The Impact of Social Media Influencers Raffi Ahmad and Nagita Slavina on Tourism Visit Intentions across Millennials and Zoomers Using a Hierarchical Likelihood Structural Equation Model

Abstract

Background: In this paper, we examine how social media influencers can influence visit intention, especially in the case of Raffi Ahmad and Nagita Slavina, a top influencer who by 2 September 2021 had reached 21.3 M subscribers on YouTube and 54.9 m followers on Instagram with an engagement rate of 0.42%. The focus of this study is Generation Y or Millennials (born 1981–1996) and Generation Z (born 1997–2012). Design/methodology/approach: Snowball sampling was performed to arrive at a representative group of Millennials. Data analysis was performed using hierarchical likelihood via structural equation modeling. Findings: The study results are helpful for a comprehensive understanding of factors affecting visit intention. Effects of the study results summary, tourists from Generations Y and Z are thriving within the internet of things and the digital age, an era in which information can be accessed via various forms of technology across multiple platforms. Practical implications: We discuss and identify the relative importance of each factor through the use of logistics with variational approximation and structural equation models using hierarchical likelihood. Originality: The technique we use is an integrated and extended version of the structural equation model with hierarchical likelihood estimation and features selection using logistics variational approximation.

1. Introduction

Tourism is an industry that continues to proliferate. At present, tourism activities have become a necessity in modern society. Several community groups have included tourism activities as a mandatory agenda item [1]. For example, in sports tourism [2], social exchange [3], and academic conferences [4]. Tourism encompasses a variety of tourist activities supported by various facilities and services provided by the community, government, business, and local government [5,6]. Indonesia’s cultural diversity is something that cannot be denied, and it can be said to be unique [7], e.g., tourism between Bali, Indonesia and Bangkok, Thailand [8], and the sustainability of coastal tourism between Indonesia, Germany, and Lithuania [9]. Nowadays, Indonesia leads as a halal tourism destination in Asia [10].

Indonesia has a complete and varied cultural portrait reflected in daily life [11]. Indonesian culture includes sustainable livelihoods [12], traditional houses [13], traditional culture [14], traditional ceremonies [15], traditional dances [16], traditional clothing [17,18,19], folk songs [20], regional music and musical instruments [21], drawing and other arts [22], sculpture [23], and the typical foods of each region [24,25,26]. In short, each region has different cultural characteristics. These characteristics make each regional culture unique and attractive. Regional culture is the root of national culture [27,28,29]. Meanwhile, Indonesia, a country with a prominent Muslim community, understands the importance of modern economic development through the halal tourism industry [30,31,32]. The influence of sustainable tourism continues to drive Indonesia to promote the industry to push micro and small enterprises (MSMEs) via financial gains from the tourism industry, developing resilience and ways of achieving high growth rates [33,34,35,36,37].

It is undeniable that the digital era in Indonesia is developing rapidly [38]. This is an era where everyone is becoming more fluent in digital media [39]. The digital age is an epoch in which Indonesian people, especially young people, live and develop each day with the internet [40]. This certainly changes the mindset of business people and marketers in carrying out promotions, one means of which is by using a digital marketing strategy [41]. Changes in social media, along with developments in how people access the internet using gadgets or smartphones that can easily be accessed anytime and anywhere, means that every day people encounter promotions on social media carried out by marketers [42,43,44]. This emerging potential will be even more tremendous when targeting and reaching the generation of Millennials and Gen Xennials through online channels or via the internet [45,46].

The term ‘Millennials’ comes from the American generation experts Neil Howe and William Strauss (1991) [47]. Baby boomers were born between 1946 and 1964. Gen X was born between 1965 and 1979/80. Gen Y or Millennials were born between 1981 and 1994/1996. Gen Z is the newest generation, born between 1997 and 2012/2015. This Millennial generation is exceptionally comfortable with mobile devices and using computers to purchase. They typically have multiple social media accounts. Nevertheless, Generation Y and Generation Z were the early adopters of new technology products and services [48,49].

Compared to baby boomers and Generation X, the most distinctive feature of Millennials is their lifestyle based on digital media. This is because, even before digital media emerged, the current generations had a fixed lifestyle pattern. We must strive to understand the Millennial generation because its characteristics coincide with the direction of change that our society will face in the future, such as the fourth industrial revolution and digitalization. Organizations will fail if they fail to bring out the capabilities of those who can, at the same time, be both members of the organization and consumers.

In terms of technology adoption, Generation X and Millennials’ disposable income expanded along with technology and better direct connection to information; however, this has impacted spending habits differently within each generation [49,50,51]. Although Generation X maintained the same budgetary proportion regarding their purchased items, Millennials have increased the amount spent on concessional personal enjoyment [52]. The above outcome is related to the stage of life, as individuals in Generation X seem to be highly likely to have more family commitments and might be starting to build wealth [53,54,55].

The increasing power of social media influencers (SMIs) is driving increased market trends in the business world in the XXIst century [44,56,57]. Business people and marketers have been using digital marketing strategies for their promotional activities for a long time. Social Media Influencer (SMIs) in Instagram and YouTube are two examples of media marketers use to carry out digital-based promotional activities or what we usually call Social Media Influencers. Social media influencers express their opinions and experiences in a broad and focused range. The growth potential of influencers is deemed to be a reliable determinant of taste or trendsetter in one or several specific groups [58,59]. Influencers are not always synonymous with a large number of followers. Social media influencers are people who influence a particular audience of target or online media that companies can use to sponsor their products with their content or interactions with their audience to increase reach, sales, and engagement through positive interactions [60,61,62,63].

The main component of expertise and attractiveness, including likeability, familiarity, similarity, brand attitude, and brand admiration, are synchronously moderated by purchase intention [64,65]. Likewise, the positive effect of influencer likeability on the number of followers is moderated one after the other by the perception of celebrity status and an attributed influence [66]. Language similarity, interest similarity, and self-disclosure all positively affect followers’ perceptions of quality of life, which is strongly influenced by a perspective of friendship [67]. The direct view of sponsorship release of product attributes is moderated consecutively by increased recognition, demand-made scepticism, and lower perceptions of influencer trustworthiness [67].

Nowadays, a company’s marketers use social media influencers to increase consumer interest and sales. Using social media influencers to do marketing can also boost brand awareness. Even if you look at the facts in the field, social media influencers who have a smaller number of followers are considered more effective in connecting with their followers. Social media influencers with a smaller number of followers are known as micro-influencers. Micro-influencers influence more consumer trust and purchase an interest in a brand than social media influencers with large followers [68,69,70]. Interestingly, micro-influencers tend to be more trustworthy because their engagement with their followers is more intense and interactive. Research says that even though they have fewer followers, micro-influencers are more profitable for brands, especially tourism issues and visit intention [71,72,73].

Structural Equation Modeling (SEM) is a multivariate statistical analysis method that may address the relatively complex relationship structures involving many variables [74,75,76,77]. SEM can define a model that involves multiple relationships by explaining the overall causality relationship between latent variables and taking into account measurement errors in the estimation process [78]. In line with this, the SEM construction consists of two parts. First, the structural part connects the latent variables through a simultaneous equation. The measurement connects the observed variables or indicators with latent variables through the confirmatory factor model [79,80,81]. During last decade, the most widely used SEM approach is frequentist [82], Bayesian [83,84,85], partial least squares [86,87,88], covariance-based [89,90], robust non-linear [91], neural network [92,93,94].

The statistical inference and estimation are based on the Lee and Nelder in 1996 are hierarchical likelihood (h-likelihood) [95]. The significant advantage of using h-likelihood is that we can develop fast computation time algorithms for fitting advanced models, infer unobservables, and predict future events from models including unobservables [96]. Furthermore, we use a model-checking plot for regression analysis and generalized linear models, which makes assumptions throughout all parts of a hierarchical model checkable [97,98,99,100]. In line with this, we extend the SEM study via hierarchical likelihood [101,102,103].

In this paper, we examine how the top SMIs in Indonesia are Raffi Ahmad and Nagita Slavina (@raffinagita1717) in influencing visiting intention towards Generation Y and Generation Z. In Section 2; we discuss the literature review. Section 3 covers the experiment conducted, and then we discuss the critical variables for using the logistics variational approximate hierarchical likelihood structural equation model. The results are discussed in Section 4. In a nutshell, the conclusion is given in Section 5 and the recommendation and practical implications in Section 6.

2. Literature Review

2.1. Intention to Visit

Visitors’ intention to visit a destination has regularly been the midpoint of tourism research. In marketing research, intention refers to a situation when a customer behaviorally intent to purchase a product [104]. Similarly, in tourism, visit intention is conceptualized as the visitors’ behavioral interest, including revisiting the destination, recommending the goal to colleagues and kinfolk, and saying positive things about the destination [105]. For instance, visit intention can be defined as the possibility that the visitors will see the goal in the future [104].

In contemporary tourism marketing, tourist destination depends not only on the conventional marketing strategy but also on the online marketing strategy, e.g., SMI [106]. The influencers affect the consumer decision [107] and unconsciously encourage the potential visitors to visit the destination [24]. This strategy has been proven effective in stimulating customers’ intention to visit or even revisit the goal [108]. As a vital destination variable, potential visitors’ perception of the SMI may impact visitors’ attitudes and appraisal regarding the destination [43].

Literature exhibits that the customers’ intention to visit could be resulted from, for example, brand image and brand identity [109], reputation [110], and social media [106]. For tourism sites, an exact value shown by the destination through advertising enhances the visitors’ positive thoughts about the sites, fashioning positive word of mouth and generating an intention to be on sites [111]. Thus, from digital marketing, social media is an emerging channel in tourism that inspires cognitive response that promotes the customers’ affective response [112]. For instance, an influencers’ social media account could be the source of intention to visit potential visitors. In other words, SMIs will affect the customer visit intention.

Millennials’ and Zoomers’ attachment to social media is very high [113]. According to [114], mobile phone usage (i.e., social media) may trigger users’ cognitive feelings about what they perceive from social media. In the context of SMIs and their followers’ relationships, the act publicized in the social media account will impulsively promote the followers to consume. However, the followers’ emotional attachment to SMIs boosts their behavioral intention. Logically, the more considerable number of followers of an SMI account will be perceived more by people and trigger more peoples’ cognitive feelings [115]. Therefore, it is argued that Millennials’ and Zoomers’ mental feelings could be derived from social media usage that concurrently enhances their intentions to visit.

An issue requiring special attention is when residents prefer visiting abroad rather than vacationing to local tourist attractions. This issue will have a directly impacted effect on economic growth on both national and local levels. One of the alternatives that can be conducted by increasing tourist attraction is travel attractiveness. Travel attractiveness is all things that can bring a feeling of interest, pleasant, positive value to be visited and revisited. The tourist attraction is a tourism product in unity, including not only the natural beauty of the destination but also other attributes such as attractiveness, facilities when travelling, and access to these tourist attractions, following the natural attractions, which include natural landscapes of land, ocean views, beaches, climate, or weather. In addition, building attraction includes historical architecture, modern architecture, and archaeology. Meanwhile, visitor attractions also include industrial heritage sites [116,117].

At the same time, the cultural attractions include theatres, museums, homesteads, customs, religious places, special events such as festivals and historical dramas, pageants, and heritages such as cultural heritage—additionally, social tourism attractions such as the lifestyle of residents in tourist destinations. The elements of the appeal of a tourist destination are visitors’ choices, which encourage visitors to make tourist visits. Revisit intention is the possibility for tourists to repeat activities or revisit a goal. The four impacts can lead to returning to visit: travel motivation, experience, perceived context, and attitude [118,119].

The intention to return has been recognized as a critical topic for tourism destination operations. It is a cognizable situation that always leads to a visitor’s plan to revisit a tourism industry within a specified time frame [120,121,122]. The tourist experience is defined as the experience that consumers receive, either directly or indirectly, regarding the service process, company, facilities, and how a consumer interacts with the company and with other consumers [123]. This, in turn, will create cognitive, emotional, and behavioral responses to consumers and leave consumers or visitors memories of experiences during and after visiting this destination [124,125,126,127,128,129].

2.2. The Power of Social Media Influencers

Internet users in Indonesia in early 2021 reached 202.6 million people. This number increased by 15.5 percent or 27 million people compared to January 2020 [130]. Based on the results of the National Socio-Economic Survey (SUSENAS) and the Village Potential Statistics (PODES), as well as secondary data from related agencies such as the Ministry of Communication and Information Technology, Republic of Indonesia (KEMENKOMINFO), and telecommunications operator companies, it was found that the fast expansion of digital devices is inextricably linked to the high number of internet users in Indonesia, as 89.09 percent of Indonesian households had/possessed at least one cell phone number [131].

Today’s technological developments in communication media greatly facilitate people to develop their potential, especially self-presentation. The increasing number of social media users is due to technological developments. Each social media has its appeal depending on how information providers place content and the target audience they want to obtain. Monetization in the range also affects how SMIs show existence [68,70]. After watching the content, social media influencers can be more convenient than conventional marketing, especially customer behavior. Social media influencers are excellent for reaching out to hard-to-reach communities [69]. Users are watching and ingesting fewer and fewer traditional media, such as TV or print ad campaigns. Furthermore, the fame of the SMI or the brand they promote has been shown to influence an audience [69]. Besides that, an SMI’s popularity and thus attractiveness can influence viewer numbers’ perspective about them as media figures [132].

3. Material and Methods

This section discusses the dataset; the methods used include the variational logistics approximation and the step construction of hierarchical likelihood towards structural equation models. The sampling technique used is snowball or chain referral by considering Generation Y and Generation Z who have and actively use social media and follow SMIs @raffinagita1717. Additionally, some respondents have tourism and hospitality education fields or are students and faculty members of the Tourism Polytechnic of Lombok, Lombok, West Nusa Tenggara, shown in Figure 1.

The snowball, or chain referral, is the nonprobability method of survey sample selection that is commonly used to locate hidden populations. [133]. In addition, the snowball sampling method entails identifying an initial set of relevant respondents and then asking them to suggest other research respondents who share common values and have some value and importance to the area of study [134].

Table 1. Type of Social Media Users.

Generation	Facebook	Instagram	TikTok	Twitter	YouTube	Others	Total
Millennials	15 (3.75%)	143 (35.75%)	8 (2%)	13(3.25%)	16 (4%)	5 (1.25%)	200
Zoomers	10(2.5%)	123(30.75%)	20(5%)	6 (1.5%)	24 (6%)	17(4.25%)	200
Total	25	266	28	19	40	22	400

explains the types of social media that are most widely used, and it can be seen that Instagram is the most favorite social media platform by Millennials and Zoomers. At the same time,

Table 2. Daily social media usage across generations.

Generation	<1 h	1–2 h	2–4 h	4–6 h	>6 h	Total
Millennials	18 (9%)	49 (24.5%)	57 (28.5%)	35(17.5%)	41(20.5%)	200
Zoomers	17 (8.5%)	31 (15.5%)	70(35%)	36 (18%)	46(23%)	200
Total	35	80	127	71	87	400

shows that most Millennials and Zoomers spend 2–4 h per day accessing the internet.

3.1. Dataset

The questionnaires contain seven case-based psychometric scales ranging from [135] to assess the reputation of influencer marketing. Since the respondents were autochthonous Indonesian citizens, this same questionnaire survey was translated into Indonesian.

Table 3. Coding Criteria Variables.

Construct	Indicator	Measure
Basic Information	Q1	Do you intend to take this survey?; Yes (1); No (0)
	Q2	How many social media accounts do you have?
	Q3	What social media do you use frequently every day? 1 = Facebook, 2 = Instagram, 3 = TikTok, 4 = Twitter, 5 = YouTube, 6 = Others
	Q4	How much time do you spend on social media every day? 1 = <1 h, 2 = 1–2 h, 3 = 2–4 h, 4 = 4–6 h, 5 => 6 h
	Q5	Do you know/ have you heard about Social Media Influencers (SMIs) before?
	Q6	Do you follow SMIs on social media? Such us: Raffi Ahmad and Nagita Slavina
	Q7	How many SMIs do you follow on social media?
	Income	1 = <Rp.1.000.0000; 2 = Rp.1.000.001-Rp.2.500.000; 3 = Rp.2.500.001-Rp.4.000.000; 4 = Rp.4.000.001-Rp.5.500.000; 5 => Rp.5.500.001
	Status Tourism Student	Yes (1); No (2)
	Education	1 = Junior school; 2 = High school; 3 = Diploma/Bachelor; 4 = Master; 5 = PhD
	Occupation	1 = Civil servants; 2 = State-owned company; 3 = Private-owned company; 4 = Bussinessman/Entrepreneurs; 5 = Student; 6 = Others
Attractiveness	ATR1	In my opinion, (SELECTED SMI) … (attractive–unattractive)
	ATR2	In my opinion, (SELECTED SMI) … (classy–not classy)
	ATR3	In my opinion, (SELECTED SMI) … (beautiful/handsome–ugly)
	ATR4	In my opinion, (SELECTED SMI) … (elegant–plain)
	ATR5	In my opinion, (SELECTED SMI) … (sexy–not sexy)
Trustworthiness	TRS1	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … dependable–undependable.
	TRS2	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … honest–dishonest
	TRS3	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … reliable–unreliable
	TRS4	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … sincere–insincere
	TRS5	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … untrustworthy–trustworthy
Expertise	EXP1	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … expert–not an expert
	EXP2	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … experienced–inexperienced
	EXP3	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … knowledgeable–unknowledgeable
	EXP4	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … qualified–unqualified
	EXP5	I think, when delivering tourism destination information through his/her social media account, (SELECTED SMI) … skilled–unskilled
Visit Intention	VIT1	When (SELECTED SMI) delivers tourist destination information through his social media, I will find out the information about the tourist destination (strongly disagree–strongly agree)
	VIT2	When (SELECTED SMI) delivers tourist destination information through his social media, I will consider visiting the tourist destination (strongly disagree–strongly agree)
	VIT3	When (SELECTED SMI) delivers tourist destination information through his social media, I will visit the tourist destination (strongly disagree–strongly agree)

represents the coding criteria of variables. Meanwhile, this study used 400 respondents from 18 provinces in Indonesia, and most of them came from West Nusa Tenggara and South Sulawesi, as shown in Figure 2.

The case study of SMIs used in this research is Raffi Ahmad and Nagita Slavina (@raffinagita1717). Raffi Ahmad, born in Bandung, West Java, Indonesia, on 17 February 1987, is an Indonesian actor, presenter, singer, entrepreneur, and producer of Sundanese and Pakistani descent. Nagita Slavina Mariana Tengker, better known as Nagita Slavina, born in Jakarta, Indonesia, on 17 February 1988, is an Indonesian actress, presenter, and singer of mixed Minahasan Javanese and Minangkabau descent. However, on 17 October 2014, Nagita Slavina married Raffi Ahmad. This celebrity couple had a remarkable impact on Indonesia’s industry and the showbiz industry. Raffi Ahmad and Nagita Slavina manage a YouTube channel called RANS Entertainment. The name RANS is the merging of the initials Raffi Ahmad (RA) and Nagita Slavina (NS), according to a report generated on 2 September 2021 on the HypeAuditor website until September 2021, and the couple has 54.9 m followers on Instagram with an engagement rate of 0.42% and 21.3 M subscribers on YouTube.

However, the audience interest of @raffinagita1717 is about 57.34% Business and Careers, 51.78% Photography, 47.96% Travel and Tourism, 46.34% Entertainment, 42.48% Art and Design, 41.64% Music, 41.28% Restaurants, Food and Grocery 41.16%, Sports 36.36%, Movies and TV 32.28%, Fitness and Yoga 31.74%, and Luxury Goods 31.44%. At the same time, RANS expanded its business, including food and beverage, RANS living furniture, and RANS Cilegon’s Football Club is an Indonesian football club based in Cilegon, Banten. RANS Cilegon’s FC is currently playing in Liga 2.

3.2. Feature Selection Using Logistics Variational Approximation

This section introduces representative feature selection techniques using logistics variational approximation. Feature selection has been actively used in machine learning and statistic techniques, such as data mining [136,137,138,139]. It is crucial to choose a set group with relevant features for constructing prediction models. There are many benefits of feature selection: improving model interpretability and overcoming the curse of dimensionality to boost the prediction model’s efficiency and data visualization, including processing cost and time [140,141,142,143]. This section describes the basic model setting to derive the variational method. We begin with a regression model [144] given by

$$P\left(y=1|x,\theta \right)=g\left({\theta }^{T}x\right)$$

where $g\left(x\right)={\left(1+\exp \left(-x\right) \right)}^{-1}$ is the logistic function; $y$ is the binary response variable; $x={\left(x_1,\dots ,x_r\right)}^T,x$ is a vector of explanatory factor. Prior distribution $\pi \left(\theta \right)$ is assumed Gaussian with a possibly full covariance structure [145,146,147,148,149,150]. Our predictive likelihood is therefore

$$P\left(y|x\right)=\int P\left(y|x,\theta \right)P\left|D\right)d\theta$$

where $D=\left\{D^1,\dots , D^T\right\}$ and $D^t\{y^t$,$x_1^t,\dots ,x_r^t\}.$ To obtain predicive value of $y$ given $x$, we must know the posterior distribution

$$P\left(\theta |D\right)=P\left(\theta |D^1,D^2,\dots , D^T\right)$$

where $D^t\left\{y^t, x_1^t,\dots ,x_r^t\right\}$ is a complete observation. Furthermore, we still obtain approximated posterior distribution and posterior predictive likelihood through the variational method. Here, we will use the idea described above so that logistic likelihood transformed into the variational can satisfy the conjugate property with Gaussian prior to manipulating variational approximation; let us take the following steps:

Step 1: Transforming the logistic Regression Function

$$\begin{gathered} \log g\left(x\right)=-\log \left(1+\exp \left(-x\right)\right) \\ =\frac{x}{2}-\log \left(\exp \left(\frac{x}{2}\right)+\exp \left(-\frac{x}{2}\right)\right) \end{gathered}$$

Remark:$\log g\left(x\right)$ is a convex function of $x^2$. You can verify this by taking the second derivative of $\log g\left(x\right)$ regarding $x^2.$

$$\begin{gathered} \mathbf{Step} \mathbf{2}: \mathbf{Obtain} \mathbf{the} \mathbf{tan}\mathbf{gent} \mathbf{line} \mathbf{of} \mathbf{log}\mathit{g}\left(\mathit{x}\right) \\ f\left(x\right)\ge f\left(\xi \right)+\frac{\partial f\left(\xi \right)}{\partial {\xi }^2}\left(x^2-{\xi }^2\right) \\ =-\frac{\xi }{2}+\log g\left(\xi \right)-\frac{1}{4\xi }\tanh \left(\frac{\xi }{2}\right)\left(x^2-{\xi }^2\right) \end{gathered}$$

Remark:$f\left(x\right)=-\log \exp \left(\frac{x}{2}\right)+\exp \left(-\frac{x}{2}\right)$ is a convex function in the variable $x^2.$ The tangent surface that is on the right-hand side is a gloval lower bound of $f\left(x\right).$ In the other words, we can lower bound $f\left(x\right)$ globally with a first-order Taylor expansion, including an example tangent line.

Step 3: Derivation of variational transform

$$\begin{gathered} P\left(y|x,\theta \right)=g\left(H_y\right) \\ \ge g\left(\xi \right)\exp \left\{\frac{H_y-\xi }{2}-\lambda \left(\xi \right)\left(H_y^2-{\xi }^2\right)\right\} \\ P\left(y|x,\theta \right)\equiv g\left(\xi \right)\exp \left\{\frac{H_y-\xi }{2}-\lambda \left(\xi \right)\left(H_y^2-{\xi }^2\right)\right\} \end{gathered}$$

Remark: combining Step 1 and Step 2 results, exponentiating yields is the variational transformation of the logistic function $g\left(H_y\right).$

$$H_y=\left(2y-1\right){\theta }^{T}x \mathrm{and} \lambda \left(\xi \right)=\tanh \frac{\frac{\xi }{2}}{4\xi }$$

$P\left(y|x,\theta ,\xi \right)$ is the variational lower bound on the logistic function $g\left(H_y\right).$

$P\left(y|x,\theta ,\xi \right)$ is the $\xi$ transformation of the conditional probabilty.

$P\left(y|x,\theta ,\xi \right)$ follows Gaussian with respect to $H_y$. For each value of $H_y,$ we can approximate $P\left(y|x,\theta ,\xi \right)$ to $H_y$.

Step 4: Deriving unnormalized posterior approximation

$$P\left(y|x,\theta \right) P\left(\theta \right)\ge P\left(y|x,\theta ,\xi \right) P\left(\theta \right)$$

Remark: The equation above is clear by the inequality of Step 3.

Step 5: Deriving normalized posterior distribution via the conjugacy

$$\frac{P\left(y|x,\theta ,\xi \right)}{\int P\left(y|x,\theta ,\xi \right)d\theta }~N({\mu }_{pos},{{\displaystyle \sum }}_{pos}^{-1})$$

Remark: given that $\theta ~N\left(\mu ,\mathsf{\Sigma }\right)$ and Gaussian variational form $P\left(y|x,\theta ,\xi \right)$, the normalized variational posterior distribution is Gaussian.

$$\begin{gathered} {{\displaystyle \sum }}_{pos}^{-1}={\mathsf{\Sigma }}^{-1}+2\lambda \left(\xi \right)x{x}^T \\ {\mu }_{pos}={\mathsf{\Sigma }}_{pos}\left[{\mathsf{\Sigma }}^{-1}\mu +\left(y-\frac{1}{2}\right)x \right] \end{gathered}$$

For a single observation $\left(y,x\right)$, where $x=\left(x_1,\dots ,x_n\right)$. Although, $P\left(y|x,\theta ,\xi \right)$ is a lower bound on the true conditional probability, our variational posterior approximation is a proper density and thus no longer abound. Successive observations can be incorporated into the posterior by applying the updates recursively.

$$\mathbf{Step} \mathbf{6}: \mathbf{Tuning} \mathbf{the} \mathbf{variational} \mathbf{parameter} \mathsf{\xi }$$

As we mentioned, the efficiency of the variational approximation depends on the value of variational parameter $\xi .$ Thus, we have to tune the parameter $\xi .$ How can we choose the value of $\xi ?$ The idea is simple. We choose a value of $\xi$ that yields a tighter lower bound in Step 4. In this, we can see the problem of inference changes to that optimization. By the way as the value of $H_y$ changes, the optimal value of $\xi$ changes at the same time. Therefore, we think of the method to measure the efficiency of $\xi .$ The following measure such that find $\xi$ so that maximizes the right-hand side of

$$\int P\left(y|x,\theta \right)P\left(\theta \right)d\theta \ge \int P\left(y|x,\theta ,\xi \right)P\left(\theta \right)d\theta$$

To achieve this object, we obtain the following EM optimization to update the procedure for $\xi$ parameter

$${\xi }^2=E\left\{{\left({\theta }^{T}x\right)}^2\right\}=x^T{\mathsf{\Sigma }}_{pos}x+{\left(x^T{\mu }_{pos}\right)}^2$$

where the expectation is taken with respect to $P\left(\theta |x,{\xi }^{old}\right)$. Owing to the EM-formulation, the update procedure makes $\int P\left(y|x,\theta ,\xi \right)P\left(\theta \right)d$ increases for each iteration. Until now, we have studied how to use the variational method to obtain a posterior approximation. In this paragraph, I will talk about the variational approach to obtaining a posterior predictive distribution. In Bayesian statistics, we define predictive likelihood like this

$$P\left(y|x,D\right)\int P\left(y|x,\theta \right)P\left(\theta \right)d\theta$$

Predictive likelihood means the probability of response variable $y$ given explanatory variables vector $x$ and past data $D$, where $D=\left\{D^1,D^2,\dots , D^T\right\}$ and $D^t\{y^t$,$x_1^t,\dots ,x_r^t\}$. Posterior distribution $P\left(\theta |D\right)$ comes from making a single pass through the dataset $D.$ The predictive lower bound $\log P(s^t|X^t,D)$ take the form:

$$\log P(s^t|X^t,D)=\log g({\xi }_t)-\frac{{\xi }_t}{2}+\lambda \left({\xi }_t\right){\xi }_t^2-\frac{1}{2}{\mu }^T{\mathsf{\Sigma }}^{-1}\mu +\frac{1}{2}{\mu }^T{\mathsf{\Sigma }}^{-1}\mu +\frac{1}{2}\log \frac{\left|{\mathsf{\Sigma }}_t\right|}{\left|\mathsf{\Sigma }\right|}$$

For any complete observation $D^t,$ where $\mu$ and $\mathsf{\Sigma }$ express the parameter in $P\left(\theta |D\right)$ and the subscript $t$ refer to the posterior $P\left(\theta |D,D^t\right)$ found by augmenting the data set to include the point $D^t$.

3.3. Structural Equation Modelling Using Hierarchical Likelihood

This study adopts hierarchical likelihood structural equation modeling to assess the hypotheses. Meanwhile, its recent advances and applications of the h-likelihood method have been used in survival data, competing risk models with frailty, joint models for longitudinal and survival outcomes, sparse high-dimensional multivariate analysis, spatial analysis, and multiple testing [96,98,99,102]. HGLM can be further extended by allowing additional random effects in their various components. Lee and Nelder [151] introduce a class of double HGLMs (DHGLMs) in which random effects can be specified in both the mean and the residual variances. Heteroscedasticity between clusters can be modeled by introducing random effects in the dispersion model as heterogeneity between clusters in the mean model. With DHGLMs, it is possible to have robust inference against outliers by allowing heavy-tailed distribution. Many models can be unified and extended further by the use of DHGLMs. Models can be further extended by introducing random effects in the variance terms. Suppose that conditional on the pair of random effects $\left(a, u\right)$, the response $y$ satisfies

$$E\left(y|a, u\right)=\mu \mathrm{and} var\left(y|a, u\right)=\varphi V\left(\mu \right).$$

The key extension is to introduce random effects into the component $\varphi$.

1.

Given $u$, the linear predictor for $\mu$ takes the HGLM form

$$\eta =g\left(\mu \right)=X\beta +Zv,$$

where $g()$ is the link function, $X$ and $Z$ are model matrices, $v=g_M\left(u\right)$ for some monotone function $g_M\left(u\right)$ are the random effects, and $\beta$ are the fixed effects. Moreover, dispersion parameters $\lambda$ for $u$ have the GLM form

$${\xi }_M=h_M\left(\lambda \right)=G_M{\gamma }_M,$$

where $h_M()$ is the link function, $G_M$ is the model matrix, and ${\gamma }_M$ are fixed effects.

2.

Given $a$, the linear predictor for $\varphi$ takes the HGLM form

$$\xi =h\left(\varphi \right)=G\gamma +Fb,$$

where $h()$ is the link function, $G$ and $F$ are model matrices, $b=g_D\left(a\right)$ for some monotone function $g_D\left(a\right)$ are the random effects, and $\gamma$ are the fixed effects. Moreover, dispersion parameters $\alpha$ for $a$ have the GLM form with

$${\xi }_D=h_D\left(\alpha \right)=G_D{\gamma }_D,$$

where $h_D()$ is the link function, $G_D$ is the model matrix, and ${\gamma }_D$ are fixed effects. Here, the labels $M$ and $D$ stand for mean and dispersion respectively.

Structural models describe the relationships that exist between latent variables. The relationship is generally linear [152,153,154,155,156]. The parameters indicating regression of endogenous variables on exogenous latent variables are labeled with Greek letter γ (gamma), while regression of endogenous latent variables on other endogenous latent variables is labeled with Greek letter β (beta). The following are examples of structural models in equation form and matrix form [82,157,158].

$$\begin{gathered} {\eta }_1={\gamma }_{11}{\xi }_1+{\gamma }_{12}{\xi }_2+{\zeta }_1 \\ {\eta }_2={\beta }_{21}{\eta }_1+{\zeta }_2 \\ {\eta }_3={\beta }_{31}{\eta }_1+{\gamma }_{32}{\xi }_2+{\zeta }_3 \\ \left[\begin{array}{c}{\eta }_1\\ {\eta }_2\\ {\eta }_3\end{array}\right]=\left[\begin{array}{ccc}0& 0& 0\\ {\beta }_{21}& 0& 0\\ {\beta }_{31}& 0& 0\end{array}\right]\left[\begin{array}{c}{\eta }_1\\ {\eta }_2\\ {\eta }_3\end{array}\right]+\left[\begin{array}{cc}{\gamma }_{11}& {\gamma }_{12}\\ 0& 0\\ 0& {\gamma }_{32}\end{array}\right]\left[\begin{array}{c}{\xi }_1\\ {\xi }_2\end{array}\right]+\left[\begin{array}{c}{\zeta }_1\\ {\zeta }_2\\ {\zeta }_3\end{array}\right] \end{gathered}$$

Structural model notation can be written as follows:

$$\eta =\mathsf{{\rm B}}\eta +\mathsf{\Gamma }\xi +\zeta$$

The above formula is a general matrix that represents the structural equation of the latent variable model. Therefore, we address the likelihood for fitting SEMs that support various combinations of different distributions for response variables. In DHGLMs, the h-likelihood can be defined by the logarithm of the joint density of the response $y$ and the unobserved vectors of random effects $v,p,$ and $q$ given by:

$$\begin{gathered} h=h\left(\beta ,\gamma ,\delta ,v,p,q;y\right) \\ h=\log f_{\beta ,\gamma ,\delta }\left(y|v,p,q\right)+\log f_{\delta }\left(v|q\right)+\log f_{\alpha }\left(p\right)+\log f_{\xi }\left(q\right) \end{gathered}$$

For estimation, we use $h$ for $v,p_v\left(h\right)$ for $\beta ,p_{v,\beta }\left(h\right)$ for $\left(\gamma ,\delta ,p,q\right),p_{b,\beta ,\gamma ,p}\left(h\right)$ for $\alpha$ and $p_{v,\beta ,\delta ,q}\left(h\right)$ for $\xi$ [102].

4. Results and Discussion

4.1. Finding Best Feature towards Visiting Intention

This paper incorporates an integrated and extended version of SEMs to identify and validate the research framework, which is discussed in Section 3. The approach comprises of SEM with hierarchical likelihood and variational logistics approximation also already appropriately discussed. Our new concept helps in identifying the multiple relationships simultaneously and testing the hypotheses associated with them and will be discussed clearly in this section. Meanwhile, HSEM assists in the simultaneous identification of various associations and the testing of theories related to them. According to this, we use the hierarchical likelihood for testing hypotheses and validating proposed frameworks, while neural network ability to detect nonlinear relationships as well as its machine learning implementation methodology make it inappropriate for theoretical frameworks.

Table 4. t-test.

Variables	t-Value	df	p-Value	Mean Difference	se	Lower CI	Upper CI	Results
Generation and Q3	−16.8178	399.0000	0.0000	−1.1225	0.0667	−1.2537	−0.9913	Significant difference
Generation and Q4	−26.5129	399.0000	0.0000	−1.7375	0.0655	−1.8663	−1.6087	Significant difference
Generation and Q5	12.6298	399.0000	0.0000	0.3575	0.0283	0.3019	0.4131	Significant difference
Generation and number of following	−36.7967	399.0000	0.0000	−3.2775	0.0891	−3.4526	−3.1024	Significant difference
Generation and Q4	−26.5129	399.0000	0.0000	−1.7375	0.0655	−1.8663	−1.6087	Significant difference

describes the t-test result. In general, among cross generations, there are significant differences in the frequency of social media use, the type of social media used, knowledge profile information about SMIs Raffi Ahmad and Nagita Slavina, and also the number of SMIs that the respondents are following on social media.

Figure 3 explains the variable information of the correlation from the dataset, and it can be seen that occupancy, birth year (generation), education, gender, information of respondents including the number of social media (Q2), type of social media (Q3), time spend (Q4), SMIs information (Q5), and a number of following SMIs (Q7) have a weak relationship. However, the expertise variable by SMIs Raffi Ahmad and Nagita Slavina has an important role in visiting tourism intention. In addition, the trustworthiness variable is also very important for Generation Y and Generation Z.

The selection technique using logistics variational approximation is very good to use because the target data for the dependent variable is continuous. One limitation of the technique is that it is only used for target binary responses. Based on

Table 5. Best Feature Selection.

Generation	Variable	Mean Importance	Median Importance	Minimum Importance	Max Importance	Norm Hits	Decision
Generation Y	Q2	−0.665	−0.830	−3.266	2.159	0.010	Rejected
	Q3	5.295	5.209	2.738	7.614	0.929	Confirmed
	Q4	1.397	1.387	0.242	2.955	0.000	Rejected
	Q5	0.742	0.809	−0.204	2.181	0.000	Rejected
	Q6	0.000	0.000	0.000	0.000	0.000	Rejected
	Q7	0.593	0.452	−0.273	2.257	0.000	Rejected
	ATR_RA1	11.630	11.653	9.854	1.353	1.000	Confirmed
	ATR_RA2	4.624	4.662	2.280	6.689	0.889	Confirmed
	ATR_RA3	5.080	5.018	3.012	7.706	0.960	Confirmed
	ATR_RA4	6.422	6.400	4.328	8.360	0.990	Confirmed
	ATR_RA5	5.588	5.657	3.462	7.906	0.980	Confirmed
	TRS_RA1	9.128	9.088	7.595	10.585	1.000	Confirmed
	TRS_RA2	11.865	11.861	10.324	13.260	1.000	Confirmed
	TRS_RA3	9.700	9.727	8.141	11.606	1.000	Confirmed
	TRS_RA4	10.499	10.565	9.036	12.244	1.000	Confirmed
	TRS_RA5	11.999	12.028	9.662	14.513	1.000	Confirmed
	EXP_RA1	7.817	7.763	6.004	9.980	1.000	Confirmed
	EXP_RA2	9.991	10.001	8.211	11.705	1.000	Confirmed
	EXP_RA3	8.298	8.265	6.183	10.275	1.000	Confirmed
	EXP_RA4	10.046	10.030	8.564	11.446	1.000	Confirmed
	EXP_RA5	6.924	6.997	4.913	8.839	1.000	Confirmed
	GENDER	2.833	2.786	0.147	4.857	0.495	Tentative
	EDUCATION	3.426	3.588	0.729	6.190	0.646	Tentative
	OCCUPATION	3.369	3.458	1.241	5.643	0.667	Tentative
	INCOME	2.844	2.931	0.169	5.230	0.505	Tentative
	Status Tourism Student	−1.931	−2.295	−2.859	−0.027	0.000	Rejected
Generation Z	Q2	0.220	0.239	−1.108	1.986	0.000	Rejected
	Q3	1.248	1.181	0.110	2.695	0.000	Rejected
	Q4	1.121	0.881	−0.156	3.483	0.000	Rejected
	Q5	−1.174	−1.542	−2.500	1.046	0.000	Rejected
	Q6	0.000	0.000	0.000	0.000	0.000	Rejected
	Q7	0.689	0.322	−1.371	3.225	0.020	Rejected
	ATR_RA1	9.160	9.089	7.546	10.771	1.000	Confirmed
	ATR_RA2	7.365	7.382	5.048	9.142	1.000	Confirmed
	ATR_RA3	9.817	9.692	7.365	11.684	1.000	Confirmed
	ATR_RA4	12.566	12.533	10.556	14.975	1.000	Confirmed
	ATR_RA5	5.585	5.510	3.426	7.480	0.970	Confirmed
	TRS_RA1	6.763	6.812	4.021	8.714	1.000	Confirmed
	TRS_RA2	4.338	4.279	1.963	6.615	0.869	Confirmed
	TRS_RA3	11.119	11.049	9.673	13.061	1.000	Confirmed
	TRS_RA4	8.349	8.403	6.806	10.674	1.000	Confirmed
	TRS_RA5	12.338	12.327	10.538	13.942	1.000	Confirmed
	EXP_RA1	6.940	7.102	4.870	8.811	1.000	Confirmed
	EXP_RA2	6.036	6.207	3.990	7.717	1.000	Confirmed
	EXP_RA3	11.598	11.594	9.765	13.261	1.000	Confirmed
	EXP_RA4	10.585	10.614	9.279	12.706	1.000	Confirmed
	EXP_RA5	11.985	12.045	9.478	14.414	1.000	Confirmed
	GENDER	−0.129	−0.074	−2.154	1.118	0.000	Rejected
	EDUCATION	−0.164	−0.594	−1.651	1.948	0.000	Rejected
	OCCUPATION	1.829	1.982	−0.591	4.057	0.061	Rejected
	INCOME	−1.022	−0.815	−2.047	−0.085	0.000	Rejected
	Status Tourism Student	2.282	2.289	0.065	4.793	0.364	Tentative

and Figure 4, it can be seen that, in general, Generation Y tends to consider that the type of social media affects the intention to visit. Generation Y has a tendency that attractive (Attribute 1), TRS (attribute 2), TRS (attribute 5), and Expertise (Attribute 4) have the highest contribution to the decision to visit a tourist spot. It is different that this generation does not pay attention to the duration of social media use, and also the attributes of gender, acceleration, occupation, and level of education are tentative variables. Generation Y pays attention to the influence of SMIs themselves, namely the attractiveness, trustworthiness, and professionalism of Raffi Ahmad and Nagita Slavina.

Nevertheless, different from Generation Y, Generation Z has a very high tendency towards the appearance of Raffi Ahmad and Nagita Slavina when compared to other variables. Interestingly, variables that are considered fundamental are not important, such as internet usage time, type of social media, SMI information, occupation, education, income, and gender. Generation Z is also known as a new generation. This is what differentiates Generation Z from its predecessor generation. For Generation Z, the internet is not something new because they were raised in an era where the development of the internet is so fast that the internet is a part of everyday life, which also causes the type of social media to not affect the decision to visit intention.

4.2. Addressing the Structural Equation Models Using Hierarchical Likelihood

The multivariate hierarchical generalized linear model (multivariate HGLM) is a hierarchical likelihood concept for dealing with multiple endpoints at the same time. Youngjo Lee and Nelder in 1996 proposed a class of models called the double hierarchical generalized linear model (double HGLM) in which random effects can be specified for both the mean and dispersion [151,159,160,161]. This class enables the use of models with heavy-tailed distributions.

The theory of likelihood implies that likelihood methods offer an efficient analysis if the model is correct. Thus, it is important to validate the model to verify the analysis and interpretation of the results. Thus, model checking plots and related model selection procedures are important. For binary data, it is challenging to check model assumptions such as the distribution of random effects so that various model selection criteria have been proposed [96].

The use of heavy-tailed distribution for random effect via DHGLMs provides robust analysis for distributional misspecification. The random effect for residual variance gives robust analysis against the outlier. During the simulation, using HSEM’s normal assumption in binary GLMMs can give non-ignorable biases if the normal assumption on random effects is not right. The likelihood inferences from the normal models are considered to be susceptible to the presence of outliers or model misspecification [98,99,100,102,148]. If the size of the data is limited, we can review the data carefully to detect outliers. Still, it can be difficult for large-scale data to identify outliers or degraded data. For such data, heavy-tail models are useful because they automatically give smaller weights during analysis to outliers or degraded data [97,102].

To begin with, question Q6 is a confirmation of whether the respondent follows Raffi Ahmad and Nagita Slavina or not. Then this variable is excluded for analysis using SEM-HL. The first stage is to analyze the suitability test using the Kaiser–Meyer–Olkin factor adequacy. Suppose the measurement model fit test has passed. In that case, the testing process can be conducted by testing the existing structural model, testing the fit of a measurement model, then testing the structural model, which includes two main parts. First, we should test the overall fit model of the structural model. Second, we should be testing the structural parameter estimate, which is the relationship between constructs or independent variables in the structural model. Although they have the same components, there are major differences between the measurement and structural models. The structural model is a relationship between constructs that has a causal relationship (cause-effect). Thus there will be an independent variable and a dependent variable. In contrast, in the measurement model, all variables are treated as independent variables. Suppose the overall fit model test is all considered good. In that case, the next process is to see a significant and close relationship between the independent and dependent variables.

Table 6. Variables evaluation.

Information	Attractiveness	Trustworthiness	Expertise	Visit Intention
Cronbach’s alpha	0.82024	0.94165	0.94325	0.90909
Composite Reliablity	0.76101	0.92596	0.92458	0.87342
Average Variance Extracted	0.81294	0.94200	0.94443	0.90977

shows the evaluation variable using Cronbach’s alpha, composite reliability, and average variance extracted. All of the variables are more than 0.7. At the same time,

Table 7. Factor loadings using hierarchical likelihood SEM.

No	Variable	Attribute	Estimate	Std. Err	t-Value	p-Value
1	Attractiveness	ATR_RA1	1.00000	0.00000	NA	NA
2	Attractiveness	ATR_RA2	0.84986	0.04825	17.61520	0.00000
3	Attractiveness	ATR_RA3	0.76237	0.04476	17.03368	0.00000
4	Attractiveness	ATR_RA4	0.91252	0.05616	16.24887	0.00000
5	Attractiveness	ATR_RA5	0.53298	0.07047	7.56301	0.00000
6	Trustworthiness	TRS_RA1	1.00000	0.00000	NA	NA
7	Trustworthiness	TRS_RA2	0.95955	0.04334	22.14259	0.00000
8	Trustworthiness	TRS_RA3	0.97795	0.03987	24.53103	0.00000
9	Trustworthiness	TRS_RA4	0.94667	0.04213	22.46859	0.00000
10	Trustworthiness	TRS_RA5	0.98924	0.04017	24.62774	0.00000
11	Expertise	EXP_RA1	1.00000	0.00000	NA	NA
12	Expertise	EXP_RA2	0.98271	0.03864	25.43153	0.00000
13	Expertise	EXP_RA3	0.97099	0.03940	24.64362	0.00000
14	Expertise	EXP_RA4	0.94420	0.03957	23.86005	0.00000
15	Expertise	EXP_RA5	0.91606	0.03726	24.58705	0.00000
16	Visit Intention	VIT_RA1	1.00000	0.00000	NA	NA
17	Visit Intention	VIT_RA2	1.05483	0.04481	23.53866	0.00000
18	Visit Intention	VIT_RA3	1.08315	0.04799	22.56885	0.00000

shows the value of factor loadings using HSEM, and all the variables are significant with p-$\mathrm{value}<\alpha =5\%$.

Table 8. Kaiser–Meyer–Olkin Factor Adequacy.

Basic Information	Score	Basic Information	Score	Basic Information	Score
Q2	0.51	ATR_RA1	0.95	TRS_RA1	0.96
Q3	0.62	ATR_RA2	0.95	TRS_RA2	0.96
Q4	0.65	ATR_RA3	0.95	TRS_RA3	0.97
Q5	0.61	ATR_RA4	0.93	TRS_RA4	0.95
Q7	0.87	ATR_RA5	0.84	TRS_RA5	0.95
EXP_RA1	0.95	VIT_RA1	0.95	EDUCATION	0.7
EXP_RA2	0.94	VIT_RA2	0.93	OCCUPATION	0.66
EXP_RA3	0.96	VIT_RA3	0.92	INCOME	0.66
EXP_RA4	0.97	VIT_RA4	0.97	Status Tourism Student	0.76
EXP_RA5	0.97	VIT_RA5	0.97	Average	0.94

shows the average value of KMO is 0.94 with value $X^2=7280.672,$ degree of freedom = 378, and significant with a p-value = 0.000. However, we will not use the variable with a score less than 0.50. If we run the analysis across Generation Y and Generation Z, there is a tendency that the duration of the use of social media does not have a significant effect on the decisions and intention to visit. Based on

Table 9. Intercept and variance for the response.

Intercepts for Responses					Variance for Responses
Variables	Estimate	Std. Err	t-Value	p-Value	Variables	Estimate	Std. Err	t-Value	p-Value
ATR_RA1	4.27	0.52	74.6599	0.0000	ATR_RA1	5.37	0.49	0.75	0.0000
ATR_RA2	6.84	0.44	93.0847	0.0000	ATR_RA2	4.02	0.36	1.52	0.0000
ATR_RA3	4.77	0.41	96.0187	0.0000	ATR_RA3	3.73	0.32	4.32	0.0000
ATR_RA4	3.33	0.51	74.4612	0.0000	ATR_RA4	6.40	0.53	7.35	0.0000
ATR_RA5	5.75	0.58	45.6947	0.0000	ATR_RA5	2.58	1.18	6.52	0.0000
TRS_RA1	4.90	0.47	84.1567	0.0000	TRS_RA1	3.61	0.29	3.72	0.0000
TRS_RA2	3.33	0.44	85.3341	0.0000	TRS_RA2	3.03	0.25	2.55	0.0000
TRS_RA3	3.72	0.43	89.4868	0.0000	TRS_RA3	1.89	0.17	0.79	0.0000
TRS_RA4	4.44	0.43	89.7800	0.0000	TRS_RA4	2.77	0.23	1.70	0.0000
TRS_RA5	5.33	0.43	92.3947	0.0000	TRS_RA5	1.89	0.17	7.18	0.0000
EXP_RA1	4.13	0.47	82.7218	0.0000	EXP_RA1	3.04	0.26	6.31	0.0000
EXP_RA2	5.47	0.45	89.4320	0.0000	EXP_RA2	2.31	0.21	1.92	0.0000
EXP_RA3	3.82	0.45	85.4615	0.0000	EXP_RA3	2.61	0.22	4.67	0.0000
EXP_RA4	4.51	0.44	88.0473	0.0000	EXP_RA4	2.82	0.24	6.93	0.0000
EXP_RA5	6.23	0.42	96.1799	0.0000	EXP_RA5	2.34	0.20	4.85	0.0000
VIT_RA1	3.16	0.49	76.7331	0.0000	VIT_RA1	3.72	0.35	5.01	0.0000
VIT_RA2	2.53	0.49	75.3959	0.0000	VIT_RA2	2.60	0.30	4.34	0.0000
VIT_RA3	0.82	0.52	68.0540	0.0000	VIT_RA3	3.60	0.36	6.41	0.0000

, the covariance value shows that all variables give the result with the p-value < $\alpha =5\%$. Table 6 explains that Attractiveness contributed to Trustworthiness by 54.9% and Expertise by 49.7%, respectively. However, Trustworthiness contributed 57.8% to expertise.

The scaled deviance for the hierarchical generalized linear model (HGLM) in which random effects were specified based on the conditional log-likelihood can be expressed by $l_1\left(\mu ;y|v\right)=\log g(y|v;\mu )$ for $y$ given random effect $v$ [151]. Therefore, all the variables can be used for further construction using SEM H-Likelihood. To measure this model, we use scaled deviance = 501.30040, degree of freedom = 129, cAIC = 759.3004, CTI = 0.84354, TLI 0.93304, and RMSEA = 0.08509 since we have continuous outcomes. Figure 5 represents the path analysis, and Figure 6 shows the model checking plot of HSEM. The following section considers mixed types of dependent variables, including H continuous variables, N ordinal variables, and G nominal variables. The multiple dependent variables are endogenously correlated with other dependent variables and affected by latent psychological variables where social interactions take place. There are $H$ continuous outcomes with index $continuous$ variable of individual $q, y_{qh}$. It can be defined as:

$$y_{qh}=\left(y_h^{\prime }+\left(d_h{z}_q^{*}\right)\right)x_q+{\epsilon }_{qh}$$

where $x_q$ is a $\left(A\times 1\right)$ vector of explanatory variables, including observed exogenous variables, as well as other endogenous dependent variables. $y_h^{\prime }$ is a $\left(A\times 1\right)$ vector of corresponding coefficients of $x_q$ captures the pure effect of explanatory variables. $d_h$ is a $\left(A\times L\right)$ matrix that captures the effect of $x_q$ through a latent variable with visit intention. However, ${\epsilon }_{qh}$ is a normally distributed random error term.

5. Conclusions

To increase enthusiasm in visiting local tourist attractions, several aspects need to be considered, such as something to see, which refers to special attractions tourists can see. However, the tourist can do the activities by visiting the tourist attraction and buying souvenirs. Some Millennials are much comfier making this decision as an employee in a company, while others are bold enough to start their businesses as businesspeople. This employee motivation will be even stronger in Generation Z because they have seen numerous multiple facets of entrepreneurial behavior, and in general, they desire a more self-governing and far less restrictive working atmosphere.

This is also a justifying reason for the variables education, occupation, and income to be tentative variables. Millennial people are more interested in having saved or investing money than today’s youth if it was at the same age. This same Gen Z group is more eager to spend money on items that have high value and quality. However, the Millennial generation is more interested in the buying process itself and likes to try new things by having to travel.

Millennial and Generation Z tourists thrive in the Internet of Things (IoT) era. An era in which information can be accessed via various forms of technology across multiple platforms. The Internet of Things era directly impacted the emergence of digital transformation, which was the catalyst for the Tourism 4.0 trend. From this research, it was also found that Instagram is an excellent medium for promotion. From the user’s point of view, Instagram is very closely related to their cycle of looking for inspiration to share moments on their travels. Most Millennials and Generation Z mention that this platform is important for their personal use. Millennials and Generation Z consider relevant posts to motivate them to travel, increase their desire to visit, and even influence their opinion on tourism brands. Again, this shows the importance of Instagram in influencing Millennial decisions. According to these primary highlights and interests, Instagram is able to play a significant role in travel marketing by focusing on the visual aspect (photos and videos). As more than just a side benefit, this platform is highly recommended as part of a marketing strategy, particularly for tourist destinations.

6. Recommendation and Practical Implications

The limitation of this study is that we only focus on Millennials and Zoomers generations. Long story short, each generation has unique characteristics that are different from one another. From the analysis using HSEM, this method can obtain information on which variables are important to Generation Y and Generation Z regarding attractiveness, trustworthiness, expertise variables possessed by social media influencers towards tourism visit intention. Raffi Ahmad and Nagita Slavina are influencers who already have brand personalities, making it easier for them at the beginning of their careers as actors in television media programs. Currently, they are now very active on the YouTube channel. To support performance in building a loyal community and be notified by Penta-Helix contributors, they only need to enrich their knowledge management. In addition, they have started to be concerned with data analytics and data science for social media engagement, which is described in Figure 7a.

This step will be useful in setting goals efficiently for decision making. Figure 7b describes the stages of how newcomer SMIs can obtain a loyal community and are noticed by Penta-Helix contributors. The newcomer SMIs actors need to be creative in preparing content, identifying audiences, and setting goals. Therefore, the brand personality will be well-formed. Newcomer SMI actors also need to be sensitive and understand knowledge management and open innovation. At this stage, the data analytics skills are also needed to evaluate whether the content being worked on is good enough in shaping brand personality so that a loyal community will be formed and can be noticed by Penta-helix contributors [102]. The contribution of Penta-Helix is needed involving the public, private industry, academia, civil society, and NGOs.

Understanding the use of data analytics and data science towards big data as a recommendation system is very important [162,163,164,165]. At this stage, it can be used to evaluate and assess the effectiveness of the content created and the target audience and market segmentation. Then as an evaluation material for the formation of an appropriate brand personality. Thus, a community that is loyal to SMIs actors is formed. Contribution and Cooperation in Penta-Helix are needed involving the public, private industry, academia, civil society, and NGOs [102,166].