The influence of readability of demand letters on debtor response and the moderating role of debtor characteristics An exploratory research

The influence of readability of demand letters on debtor

response and the moderating role of debtor characteristics

An exploratory research

2

The influence of readability of demand letters on debtor

response and the moderating role of debtor characteristics

An exploratory research

Master thesis, Marketing Intelligence and Marketing Management.

University of Groningen,

Faculty of Economics and Business.

Department of Marketing

PO Box 800, 9700 AV Groningen (NL)

June 20, 2016

Esmée Heesters

s2208695

Korreweg 256a

9715 AP Groningen

tel.: +31 (0) 653897216

e.l.heesters@student.rug.nl

First supervisor/ University of Groningen

Dr. M. Keizer

Second supervisor/ University of Groningen

Dr. J. T. Bouma

Supervisor/ Field of study

3

Management summary

A long lasting problem in the debt collection industry is the lack of response, in payments or creditor engagement, of debtors (Lea, Webley and Walker, 1995; Mewse, Lea and Wrapson, 2010). Creditors try to motivate debtors to respond by sending them demand letters. However, currently 37% of the Dutch population has difficulties comprehending the average written communication used by companies and government (Bohnenn, Ceulemans, van de Guchte, Kurvers and van Tendeloo, 2004). The readability of a written text is ‘the ease with which a reader can read and understand a text’ (Oakland and Lane, 2004). The influence of a more readable text has not yet been scientifically tested in the debt collection industry. To get more insight in this problem the potential moderating effect of the debtor characteristics age, social class, amount of debt and gender were also explored.

A comparison was made between the demand letters of 2014 and 2016. Each year was divided into separate analyses for two subsequently sent demand letters, a charge free demand letter and a first demand letter. For the analysis of first demand letter the additional variable, whether or not a charge free demand letter was received by the debtor was included. Data of 10.084 records from a healthcare insurer was collected by a nationwide operating creditor and analysed with a type-II tobit model and a logit model.

The results of these models revealed a negative effect of a more readable demand letter. In addition, only the amount of debt, gender and receiving a charge free demand letter had moderating effects. Without considering the readability, it can be determined that age, amount of debt, social class and gender have significant main effects. According to these results, using a more readable letter cannot be recommended. However, it is advised to investigate how creditors should approach male and female debtors with less or more debt regard to these letters. With regard to the main effects, the creditor is advised to segment the debtor population based on the characteristics that were found to have significant influence. While the validity of the study can be questioned due to limitations in the study design and data, this study is an important first step towards more insight in the relation between readability and debtor response.

Key-words: readability, debtor payments, creditor engagement, debtor response, debtor

4

Preface

During a research project I was introduced into the world of debtors and debt collection. While this got my attention I decided to write my master thesis for a debt collection agency. I am grateful to this debt collection agency for giving me the opportunity of writing my thesis in a dynamic organisation. I feel that I had the opportunity to learn a lot about the debt collection process and how to carry out an academic research project in a practical environment.

First, I want to thank my colleagues and supervisor of the debt collection agency for all their support and dedication during this process. Secondly, I would like to thank my academic supervisor Dr. M. Keizer for all his valuable feedback and reassurance during the process of writing this thesis. Thirdly, I am grateful to Dr. J. T. Bouma for helping me find the internship and for his feedback during the start of my research project. Furthermore, I want to thank Dr. H. Risselada for providing me with feedback about the analysis. Lastly, I would like to thank my friends, family and especially my roommate Joana, brother Marijn and my boyfriend Erik for their feedback and support during the last five months.

5

Table of Contents

Management summary ... 3 Preface ... 4 1. Introduction ... 6 2. Theoretical framework ... 8 2.1 Debtor response ... 8 2.2 Readability ... 11 2.3 Debtor characteristics ... 14 2.4 Conceptual model ... 17 3. Method ... 18 3.1 Study design ... 18 3.2 Subjects ... 21 3.3 Variables ... 22

3.4 Plan of analysis and model specification ... 27

3.5 Model criteria ... 30 3.6 Estimation ... 31 4. Results ... 32 4.1 Pre- study ... 32 4.2 Exploratory analysis ... 32 4.3 Payments ... 34 4.4 Debtor Engagement ... 40 5. Discussion ... 42 5.1 Readability ... 43 5.2 Debtor characteristics ... 45 5.3 Implications ... 49

5.4 Limitations and future research directions ... 50

5.5 Conclusion ... 52

6

1. Introduction

When debtors do not pay their debts they receive demand letters from debt collection agencies, the creditors. In these letters the creditor asks them to pay their debts or contact them when they do not have the means to pay straight away. However, these letters are often written in an official and complex language containing a lot of terminology. The next situation may follow: the debtor does not understand what is asked of him or he does not comprehend the consequences when he does not comply with the request. Therefore, the payment is postponed and the creditor is not contacted. This will ultimately lead to a higher debt due to increasing collection costs without revenue for the creditor.

The situation described above indicates the importance of clear communication between the creditor and the debtor. According to the webpage of BureauTaal 60% of the Dutch people cannot comprehend the commonly used language in the written communication of the government and companies. The Common European Framework of References of Languages developed a measurement scale for reading comprehension levels of written text, ranging from very basic reading skills (level A1) to near-native written language fluency (level C2) (Alderson, Figueras, Kuijper, Nold, Takala and Tardieu, 2004). When measuring the reading comprehension level of the Dutch population an average of B1 is recorded, while for the average text of the government or companies the much higher C1 comprehension level is needed. This implies that a large part of the population has difficulties understanding the average communication of the government and companies (BureauTaal; Velleman and van der Geest, 2014).

7 The Common European Framework of References and Languages can only be used to measure reading comprehension levels of people and is not suitable for measuring the language level of written communication (Jansen, 2005). Therefore, the term readability will be used in this study. The readability of a written text is ‘the ease with which a reader can read and understand a text’ (Oakland and Lane, 2004). Evidently, a more readable text indicates an easier text.

The influence of a more readable text is not yet scientifically tested in the debt collection industry. Such scientific testing is of interest for creditors because when debtors have more understanding of the communication, it could mean that creditors receive more payments (sooner). In addition other companies with written communication that is difficult to read could benefit. For example, financial or legal firms often use complex language with a lot of terminology (AFM, 2012). A more readable text would increase the understanding (Jansen, 2005). Moreover, debtors may also benefit from this research. When debtors reduce their debt by responding to the creditor, they will also reduce the risk of negative consequences of this debt like psychological distress (Bridges and Disney, 2010; Dunn and Mirzaie, 2016; Jenkins, Bhugra, Bebbington, Brugha, Farrell and Coid, 2008) and physical distress (Lavrakas and Tompson, 2009; Lyons and Yilmazer, 2005).

To make sure debtors pay their debt, creditors contact debtors with the demand letters. Yet, debtors often do not respond to this communication (Lea, Webley and Walker, 1995; Mewse, Lea and Wrapson, 2010). Based on the statements listed above, the debtors misunderstanding of the written communication could be a contributing factor to why debtors do not respond to the written communication. In this scientific research the following question will be investigated:

RQ1: What is the influence of more readable written communication sent out by creditors on the level of debtor response?

8

RQ2: Does the number of debtors who respond to a more readable letter, differ between debtors with different characteristics?

In this paper a conceptual model is presented along with definitions of the relevant constructs. A theoretical framework will be introduced after which the research design and results will be presented. Finally, a discussion, implications and limitations of these results will be discussed.

2. Theoretical framework

In the following section the different relevant constructs will be discussed. First the dependent variable, debtor response, is defined and a relevant theoretical background is given. Next, readability is defined and a hypothesis is provided. This is followed by a discussion of the debtor characteristics and control variables. Lastly, a visual representation of the conceptual model is presented.

2.1 Debtor response

When debtors have a higher understanding of the written communication of creditors the debtor response could be affected. Following MacDermott (2008) and Mewse, Lea and Wrapson (2010), in this study debtor response is defined as: the debtors’ decision to pay or engage with the creditor. Making a payment is important information for the creditor as that is their core business. However, this research does not only focus on the payments that are made but also whether the creditor was contacted, creditor engagement. The importance of contacting the creditor is stressed by The Dutch Ministry of Social Affairs and Employment (2008). They recommend debtors to first get insights in their financial situation. Debtors can accomplish this by making an overview of their earnings and expenses and identifying their total amount of debt. The next step is to contact the creditors, when they cannot pay the whole debt at once, to make an agreement about a payment plan and follow it. The importance of this creditor engagement on the way to becoming debt-free is also stressed by Ford (1988) and MacDermott (2008).

9 inform the debtor that he has to pay his debt or contact the creditor within a certain period of time. The creditor asks the debtor to contact the creditor when the debtor does not have the means to pay the debt at once immediately. When the debtor contacts the creditor they can make an agreement about a payment plan, so the debtor can pay in multiple instalments. When the debtor fails to pay or contact the creditor after the first letter, his debt will increase due to collection costs. When the debtor fails to pay or contact after multiple letters the court becomes involved. Multiple legal stages will follow which are accompanied by extra charges. For example, charges for drafting and handing over the summons, which is a written call that the debtor has to appear in court. Thus, by not responding to the creditor the initial amount of debt increases. It can therefore be concluded that an early response is also greatly beneficial to debtors.

Thus while response is important, previous research shows that debtors do not automatically respond to creditors when they have financial difficulties. Based on the fact that an advice to contact the creditor needs to be given, it could be derived that debtors do not automatically contact the creditor (Mewse, Lea and Wrapson, 2010). In addition, Lea, Webley and Walker (1995) indicated that debtors were restrained in approaching creditors and reported them as one of the less useful sources of help. Thus, it can be concluded that debtors are restrained in engaging with the creditor.

10 and Tofigbakhsh, 2000) or when the ink colour makes the statement easier to read (Reber and Schwarz, 1999).

Of these two metacognitive experiences, bringing information to mind and processing information, this study focuses on processing fluency, the subjective experience of ease with which people process information (Alter and Oppenheimer, 2009). Processing fluency can be influenced by a large number of variables. Including the duration of a presentation, clarity with which a stimulus is presented, or the figure- ground contrast of a stimulus (Schwarz, 2004). An improvement in this processing fluency is associated with multiple positive dimensions like higher judgement of truth (Reber and Schwarz, 1999), confidence (Norwick and Epley, 2002), frequency (Tversky and Kahneman, 1973), fame (Jacoby, Kelley, Brown, and Jasechko, 1989, and liking (Reber, Winkielman, and Schwarz, 1998). This suggests that adjusting the processing fluency can lead to positive judgements.

According to Song and Schwarz (2008) processing fluency is also related to effort prediction and motivation. In their study they showed that their participants used the experienced processing fluency as a predictor for the necessary effort needed for a specific task. Participants that read the instructions in the low processing fluency condition predicted higher levels of required effort and were less motivated to engage in the task. According to Dreisbach and Fischer (2011) the results of Song and Schwarz (2008) fit perfectly with the rather historic idea of law of least effort (Hull, 1943). This idea proposes that when people find themselves in a difficult processing situation they will switch activities and will avoid this activity in the future.

11 are more likely to engage in the desired behaviour when the perceived effort needed for this behaviour is lower. Consequently, this perceived effort can be lowered by improving the processing fluency. Although no research is conducted in this field, it is likely that debtors will be more inclined to respond when the processing fluency of the demand letters is lower. According to Oppenheimer (2006) one way to increase the processing fluency is by writing an easier text. This would suggest that debtors are more likely to respond to a more readable letter.

2.2 Readability

The readability of a text is generally viewed as the perceived difficulties within a text and the ease with which a reader can read and understand a text (Oakland and Lane, 2004; Wray and Janan, 2013). A more readable text includes less perceived difficulties in the text and is easier to understand. The readability of a text consists of multiple aspects. According to Oakland and Lane (2004) the classical method to assess readability consists of two surface level features: vocabulary and syntax. The vocabulary can be defined as ‘the meaning of words’. This can be assessed by word familiarity and/or the number of letters within a word. Syntax is the language rules whereby words or other elements of sentence structures are combined to promote reading. The syntax is typically assessed by sentence and paragraph length and/or sentence and passage complexity. The number of letters within a word and the number of words in a sentence can explain around 60% of the variance in comprehension (Jansen and Boersma, 2013). Besides, according to Chall and Dale (1995) these components are sufficient to make good predictions about readability. Thus, by adjusting the vocabulary and the syntax the readability of a written text can be influenced.

12 However, some studies suggest that a more readable letter may also have no effects or adverse effects on debtor response. First, according to Schwarz (2004) and Alter and Oppenheimer (2009) individuals are not influenced by processing fluency once they are aware of the source and content. In the study of Rennekamp (2012) findings suggest that not all people react the same to processing fluency. According to Rennekamp subjective perceptions of processing fluency vary with knowledge and experience. This leads to potential differences in how experienced and in- experienced individuals react to the processing fluency of certain information. It could be that a debtor often receives demand letters, so he knows what will be written in the demand letter. This makes him more experienced with the demand letters and this could lead to no influence of the more readable demand letter.

Second, the study of Atkinson and Carskaddon (1975) suggest a possible negative effect of a more readable letter. Their study took place in the mental health context and reported that the use of technical jargon of counsellors improved credibility and facilitated behaviour change. The counsellors who used technical jargon were perceived as having greater knowledge of their field than counsellors who used layman’s language. Subjects may have felt that, although they did not understand the jargon, the counsellors using this jargon had more knowledge about psychology than they did themselves. With the use of layman’s language, they may have felt that the knowledge of the counsellor was no greater than their own. In addition Givens (2015) states that technical jargon offers increased credibility, enhanced authority and greater power for counsellors.

13 receive a letter with a more complex language, are more likely to follow the instructions in the letter. Debtors could have the feeling that the knowledge of the person writing the letter is greater than their own and will therefore follow the instructions. This suggests that a more readable letter may decrease the debtor response.

Although awareness of source and content or authority could play a role, there is another principle which supports a positive effect of a more readable letter for the population of debtors, the principle of scarcity. According to Mullainathan and Shafir (2013:p. 8) scarcity is ´the subjective sensation that one has less of something than one needs, that captures the mind´. For example, this could be a lack of money, time, friends or food. Scarcity influences human cognition, choices and behaviour because the mind orients automatically and powerfully to unfulfilled needs (Mullainathan and Shafir, 2013). Research of Mani, Mullainathan, Shafir and Zhao (2013) demonstrate a relation between poverty, measured in budgetary preoccupations, and cognitive function. People in poverty have access to less cognitive resources because the context of poverty occupies and hinders cognitive capacity. ‘Poverty-related concerns consume mental resources, leaving less for other tasks’ (Mani, et al, 2013:p. 978). This cognitive impact is about the same size as losing a full night of sleep. Therefore, Mani et al. (2013) propose that policy makers should avoid composing cognitive load on the poor. Debtors who receive a demand letter will have budgetary preoccupations, and therefore limited cognitive resources. To avoid composing cognitive load on the debtors, the demand letters should be easy to understand. These letters need a high processing fluency which, as previously stated, could be accomplished by a more readable letter. Therefore, this study expects a positive effect of a more readable letter. The following hypothesis is proposed:

H1: A more readable demand letter will lead to more debtor response.

14 comply better with the needs of the debtors (Athanassopoulos, 2000). Therefore, the next section will focus on debtor characteristics.

2.3 Debtor characteristics

In segmentation homogenous groups of customers are identified with similar needs and preferences, which can be targeted in the same manner (Webel, 2002). An old but still accurate definition according to Webel (2002) is that of Smith (1956). Smith defines market segmentation as ‘viewing a heterogeneous market as a number of smaller homogeneous markets, in response to differing preferences, attributable to the desires of customers for more precise satisfactions of their varying wants’ (p.4). Thus to better react to the desires of the customers, companies divide the market into different groups. Segmentation has proven to be a very useful concept to managers and they seem comfortable with the idea of market segments (Webel, 2002). In this study, debtor characteristics can be used to segment the market.

2.3.1 Debtors compared to non-debtors

Before segmentation, the profile of the prototypical debtor is identified to get an impression of the population and to see whether debtors often have similar characteristics compared to non-debtors. In table 1, an overview of the demographic characteristics of the prototypical debtor, as indicated by different researches, is presented. First, two recent studies of the Dutch population are displayed. According to Panteia (2015) the Dutch debtors can be characterized based on four characteristics: net household income, household composition, presence of children and homeownership. Age, gender and education level did not differ significantly between debtors and non-debtors. The second study of Nibud (2015) indicated differences in age, gender, education, children, income, homeownership and household composition. Due to the contradictions in these findings and to increase the insights about the prototypical debtor a less recent third study among the Dutch population is displayed. According to Webley and Nyhus (2001: p. 424) ‘the prototypic debtor is a young, single parent living in a rented accommodation’. In their research they found that debtors have lower incomes, are less likely to own their own houses, are less likely to have a partner, have more children, and tend to be younger.

15 non-debtors. This makes segmentation even more relevant because debtors are clearly a heterogeneous group.

Table 1: Characteristics of the prototypical debtor

Panteia (2015) Nibud (2015) Webley and Nyhus

(2001)

Age - 45-64 Young

Gender - Men -

Education - Low -

Children No children With children More children

Income Low Low Low

Homeownership Rental Rental Rental

Household composition Cohabit Single Single

2.3.2 Engaging debtors compared to non-engaging debtors

16 discussed in this study. However, the variables age, education, income and amount of debt will be discussed.

2.3.3 Segmentation criteria

In this section the segmentation criteria age, education, income and amount of debt are discussed.

Age

According to Duyan-Bump and Grant (2009) and Sullican and Fisher (1988) the number of payments to the creditor increases with age. The study of Canner and Luckett (1991: p. 59) provided similar results: ‘households headed by people under age 35 were nearly four times as likely to report payment problems as were those headed by an individual at least 55 years of age’. Unfortunately, there is no indication in the literature that age has an influence on the decision whether to engage with the creditor. While age has an influence on payments, it is included in this study to examine whether the effects of a more readable letter are influenced by the age of the debtor.

Education

Payments are more likely to occur among households with higher-educated heads (Sullivan and Fisher, 1988). In addition, the research of Mewse, Lea and Wrapson (2010) states that debtors with higher educational qualifications engage more with the creditors than debtors with lower educational qualifications. So higher educated debtors are more likely to respond. In addition, education is included in the research while it would be expected that a more readable letter could help motivate the lower educated debtors to respond. More intelligent authors are more inclined to use complex vocabulary (Pennebaker and King, 1999) which would indicate they have a higher reading level. Lower educated debtors would have a lower reading level and would therefore benefit more from a more readable letter.

Income

17 income will be incorporated in this study to see whether differences in income affect the relation between readability and response.

Due to data limitations it is not possible to look at both education and income separately. Therefore, these two variables will be combined in the concept of social class. Social class is defined as ‘the position of a person or group of persons occupies in society and it designates access to social and economic resources and to valued life experience’ (Daniel, 1984: p. 218). Education and income are commonly chosen indicators of social class (Daniel, 1984). Therefore, social class will be incorporated to see if differences have influence on the relation between readability and debtor response.

Amount of debt

The research of Livingstone and Lunt (1992) shows that payments are predicted by the money owed. They state that the more one owes, the more one pays, provided they have the resources to do so. As is established in table 1, debtors often have a low income and thus limited resources to do so. In addition, according to Mewse, Lea and Wrapson (2010) debtors with lower amounts of debt engage less with the creditor than debtors with higher amounts of debt. Therefore, it would be interesting to see whether amount of debt has an influence on the relation between readability and debtor response.

Based on the findings above, it can be concluded that there are behavioural differences among the debt population. Not every debtor reacts to their debt in the same way. It is likely that there will be different responses on the more readable demand letter and therefore these debtor characteristics should be considered. Additionally, according to the literature gender does not influence the debtor response but will be included in this research as a control variable.

2.4 Conceptual model

18

3. Method

This section starts with an elaboration about the design of the study. After which the subjects, the measurement of the variables, the plan of analyses, model criteria and model estimation will be discussed.

3.1 Study design

The study took place in the debt collection industry. The field experiment was conducted in collaboration with a nationwide-operating creditor based on the readability of demand letters send to debtors of Dutch healthcare insurer. A comparison was made between demand letters of 2014 and 2016. Each year was divided into two separate analyses: one for the charge free demand letter and one for the first demand letter, explained further on. Consequently, it could be explored whether the effect of readability and the debtor characteristics differ between the charge free demand letter and first demand letter. This created more insight in whether the approach of the creditor should be the same for both letters. This quantitative study was based on the debt collection data collected by the creditor.

Processing fluency Debtor response Readability demand letter + + H1: + Debtor characteristics Age Social class Amount of debt

Gender (Control variable)

19

3.1.1 The debt collection process

The debt collection process consists of two consecutive phases: the amicable phase and the legal phase. This study focuses on the effects of changes in readability in the first phase of the process. The amicable phase was chosen because this is where the creditor can generate the most profit. The creditor can immediately charge the collection costs to the debtors, otherwise they have to pay these costs in advance themselves. Thus, an earlier response is in the interest of both the creditor and the debtor. While as mentioned previously, if debtors respond before involving the court this can prevent the debt from increasing.

During the amicable process three demand letters were sent to the debtors, see figure 2 and 3. The demand letters as sent to the debtors can be found in appendix 1 and 2. Of the three letters sent, the first letter was free of charge (charge free demand letter) in which the debtors were asked to pay their debt or to contact the collection creditor and if they did so within 14 days there were no collection charges. If they did not pay or contact the creditor, they received a second letter (first demand letter) in which they were asked to pay their debt and the collection costs within 8 or 14 days. When there was still no response they received a third and final letter (second demand letter) telling them to pay or contact within 5 days. When they did not respond after three letters the court became involved and the legal phase started.

Figure 2: Process of 2016; * indicates the payment terms.

Figure 3: Process of 2014; * indicates the payment terms.

20

3.1.2 Unit of analysis

The data from the more readable demand letters (2016) was compared to the data of demand letters of two years ago (2014), for a period of two months. The letters of 2014 were chosen because in 2015 the creditor already tried to make the letters more readable. The difference in readability between 2014 and 2016 was higher and therefore more interesting to analyse the effects. In both years the letters sent in the period of March and April were analysed. The same months were chosen because this eliminates the effect that different times of the year can have on spending behaviour (Nibud, 2013).

From figure 2 and 3 can be derived that the process changed after the first demand letter was sent. In 2014 only a text message was sent after sending the first demand letter while in 2016 a text message and an email were sent. To account for this difference in the process, the time frame of the analysis was adjusted, only the debtor response of the first two letters was analysed.

21 While the data was collected until the 8 May the charge free demand letters that were sent from 1 March until 24 April for both 2014 and 2016 were included in the analysis. This limit is set 14 days before 8May to give the debtors enough time to respond. In addition, the first demand letters that were sent from 1 March until 30 April 2014, nine days before 8 May, and 1 March until 3 May 2016, five days before 8 May, were included for the same reasons.

As mentioned previously, separate analyses for the charge free demand letter and the first demand letter were performed. The charge free demand letter should be sent 14 days before the first demand letter however, this is not always the case. Sometimes the charge free demand letter was sent as much as a year ago. Not all the debtors who received a first demand letter also received the charge free demand letter in the specified time frame as just introduced. Therefore, an additional debtor characteristic for the analysis of the first demand letter is used, namely whether or not a debtor received a charge free demand letter.

3.2 Subjects

The target population of this study were debtors in the Netherlands who receive demand letters. The subjects for the sample were chosen with purposive sampling, they were the debtors of a health insurer in the Netherlands. All the debtors that received a charge free demand letter in the period from 1 March until 24 April 2014 or 2016 or a first demand letter in the period of 1 March until 30 April 2014 or 3 May 2014 were analysed. This sample consisted in total of N=10,084 records and 9,836 debtors, while some debtors have multiple records. Having multiple records indicated that some debtors had multiple debts or debts in both 2014 and 2016.

Figure 4: Distribution of debtors paid in 2016 within 8 days Figure 5: Distribution of debtors paid in 2014 within 14 days

22 3.3 Variables

In the following section the measurement of the different variables is explained. First, the dependent variable debtor response is discussed, followed by the independent variable readability and the moderator debtor characteristics. Lastly, an overview of the variables is displayed (see table 5).

3.3.1 Debtor response

The debtor response is measured based on multiple indicators. The first one is payments, which was measured based on whether a payment is made, yes or no, and the amount of the payment in euros. The second indicator is creditor engagement. This is measured by whether the debtor had contacted the creditor, yes or no.

In table 2, the number of letters send and the debtor response is displayed. In this table the average amount that the debtors paid to the creditor, the percentage of dossiers were a payment was made and percentage of dossiers were the creditor was contacted for the charge free demand letter and the first demand letter of 2014 and 2016 is presented. The total letters column shows the average amount paid, how often a payment was made and creditor engagement for both letters. The total records column shows the percentage of records in which a payment or creditor engagement occurred, after either the charge free demand letter or the first demand letter. From the table can be derived that in all cases the debtor response is higher in 2014.

Table 2: Overview of debtor response

23

3.3.2 Readability

To create more readable demand letters, the creditor adjusted the letters of 2016. The studies of Oakland and Lane (2004) and Meyer (2003) name text factors to create a more readable text and several of these were used by the creditor. The syntax was adjusted by decreasing the sentence length. The vocabulary was adjusted by using less terminology and less complex words. In addition, headings were included to create a more structured text. According to the Securities and Exchange Commission (SEC) (1998) and as used in the research of Rennekamp (2012) including headings will increase the readability of a text.

To ascertain that the adjustments create more readable letters, the readability level was determined with help of readability formulas. These formulas provide an easy and quick way to predict readability (Meyer, 2003). There are over 30 formulas developed to estimate text difficulty. The majority of these formulas are based on the vocabulary and the syntax as explained previously and developed for text written in English. However, there are also some formulas available that can be applied to texts written in other European languages. These are formulas which are not based on syllables but on the number of letters or characters. Formulas based on the number of syllables designed for English text are not reliable to use for other languages. The reason behind this, is that in the Dutch language, words often have more syllables than words in the English language (Douma, 1960). Therefore, in this study the Coleman- Liau Index and the Automated Reliability Index were used.

Coleman- Liau Index

The Coleman- Liau Index (CLI) is based on the number of characters per word and the number of words per sentence (Coleman and Liau, 1975). Which led to the formula displayed in equation 3.1.

𝐶𝐿𝐼 = 0.0588𝐿 − 0.296𝑆 − 15.8 (3.1)

24 formulas for Dutch texts. The results of this research indicated that the Coleman- Liau Index performs better as a predictor of readability than these online tools.

Automated Reliability Index

In addition to the Coleman- Liau Index, the Automated Reliability Index (ARI) was used. This formula was included to improve the reliability of the results by including a second metric. The ARI is the only other frequently used metric that does not take the number of syllables into account. The ARI is based on the number of characters per word and the number of words per sentence (Senter and Smith, 1967). The output of the ARI is also related to the U.S. grade level and therefore a higher score indicates a more difficult text (see appendix 3). The formula is displayed in equation 3.2.

𝐴𝑅𝐼 = 4.71 4567689:7;_<=7>; + 0.5 _@:A9:A8:;<=7>; − 21.43 (3.2)

At which Characters is the number of number of letters and numbers, Words is the number of spaces, and Sentences indicates the number of sentences.

In the seventies of the last century the classical readability formulas increased in popularity, which was soon followed by an increase of critique from the academic world. There are two often mentioned problems (Anderson and Davison, 1988; Connaster, 1999). The first one is that readability formulas give the impression that a text can be improved by adjusting simple text characteristics like word- and sentence length. For example, simply cutting one sentence into two at the comma does not necessarily means that the text is more readable (Land, 2009). Secondly, readability formulas only take the word- and sentence level into account. They ignore characteristics like coherence and structure. However, this does not mean that the readability formulas cannot be used. Results of multiple studies show that using word and sentence length as predictors of readability explain 60% to 70% of the variance in comprehension (Chall and Dale, 1995; Jansen and Boersma, 2013; Kintsch, 1979; Kintsch and Vipond, 1979). Consequently, while no perfect metric is developed, the readability formulas are a viable option to assess readability (Jansen and Boersma, 2013).

25 more understandable language, the readability formulas are more suitable to rank text than to give exact predictions. Therefore, this study only made statements about the relative readability level and not about the absolute values that were provided by the formulas.

Readability scores

To predict the readability of the demand letters with these formulas an online tool is used. De Clercq, Hoste, Desmet, van Oosten, De Cock and Macken (2012) of the University of Ghent developed a tool to calculate different readability scores. The parts of the letters as presented in appendix 1 and 2 are included in the calculation. The address, date, and overview of the claim are not included because the length of these words and sentences will differ in each letter. They will influence the scores but not make the text easier or harder to read.

The results of the calculations can be found in table 3. As can be derived from this table, the CLI and the ARI values of the letters sent in 2016 were lower than the values of the letters sent in 2014. According to the CLI, the letters in 2014 and the charge free demand letter of 2016 were above the maximum value of 10 (which indicates ninth grade). However, the ARI only showed values below 10 for all the letters in 2014 and 2016. Both metrics indicate that the text of the letters sent in 2016 had a lower readability score than the text of the letters sent in 2014. Therefore, it can be concluded that the letters of 2016 are more readable than the letters of 2014.

Table 3: Readability scores

CLI ARI

2014 2016 2014 2016

Charge free demand letter 11.20 10.84 7.10 6.83

First demand letter 12.76 8.90 8.50 5.20

26

3.3.3 Debtor characteristics

As mentioned previously, the debtor characteristics that were included in this study were age, social class, amount of debt, gender and whether a charge free demand letter was received. The age of the debtor was measured based on the birthdate of the debtor. A value for age was missing in 60 cases. In these cases, the missing values were replaced by the mean of that year. Correspondingly, the missing values of 2016, 36 cases, were replaced by 39.642 and the missing values in 2014, 24 cases, were replaced by 40.895.

The social class of the debtor was measured based on a status score. This was a score developed by the Social and Cultural Planning Creditor (Sociaal en Cultureel Plan bureau). This is a scientific institute which conducts social and academic research in the Netherlands. The status score was based on zip codes in the Netherlands. They indicated the social status of a district relative to other districts in the Netherlands. The status score was based on the average income, the percentage of people with a low income, the percentage of low educated people, and the percentage of unemployed people. This means that income and education are incorporated in this measurement. A higher status score indicated a higher social class. On average in the years 1998 until 2014 the status score in the Netherlands was 0. Therefore, the scores indicated how much the specific zip code was above (a positive number) or below (a negative number) this average of 0. In this study the lowest score was -5.38 and the highest score was 3.02 and the most recent status score, from 2014, was used.

27

Table 4: Descriptive statistics debtor characteristics

Variable 2014 2016 Age M: 40.895 SD: 14.252 M: 39.642 SD: 14.051 Social Class M: -0.2870 SD: 1.3121 M: -0.2656 SD: 1.3116 Amount of debt

Charge free demand letter

M: 353.06 SD: 552.12

M: 298.79 SD: 400.03 Amount of debt

First demand letter

M:392.09 SD:655.34 M: 441.21 SD: 765.41 Gender Male: 57% Female: 43% Male: 53.4% Female: 40.7% Received charge free demand letter Yes: 90%

No: 10%

Yes: 84% No: 16%

Table 5: Variable overview

Variable Variable description Notation

Readability level The readability level of the demand letters that the debtor received (0= _{2014, 1=2016)} 𝑅C,E

Age Age of the debtor (mean centred) 𝐴C,E

Social class The status score (mean centered) 𝑆𝐶C,E

Principal amount of debt The total amount of debt in euro’s the debtor has with this health _{insurer (mean centred and log)} 𝐷C,E

Gender Gender of the debtor (0=male, 1=female) 𝐺C,E

Received Charge free demand letter

Has the debtor who receives a first demand letter also received a

charge free demand letter (0=no, 1=yes) 𝐶𝐹C,E

Payment made Has the debtor made a payment (0=no, 1=yes) 𝑌_C,E

Amount of payment Amount of payment in euros 𝑦C,E∗

Creditor engagement Has the debtor contacted the creditor (0=no, 1=yes) 𝐶C,E

Note: 𝑖= record, 𝑙= the demand letter received (1= charge free demand letter, 2= first demand letter)

3.4 Plan of analysis and model specification

To analyse whether a more readable demand letter has a positive influence on the debtor response and the role of debtor characteristics in the relation, different statistical methods are used. This section elaborates on the motivation for using a type-II tobit model and a logit model and a specification of these models.

3.4.1 Type-II Tobit model

28 are observed. Therefore, using a regular regression would lead to biased estimates (Franses and Paap, 2001). Consequently, a censored regression model, like the tobit model, should be applied. While there are multiple types of the tobit model, in this case the type-II tobit model was most suited. Namely, the type-II tobit model allowed for variations between the influence of the independent variables on the decision to pay and how much to pay. This gave additional information with reference to the different influences of the independent and moderation variables on the dependent variables.

The type-II tobit model consisted of two stages. In the first stage the model explains whether a debtor pays or not, using a probit model. In the second stage, the amount of payment in euros, given that the debtor pays, is modelled with a truncated regression formulation. In this section two separate tobit models will be displayed for the charge free demand letter and the first demand letter.

Regarding the charge free demand letter, the payment incidence for each record 𝑖 denoted is as 𝑌_C,N. When the variable equals 1, the debtor has made a payment and when the variable equals 0, no payment was made. The second part of the model, the amount of payment ( 𝑦_C,N∗ ₎

is modelled for the positive values of 𝑌_C (so when a payment was made). The type-II tobit model should include different parameters and error terms in both models, which allows the influence of the explanatory variables to vary. Therefore the average effect of each of the parameters (𝛽N _{and 𝛽}P_{) and the error terms (𝜀}N _{and 𝜀}P_{) have a different notation in the two}

parts of the model. This leads to the following model:

𝑌C,N= 0 𝑖𝑓 𝑦C,N∗ = 𝛽NS+ 𝛽NN𝑅C,N+ 𝛽NP𝐴C,N+ 𝛽NT𝑆𝐶C,N+ 𝛽NU𝐷C,N+ 𝛽NV𝐺C,N+ 𝛽NW𝑅C,N𝐴C,N+ 𝛽NX𝑅C,N𝑆𝐶C,N + 𝛽N Y𝑅C,N𝐷C,N+ 𝛽NZ𝑅C,N𝐺C,N + 𝜀NC,N≤ 0 𝑌C,N= 𝛽P_S+ 𝛽P_N𝑅C,N+ 𝛽P_P𝐴C,N+ 𝛽P_T𝑆𝐶C,N+ 𝛽P_U𝐷C,N+ 𝛽P_V𝐺C,N+ 𝛽P_W𝑅C,N𝐴C,N+ 𝛽P_X𝑅C,N𝑆𝐶C,N+ 𝛽P_Y𝑅C,N𝐷C,N + 𝛽P Z𝑅C,N𝐺C,N + 𝜀PC,N 𝑖𝑓 𝑦C,N∗ = 𝛽NS+ 𝛽NN𝑅C,N+ 𝛽NP𝐴C,N+ 𝛽NT𝑆𝐶C,N+ 𝛽NU𝐷C,N+ 𝛽NV𝐺C,N+ 𝛽NW𝑅C,N𝐴C,N+ 𝛽NX𝑅C,N𝑆𝐶C,N + 𝛽N Y𝑅C,N𝐷C,N+ 𝛽NZ𝑅C,N𝐺C,N + 𝜀NC,N> 0 (3.3) where the meaning of the explanatory variables is represented in table 5 and the 𝛽N _{is the}

average effect of each of the parameters and 𝜀N

C,N is the error term. For the first demand

29 first demand letter and 𝑦_C,P∗ _{indicates the amount paid after the first demand letter. Only the}

additional variable, 𝐶𝐹_C,P, whether or not the debtor received a charge free demand letter before receiving the first demand letter (0=no, 1=yes) differs from the type-II tobit model for the charge free demand letter. This leads to the model displayed in equation 3.4.

𝑌C,P= 0 𝑖𝑓 𝑦C,P∗ = 𝛽NS+ 𝛽NN𝑅C,P+ 𝛽NP𝐴C,P+ 𝛽NT𝑆𝐶C,P+ 𝛽NU𝐷C,P+ 𝛽NV𝐺C,P+ 𝛽NW𝑅C,P𝐴C,P+ 𝛽NX𝑅C,P𝑆𝐶C,P + 𝛽N Y𝑅C,P𝐷C,P+ 𝛽NZ𝑅C,P𝐺C,P + 𝛽NNS𝐶𝐹C,P+ 𝛽NNN𝑅C,P𝐶𝐹C,P+ 𝜀NC,P≤ 0 𝑌C,P= 𝛽PS+ 𝛽PN𝑅C,P+ 𝛽PP𝐴C,P+ 𝛽PT𝑆𝐶C,P+ 𝛽PU𝐷C,P+ 𝛽PV𝐺C,P+ 𝛽PW𝑅C,P𝐴C,P+ 𝛽PX𝑅C,P𝑆𝐶C,P+ 𝛽PY𝑅C,P𝐷C,P + 𝛽P Z𝑅C,P𝐺C,P 𝛽PNS𝐶𝐹C,P+ 𝛽PNN𝑅C,P𝐶𝐹C,P+ 𝜀PC,P 𝑖𝑓 𝑦_C,P∗ _{= 𝛽}N S+ 𝛽NN𝑅C,P+ 𝛽NP𝐴C,P+ 𝛽NT𝑆𝐶C,P+ 𝛽NU𝐷C,P+ 𝛽NV𝐺C,P+ 𝛽NW𝑅C,P𝐴C,P+ 𝛽NX𝑅C,P𝑆𝐶C,P + 𝛽N Y𝑅C,P𝐷C,P+ 𝛽NZ𝑅C,P𝐺C,P + 𝛽NNS𝐶𝐹C,P+ 𝛽NNN𝑅C,P𝐶𝐹C,P+ 𝜀NC,P> 0 (3.4) Although it is not necessary to include the same explanatory variables in both parts of the type-II tobit model (Nierop, Leeflang, Teerling and Huizingh, 2011) it was done in this study as a starting point to assess the effects of all the explanatory variables. Though there is a potential for dependence between the two parts of the model, since it is possible that one part of the model uses all the variance which leaves no variance to be explained for the second part of the model, which could lead to insignificant variables. Therefore, it is important to be careful while incorporating the variables in the type-II tobit model. Especially for this study this is important, due to the low percentage of debtors who actually paid compared to the large percentage of debtors who did not pay. When including all the main variables and interaction variables in both parts of the type-II tobit model, it may jeopardize the validity of the available data. To account for this problem a probit and a linear regression model were estimated to get an impression of which variables are important to include in the type-II tobit model. Running a separate probit model followed by a regression is called the two-part model (Manning, Duan and Rogers, 1987). In the two-part model the same variables were included for the charge free demand letter and the first demand letter as specified in equation 3.4 and 3.5.

3.4.2 Logit model

30 variable, two models can be used, a logit or probit model. Both models use the cumulative distribution function and give similar results. In this study the logit model was chosen because it is easier to interpret than the probit model (Franses and Paap, 2000).

The logit model uses probabilities to predict creditor engagement. To predict the probability for creditor engagement 𝐶_C,N, after the receiving the charge free demand letter, the model in equation 3.5 is used.

𝐹(𝐶C,N)

= exp (𝛽S+ 𝛽N𝑅C,N+ 𝛽P𝐴C,N+ 𝛽T𝑆𝐶C,N+ 𝛽U𝐷C,N+ 𝛽V𝐺C,N+ 𝛽W𝑅C,N𝐴C,N+ 𝛽X𝑅C,N𝑆𝐶C,N+ 𝛽Y𝑅C,N𝐷C,N+ 𝛽Z𝑅C,N𝐺C,N) 1 + (𝛽S+ 𝛽N𝑅C,N+ 𝛽P𝐴C,N+ 𝛽T𝑆𝐶C,N+ 𝛽U𝐷C,N+ 𝛽V𝐺C,N+ 𝛽W𝑅C,N𝐴C,N+ 𝛽X𝑅C,N𝑆𝐶C,N+ 𝛽Y𝑅C,N𝐷C,N+ 𝛽Z𝑅C,N𝐺C,N)

(3.5) To predict the probability for creditor engagement 𝐶_C,P after the receiving the first demand letter the model in equation 3.6 is used.

𝐹(𝐶C,P) =exp (𝛽S+ 𝛽N𝑅C,P+ 𝛽P𝐴C,P+ 𝛽T𝑆𝐶C,P+ 𝛽U𝐷C,P+ 𝛽V𝐺C,P+ 𝛽W𝑅C,P𝐴C,P+ 𝛽X𝑅C,P𝑆𝐶C,P+ 𝛽Y𝑅C,P𝐷C,P+ 𝛽Z𝑅C,P𝐺C,P 1 + (𝛽S+ 𝛽N𝑅C,P+ 𝛽P𝐴C,P+ 𝛽T𝑆𝐶C,P+ 𝛽U𝐷C,P+ 𝛽V𝐺C,P+ 𝛽W𝑅C,P𝐴C,P+ 𝛽X𝑅C,P𝑆𝐶C,P+ 𝛽Y𝑅C,P𝐷C,P+ 𝛽Z𝑅C,P𝐺C,P + 𝛽NS𝐶𝐹C,P+ 𝛽NN𝑅C,P𝐶𝐹C,P) + 𝛽_NS𝐶𝐹_C,P+ 𝛽_NN𝑅_C,P𝐶𝐹_C,P) (3.6) In both models the meaning of the explanatory variables is represented in table 5 and the 𝛽 ‘s are the average effect of each of the parameters.

3.5 Model criteria

31 moderating effect the model cannot be considered to be complete. However, because the model has to be simple, the most important characteristics are incorporated based on the literature. In addition, this research is explanatory and therefore these variables are an acceptable starting point. The fourth criterion is that the model should be adaptive. Although this is an explanatory research and the main focus is to explain debtor response, not to predict it, it is possible to adapt the parameters. The last criterion is robustness. The current models are robust as they always give sensible numbers. The probit part of the type-II tobit model and the logit model provide probabilities whereas the linear regression part of the type-II tobit model always gives a positive number. In conclusion, the model is not perfect but it is a sensible starting point for explanatory research.

3.6 Estimation

To estimate the models, the natural log was taken from the amount of debt variable. This was necessary because the absolute amount of variance in the other variables included in the model is much lower than the variance of the amount of debt. For example, the variance of age in 2016 is 197.43 while the variance for the amount of debt of the charge free demand letter in 2016 is 160024. So to create a more equal distribution, the natural log was taken. In addition, the age and amount of debt variable are mean centred. Hereby, the main effect of the variables included in the interactions contain a meaningful value (Shieh, 2011).

32 and can therefore be estimated separately. When the null-hypothesis of rho=0 is rejected, the two parts share common parameters and the Heckman selection model is the better approach. The logit model was estimated with the maximum likelihood (Franses and Paap, 2001). The parameters are interpreted with the odds ratio. This ratio indicates the likelihood of an event, creditor engagement and happening versus not happening. For both the type-II tobit model the BIC, AIC and, the Wald chi and the likelihood ratio are used for model selection (Franses and Paap, 2001). For the logit model the AIC and the likelihood ratio are used for model selection (Franses and Paap, 2001).

4. Results

4.1 Pre- study

To analyse whether the more readable letters increases the processing fluency a small pre-study is conducted, see appendix 4. This pre-pre-study consisted of a survey which was distributed online. The survey was sent in four different conditions. The charge free demand letter of 2014, the first demand letter of 2014, the charge free demand letter of 2016, and the first demand letter of 2016. The survey consisted of questions about the effort it took to read the letters, how easy the letters were to read, how credible they thought the creditor was and how likely and willing they were to pay when they had (not) enough financial means to do so. A total of N=40 respondents completed the survey. The results show that the letters of 2016 are easier to read and need less effort to process than the letters of 2014. The respondents were also more likely and willing to pay when receiving the letters of 2016 when they did not have enough financial means. When they had enough financial means there was no difference between the letters. In addition, the creditor was not seen as more credible or authoritarian when the respondents received the letters from 2014.

4.2 Exploratory analysis

33 The results of the correlation as displayed in table 6, indicate that there are some significant correlations between readability and age (p<.01), amount of debt for the charge free demand letter (p<.01) and the first demand letter (p=.001) and whether or not the charge free demand letter was sent (p=.002). These correlations indicate that there are some minimal differences between age (r=-.042), amount of debt of the charge free demand letter (r=-.049), amount of debt of the first demand letter (r=.048) and sending the charge free demand letter (r=.034) between the years of 2014 and 2016. Since these correlations are rather small, below .05, they represent no big threat to the validity of this study.

In addition, no significant high correlations can be discovered except the correlation between amount of debt of the charge free demand letter and amount of debt of the first demand letter (r=.956, p<.01), and the correlation between amount of debt of the first demand letter and sending the charge free demand letter (r=-.231, p<.01). The correlation between the two variables of amount of debt is quite logical, while the amount of debt stated on the first demand letter is the amount of debt that is stated on the charge free demand letter plus the debt collection costs minus the amount already paid. While two different models for the charge free demand letter and first demand letter were estimated this will not cause a problem. The second relatively high correlation between amount of debt of the first demand letter and sending the charge free demand letter is quite remarkable. Sending the charge free demand letter is accompanied by lower debts. It would be expected that a debtor who received a charge free demand letter has a higher debt because they already received one letter. While this correlation is remarkable it will not have any influence on the current research whereas the correlation is not high enough to cause multicollinearity.

Table 6: Correlation analysis for the independent variable and moderators.

Readability Age Social Class Amount of Debt CFL Amount of Debt FDL Gender Charge free demand letter received Readability 1 Age -.042** 1 Social Class .005 .071** 1 Amount of Debt CFL -.049** .050** .010 1 Amount of Debt FDL .048** .022 .022 .956** 1 Gender .002 -.019 -.020 -.022* -.038** Charge free demand letter received .034** -.037** -.015 - -.231** .023* 1

Note: CF= charge free demand letter, FD= first demand letter

* p<0.1 ** p<0.05

34 4.3 Payments

To analyse whether the debtors paid their debt and how much they paid after receiving the charge free demand letter and the first demand letter, two different type-II tobit model were estimated. In table 7, 8 and 9, the payment incidence column displays the variables that determine whether or not someone had made a payment to the creditor. The amount column displays which variables determine how much the debtor paid to the creditor.

In table 8, the results of the two-step model are displayed. Based on these results the starting point of the type-II tobit model was determined. The results of the probit of the charge free demand letter indicate that readability (p<.01), age (p=.001), amount of debt (p<.01) and the moderator of amount of debt (p=.021) have a significant influence on the payment incidence.The linear regression shows only a significant effect of the amount of debt (p<.01) and the moderator age (p=.058) on the amount paid. The results of the probit model of the first demand letter indicate that the main effects of readability (p<.01), age (p=.061) social class (p=.089), sending charge free demand letter (p=.069) and the moderating effect of gender (p=.059) could have an influence on the payment incidence. The linear regression showed significant influence on payment amount for the main effects of readability (p=.029), amount of debt (p<.01) and for the moderator, receiving the charge free demand letter (p=.066). These values will not be interpreted but are the basis of the type-II tobit model.

Charge free demand letter First demand letter

Probit Payment Inc. Linear regression Amount Probit Payment Inc. Linear regression Amount Intercept 1.054*** 235.635 1.811*** 261.158*** Main effects Readability -.177*** -15.169 -.664*** -227.401** Age -.006*** -.341 -.008* -.321 Social Class .003 4.134 .081* 7.860 Amount of Debt .131*** 140.479*** -.020 58.960*** Gender -.046 -11.176 -.115 -27.000

Charge free demand letter

received .403* -19.027 Moderation effects Readability*Age .001 .957* .007 2.372 Readability*Social Class .006 3.156 -.076 -6.135 Readability*Amount of debt -.070** .755 .056 -26.930 Readability*Gender .011 7.510 .271* -80.466 Readability*Charge free demand letter -.353 195.979*

Table 7: Parameter estimates for the probit and linear regression for the payment incidence and amount(N=5202).

Note: * p<.1

35

4.3.1 Charge free demand letter

In table 8 the estimation results for the parameters 𝛽N_{, payment incidence and 𝛽}P _{, payment}

amount, are provided for the charge free demand letter. In this table three different models are presented. In model 1 the significant values from the two-part model, as discussed before, are included. As can be seen not all variables that were significant in the two-part model from table 8 are significant or have the same effect size in this type-II tobit model. This stresses the importance of which of these models is used. Additionally, the full type-II tobit model is estimated as presented in the conceptual model. Although all three models predict better than the null- model according to the significant Wald chi-test (p<.01), model 3 is chosen as the final model. Model 3 has the lowest AIC (30168.26) and BIC (30254.13) which indicates that this model is the best fit for the data. Although this is an exploratory research the model with the best model fit is interpreted instead of the full model. In addition, model 3 has a significant rho of .9916 (p<.01). This indicates that the Heckman selection model should be used and not a two-part model.

Payment incidence

From table 8, many conclusions can be drawn. First, it can be derived that readability has a negative influence on the payment incidence when the amount of debt is average (𝛽N

N=-.185,

p<.01). It can therefore be concluded that debtors are less likely to pay when they receive the

charge free letter from 2016 compared to receiving the charge free letter of 2014, which has the largest influence of all the variables included. Age (𝛽N

P=0.005, p<.01) and the amount of

debt (𝛽N

U=.014, p=.059) display a positive sign. Therefore, it can be concluded that debtors

who have a higher amount of debt are more inclined to pay, when they have received the letter from 2014. In addition, debtors who are older are also more likely to pay. Although these effects are smaller compared to the effect of readability. The social class and gender did not have a significant effect and were therefore excluded from the final model. Only the moderator amount of debt shows a small significant negative effect (𝛽N

Y=-0.090, p<.01).

36

Payment amount

The estimation results of the amount of payment conclude that readability does not significantly influence the amount that is paid (p=.200). On the contrary, age (𝛽P

P=1.441,

p<.01) and social class (𝛽P

T=3.859, p=.027) did have a significant positive influence on the

amount paid. When debtors are 1 year older than average they pay €1.44 more. When a debtors’ status score increases with 1 more than average he pays €3.86 more. In addition, a higher amount of debt positively influenced the amount paid (𝛽P

U=49.597, p<.01). A 1%

increase in the average amount of debt increases the payment amount with €49.69, which is a quite large effect. No moderation effects had a significant influence in the final model.

Table 8: Charge free demand letter: parameter estimates and model fit for the type-II tobit model for payment incidence and amount (Eq. 3.3) (N=9472)

Model 1 Model 2 Model 3

Payment Inc. Amount Payment Inc. Amount Payment Inc. Amount 𝜷𝟎 Intercept -1.146*** -488.489*** -1.225*** -514.120 -1.268*** -534.333*** Main effects 𝜷𝟏 Readability -.050 -63.529 -.244 52.816 -.185*** -49.568 𝜷𝟐 Age .005*** 1.339 .005*** 1.137*** .005*** 1.441*** 𝜷𝟑 Social Class -.006 .468 3.859** 𝜷𝟒 Amount of Debt .014 41.846*** .037** 49.462 .014** 49.597*** 𝜷𝟓 Gender -.046 -16.988 Moderation effects 𝜷𝟔 Readability*Age .336 .002 .772 𝜷𝟕 Readability* Social Class .006 6.098 𝜷𝟖 Readability* Amount of debt -.026** -.092*** -25.478** -.071*** 𝜷𝟗 Readability* Gender -.012 1.877 Model fit BIC 30257.97 30332.92 30254.13 AIC 30172.1 30175.48 30168.26 0% 5% 10% 15% 20% 25%

Low amount of debt High amount of debt

P rob ab il it y to pay 2014 2016

Figure 6: Charge free demand letter: moderation effect amount of debt for payment incidence, low amount of debt= a lower debt than average (mean), high amount of debt= a higher debt than average (mean)

Note: * p<.1

** p<.05

37

4.3.2 First demand letter

For the first demand letters again three different type-II tobit models were estimated (see table 9). The estimates for the parameters 𝛽N_{, payment incidence and 𝛽}P _{, payment amount,}

are provided. Model 1 includes the significant values from the two-part model (see table 7). Again not all variables that were significant in the two-part model are significant in the type-II tobit model. In addition, the effect sizes also differ. In model 2 the variables from the proposed conceptual model are presented. In this model social class is not included in the amount paid part of the type-II tobit model. Due to limitations in the variance of the data, as discussed previously, incorporating all variables in both parts of the type-II tobit model gave an incorrect image with unusually high standard errors. To account for this problem variables were included step by step following the two-part model. Including social class in both parts of the tobit model was the cause of the incorrect image and therefore only included in the probit part while according to table 7 social class did not have an effect.

When comparing the three models, the BIC (9385.548) indicates model 1 as the best model while the AIC (9289.009) indicates model 3 as the best model. Both models predict better than the null model therefore, the log likelihood ratio test is used to determine which model is better. The log likelihood ratio test between model 1 and model 3 was significant (LR=12.748, p<.05). Thus it can be concluded that model 3 fits the data better. Therefore, this model is interpreted. In addition, model 3 has a significant rho of .9995 (p<.01) which indicates that the Heckman selection model is the right approach compared to the two-part model.

Payment incidence

The most important conclusion that can be derived from table 9 is that the payment incidence is negatively influenced by the readability of the letter (𝛽N

N=-.747, p=.001), which has the

largest effect of all the variables. Debtors who received the letter from 2014 were more likely to pay than debtors who received the letter from 2016. Whether or not a payment was made is also influenced by the gender of the debtor. Female debtors are more likely to pay than male debtors (𝛽N

V= .121, p=.021). Two moderation effects are present, gender and sending the

charge free demand letter. Gender has a marginal negative moderation effect on the relation between readability and payment incidence (𝛽N

Z= -.265, p=.059). In figure 7 the effect of

38 receiving the charge free demand letter is displayed in figure 8. Receiving the charge free demand letter has a marginal significant positive effect (𝛽N

NN= .415, p=.061). Receiving the

charge free demand letter has a stronger positive effect for the letters from 2016 than for 2014.

Payment amount

The amount paid is also negatively influenced by the readability of the demand letters. When debtors received the letter from 2016 they paid less than debtors who received the letters from 2014 (𝛽P

N =-289.476, p=.004). Male debtors with an average amount of debt that

received the letter from 2016 paid €289.48 less than the male debtors with an average amount of debt who received the letter from 2014. The amount of debt (p=.354) and sending the charge free demand (p=.335) do not significantly influence the amount paid. Gender has a positive effect on the amount paid, which indicates that females pay €44.30 more than males when receiving the letter from 2014 (𝛽P

V= 44.302, p=.049). Two moderation effect were

marginally significant, amount of debt and gender. Amount of debt has a negative moderation effect (𝛽P

Y= -7.074, p=.049) and is displayed in figure 9. When a debtor has a higher amount

of debt the payment amount increases less for the 2016 letter than for the 2014 letter. Gender also has a negative moderation effect (𝛽P

Z= -108.923, p=.070). The moderation effect is

displayed in figure 10, begin female has a stronger negative effect for the letters from 2016 than for the letters from 2014. Sending the charge free demand letter does not have a significant moderation effect (p=.172).

Figure 8: First demand letter: moderation effect receiving charge free demand letter

Figure 7: First demand letter: moderation effect gender

10% 12% 5% 3% 0% 2% 4% 6% 8% 10% 12% 14% Male Female P rob ab il it y to pay 2014 2016 10% 11% 2% 5% 0% 2% 4% 6% 8% 10% 12% 14% No charge free

demand letter senddemand letter sendCharge free