• No results found

The path towards a healthy user base

N/A
N/A
Protected

Academic year: 2021

Share "The path towards a healthy user base"

Copied!
63
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

The path towards a healthy user base

Predicting continued app usage for a health initiative

(2)

The path towards a healthy user base

Predicting continued Health Initiative app usage

Master Thesis By Tjalling de Jong

University of Groningen Faculty of Economics and Business MSc Marketing Management / Intelligence

January 12, 2020

1st supervisor: Prof. Dr. Ir. K. van Ittersum

(3)

Management summary

The aim of Health Initiative is to make a positive and sustainable impact on the lifestyle of its users. Mhealth apps have shown to possess the ability to positively influence the health of the user (Anderson, Burford, & Emmerton, 2016; Free, Phillips, Galli, 2013; Gagnon et al., 2016; Steinhubl, Muse, & Topol, 2015). One requisite to utilize this ability is continued mhealth app usage.

This research looked into the factors that have a potential impact on Health Initiative app usage. The Health Initiative app was first introduced by A health insurance company, but now operates as a stand-alone entity, being freely available for everyone. The data for this study was provided by A health insurance company / Health Initiative. The goal of the research was to find significant predictor variables for continued Health Initiative app usage.

Through literature study, fourteen explanatory variables were established and the accompanying hypotheses regarding their effect on the dependent variable (continued Health Initiative app usage) were measured by means of a logistic regression model. The variables can be broadly subdivided into the categories App Behavior, Customer Characteristics and Seasonality.

Variables that reflect app behavior that requires true activity by the user showed to have positive effect on continued Health Initiative app usage. The level of login, level of activity, level of point collection, level of challenge participation, weekgoal participation and weekgoal completion all have a positive effect on continued app usage. Variables that reflect app behavior that does not require much real-life activity form the user showed a negative relationship with continued app usage. The level of interactivity (number of coach questions answered) showed a negative effect on continued app usage. The same holds for the level of reward redemption, which means that when a user places an order (reward) in the webshop, he/she is less likely to continue to use the app in the next month.

In terms of app characteristics and seasonality it was found that being a customer at A health insurance company, Age and the lifetime of a person as a Health Initiative app user positively influence continued app usage. Decreased probabilities for continued app usage were found for users who have unsubscribed from the Health Initiative newsletter and during the months of November and December.

(4)

Acknowledgements

(5)

Table of Contents

MANAGEMENT SUMMARY ... 3

1 INTRODUCTION ... 6

2 THEORETICAL FRAMEWORK ... 9

2.1CONTINUED APP USAGE (DV) ... 9

2.2INDEPENDENT VARIABLES (IV’S) ... 9

2.2.1 App behavior ... 10 2.2.3 User characteristics ... 13 2.2.3 Seasonality ... 14 3 METHODOLOGY ... 16 3.1DATA ... 17 3.1.1 Missing values ... 17 3.1.2 Outliers ... 17 3.1.3 Data transformation ... 18 3.2DESCRIPTIVE STATISTICS ... 19 3.3RESEARCH METHOD ... 20 4 RESULTS ... 21

4.1SIGNIFICANT MAIN EFFECTS ... 21

4.2DRIVERS OF CONTINUED APP USAGE ... 22

4.2.1 Health Initiative App behavior ... 22

4.2.2 Customer characteristics ... 23 4.2.3 Seasonality ... 24 4.3INTERACTION EFFECTS ... 25 4.3.1 Multicollinearity ... 25 4.4VALIDATION ... 26 4.4.1 Validity metrics ... 26

4.4.2 Main Effects Model ... 27

4.4.3 Significant Main Effects Model ... 27

4.4.4 Interaction Model ... 27 4.4.5 Null Model ... 28 4.4.6 Model selection ... 28 5 DISCUSSION ... 30 5.1CONCLUSIONS ... 31 5.1.1 App behavior ... 31 5.1.2 Customer characteristics ... 32 5.1.3 Seasonality ... 33

5.2LIMITATIONS AND FUTURE RESEARCH ... 34

5.3MANAGERIAL IMPLICATIONS ... 35

APPENDICES ... 38

APPENDIX 1A:HYPOTHESES ... 38

APPENDIX 1B:RESEARCH VARIABLES ... 38

APPENDIX 2:DESCRIPTIVE STATISTICS ... 39

APPENDIX 3:MODELS OUTPUT ... 40

APPENDIX 4:VIF SCORES ... 41

APPENDIX 5:MODEL VALIDATION ... 42

APPENDIX 6:HYPOTHESES REVISITED ... 44

APPENDIX 7:R-CODE ... 46

(6)

1 Introduction

Personal health is a factor that has a tremendous impact on the quality of life. Scientific studies consistently show a positive relationship between happiness and health on a personal level (Graham, 2008). Collectively, the age expectancy of the Dutch population is expected to increase, while not giving in on perceived quality of personal health during those extra years (RIVM, 2018). However, the number of people with chronic diseases will grow due to the aging society. At the same time, the population is faced with an increase in societal pressures and work-related stress. The fact that unhealthy behavior is responsible for almost 20% of the disease burden indicates that lifestyle choices have an effect on individual- and collective health (RIVM, 2018).

The technological revolution has a tremendous impact on the way we live our lives. Smartphones have become ubiquitous and its functionalities are ever expanding, resulting in increased communication, interactivity and health information availability (Vaghefi & Tulu, 2019). Consumers have their smartphone within arm’s reach for the majority of the day (Dey et al., 2011). Approximately two hours of the consumers’ day are spent on apps, suggesting that apps can be an appropriate tool for behavior change (Flaherty et al., 2019). An emerging app domain is that of Mobile Health (mhealth). Mhealth apps have the ability to implement interventions that support users in weight management, smoking cessation, stress

management and dealing with mental conditions (Anderson, Burford, & Emmerton, 2016; Free, Phillips, Galli, 2013; Gagnon et al., 2016; Steinhubl, Muse, & Topol, 2015).

Nudging consumers towards a healthier lifestyle does not only have potential benefits for the individual user, but might also provide benefits on a global scale. In 2018, a total amount of 100 billion euros was spent on healthcare in the Netherlands, which translates to almost 10% of the Gross Domestic Product (GDP) (CBS, 2019). These expenditures are expected to increase up to 174 billion in 2040, due to an ageing society and technological advancements (RIVM, 2018). Thus, in a broader sense, mhealth provides an opportunity to mitigate expenditures in healthcare on the long term through its ability to deal with unhealthy behavioral habits of consumers (Anderson, Burford, & Emmerton, 2016; Free, Phillips, Galli, 2013; Gagnon et al., 2016; Steinhubl, Muse, & Topol, 2015).

(7)

company customers for making healthy decisions. In 2017, A health insurance company introduced a mobile app version of the program and made it freely accessible for everyone, thus moving away from the sole-focus on customer loyalty and moving towards the broader perspective of increasing general consumer well-being. The main mission of Health Initiative is to make a positive and sustainable impact on the users’ lifestyle and ultimately help them unravel the path towards a healthy future. This mission statement is an expression of A health insurance company’ belief in Corporate Social Responsibility (CSR). The program

anticipates the expected growth in healthcare expenses by putting into practice the adagium ‘prevention is better than cure’. Health Initiative allows users to acquire points through a multitude of healthy behaviors. The points collected can be exchanged against products, services, discounts and/or philanthropic initiatives. In this way, users of the Health Initiative app can acquire rewards while simultaneously improving their health.

In the rapidly growing mhealth market, more than 300,000 health apps are currently available (Lancet, 2019). Thus, Health Initiative operates in a crowded and competitive market, resulting in volatile consumer behavior (Ding & Chai, 2015). Customer acquisition can be up to twelve times more expensive in comparison with customer retention (Torkzadeh, Chang, & Hansen, 2006). Thus, from an economic perspective it is more beneficial to retain users than having to acquire new members. In terms of mhealth adoption, it was reported that almost 20% of smartphone owners make use of such apps (Cho, 2016). According to Ding & Chai (2015), about one third of app users will continue to do so after a one-month trial period and there is a retention rate of 4% after a whole year. Other reports state that a quarter of mhealth apps are only used once after installation (Kayyali et al., 2017; Vaghefi & Tulu, 2019). These statistics show the volatility in consumer behavior with regard to apps. An app is often just as easily deleted as it is downloaded. It is believed to be unlikely that the

intended benefits from mhealth app use will be realized when the usage is short-lived (Michie et al., 2017; Vaghefi & Tulu, 2019). This challenges the mission of Health Initiative.

(8)

apps. In particular, this research will focus on the Health Initiative app by A health insurance company.

This introductory section is followed by a literature review section, which defines factors that potentially influence continued app usage in the Health Initiative app domain and the accompanying hypotheses for this research. After that, the methodology of the research and data issues are described. The outcomes of the model are reported in the results section, which is followed by a discussion of the results as well as managerial implications. The study concludes with a reflection on the limitations of the research and potential avenues for future research.

(9)

2 Theoretical Framework

In this part of the report the theoretical foundation of the research is described, dependent (DV) and independent (IV) variables are defined and hypotheses presented and explained. The goal of this research is to establish insights in the effects of several factors on continued app usage for the Health Initiative app.

2.1 Continued App Usage (DV)

Retention, attrition and churn are closely related constructs. Attrition/churn rates define the percentage of consumers that stop being a customer of a company, while the retention rate constitutes the percentage of consumers who continue to be a customer of a company. They are important indicators for organizations, because when a customer leaves, extra efforts have to be put into acquisition, which ultimately lead to negative financial consequences (Ascarza, Iyengar, & Schleicher, 2016). Thus, valuable information resides in the probabilities of whether an individual customer is going to churn or is going to be retained. Establishing insights in the antecedents of this behavior can be helpful in the retention process of

customers in the future. In the mhealth business it is important that app users stay active after the adoption phase in order to bring about positive effects in terms of lifestyle change and positive overall health outcomes (Abebe et al., 2013; Free et al., 2013; Tomlinson et al., 2013).

In the context of this study, the dependent variable consists of continued Health Initiative app usage. There is no official pre-set moment in time when it becomes abundantly clear that a Health Initiative app user has churned, because the transaction takes place in a non-contractual setting and the transaction is continuous. The industry generally uses one month of inactivity as a cut-ff moment (Research2Guidance, 2018). This makes sense,

because most companies refer to their members/users as Monthly Active users (MAU’s). The research measures the effect of different factors on the probability of consumers to be

retained as MAU’s. Thus, a customer is retained as a MAU when at least one login per month occurs for this user.

2.2 Independent Variables (IV’s)

(10)

Initiative app. The factors are broadly categorized in the segments App Behavior, User Characteristics and Seasonality.

2.2.1 App behavior

A user of an app has a subjective experience while using an app. Thus, the user will evaluate the app based on its characteristics. Vaghefi & Tulu (2019) found user experience to have an effect on the continued use of mHealth apps. When the evaluation of the capabilities of the app is positive, customers are more likely to continue to use the app.

According to the Post Acceptance Model (PAM), the capability usefulness is an influencer of satisfaction and continuance intention (Bhattacherjee & Barfar, 2011). The perceived usefulness of an app increases when the attributes of the app match the attributes that the user seeks (Vaghefi & Tulu, 2019). In this case, an app is useful when it helps the user attain his/her goals. Thus, the app serves as a means to an end (goal of the user). When the app offers users an opportunity to achieve their health goals, users may extend and continue their use of the app beyond the first few interactions (Vaghefi, 2019). This is also proposed in the extended TAM model by Venkatesh & Bala (2008), who claim that relevance and output quality are drivers of perceived usefulness. In the Health Initiative app, users can participate in weekgoals. We argue that a user will participate in a weekgoal when the user values this goal and therefore a match between the goals of the user and the attributes of the app exists. Thus, when a user engages in a weekgoal, we argue that this user perceives the app as more useful than a user who does not engage in a health goal.

H1: Weekgoal participation has a positive effect on continued app usage. H2: Weekgoal completion has a positive effect on continued app usage.

Customer Engagement (CE) is a construct that has been applied into a wide variety of domains. Flaherty et al. (2019, p. 313) proposed that CE ‘reflects an individual-specific, context-dependent, psychological state that emerges through two-way interaction with an object’. Flaherty et al. (2019) distinguish between three types of engagement: behavioral, cognitive and emotional engagement.

(11)

a vital role in continuing to use an mhealth app’ and that ‘having persistence at intended health goals and being able to pursue them despite the likely challenges appeared to play a key role toward continued use of mHealth apps’. In addition, according to Van der Heijden (2004),people who are not internally motivated are less likely to persist in health app usage, especially in cases where there is no obligation to usage continuation (Karahanna et al., 1999). We argue that intrinsic motivation and persistence are intertwined with behavioral engagement. The level of activity of a user in the Health Initiative app domain increases when the user puts in extra time, effort and persistence. It is proposed that the number of monthly activities of a Health Initiative user is an indication of his/her behavioral

engagement and therefore has a positive effect on continued app usage. In the same way it is expected that the number of days logged into the Health Initiative app in a particular month (M) has a positive effect on continued app usage in the next month (M+1).

H3: The level of activity has a positive effect on continued app usage.

H4: The number of days logged into the Health Initiative app in the previous month

has a positive effect on continued app usage in the next month

The ultimate goal of an LP is to foster loyalty, which it tries to encourage through rewards. According to Dorotic et al. (2011) an LP should reward customers for their loyal behavior. This can happen in the form of discounts, services, etc. Although Health Initiative moved away from solely being an LP, it still aims for loyal behavior in their user base.

A user of the Health Initiative app can collect points, which in turn can be used to attain rewards. According to previous research, the closer a person is to reaching a certain reward (through a higher number of points accumulated as opposed to a lower number), the more effort he or she is willing to exert into obtaining this goal (Dorotic, Verhoef, Fok, & Bijmolt, 2014; Kivetz, Urminsky, & Zheng, 2006; Nunes, 2006; Taylor & Neslin, 2005). This phenomenon is labeled the points-pressure mechanism. Based on this mechanism, we expect that the number of points collected is positively related to the retention of Health Initiative app users.

H5: The level of point collection has a positive effect on continued app usage.

The points collected by a member of an LP can be used for a variety of purposes. The

(12)

disxount. In the situation when members redeem their points for a reward, their response is affected both attitudinally and behaviorally, ultimately leading those individuals to be more attached to the firm, the so-called rewarded behavior mechanism (Palmatier et al., 2009; Taylor & Neslin, 2005). We hypothesize that this attachment to the firm / app leads to increased retention of Health Initiative app users.

H6: The level of reward redemption has a positive effect on continued app usage

Aristotle thought of humans as social animals by nature. More recent arguments for his view are that humans are hardwired for social interaction because of the fact that our lives

depended on the herd for a large part of our existence. Social influence is also ingrained in the Unified Theory of Acceptance and use of Technology (UTAUT), where it is expected to have a direct effect on behavioral intentions of individuals (Venkatesh, Morris, Davis, & Davis, 2003).

Ding & Chai (2015) propose social elements to enhance emotional attachment to an app and that ‘emotional attachment to the social community is the most valuable to the users compared to many features offered by these apps such as recording daily activities and diet’ (Ding & Chai, 2015, p.847). Social competition is one of the mechanisms that is part of the human social behavior. Seeing other people using the app and sharing own behavioral data in a health app is one of the external motivators (Peng et al., 2016). Health Initiative app users have the possibility to participate in challenges, where their rank among their peers is based on their achievements. We expect that social competition kicks in when users participate in challenges and hypothesize that challenge participation has a positive effect on continued app usage.

H7: participation in a challenge has a positive effect on continued app usage.

(13)

physical activity (Hosseinpour & Terlutter, 2019). We hypothesize that the users who choose to respond to the coach questions have a higher level of interactivity than users that who do not answer coach questions. Therefore, it is expected that interactivity with the coach increases the likelihood that of this user to continue to use the Health Initiative app.

H8: The level of interactivity with the Health Initiative coach is positively related to

continued app usage.

2.2.3 User characteristics

The user base of the Health Initiative app is formed by a heterogenous group of people. Users tend to differ in terms of demographics and personal characteristics. This paragraph proposes hypotheses for these variables. These variables are distinguished from the variables in section 2.2.2 because they do not require any active behavior from the user. These more passive variables include age, membership length at Health Initiative (web/app), whether someone is a A health insurance company customer or not and whether a user has unsubscribed for the Health Initiative newsletter or not.

First, the act of taking part in physical activity has been shown to have a negative relation with age (Krug, 2018). Next to that, research by Morris & Venkatesh (2000) has proposed that increased age is associated with having difficulties in processing complex stimuli, which may very well be the case in the process of Health Initiative app use. This finding is in line with earlier research that established a negative relationship of smartphone usage and adoption of mobile technologies with age (Klenk & Reifegerste, 2017). Thus, age is expected to have a negative relationship with continued app usage.

H9: Age has a negative relationship with continued app usage.

As was already mentioned in the introductory section, the continued usage of mhealth apps is sparse. Research shows low retention rates for mhealth app users and it is reported that a majority of mhealth apps are used only once after installation (Ding & Chai, 2015; Kayyali et al., 2017; Vaghefi & Tulu, 2019). Based on these findings, we argue that the longer a

(14)

positive effect on continued app usage. In the same way, the lifetime as a user of the Health Initiative app is expected to have a positive effect on continued app usage.

H10: lifetime as a member of Health Initiative has a positive effect on continued app

usage.

H11: lifetime as a user of the Health Initiative app has a positive effect on continued

app usage.

The Health Initiative program is freely accessible for all consumers. This means that a user of the Health Initiative program can be a A health insurance company customer, although this is not necessary. However, this does provide a distinction between the users of the program. Namely, a user can be a A health insurance company customer or have their health insurance organized elsewhere. We argue that a user who is a customer of A health insurance company has a higher behavioral engagement than a user who is insured elsewhere. Therefore, we hypothesize that being a customer of A health insurance company has a positive effect on continued app usage.

H12: Being a A health insurance company customer has a positive effect on

continued app usage.

Users of the Health Initiative app can indicate whether they do or do not want to receive e-mails and/or newsletters from Health Initiative. According to Hollebeek (2011), sharing information in such a way leads to an increase in cognitive engagement. Other research has linked continued app usage to the provision of relevant content (Ball et al., 2014). Therefore, we expect app users who unsubscribe for such Health Initiative content to be less cognitively engaged than app users who did not unsubscribe and thus still receive the content.

H13: Unsubscribing for the Health Initiative newsletter has a negative effect on

continued app usage.

2.2.3 Seasonality

(15)

year when they can switch to another health insurance company. In the Netherlands, this period ranges from November until January. Customers can apply for another health insurer and thus churn during the months November and December and this change is then

formalized in January. Since Health Initiative was previously an LP by A health insurance company, it is expected that consumers do still tie Health Initiative and A health insurance company together. We therefore hypothesize that the months November and December will show a negative effect on continued app usage.

H14: The months November and December have a negative effect on continued app

usage.

An overview of the total number of hypotheses to be tested is presented in table 1. This table indicates the hypothesis statement as well as the direction of the expected effect.

# Hypothesis +/-

1 Weekgoal participation has a positive effect on continued app usage. + 2 Weekgoal completion has a positive effect on continued app usage. + 3 The level of activity has a positive effect on continued app usage. + 4 The number of days logged into the Health Initiative app in the previous month has a

positive effect on continued app usage in the next month + 5 The level of point collection has a positive effect on continued app usage + 6 The level of reward redemption has a positive effect on continued app usage + 7 Participation in a challenge has a positive effect on continued app + 8 The level of interactivity with the Health Initiative coach is positively related to continued

app usage.

+ 9 Age has a negative relationship with continued app usage. - 10 Lifetime as a member of Health Initiative has a positive effect on continued app usage. + 11 lifetime as a user of the Health Initiative app has a positive effect on continued app usage. + 12 Being a A health insurance company customer has a positive effect on continued app

usage. +

13 Unsubscribing for the Health Initiative mail / content has a negative effect on continued app usage.

(16)

3 Methodology

The goal of this research is to find variables that influence continued app usage in the context of the Health Initiative app by A health insurance company. Ultimately, the goal is to

estimate the effects of different predictor variables on continues app usage. Different types of factors serve as IV’s in order to predict the DV. Table 1 provides an overview of all the variables used in this study.

Variables Description Scaling

Weekgoal participation Whether a user participated in a Health Initiative weekgoal in a particular month

Binary: 0 = no, 1 = yes

Weekgoal completion Number of weekgoals completed by the user in a particular month

Numeric (0-5)

Level of activity The number of activities of the user in a particular month

Numeric (0-200)

Level of login The number of days a user has logged into the Health Initiative app during a particular month (M)

Numeric (1-31)

Level of point collection Number of points collected by user

in a particular month Categorical: o 0 o 1 – 2.000 o >2.000

Level of reward redemption Number of orders in webshop by

user in a particular month Numeric (0-20)

Level of challenge participation Number of challenges a user has participated in during a particular month

Numeric (0-7)

Level of interactivity Number of interactions with the Health Initiative health coach in a particular month

Numeric (0-200)

Age The age of the user Numeric (16-119)

User lifetime (Web) The lifetime of the user in the

Health Initiative domain in years Numeric (0-10)

User lifetime (App) The lifetime of the user in the

SamenGezoind app domain Categorical: o 0-6 months o 6-12 months o 12-18 months

A health insurance company customer

Whether a Health Initiative user is a customer at A health insurance company

Binary: 0 = no, 1 = yes

Unsubscribed Health Initiative newsletter

Whether an individual has unsubscribed for the newsletter of Health Initiative

Character: yes/no/unknown

Month The particular month Character (August-December)

Continued app usage (DV) Whether a Health Initiative app user has used the app at least once in the next month (M+1)

Binary: 0 = no, 1 = yes

(17)

3.1 Data

Data is needed in order to facilitate this study. The necessary data about the Health Initiative app is provided by A health insurance company. The Health Initiative app currently has approximately 57.969 users. The observation period used for this research ranges from July 2019 until December 2019. The data from the month of December is solely used as input for the dependent variable. The data from the months of July until November are as input for the dependent variable as well as the independent variables. As can be seen in table 1, this research utilizes a total of fifteen variables, of which fourteen function as independent variables to predict the one dependent variable. In order to construct a usable dataset that is ready for statistical analysis some preparatory steps have been taken. The following sub-paragraphs explain the way this study handled missing values, outliers and other factors that forced data transformation.

3.1.1 Missing values

The selected data contained numerous missing values. These NA’s (Not Available) get in the way of conducting proper statistical analysis because these non-values are not interpretable. The important question to ask before transforming such observations is whether the NA’s are truly unavailable or if they are perhaps disguised observations of zero values. Simply deleting NA’s would result in a loss of statistical accuracy, since in that case the numerous

observations containing NA’s would be excluded from the model, which would result in a much smaller sample. The solution for this problem lies in the transformation of the NA’s into values of zero or into “unknown”. For this study it sufficed to transform the data into a “0” or simply into an “unknown”, depending on the nature of the particular variable.

3.1.2 Outliers

The same way data analysis can be corrupted by including missing values into the data, outliers can deteriorate the accuracy of a model. Outliers can be defined as observations that lie outside of the overall pattern of a distribution (Cateni, Colla, & Vannucci, 2008). Thus, outliers are mostly observations that are either too low or too high in order to fit the

(18)

no or a small number of activities, while a minority shows a tremendous number of activities. According to the data, one particular user managed to perform 926 activities during one single month, which roughly translates into 30 activities a day. The Health Initiative data contain heaps of such fraud-reeking examples. This behavior is something that the Health Initiative department deals with on a daily basis. In order to prevent these potentially fraudulent cases from messing with the accuracy of the estimates, a cut-off point has to be set. The fact that the majority of users is fairly inactive results in a distribution that is severely skewed to the right. This makes it hard to use the general rules of thumb, because this would result in a very low cut-off point for the number of activities, which means that a lot of “true” data would be lost. When common logic applied to come up with a reasonable amount of activities that can be performed in a single month, a cut-off point of 200 activities per month is chosen. In the same way the number of interactions with the coach is cut off at 200 and the number of goals achieved at fifteen per month. In this way, the extreme values with regard to point collection are automatically filtered out as well, because these

observations depended on a high number of activities.

In terms of age, people younger than sixteen years old officially are not allowed to make use of the Health Initiative program, so these people have been filtered out as well. It was also observed that a fairly large number of people was above the age of 119, which did not seem plausible due to the active nature of the app, so these observations were filtered out as well. Next to that, deceased cases were excluded from analysis.

3.1.3 Data transformation

Not only did the initial data contain missing values and outliers, also the target variables were presented in ways that were not always as useful for statistical analysis. For example,

birthdates and application dates had to be translated from a date into a certain duration of time (age or relationship length). The reason for this is simply that it makes the

(19)

3.2 Descriptive statistics

In order to get a feel for the overall pattern within the available data, this paragraph presents a brief outline of the observations in the dataset. The picture that is presented is one where the anomalies (missing values and outliers) have already been handled through exclusion / imputation. An overview of the descriptive statistics can be found in appendix 2.

The dataset contains a total of 110.155 observations. It has to be noted that this number does not reflect the number of users, since an individual user can occur multiple times in the dataset. This form of repeated measures occurs because of the fact that we try to shed light on the potential predictors of continued app usage on a monthly basis. Therefore, we include every single month of usage by an individual user into the dataset. The dependent variable consists of the observation whether or not that same individual has also used the Health Initiative app at least once in the next month. The dataset consists out of 40.004 unique users. The maximum amount of times that a user occurs is five times, since no more months are included.

The average age of the Health Initiative app user in the dataset is 47.1 years old. 40.706 users are male and 69.211 are female, while the gender 238 users are unknown. The average number of logins per month is 13.3. On average, users showed 10.8 activities per month, with a minimum of 0 and a maximum of 200. The maximum number of orders by an individual user in a particular month was 20, while the average number of orders per month is 0.2. The third percentile of the number of orders per month has a value of 0, which shows that the data for this variable is positively skewed. The average number of points earned is 438, with a maximum of 25.000 points. 80% of the dataset is a customer of A health insurance company, while 20% is not.

(20)

3.3 Research method

A binary logistic regression (logit) model is chosen to analyze the research question. A logit allows us to infer the unobserved (latent) utility of a choice by the estimated probability (Leeflang et al., 2015). In more specific terms, a Generalized Estimating Equation (GEE) logit model is constructed in order to account for the repeated measures that occur in the panel data. Figure 2 shows a graphical representation of the traditional logit building blocks.

Figure 1: graphical view Logit

Because of the fact that we want to estimate and predict the choice of the consumer, we need to define the latent utility variable. The latent utility is defined according to equation 1.

(1)

Eventually we want to predict choice with utility, so these variables have to be linked in an accurate way. The linking method between utility and choice is defined according to equation 2.

(2)

We assume that the error term of equation 1 is independently distributed according to the Cumulative Distribution Function (CDF). The estimation of the logit according to the CDF provides us with the logistics distribution (equation 3). The traditional logit makes use of Maximum Likelihood Estimation (MLE), a statistical procedure will try to find for what parameter estimates the model fits the data best. GEE models make use of Quasi Likelihood Estimation (QLI) for the very same purpose. The outcome is invariably the same in terms of parameter estimates, although these cannot be estimated for specific subjects in the case of QLI. However, GEE models do provide the population-averaged estimates (Halekoh, Højsgaard, & Yan, 2006).

(21)

4 Results

This chapter of the report contains relevant information with regard to the outcomes of the statistical modelling procedures described in the previous section. The subsections of this chapter present the (significant) main effects of independent variables as well as the potential interaction effects between those same variables. Furthermore, the predictive validity of the outcomes is treated. The parameter estimates of the significant main effects including interactions model are interpreted since they provide us with the most accurate coefficients.

4.1 Significant Main Effects

The parameter estimates of the main effects model are presented in table 2.

Variable Std. Err. P-value Exp(B)

App behavior Nr. of logins 0.133 < 0.001*** 1.222 Level of activity 0.003 < 0.001*** 1.017 Level of point collection

o None 0.126 < 0.001*** 0.518

o 1 – 2000 0.126 0.003** 0.690

o >2000 (reference category)

Level of reward redemption: 0.022 0.003** 0.936 Level of challenge participation 0.072 < 0.001*** 1.493 Level of interactivity 0.001 0.011** 0.997 Weekgoal participation 0.023 < 0.001*** 1.720 Weekgoal completion 0.030 < 0.001*** 1.536

Customer Characteristics

Health Initiative App membership length

o 6 – 12 Months 0.029 < 0.001** 1.611 o 12 – 18 Months 0.028 < 0.001** 2.215 o 0 – 6 Months (reference category)

A health insurance company customer 0.027 < 0.001*** 1.435 Health Initiative newsletter unsubscribed:

(22)

Table 3: Parameter estimates (Signif. codes: 0.001 "****", 0.01 "**", 0.05 "*")

4.2 Drivers of continued app usage

The parameter estimates of the main effects model are interpreted by means of odds ratio calculation. The odds ratio is calculated by dividing the probability of continued app usage (1) by the probability of observing churn (0). In the case of an odds ratio assuming the value 2, the probability of observing continued app usage is twice as high as opposed to observing churn (Leeflang et al., 2015).

4.2.1 Health Initiative App behavior

The number of days logged into the Health Initiative app has a significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. By interpretation of the odds ratio we infer that of every one-unit increase in the number of days that a user has logged in, the odds of continued app usage increase with factor 1.222.

The level of activity has a significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. It can be inferred from the odds ratio that for every additional activity of the user, the odds of continued app usage increase with factor 1.017. Let us say that a user shows 30 activities in a particular month, then the odds of continued app usage increase with factor 1.66 (1.017^30 = 1.66). In this way, the effect of the level of activity can be calculated by filling in the number of activities.

The level of point collection is divided into three categories. The interpretation of the odds ratio is somewhat misleading here, because it shows a negative factor (exp(B) <1), but it grows closer to one when the level of point collection increases. Thus, an increase in level of point collection ensures a less strong negative effect. Therefore, the level of point collection has a positive effect on continued app usage. The odds ratio shows that for users who have not collected any points, their odds of continued usage decrease with factor 0.518. The odds for continued app usage of users whose level point collection ranges from one until 2000 points decreases with factor 0.690. This is in reference to a level of point collection of more than 2000 points, which is in this case has the least negative effect on continued app usage.

The level of reward redemption, defined by the number of orders per month, is

o November 0.032 0.004** 0.913

o December 0.031 < 0.001*** 0.832

(23)

the calculated odds ratio we infer that for every additional order by an individual app user, the odds of continued app usage decrease with factor 0.936.

The level of challenge participation has a significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. By looking at the odds ratio we see that for one additional challenge participated in by the app user, the odds of continued app usage increase with factor 1.493.

The level of interactivity, consisting of the number of coach questions answered by the app user per month, has a significant (p < .05) and negative (exp(B) < 1) effect on

continued app usage. The interpretation of the odds ratio leaves us to believe that the odds of continued app usage decrease with a factor of .997 for one extra coach question answered.

The variable weekgoal participation has a significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. This variable binary scaled, meaning that users either do or do not participate in weekgoals. The odds ratio shows the strength of the effect in the case that a user is a weekgoal participant. Thus, we infer that by being a weekgoal participant the odds of continued app usage increase with factor 1.72.

Depending on whether a user is a weekgoal participant, the user can succeed or fail in achieving these weekgoals. The variable weekgoal completion shows the number of

weekgoals completed by a user in a particular month. The effect of this variable is significant (p < .001) and positive (exp(B) > 1). The odds ratio indicates that for one weekgoal

completed in a particular month, the odds of continued app usage increase with factor 1.536. In the case that a user would complete the maximum of 5 weekgoals in a particular month, the odds of continued app usage in the next month would increase with factor 8.55 (1.536^5 = 8.55)

4.2.2 Customer characteristics

(24)

significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. The odds for continued app usage of a user in this category increase with factor 2.215. Interpretation of this odds ratio tells us that the odds for continued app usage are twice as high for users in this last category in comparison with people in the reference category (0-6 months).

The variable Health Initiative newsletter unsubscribed indicates whether a Health Initiative app user has or has not unsubscribed from the general Health Initiative newsletter. For some users it is unknown whether they have or have not unsubscribed from the general Health Initiative newsletter. The effect of having unsubscribed for the newsletter is

significant (p < .001) and negative (exp(B) < 1). The odds ratio indicates that when a Health Initiative app user has unsubscribed for the general Health Initiative newsletter, the odds of continued app usage decrease with factor 0.547. The effect of it being unknown whether a user has unsubscribed or not also has a significant (p < .05) and negative (exp(B) < 1) effect on continued app usage. According to the odds ratio, the odds for continued app usage

decrease with factor 0.751 when it is unknown. Both categories are interpreted in reference to the “no” category.

The age variable indicates the age of the Health Initiative app user in years. Age has a significant (p < .001) and positive (exp(B) > 1) effect on continued app usage. The odds ratio indicates that when a Health Initiative app user ages with one year, the odds for continued app usage increase with factor 1.009.

4.2.3 Seasonality

The month variable indicates the effect that a particular month has on continued app usage. The interpretation of this variable estimate is per month. For the month of September, the effect is significant (p < .001) and positive (exp(B) > 1). The odds ratio indicates that the odds for continued app usage are increased with factor 2.117 in the month of September.

The month of October has a significant (p < .001) and positive (exp(B) > 1) effect. The odds ratio indicates that users’ odds for continued app usage increase with factor 1.189 in the month of October.

(25)

For the month of December, the effect is significant (p < .001) and negative (exp(B) < 1). From reading the odds ratio, we infer that the odds for continued app usage decrease with a factor 0.832 in the month of December.

4.3 Interaction effects

It can be the case that the effect of an independent variable to a certain degree depends upon the value of another independent variable. Estimating such interaction effects improves interpretability of a model. Table 3 presents the significant interaction effects that have been found in the model.

Variable Std. Err. P-value Exp(B)

Level of Login * Age 0.001 < 0.001*** 1.001

Level of Activity * Age 0.001 0.002** 1.000

Level of Weekgoal Completion * Age 0.002 0.005** 0.994 Table 4: significant interaction effects (Sig. codes: 0.001 "****", 0.01 "**", 0.05 "*")

As shown in table 3, the age variable interacts with the variables level of login, level of activity and level of weekgoal completion. The interaction between age and the level of login is significant (p < .001) and positive (exp(B) > 1). Thus, the effect of the variable level of login on the odds for continued app usage increase with 0.1% when the age of a user increases with one year.

The interaction between age and level of activity is significant (p = .002), but does not show any kind of additional predictive power to the model (exp(B) = 1). The effect of the level of activity remains the same as it was when the age of a user increases with one year.

The interaction effect between age and the level of weekgoal Completion is significant (p = .005) and negative (exp(B) < 1). This means that the effect of the level of weekgoal Completion on the odds for continued app usage decrease with 0.6% when the age of a user increases with one year.

4.3.1 Multicollinearity

(26)

This calculation is performed for all models. The outcomes of the VIF calculations are presented in appendix 7. Generally, a VIF score higher than four is problematic (Miles & Shevlin, 2001). Such values (>4) were only to be encountered in the in the interaction model, where they are a logical side effect of the intended incorporated interaction effects, thus where they do not cause harm to the interpretation.

4.4 Validation

An important step in the model building process is the validation of the model. In this

paragraph, different models are compared to each other in order to test the predictive validity and ultimately to select the most valid model. The models to be compared are a main effects model, significant main effects model, interaction model and a null model.

4.4.1 Validity metrics

The measures taken into account are the Hitrate (out-of-sample), Top Decile Lift (TDL), GINI coefficient and the Mean Squared Error (MSE). Table 4 presents the outcomes of the first four validity measures between the models.

The out-of-sample hitrate of a model constitutes the amount of “hits” a model generates when it is confronted with new data. This means that a model is trained with a certain part of the data (training set), and tested for its performance with new data (test set). In this study, the data was divided into a training set consisting of 75% of the data and a test set consisting of the remaining 25% of the data. Thus, the hitrate in table 4 shows the out-of-sample predictive performance of the different models.

The Top Decile Lift (TDL) is a metric that shows the predictive performance of a model on the top decile. The top decile consists of the 10% of observations in the sample with the highest model predictions. Berry and Linoff (2004) expect to find a TDL of 1 for a random model, whereas a TDL of 3 would indicate a tripled number of correctly predicted cases when compared to the random model.

(27)

The Mean Squared Error (MSE) gives an indication of the amount of error in the predictions of a model. In the context of this study we are looking for a model with a low MSE, because a model with a low MSE implies low average error in the predictions produced.

Main Effects

Sig. Main Effects Interaction Model Null Model Hitrate (out-of-sample) 85.1% 85.1% 85.4% 81.6% TDL 1.22 1.22 1.22 1.06 GINI 0.263 0.263 0.264 -0.0134 MSE 0.1 0.1 0.1 0.154

Table 5: Hitrate, TDL, GINI and MSE across models

4.4.2 Main Effects Model

The main effects model contains all the variables that were to be tested according to the hypotheses (see chapter 2). Therefore, this model contains an insignificant parameter. The out-of-sample hitrate of the main effects model is 85.1%. This means that the model predicts 85.1% of the new data correctly. The TDL for this model is 1.22, indicating that it makes better estimates for the top ten per cent of the population with the highest probability of continued app usage than a random model. The GINI value of 0.263 shows that there exists variability within the data. The MSE for this model is 0.1.

4.4.3 Significant Main Effects Model

The significant main effects model contains all the variables from the main effects model that proved to be significant (p < .05). Because of this it contains one parameter less than the main effects model. Nevertheless, the validity measures for the significant main effects model are exactly the same as for the main effects, thus showing no significant improvement in

(predictive) validity.

4.4.4 Interaction Model

(28)

slightly higher (0.264). The MSE of the interaction model has the same value as the other models.

4.4.5 Null Model

The null model is a model that solely exists of an intercept, thus not containing any

independent variables. The null model has a hitrate that is somewhat lower than that of the models that include independent variables (81.6%). However, this fairly high predictive power of the null model is easily explained. The data is distributed in a way that roughly 80% of the observations continue to use the Health Initiative app, resulting in a 80/20 distribution in terms of the dependent variable. The intercept of the model is higher than 0.5, thus it predicts a 1 (continued app usage) for every observation it encounters. This automatically results in a relatively high hitrate. However, the difference in predictive validity with the other models comes to show in the values for the TDL (1.06), GINI (-0.0134) and MSE (.0154), which are clearly worse measures.

4.4.6 Model selection

Based on the outcomes of the validity metrics (table 4), the interaction model shows the highest predictive validity. Thus, the interaction model seems like a logical choice to use for parameter estimate interpretation and ultimately prediction. A double check is performed by means of a Likelihood Ratio (LR) anova test, which chooses one best performing model from a selection of two. The anova test returns a Chi-Squared statistic, also known as the

(29)

Main Effects Sig. Main Effects

Sig + Interaction Null

Main Effects (p > .05) Sig + Interaction (p < .01**)

Main (p < .001***)

Sig. Main Effects (p > .05) Sig + Interaction (p < .01**)

Sig. Main Effects (p < .001***)

Sig + Interaction Sig + Interaction (p < .01**) Sig + Interaction (p < .01**) Sig + Interaction (p < .001***) Null Main (p < .001***)

Sig. Main Effects (p < .001***)

Sig + Interaction (p < .001***)

(30)

5 Discussion

The goal of this research was to find significant factors that influence continued app usage. Multiple significant effects have been established. The objective of this chapter is to interpret the results more in-depth. First, table 6 shows whether the findings support or reject

hypotheses from paragraph 2.2. In the case that a hypothesis is rejected, a note is added for the true effect that was encountered. From the fourteen initial hypotheses, ten were supported by the findings. Of the remaining four hypotheses, one could neither be supported nor

rejected due to a lack of statistical significance (p > .05). The variables of the other three hypotheses were statistically significant, but showed an effect in the opposite direction.

Hypothesis Verdict Note H1: Weekgoal participation has a positive effect on

continued app usage. Supported

H2: Weekgoal completion has a positive effect on

continued app usage.

Supported

H3: The level of activity has a positive effect on continued

app usage. Supported

H4: The number of days logged into the Health Initiative

app in the previous month has a positive effect on continued app usage in the next month

Supported

H5: The level of point collection has a positive effect on

continued app usage Supported

H6: The level of reward redemption has a positive effect

on continued app usage

Rejected The level of reward redemption has a significant and negative effect on continued app usage

H7: participation in a challenge has a positive effect on

continued app usage. Supported

H8: The level of interactivity with the Health Initiative

coach is positively related to continued app usage.

Rejected The level of interactivity has a significant and negative effect on continued app usage

H9: Age has a negative relationship with continued app

usage. Rejected Age has a significant and positive effect on continued app usage

H10: lifetime as a member of Health Initiative (in years)

has a positive effect on continued app usage. Rejected No significant effect was found (p > .05)

H11: lifetime as a user of the Health Initiative app has a

positive effect on continued app usage. Supported

H12: Being a A health insurance company customer has a

positive effect on continued app usage. Supported

H13: Unsubscribing for the Health Initiative mail /

(31)

H14: The months November and December have a

negative effect on continued app usage.

Supported

Table 7: hypotheses revisited

5.1 Conclusions

This paragraph draws conclusions from the results chapter and ties them together with the hypotheses that were proposed earlier in the report. At the same time, we reflect on the theories on which the hypotheses were based and will accommodate our views in order to also extract insights from the findings that were contrary to expectation.

5.1.1 App behavior

Six out of eight hypotheses of this subsection were supported by the model. Next to that, all of them showed a statistically significant effect. Weekgoal participation was expected to have a positive effect on continued app usage. This hypothesis is supported by the findings. Thus, users who participate in weekgoals are more likely to continue the app in the next month. From this, we can infer that users value the weekgoals or the outcomes that accompany them. We conclude that users who participate in weekgoals to some extent perceive the Health Initiative app to be more useful than users who do not participate. The second hypothesis stated that weekgoal completion has a positive effect on continued app usage. This statement was also supported by the results. It is concluded that weekgoal completion ensures an increased likelihood of continued app usage.

The level of activity was expected to have a positive effect on continued app usage and this hypothesis was supported by the findings. This hypothesis originated from Flaherty’s (2019) behavioral engagement construct. Users who show more activities in one month spend more time and effort and are therefore more likely to continue to use the app in the next month. The level of login was likewise expected to have a positive effect on continued app usage and was also supported by the findings.

(32)

The level of reward redemption was hypothesized to have a positive effect on

continued app usage. This hypothesis was rejected by the model, which showed a significant and negative between level of reward redemption and continued app usage instead. It was expected that the rewarded behavior mechanism would positively affect consumer responses behaviorally as well as attitudinally (Palmatier et al., 2009; Taylor & Neslin, 2005). The findings suggest that the level of reward redemption has a negative effect on continued app usage. A possible explanation is that users of the Health Initiative app might be more outcome-oriented than people who are more intrinsically motivated and there is sample-selection bias at play (Heckman, 1979; Leenheer et al., 2007). Achieving the reward might then lead to satiation and consequentially remove the incentive for continued app usage.

Level of challenge participation was expected to have a positive effect on continued app usage. Based on literature in the domain of social influence, we expected that when a user gets confronted with a social community and social competition, this user is more likely to continue to use the app than users who do not participate in such challenges (Ding & Chai, 2015; Peng et al., 2016). This hypothesis was supported by the model, meaning that a

positive and significant effect of the level of challenge participation on continue app usage was found.

The level of interactivity was expected to have a positive effect on continued app usage. This direction of this effect was based on literature that proposed interactivity to lead to increased emotional engagement (Ding & Chai, 2015; Flaherty et al., 2019; Vodanovich et al., 2010). Nevertheless, the findings suggest otherwise. The model suggests a significant and negative effect of the level of interactivity on continued app usage. A potential explanation might be that sample-selection bias also plays a role here. Answering coach questions is also a way to obtain (a marginal amount of) points and this feature thus might fall prey to users who want to gain a quick buck instead of truly valuing the feedback.

5.1.2 Customer characteristics

Age was hypothesized to have a negative effect on continued app usage, indicating that the older the user, the less likely continued app usage becomes. The model rejected this

(33)

2018). The data suggests that older people are more likely to continue to use the app.

Incorporating these new findings with existing literature leaves us to believe that older people are less likely to adopt new technologies, but are also less likely to show volatile behavior. Once they are used to a technology, they are less likely to switch to another technology.

The total lifetime of a consumer as a user of Health Initiative was expected to have a positive effect on continued app usage. This hypothesis was not supported by the model due to a lack of statistical significance. It is therefore concluded that the total lifetime of a consumer as a user of Health Initiative does not have an effect on continued app usage. However, the lifetime of a consumer as a Health Initiative app user does prove to be

statistically significant. It was hypothesized that the relationship length as a Health Initiative app user has a positive effect on continued app usage. This hypothesis was supported by the model. This is in line with previous research that found continued app usage to be sparse for mhealth apps, especially short after installation (Ding & Chai, 2015; Kayyali et al., 2017; Vaghefi & Tulu, 2019). This means that people who have been using the Health Initiative for a longer period of time are more likely to continue to use the app.

Being a customer of A health insurance company for health insurance while also making use of the Health Initiative app was hypothesized to have a positive effect on continued app usage due to increased behavioral engagement (Flaherty et al., 2019). This hypothesis was supported by the findings, indicating that A health insurance company customers are more likely to continue to use the Health Initiative app in comparison with users who have arranged their health insurance elsewhere.

It was hypothesized that users who have unsubscribed from Health Initiative content / mail are less likely to continue to use the app. This expectation finds its roots in research on information provision, ultimately leading to cognitive engagement (Ball et al., 2014;

Hollebeek, 2011). The outcomes of the model supported this hypothesis, thus concluding that unsubscribing for the Health Initiative content / mail has a negative effect on continued app usage.

5.1.3 Seasonality

(34)

5.2 Limitations and future research

The current study naturally has its limitations and it is wise to address these shortcomings, so other researchers in this field of study can learn from the issues that were dealt with. Thus, the aim of this paragraph is to clearly summarize the shortcomings of the current study and to provide potential avenues for future research.

First, this study was performed in coordination with A health insurance company. Data was provided by A health insurance company, resulting in a sample that solely consists of behavioral data for Health Initiative users. This study does not aim to predict or estimate the effects of certain factors for other apps than Health Initiative and is therefore not broadly generalizable. Future research might provide more generalizable finding by working with data from multiple sources. Also, the data sample ranged from the month of July until the month of December. This accounts for almost half a year. Therefore, the findings cannot be compared to the rest of the year while it might just be that the effects differ over time. This is something that cannot be tested by the current model, so we advise future researches to include data for at least one whole year.

Second, the way the data was handled did provide some downsides. The data was arranged in a way that made it possible to predict continued app usage in one month based on the values of variables in the preceding month. This means that the model did not take

behavioral patterns over longer periods of time into account. Future research might choose to arrange the data in such a way that it can include these variables over time.

Third, the definition of continued app usage is self-determined and therefore might exclude certain observations from being correctly categorized as continued app usage or churn. For example, the model does not distinguish between users that have permanently churned and users that have gone through a period of inactivity after which they continued to use the app. Future studies might want to detect such dormant users.

(35)

5.3 Managerial implications

Previous research has shown that low retention rates exist for mhealth apps. For A health insurance company and Health Initiative this means that it is hard to establish a healthy user base. The findings in this report can be used to pave the first steps in overcoming this problem. This paragraph highlights some outcomes in terms of practical use.

Multiple factors have been found to affect continued Health Initiative app usage. The objective for Health Initiative is to develop the app in such a way that consumers use the app for a prolonged time period. Practically, this can be done by trying to diminish the negative effects of certain factors or by increasing the positive effects of others.

Firstly, when the factors that positively influence continued Health Initiative app use are considered, the app behavior of the user is handled. In general, six out of eight of these variables showed to have a positive effect on continued app usage. The level of login resembles the amount of days of a month during which a user has logged into the Health Initiative app. The fact that this has a positive effect is an indication that users who are consistently logging into the app during a month are more likely to continue to do so in the next month. Thus, it seems fruitful for Health Initiative managers to stimulate such habitual behavior within their user base.

The Level of Activity, Point Collection, Challenge Participation, Weekgoal Participation and Weekgoal Completion all show a positive and significant effect on

continued app usage. These are factors that require somewhat more “active behavior” of the user in comparison with the level of login. These factors are an indication that a user is more inclined to keep using the app when he/she is in the middle of an active process. Current activity functions as a catalyst for future activity.

(36)

reward itself. The difficult task for Health Initiative managers is to distinguish the users who are intrinsically motivated and who value the processes that the Health Initiative app offers from the users who are motivated by external factors like rewards. The findings of this report provide helpful insights for the task at hand.

In terms of Customer Characteristics also some valuable information has been found. The longer a consumer makes use of the Health Initiative app, the more likely this user is to continue using the app. For example, a person who has been using the app somewhere

between twelve and eighteen months is more than twice as likely to continue to use the app in the next month comparison with a person who has been using the app between zero and six months. Thus, continued app usage has a self-reinforcing effect. One way to start this

virtuous cycle is to incentivize prolonged app usage. We have observed that material rewards (number of orders) do not necessarily create the effect that we seek here, so it would be wise to focus on more process-oriented incentives. For example, by unlocking special app features after a certain period of usage.

Users who have unsubscribed from the Health Initiative newsletter their chances of continued app usage are almost cut in half in comparison with users who still receive it. Theories on information provision claimed that cognitive engagement arises when relevant information is offered (Ball et al., 2014; Hollebeek, 2011). Some users who have

unsubscribed probably just do not want to be bothered with any information. On the other hand, it might be the case that other users have unsubscribed because they do not want to receive all information, but might consider to subscribe to a content that is relevant to them. Relevance is in essence a subjective criterium. Offering the users more options to receive Health Initiative content instead of the current binary option (yes/know) can increase cognitive engagement.

(37)
(38)

Appendices

Appendix 1a: Hypotheses

# Hypothesis +/-

1 Weekgoal participation has a positive effect on continued app usage. + 2 Weekgoal completion has a positive effect on continued app usage. + 3 The level of activity has a positive effect on continued app usage. + 4 The number of days logged into the Health Initiative app in the previous month has a

positive effect on continued app usage in the next month + 5 The level of point collection has a positive effect on continued app usage + 6 The level of reward redemption has a positive effect on continued app usage + 7 Participation in a challenge has a positive effect on continued app + 8 The level of interactivity with the Health Initiative coach is positively related to continued

app usage. +

9 Age has a negative relationship with continued app usage. - 10 Lifetime as a member of Health Initiative has a positive effect on continued app usage. + 11 lifetime as a user of the Health Initiative app has a positive effect on continued app usage. + 12 Being a A health insurance company customer has a positive effect on continued app

usage.

+ 13 Unsubscribing for the Health Initiative mail / content has a negative effect on continued

app usage. -

14 The months November and December have a negative effect on continued app usage. - Table 1: Hypotheses

Appendix 1b: Research variables

Variables Description Scaling

Weekgoal participation Whether a user participated in a Health Initiative weekgoal in a particular month

Binary: 0 = no, 1 = yes

Weekgoal completion Number of weekgoals completed

by the user in a particular month Numeric (0-5)

Level of activity The number of activities of the

user in a particular month Numeric (0-200)

Level of login The number of days a user has logged into the Health Initiative app during a particular month (M)

Numeric (1-31)

Level of point collection Number of points collected by user

in a particular month Categorical: o 0 o 1 – 2.000 o >2.000

(39)

Level of challenge participation Number of challenges a user has participated in during a particular month

Numeric (0-7)

Level of interactivity Number of interactions with the Health Initiative health coach in a particular month

Numeric (0-200)

Age The age of the user Numeric (16-119)

User lifetime (Web) The lifetime of the user in the

Health Initiative domain in years Numeric (0-10)

User lifetime (App) The lifetime of the user in the SamenGezoind app domain

Categorical: o 0-6 months o 6-12 months o 12-18 months

A health insurance company customer

Whether a Health Initiative user is a customer at A health insurance company

Binary: 0 = no, 1 = yes

Unsubscribed Health Initiative newsletter

Whether an individual has unsubscribed for the newsletter of Health Initiative

Character: yes/no/unknown

Month The particular month Character (August-December)

Continued app usage (DV) Whether a Health Initiative app user has used the app at least once in the next month (M+1)

Binary: 0 = no, 1 = yes

Table 2: Variables overview

Appendix 2: Descriptive statistics

(40)

Appendix 3: Models output

Table 9: Output Main Effects Model

(41)

Table 11: Output Interaction Model

Appendix 4: VIF scores

Table 12: VIF scores Main Effects Model

(42)

Table 14: VIF scores Interaction Model Appendix 5: Model validation

Figure 2: Lift curve, TDL and GINI (Main Effects Model)

(43)

Figure 3: Lift curve, TDL and GINI (Interaction Model)

Figure 4: Lift curve, TDL and GINI (Null Model)

Main Effects

Sig. Main Effects Interaction Model Null Model Hitrate (out-of-sample) 85.1% 85.1% 85.4% 81.6% TDL 1.22 1.22 1.22 1.06 GINI 0.263 0.263 0.264 -0.0134 MSE 0.1 0.1 0.1 0.154

Referenties

GERELATEERDE DOCUMENTEN

The good memories of music tours provided them with a desire to experience the exhilaration of performing and listening, a desire to make meaningful connections

Driving performance measures do indicate that novice drivers are ill prepared and smooth, error free performance is not achieved after formal driver

Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of

Five factors that might have an effect on customer satisfaction and purchase intent, which drive people to use mobile applications, were chosen from the literature (i.e.

In conclusion this study found statistically significant differences in maternal ages between the two groups, with younger women delivering by CS; gravidity and parity was lower in

Keywords: Appreciative Inquiry; Generative Change Process; Alteration of Social Reality; Participation; Collective Experience and Action; Cognitive and Affective Readiness

Five constructs: (1) Facebook Intensity, (2) Electronic word-of-mouth, (3) Perceived valence of information, (4) User-generated content sensitivity and (5) Perceived

website including face value, regular price, redemption location, type of product and expiration date of coupons were used.. 3.1 Methodology