Do quantitative cues and effort level influence people’s ability to detect deception and trustworthiness? : evidences from informal written messages

(1)

UNIVERSITEIT VAN AMSTERDAM

Do quantitative cues and effort

level influence people’s ability to

detect deception and

trustworthiness?

-Evidences from informal written messages

Name: Caixin Zhou

Student Number: 10829032

Study Track: Organization Economics

Number of ECTS: 15

Abstract

An increasing number of people use informal written messages to communicate with online sellers nowadays. Trust, as the essential of both traditional and online business, has been the research object of relevant studies. However, Chen and Houser’s study in 2013 is the first paper which focuses on the quantifiable cues in this kind of information. In this paper, I analyses the data from an experiment which only allows people communicate via online chat tool to test their conclusion. Evidence shows that the number of money-related words in the messages and sellers’ words length have significant positive impacts on sellers’ trustworthiness and buyers’ ability to detect deception. Many previous studies show that salespeople’s success is related to their hard working. Results from my survey suggested that people tend to trust those who work hard.

(2)

1

Statement of Originality

This document is written by Student Caixin Zhou who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

(3)

2

1 Introduction

Online shopping has already become one of the most popular forms of shopping for ordinary people in modern society. E-commerce has made rapid progress in the past few decades, whereas the development of it is even more promising in the future. For instance, China’s online shopping users reached 413 million in 2015 (361 million in 2014). The numbers for USA and UK are 205 million and 41 million, respectively. Trust which is the foundation of sales success (Farber 2007) in traditional business is also the cornerstone of the diffusion and acceptance of electronic commerce (Grabner-Kraeuter, 2002).

However, the buyer-seller negotiation in shopping online has a very special environment. Generally speaking, buyers are not able to observe the actual commodities before their purchasing action. And the communication between two sides online mostly through informal written messages such as Aliwangwang(an instant chat tool provided by the biggest online shopping platform—Taobao), messages from Facebook, the descriptive text posted on the product page and Q&A section online. In this situation, online deception detection turns out to be very different from the traditional studies which mostly focus on face-to-face communications (Chen, 2013). Given the unique context of the online communication process, it would be interesting to see what factors have an influence on both sellers’ trust worthiness and buyers’ ability to detect deception.

Deception and Trust are involved in many economic and social situations, especially the seller-buyer relationship building activity (Mazar and Ariely, 2006). Although the lies detection attracted people’s attention about 120 years ago (Lpmbroso, 1895), detecting deception on informal written messages was rarely discussed. Chen and Houser, as the first researchers working in this field from an economic perspective, run a “Mistress Game” which included 52 people and all their written messages. They suggested that there are three potential quantifiable objective cues in informal written words can be related to trust and deception detection. First, using encompassing terms can make people trust you more. Secondly, the trustworthiness of sellers whose words are relatively longer is higher. Lastly, messages contain more money or benefit related words are more likely to be considered as a promise can be trusted. However, this also is a sign of lying. Do they have the right

(5)

4

conclusion? If they do, I want to have a further step to research the impact of the number of these objective cues.

To answer this question, in this research, I analyse the data of the objective cues from 1647 chats from an experiment that simulates the online shopping process. In this experiment, each chat includes one buyer and one seller, and they only able to talk via an online chat tool. Asymmetric information (only sellers know the actual quality of the product) also exists in the game as the real business world.

Furthermore, there are some previous studies shows that salespeople with high-performance usually are those who are hard-working (Churchill et al.1985, Sujan et al. 1994, Hu 2009 etc.). I start to wonder, is that because the hard-working sellers are more trustworthy to buyers? Meanwhile, I cannot find any paper focus on the relation between effort levels and trustworthiness of sellers.

To figure out the relationship between effort level and reliability, I distribute 100 questionnaires includes original dialogues context from simulating sellers-buyers chats. The effort level is hard to have the same standard to evaluate for everyone, so in this survey, I ask respondents to assess both how hard the seller is selling and how much are sales people in the dialogue can be trusted.

The main results from this paper are based on Mann-Whitney rank sum tests and logistical regressions. The findings show that buyers do trust sellers more when their words are longer and contain more money-related words. However, more benefit-related words in sellers’ words also are the signals of lying vendors. Encompassing terms, as an important cue in informal written messages mention by Chen and Houser (2013), do not show a significant effect on seller’s trustworthiness, although this kind of words decreases the accuracy of buyer’s deception detection. For the effort levels which are evaluated subjectively by survey respondents, there is a significant positive relation between effort level and sellers’ reliability. People tend to trust those who are hard working.

These conclusions can be related to our daily life situations include informal written messages such as online purchasing activates and online dating. For example, sellers who want to be trusted can include more words in their messages and advertisements in texts for.

(6)

5

Salespeople online also can find the right way to show that they are really working hard. On the hard, buyers can be careful about sellers who talk a lot about money.

Also, as the first paper which verifies Chen’s conclusion and tests the effort levels’ effect on trustworthiness, this research tries to draw more attention to this topic.

The structure of this paper is organized as follows. In section 2, I discuss some related literatures, and my hypotheses are showed in part 3. The experiment and survey design are presented in section 4. Section 5 explains the results. Conclusion, whilst further discussion of this paper is in section 6.

2 Literature Review

For a better understanding of my research topic, the first part of the literature review is about the definition of deception and how the general fraud detector developed in the past. The second part of this review is about the dramatic growth of e-commerce and the online trust issues. Lastly, because of the unique communication environment online, I will introduce some previous studies about the deception detection in informal written messages and the limitation of the economic attention on this topic.

2.1 Deception Detection and Trustworthiness

Based on the comprehensive deception theory from Mitchell and Thompson (1986), Thagard (1992) and Johnson et al. (2001), a deception is defined as facing a conflict benefits situation, there is a cogitative inter-individual behaviour between two sides. The one side, the lie-teller, actively transfers their messages and information to the other party, the respondent, and make them draw a wrong conclusion.

In ordinary people’s lives and general society, deception detection always plays a vital role (Kleinmuntz and Szucko, 1984) .Thus, this topic has attracted research interest from a lot of different aspects such as Psychology(Lykken, 1978;Waid and Orne 1981 etc.), Information Systems (Xiao and Benbasat, 2011; Zhou 2012 etc), Classical Economics (Horton et al., 2011), Behavioral Economics (Belot and van de Ven, 2014),Organizational behaviour (Gefen et al., 2003) and Sociology (Toma et al., 2008 ).

(7)

6

Psychologists showed the most and the earliest interest in ability to detect lies so they have a considerable number of related researches. As a summary, Bond Jr. and DePaulo (2006) analyzed 206 papers and 24483 observations from previous studies and found out that people do have the ability to detect lies, although the accuracy of judgement is just slightly higher than the random chance (54%).

In order to make people improve the capacity to detect lies, scientists from different background started related researches very early. An Italian criminologist Lpmbroso(1895) argued that emotions can be an index of lying and later, Benusi(1914), a psychologist, suggested that respiratory rate can be another cue for deception detection. And Marston, a lawyer and psychology student at Harvard, showed that the blood pressure is an important indicator for lie-detect in 1938. All these three factors are now used in the modern polytrophic machine (Kleinmuntz and Szucho, 1984). Furthermore, Meservy et al. (2005) conduct research on deception detection through automatic, unobtrusive analysis of nonverbal behaviour. They design an automated deception detection system, where the polygraph, micro-momentary expressions, and movements and behaviour patterns are analysed for both written and oral communication.

In the development of lie detection, people also consider a lot of other factors such as movements, behavioural patterns (Meservy, Jensen, Kruse and Burgoon etc., 2005), the relationship between the sender and receiver (Buller, White, Afifi and Buslig, 1999), social status (Miller and Stiff, 1993) and so on. The confidence level is also taken into account when people try to find the factor can be related to deception detection by DePaulo et al. (1997).

There are already some useful models such as Criteria Based Content Analysis(CBCA) and Reality Monitoring (RM) that can improve the success lie detection rate significantly from the guess rate to 80% and 64%, respectively (Edward, Roberts and Bull, 2000).

However, these models are more suitable for a criminal investigation with well-trained officers. And most of previous studies about lie detection are in psychology area. Limited studies are related to the economy and tell ordinary people how we should use the lie detection knowledge in our daily business activities like choosing an online store or a seller.

(8)

7

2.2 Effort Level

There are many reasons besides trust for salespeople’s success such as effort level, product type and quality, organizational/environmental factors and so on (Walker, Jr et al. 1985, Oakes 1990 etc.). Particularly, effort level seems to be the most hard to evaluate objectively factor and many researches show interests in the relationship between hard-working and successful sales.

Cole (2009) suggests that compared to work smart, working hard has a great positive impact on sellers’ performance and this factor has no difference between men and women. Hu and others (2009) try to research on this topic in the Chinese culture, which is very different from most previous studies in western cultures. However, they still get the conclusion that hard work has a statistically significant correlation with adapting sellings.

2.3 Online Shopping

2.3.1 Online Shopping Development

With the development of internet technology, shopping online has already become one of the most popular shopping methods in many countries. Moreover, the explosive growth of the electronic commerce since the 1990s makes a great impact on the world economy. (Chen and Chang, 2003)For example, the total annual turnover of Taobao, the leader of online C2C(Customer to Customer) platform providers in China with more than 70% of market shares, reached 4600 billion U.S. dollars(3 trillion RMB) in the 2015 financial year (Apr. 2015 to Apr.2016).

For the great number of customers online, e-vendors realise that trust is the most important attribute which consumers respond to()

2.3.2 Online Shopping Trust Issues

It is without doubt that trust is the centre of online business service providers like eBay. Roman (2010) argues that consumer’s perceptions of online retailer’s deceptive activities have a negative impact towards the satisfaction and loyalty intentions based on a sample of 398 real customers. Li and Zhang (2002) conduct research in online shopping customers’ attitudes and behaviours from IS (information system) perspective. Based on 35

(9)

8

peer-reviewed publications in IS, they perform a conceptual model of online shopping and find out that deception detection is at the centre of customer satisfaction.

Although people start to see the importance of trust online, trust issues within the online communities have been debatable for decades. While the research in individual trust in online firms started early in 2002 (Bhattacherjee, 2002), the trust and deception detection related issues already existed for a long period since the online business model starts. While lots of researchers cover the psychological factors for deception detection, it is important for us to recognise the uniqueness that the online market contains. For example, the customers who trade online typically do not talk to the sellers face to face.

As the traditional psychologists want to test whether people have the ability to detect lie on face-to-face communication, Grazioli and Wang (2001) try to understand whether consumers can detect Internet deception or not. Based on the deception, trust, and risk (DTR) model proposed by Grazioli and Jarvenpaa, they suggest that consumers usually struggle in establishing the trust with online sellers. While on the other hand, lots of Internet users and online shoppers could have a hard time in deception detection.

Given the increasing attention to the online shopping trust problems, people are trying to know more influence factor of Internet lie deception, and some researchers have impressive results. Xiao and Benbasat (2011) investigate the fraud in e-commerce, especially in business to customers (B2C) model, from a theoretical perspective, and they conclude that the online deception contains concealment, equivocation, and falsification. Riedley et al. (2010) conduct research about trust issues from eBay offers based on an fMRI method. Evidences from their research support that the gender difference has an effect on online trust issues.

We also can have some signals from e-vendors’ side for detecting online fraud. You et al. (2011) discover that in the Chinese C2C market, the trustworthy sellers will become harder to be identified when they are building the reputation. Another factor can be the posting cost online; Ott et al. (2012) find out that online communities with lower posting cost will have more deception activities.

Since people seem to lack the ability to detect online deception, economists, and other researchers try to find what factors can affect the buyer’s judgement. Lee et al. (2005)

(10)

9

suggest the problem of trust and deception detection related issues can be viewed as a lemon on the Web. They point out that asymmetric information is the key matters that involve deception detection and trust. Also, they identify that there are some pieces of information can be viewed as the signals of trustworthiness. For example, the brand can be a powerful message. People tend to have more trust towards well-known brands compared to unknown brands. On the other hand, the privacy policy that is stated on owner’s website is also another signal of trustworthiness as a lot of users will have concerns about sellers’ integrity.

2.4 Deception Detection in Written Messages

As discussed above, we can see how important the online business in the world economy and the vital role of the trust and deception detection in e-commerce. Also, because the informal written messages are the primary method for buyers and sellers online to communicate, this part of literature review will introduce some development of deception detection in written words.

Chen and Houser (2013) suggest that in natural language written messages, people will usually respond to cheap talk, and it is among the most efficient ways to communicate with sellers online. Thus, the research about deception detection in informal written messages can contribute to both economy and people’s daily purchasing activities.

According to Masip, Bethencourt et al. (2011), very few studies are focusing on deception detection of written messages. No peer-reviewed paper has been reported on the cues for receivers to detect lies in written accounts. Thus, they had an experiment which contains 78 participants who need to write either a truthful or deceptive story and argued that only verbal fluidity among all the verbal skills can be a sign of successful communication. They also suggested that subject cues such as the number of details, emotion words and consistency of language (signals that receivers think can be useful for lie detection) have no relation to the success judgment of the story’s truth.

Besides psychologists, most of the previous studies about trust issues through written messages are in Information System view. For instance, Gefen et al. (2008), from an IS perspective, suggest that online trust is highly related to the concept of perceived risk.

(11)

10

In order to create a computerised tool to help people detect lies from the written messages, Zhou et al. (2003) conducted an exploratory study into deception detection in text-based computer-mediated communication which is very close to the research topic. Although they realise the textual nature of online communication may raise an additional challenge in detecting deception, it is suggested that deception detection is related to the following eight possible objective cues: offensive usage, incomplete sentence, ellipsis, wordy, sentence variety, jargon usage, run-on sentences, second person and possessive from. While some signals(ellipsis, wordy, passive voice, second person address and possessive from) shows the significant difference from truthful senders’ group to deceptive senders, the rest shows no difference between these two groups. They also test some other possible factors such as diversity of language, complexity, affect and so on. The main conclusion from their research is that deceptive messages are significantly higher on quantity cues and lower on diversity than messages from truth-tellers.

However, their findings have a big difference from previous studies and it may suggest that objectives cues in written messages are possible to differ in terms of the deceivers.

From an economic perspective, Chen and Houser (2013) devise a special three-person experiment to investigate cues of deception and trustworthiness in simple written messages. The results of this research suggest that people do recognize the cues that correlate with deception and trustworthiness, but they do not always use them right. Specifically, they found out that people are more likely to trust senders who sent messages which contained encompassing words, included more words and mentioned money, while information contains money is more likely to be a lie.

From the discussion so far, it becomes clear that there are still theoretical gaps from the previous literature review. First of all, most of the deception detection related research from written messages are from IS research. There is not enough research that is from an economics perspective. Secondly, no research specifically focuses on informal written messages from conversation included both sellers and buyers which is the most common situation in real online shopping communication phase. Also, there is no literature about the relation between the effort level in other people’s mind and trustworthiness.

(12)

11

Therefore, this paper will try to fill this gap by analysing quantitative objective cues in written messages from an economic game and contacting a survey to test the effort level’s effect. This experiment will be very close to the real online shopping environmentbecause1) most of online communication process does not contain physical meeting, primarily based on written information; 2) information asymmetry exists between the buyer and seller about the quality of the product online since normally the seller write and post all the information about the product online.

3 Hypotheses

Chen and Houser (2013) design a novel three-person trust game to test the objective quantifiable cues for trust and lie in the informal written messages from an economic perspective. It is identified that messages contain encompassing terms and a higher number of words are significant to be viewed as trustworthy promises. It is also identified that messages that mention money are to be trusted significantly more even people whose words contain money are more likely to broken the promise. Are they right?

Besides trust, the essential of an economy (Belot et al. 2012), previous studies (Hu et al. 2009, Williams 1991) suggest that successful salespeople are also those who are working hard. Does the effort level connected to trust?

Thus, my research question is as follows:

Research Question:

Do quantifiable cues and effort level in informal written messages influence buyers’ deception detection and sellers’ trustworthiness?

We can see from the research question that I actually want to test two different types of cues in informal written messages. One is the objective, quantifiable signals which contain word length, the number of encompassing terms, and money mention times; the other kind of factor is subjective evaluation includes the buyers or others’ personal view of seller’s effort level to sell.

(13)

12

For the quantifiable objective cues, Zhou and Twitchell (2002) examine some potential possibilities in emails from system information perspective but they have a different opinion from Chen (2013). Evidences from Zhou (2002)’s research argued that a larger number of words can be a sign of a lie and less trustworthy.

To verify and expand Chen and Houser’s results, and also because our experiment is from the economy perspective as them, my first hypothesis will follow their key findings.

My hypothesis one is as follows:

Hypothesis 1: People would tend to trust sellers when their words include more objective cues (money mention, encompassing words mention, and more words).

For the subjective evaluation factors for trust, DePaula (1997) argued that confidence can affect people’s trust choices. Except this factor, there are few studies tried to find the relation between the subjective cues and trustworthiness. There are some previous studies show that the positive relation between hardworking and sales success. (Williams 1991, Hu 2009) Because of the importance of trust in all kinds of business, the reason behind this result can be the higher trustworthiness for those who work hard. Is that true?

Based on what I discussed above, I expect that the effort’s level has a positive impact on sellers’ trustworthiness.

My second hypothesis is as follow:

Hypothesis 2: When buyers think sellers are really trying to sell a product (the effort level is high), purchaser are more likely to trust vendors.

4 Methodology

There are only very few studies that focus on detecting informal written messages from sellers influence buyers’ deception detection. As stated in the literature review, mixed approach in this paper combined with the design of experiment and survey will gain a better idea of the trustworthiness and deception detection of informal written data during the transaction.

(14)

13

4.1 Experiment

4.1.1 Experiment Design

The design of an experiment is a research method that systematically understands and determines the relationship between different factors that can potentially affect the processes (Kempthorne, 1952). The experiment aims to predict the outcome of the process by introducing the predictor. There are several considerations that researchers need to pay attention to, which are validity, reliability, and replicability (Condra, 2001).

In the experiment, there is an even number of participants that would be paired randomly into buyers and sellers. This means at the beginning of the game, half of the participants will become buyers and the other half will become sellers. Then for each pair, which consists of one buyer and one seller, a card will be presented to the seller. This card represents any service or goods that sellers want to sell while buyers are willing to buy. Sellers will have equal probability to receive a green card or a red one. A green card indicates the product is in good condition while the red one indicates the product is bad. The colour of the card is considered confidential to the buyer and will not be reviled to the buyer during the purchase. The seller claims the colour of the card to the customer after he/she gets the card. After the claim, there would be an opportunity for buyers and sellers to communicate for 2 minutes. During this phase, seller and buyer will be able to use informal written messages for their dialogue. After the communication, buyers will need to make decisions on whether or not to purchase the card. Customers will not purchase the card if they think the card is red, based on the communication and information gathered during the two-minute dialogue but they will buy it if they believe that the card is green.

The buyers’ decision will be the last step of this game. After this, sellers and buyers will get their payoff according to the situation they are in. There are four different situations in this game. First of all, the card which the seller gets is green, and the buyer purchases it because he/she thinks the colour of the card is green. In this case, both seller and buyer will get the highest payoff in the game. The second case is the product is in good condition (Green), but the purchaser does not buy it which means both seller and buyer gets the lower from this game. The third situation is the buyer purchases the card, but the product is in poor condition, in this case, the seller gets the higher payoff and the buyer gets nothing. The

(15)

14

last situation is the buyer thinks the card is red and does not buy it and the card which the seller has is red. Then the buyer gets the better payoff and the seller gets the lowest payoff.

Table 1 shows three different payoff structures in this experiment. In this chart, the left number represents how much money (Euros) does the seller get in the given situation and the number on the right is the benefit which the buyer would get. For instance, in Payoff Matrix 2, where the case shows the seller’s card is green but the buyer’s choice is red. The payoff results for this pair in this situation is”12,0”. And then the seller gets 12 euros but the buyer gets nothing.

Table 1: Three Different Payoff Situationsinthe Experiment

As stated above, buyers will be better off if they can obtain the actual information during the two-minute communication. Thus, in the chat part, the buyer has the biggest interest to find out what colour is the real colour of the seller’s card. From vendors’ perspective, no matter what colour of the cards they receive, it is important to make the buyer believes the card they have is green and purchases it. Thus, sellers should try their best to convince customers that the product is in good condition in any payoff situation.

We can treat the environment of this experiment as a simulation of purchasing experience during the online transaction, particularly in the customer-to-customer, or consumer-to-consumer (C2C) transaction within the online marketplace like Amazon.com, eBay and Alibaba(King and Lee, 2000). In which cases, customers have very limited

Payoff Matrix 1 Buyer's Choice

Green Red

Seller's Card (Random dram)

Green 30,30 0,0

Red 30,0 0,30

Green Red

Green 18,30 12,0

Red 18,0 12,30

Green Red

Green 17,30 13,0

(16)

15

information about the actual products and usually, they can only communicate with sellers via different online chat tools. On the other hand, it depends on the sellers to disclose the true information of the product or service. Other than checking and reading the product information(the claim of the seller), most of the marketplaces like Amazon C2C and eBay will allow customers to ask questions to vendors so that sellers can reply that in the written form. The design of this experiment is a simplified version of the actual transaction in the real business setting. By conducting the research, the patterns and behaviour information from both buyers and sellers will be recorded.

Once the data is observed and recorded, statistical analysis will be conducted to understand the patterns behind the data further. First of all, the length of words during the communication from sellers will be collected. Secondly, the frequency of money-related words will be counted and calculated. Thirdly, the frequency of encompassing words will be calculated. Finally, the facts 1) whether sellers are lying or not; 2) whether buyers believe that or not will be recorded. Wilcoxon rank sum test will be conducted. Key statistical indicators like Z-value and p-value will be calculated.

4.1.2 Experiment Procedures

This experiment and all the data collection are conducted in the Centre for Research in Experimental Economics and Political Decision Making(CREED) Laboratory in University of Amsterdam. Before the experiment begins, participants are asked to turn off their cell phones and have no other kinds of communication except the online chat tool. The whole process is computerized. Before the experiment starts, every subject gets the same instruction about the structure and process of the game and their earning situation for part 1(Payoff Matrix 1). They all have to pass a quiz to ensure that everyone understands the game and then everyone gets a print piece of the instruction. Then the experiment starts.

We have this experiment for 16 sessions which contain 2 parts. For each session, there are 16 participants. Half of them are randomly assigned to be sellers by computer and the rest are buyers. The role of a subject maintains the same for the whole session. And in each part, every seller would be matched to one new buyer until they met all the buyers, so there are 8 rounds in one part. In the part one at the beginning of each round, a card is distributed randomly to the seller. The card can be green or red, i.e. a product in good or bad

(17)

16

condition. Then the seller claims the colour of the card to the buyer. After the announcement, the communication phase will open to the pair of subjects for two minutes. In the chat section, the format of talking is not regulated but offensive languages are forbidden. In this section, not only the seller can provide information to the buyer, but the buyer can also ask relevant questions to obtain some cues from the seller. During the communication, the colour of the card remains confidential to the purchaser. When the dialogue ends, the buyer will have a guess of the condition of the card. Then both the seller and buyer have to fill in a scale from one to ten for their trustworthiness. The customer estimates the reliability of the vendor and the seller evaluates in which trust level does the buyer believe in him/her. And this is the end of one round. There is no feedback and they move to next round when everyone finishes the first round. After 8 rounds in part 1, half of the pairs will get a change in their payoff structure (Payoff Matrix 2) and the other half will have a different adjustment on their payoff (Payoff Matrix 3). The process of part 2 stays the same as part 1.

After these two parts, all participants fill in a questionnaire includes multiple questions such as their strategy and their personal information. And all the subjects get their gains at the last step of this experiment. Their earnings will be the payoff of one random round picked by the software and adding a 5 euros show-up fee.

4.2Survey

The second stage of the research contains survey methodology that further investigates the research question. Based on the design of this experiment, it appears that most of the information obtained so far is objective information such as the number of one particular type of words. Survey methodology is applied to investigate the relationship between the intentions further to purchase the product. Moreover, the survey methodology is a quantitative method that can potentially understand the cognitive processes during the

decision-making process (Sudman et al., 1996). Overall, the purpose of this survey is to study the relationship between deception detection, trustworthiness from sellers, and effort level’s impact on buyers.

There are several benefits from using the survey method. First of all, it is relatively easier to dministrate the questionnaire process compared with experiment. In the

(18)

17

experiment process, it is required that all participants must be present in the same location. The time and geographic restrictions could potentially limit the sample size of the study. However, the survey will be relatively easier and less expensive. It does not have restrictions on time and locations. As long as participants can finish it and return the survey within a timeframe, the data will be valid. Secondly, the survey process can be automated by the usage of information systems. For example, the study can be conducted via a website and the results will be collected automatically which are available to the researchers. Thirdly, the experiment typically takes more time like about an hour, while the survey usually lasts for about 10 minutes. Thus, it utilizes fewer resources to conduct survey compared with other methods. Fourthly, the fewer restrictions further reduce the geographical dependence of the data. By asking different questions from the same subject, this provides more flexibility in the data analysis section later. Finally, the richness of the data can be enhanced by spreading survey among people from very different backgrounds.

However, it is also important to note that there are several potential disadvantages of using survey as a research methodology. For example, it might be the case that participants are not entirely aware of the reasons that they pick a specific answer. Some participants may not fully read the questions, and they just want to finish issues and tasks by randomly selecting the answers, which do not reflect their intention. Also, care must be paid towards the design of the survey. For example, survey question answers could lead to miscommunication and misunderstanding by different participants. “Somehow agree” may have different meanings. While some participants may consider this means “Yes,” some participants may take into account as a “no.” Also, it might be the case that survey itself designed by the researchers may contain some types of errors. The later section will describe the process of survey designing.

4.2.1 Survey Design

In this study, a set of survey questions is designed to understand further the potential factors which can influence the ability to detect deception. It is suggested that a set of ten sections will be asked to people from different backgrounds. By doing so, it helps overcome the bias that most of the participants from the experiments are students.

(19)

18

This survey contains 2 main parts; the first part is the introduction section which gives the respondents a brief looking at the basic structure of the experiment about and overall payoff data for seller and buyer. And the second part includes 10 sections which contain 10 random and original dialogues from the previous experiment. In each section, there are 2 questions following the chat. In addition, because of the featrue of the experiment, the conversations between subjects happen both in English and Dutch. Therefore, I have 2 versions of the survey; one is entirely in English, and the other consists of half in English and half in Dutch.

You can see a sample section of Part 2 below.

Figure 1 Example of one section in Part 2 from the survey

CHAT X

Question 1: In your opinion, how hard did this seller try to convince the buyer that the card

was green (i.e., the product is in good condition)? 1 (Not trying at all) 2 3 4 5 (Trying very hard)

(20)

19

Question 2: How convinced are you that this seller has a green card (i.e., the product is in

good condition)?

In the questions part of my survey, 10 independent chats are chosen for each version. And 2 same questions as showed above followed every single conversation.

I choose the Likert Scale (5-point scale) as the answer structure in my survey because Likert Scale is the most widely applied method to research on people’s subjective thoughts. Also, the 5-point scale is adopted as it is easier to interpret the data gathered from participants (Gliem and Gliem, 2003).More specifically, the purpose of this survey is to collect quantitative data about how people evaluate the relation between effort’s levels and trustworthiness for later statistical analysis. And a recent empirical research shows that there is tiny difference in mean, Skewness and kurtosis among 5- point, 7- point and 10- point scales (Dawes, 2008). Thus, from a statistical view, the analysis data from a five-point scale has no significant difference compared to other scales. Another advantage of Likert Scale is that the Evaluation criteria are consistent for every question, so it is easy for us to compare the average number.

4.2.2 Survey Sample

This survey’s carrier is the online questionnaire platform—Qualtrics. For the diversity of the respondents, besides the social media such as Facebook, Wechat, Whatsapp and other online social tools, I also sent out via Philips internal channel to get more different views from a real business world. In total, 100 respondents filled in this survey and 50 of them are for the English version and the restare for the English & Dutch version. Because of the language nature, most people who fill in the version includes Dutch conversation are from Netherlands. 1 (Almost certainly a red card) 2 3 4 5 (Almost certainly a green card)

(21)

20

Specifically, 51% of the respondents are female. Most respondents are from China (36%) and Netherlands (40%).And the average age of all participants is 30, ranging from 18 to 62.

Some background indicators show in below table.

Table 2Background factors by survey version

Factors English Version English&Dutch

Version Total Gender (Female) 56% 48% 51% Age (Average) 26.38 32.52 29.45 Nationality Netherlands 4% 80% 42% China 60% 12% 36% Others 36% 8% 22%

In the later section of this paper, survey and quantitative analysis will be conducted to assist answer the research question. First of all, I will collect effort level and trust level data for every single chat per respondent in the questionnaire. Then Mann-Whitney rank sum test will be conducted and key statistical indicators like Z value, p-value will be calculated based on the trust level. Finally, combined with the three top objective cues included money mention, encompassing words and number of words, the regression analysis will be conducted to see whether all these factors mentioned before have a considerable impact on trustworthiness or deception detection.

(22)

21

5 Data Analysis and Results

5.1 Overview of effective data

We conduct the experiment showed in the methodology part of this paper 14 times. Thus, there are 14 sessions. Under per session, 16 subjects participate, and half of them are assigned to buyers and the rest of them are sellers, so there are 8 different groups during the whole game. In each round, we let every seller rematch to a new buyer until every buyer has the chance to talk to every vendor and do it again after the incentives structure change (once) in the game. Thus, each different pair contains one buyer, and one seller do the communication twice, and we have 64 (8 sellers X 8 buyers) different pairs per game.

We are supposed to have 1792(number of sessions (14) x different pairs per session (64) x communication times per pair(2)) conversations and results between one buyer and one seller. However, some original conversations do not have words from buyers or sellers because the subjects give up the chance to communicate. In addition, some buyers do not make their choice after the communication.

Therefore, after selection, I have 1647 sets of valid data from the experiment which includes sufficient and complete information and data I need to do the following analysis to see the effect of cues.

In this 1647 sets of data, 813 groups have green card i.e. the product in good condition and 834 cards are red i.e. the product in bad condition. This distribution of card colour basically follows a random distribution. Furthermore, in the situation where the seller holds a green card, 100% of them choose to tell the truth, but only 32% of sellers who carry a red card do not lie to the seller. For buyers, they do show the ability to detect lies since they have a bigger success rate on lie detection (66.91%) than the random chance (50%). However, the accuracy for people who are told by truth(80.66%) nearly double the accuracy rate for people who are under deception situation(40.63%).

For the data from the survey, we have 100 sets of complete data like mentioned in the survey sample part.

(23)

22

5.2Objective Cues from Experiment

As mentioned before, my first hypothesis as below:

Hypothesis 1: People are more likely to trust sellers when their words include more objective cues (money mention, encompassing words mention, and words length).

To test hypothesis 1, I first define the “Trust” as the buyer’s choice of the card’s colour after the communication in the experiment is guessing the same colour which the seller claims in their conversation (e.g. the seller claims the card he/she holds is green and the buyer’s guess is green) and “Not Trust” as the buyer’s choice is not the colour which the seller tell in their chat (e.g. the seller said he/she has a red colour but the buyer’s guess for the card colour is green). Furthermore, I define “Lie” as the seller do not tell the buyer the true condition of the product and “Truth” as the vendor claims the actual colour of the card. Finally, I define “Success” in lie detection in this paper as the buyer choose the right colour of the card of 2-minute chat. Otherwise, the choice of this customer is defined as “Fail”.

For the quantifiable cues, because the experiment in this paper happens in Netherlands and the contents of the communication contain both English and Dutch, when I have the definition for the keywords, both English and Dutch words are taken into account.

There are three main potential factors which can influence the trustworthiness in my hypothesis one. In the following sections, I will first analyze them separately and then see the correlation relationship and how they affect the deception detection together.

5.2.1 Money Mention

I define the “Money Mention” in each chat as how many times the money-related words such as“30, 35, 5, 12, 18, 0, 6, euro/euros, money/geld,” showed in the conversation between one seller and one buyer in the experiment.

The summary of Money Mention as the table below:

Variables N Mean Standard Errors _Min _Max

(24)

23

The table above shows the number of observation, mean, standard deviation, minimum and maximum of money mention in the effective datasets. We can see on average, 1.16 money-related words are mentioned in each communication phase. And there are some chats which do not mention this kind of words at all and some people talk about payoff as many as 13 times.

I want to test the difference between the numbers of money-related words in different trustworthiness situations. More specifically, there are 3 different situations for each dialogue. First of all, the seller’s claim decides whether the situation is “Lie” or “Truth” (see above definitions). Secondly, the buyer’s choice determines he/she is in a “Trust” or “Not Trust” situation. Finally, we can see from the seller’s claim and the buyer’s choice to decide the deception detection is “Success” or “Fail”. To do this, I split the original data according to the buyer’s trust choice after the communication into two groups: “Trust” group and “Not Trust” group. Then I calculated the means and standard deviation of the “Money Mention” in two different groups because I want to see whether more payoff related words make buyers more likely to trust sellers. The last step of this part is that I use the two-sided Mann-Whitney tests to test whether the means of two groups have a significant difference. The buyer’s trust choice is the dependent variable. And the significant level of this test is 5%. The Mann-Whitney test is suitable for this situation for following reasons:

1 The contents in chats in this game are independent due to the experiment design.

2 In this case, I do not assume that the dependent variables conform to any certain distribution.

After the analysis of the “Trust” situation, I repeat the process for both “Lie/Truth” and “Success/Fail” situations. And the null hypothesis in the Mann-Whitney tests for different groups is

Money Mention in “Trust/Lie/Success” group= Money Mention in “Not Trust/Truth/Fail” group.

(25)

24

Table 3: Comparison of the “Money Mention” in Different Groups

Observations Mean Z stat

Group Trust Not Trust Trust Not Trust

Money Mention ₁₂₀₈ ₄₃₉ _1.27 (1.84) 0.87 (1.47) -3.905*** 0.0001

Group Lie Truth Lie Truth

Money Mention ₅₆₆ ₁₀₈₁ _1.02 (1.61) 1.23 (1.83) -1.985** 0.0471

Group Success Fail Success Fail

Money Mention ₁₁₀₂ ₅₄₅ _1.24 (1.86) 0.99 (1.53) -2.333** 0.0196 Stand errors are reported in the parentheses. The Z statistic derives from two-sided Mann-Whitney tests. ***indicate p<0.01 two tailed tests and ** indicate p<0.05 two tailed tests.

As shown in Table 1, we have evidence that buyers are more likely to trust sellers when they talk about money more and it makes sense because liar talks significant less about the payoff and benefit. More specifically, around 0.4 more money-related words show in the situation where customers trust vendors than the situation that sellers do not get trust. And this result is statistically significant at 1% significance level (two-sample Mann-Whitney U test, Z= -3.905 and P=0.0001). In addition, in those conversations where the buyer at the end of the game detect the deception successfully, the money-related words are mentioned significantly (Z=-2.333, P<0.05) more than fail situation too. However, we also can see from the table 1, when sellers tell the truth, the communication involves about 0.21 more earning-related words (Z=-1.985, p<0.05).

This result may indicate that when vendors talk more about financial benefits, buyers are more likely to believe what the sellers tell them. And the number of money-related words in one conversation can be a suitable cue to help buyers for deception detection. The possible explanation behind this maybe is when sellers tell the lie, they tend to talk less about money-related words.

(26)

25

5.2.2 Encompassing Words Mention

“Encompassing Words” is defined as in one single dialogue of the experiment in this paper, how many times have encompassing words such as “we/wij, us/ons, our/onze, lets/let’s/laten, both/beide/allebei, together/samen” been mentioned.

Similar to the analysis for “Money Mention”, I first see the descriptive statistics of the encompassing words.

Variables N Mean Standard Errors Min Max

Encompassing Words 1647 1.17 1.37 0 8

We can see people’s chats involve an average of 1.17 encompassing words in this game. And the times they talk about “we or us “and other similar words vary from 0 to 8.

To test the second factor’s influence on different situations, I do the similar analysis process as money mention of encompassing words. Firstly, I group all the 1647 conversations according to “Trust/ Not Trust” ,“Lie/Truth” and “Success/Fail” results separately. Then I get the means and standard deviation of encompassing words in different groups. Lastly, I run the two-sample Mann-Whitney U test to see whether there is a significant difference at 5% significance level in different situations because I do not assume this variable submit any certain distribution.

(27)

26

Table 4: Comparison of the “Encompassing Words Mention” in Different Groups

Stand errors are reported in the parentheses. The Z statistic derives from two-sided Mann-Whitney tests.

As shown in Table 4, it is difficult to say that more encompassing words make buyers trust sellers more although the mean of this kind of words in the trust group is slightly higher(0.02 more) than the not trust group. The occurrence times of encompassing words are statistically identical (Two-sided Mann-Whitney tests, Z= 0.034, P=0.9727) in these two groups. We also can see from Table 2 although the means have a bigger difference (0.13 compared to 0.02) in “Lie/Truth” groups and “Success/Fail” groups than “Trust/Not Trust” groups, the mention times of encompassing words are still insignificantly different at 5% significant level. In two-sample Mann-Whitney test, the Z-stat value is 1.019 and the p-value is 0.3081(>0.05) in the “Lie/Trust” groups and the in the “Success/Fail” groups, Z-stat value is 1.384 and p=0.1665(>0.05).

Based on the results above, the mention times of encompassing words seems cannot be the cue for buyers to detect whether sellers tell the truth or not. They also show no difference between the “Trust” and “Not Trust” groups. The potential explanation is that in this experiment environment which is a certain university, and subjects may already have encompassing feelings because they are of similar age and from the same university. Then the effect of encompassing words is small.

Encompassing Words 1208 439 1.18 (1.17) 1.16 (1.35) 0.034 0.9727

(28)

27

5.2.3 Word Length

I define “Word Length” as how many words the seller type in one two-minute informal conversation via online communication tool.

We can see from the below table for a summary of the word length.

In these 1647 dialogues, the mean of the words for every seller is 38.80. Sellers’ word lengths vary from 0 to 104 and the large standard deviation (18.35) indicates that there are big inter-group differences in their word length.

The next step is to do the same analysis as what I do for the previous two factors. Again I calculate the means and the standard deviations for the word length in different groups and then run the two-sample Mann-Whitney test to test whether the word lengths have a significant difference between groups at the 5% significance level.

Table 5: Comparison of the “Word Length” in Different Groups

Word Length 1208 439 39.58 (18.58) 36.65 (17.55) -2.774*** 0.0055

Word Length 566 1081 36.95 (17.79) 39.77 (18.57) -2.760*** 0.0058

Word Length 1102 545 38.76 (18.33) 38.89 (18.42) 0.114 0.9089 Stand errors are reported in the parentheses. The Z statistic derives from two-sided Mann-Whitney tests. ***indicate p<0.01 two tailed tests.

Variables N Mean Standard Errors Min Max

(29)

28

As shown in table 5, people really tend to believe sellers who talk more since the mean of word length in the trust group is significantly larger than it in the not trust group (two-sided Mann-Whitney tests Z=-2.774 and p-value= 0.0055). Meanwhile, when sellers talk more, it seems that they are telling the truth. We can see it from the mean of word length in the Truth group is significantly bigger than the Lie group at 1% significance lever(Z=-2.760,p=0.0058). However, the word length fails to help buyers to have an accurate judgement on the deception. The word lengths in the Success/Fail group are statistically the same.(Z=0.114,p-value>0.05)

People do trust others when they are talking a lot. It seems to be the right choice for buyers to judge whether the seller lies or not by how many words the seller tells in one chat. Sellers tend to tell more when they are telling the truth and the potential reason can be they are more confident when they do not lie. On the other hand, although sellers do talk more when they tell the truth, the word length is not helping buyers to detect deception in our experiment. The reason behind this can be buyers’ ability to detect lie become weaker when they get lots of information or there are some other factors in the conversation influence the buyer’s choice. We will see further combine analysis in later sections.

5.2.4 Correlation Analysis

From the results we have in former tables, we can have brief look at the three main factors which are mentioned in the hypothesis one from the graph one.

(30)

29

People are really tend to believe words where contains more money-related words (Z= -3.905, P< 1%) and longer words (Z=-2.744, p<1%) but the mention of encompassing words (P>10%) seems not have a significance enough effect on people’s trust decision.

Trustworthiness

We already see the difference on objective factors of written messages between “Trust” and “Not Trust” situation. In this part of the section, we attempt to find out if the quantitative cues discussed above have a strong impact on sellers’ trustworthiness , the extent to which that the buyer trusts the seller. To see the extent, my analysis is based on Tobit regression as Chen and Houser (2013)’s analysis in their research.

To begin, I define the dependent variable is the buyers’ trust decision with that 1 means buyers trust seller and 0 means sellers are not trusted and the three above-mentioned factors as the independent variables in this model. Lastly, there will be one constant to indicate the error term.

𝑌_𝑖 = 𝑋_𝑖′𝛼 + 𝜇

In the model above, Y (all the variables range from 0 to 1) is the buyers’ trust choice; 𝑋𝑖 is the vectors of observable quantitative cues of messagei, α indicates the average

response to the factors 𝑋𝑖 in message I and µ is the error term.

Trus t Selle rs Not Trus t Money Mention 1.27 0.87 Encompassing Words 1.17 1.16 Word Length 39.58 36.65 35.00 35.50 36.00 36.50 37.00 37.50 38.00 38.50 39.00 39.50 40.00 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 Mean of Word Length Mean of the Money Mention/Encom passiong words Groups

(31)

30

Before the regression starts, it is important to see that whether there are strong correlations between dependent variables for testing multicollinearity problem. If multicollinearity problem exists between any two dependent variables, which factor is really working on the independent variable will become very hard to say. The other consequence of this problem is that even there is a high general significance for this model; no single independent variable shows the significant effect in the model.

Specifically on our case, these three factors in one conversation are possible to affect each other simultaneously in reality. For example, written messages include more words have a better chance to mention money and encompassing words more times than messages have fewer words. Thus, we need to do a correlation analysis for independent variables before the regression.

The below correlation matrix shows the correlation rate.

Table 6: Correlation Matrix for 3 objective cues

Words Length Money Mention Encompassing Word

Words Length 1.0000 Money Mention 0.2635*** (0.000) 1.0000 Encompassing Words 0.2778*** (0.000) 0.1767*** (0.000) 1.0000

N=1647, the p-values of the correlation coefficient are reported in the parentheses. ***indicate that p<0.01 two tailed tests.

We can see from the table 6 that these variables have significant positive correlations. However, all the absolute values of the correlation coefficient are below 0.3 so we can consider they are not related to each other. The occurrence chance for multicollinearity problem in this model will be very low, so we can continue our regression process.

(32)

31

Dependent Variable: Buyer’s Trust Choice

Money Mention 0.0307*** (0.01) Encompassing Words -0.0110 (0.01) Word Length 0.0018** (0.00) Number of Observation 1647

Standard errors are in parentheses, ** and *** indicated significance at 5% ,1% level

We can see from the results that when people mention money once more in their dialogue , buyers are 3% more likely to believe in what sellers tell.(p-value=0)And when sellers speak 10 more words in one chat, buyers are 1% more likely to trust sellers. (P-value=0.04)In contrast, mention of encompassing words such as we/us and together do not have a significant impact (P=0.329) on buyer’s thoughts.

Deception Detection

Similar to the regression above, we also want to see whether these factors have significant power to affect buyers’ ability to detect deception, if so, to which extent. Thus, we keep all the design from the previous regression for trust except the dependent variable. I define the success of deception detection as dependent variable with that 1 indicates buyers’ success on lie detection and 0 means that buyers fail to detect deception.

Table 8 Quantitative Objective Cues and Deception Detection

Dependent Variable: Buyer’s Trust Choice

(33)

32 Money Mention 0.0312*** (0.01) Encompassing Words -0.0295** (0.01) Word Length -0.0003 (0.00) Number of Observation 1647

Standard errors are in parentheses, ** and *** indicate significance at 5% ,1% level

As shown in table 6, how many words does the seller tell in one chat seems do not help the buyer to find out the truth. However, mention of money words has a great positive impact on buyers’ ability to detect lies again. When the chat includes one more money-related word, buyers are 3% more likely to detect the truth successfully. On the other hand, mention of encompassing words has a negative effect on buyers’ detection ability. For example, messages with 5 more this kind of words make buyers’ success rate of deception detection decrease 14.75%.

In summary, people do have the ability to detect deception and they realize the function of these above-mentioned objective quantitative cues. However, not all of them have a great help on detect sellers’ deception or trustworthiness. Below are the results I conclude from all these effective data from experiment:

Result 1: People do trust sellers more when their conversation included more money-related words and when sellers have longer words but encompassing words do not really affect the sellers’ trustworthiness.

Results 2: Mention of money increases people’s ability to detect lie but the mention of encompassing words decreasing this ability. Unlike the case in trust, the number of words form seller does not affect buyers’ deception detection.

(34)

33

5.3 Effort Level

5.3.1 The Descriptive Statistics

The effort’s level evaluations of selected 15 original conversations in the experiment (10 are in English and 5 are in Dutch) are from 100 respondents who filled the survey mentioned before. Every respondent gives their thoughts on effort level and how much the seller of the chat can be trust in each conversation. Chat 1 to chat 5 are chats are in English and show in both two versions , and chat 6 to chat 10 are in English and only show in the English version. Chats 11 to 15 are in Dutch and they only show in the Dutch &English version.

You can see the descriptive statistic from the questionnaire data.

Table 9 : The Means For Effort and Trust in Each Chat

Effort Level Trust Extent N

Mean Standard Error Mean Standard Error

Chat 1 2.60 1.15 2.50 0.92 100 Chat 2 3.03 0.97 2.72 0.85 100 Chat 3 3.33 1.06 3.10 0.95 100 Chat 4 3.60 0.94 3.30 0.83 100 Chat 5 2.71 1.05 2.77 0.96 100 Chat 6 3.50 1.02 3.02 0.91 50 Chat 7 3.20 1.05 2.94 1.06 50 Chat 8 3.48 1.03 2.94 1.02 50 Chat 9 3.58 1.18 3.08 1.01 50 Chat 10 2.88 1.02 2.56 0.95 50 Chat 11 2.92 1.05 2.80 0.95 50 Chat 12 3.18 0.77 2.76 0.82 50 Chat 13 2.96 1.18 3.14 1.03 50 Chat 14 3.34 1.08 3.08 0.99 50 Chat 15 3.50 0.95 3.38 1.03 50

(35)

34

5.3.2 Correlation Analysis

The second hypothesis in this paper is as below.

Hypothesis 2: When buyers think sellers are really trying to sell a product (the effort level is high), purchaser are more likely to trust vendors.

To test this hypothesis, I first define the trust extent as the respondents’ answer to question”How convinced are you that this seller has a green card?” (1 is not trust at all and 5 is fully trust) and then I define effort level as the respondents’ answer from the survey to question one “In your opinion, how hard did this seller try to convince the buyer that the card was green?”(1 is the lowest effort level and 5 is the highest)

Then, when I try to find out the relation between effort level and trust extent, I do not group the data by their original dialogue like what I do for objective cues. In addition, I separate all the data by the trust extent 1 to 5. After grouping, I calculate all the means of the effort level for each group. The last step of the analysis is to run the Kruskal-Wallis test to see is there any significant difference.

Then I get the following results:

Table 10: Kruskal-Wallis equality-of-populations rank test Results by trust extent

Trust Extent Observation Mean of effort level Ranksum

1 78 2 (1.42) 19231.00

2 231 2.62(0.96) 82347.50

3 426 3.19( 0.83) 214775.50

4 219 3.81(0.83) 148239.00

5 46 4.24(0.92) 35907.00

The standard errors are reported in the parentheses

From the p-value (<1%), we can see there are really some significant differences on effort level from the groups by trust extent. However, only from Kruskal-Wallis e rank test, we cannot tell the specific difference in these groups of data. Thus, to see the detailed difference, we still need to run the Mann-Whitney test for every two groups.

(36)

35

Table 11: Z-value among different trust groups

Trust 1 Trust 2 Trust 3 Trust 4 Trust 5

Trust 1 / / / / /

Trust 2 -5.421*** / / / /

Trust 3 -8.198*** -7.697*** / / /

Trust 4 -9.118*** -11.967 -8.820*** / /

Trust 5 -6.981*** -8.322*** -7.629*** -3.625*** /

***indicate p<0.01 two tailed test.

Combined these two tables, we can tell that people are more likely to trust sellers who seem to try harder than others from significant Z-value and P-value. The potential reason behind this can be that we connect two good characters (honest and hard-working) of one person together. And this positive relations partly explain that the improvement on effort level increase salespeople’s performance. (Churchill et al.1985)This also explains why better-performing salespeople are generally those who work very hard at their tasks (Churchill et al. 1985; Sujan et al. 1994) Purchasers tend to trust sellers who work harder than others so there is a higher success rate in these cases.

Result 3: Buyers tend to trust sellers when they think the vendors are really trying hard to sell the products.

5.4 Combination Analysis with Objective Cues and Effort Level

In total, we have 15 samples which have data about both objective cues and effort levels. Effort level is the mean of the survey respondents’ evaluation for this chat. nword_S in the table means the number of words the seller tell. The chat contains how many encompassing words and money-related words show in the table as E_words and Money_Mention. Trust 1 is defined as whether the buyer (from the experiment) trusts the seller (show as 1 in the table) or not (0). Similar to this, I define Detection as whether the buyer detects the truth successfully (1) or not (0).Lastly, I define Trust 2 as the respondents’ average trust extent for each chat.

(37)

36

Table 12: Detailed data of 15 samples

Chat ID Effort Level

nwords_S E_words Money_ Mention

Trust 1 Detection Trust 2

1 2.6 40 0 0 1 0 2.5 2 3.03 31 1 4 1 1 2.72 3 3.33 61 3 6 1 0 3.1 4 3.6 63 3 6 1 1 3.3 5 2.71 15 0 0 1 1 2.77 6 3.5 76 3 0 0 0 3.02 7 3.2 19 1 0 0 1 2.94 8 3.48 34 0 0 0 1 2.94 9 3.58 47 6 0 1 0 3.08 10 2.88 38 2 1 0 1 2.56 11 2.92 25 0 0 0 1 2.8 12 3.18 42 4 2 0 0 2.76 13 2.96 38 2 0 0 1 3.14 14 3.34 51 0 0 1 0 3.08 15 3.5 44 2 3 1 1 3.38

Limited by the diversity of sample (15) and the number of variables (6), it is hard to do a statistical analysis. Normally, sample number for the regression’s should at least bigger than 50 (B. Green,1991, Nunnally1978)

Thus, I simply list the detailed data for all the 15 dialogues for future researches.

6 Conclusion and Discussion

The results from Mann-Whitney test for the different means and the Tobit Regression for all factors’ comprehensive influence degree indicate the great power of mention of money-related words. In informal written messages, mention of benefit-related words has a significantly positive impact on both buyers’ deception detection ability and seller’s trustworthiness. This result is consistent with Chen and Houser’s study in 2013. People are proved to be sensitive about the discussion of monetary factors. Sellers can add this kinf of words e on their advertisements to improve the trust from customers.

However, for the other two objective cues, my result shows some differences from the previous study. Buyers’ trust decision is not influenced by the mention of encompassing

Do quantitative cues and effort level influence people’s ability to detect deception and trustworthiness? : evidences from informal written messages

UNIVERSITEIT VAN AMSTERDAM