Don’t say Sorry, say Thank You!

(1)

Don’t say Sorry, say Thank You!

An empirical study on the effectiveness of online service recovery strategies on

Twitter and the moderating effect of companies’ overall satisfaction level on

complainant satisfaction.

By: Damianos Michailidis University of Groningen Faculty of Economics and Business

(2)

1

Acknowledgements

(3)

2

Abstract

This thesis addresses the effects the different service recovery strategies have on customer sentiment in an online service recovery context and also examines the moderating role of a company’s overall satisfaction level in the above relationships. The study contributes to the currently expanding online service recovery literature by examining Twitter data obtained by customer support pages of various companies on Twitter. From the strategies that were examined, expressing gratitude appears to be the most effective one, compared with apology and willingness to listen. Response time is found to have an inverted U-shaped relationship with sentiment and response length has a negative relationship with it. Finally, the overall satisfaction level of a company is found to have a positive interaction effect with all the examined strategies and a negative one with response time and length.

(4)

3

1. Introduction

Nowadays, people are communicating and interacting with each other in a very different way than a couple of decades ago. Digital information exchanges were happening already back in the 1970s, the Arpanet network was available for use since 1983 and common individuals started to use them from 1993 (Deighton & Kornfeld, 2009). This gave individuals, more than ever before, the ability to communicate and exchange information effortlessly from everywhere around the world in real time. This can be extended to the way customers communicate and exchange information regarding products and services and interacting with companies (Hanssens et. al., 2009). The appearance of new digital media like Facebook, Myspace, eBay, Amazon, YouTube and many others gave customers the ability to “…talk back and talk to each

other” (Deighton & Kornfeld, 2009, p.4) in regards to products and services. This development

and spread of many social media platforms, old and new, have disrupted the status quo and gave consumers the power to question and openly comment on the behaviors of companies and brands online sifting the power balance (Deighton & Kornfeld, 2009).

(6)

5 valuable than in a “traditional” setting (Einwiller & Steilen, 2015). If the customer is satisfied after complain handling, then this can result in increased positive WoM and even increase purchase intention (Kau & Loh, 2006). Further, a reasonable complaint resolution is also read/seen by other social media participants and can create positive spillovers. It is understandable then why big emphasis needs to be given in online service recovery techniques and strategies and why it is crucial for the company that this will be done right.

Service Recovery

According to Spreng et. al. (1995), when a service failure occurs the satisfaction of the customer is going to be affected by two different factors. The first factor is going to be the characteristics of the specific service failure that the customer faced and the second being the attributes of the service recovery process. With this in mind service recovery can be defined as all the actions a

company does in order engage with the customer and address a complaint (Gronroos, 1988). An example of this would be that in an i-pad, an app is not working properly after the last software update was implemented (first factor) and the way that the customer can contact the company in order to resolve this (second factor). Even though service recovery is an outcome of a service failure it can often has positive results in dimensions like customer satisfaction and customer loyalty (Hart et. al., 1990; Kelley & Davis, 1994). It is understandable so why an emphasis is put from a lot of companies to facilitate this process as good as possible, since “every customer’s problem is an opportunity for the company to prove its commitment to

service – even if the company is not to blame” (Hart et. al., 1990, p.151).

Twitter

(7)

6 without needing approval from them, unlike Facebook that if you want to follow someone this person needs to accept your friend request (Cha et. al. 2010). These and other properties of the platform make Twitter an avenue that is considered to be very effective in regards to communication, industrial use and marketing (Know et. al., 2014). Last year the company’s revenues reached 3.04 billion US dollars, with most of them coming from advertising. In fact, in 2018 the number of monetizable daily active Twitter users reached 126 million worldwide (Twitter, 2019).

With the above statistics in mind, it is easy to see why Twitter has the potential to be a very useful tool for companies that can be used for service recovery purposes. Complainants are already on Tweeter, a lot of times being vocal about their problems and using hashtags that can go viral really fast. A lot of companies, especially in the U.S. where the use of the platform is even more prominent, encourage the service recovery attempts though Twitter, considering it a great opportunity to win the customer back (Abney et. al. 2017). A lot of big companies like Amazon, AT&T, Apple etc. have created their own Twitter pages dedicated to customer support, like @AmazonHelp, @ATT, @AppleSupport.

There is a growing academic interest regarding the effectiveness of the aforementioned service recovery services. Unfortunately, there is still a lack of understanding on how customers react and evaluate the corporate service recovery efforts and strategies, in online settings like Twitter (McCollough et al., 2000; Holloway and Beatty, 2003; Abney et. al. 2017).

(8)

7 the actions (in terms of complaint satisfaction in Twitter) they take. In the following chapters a review of the relevant literature for the topic is discussed and specific hypotheses are formulated. Next, the results of the study regarding the hypotheses are shown and in the end a discussion about the results and managerial implications are provided. From customers’ perspective, they seem to react better when gratitude is shown to them and not that well, when they get an apology or when a lot of information it is given to them. Gratitude has a positive effect on their sentiment making them essentially feel better and it is making them feel that the person taking them respects and understands them. On the other hand, apologies and explanations negatively affecting customer sentiment since they can be perceived as not honest and ineffective at resolving issues.

(9)

8

2. Theoretical Background

This part of the thesis includes the essential literature needed in order to have a meaningful analysis and discussion of the topic. The theoretical background includes specific information used to define the key variables of this study, formulate hypotheses that explain the relationships between them and form a conceptual model.

2.1 Social media complains

Social media platforms like Twitter provide consumers with an easy way to voice their complaints in case a product/service failure. This has put added complexity in traditional complaint management procedures, shifting the power balance between consumer and company towards the side of the consumer, since comments and issues are public and companies need to be very careful in how they respond and deal with them. People expect brands to be honest, accepting and open towards their mistakes (Xia, 2013). Appropriate response can have a positive impact on customer satisfaction, brand perception, higher purchase intention and better word of mouth (WoM) (Xia, 2013).

Addressing customer complains on social media is something that can be described as a service recovery (Kau & Loh, 2006). The customer has an issue with the product/service and contacts customer support to resolve it. Generally, the responsiveness to complains on social media is considered moderate to low, as shown by Martiz (2011) that found a 33 percent response rate on complaints voiced through Twitter. There is not yet a theory that includes all the facets of service recovery, according to the review by Abney et al. (2017).

(10)

9 needs to be given in attentiveness, since it is a complex dimension that refers to the interaction between the company and the customer (Davidbow, 2003). This dimension gives more weight to the human interaction between the representative and the complainant. Willingness to listen, empathy and extra effort from the representative can all be accounted here. In their article that examined complains on social networks, Einwiller & Steilen, (2015) split this dimension in three different variables, respect, empathy and willingness to listen. The above theoretical framework is going to be used in order to define and formulate the hypotheses mentioned on this section.

2.2 Customer sentiment

Sentiment analysis is a natural language processing task and specifically a part-of-speech tagging (POS) (Agarwal et al., 2011, Kumar et al., 2016). According to the work of Saif et al. (2012) examining sentiment on Twitter has a lot to offer to a company, since it is an effective way to see the publics’ attitude towards their company, specific brands, business practices, etc. Twitter sentiment analysis specifically, is a very interesting case, characterized on average by short length of data and irregular structure (semi-unstructured).

(11)

10 does not include the word not (Naldi, 2019). It becomes apparent that choosing the right tool that can mitigate these issues is crucial for obtaining high-quality results.

In this study, we look in customer support data from the official support Twitter page of various companies and we are identifying the sentiment of the customers’ first response, after the company has responded to their questions. We use the sentiment by command from the sentimentr package in R, to extract the sentient of the tweet, scoring the tweet in regards of positive and negative properties and giving us a final overall score. This package was chosen in order to address the challenges discussed above. It uses a sophisticated way to score each sentence in a continues scale, taking into account the balance shifters. More details are going to be provided in the Methodology section.

Websites like Twitter are considered microblogs and they can hold a lot of different kinds of information. In this setting customers can react and comment in real time for the things they like (positive sentiment) and don’t like (negative sentiment). It is an established practice for companies to use these blogs and look for customer reactions to specific actions they take. A lot of times they actually reply/help their customers through these platforms (Agarwal et al., 2011), doing things like customer support, advertising and damage control. Keeping customers satisfied especially in social media has the added benefit to prevent negative WoM that it is really easy to spread in platforms like Twitter or Facebook (Kwon et. al., 2014; Einwiller & Steilen, 2015).

(12)

11

2.3 Company Response time

In this and the next parts of the Theoretical background the formulation of specific hypothesis is going to be discussed. Different aspects of service recovery will be investigated and the relationship between them and the DV will be examined. Let’s start with how fast τhe company responds to a complaint.

In the literature the word timelessness in used to signify how fast a company is able to respond in customer complains (Einwiller & Steilen, 2015). Older empirical studies were supportive of the idea that response speed can have a lot of positive benefits like increase repurchase intentions and decreased amount of negative WoM (Conlon and Murray, 1996). This make sense, since the faster the company responds the safer the customer can feel, thinking that someone is always there, ready to help.

Despite these earlier findings, later papers like Davidbow (2000) and Einwiller & Steilen (2015) that examined timelessness in the context of customer support satisfaction, concluded that it has no effect on complainant satisfaction. Thinking across these lines, this probably happens because the most important factor for complainants to be satisfied with the customer support is if they were able to solve their problem. A fast response can often be an automated response, just asking them to send a private message in another person. This has been shown to actually have negative results in customer satisfaction (Davidbow, 2003).

Since conflicting pieces of evidence seem to exist in regards to the effects of time in company responses, a negative effect is going to be assumed in order to create a hypothesis eligible for testing. Having said that, the following hypotheses can be formed.

H1: Company response time in a complaint have a negative impact on customer

sentiment.

2.4 Tweet length

(13)

12 the complaint and shows a level of care for the customer. Various researchers have found positive correlations between these attributes and customer satisfaction (Coombs and Holladay 2008, Davidbow 2003). This helps in creating an environment that is perceived by customers as more responsive and attentive, thus making them more likely to express a complaint or problem through the company channels (Davidbow, 2000).

On the other hand, Einwiller & Steilen (2015) showed in their research that customers whose complains received more explanation were not more satisfied than others. In some cases, they were even a bit less satisfied. Hence, it is very interesting to see the effect that the length of the response will have on sentiment. So, according to the above empirical findings the following hypotheses were developed:

H2: The length of the company’s initial response tweet has a positive impact on

customer sentiment.

2.5 Apology Strategy

(14)

13 Finally, Know et al. (2014) found that apologizing to the consumer doesn’t have a significant effect on complaint satisfaction. A lot of times an apology can create the impression that nothing else is going to be done. Therefore, the following hypotheses can be formed:

H3: Apologizing to the customer has a negative impact on customer sentiment.

2.6 Expressing gratitude

Another way to handle customer requests, is trying to be sympathetic to them and the issue they are facing. Instead of accepting responsibility and apologizing for a situation you can express sympathy towards them and make them feel that you understand them (Coombs and Holladay, 2008).

Gratitude is connected with the respect and empathy aspect of attentiveness that Einwiller & Steilen (2015) mention in their research. It can both improve or worsen a situation; it is a complex interaction between the company representative and the complaint that needs to be done right. Saying “thank you” can drastically help in this direction and be part of an attentive answer that is making the person feel better. If that happens, showing gratitude has been shown to significantly affect complaint satisfaction, more so that other response ways and can also decrease the potential negative word-of-mouth. (Einwiller & Steilen, 2015, Davidbow, 2000). Based on these we can form two new hypotheses:

H4: Expressing gratitude has a positive effect on customer sentiment.

2.7 Inquiring further information – Willingness to listen

Complainants are often not specific enough when expressing a problem in customer support. Further information is a lot of times necessary in order for the issue to be resolved. Asking is also showing a willingness from the company to listen and this attribute is connected with an overall attentive response strategy (Davidbow, 2003) and specifically with the willingness to listen aspect (Einwiller & Steilen, 2015) meaning that positive effects can be expected.

(15)

14 H5: Inquiring more information has a positive effect on customer sentiment.

2.8 Moderating role of company’s overall customer satisfaction level

Customer satisfaction is a very important aspect of form performance. If a company manages to create satisfied customers then this enchases positively many other corporate metrics. It is proven that customer satisfaction is an indicator of overall firm value and can have positive impact on customer loyalty, usage behavior, purchase intentions etc., and many other metrics that improve overall firm performance (Peng et al., 2015). Having high customer satisfaction can give the company can also have positive impact on finance metrics like cash flows and stock returns giving a long-term competitive advantage to the firms (Hanssens et al., 2009)

Customer satisfaction is created and affected by both economic and non-economic (psychological) facets of the communication between the company and the individual. In their article Geyskens et al. (1999) distinguish the differences in satisfaction between these two terms. They define economic satisfaction as the members (customer) positive economic rewards towards the company, such as sales volume and margins and the non-economic satisfaction as the positive effects of the member in the psychological aspects of the relationship, meaning that s/he is respectful, concerned and willing in exchanging ideas.

(16)

15 This makes the premise of studying the influence of overall satisfaction level all the more intriguing and adds an originality factor in our findings. What it is interesting is to see if firms with higher satisfaction, these companies are perceived as good companies from the consumers, are getting a “free pass” let’s say, in regards to issues that might come up or if customers will be even more critical if a company with high quality perception fail. Both of these can base of intuition be right. For example, a customer can give a pass to Apple if their iMac has a problem with the most recent update because they are overall very satisfied with apple products. But they can also be more critical and upset as they would usually be with an issue like that because they have paid a premium for a product from a good company and they expect a certain level of quality. This can have important managerial implications, since different response strategies might by more or less effective in different companies.

Based on the literate evidence that customer satisfaction has mostly positive effects and the review of the literature discussed in the previous part of the chapter, the next set of hypotheses is formed:

H6: Customer satisfaction has a negative effect on the relation between customer

sentiment and a) response time and b) apology strategy.

H6: Customer satisfaction has a positive effect on the relation between customer

sentiment and c) response length, d) expressing gratitude, e) willingness to listen.

2.9 Conceptual Model

(17)

16

(18)

17

3. Methodology

This chapter of the thesis contains all the elements used to design and conduct this study. First the characteristics of the dataset are analyzed, giving an overview of the raw materials that are available at our disposal. Secondly, because of the nature of the data, the structuring process is explained, showing all the steps that were taken in order to create a tidy-format dataset. Here the information extraction and text categorization processes are also explained, showing how we transform the qualitative data to quantitative ones, in order to be able to run the necessary statistical analyses. In the chapter’s final part, discussion of model assumptions is necessary for clarification and justification of the statistical methods that were used for result generation.

3.1 Dataset description

To test the hypotheses made in the previous part, a pre-existing dataset was used. This dataset was available through Kaggle’s website (www.kaggle.com) and it was created by author Stuart Axelbrooke. This dataset was created by periodic searches for mentions of company names on Twitter and it was last updated at 04/12/2017. The data present here represent tweets from 08/05/2008 till 03/12/2017, although 2017 tweets take the biggest balk of the data, 99.91%. It contains 2.811.384 entries (rows) and different variables (columns) that are explained in table 1, below. From these entries, 1.537.590 (54.69%) are customer tweets, 1.273.730 (45.30%) are company tweets and 64 (0.01%) of them contain random texts in random column indicating that something went wrong in the data extraction and conversion process.

[Insert Table 1 here]

(19)

American-18 based multi-national corporations, but there are also companies based in other countries like the United Kingdom.

3.2 Data mining

Text data mining, according to Tang and Guo (2015, p.68) is “the analysis of text data in order

to discover hidden patterns, traits, and relationships”. The main goal of text mining is to

transform unstructured and semi-unstructured data (like Twitter data) into a structured form, so it can be analyzed with traditional KDD, knowledge discovery in databases, and statistical techniques. Text analysis has been used to conduct both qualitative and quantitative studies, mainly in psychology, but also in other social sciences and marketing.

In the marketing field, text mining has been used to research topics like: important elements of eWoM (Agarwal et al. 2009), run sentiment analysis (Saif et al., 2012, Agarwal et al., 2011), complaint handling (Einwiller & Steilen, 2015) and topics in many other areas of the field. It is a widely accepted practice, suitable for achieving the goals of the project in the present study.

Text mining can include the following processes (Tand and Guo, 2015): 1. Information extraction (identify important marker items and themes) 2. Text categorization (classify text into pre-determined categories) 3. Text clustering (group similar documents together)

4. Document summarization (summarize the important concepts of a text) 5. Association analysis (finds associations between terms)

(20)

19

3.3 Data cleaning and structuring procedure

A lot effort was dedicated to examining and structuring the dataset in a way that meaningful analysis can be conducted in order to identify relationships between variables. First step was to delete the 0.01% of the data that were not in a usable form. Since these data represent an extremely small percentage of the data and given that we are working with big data the lost in statistical power by this deletion should be insignificant.

Then, it is necessary to create a framework in which questions and responses can be matched with each other, recognizing the complainants and the company’s response to them. This was achieved by splitting the data into company tweets (tweets that were not inbound) and first customer tweets (tweets that had the in_response_to_tweet_id empty, meaning that these tweets were not replies to existing ones but tweets that were direct comments to the customer support page voicing a complaint). Then, these two datasets were merged using a left joint operation that matched the tweet_id from customer tweets with the in_response_to_tweet_id from the company tweets data frame. This created a new dataset that contained 733.350 pairs of questions and company responses.

(21)

20 At this point our data are structured in a similar way with the data used by Einwiller & Steilen (2015) for analyzing Facebook and Twitter complains. They also had their data in a thread format, starting with an independent post by the complainant, continue with company response and then follow up posts. The main difference is in the volume of the data. In our sample there are more than 8 hundred thousand threads while in theirs only 5.023 since they were downloading their data manually from their selected companies adding data every two weeks and categorizing their responses manually. They of course researched customer satisfaction and this study researches sentiment, but method wise there are similarities and it will be interesting to compare results.

3.4 Measures

The next step in the process is to create quantitative values for the DVs, IVs and moderator. In order to extract customer sentiments from their Tweets, a lexicon-based text categorization process was performed using R’s sentimentr package. This package is the only package in r that can successfully account for the existence of negators in sentences, according to the study by Naldi (2019) were he compared all the different sentiment analysis tools available in R. Accounting for them it is very important since they and other valence shifters are very common in social media comments and can affect the meaning that polarized words (words classified as positive or negative by a lexicon). The table below shows how common they are, using a sample of various texts.

[Insert Table 2 here]

(22)

21 sentence or a bulk of text overall, calculating the polarity of the text and giving us a continues output that ranges between -1 and 1.

Time differences were calculated using information already present in the data, when

each tweet was created, and tweet length is calculated by a word count that excludes words with 3 letters or less, taking out of the equation the effect of miscellaneous words like “the” and “a” for example.

To evaluate the effect of the different response strategies binary variables of 0 and 1 were created, one for each strategy. These were created by applying text information extraction processes using handcrafted lexicons for each strategy in order to determine with what category/ies each tweet can be associated with.

[Insert Table 3 here]

In order to obtain customer satisfaction measurements, we turned to ACSI, meaning the American Customer Satisfaction Index. This information is publicly available through the ACSI web site (www.theacsi.org). The index was created by the National Quality Research Center and it measures the perceived quality of goods and services that were produced and purchased in the USA. This index is considered a very good proxy for overall customer satisfaction and it has been used repeatedly in various studies (Peng et al., 2015). The index gives a score between 0 and 100 of more than 230 companies. These companies provide a good representation of the overall industry in the USA (Aksoy et al., 2008).

3.5 Methods

(23)

22 in order to give values, longer answers would inevitably have had a bigger impact, disrupting the results. 𝑀𝑜𝑑𝑒𝑙1: 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟. 𝑠𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑖 = 𝛽0+ 𝛽1𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖+ 𝛽2𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖+ 𝛽3𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖 + 𝛽4𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖+ 𝛽5𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖+ 𝜀𝜄 𝑀𝑜𝑑𝑒𝑙2: 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟. 𝑠𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑖 = 𝛽0+ 𝛽1𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖 + 𝛽2𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖 + 𝛽3𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖 + 𝛽4𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖 + 𝛽5𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖 + 𝛽1𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝛽2𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽3𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝛽4𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽5𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝜀𝜄

Another thing that was considered was the fact that in order to be able to explain the interaction effects in the models that include the moderator, the variables need to be mean centered. So, for interpretation purposes, the variables for satisfaction, response time and response length have been mean-centered for the latter two models.

GLM models are estimated using the Maximum Likelihood Estimation (MLE) instead of the OLS that linear models would use. Maximum Likelihood is a technique that calculates the most likely values of the parameters β, given the data that were observed. For the MLE the variables of the N subjects are assumed to be independent (Leeflang et. al., 2016). The likelihood in this case is expressed as follows:

𝐿(𝜃) = ∏ 𝑓(𝑦𝑖)|𝜃) 𝛮

𝑖=1

(24)

23 A Wald Chi-square test was used to evaluate the goodness of the model. The test returned values less than the critical for all the models, meaning that the assumption that the models predict better than the null model holds and the models are adequate. In all the cases the residual deviance is smaller than the critical value for the corresponding degree of freedom as can be seen in Table 4, below.

[Insert Table 4 here]

Multicollinearity

Multicollinearity for the variables used was also addressed. A VIF scores test was used in order to be sure that multicollinearity was not an issue for these variables. This allows us to correctly interpret our results. In Table 5 the vif-scores for the variables of Model 1 were calculated and it can be seen that since all of them have values less than 3 there is not a multicollinearity issue. The same is true for the same variables in all the models they were calculated.

(25)

24

4. Results

In this part the outputs of the analysis are going to be displayed. In total seven different glm regressions have been conducted, six parametric and 1 non-parametric. The first two explain the base direct effects of the IVs and DV and the interaction effect that our moderator, the company’s overall satisfaction level, has in those relations. This model is going to be used for the initial confirmation or disproval of the aforementioned hypotheses. The other models were created to see if the effect of some variables can be explained better and also for robustness purposes. A non-parametric model was created to examine possible non-linear effects and then two more glm models were created including the squared term of response time. Finally, the two last models include a different version of our DV, sentiment.diff, that represents the increase or decrease of the customer between his first and second tweet.

4.1 Direct effects on customer sentiment

Here the customer’s sentiment score acts as the dependent variable. The model examined is the base one without any non-linear effects. In this model Response time has a significant positive effect (p < 0.001, β = 2.124e-07) on sentiment. This means that a longer response time can increase customer sentiment, indicating that a fast response doesn’t guarantee a more satisfied customer, but on the contrary, can be damaging if it is not meaningful. As mentioned in the theory part, faster responses are in a many cases just automated responses that do not help the complainant. This finding contradicts the first hypothesis (H1) that response time has a negative effect on customer sentiment and rejects it.

Looking at Table 6 below, it is noticeable that Response length has a significant negative relationship with the DV (p < 0.001, β = -3.103e-04). showing that a longer response will make the sentiment of the complainant worse, underlining the importance of efficiency in answers. This disproves our initial hypothesis (H2) that response length would have a positive effect on sentiment.

(26)

25

Apology strategy seems to also have a significant and negative relationship with

sentiment (p < 0.001, β = -6.985e-03), confirming this time the hypothesis that was formed (H3) in the theoretical section. Expressing gratitude on the other hand has a significant positive effect (p < 0.001, β = 5.491e-03) on customer sentiment confirming again our hypothesis (H4). Finally, willingness to listen appears to have a negative significant impact on the dependent variable (p < 0.001, β = -8.074e-03) which contradicts the hypothesis previously stated (H5).

4.2 Interaction effect of satisfaction on customer sentiment

In moderation analysis, in order to say that the effect of the moderator is significant the interaction effect between the two variables (moderator and IV) needs to be significant. As presented in table 7, response time has a significant negative interaction effect with the moderator (p < 0.01, β = -1.277e-07), meaning that customer services of better companies (higher ACSI index) that took more time to address a customer’s complaint had smaller positive effects. This finding confirms the hypothesis H6a that this effect would be negative. Company satisfaction level has a significant negative interaction effect with response length (p < 0.001, β = -5.666e-05), meaning that the moderator re-enforces the negative effect of the base relation. This is making us reject the hypothesis (H6c) that response length has a positive interaction effect with the moderator.

Apology has a significant positive interaction effect with satisfaction (p < 0.001, β =

7.430e-04) rejecting the initial hypothesis (H6b) and showing that an apology coming from a high-valued company holds bigger meaning. Gratitude strategy also seems to have a stronger effect when coming from a high-valued company (p < 0.001, β = 7.887e-04) confirming the initial hypothesis (H6d). Last but not least, willingness to listen also has a significant positive interaction effect with the moderator (p < 0.05, β = 1.520e-04) confirming hypothesis H6e.

(27)

26

4.3 Robustness checks

Robustness is one of the main attributes that a good model should have. According to Leeflang et. al. (2016) robustness refers to quality characteristics that make the model difficult to produce bad answers. A robust model should have minimal changes when small changes happen in the way some measurements work and if some parts of the model’s specifications change (like adding new variables for example). In this chapter different models are going to be computed, with the purpose of complementing the base models and give a more detailed picture of specific effects.

4.3.1 Non-linear effects

A scatterplot examination of the relation between the dependent variable and two of the independent variables, response time and response length showed that the effects of the variables are probably non-linear. In order for us to check that we calculated the same model again including non-parametric relations for the two variables. To do this, the gam() function from the mgcv package in R was used. This package can fit a generalized additive model (GAM)

to data1 and it is going to show us if non-linearity for the above two variables can be assumed. For this model the two variables were mean-centered in order to obtain a better image for the non-linear effects. After calculating the model (table 8, below) and plotting the effects (see

figures 2 and 3 below), it becomes clear that non-linearity can be assumed for the variable

response_time since there is an inverted U-shaped effect on the left part of the plot where almost all of the observation are, but not for the variable response_length since there is not a great departure from linearity.

[Insert Table 8 here]

[Insert Figures 2 and 3 here]

1_{From the package info in the RDocumentation page:}

(28)

27 So, another model, Model3, was created that included non-linear effects for response time. This new model has a lower AIC value and also a lower residual deviance, meaning that it is indeed a better version of the previous model.

[Insert Table 9 here]

The next step is to also compute the moderating model again, including this non-linear effect for response time. The new model, let’s call it Model4, is also a bit better according to the AIC criterion. This re-enforces again our choice to include non-linear effects.

[Insert Table 10 here]

Finally, the equations for models 3 and 4 can be written as follows: 𝑀𝑜𝑑𝑒𝑙3: 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟. 𝑠𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑖 = 𝛽0+ 𝛽1𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖+ 𝛽2𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖2+ 𝛽3𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖 + 𝛽5𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖+ 𝛽6𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖+ 𝛽7𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖+ 𝜀𝜄 𝑀𝑜𝑑𝑒𝑙4: 𝑐𝑢𝑠𝑡𝑜𝑚𝑒𝑟. 𝑠𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑖 = 𝛽0+ 𝛽1𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖 + 𝛽2𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝛽3𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖2 + 𝛽4𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖 + 𝛽6𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖+ 𝛽7𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖 + 𝛽8𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖+ 𝛽9𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽10𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑡𝑖𝑚𝑒𝑖2𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽11𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒_𝑙𝑒𝑛𝑔𝑡ℎ𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽13𝑎𝑝𝑜𝑙𝑜𝑔𝑦𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝛽14𝑔𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖 + 𝛽15𝑤𝑖𝑙𝑙𝑖𝑛𝑔𝑛𝑒𝑠𝑠_𝑡𝑜_𝑙𝑖𝑠𝑡𝑒𝑛𝑖𝑎𝑐𝑠𝑖_𝑠𝑐𝑜𝑟𝑒𝑖+ 𝜀𝜄

(29)

28 squared term of the variable has a significant negative relationship (p < 0.001, β = -1.421e-12) with the DV. This means that a longer response time can increase customer sentiment, but this effect is getting smaller and smaller as time passes and eventually becomes negative. This proves the inverted U relationship of time with sentiment. Looking at the model with the moderation now, the interaction effect is again negative and significant as in the previous case. The interaction between the moderator and the squared term of response time it is not significant.

[Insert Tables 11 and 12 here]

If you look at all the other variables it becomes clear that both of the new models give us the same insights regarding our hypotheses as the models without the non-linear effects. For most of the variables estimates and significances are almost or even exactly identical.

4.3.2 Satisfaction difference model

In order to even further test model robustness, two new models were computed using the difference of sentiment between the first and the second customer tweet as DV. This value shows how much the sentiment of specific customer was changed, increased or decreased, after the company responded to him. The expectation is that if our model is robust enough, this new model will give again similar coefficients, with small differences in the values and definitely same directions and significances of the effects.

One alternative model was calculated for Model1 and Model3 (see Tables 13-14). In all the different occasions the effects maintain their original direction, positive or negative, and they also remain significant, meaning that the results of our model remain consistent and they provide a good representation of the way the examined phenomena work.

(30)

29

5. Discussion

In general, the analysis gave us mixed results in regards to what was expected. Time seems to be an important factor in regards to customer satisfaction, making sentiment more positive as it goes up, but with decreasing rate. This means that a customer is willing to wait a bit in order to get a response that actually solves the problem s/he is facing. The positive relation that the variable has is indicating that companies that rushed to answer fast or just had an automated response that didn’t really solved anything, made the complainants more negative. Of course, the inverted U relationship that the effect has, shows that there is a limit on how long customers are going to understand and forgive a time delay in the answer before their sentiment starts to drop.

This underlines the importance of an effective online customer support function for companies that can respond to questions and complaints, giving effective and useful answers in a reasonable time frame. This finding seems to contrast earlier studies, like the one from Conlon and Murray (1996), that found timelessness to have a positive relation with customer satisfaction, but also the works of Davidbow (2000) and Einwiller & Steilen (2015) that found timelessness to have no impact in customer satisfaction. Combing now this effect with the overall satisfaction level that a company has provided some interesting results. It is shown that people are going to be less patient while waiting for a response from a high-valued company. This indicates that companies, especially the more successful ones with strong brand names, should be extra careful in this regard.

Response length seems to negatively affect sentiment, showing that a shorter and more

(31)

30 The moderating interaction with response length, is indicating that satisfaction with the company can play the role of a conductor, amplifying the negative effect of an unnecessary long response on the sentiment, more so when this happens from a company that is perceived as “good”.

These overall characteristics seem to disagree with papers like Coombs and Holladay (2008) and Davidbow (2000, 2003) that supported that customers will be more satisfied when companies are giving them more attention. The results are more aligned with the paper of Einwiller & Steilen (2015) that found that customers whose complains are receiving more explanation are not more satisfied and, in some cases, they are even a bit less satisfied.

Comparing the different response strategies also leads to interesting implications. The

results of this thesis indicate that apologizing can actually increase negative sentiment reinforcing the claim from Know et al. (2014) that apologizing can give to the complainant the impression that the problem is not going to be solved. It is also contradicting older researches that indicated that an apology strategy can be effective in handling complaints and crises (Goodman et. al. ,1987; Bratford & Garret, 1995; Davidbow, 2000, 2003; Dean, 2004; Wirtz and Mattila, 2003). This could be indicating that the social media environment has its own unique characteristics and properties, so not everything that would be effective in a physical setting is also effective on Twitter. The fact that customers do not appreciate apologies re-enforces the opinion of Davidbow (2000, 2003) that customers often have a negative stance towards apology, since a lot of times they cannot tell the difference between an empty apology and a sincere one.

From the firm perspective now, the findings are also strengthening the argument made by Coombs and Holladay (2008) that in a lot of cases it is better to not use an apology strategy, but some other, like sympathy or compensation, since the first can potentially be much more expensive. Since the results in our case are be negative, apology strategy seems to be a risky one that companies should make very careful use.

(32)

31 by apologizing and showing that they are taking ownership of their mistake and act upon it. For example, if the complainant receives an apology that is perceived as meaningful it will increase his personal satisfaction more and reinforce the good image that the company has in his mind. The other major response strategy that was examined, expressing gratitude or sympathy strategy, seems to be more effective than apology strategy. Not like apology, it can actually increase the complainant’s sentiment and it is not costly at all. This effect is also positively affected by overall satisfaction, so saying thank you and making customers of the relatively better companies feel that the company is even more sympathetic to their problems. Hence, saying “thank you” is always going to be effective at mitigating crisis and help ease the customers negative feelings.

Overall, it is important for managers to keep in mind that a gratitude strategy, it is generally the way to go in regards to complaint handling and not an apology strategy. Not only our findings show that apology is not effective but it has been argued by Coombs and Holladay (2008) that it can be very expensive, since it involves accepting responsibility for the incident. Apologies in official statements have been used in the past to win lawsuits in court. So, using another strategy in response to a crisis, even a low level one like customer complains on Twitter, is preferable if it can be as effective. They conclude in their research that sympathy, compensation and apology strategies seem to have very similar effects, especially in the case of low to moderate level crises. Combining this with our finding, companies do not seem to have a reason to use apologies, since they are both ineffective and potentially expensive. The only exception to this appears to be in cases that companies know that they are at fault. In these cases, apology is advised since it can have positive effects and it would also be unethical for the company to deny responsibility in this situation (Coombs and Holladay, 2008).

(33)

32 their issues regarding the service this re-enforces the good image that the customers have in their minds from the company. We are seeing that a service recovery opportunity can indeed be a chance to win the customer back and improve satisfaction and loyalty.

On the other hand, all it is not peaches and roses from companies with an overall higher satisfaction score. They need to be careful, since customers appear to be less forgiving to them, in regards to timelessness and response effectively. This indicates that complainants have the expectation that the customer support will response fast and effective to their problem, more so from the better firms, and they can be more forgiving towards less reputable firms.

So, the satisfaction index that a company has can be compared with a catalyst that strengthens the effectiveness of all service recovery attempts but also penalizing delays and non-effectiveness in responses.

Limitations and future research

This study has of course some limitations. First of all, only the sentiment of the first customer respond was examined in order to obtain sentiments. Future research could also evaluate the effectiveness of the different strategies in other stages of the company-customer interaction and see if they the effects remain consistent as the dialogue unfolds. Secondly, in this study there was not the possibility to observe any compensation or redress strategies. These strategies have been proven effective in previous researches (Davidbow, 2000; Coombs and Holladay, 2008 and more) but this information was not available in the examined dataset. Give compensation for something, like for example, receiving a voucher as compensation for a trip delay, is not something that the company will publicly announce on Twitter or other platforms but rather it will be provided to the complainant through a personal e-mail.

(34)

(35)

34

Table 1:

Customer Support on Twitter dataset

2

Column name explanation

tweet_id The unique ID for this tweet

author_id The unqiue ID for this tweet author (anonymized for non-company users)

inbound Whether or not the tweet was sent (inbound) to a company created_at When the tweet was created

text The text content of the tweet

response_tweet_id The tweet that responded to this one, if any in_response_to_tweet_id The tweet this tweet was in response to, if any

Table 2:

Occurrence of valance shifters

3

Text Negator Amplifier Deamplifier Adversative

Cannon reviews 21% 23% 8% 12% 2012 presidential debate 23% 18% 1% 11% Trump speeches 12% 14% 3% 10% Trump tweets 19% 18% 4% 4% Dylan songs 4% 10% 0% 4% Austen books 21% 18% 6% 11% Hamlet 26% 17% 2% 16%

Table 3:

Words used for applying text information extraction process

Service recovery strategy Handcrafted lexicon search

Apology strategy “sorry”, “apologize”, “apologies”, “excuse”, “afraid”

Expressing Gratitude strategy “thank”, “thanks” Inquiring further information - Willingness

to listen “please”, “?”

Table 4:

Wald Chi-square test

Residual deviance DF Critical value for Chi-square test (p < 0.05)

Model 1 1089.10 * 296933 298201.7

Model 2 470.96 * 169918 170878.0

2_*_{column descriptions as mentioned in the dataset’s page on kaggle.com}

(36)

35

Table 5:

Vif-scores

response_time response_length apology gratitude information 1.000840 1.052192 1.041424 1.015370 1.007270

Table 6:

Model1 - Base GLM Model

Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.448e-02 3.220e-04 76.029 < 2e-16 *** respone_time 2.124e-07 4.534e-08 4.685 2.8e-06 *** response_lenght -3.103e-04 2.386e-05 -13.006 < 2e-16 *** apology1 -6.985e-03 2.996e-04 -23.315 < 2e-16 *** gratitude1 5.491e-03 3.727e-04 14.732 < 2e-16 *** information1 -8.074e-03 2.246e-04 -35.950 < 2e-16 *** Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(37)

36

Table 8:

Model with non-parametric relations

Parametric Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 0.0207570 0.0001785 116.28 < 2e-16 *** apology1 -0.0068881 0.0003006 -22.91 < 2e-16 *** gratitude1 0.0054025 0.0003739 14.45 < 2e-16 *** information1 -0.0080361 0.0002248 -35.75 < 2e-16 *** Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Approximate significance of smooth terms:

edf Ref.df F p-value s(response_time.mc) 8.910 8.997 32.05 <2e-16 *** s(response_length.mc) 6.769 7.461 23.96 <2e-16 *** Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

R-sq.(adj) = 0.00949 Deviance explained = 0.955% GCV = 0.0036647 Scale est. = 0.0036645 n = 296939

Table 9:

Model comparison

AIC Residual Deviance

Model 1 (Base Model) -822598 1089.10

Model 3 (Non-linear effect inclusion) -822668 1088.80

Table 10:

Model comparison - Moderation

AIC Residual Deviance

Model 2 (Base Moderation) -518344 470.96

(38)

37

Table 11:

GLM Model – Inclusion of non-linear effect

Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.080e-02 1.784e-04 116.618 < 2e-16 *** response_time 8.894e-07 9.153e-08 9.717 < 2e-16 *** I(response_time^2) -1.417e-12 1.665e-13 -8.513 < 2e-16 *** response_lenght -3.106e-04 2.386e-05 -13.018 <2e-16 *** apology1 -6.926e-03 2.996e-04 -23.114 < 2e-16 *** gratitude1 5.377e-03 3.729e-04 14.421 < 2e-16 *** information1 -8.088e-03 2.246e-04 -36.016 < 2e-16 *** Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Table 12:

GLM Model with Moderative effect of overall satisfaction level with

non-linear effects

(39)

38

Table 13:

Model with Sentiment diff. between first and second customer tweet as

DV (Robustness check)

Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 1.968e-02 3.465e-04 56.817 < 2e-16 *** respone_time 1.389e-07 4.877e-08 2.847 0.00441 ** response_lenght -1.396e-04 2.567e-05 -5.437 5.44e-08 *** apology1 -2.621e-03 3.223e-04 -8.131 4.28e-16 *** gratitude1 2.092e-03 4.009e-04 5.216 1.82e-07 *** information1 -3.401e-03 2.416e-04 -14.077 < 2e-16 *** Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Table 14:

Model with Sentiment diff. between first and second customer tweet as

DV with non-linear effects (Robustness check)

(40)

39

Table 15:

Supported hypotheses

Hypotheses Supported/Not

supported H1: Company response time in a complaint has a negative impact on

customer sentiment. Not supported

H2: The length of the company’s initial response tweet has a positive

impact on customer sentiment. Not supported

H3: Apologizing to the customer has a negative impact on customer

sentiment. Supported

H4: Expressing gratitude has a positive effect on customer sentiment. Supported H5: Inquiring more information has a positive effect on customer

sentiment. Not Supported

H6a: Customer satisfaction has a negative effect on the relation between

customer sentiment and response time. Supported H6b: Customer satisfaction has a negative effect on the relation between

customer sentiment and apology strategy. Not supported H6c: Customer satisfaction has a positive effect on the relation between

customer sentiment and response length. Not supported H6d: Customer satisfaction has a positive effect on the relation between

customer sentiment and expressing gratitude. Supported H6e: Customer satisfaction has a positive effect on the relation between

(41)

40

Figure 1: Conceptual model

(42)

41

(43)

42

References

Abney, A. K., Pelletier, M. J., Ford, T. R. S., & Horky, A. B. (2017). # IHateYourBrand: adaptive service recovery strategies on Twitter. Journal of Services Marketing, 31(3), 281-294.

Agarwal, A., Xie, B., Vovsha, I., Rambow, O., & Passonneau, R. (2011, June). Sentiment analysis of twitter data. In Proceedings of the Workshop on Language in Social Media

(LSM 2011) (pp. 30-38).

Aksoy, L., Cooil, B., Groening, C., Keiningham, T. L., & Yalçın, A. (2008). The long-term stock market valuation of customer satisfaction. Journal of Marketing, 72(4), 105-122. Bradford, J. L., & Garrett, D. E. (1995). The effectiveness of corporate communicative responses to accusations of unethical behavior. Journal of business ethics, 14(11), 875-892. Cha, M., Haddadi, H., Benevenuto, F., & Gummadi, K. P. (2010, May). Measuring user influence in twitter: The million follower fallacy. In fourth international AAAI conference

on weblogs and social media.

Chen, J., Chai, A., Goel, M., Lieu, D., Mohamed, F., Nahm, D., & Wu, B. (2015). Predicting stock prices from news articles. The Undergraduate Statistics

Association-Project Committee journal. APPENDICES.

Conlon, D. E., & Murray, N. M. (1996). Customer perceptions of corporate responses to product complaints: The role of explanations. Academy of management journal, 39(4), 1040-1056.

Coombs, W., & Holladay, S. (2008). Comparing apology to equivalent crisis response strategies: Clarifying apology's role and value in crisis communication. Public Relations

Review, 34(3), 252-257. doi:10.1016/j.pubrev.2008.04.001

Davidow, M. (2000). The bottom line impact of organizational responses to customer complaints. Journal of Hospitality and Tourism Research, 24(4), 473-490. doi:10.1177/109634800002400404

Davidow, M. (2003). Organizational responses to customer complaints: What works and what doesn’t. Journal of service research, 5(3), 225-250.

Dean, D. H. (2004). Consumer reaction to negative publicity: Effects of corporate reputation, response, and responsibility for a crisis event. The Journal of Business

(44)

43 Deighton, J., & Kornfeld, L. (2009). Interactivity's unanticipated consequences for marketers and marketing. Journal of Interactive marketing, 23(1), 4-10.

Einwiller, S. A., & Steilen, S. (2015). Handling complaints on social network sites–An analysis of complaints and complaint responses on Facebook and Twitter pages of large US companies. Public Relations Review, 41(2), 195-204.

Feine, J., Morana, S., & Gnewuch, U. (2019). Measuring Service Encounter Satisfaction with Customer Service Chatbots using Sentiment Analysis.

Geetha, M., Singha, P., & Sinha, S. (2017). Relationship between customer sentiment and online customer ratings for hotels-An empirical analysis. Tourism Management, 61, 43-54. Geyskens, I., Steenkamp, J. B. E., & Kumar, N. (1999). A meta-analysis of satisfaction in marketing channel relationships. Journal of marketing Research, 36(2), 223-238.

Goodman, J. A., Malech, A. R., & Boyd, S. (1987). Danger! Angry customer!. American

Bankers Association. ABA Banking Journal, 79(1), 63.

Gronroos, C. (1988). Service quality: The six criteria of good perceived service. Review of

business, 9(3), 10.

Hanssens, D. M., Rust, R. T., & Srivastava, R. K. (2009). Marketing strategy and Wall Street: nailing down marketing's impact. Journal of Marketing, 73(6), 115-118.

Hart, C. W., Heskett, J. L., & Sasser, J. W. (1990). The profitable art of service recovery. Harvard business review, 68(4), 148-156.

Hennig-Thurau, T., Malthouse, E. C., Friege, C., Gensler, S., Lobschat, L., Rangaswamy, A., & Skiera, B. (2010). The impact of new media on customer relationships. Journal of

service research, 13(3), 311-330.

Holloway, B. B., & Beatty, S. E. (2003). Service failure in online retailing: A recovery opportunity. Journal of service research, 6(1), 92-105.

Ikoro, V., Sharmina, M., Malik, K., & Batista-Navarro, R. (2018, October). Analyzing sentiments expressed on twitter by uk energy company consumers. In 2018 Fifth

International Conference on Social Networks Analysis, Management and Security (SNAMS) (pp. 95-98). IEEE.

(45)

44 Kau, A., & Loh, E. (2006). The effects of service recovery on consumer satisfaction: A comparison between complainants and non-complainants. Journal of Services

Marketing, 20(2), 101-111.

Kelley, S. W., & Davis, M. A. (1994). Antecedents to customer expectations for service recovery. Journal of the Academy of Marketing Science, 22(1), 52-61.

Kim, S. M., & Hovy, E. (2004, August). Determining the sentiment of opinions. In Proceedings of the 20th international conference on Computational Linguistics (p. 1367). Association for Computational Linguistics.

Kumar, A., Irsoy, O., Ondruska, P., Iyyer, M., Bradbury, J., Gulrajani, I., ... & Socher, R. (2016, June). Ask me anything: Dynamic memory networks for natural language processing. In International conference on machine learning (pp. 1378-1387).

Kwon, S., Park, E., & Kim, K. (2014). What drives successful social networking services? a comparative analysis of user acceptance of facebook and twitter. The Social Science

Journal, 51(4), 534-544.

Leeflang, P., Wieringa, J. E., Bijmolt, T. H., & Pauwels, K. H. (2016). Modeling markets. Springer-Verlag New York.

Maritz Research. (2011). Maritz research and evolve24 – Twitter study. Retrieved from http://www.maritzresearch.com/∼/media/Files/MaritzResearch/e24/ExecutiveSummaryT witterPoll.pdf

McCollough, M. A., Berry, L. L., & Yadav, M. S. (2000). An empirical investigation of customer satisfaction after service failure and recovery. Journal of service research, 3(2), 121-137.

Naldi, M. (2019). A review of sentiment computation methods with R packages. arXiv

preprint arXiv:1901.08319.

Negash, S., Ryan, T., & Igbaria, M. (2003). Quality and effectiveness in web-based customer support systems. Information & Management, 40(8), 757-768. doi:10.1016/S0378-7206(02)00101-5

Peng, C. L., Lai, K. L., Chen, M. L., & Wei, A. P. (2015). Investor sentiment, customer satisfaction and stock returns. European Journal of Marketing, 49(5/6), 827-850.

(46)

45 Saif, H., He, Y., & Alani, H. (2012, November). Semantic sentiment analysis of twitter. In International semantic web conference (pp. 508-524). Springer, Berlin, Heidelberg. Spreng, R. A., Harrell, G. D., & Mackoy, R. D. (1995). Service recovery: impact on satisfaction and intentions. Journal of Services marketing, 9(1), 15-23.

Tang, C., & Guo, L. (2015). Digging for gold with a simple tool: Validating text mining in studying electronic word-of-mouth (eWOM) communication. Marketing Letters, 26(1), 67-80.

Tumasjan, A., Sprenger, T. O., Sandner, P. G., & Welpe, I. M. (2010, May). Predicting elections with twitter: What 140 characters reveal about political sentiment. In Fourth

international AAAI conference on weblogs and social media.

Weissman, G. E., Ungar, L. H., Harhay, M. O., Courtright, K. R., & Halpern, S. D. (2019). Construct validity of six sentiment analysis methods in the text of encounter notes of patients with critical illness. Journal of biomedical informatics, 89, 114-121.

Wirtz, J., & Mattila, A. S. (2003). The effects of consumer expertise on evoked set size and service loyalty. Journal of Services Marketing, 17(7), 649-665.

Xia, L. (2013). Effects of companies' responses to consumer criticism in social media. International Journal of Electronic Commerce, 17(4), 73-100.

Zerfaß, A., Tench, R., Verčič, D., Verhoeven, P., & Moreno, A. (2014). European

communication monitor 2014: excellence in strategic communication-key issues, leadership, gender and mobile media: results of a survey in 42 countries.

BrusselsEACD/EUPRERA, Helios Media.

Web-sources:

ACSI website: https://www.theacsi.org/ Kaggle web-site: https://www.kaggle.com/

(47)