Recalculating the Net Promotor Score - Tackling the Cultural Bias in the Likelihood of

(1)

MSc Marketing Intelligence Thesis

Recalculating the Net Promotor Score - Tackling the Cultural Bias in the Likelihood of

Recommendation Question

University of Groningen Faculty of Economics and Business

Department of Marketing PO Box 800, 9700 AV Groningen (NL)

Maylen Romee de Koning S2541742

Prinsesseweg 54 9717 BK Groningen

+31623651354

maylendekoning@gmail.com

First supervisor: dr. F. Eggers F.Eggers@rug.nl

Second supervisor: dr. J.T. Bouma J.T.Bouma@rug.nl

(2)

2

Abstract

In recent years, the Net Promoter Score (NPS) has been acknowledged as a widely used measure for customer satisfaction and loyalty. However, questions have arisen concerning the subjectivity of service evaluations and international service providers are confronted with divergent NPS scores across countries. This thesis focuses on explaining this phenomenon by assessing cultural differences in scale usage, which is reflected in a country’s average response style. The levels of extreme and midpoint response style are expected to have a mediating effect in the relationship between national culture, distinguished by Hofstede’s cultural values, and the NPS. A data set generated from a customer satisfaction survey of a large international service provider over the year 2018, consisting of 361,685 respondents with 82 countries of nationality, is used to estimate multiple linear regressions and a Statistical Mediation Analysis. The results of the two methods indicate that positive extreme response style significantly mediates the relationship between cultural values and NPS. The variance in this response style is found to be significantly caused by a country’s levels of individualism and uncertainty avoidance. Overall, the analyses prove that the cultural bias in NPS is not easily captured by the cultural values but rather is a product of countries’ response style throughout a customer survey. Based on this finding, correction weights for the 82 countries are derived, which can be directly applied as a practical guideline to reduce the cultural bias in NPS and make the metric more comparable across countries, thereby improving internal service providers’

understanding of customer satisfaction.

Keywords: Net Promoter Score, extreme response style, midpoint response style, Hofstede, cultural values, customer satisfaction, service quality

(3)

3

Introduction (Chapter 1)

A solid understanding of the cross-cultural differences of customers and how customers with particular cultural backgrounds evaluate a provided service is crucial for international service companies to better understand firm performance. Many companies handle customer surveys to collect feedback data, typically including questions on perceived quality and satisfaction regarding specific aspects of the provided service (Hayes, 1997). Feedback gathered is then used to improve their service and ultimately customer satisfaction and loyalty (Voss et al., 2004). Numerous studies have been conducted on different customer feedback metrics, such as overall satisfaction (De Haan, Verhoef, & Wiesel, 2015) and the SERVQUAL instrument (Parasuraman, Zeithaml, & Berry, 1988).

In 2003, Fred Reichheld introduced a metric that is now widely used by international companies: the Net Promoter Score (NPS). The NPS can be calculated by asking customers how likely they are to recommend the company to friends or colleagues on a scale from 0-10. Despite criticism (Keiningham et al., 2007; Kumar & Shah, 2004; Morgan & Rego, 2006; Schneider et al., 2008), this metric has proven to be correlated to firm growth (Mercedy, Feetham, & Wright, 2016; Van Doorn, Leeflang, & Tijs, 2013) and customer retention (De Haan, Verhoef, & Wiesel, 2015). This likelihood of recommendation question is often included in customer surveys to summarize both customer satisfaction (Krol et al., 2015) and loyalty (Reichheld, 2003). Furthermore, it is used as a metric in company industry rankings (Garrow & Ferguson, 2008) and employee performance evaluation (Peterson & Wilson, 1992).

Yet, international service providers are confronted with divergent NPS scores across countries.

Country managers are required to reach a certain NPS goal and are being discredited if this goal cannot be attained. In recent years, questions have arisen concerning the subjectivity of service evaluations. The scores on a measurement instrument should be structurally equivalent and have no construct bias for them to be comparable across cultures. However, numerically equal scores on customer surveys might reflect higher or lower levels of intensity in one culture than another (Van Herk, Poortinga, & Verhallen, 2004). Laroche et al. (2004) confirmed that most service quality and satisfaction measures are nonequivalent across cultures. This thesis focuses on explaining this phenomenon by assessing cultural differences in scale usage, which can be captured in a respondent’s response style (Kemmelmeier, 2016). Even though the growing homogeneity of customers all over the world is addressed in several marketing studies (e.g. Carpenter et al., 2013;

Cleveland & Laroche, 2007; Merz, He, & Alden, 2008; Özsomer & Altaras, 2008), the response style

(6)

6

differences between countries, which were already studied in 1967 by Zax & Takahashi, are very stable across time (Harzing, 2006). The most pervasive and frequently studied types are extreme response style and its counterpart, midpoint response style (Baumgartner & Steenkamp, 2001; De Jong et al., 2008; Greenleaf, 1992; Paulhus, 1991). While previous research encompassed only a limited number of countries and few multi-country studies are conducted (e.g. Baumgartner &

Steenkamp, 2001; Smith, 2004), Johnson et al. (2005) have studied extreme and midpoint response styles across 19 nations on five continents and linked four of Hofstede’s cultural dimensions to them as distinctive measures of national culture (Hofstede, 1984). They concluded that respondents from cultures with high masculinity and high power distance are more likely to make use of extreme responses. Harzing (2006) corroborated these results and additionally found individualism and uncertainty avoidance to be related to an increased usage of extreme response styles. Her research covered 26 countries.

The extent to which respondents employ extreme and midpoint response styles are expected to have a substantial impact on Net Promoter Score, due to the way this metric is calculated. This can be explained by a concrete example of an international service provider. Imagine a specific service is delivered to customers characterized by extreme response styles who eventually give a score of 10 on the likelihood of recommendation question because they are very satisfied with the service, making them ‘promoters’. This will result in a Net Promoter Score (percentage promoters minus percentage detractors) of 100. Now, if the same service with the exact same quality level is delivered to customers having a midpoint response style, their ratings will likely be close to the midpoint of the scale. Let us assume they all rate the likelihood of recommendation question with a 7 because they are very satisfied, but are not prone to giving extreme ratings. This will lead to a Net Promoter Score of 0 since those scoring a 7 will be labeled as ‘neutral’. The NPS scores of the two groups are far apart even though both groups are very satisfied which is a result of divergent response styles.

The current research is aiming to elaborate on the existing literature regarding the effect of Hofstede’s cultural dimensions on these response styles with an even more extensive set of countries. Also, the effect of cultural differences on Net Promoter Score has not yet been studied, which is surprising because NPS is a widely used measure for international service providers to evaluate firm performance. Nonetheless, there is a gap in the literature on how extreme and midpoint response styles influence Net Promoter Score, a metric that is not reflecting a mean value and due to its calculation is expected to be more vulnerable to response styles. The NPS might simply reflect differences in the way people respond to surveys rather than picking up the true service performance across countries.

(7)

7

The first research question is formulated as follows:

1. Is there a relationship between national culture and Net Promoter Score, which is mediated by response style?

To answer this question, the differences between national cultures are reflected in positive extreme, midpoint, and negative extreme response styles. These response styles have an effect on a customer’s score on the likelihood of recommendation question. Would a manager not account for cultural differences, and then the Net Promoter Score is likely to reflect the composition of respondents. Practically, if country 1 has a high level of positive extreme response style, one can expect high scores on the likelihood of recommendation question. On the contrary, would country 2 have the tendency to use the midpoint of a response scale, it can be assumed that this would lead to a lower NPS. Consequently, the question arises how the two scores can be compared. This research provides a practical solution for a reduction of the cultural bias by a recalculation of the Net Promoter Score.

The second research question is therefore formulated as follows:

2. How can the Net Promoter Score be adjusted to account for cultural differences?

The remainder of this paper is structured as follows: in chapter 2, an overview of the theoretical background and previous studies regarding Hofstede’s cultural dimensions, response styles, and NPS is provided, as well as the relationship between these three concepts. In chapter 3, the methodology of the study is explained, followed by the results in chapter 4 including the recalculation of NPS. The results are elaborated on and a conclusion is formed in chapter 5. This chapter also provides managerial implications and limitations of the current study.

(8)

8

Theory (Chapter 2)

2.1 Response styles

Response style is defined as the tendency to respond systematically to questionnaire items in a manner other than what the items were specifically designed to measure (Paulhus, 1991). This implies that respondents have the urge to answer questions in a particular way, regardless of the content (Baumgartner & Steenkamp, 2001).

Such habitual behavior can result in seriously biased conclusions in cross-cultural marketing research (Clarke, 2000; Hui & Triandis, 1989). This confirms Bearden and Netemeyer’s (1999) finding that many commonly used scales used in marketing fail to control adequately for response style and thus threaten the validity of conclusions from marketing. Not to mention, these seem most problematic in attitude and survey research (Bentler, Jackson, & Messick, 1971; Schuman & Presser, 1996).

The two most widely studied types of response styles in survey scales are extreme response style and midpoint response style (De Jong et al., 2008). Extreme response style refers to the tendency to use the endpoints of a rating scale (Crandall, 1982; Greenleaf, 1992; Hamilton, 1968; Paulhus, 1991).

This way of responding can be distinguished into positive and negative extreme response style, depending on the linked question to be answered and corresponding use of the scale. Midpoint response style is considered as the counterpart, referring to the tendency to avoid the highest and lowest response categories (Hurley, 1998) and overuse the midpoint of a rating scale (Baumgartner

& Steenkamp, 2001). Because of their complementary nature and strong theoretical foundation, these are the main response styles examined in this research.

2.2 Hofstede’s cultural dimensions

Culture has gained increased attention in marketing research as it is an influential factor that shapes individual values and affects perceptions and behavior (Kassim & Asiah Abdullah, 2010; Zeithaml, Bitner, & Gremler, 2002). However, it is found difficult to research the influence of culture on international customer behavior and marketing due to the complexity of the concept (Soares, Farhangmerh, & Shoham, 2007). In this paper, culture is defined as the collective programming of the mind which distinguishes the members of one group or category of people from those of another, following Hofstede and Bond (1988).

(9)

9

Hofstede (1984, 1991, 2001) has devoted multiple pieces of research to defining culture and how cultures can be distinguished from one another. He developed a framework that embodies four cultural values along which countries can be categorized based on factor analysis: individualism, uncertainty avoidance, power distance, and masculinity. A fifth dimension, long-term orientation, was added in 1991 and in 2008 a sixth dimension, indulgence, was established by Michael Minkov (Hofstede & Minkov, 2010). Hofstede’s method provides a way to observe fact-based differences that exist between nations and demonstrates that due to similarities among people within a nation, culture is a national character (Kassim & Asiah Abdullah, 2010). These similarities can be found in history, language, politics, laws, and education, among others. There is support for between-country differences (Hofstede, 1984; Steenkamp, 2001), although Hofstede does not claim that there is complete within-country homogeneity. However, there is still a meaningful degree of within-country commonality to which researchers argue that culture can be conceptualized at a national level (Sivakumar & Nakata, 2001; Steenkamp, 2001).

Despite criticism concerning the generalization of the research setting (Keiningham et al., 2007;

Kumar & Shah, 2004; Morgan & Rego, 2006; Schneider et al., 2008), Hofstede’s tool has gained popularity over the years because of its clarity, parsimony, and resonance with managers. It is perceived useful in formulating hypotheses and is by now the most commonly used cultural framework in a wide range of empirical cross-cultural studies (Kirkman, Lowe, & Gibson, 2006;

Sivakumar & Nakata, 2001; Soares, Farhangmehr, & Shoham, 2007; Steenkamp, 2001). In their literature research, Soares, Farhangmehr, and Shoham (2007) have confirmed that the cultural dimensions of Hofstede are relevant for cross-national research in international marketing and customer behavior, and therefore is used to distinguish cultures on a national level in this study.

2.3 The effect of culture on response styles

Differences in response patterns have been found between a variety of countries, confirming the relationship between culture and the response bias (Kemmelmeier, 2016). To assess the effect of culture on extreme response style and midpoint response style, four of Hofstede’s cultural dimensions are examined: individualism, uncertainty avoidance, power distance, and masculinity.

These four cultural dimensions have previously been proven to explain 59% of the between-country variance in extreme response style (De Jong et al., 2008) and will be explained one by one.

(10)

10

2.3.1 Individualism

This dimension describes the social behaviors and types of relationships within a group. In individualistic societies, the ties between individuals are loose and every person is expected to only look after him- or herself and immediate family members (Hofstede, 1991). Personal freedom is highly valued and the focus lies on individual rights, characteristics, self-reliance, and creativity, among others. As people from such cultures are less concerned with the consequences of expressing strong opinions, combined with the habit to achieve clarity in individual explicit verbal statements (Chen, Lee, & Stevenson, 1995), a higher tendency towards extreme response style is expected (Chen, Lee, & Stevenson, 1995; De Jong et al, 2008; Harzing et al., 2012; Johnson et al., 2005; Peterson, Rhi-Perez, & Albaum, 2014; Smith & Fischer, 2008). On the contrary, citizens of a collectivist society are integrated into a strong, cohesive group in which all members look out for one another (Hofstede, 1991). Decisions are approached from a ‘we’ standpoint and the emphasized values include obedience, duty, cooperation, and sacrifice for the group. With the diminished emphasis on individual opinions, collectivist cultures are argued to have a higher level of midpoint response style (Harzing, 2006).

H1a. The higher a country’s level of individualism, the more associated with positive extreme response style.

H1b. The lower a country’s level of individualism, the more associated with midpoint response style.

H1c. The higher a country’s level of individualism, the more associated with negative extreme response style.

2.3.2 Uncertainty Avoidance

This dimension concerns to what extent a social group minimizes uncertainty to cope with anxiety.

The higher in uncertainty avoidance, the greater people feel threatened by uncertainty and ambiguity and take actions to stay away from such situations (Hofstede, 1991). A reflection of this intolerance of ambiguity and the tendency to adopt rigid attitudes and rules (Hamilton, 1968) are likely to be reflected in a more extreme response style as the endpoints of a scale can be interpreted as more definitive and will help achieve a greater degree of structure (De Jong et al., 2008; Harzing, 2006; He et al., 2014; Johnson et al., 2005; Van Dijk et al., 2009). In contrast, citizens from a lower uncertainty avoidant culture are less susceptible to uncertainty and thus are in a lower need for definitive decisions, resulting in an increased tendency towards midpoint response styles (He et al., 2014).

(11)

11

H2a. The higher a country’s level of uncertainty avoidance, the more associated with positive extreme response style.

H2b. The lower a country’s level of uncertainty avoidance, the more associated with midpoint response style.

H2c. The higher a country’s level of uncertainty avoidance, the more associated with negative extreme response style.

2.3.3 Power Distance

Measures in power distance are characterized by power inequality and authority relations, or simply put the extent to which unequally distributed power is accepted. The higher in power distance a society is, the better the acceptance of the differences between more powerful and less powerful people in terms of social class, education level, and occupation (Hofstede, 1991). Also, such a society fosters greater decisiveness and definitiveness in communications, which will promote extreme response style behavior (He et al., 2014; Peterson, Rhi-Perez, & Albaum, 2014). Countries with a lower degree of power distance are more likely to emphasize modesty as a value (Nelson & Shavitt, 2002), which can be associated with an increase in midpoint response style (He et al., 2014).

However, multiple studies resulted in inconsistent findings concerning the effect of power distance (e.g. Johnson et al., 2005; Smith & Fischer, 2008; Van Dijk et al., 2009). As a result, an increase in power distance was argued to have a null effect as decisiveness and definiteness in communications is demanded by superiors, while subordinates should respond modestly (De Jong et al., 2008;

Harzing, 2006). Nonetheless, based on previous argumentation the following hypotheses are established.

H3a. The higher a country’s level of power distance, the more associated with positive extreme response style.

H3b. The lower a country’s level of power distance, the more associated with midpoint response style.

H3c. The higher a country’s level of power distance, the more associated with negative extreme response style.

2.3.4 Masculinity

The fourth dimension indicates the extent to which sex roles are dominant within society and how these roles are socially distinct. The higher the level of masculinity, the greater the expectancy to be

(12)

12

tough and focused on material success (Hofstede, 1991). It is therefore argued that the behaviors evoked in more masculine cultures, such as being more assertive, decisive, and daring, will be reflected in the tendency to select the strongest available choices for representing their opinions, i.e.

handling an extreme response style (De Jong et al., 2008; Johnson et al., 2005; Van Dijk et al., 2009).

More feminine cultures appreciate relationships and altruism to a higher extent. Their increased value for modesty and harmony can be shown in a midpoint response style.

H4a. The higher a country’s level of masculinity, the more associated with positive extreme response style.

H4b. The higher a country’s level of femininity, the more associated with midpoint response style.

H4c. The higher a country’s level of masculinity, the more associated with negative extreme response style.

2.4 The Net Promoter Score

Companies typically make use of customer surveys for gathering customer feedback data in order to measure and improve customer satisfaction and loyalty, as well as learn about their customers (Schneider et al., 2008). The overall satisfaction score, number of complaints, and the Top 2 customer satisfaction score are common measures for customer satisfaction and have proven to be related to business performance (Morgan & Rego, 2006). Examples of loyalty measures can be divided into loyalty intentions, such as repurchase likelihood and likelihood to recommend, and actual behavior, such as making recommendations and customer retention (Morgan & Rego, 2006;

Rust, Zahorik, & Keiningham, 1995). However, according to Reichheld (2003), retention is a poor measure for customer loyalty and, together with overall customer satisfaction, is not truly connected to actual customer behavior and increasing sales. “The one metric you need to grow” according to Reichheld (2003) is Net Promoter Score (NPS). This popular metric is used by companies to gauge customer satisfaction and loyalty. The calculation of Net Promoter Score is promoted by asking your customers “How likely is it that you would recommend our company to a friend or colleague?” and thus is a measure of recommendation likelihood. From the answers to this one question, the ratio of promoters to detractors can be derived. “Promoters” are the group of customers who answered this question with a 9 or 10, based on a 0 to 10 rating scale, and “detractors” are those who responded with a 0-6 rating. Customers who rated in between these two groups, thus with a 7-8 rating, are passively satisfied and therefore called “neutrals”. When you subtract the percentage of detractors from the percentage promoters, you will get the net-promoter score of your company. This

(13)

13

percentage can be positive, zero, or negative. A graphical visualization of the Net Promoter Score is shown in Figure 1.

Figure 1. Net promoter score (Marielle, 2018)

Many companies, e.g. service providers and financial institutions, use this index as a central metric because it is known for its simplicity and its ease of measurement (Gupta & Zeithaml, 2006). This one question also helps to understand one’s position compared to other regions, segments, or competitors (Reichheld, 2003). Reichheld (2003) argues that Net Promoter Score is strongly correlated with a company’s growth rate and loyalty, with notable results in the company’s industry.

This strong effect is due to a customer’s willingness to promote a company, which results in customers putting their credibility and reputations on the line. This is a risk one would only take in case of strong loyalty (Reichheld, 2003). Moreover, recommendation intention is correlated with actual recommending behavior, word-of-mouth (Keiningham et al., 2007; Raassens & Haans, 2017), and customer retention (De Haan, Verhoef, & Wiesel, 2015). This statement was evaluated and criticized by Keiningham et al. (2007), Kumar and Shah (2004), Morgan and Rego (2006), and Schneider et al. (2008), whose studies mainly questioned the data used for deriving these conclusions. However, research has indicated that NPS is still a reliable metric concerning correlations with company growth rates (Mercedy, Feetham, & Wright, 2016; Van Doorn, Leeflang,

& Tijs, 2013).

(14)

14

2.5 The effects of response styles on Net Promoter Score

One should account for the different response styles when measuring NPS to prevent two major problems: the threatening of scale validity and reliability and, more importantly for company performance measurement practices, the comparison of customer groups will be jeopardized (Smith

& Reynolds, 2002). Extreme and midpoint response styles have a biasing effect on both the mean level of responses and the correlation between marketing constructs by either inflating or deflating respondents’ scores on measurement instruments (Baumgartner & Steenkamp 2001; Greenleaf 1992).

These two types of response styles, therefore, are expected to generate an even greater bias on NPS than on mean scores since the percentage detractors is subtracted from the percentage promoters.

When customers have a tendency to make use of the endpoints of a scale, they are more likely to be either promoters or detractors, depending on the valence of their opinion. The use of the midpoints of the scale is argued to be a reflection of a neutral opinion.

H5: Positive extreme response style is positively related to being a promoter.

H6: Midpoint response style is positively related to being neutral.

H7: Negative extreme response style is positively related to being a detractor.

The conceptual model of this study is presented in Figure 2. The colors represent the hypotheses specified in the previous sections.

Figure 2. Conceptual Model

(15)

15

Design (Chapter 3)

3.1 Data description

For this thesis, an international service provider has provided a large set of customer data generated from surveys. Approximately 75,000 respondents from all over the world digitally fill in these subjective questionnaires each month. The survey is continuously generating a large and growing database of customer information linked to the respondents’ service characteristics and their subjective service experience. The data is used internally to measure and improve customer satisfaction, and externally for benchmarking purposes, with NPS as the most important metric.

After selecting the preferred language for the survey, the first question provided is the likelihood of recommendation question: “First of all, based on your opinion and experience: How likely would you be to recommend (‘Company name’)?” Respondents can answer this question on a scale ranging from 10 (definitely would recommend) to 0 (definitely would not recommend). Then, after answering demographic and service-specific questions, the majority of the questionnaire contains satisfaction ratings on specific subjects with a five-point Likert scale ranging from “poor”, “fair”,

“good”, “very good”, to “excellent” with the possibility to elaborate on extremely high or low ratings.

For the purpose of relationship testing with the assumption that service level is comparable throughout the complete dataset, the models are estimated with data from this company’s customers only. The year 2018 is under investigation with customer survey data ranging from January 1^st up to December 31^st. The dataset contains 361,685 respondents in total after excluding respondents from countries that have less than 50 respondents, for statistical power purposes, and from countries for which the cultural values by Hofstede are not known. The final dataset contains 82 different nationalities and gender distribution of 57% male and 43% female.

3.1.1 Weighting Factor

In order to increase the reliability and validity of the measures generated from the customer surveys, it was decided to weight the data as a result of known biases within the company. The weighting factor can be calculated by the following formula:

𝐹 = ^!_!^!

! / ^!_!^!

! (Eq. 1)

where

(16)

16

𝐹 = weighting factor

𝑃_! = number of respondents in the specific layer of the dated service process 𝑃_! = number of customers in the specific layer of the dated service process 𝑆_! = number of respondents in the dated service process

𝑆_! = number of customers in the dated service process

After weighing the data of the 361,685 respondents in order to enable statistical significant tests, the dataset was aggregated on country-level. This procedure was performed by averaging the variables of interest, which will now be discussed.

3.1.2 Dependent variable: Net Promoter Score

The purpose of this study is to explain variations in NPS across countries, making this metric the dependent variable in the model. The Net Promoter Score is derived from the likelihood of recommendation question in the customer survey: “How likely would you be to recommend (‘Company name)?” The NPS variable on individual level contains the values -100 for detractors (rating 0-6), 0 for neutrals (rating 7-8) and 100 for promoters (rating 9-10). By taking the mean of this variable, aggregated on the country of nationality, the Net Promoter Score is given for each country in an easy manner. A country’s NPS can range from -100 as the lowest to 100 as the highest score. One outlier was detected at the low end of the distribution, which is Japan with an NPS of 1.074. For the sake of testing the cause of differences in NPS, this outlier was not removed from the dataset. Also, the Japanese are of high relevance for the company as they form a relatively great share of their customers. The descriptive statistics of the NPS for the company in the year 2018 can be found in Table 1. Venezuela has the highest average NPS throughout 2018, with a score of 71.12.

Figure 3 illustrates the uneven distribution of the NPS with peaks at 40-42, 49-50, and 67-68.

Variable Mean SD Median Min Max

Net Promoter Score 42.03 70.92 100 1.074 Japan 71.116 Venezuela

Table 1. Descriptives of the Net Promoter Score

(17)

17

Figure 3. Distribution of the Net Promoter Score

3.1.4 Mediators: Response styles

In order to measure the response style of the individual respondents, a set of questions needed to be selected from the survey. These questions should be least correlated with each other and with the dependent variable in order to measure response style properly (Baumgartner & Steenkamp, 2001). A pre-selection was made of the 12 overall touchpoint ratings, as the sub-questions related to these touchpoints are highly influencing the overall ratings. After checking their correlations, 9 questions with the lowest correlation with each other and with the dependent variable were chosen.

In Appendix A, the correlation matrix can be found, showing correlations below .6. An overview of the question variables and their descriptions can be found in Table 2. q69_71_1 is an average of q69 and q71 combined, which are both ratings on the same topic.

Variable Question asked Mean SD Median

q38_1 --- 3.35 1.22 3 q38_2 --- 3.73 1.08 4 q42 --- 3.44 1.07 4 q45_1 --- 3.84 1.17 4 q62_1 --- 3.5 1.05 4 q65_1 --- 4.02 0.97 4 q69_71_1 --- 3.18 1.14 3 q72_1 --- 3.32 1.07 3 q77_1 --- 3.55 1.06 4 Table 2. Descriptives of the survey questions

Answers to the 9 questions are given on a five-point Likert scale (either “most definitely”,

“definitely”, “somewhat”, “not”, “not at all”; or “excellent”, “very good”, “good”, “fair”, “poor”) and the option “Don’t know / No opinion” is included. The answers are coded as a score from 1

(18)

18

(“excellent” / “most definitely”) to 5 (“poor” / “not at all”) and 6 as no opinion option. For measuring positive extreme response style (positive ERS), the answers scales are recoded as 1 0 0 0 0 0 and summed for the 9 questions for each respondent. Negative extreme response style (negative ERS) is measured in a similar way with recoded scores 0 0 0 0 1 0. This simple calculation method for extreme response style is proven to be statistically best compared to other methods and therefore used in this study (Peterson et al., 2014). Midpoint response style (midpoint RS) was calculated in a similar way. To measure this variable, the customer touchpoint questions were recoded as 0 0 1 0 0 0. In this manner, every respondent will have a score ranging from 0 to 9 on each of the three response styles indicating the extent to which this person applies a specific style of answering satisfaction-related questions. The individual values were averaged across countries to generate country-level scores on the three response styles. Pairwise deletion excluded the missing ratings for each question independently, without excluding the complete case and losing valuable insights. This resulted in two outliers on the high end of the distribution for negative ERS: United Arab Emirates (.877) and Turkey (.843). A possible explanation of these outliers could be high service expectations, as these countries typically pride themselves in high service quality scores (---).

Service expectations, however, are not measured in this study and thus cannot be statistically proven to influence response styles. Therefore, it was decided not to remove these outliers. The descriptive statistics of the three response styles can be found in Table 3. Figure 4 illustrates the distribution of response styles. Positive extreme response style is rather evenly distributed on the middle right side of the scale, while negative extreme response style is clustered on the left with a peak at 0.25 and 0.45. This indicates that, overall, customer of the company are rating the touchpoints more positive than negative. Midpoint response style is clustered on the middle-right of the scale.

Variable Mean SD Median Min Max

Positive extreme response style 2.23 0.59 2.35 Hong Kong 1.002 Venezuela 3.517 Midpoint response style 2.08 0.37 2.03 Serbia 1.392 The Netherlands 3.141 Negative extreme response style 0.43 0.17 0.40 Vietnam 0.099 United Arab Emirates 0.877 Table 3. Descriptives of the response styles

(19)

19

Figure 4. Distributions of the response styles

3.1.3 Independent variables: Cultural values

Geert Hofstede has shared the cultural values generated from his research (Hofstede, Hofstede, &

Minkov, 2010). The country scores on individualism, power distance, uncertainty avoidance and masculinity are available for 104 regions and were matched to the 82 countries in the dataset. The value of this variable has a range from 0 to 100, indicating the level of the corresponding cultural value. An overview of the descriptives of the cultural values is presented in Table 4. Figure 5 shows the distribution of the cultural values, which all seem to vary along the scale. Individualism has a peak at a value of 25. For uncertainty avoidance, this peak is around 50, and for power distance around 70. Lastly, masculinity shows a peak at values of 35-40. Masculinity contains two outliers on the high end of the distribution: Slovakia with a level of 100 and Japan with a level of 95. Considering the objective of the research, no outliers were removed here.

Variable Mean SD Median Min Max

Individualism 41.68 22.92 35.00 6 Guatemala 91 United States

Uncertainty avoidance 66.55 21.36 67.50 8 Singapore 100 Greece

Power Distance 62.45 21.83 66.50 11 Austria 100 Slovakia

100 Malaysia

Masculinity 46.98 19.36 45.50 5 Sweden 100 Slovakia

Table 4. Descriptives of the cultural values

Figure 5. Distributions of the cultural values

3.2 Methods

The effect of cultural values on the NPS, mediated by response style can be assessed using several methods. In this study, two methods have been applied and compared. The first relationship test is performed by multiple linear regressions, followed by a Statistical Mediation Analysis created by Andrew Hayes (2009).

(20)

20

3.2.1 Multiple linear regressions

The purpose of this study is to test the mediating effect of response style on the relationship between cultural values and NPS. Multiple linear regressions were conducted to estimate the four relationships between the included variables as illustrated in Figure 6. Model 1 measures the effect of positive extreme response style on the four cultural values, Model 2 measures the effect of midpoint response style on the cultural values. Model 3 tests the relationship between negative extreme response style and cultural values. Model 4 measures the relationship between the response styles and NPS, and finally Model 5 tests the relationship between the cultural values and NPS. The advantage of this method is that meeting the linear regression assumption required can be done rather easily by transforming the functional form of the models. This makes it right for testing the hypotheses properly. However, multiple linear regressions do not allow capturing all effects, i.e.

direct, indirect and mediation effects, in one model.

Figure 6. Visualization of the models for multiple linear regressions

The initial models can be formulated as follows:

Model 1:

𝑃𝑜𝑠_! = 𝛼 + 𝛽_!𝐼_!+ 𝛽_!𝑈_!+ 𝛽_!𝑃_!+ 𝛽_!𝑀_!+ 𝜀_! (Eq. 2)

(21)

21

Model 2:

𝑀𝑖𝑑_! = 𝛼 + 𝛽_!𝐼_!+ 𝛽_!𝑈_!+ 𝛽_!𝑃_!+ 𝛽_!𝑀_!+ 𝜀_! (Eq. 3)

Model 3:

𝑁𝑒𝑔_! = 𝛼 + 𝛽_!𝐼_!+ 𝛽_!𝑈_!+ 𝛽_!𝑃_!+ 𝛽_!𝑀_!+ 𝜀_! (Eq. 4)

Model 4:

𝑁𝑃𝑆_! = 𝛼 + 𝛽_!𝑃𝑜𝑠_!+ 𝛽_!𝑀𝑖𝑑_!+ 𝛽_!𝑁𝑒𝑔_! + 𝜀_! (Eq. 5)

Model 5:

𝑁𝑃𝑆_! = 𝛼 + 𝛽_!𝐼_!+ 𝛽_!𝑈_!+ 𝛽_!𝑃_!+ 𝛽_!𝑀_!+ 𝜀_! (Eq. 6)

where

𝑃𝑜𝑠_! = level of positive extreme response style of country c 𝑀𝑖𝑑_! = level of midpoint response style of country c

𝑁𝑒𝑔_! = level of negative extreme response style of country c 𝑁𝑃𝑆_! = Net Promoter Score of country c

𝛼 = constant

𝛽_!, … , 𝛽_! = model parameters

𝐼_! = level of individualism of country c

𝑈_! = level of uncertainty avoidance of country c 𝑃_! = level of power distance of country c 𝑀_! = level of masculinity of country c 𝜀_! = error term

(22)

22

3.2.2 Mediation analysis

An alternative method to measure the relationships between the cultural values and the Net Promoter Score mediated by response style is a Statistical Mediation Analysis (Hayes, 2009). An important advantage of this method is that it accounts for partial mediation by estimating the direct effects of the independent variable on the outcome variable while accounting for the mediation effect, which is not measured by the multiple linear regressions. A limitation of this procedure is that all relationships are assumed to be linear. In case a relationship violates one of the assumptions, it is more difficult to account for reformulated models. In the results (chapter 4) this limitation will be elaborated on. The statistical mediation method is applied by the PROCESS macro in SPSS 25, using bootstrapping to create 10,000 samples.

Figure 7 shows the separate regression analyses that are performed by the method. The a-paths are the coefficients for the cultural values in a model predicting response styles from the cultural values.

The b- and c’-paths are the coefficients in a model predicting NPS from both the response styles and cultural values, respectively. The c’-paths quantify the direct effect of cultural values, whereas the products of a and b quantify the indirect effect of the cultural values on NPS through the response styles.

Figure 7. Visualization of the models for Statistical Mediation Analysis

Results (Chapter 4)

(23)

23

This chapter presents multiple regressions conducted to describe the behavior of the response styles as mediators caused by the four cultural values as independent variables, as well as to describe the mediation effect on the change in Net Promotor Score being the dependent variable. For each of the models, the linear regression assumptions that have to be met before one can perform linear regressions are first described. Based on this, modifications are made where needed leading to improved models. Then, the model fit of the regressions is discussed, followed by the parameter results. Finally, In order to account for the mediation and test for partial mediation, a Statistical Mediation Analysis is conducted additionally. The results of this method are presented together with robustness checks.

4.1 Assumptions & model fit

Before performing the multiple linear regressions, four assumptions should be met for the results to be trustworthy. These assumptions entail linear relationships; no multicollinearity among the predictor variables; homoscedastic residuals, and finally; normally distributed residuals of the models. Once these assumptions are met, the fit of the models is examined.

4.1.1 Linear relationships

The first assumption to be met before performing a linear regression is one of a linear relationship.

Initially, the residuals of the multiple linear regressions were plotted against the predicted values shown in Appendix B. A horizontal LOESS-fitted line in these plots would indicate a linear relationship. Intuitively, no obvious relationships other than linear can be identified. However, by these methods, the relationships were not easily interpretable and therefore the specification of the correct functional form was tested by using the Ramsey RESET-test (Ramsey, 1969). The null hypothesis of this test states that the model is specified by the correct functional form. For Model 1, 2, 3, and 5 the RESET-test was not significant, indicating that the cultural values do not have a different relationship with the dependent variable other than linear, and therefore there is no need to add the powers of the variables to the model (Ramsey, 1969). This test was significant for Model 4, however, indicating that powers should be added to the predictor variables. After doing this, the predictor variables were mean-centered by subtracting the overall mean from the individual data points in order to reduce collinearity among the variables. In a model with mean-centered variables, the constant terms of the model refer to the expectation of the response at the mean of the predictors. The model was modified using trial and error with the goal of increasing model quality examined by the coefficient of determination (R²), the Akaike Information Criterion (AIC) and

(24)

24

Bayesian Information Criterion (BIC). The R²indicates the proportion of total variance in the dependent variable explained by the model (Leeflang et al., 2015) and should be maximized. The AIC (Akaike, 1974) and BIC measure the relative amount of information lost by the model and therefore, the model with the lowest criteria is preferred. BIC accounts for a larger penalty term, which includes the number of parameters in the model. For every model, the p-value of the RESET-test is presented in Table 5, indicating the need to add powers to the model when significant (p < .05). This was not significant anymore once the squared term of negative extreme response style was added to the model.

Model RESET-test p-value R² AIC BIC

Linear (Eq. 5) .017 .7615 575.52 587.55

+ Pos² .021 .7668 575.66 590.10

+ Mid² .019 .7629 577.04 591.48

+ Neg² .696 .7911 566.65 581.09

+ Pos² + Neg² .844 .7954 566.94 583.79

+ Pos² + Neg² + Mid² .867 .8017 566.37 585.62

Table 5. Model evolution results

The model including only the added squared terms of negative extreme response style shows the lowest BIC. Nonetheless, the fully squared model has the highest R²as well as the lowest AIC. Since the BIC penalizes more complex models more heavily than AIC does, this last model with the lowest AIC is chosen. This choice is also motivated by the consistency of the measurement form of the predictor variables, which the last model provides. The model for NPS is now linear in parameters, but not in variables (Leeflang et al., 2015) and is reformulated as follows:

𝑁𝑃𝑆_! = 𝛼 + 𝛽_!𝑃𝑜𝑠_!+ 𝛽_!𝑃𝑜𝑠_!^!+ 𝛽_!𝑀𝑖𝑑_!+ 𝛽_!𝑀𝑖𝑑_!^!+ 𝛽_!𝑁𝑒𝑔_!+ 𝛽_!"𝑁𝑒𝑔_!^!+ 𝜀_! (Eq. 7)

where

𝑁𝑃𝑆_! = Net Promoter Score of country c

𝛼 = constant

𝑃𝑜𝑠_! = mean-centered level of positive extreme response style of country c 𝑀𝑖𝑑_! = mean-centered level of midpoint response style of country c

𝑁𝑒𝑔_! = mean-centered level of negative extreme response style of country c 𝛽_!, … , 𝛽_!" = model parameters

(25)

25

𝜀_! = error term

4.1.2 No multicollinearity

The second assumption to be tested is the absence of multicollinearity, a phenomenon that occurs when there is a high correlation of at least one predictor variable with a combination of the other predictor variables (Leeflang et al., 2015). In this study, the predictor variables are the same for Model 1, 2, and 3. Multicollinearity can be detected in the correlation matrix of the predictor variables (Appendix C) that, for the first three models, shows relative high positive correlations up to -0.68 and up to -0.88 for the predictor variables. For this reason, another method for testing multicollinearity was applied. The variance inflation factor (VIF) regression was performed for which values should be <10, suggested by Hair et al. (2010) as indicative of inconsequential collinearity. The low variance inflation factors (VIFs) presented in Table 6 show that collinearity is not a problem for all five of the models.

Model Predictor variable VIF

1 (Eq. 2), 2 (Eq. 3), 3 (Eq. 4) Individualism 1.97

Uncertainty avoidance 1.07

Power distance 1.97

Masculinity 1.06

4 (Eq. 7) Positive ERS 5.55

Positive ERS² 1.41

Midpoint RS 5.78

Midpoint RS² 1.66

Negative ERS 1.63

Negative ERS² 1.43

5 (Eq. 6) Individualism 1.97

Uncertainty avoidance 1.07

Power distance 1.97

Masculinity 1.06

Table 6. VIF-values for independent variables in the models

4.1.3 Homoscedasticity

The third assumption states that the error terms should have the same variance in all cases, i.e. be homoscedastic (Leeflang et al., 2015). This can be examined by the Goldfeld-Quandt test (Goldfeld &

Quandt, 1965) or the Breusch-Pagan (Breusch & Pagan, 1979) test for homoscedasticity. A significant result (p < .05) from these tests would indicate that the residuals are heteroscedastic and the

Recalculating the Net Promotor Score - Tackling the Cultural Bias in the Likelihood of

MSc Marketing Intelligence Thesis