1
The effect of personalized text messages on the log in
and repayment behavior of consumers at a health
insurer.
Master Thesis
Rijksuniversiteit Groningen
Faculty of Economics and Business
July , 2018
2 Abstract
This research concerns itself with the question whether a personalized text messages can help improve log in and repayment behavior of consumers. To test this a field experiment was performed where on two occasions personalized text messages were send out to consumers. In March 1675 consumers received a text message and in June another 437 consumers received a text message. Two different types of personalization were used in this research to study which type would lead to the best results. In March the personalization of the text messages was based on the BSR color model. Where red/yellow consumers received a more informal message and blue/green consumers a more formal message. In June the personalization of the text messages was based on the debt characteristics of the consumers. Where consumers that had been defaulters before received messages that emphasized negative consequences and where consumers that hadn’t been defaulters before received messages that emphasized positive consequences. Consumers that received a text message had a higher probability to log in or repay their debts compared to consumers that didn’t receive a text message. The
personalization on the basis of the debt characteristics didn’t have an effect on the probability that a consumer would log in or pay. For the BSR color model there was no effect of
personalization for the red/yellow consumers. But there was a positive effect of personalization for blue/green consumers on repayment.
Key words: Text message reminder, personalization, direct marketing, debt collection, health
3 Preface
This master thesis is the final requirement for the graduation for my master in marketing intelligence. I knew beforehand that I wanted to write my thesis at a company so that I could finish my studies and also experience doing research at a company. Now that I am finished with my thesis I am thankful the health insurer gave me this opportunity. This because in the 6 months at the company I learned a great deal of new skills.
The realization of this study wouldn’t have been possible without a view important people that I would like to thank. First I would like to thank Dr. J. T. Bouma for suggesting the company to write the thesis at and for making the first contact with the company. Secondly I want to thank my first supervisor Dr. H. Risselada who provided me with valuable feedback and also helped me when there were some setbacks in the writing process. Next to that I want to thank all my colleagues of team data that helped me wherever they could. Especially I would like to thank my supervisor at the company Frédérique Kuiper who helped me with all aspects of my thesis and from which I learned a lot. Lastly I would also like to thank all my friends and family who supported me during the writing process.
4 Table of contents
1. Introduction 5
2. Literature review 8
2.1 Sending an extra text message 8
2.2 Segmentation, Targeting and positioning 12
2.3 Segmentation on debt characteristics 14
2.4 Psychographic segmentation 15 2.5 Conceptual model 17 3. Method 17 3.1 Design 17 3.2 Sample 18 3.3 Process 18 3.4 Measures 19 3.5 Control variables 19 3.6 Psychographic segments 21
3.7 Segmentation based on debt characteristics 22
3.8 Statistical methods 24
4. Results 26
4.1 Model free evidence 26
4.2 Duration models 31
4.3 Duration models and log in behavior 32
4.4 Duration models and repayment behavior 36
5. Discussion 39
5.1 Text message reminders 39
5.2 Personalization based on the debt characteristics 40
5.3 Personalization based on the BSR color model 41
5.4 General reasons of the lack of effect of personalization 41
5.5 Managerial implications 42
5.6 Limitations and future research 43
5.7 Conclusion 44
6. References 45
7. Appendix 51
1. Communication guide 51
2. Text messages based on psychographic segmentation 52
3. Text messages based on the debt characteristics 52
4. Clustering 53
5. Proportionality tests 54
6. Overview log in behavior 54
5 1. Introduction:
According to Mewse, Lea and Wrapson (2012) the credit use of UK consumers has more than doubled over the last decade. Consumers nowadays seem to be less hesitant in using credit to make their purchases. This trend is also visible in the Netherlands where one in five families has risky or problematic debts (Westhof, De Ruig & Kerckhaert, 2015). This increased use of credit by consumers can create problems for both the debtors and the creditors. Debtors have stress related to whether or not they can pay their debts. While companies have uncertainty whether or not they will receive the funds from their consumers (Keizer, 2016). Next to the problems of the debtors and creditors, the debts can also have a negative impact on the society as a whole. Research of the Nibud shows that when debts are not resolved it could cost
society 100.000 euro per household (Madern, 2014). Therefore, it is essential and in everybody’s interest to resolve the debts of consumers quick and in the best way possible. A great deal of research has been done on the size of the debts and the causes of the debts. One of the most important causes found of these debts is that consumers don’t always act rationally when they have debts (Mewse, Lea and Wrapson, 2012). Consumers are not opening their mail and are not paying their debts directly to minimize the negative
consequences. Creditors often have to motivate and activate their consumers to seek contact and to pay their debts (Keizer 2016). Because these debts have so many negative
consequences for everyone involved it is interesting to study how companies can improve repayment.
6
debts. This because consumers have limitations regarding attention and memory which leads to suboptimal decisions regarding their long-term self-interests (DellaVigna, 2009).
From the study of Keizer (2016) it becomes clear that personalizing messages could help improve repayment rates of consumers. Which is also confirmed by studies that state that personalized messages are helpful in activating consumers (Kaptein et al. 2012 ; Autoriteiten Financiële Markten, 2018). However the literature is not completely one sided regarding personalization and sending extra messages. Other studies in different settings had mixed findings regarding personalization and consumers engagement (Aguirre et al., 2016; Gerber et al., 2009; White et al., 2007). The study of Gerber et al. (2009) didn’t find an overall effect of personalized text messages and stated that messaging fatigue could be a cause for this. If consumers receive to many messages they won’t take action. These mixed findings make it interesting to study the exact effect of personalization of messages in different contexts. A solution for the drawbacks of the study of Keizer (2016) could be to send personalized text messages instead of demand letters with personalized cards. Text messages can be used to overcome the fact that the personalized cards are not always seen. This because text messages have the advantage of being at the right place at the right time and are unlikely to be ignored (Kass, 2007; Dale & Strauss, 2009). Additionally, according to Fortin (2000) consumers recall the information in text messages better compared to other channels. Text messages can also include links and references to the online environment of the company. According to the autoriteit financiële markten (2018) and MacDermott (2008) taking action and taking this first step is essential for consumers that want to repay their debts. The online environment is especially suitable for this because consumers can easily assess and pay their debts there. The research of the autoriteit financiële markten (2018) found that a text message which
mentioned the online environment led to an increase in the access of the online environment. So text messages could be useful as an alternative because they can overcome the drawbacks of personalized cards and can improve the amount of consumers that take action and access the online environment.
7
Bandura and Locke (2003) and Block and Williams (2002). The personalization on
psychographic segmentation will be done with the help of the brand strategy (BSR) model which divides consumers in four types of personas (Dam, Hattum & Schieven, 2013). The effectiveness of these two segmentation types will be tested to check which segmentation is most suitable for addressing consumers that have late payments.
To test the effectiveness of an extra personalized message the following research questions are formed:
- Is a personalized text message effective in increasing the access to the online
environment compared to no text message or a neutral text message?
- Is a personalized text message effective in reducing the repayment times
compared to no text message or a neutral text message?
o Which segmentation method is best in creating the personalized messages and reducing repayment times/increasing access to the online environment? These research questions will be tested by sending an extra text message after the consumers received a demand letter to remind them to pay their debts. The research will be done in the form of a field experiment. The text messages will be send to real consumers of a health insurer. An advantage of a field experiment is that the results are obtained in a natural setting which gives a true indication of the effect of sending text messages.
The findings of this study will expand the literature on how consumers can be activated to pay their debts or face their debts. The autoriteit financiele markten (2018) stated that additional research is much needed on how to activate financial vulnerable consumers. This study tries to do that by investigating the effectiveness of personalized text messages as activation strategies. This study will also expand the literature by sending text messages to consumers that already missed a payment instead of consumers that have to pay in the future. By doing this study in a different context than the previous studies the knowledge will be expanded on the subject of consumer activation and personalization.
The managerial implications of this study lie with the fact that managers could use these activation techniques to reduce the debt of the company. If these text messages prove to be successful, managers could use personalized text messages to make sure vulnerable
8
useful tool for managers to increase the liquidity of their organizations. So that they can use the money available for new initiatives or opportunities.
The research is structured as follows: first all of the existing theories and concepts about motivating and activating consumers to log in and to pay their debts are discussed. After that the different methods used in this research will be discussed. Following will be the results of the study. The last part will be a discussion of the results and recommendations for further research in this field.
2. Theoretical framework
This part of the research will provide the theoretical background the research is based on. It will look at past and present literature to come up with relevant hypotheses. First, it will focus on the literature regarding direct marketing and sending an extra text message. Second, it will expand on the theories that are present on segmentation, targeting and positioning. Third, this part will look at the two different types of segmentation methods that will be used in the research. Lastly, there will be a conceptual model that will give a good overview of the expected relationships.
2.1 Sending an extra text message:
This part of the literature review is based on the process and obstacles which were proposed by autoriteit financiële markten (2018). Figure 1 shows the process and the three obstacles consumers face when they have to pay a debt. This literature review will explain how a company can help their consumers to overcome these obstacles.
Figure 1: Process of debt repayment
9
Spiller, 2006). By using direct marketing the companies have more certainty that the
consumers receive the messages that they sent. Direct marketing tactics are mainly used to get quick responses and to build a relationship with the consumer (Esteban & Hernández, 2013). Quick responses of consumers are especially valuable for the company and the consumers. This because if there is a quick response the debts can also be resolved quickly and there will be less problems for both consumers and the company.
One of the most important and most used form of direct marketing is telemarketing (Kotler & Armstrong, 2018). When companies use telemarketing consumers are directly contacted on their landlines or their mobile phones. Sending text messages to consumers is an important part of telemarketing. A company can directly send text messages with relevant information to consumers to inform them on new deals, new products or their payment status. Text
messages can be very useful for companies as it is relatively easy and inexpensive to send out a great deal of text messages at once (Karlan et al., 2016). These text messages can be used to help persuade the consumers to pay their debts. They can function as a quick reminder that a consumer has a late payment. Text messages are especially interesting as they can be used to reach a large target audience with personalized interventions (Fogg, 2003).
The first barrier for consumers that needs to be overcome is that the consumers should open their messages. Text messages are especially interesting for this purpose. At the moment consumers receive demand letters but don’t open or read the demand letters (Keizer, 2016). Text messages could overcome this problem because mobile phones play a central role in our lives nowadays. Consumers in the US spend more than an hour a day on their mobile phones on nonvoice activities(Fong, Fang & Luo, 2015). Additionally, the mobile penetration in Europe and the US is above 100% (Zaher, 2008). Thus there is only a limited chance that a consumer will not receive a text message, as almost everybody has a mobile phone and checks it daily. The fact that text messages have high response rates also indicates that consumers open and read their text messages (Zhang & Mao, 2008). Furthermore, according to Dale and Strauss (2009) text messages are unlikely to be ignored by consumers. So by sending a text message the companies could increase the amount of people that open their direct messages.
10
their exact debt is and how they can repay this in the best possible way. In reality, this step seems to be hard for consumers as they often show a lack of response on demand letters from creditors (Mewse, Lea and Wrapson, 2010). This is harmful for both the consumer and the creditor as the debt becomes larger and more problematic. One of the positive effects of sending text messages could be that consumers are less reluctant to take action. These text messages could function as a reminder for the consumers to take action regarding their debts. This could be helpful because consumers have limitations regarding attention and memory and make decisions that are not the best for their long-term interests (DellaVigna, 2009). The study of the autoriteit financiële markten (2018) looked into this and found that a text message from a bank (ING) that mentions the online environment led to a 13-percentage point increase in access of the online environment. This is valuable for a company because the consumers are able to assess the amount of debt they have and can easily pay their debts online. So an extra text message that mentions and has an explicit link to the online environment could lead to consumers taking the important first step and accessing their online environment.
The last and most important barrier for the consumers is the barrier of actually paying their debts. Text messages could be an important tool in overcoming this barrier. The research of Kast et al. (2010) studied the effect of a text message reminder on savings rates of consumers. They found that a text message reminder can improve the financial behavior of consumers. Karlan et al. (2012) found that text messages that clearly stated the company’s name
significantly and robustly improved the repayment rates of consumers in the banking industry in the Philippines. Confirming that text message reminders can improve the financial behavior of consumers. Cadena et al. (2011) extended this research by studying the effectiveness of text messages with loan repayments in Uganda. They found that an extra text message resulted in an average decline of late payments with two days per month. Next to that they also found that the text messages led to an 8% increase in the probability of paying on time. The use of text messages was equally effective as a promise of a lower interest-rate in the future or a guarantee of cash back (Cadena et al. 2011). The use of text messages in the study of Cadena et al. (2011) was seen as highly successful due to the fact the marginal costs of sending out text messages is almost zero compared to the other two methods.
11
messages led to messaging fatigue which led to people not taking action. The fact that text messages don’t always lead to activation is also partly confirmed in the research of Karlan (2012). In this research text messages which focused on framing or timing didn’t have an effect on the repayment. The research of Bracha and Meier (2015) found that text message reminders don’t work for consumers that can’t change their behavior easily. Which could be relevant in this study as it could be hard for some consumers to change their behavior and pay their debts. This because some consumer are probably not able to repay their debts as they don’t have enough funds.
Replicating the studies in different contexts is also interesting because the previous studies were contained to a specific setting of developing countries. Because of this specific setting the external validity is low and it would be hard to generalize the results across other countries (Cadena et al. 2011). Especially since in developing countries on average the creditor rights are lower (Djankov et al., 2007). These lower creditor rights combined with less financial infrastructure makes it harder to enforce the repayments of the loans (Cadena et al., 2011). Consumers in these countries can decide to just no pay and it would be incredibly hard to collect the debt. In this study it would be hard for consumers to completely avoid the payment. Because if payments are not made the consumer will eventually go to the CAK and the payment will be subtracted from their benefits, pension or salary (CAK, 2018). So there is no way to avoid the payments for consumers. According to Block & Williams (2002) and Mewse, Lea & Wrapson (2010) if consumers have unavoidable debts and receive the right message for them they have more chance of paying their debts. Thus it is expected that if the right messages are send the text messages reminders are even more successful in this study.
This research also differentiates from the other studies by focusing on consumers that have already missed a payment instead of consumers that have to pay in the future. A text message could be more effective regarding activation and repayment of consumers that already have missed a payment. This because according to Rodgers (2005) it is better to send a text message for activation during stressful periods and when the encouragement is needed the most. When a consumer has already missed a payment they need the reminder and
12
could be unpredictable. This also makes it interesting to study the exact relationship between extra text messages and repayment. Most literature indicates that there will be a positive effect and possibly a stronger effect than the previous studies but there is still enough uncertainty about the exact effect to make this study relevant.
From the literature hypotheses one and two can be formed. The hypothesis focuses on the increased access of the online environment and the repayment. The first part of the framework which focused on opening the messages isn’t included as it is not possible to measure whether or not consumers open their text messages.
H1: Sending an extra text message will increase access in to the online environment. H2: Sending an extra text message will improve the repayment rates of consumers.
2.2 Segmentation, Targeting and positioning:
A possible explanation of the mixed findings regarding the success of text messages could be that each consumer reacts differently on influence strategies. The importance of well-suited text messages is confirmed by the study of Kaptein et al. (2012) where consumer either received a tailored text message, a random text message from all the text options or a contra tailored text message. The tailored text messages led to better results than randomly chosen text messages. The randomly chosen text messages in turn led to better results than contra tailored text messages. Personalized text messages are also more likely to catch the attention of the consumer (Dijkstra, 2005). Next to that, a personalized text message can increase the propensity for consumers to take action (Autoriteiten Financiële Markten, 2018). Catching the attention and activating the consumers is especially interesting for this study. This because at the moment consumers often don’t open their demand letters or take action regarding their debts.
Adapting the text messages can be done with the help of segmentation, targeting and positioning. Segmentation is the grouping of consumers according relevant variables where newly created groups have similar needs. The different groups should be unique compared to each other and should have different responses to marketing strategies (Kotler, 2003).
13
you have targeted after the segmentation (Venter, Wright & Dibbs, 2015). In this research, all of this is done by choosing the right text messages for the right groups.
The segmentation of consumers can be done based on several different criteria. According to Avery et al. (2014) there are three main criteria on which consumer can be segmented. These criteria are behavioral, value-based and demographic characteristics of consumers. The demographic characteristics are used most to segment consumers because those are readily available for companies. Companies often have a database of their consumers regarding their age, sex, spending behavior, etcetera. While the criteria for the other two segmentation methods are harder to obtain. It is generally more difficult for a company to know how their consumers are thinking or what values they have than to know their age.
The different segmentation methods will also lead to different segments. The two
segmentation methods used in this research are psychographic (value-based) segmentation and segmentation based on the debt characteristics of the consumers. The choice for these two segmentation methods is partly based on the availability of the data for segmentation. The psychographic segmentation will be based on the brand strategy research (BSR) model and will divide the consumers into different groups according to their beliefs and values. The segmentation based on debt characteristics will be done based on information available on the debts of the consumers. The success of segmentation, targeting and positioning is largely dependent on the data that is used to create those groups. When the data is not accurate, the created groups won’t be accurate either and the positioning strategies will not be appropriate for the groups. Thus making the consumer database with all the information about the consumers one of the most important tools for marketing purposes (Trusov, Ma and Jamal, 2016). In our case this would mean that the data used for the psychographic segmentation and debt based segmentation are of vital importance of the success of the targeting of the
consumers.
The overall effectiveness of segmentation is based on selecting the right segmentation bases (Wind, 1978). Therefore, in this case it is expected that there will be a differences between the successes of the two segmentation methods. As one method will probably be better suited for this kind of situation that the other method. However it is not completely clear which
14
2.3 Segmentation on debt characteristics:
The way consumers should be approached about their debts is difficult because consumers differ in their preferences. If you for example approach consumers regarding their debts in a too aggressive way for their liking it could be that they won’t respond (Block & Williams, 2002). The consumers will act irrationally and will bury their heads in the sand and won’t take action. This is confirmed by Mewse, Lea and Wrapson (2012) that state that if consumers are addressed in a way that doesn’t fit with their debt characteristics they can be even more demotivated to pay. So the debt characteristics of a consumer are an important aspect of how you should communicate with consumers.
One of the aspects that differs per consumer is how the consequences of debt should be communicated (Keizer, 2016). Therefore, it is expected that the text messages should also differ for consumers that have different kinds of debts. For consumers with small debts and less problematic debts it would be better to emphasize the positive consequences of paying their debts (Keizer, 2016). This because when goals are attainable for consumers it is more effective to emphasize the positive consequences of reaching the goal (Bandura & Locke, 2003; Gal & McShane, 2012). When these positive consequences are emphasized the consumers also need to see their goal as being attainable (Strecher et al., 1995). Mentioning that the payment of their debts is a solution of their problems could be a way to mention positive effects.
Emphasizing the positive consequences for consumers is not always the best way to motivate consumers to pay their debts. This because consumers that have larger and more problematic debts need other incentives to be motivated to pay their debts. These consumers need an emphasis on negative consequences to be motivated to pay their debts (Keizer, 2016). For these consumers the consequences need to be large, clear and almost unavoidable (Block & Williams, 2002; Mewse, Lea & Wrapson, 2010). Next to that, consumers that receive a negative message need to know clearly what they have to do to avoid these negative consequences. This because strong fear appeals are especially effective when the
consequences are clear and people know how to avoid them (Witte & Allen, 2000). The negative consequences in this case are the fear appeals for the consumers. In our case
15
from the collection agency when they don’t pay their debts. In addition, the message should contain clear steps how they can easily pay their debts and avoid the extra costs.
H3: Sending a personalized text based on debt characteristics segmentation will increase the access in the online environment more than a neutral text message.
H4: Sending a personalized text based on debt characteristics segmentation will improve the repayment rates more than a neutral text message.
2.4 Psychographic segmentation:
The second segmentation method that will be discussed in this literature review is
psychographic segmentation. Segmentation based on debt characteristics can provide valuable insights regarding the consumers but not all of them. To get a different view of the consumers a psychographic segmentation could prove to be beneficial for a company. Psychographic segmentation could add more richness and dimensionality to the segments compared to segmentation based on debt characteristics (Plummer, 1971). Psychographic segmentation or lifestyle research has multiple definitions. A common used definition for psychographics is from Demby (1994):
“the use of psychological, sociological, and anthropological factors, self-concept, and lifestyle to determine how the market is segmented by the propensity of groups within the market - and their reasons - to make a particular decision about
a product, person, or ideology”
Wells (1975) added a broader view of psychographics, which also includes attitudes, needs, beliefs, personality traits, lifestyles and motivations. The ultimate goal of psychographic segmentation is to get a better understanding of the consumer using their opinions, interests and psychological dimensions (Ziff, 1971). Psychographic segmentation is especially interesting because according to Mcdonald & Dunbar (2000) psychographic segmentation is useful for successful market segmentation.
The psychographic segmentation will be done with the help of the brand strategy research (BSR) model. This model is based on Adler’s social psychology theory and contains valuable information about a consumer’s values and beliefs (Callebaut et al., 1999). It helps to better understand the consumers and in this case will be used to create messages that are
16
different dimensions: vitality, harmony, control and security (Dam, Hattum & Schieven, 2013). These four dimensions are created by two axes which are based on sociological values and on psychological values. Figure 2 shows how the different dimensions are formed.
Figure 5: Four dimension of the BSR model
These four dimensions provide a basis for targeting consumers and developing marketing strategies (Dam, Hattum & Schieven, 2013).
Consumers assigned to four different dimensions (Dam, Hattum & Schieven, 2013): - The vitality dimension: “consumers are adventurous, self-conscious, creative,
open-minded, passionate, energetic and always looking for the unusual”.
- The harmony dimension: “consumers are helpful, caring, kind, enthusiastic, optimistic, open and spontaneous”.
- The control dimension: “consumers are individualistic, capable, rational, wise, career oriented and competitive”.
- The security dimension: “consumers are calm, traditional, conservative and cautious”.
17
differently as they don’t have same values. However, it can also be seen that some values are similar across groups. This because the groups all share one of the four dimensions with another group. The literature leads to the following two hypotheses.
H5: Sending a personalized text message based on the BSR model will increase the access in the online environment more than a neutral text message.
H6: Sending a personalized text based on the BSR model will improve the repayment rates more than a neutral text message.
2.5 Conceptual model:
From the theory in this part the following conceptual model can be realized. The research will focus on the effect of an extra text message and how personalizing that text message
moderates that effect.
Figure 6: Conceptual model
3. Research design
This part of the study will elaborate on how the research is conducted and which methods were used to analyze the data.
3.1 design:
18
consumers. The consumers that received a demand letter in March will get a personalized text message on the basis of the psychographic segmentation. The consumers that received a demand letter in May will get a personalized message on the basis of their debt characteristic. The different segments will be divided into three different groups, where one group will get the personalized message, one group a neutral message and one group will not receive a message. This will be done to test the hypothesis whether a personalized text message is better regarding repayment and accessing the online environment compared to a neutral message or no message.
3.2 Sample:
The sample consists of consumers of a health insurer and that have received a first demand letter after they have missed a payment. It will be the consumers that have received a first demand letter in either March or May. The sample will consist of different consumers in the two months. So the consumers that received a message in March will not be included in the sample of May.The consumers that are included in this study have a late payment of at least €10. This threshold was chosen to exclude consumers that only missed a little of their payment. A small missed payment could indicate a mistake by a consumer and not that they weren’t able to pay. In March 2987 consumers were included in the study of which 1675 consumers received a text message. These messages were based on the psychographic characteristics of the consumers. For the text messages based on debt characteristics 783 consumers were included in the analysis of which 437 consumers received a text message.
3.3 Process:
19
days after the text message. This was chosen because after a week it is not sure whether or not the observed behavior is caused by the text messages. Also if the time window is too long the effect of the text message is not constant over time. Therefore a relative short time frame of seven days was chosen to analyze the data.
Table 1: Process of sending text messages
Date reminder Date of sending text Days between reminder and text
Date of measurement
06-03-2018 20-03-2018 14 27-03-2018
07-05-2018 01-06-2018 25 08-05-2018
3.4 Measures
The two aspects that need to be measured after sending an extra text message are the access of the online environment and the repayment rates of the debt. The access of the online
environment will be measured by looking at log-ins of the consumers in the online
environment. The time between the text message and the log-in will be measured in days. The log-ins were chosen as a measure for contact because they can directly be attributed to
specific consumers. The repayment rates of the debt will be measured by checking if a consumer paid their debt and how fast a consumer paid their debt. This again will be done in number of days since they received the text message.
3.5 Control variables
These variables are expected to have an impact on the repayment rates of the consumers but are not included in the research question. To control for their effects they will be added to the model to create a more clear view of the exact relationships between the repayment rate and the independent variables. Some of the variables are based on available literature while other variables are based on repayment behavior of consumers at the health insurer in the past.
Age:
20
Gender:
The gender of the consumers is another control variable that is included in the analysis. This because men and women differ in their actions and motivations (Eagly, 1987). The difference in preferences and motivations could have an effect on their log in and repayment behavior.
Defaulter before:
If a consumer has been a defaulter before in the recent past this could also influence the repayment rates. This because if a consumer has been a defaulter before it indicates that they have more problems with repaying. A consumer that had problems with repayment in the past could possibly also have problems with repayment in future.
Additional health insurances:
Whether a consumer has additional health insurances can influence their repayment and log in behavior. On average consumers that have more additional and dental insurances are more risk averse. These consumers have these insurances to cover possibly high costs in the future. It is expected that these consumers are also consumers that repay their debts quicker as they want to avoid the risk of paying extra costs.
Neighborhood:
Consumers that live in a disadvantaged neighborhood could have more problems with repayment of their debts. Disadvantaged neighborhoods are classified by Vektis which is an organization that is specialized in business intelligence in the health insurance market. They have a database with neighborhoods that have problems with paying their insurance premium. It is expected that consumers that live in these neighborhoods will have more problems with repayment and are less quick compared to other consumers.
Duration of relationship with the consumer:
21
Days after text consumer logged in:
This is a specific control variable for the model which focuses on repayment. This variable is included because consumer that log in very quickly after the text are expected to pay more often. This because consumers that log in after the text message are more likely to be activated by the message compared to consumers that log in a week later.
3.6 psychographic segments:
The psychographic segments will be based on the brand strategy model (BSR) which was shown and explained in the literature review (page 15:16). The model consists of four different categories: red , yellow, blue and green. The four categories were created by having a sociological dimension and a psychological dimension. Red represents the vitality
dimension, yellow the harmony dimension, blue the control dimension and green the security dimension. The BSR model was expended by the SmartAgent Company(SAMR, 2018) to create four personas Martijn (red), Anne (yellow), Robbert (blue) and Johan (green). They created these personas by enriching the BSR model with customer data from the health insurer. With the data from the health insurer extra quantitative analyses were performed and interviews have been conducted to confirm the personas. With the new model and personas the SmartAgent Company created a guide on how you should communicate with the personas. These personas all have do’s and don’ts on how you should communicate with them. A short summary of the communication guides is included in appendix 1.
The personas all have their unique characteristics but also have similarities. Because of the similarity and the limited options of customization in a text message the personas have been reduced to two groups. Martijn (red, vitality) and Anne (yellow, harmony) are grouped together and Robbert (blue, control) and Johan (green, security) are grouped together. Where the red and yellow group represent consumers that value more informal contact with their insurer. This group also wants less technical jargon and wants a bit more freedom than the other group. The blue and green group is almost the complete opposite of the red and yellow group. The consumers that are part of the blue and green groups prefer more formal
communication and value technical jargon when this is needed. See appendix 1 for a more elaborate guide of communicating with these consumers
22
In the red and yellow group the consumer are addressed with ‘Beste’ and ‘je’ which is a more informal way of addressing a consumer (Renkema, 2012). The closing of these text messages will be with ‘Groeten, Kim’ which is an informal way of ending the text message (Renkema, 2012). The consumers that are in the blue and green group are addressed with ‘Geachte’ and ‘u’ which is more formal compared to ‘Beste’ and ‘je’ (Renkema, 2012). The text messages in this group will be closed with ‘Met vriendelijke groet’ which is a more formal closing of a text message (Renkema, 2012). The last difference between the two text messages is related to the use of jargon. In the blue and green group more jargon is used in the text messages and ‘nota’ is used instead of ‘verzekering’. By changing these aspects of the text messages the messages become more suited for the groups. If the messages are more suited it is expected that the log in and repayment behavior of the consumers improves. The corresponding text messages for the groups can be found in the appendix 2.
3.7 Segmentation based on debt characteristics:
The debt segmentation is done on the basis of a cluster analysis of the debt characteristics of the consumers. The debt characteristics variables that were available for analysis were: the height of the overdue payment for the month, total amount of overdue payment of a consumer, number of open claims, the length of the longest claim in months and whether a consumer had been a defaulter before. After checking for correlations between the variables it became clear that several of the variables measured the same underlying principle. In the final cluster analysis the height of the overdue payment for the month, the total amount of overdue payments and whether a consumer had been a defaulter before were included.
23
The clustering was done with the help of hierarchical clustering, density-based spatial clustering of applications with noise (DBSCAN) andpartitioning around medoids (PAM) clustering. These methods were chosen because the data that was used for clustering consisted of different types of data. The data contained interval, ordinal and nominal data that needed different distance metrics to be used effectively in clusters analysis. The exact results of these clustering methods can be found in appendix 4. The clustering methods came to the same solution, which can be seen in table 2.
Table 2: Results of cluster analyses
From table 2 it becomes clear be that cluster 1 contains consumers that have been defaulters before in the recent past. These consumers also have longer claims, more total overdue payment and a higher number of open claims on average. The average age and number of years insured differ not enough from each other to adjust text messages based on these variables. It would not be possible to adjust text messages based on age for example because the consumers differ only roughly one year from each other on average. The average overdue payment for the month is higher in cluster 2. This can be attributed to the fact that the
consumers that have problems with paying have less additional insurances. These consumers don’t have the money to have extensive additional insurances. From these two clusters it can be concluded that there is one group (cluster: 1) that are the more problematic consumers and another group (cluster: 2) which can be seen as less problematic. The cluster analyses were also performed with different consumer characteristics. Additional insurances and age were also used in the analysis but this didn’t result in more meaningful clusters. Thus it was chosen to leave these variables out as they didn’t help in creating more clear clusters.
The cluster analysis was also done with different compositions of the current variables to see which variables were important in creating the clusters. Whether a consumer had been a defaulter before was the leading variable that created the two clusters. Whichever composition of variables was used in the cluster analysis the consumer were divided according to the fact
N Average overdue payment Average total overdue payment
24
whether they had been a defaulter before. Because of this the decision was made to send out the text messages based on whether or not a consumer had been a defaulter before.
The clusters indicate that there need to be two different kind of text messages. A text message that emphasized the negative consequences was created for the consumers with more
problematic debts. In this study the consumer with more problematic debts were the consumer that had been defaulter before in the last 15 months. A text which emphasized more positive consequences was created for the consumers with less problematic debts. The consumers with less problematic debts in this case were the consumers that hadn’t been defaulters before in the last 15 months.
The creation of the text messages are in line with the studies that were mentioned in the literature review (Keizer, 2016; Block & Williams, 2002; Mewse, Lea & Wrapson, 2010; Bandura & Locke, 2003; Gal & McShane, 2012). The text messages are constructed in the same way as in the study of Keizer (2016). The positive text messages mentioned that by paying the debt or seeking contact the consumers could solve their problems. The emphasis on solving problems for the consumers can be seen as a positive approach. The negative text message mentions the possible extra costs of €40 for not paying on time. This clearly shows that there are negative consequences when the debt isn’t paid on time. The text messages for the debt characteristics groups can be found in appendix 3.
3.8 Statistical methods:
25
data would lead to biased estimates (Cox, 1972). The cox proportional hazard model is able to estimate without biases with data with right censoring(Nitzan and Libai, 2011).
The dependent variable in a cox proportional hazard consists of two parts: whether an event happened or not and the time till the event happened. With the help of the cox proportional hazard model the survival function and hazard function can be estimated (Despa, 2018). The survival function gives the probability that an event didn’t happen until that time. The hazard function gives the probability that an event will happen, given that the event didn’t happen already. In the survival model the time between repayment or log in for consumer i is a random variable with a cumulative distribution function F(t) and density f(t) = F′(t).
The probability that repayment or log in has not happened at time t is provided by the survivor function:
𝑆(𝑡) = 1 − 𝐹(𝑡)
The probability that repayment or log in will happen at time t, given that it hasn’t happened yet is provided by the hazard function:
ℎ(𝑡) = 𝑓(𝑡)/𝑆(𝑡)
The hazard rate for consumer i with features that are captured by vector x is: ℎ𝑖(𝑡) = ℎ0(𝑡) exp (𝐵′𝒙
𝑖𝑡)
In the equation ℎ0(𝑡) represents the baseline hazard function that captures the longitudinal effects. The ′ in the equation indicates the effect of variable 𝑥𝑖𝑡 on the hazard rate. Four different proportional hazard model will be estimated in this study. Two models that concern itself with when a consumer logs in to their online environment are estimated. One model checks the effectiveness of a text message reminder and one model checks the effectiveness of a personalized text message. The equations below shows the models for the log in timing of the consumers.
ℎ𝑖(𝑡) = ℎ0(𝑡) exp (𝑇𝑒𝑥𝑡. 𝑚𝑒𝑠𝑠𝑎𝑔𝑒. 𝑟𝑒𝑚𝑖𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐺𝑒𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐴𝑔𝑒𝑖𝑡+ 𝐷𝑒𝑓𝑎𝑢𝑙𝑡𝑒𝑟𝑖𝑡+ 𝐴𝑑𝑑. 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡
26 ℎ𝑖(𝑡) = ℎ0(𝑡) exp (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑. 𝑡𝑒𝑥𝑡. 𝑚𝑒𝑠𝑠𝑎𝑔𝑒𝑖𝑡+ 𝐺𝑒𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐴𝑔𝑒𝑖𝑡+ 𝐷𝑒𝑓𝑎𝑢𝑙𝑡𝑒𝑟𝑖𝑡+ 𝐴𝑑𝑑. 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡
+ 𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟ℎ𝑜𝑜𝑑𝑖𝑡+ 𝑌𝑒𝑎𝑟𝑠. 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟𝑖𝑡 )
The next two equations show the models for the repayment timing of the consumers. One equation that focuses on the effect of the text message reminder an one equation that focus on the effect of the personalization of the text message.
ℎ𝑖(𝑡) = ℎ0(𝑡) exp (𝑇𝑒𝑥𝑡. 𝑚𝑒𝑠𝑠𝑎𝑔𝑒. 𝑟𝑒𝑚𝑖𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐺𝑒𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐴𝑔𝑒𝑖𝑡+ 𝐷𝑒𝑓𝑎𝑢𝑙𝑡𝑒𝑟𝑖𝑡+ 𝐴𝑑𝑑. 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡
+ 𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟ℎ𝑜𝑜𝑑𝑖𝑡+ 𝑌𝑒𝑎𝑟𝑠. 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟𝑖𝑡 + 𝐿𝑜𝑔. 𝑖𝑛. 𝑑𝑎𝑦𝑠)
ℎ𝑖(𝑡) = ℎ0(𝑡) exp (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑. 𝑡𝑒𝑥𝑡. 𝑚𝑒𝑠𝑠𝑎𝑔𝑒𝑖𝑡+ 𝐺𝑒𝑛𝑑𝑒𝑟𝑖𝑡+ 𝐴𝑔𝑒𝑖𝑡+ 𝐷𝑒𝑓𝑎𝑢𝑙𝑡𝑒𝑟𝑖𝑡+ 𝐴𝑑𝑑. 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡
+ 𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟ℎ𝑜𝑜𝑑𝑖𝑡+ 𝑌𝑒𝑎𝑟𝑠. 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟𝑖𝑡 + 𝐿𝑜𝑔. 𝑖𝑛. 𝑑𝑎𝑦𝑠)
Table 3: Description of the explanatory variables
Variable Description
Text message reminder
Indicates whether or not a consumer had received a text message reminder. (1=Yes, 0=No)
Personalized text message
Indicates whether or not a consumer had received a personalized text message. (1=Yes, 0=No)
Gender Gender of the consumer. (1=male, 0=female)
Debt The amount of debt the consumer had on which they received their demand letter. Age The age of the consumer in whole years.
Defaulter Indicates whether or not a consumer had been a defaulter before in the last 15 months. (1=Yes, 0=No)
Add Insurance Indicates whether or not a consumer had any additional health insurances at the company (1=Yes, 0=No)
Neighborhood Indicates whether or not a consumer lives in a disadvantaged neighborhood and is more at risk to miss a payment. (1=Yes, 0=No)
Years customer The number of years a consumer has been with company (in whole years)
Log in days The number of days after the text message the consumer logged in to their online environment.
4 Results
This part of the research will report on the results that were obtained after performing the analysis. First a broad overview of the log ins and repayment over time will be shown. Subsequently the effects between the different text messages will be measured with the help of simple t-tests. After the initial model free analyses several survival analyses will be
performed on the data to check if there are differences in the probability that a consumer pays or logs in during the 7 day time period.
27
This first part of the result section will give an introduction of the results of the text messages. First the cumulative repayment and log in behavior will be shown over time. This will give a an overview of when and how much the consumers log in and pay and if there is a visible difference between the different text message groups.
Graph 1: Cumulative overview of the log ins of the color text messages
Graph 1 shows the cumulative percentage of consumers that logged in after the date the text messages were sent. From the graph it becomes clear that consumers that received a text message logged in more often compared to the no text message group. After the initial shock of the text messages it seems as if the percentage of people that logs-in stays the same between the different conditions. Furthermore there isn’t a large visible difference between the different text messages and the log in behavior
Graph 2: Cumulative overview of the repayment of the color text messages
0% 5% 10% 15% 20% 25% 30% 35% 1 2 3 4 5 6 7 C um ul at ive am o unt lo gg ed in
Days after text
Cumulative percentage logged in (March)
28
Graph 2 shows the cumulative amount of consumers that have paid their debts in March. At first sight it becomes clear that consumers that received a text message paid more overall. It also shows that the consumers that received the blue/green text message did the best in paying their debts compared to the other messages. Next to that the graph indicates that during the weekend (fifth and sixth day) no consumers paid their debts. It can be seen that this difference is made up on Monday (seventh day) where the repayment increases again. Overall the
consumers that received a text message paid more but there is no large visible difference between the different text messages.
Graph 3: Cumulative overview of the log ins of the positive and negative text messages
0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 1 2 3 4 5 6 7 C um ul at ive am o unt pa id
Days after text
Cumulative percentage paid (March)
redyellow bluegreen neutral no message
0% 5% 10% 15% 20% 25% 30% 35% 1 2 3 4 5 6 7 C um ul at ive am o unt lo gg ed in
Days after text
Cumulative percentage logged in (June)
29
Graph 3 shows the log in behavior of the consumer that received text message in June. The overall log in is also lower compared to the previous month (March). This is probably due to the fact that the text messages in this month were send out later on in the month. The
consequence of this is that a relative higher percentage of consumers that are not able to pay their debts are in this sample and thus wouldn’t log in. The graph shows that the consumers that received a positive or neutral text message logged in more than the consumer that received a negative text message or no text message. It also seems that in future the no message group will overtake the consumer that received a negative text message.
Graph 4: Cumulative overview of the repayment of the positive and negative text messages
Graph 4 shows the results of the second batch of text messages which consisted of positive and negative text messages. It can be seen that overall the repayment is lower than in the other month. This can probably be attributed to the same reason as why there were less log ins made in June The second and third day show no payment and this is because this is the weekend in which no payments were received of the consumers. It also becomes clear that the neutral and positive text message are doing better than the no message and negative groups. The low percentage paid in the negative text message group could be caused by the fact that
0% 5% 10% 15% 20% 25% 1 2 3 4 5 6 7 C um ul at ive am o unt pa id
Days after text
Cumulative percentage paid (June)
30
these consumers have been defaulters before and have more trouble paying compared to the other groups.
31
Table 4: Within group comparison of personalized text messages in March Text messages March
Group: Blue & Green
Text type N Percentage
paid
Percentage logged in
Blue and Green 370 38.3%*** 31.4% ***
Neutral 383 33.7% 26.9% **
No message (control) 493 27.2% 19.3%
Group: Red & Yellow
Red & Yellow 466 33.7% ** 29.6% ***
Neutral 456 34.9% *** 34.2% ***
No message (control) 466 25.9% 20.5%
*** Significant at the 0.01 level compared to the no message category (2-tailed). ** Significant at the 0.05 level compared to the no message category (2-tailed). * Significant at the 0.10 level compared to the no message category (2-tailed).
The results show that the blue/green text message differs significantly from the no message control group. For both the number of consumers that paid and the number of consumer that logged in during the 7 day period. There was no significant difference between the neutral and the blue and green text messages. The results of the red and yellow group are similar to the blue and green group. The text messages significantly differ from the no message control group. But there is no significant difference between the red and yellow text messages and the neutral text messages.
Table 5: Within group comparison of personalized text messages in June Text messages June
Group: Defaulter before
Text type N Percentage
paid
Percentage logged in
Negative text message 94 8.5% * 0.181
Neutral 95 13.7% *** 24.2% ***
No message (control) 171 1.8% 9.4%
Group: Not been a defaulter before
Positive text message 130 20% * 30% **
Neutral 118 28% *** 33.9% ***
No message (control) 175 10.9% 16.6%
32
Table 5 shows the results of the negative and positive text messages that were send. It becomes clear that for the group that had been defaulters before that the negative and neutral text messages both are better than no message regarding repayment. For the percentage logged in only the neutral text message is significantly better than the no message group. The negative and neutral text messages didn’t significantly differ from each other. The group that hadn’t been a defaulter before and received a positive text message as their personalized message showed similar results. The personalized message and neutral message perform better than the no message control group. But there is no difference between the positive messages and the neutral messages. Overall it seems that sending a text message helps in increasing repayment and log in behavior.
4.2 Duration models
After the model free evidence this part will use duration models ,where possible, to estimate the effect of the text messages. First the models regarding log in will be discussed and after that the models regarding repayment behavior. The order of the models will be that first the results of a text message reminder will be shown and after that the effect of personalization will be shown.
Before the models were estimated all the independent variables were checked for
multicollinearity. This is done to make sure that all the variables measure different underlying principles. To check whether multicollinearity was an issue the variance inflation factors (VIF) were calculated. According to O’brien (2007) the most used threshold for VIF scores are four and ten. If the variables have a higher score than four or ten than there is an issue with multicollinearity in the data. The table below shows VIF scores of the variables.
Table 6: VIF scores independent variables
Variable VIF
Text message reminder 1.03
Gender (1=male) 1.04
Age 1.08
Defaulter before 1.02
Additional health insurances 1.08 Disadvantaged neighborhood 1.01
Years insured 1.08
33
The results show no value above four or ten so there is no problem with multicollinearity in the data. Thus regarding multicollinearity all the variables can be included in the model The independent variables in all the models were also checked for proportionality. The independent variables were checked for proportionality because in cox proportional hazard models the effect of the independent variables should be consistent over time. If a variable wasn’t consistent over time and thus didn’t satisfy the proportionality assumption they were removed from the models. If the variable of interest (text message reminder, personalized text message) didn’t satisfy the proportionality assumption than a different model form was used to estimate the effect of the text messages.
In the cox proportional hazard models the odds are calculated to get a more precise effect of the independent variables. These represent the increase or decrease in probability a consumer logs in or pays their debt in a certain week, given that they hadn’t logged in or paid yet. The odds are calculated by:
𝑂𝑑𝑑𝑠 = (exp(β)− 1)∗ 100
This equation gives the odds of something happening compared to not happening. If for example a consumer has a chance of repaying their debts of 20% and the odds increase with 35% this results in (20*1.35=27%) a 27 % chance that a consumer pays their debt.
4.3 Duration models and log in behavior
The first part of this study was to check the effect of text messages on log in behavior.
However the first model is not a duration model in this case. A different model was estimated because the proportionality assumption didn’t hold for the text message reminder variable. The results of the proportionality test can be found in appendix 5. Because the proportionality assumption didn’t hold for the independent variable that was of interest, the data had to be modelled in another way.
After checking the log in behavior of consumers that received a text message and consumers that didn’t receive a text message, it became clear that the effect of the text messages
34
Because a cox proportional hazard model couldn’t be estimated a logistic regression was estimated. This logistic model solely focuses on the number of consumers that logged in and not the timing of log ins. Both a logit and probit model were estimated but the AIC was lower in logit models or very close to the probit model. The logit models were chosen to analyze the data because these models are more suitable for interpreting the results. The AIC scores of the logit and probit models can be found in appendix 7.
Table 6: Logistic regression with log in behavior as the dependent variable
Log in behavior and text message reminder models
Model 1 Model 2 Model 3
Total March June
N 3770 2987 783
coef odds coef odds coef odds
intercept -0.641*** -47.3 -0.442* -35.7 -1.385*** -75.0
Text message reminder 0.656*** 92.7 0.582*** 79 0.881*** 141.2
Gender (1=male) -0.358*** -30.1 -0.376*** -31.4 -0.207 -18.7 Age -0.013 -1.3 -0.016*** -1.6 -0.004 -0.4 Defaulter before -0.606*** -45.5 -0.611*** -45.7 -0.581** -44.1 Additional health insurances 0.238** 26.8 0.265** 30.4 0.136 14.5
Disadvantaged neighborhood 0.091 9.5 0.159 17.2 -0.209 -18.9
Years insured -0.028 -2.7 -0.033 -3.2 -0.010 -1.0
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
AIC 4096.6 3316.5 782.13
Likelihood ratio test 186.69* 152.63*** 37.58***
35
The results show that a text message reminder increases the chances that a consumers accesses the online environment compared to no message. Thus according to these results hypothesis 1 is supported because text messages lead to more log ins.
For the log in behavior the effect of the personalization of the text messages is also assessed. In the personalization models only consumers that received text messages are included. In this case there were no problems with proportionality of the personalized text messages. The effect of the personalization of the text messages was consistent over time. Thus a cox proportional hazard model could be estimated without problems.
Table 7: Cox proportional hazard model with log in behavior as the dependent variable
Log in behavior and personalized text messages models
Model 4 Model 5 Model 6
Total March June
N 2112 1675 437
coef odds coef odds coef odds
Personalized text message -0.027 -2.6 0.017 1.7 -0.206 0.3
Gender (1=male) -0.270*** -23.7 -0.354*** -29.8 -0.006 1
Age -0.015*** -1.5 -0.003 0.7
Defaulter before -0.504** -39.6 -0.488*** -38.6
Additional health insurances 0.178* 19.5 0.158 17.1 0.185 0.4
Disadvantaged neighborhood -0.203 0.6
Years insured -0.045 -4.4 -0.083** -8 -0.038 0.5
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Concordance 0.606 0.598 0.542
Rsquare 0.039 0.035 0.007
Likelihood ratio test 83.93*** 60.04*** 3.01
The results show that the model in June is not significant in predicting the log in behavior. As the likelihood ratio test is not significant in June. A logit model was estimated with the same data to check if the model form was the cause of the insignificance. However this logistic regression wasn’t significant either. The total model and the march model are significantly better than the null model. No effect can be found for the personalization of the text message on the log in behavior. If a consumer had been a defaulter before this had a negative effect (p<0.001) on the log in behavior of the consumers. The fact whether or not a consumer was a defaulter was left out of the June model because the personalization was based on that
36
years a consumer is insured at the company has a negative effect in the March model (p<0.01).
The overall results regarding personalization and log in behavior show no effect of the personalization. To get a more clear and precise overview of the results will also be reported per text message condition.
Table 8: Cox proportional hazard model with log in behavior as the dependent variable
Log in behavior and personalized text message per condition
Model 7 Model 8 Model 9 Model 10
Blue/Green Red/Yellow Positive Negative
N 753 922 248 189
coef odds coef odds coef odds Coef odds
Personalized text message 0.227 25.4 -0.114 -10.8 -0.173 -15.9 -0.319 -27.3 Gender (1=male) -0.235 -21 -0.410*** -33.7 0.075 7.8 -0.164 -15.1 Age -0.010 -1 -0.025*** -2.5 -0.006 -0.6 0.001 0.1 Defaulter before -0.478*** -38 -0.536*** -41.5 Additional health insurances 0.097 10.2 0.233 26.2 0.125 13.3 0.219 24.5 Disadvantaged neighborhood -0.031 -3 -0.074 -7.1 -0.318 -27.2 Years insured -0.059 -5.7 -0.009 -0.9 -0.034 -3.3 Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Concordance 0.59 0.641 0.54 0.564 Rsquare 0.026 0.076 0.005 0.012 Likelihood ratio test 20.14*** 72.92*** 1.3 2.22
The results show that both the positive and negative text message models were not significantly better than the null model in predicting log in behavior. Again a logistic
regression was also estimated with the same data but this model wasn’t significant either. So the positive and negative text messages didn’t have an effect on log in behavior. In the color models the personalized text messages didn’t have an effect either. Whether a consumer had been a defaulter before did have a negative effect (p<0.001) on the log in behavior in the color models. It was excluded in the positive and negative models because there the personalization was based on whether a consumers had been a defaulter before.
37
4.4 Duration models and repayment behavior
Table 9: Cox proportional hazard model with repayment behavior as the dependent variable.
Repayment behavior and text message reminder models
Model 11 Model 12 Model 13
Total March June
N 3770 2987 783
coef odds coef odds coef odds
Text message reminder 0.470*** 60 0.365*** 44 1.055*** 187.2
Gender (1=male) -0.106 -10 -0.087 -8.4 0.083 8.6
Age 0.006** 0.6 0.004 0.4 0.012 1.2
Defaulter before -1.104*** -66.8 -1.126*** -67.6 -0.995*** -63.0
Additional health insurances 0.019 1.9 -0.019 -1.9 0.330 39.1
Disadvantaged neighborhood -0.005 -0.5 -0.006 -0.6 -0.580 -44.0
Years insured 0.042 4.3 0.102 10.8 -0.014 -1.4
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Concordance 0.664 0.659 0.715
Rsquare 0.086 0.093 0.07
Likelihood ratio test 341*** 291*** 56.8***
Table 6 shows the results of the cox regression model on the repayment behavior of the consumers. This model focused on whether or not a text message reminder was useful in improving the repayment of the consumers. The model shows the results in total and split into the two months the text messages were send. All models are predicting significantly better than the null model as the likelihood ratio test is significant in all three models.
The results in table 6 show that text message reminders and whether or not a consumers has been a defaulter before are significant (p<0.001) in all models. The text message reminder has a positive effect in all the models as the repayment coefficient is positive. The odds show that if consumers receive a text message reminder they have a 60% higher probability to pay compared to consumers that didn’t receive a text message. Next to that they have a 44% higher probability to pay compared to no message in March and a 187.2% higher probability to pay compared to no message in June. Whether or not a consumer has been a defaulter before has a significant (p<0.001) negative effect on the chance that a consumers pays their debt as the coefficient is negative. The age also has a significant (p<0.01) positive effect on repayment in the total model but this is effect is very small.
38
To check the effectiveness of personalization a model that only includes consumers that received a text messages is estimated. This way the differences between the neutral and personalized messages can be analyzed.
Table 10: Cox proportional hazard model with repayment behavior as the dependent variable.
Repayment behavior and personalized text messages models
Model 14 Model 15 Model 16
Total March June
N 2112 1675 437
coef odds coef odds coef odds
Personalized text message 0.018 1.9 0.112 11.9 -0.397 -32.8
Gender (1=male) -0.128 -12 -0.143 -13.3 0.19 21
Age 0.006* 0.6 0.005 0.5 0.015 1.5
Defaulter before
-1.070***
-65.7 -1.130*** -67.7
Additional health insurances -0.081 -7.7 -0.087 -8.4 0.211 23.5
Disadvantaged neighborhood 0.06 6.2 0.14 15 -0.889 -58.9
Years insured -0.018 -1.8 -0.033 -3.3 -0.016 -1.6
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Concordance 0.637 0.648 0.603
Rsquare 0.078 0.095 0.023
Likelihood ratio test 191.2*** 167.7 *** 10.27
Table 7 shows the results of consumers that all received a text message but where some received a personalized message and some neutral messages. The total model and the model in March are significantly better than the null model. However the model in June didn’t predict repayment better than the null model. To check if the insignificance was caused by the model choice, a logit model was estimated with the same variables. This also resulted in an insignificant model. This indicates that the variables in June were not able to predict the repayment behavior of consumers that received text messages. In the total model and the model for March personalization is also not significant. Whether a consumer had been a defaulter before is significant and has a negative effect on repayment in the total model and in March model (p<0.001). Whether or not a consumer had been a defaulter before is excluded from the analysis in June as the personalization was based on whether or not a consumer had been a defaulter before and to avoid multicollinearity.
