trustable. Also, more personal and informal conversations and messages containing the word “money” were likely to be followed by selecting the recommended colour. Nevertheless, these two cues do not belong to the identified features signalising honesty. Conversely, buyers often distrusted liars who excessively used second-person pronouns. As evidence from the investigation on the language of sellers suggests, these detected deceivers are probably sellers less comfortable and skilled in lying.
In conclusion, I make some propositions for further work on this topic. I suspect that adding feedback for both parties would bring the design of this experimental simulation closer to real settings and lead researchers to new yet undiscovered dimensions of deceptive communication. I expect that this feature would improve lie detection skills of buyers and adjust texting style of sellers in a way that they would use less recognizable cues to deception. However, which of these effects would be more substantial is questionable. I believe that this feedback would not have some general impact rather different individuals would be differently successful and effective in its implementation. Furthermore, I would experiment with different restrictions on the format of communication. Last but not least, I propose to conduct similar research in other relevant settings, such more formal communication with authorized persons or more relaxed conversation with people they know each other.
- 24 -
References
Belot, M., & Van de Ven, J. (2016). How private is private information? The ability to spot deception in an economic game. Experimental economics, 20(1), 19-43.
Belot, M., Bhaskar, V., & van de Ven, J. (2010). Promises and cooperation: Evidence from a TV game show. Journal of Economic Behavior & Organization, 73(3), 396-405.
Bok, S. (1978). Lying: Moral choice in private and public life. Pantheon, New York.
Burgoon, J. K., Marett, K., Blair, J. P., & George, J. (2004). Detecting deception in computer-mediated communication. Computers in society: Privacy, ethics and the Internet, 154-166.
Burgoon, J. K., Stoner, G. A., Bonito, J. A., & Dunbar, N. E. (2003, January). Trust and deception in mediated communication. In System Sciences, 2003. Proceedings of the 36th Annual Hawaii International Conference
on (pp. 11-pp). IEEE.
Burgoon, J., Blair, J., Qin, T., & Nunamaker, J. (2003). Detecting deception through linguistic analysis. Intelligence and Security Informatics, 958-958.
Cialdini, R. B. (1993). Influence: the psychology of persuasion. Coffman, L., & Niehaus, P. (2015). Pathways of persuasion. Mimeo.
de Haan, T., Offerman, T. J. S., & Sloof, R. (2011). Money talks? An experimental investigation of cheap talk and burned money.
DePaulo, B. M., Lindsay, J. J., Malone, B. E., Muhlenbruck, L., Charlton, K., & Cooper, H. (2003). Cues to deception. Psychological bulletin, 129(1), 74.
Dutta, S., Geiger, T., & Lanvin, B. (Ed.). (2015). Global Information Technology Report 2015. World Economic Forum. Retrived August 9, 2017, from http://www3.weforum.org.
Ekman, P. (2009). Telling lies: Clues to deceit in the marketplace, politics, and marriage (revised edition). WW Norton & Company.
Feldman, R. S., Forrest, J. A., & Happ, B. R. (2002). Self-presentation and verbal deception: Do self-presenters lie more?. Basic and applied social psychology, 24(2), 163-170.
Hancock, J. T., Curry, L. E., Goorha, S., & Woodworth, M. T. (2004, January). Lies in conversation: An examination of deception using automated linguistic analysis. In Proceedings of the Cognitive Science
Society (Vol. 26, No. 26).
Hertwig, R., & Ortmann, A. (2001). Experimental practices in economics: A methodological challenge for psychologists?. Behavioral and Brain Sciences, 24(3), 383-403.
Hyman, R. (1989). The psychology of deception. Annual review of psychology, 40(1), 133-154. Chen, J., & Houser, D. (2013). Promises and lies: An experiment on detecting deception (No. 1038).
Ickes, W., Reidhead, S., & Patterson, M. (1986). Machiavellianism and self-monitoring: As different as “me” and “you”. Social Cognition, 4(1), 58-74.
Ledeboer, E. (2016). The trade-off of one’s decision to ie : and the response of the deceived eva uated [Master’s thesis]. Retrieved from https://www.scriptiesonline.uba.uva.nl (Accession No. 610511)
Lewicki, R. J.(1984).“Lying and Deception: A Behavioral Model.”. Negotiation in Organizations, 68-90. Loewenstein, G., Cain, D. M., & Sah, S. (2011). The limits of transparency: Pitfalls and potential of disclosing conflicts of interest. The American Economic Review, 101(3), 423-428.
- 25 -
Mazar, N., & Ariely, D. (2006). Dishonesty in everyday life and its policy implications. Journal of Public Policy
& Marketing, 25(1), 117-126.
Meyer, P. (2010). Liespotting: proven techniques to detect deception. St. Martin's Press.
Navarro, J. (2012, March 15). The Truth About Lie Detection [Web log post]. Retrieved June 16, 2017, from https://www.psychologytoday.com.
Newman, M. L., Pennebaker, J. W., Berry, D. S., & Richards, J. M. (2003). Lying words: Predicting deception from linguistic styles. Personality and social psychology bulletin, 29(5), 665-675.
Niederhoffer, K. G., & Pennebaker, J. W. (2002). Linguistic style matching in social interaction. Journal of
Language and Social Psychology, 21(4), 337-360.
Pennebaker, J. W., Boyd, R. L., Jordan, K., & Blackburn, K. (2015). The development and psychometric
properties of LIWC2015.
Pennebaker, J. W., Mehl, M. R., & Niederhoffer, K. G. (2003). Psychological aspects of natural language use: Our words, our selves. Annual Review of Psychology, 54(1), 547-577.
PricewaterhouseCoopers, L. L. P. (2016). Total Retail Survey 2016. Retrieved August 15, 2017, from http://www.pwc.com.
PricewaterhouseCoopers, L. L. P. (2017). Global Entertainment & Media Outlook 2017-2021. Retrieved August 15, 2017, from http://www.pwc.com.
Schafer, J. (2016, November 19). 5 Things People Say When They´re Lying To You [Web log post]. Retrieved July 5, 2017, from https://www.psychologytoday.com.
Steiner, P. (1993). On the Internet, nobody knows you’re a dog. The New Yorker, 69(20), 61.
Sutter, M. (2009). Deception through telling the truth?! Experimental evidence from individuals and teams. The
Economic Journal, 119(534), 47-60.
Utikal, V. (2013). I am sorry: Honest and fake apologies. Mimeo.
Vrij, A. (2000). Detecting lies and deceit: The psychology of lying and implications for professional practice. Wiley.
Vrij, A., Mann, S. A., Fisher, R. P., Leal, S., Milne, R., & Bull, R. (2008). Increasing cognitive load to facilitate lie detection: The benefit of recalling an event in reverse order. Law and human behavior, 32(3), 253-265.
Wood, W., Kallgren, C. A., & Preisler, R. M. (1985). Access to attitude-relevant information in memory as a determinant of persuasion: The role of message attributes. Journal of Experimental Social Psychology, 21(1), 73-85.
World Economic Forum (2016). Digital Media and Society: Implications in a Hyperconnected Era. World Economic Forum. Retrived August 9, 2017, from http://www3.weforum.org.
Zuckerman, M., Koestner, R., & Driver, R. (1981). Beliefs about cues associated with deception. Journal of
- 26 -
Appendix A
LIWC linguistic categories relevant for this study
- 27 -
Appendix B
Lists of the most frequently used words and phrases in chats
Table 5 - Top 20 words by frequency (the absolute and relative occurrence)
- 28 -
Appendix C
Does the language of sellers differ depending on their sincerity?
- 29 -
- 30 -
- 31 -
Appendix D
Which linguistic patterns are viewed by buyers as trustworthy?
- 32 -
- 33 -
