How the Spread of User-Generated Contents (UGC) Shapes International Tourism Distribution: Using Agent-Based Modeling to Inform Strategic UGC Marketing

Social Media Channel Differentiation

Social media can be divided into two general types: relationship-based environments and topic/interest-based environments. Relationship-based environments, such as SNS (e.g., Facebook, Twitter, and Instagram), facilitate information exchanges between people sharing social connections (Sotiriadis and Van Zyl 2013). Topic/interest-based environments are publicly accessible information settings that welcome anyone to share/retrieve information on common topics/interests, such RS (e.g., TripAdvisor), blogs and collaborative communication sites (e.g., Wikipedia), and content community sites (e.g., YouTube and Flickr). Considering the extensive influence that SNS and RS exhibit on travel decisions (Casado-Díaz et al. 2020), they are adopted to represent relationship-based and topic/interest-based environments, respectively.

These two social media channels differ significantly in nature. While SNS demands social connections as a prerequisite for information access, RS allows information to be accessible to the broader public. As such RS is featured by a higher level of information centrality, which likely results in the generally perceived greater reliability of RS than SNS (Kim and Park 2017). Also, evaluations on sharers are more essential to SNS information retrieval, given SNS provides richer information than RS on sharers (e.g., travel expertise/experiences/preferences, and sharer-user social status similarity and relationship closeness) (Liu, Wu, and Li 2019).

RS also affects individual travel planning in more explicit ways (Munar and Jacobsen 2014). People are prone to actively search information from RS for travel-planning assistance, whereas SNS engagement is more passive, primarily in terms of its often subconscious impacts on tourists’ cognition/affection (Jacobsen and Munar 2012), such as the offer of travel inspiration (Brown 2017). However, the practical importance of passive SNS influences should not be underestimated. Notably, a large proportion of tourists are “lurkers” who do not directly react to social media posts or post themselves yet with choices influenced by exposure to those posts (Sotiriadis and Van Zyl 2013). Because of these nature differences, individuals normally treat these channels differently when retrieving, evaluating, and sharing UGC (Kaosiri et al. 2019), hence the need to explore the cross-channel UGC effect variations.

UGC Studies in Tourism and Hospitality

Numerous studies in tourism and hospitality fields have profiled tourists’ UGC sharing/using behaviors on social media platforms (e.g., Bigné, William, and Soria-Olivas 2020; Bronner and de Hoog 2011), identified content types likely to be shared or found useful on social media (e.g., Önder, Gunter, and Gindl 2020), and revealed factors that influence UGC sharing (e.g., trip satisfaction/dissatisfaction and previous sharing experiences) and using behaviors (e.g., UGC usefulness, trust in and perceived preferential similarity to sharers) (e.g., Oliveira, Araujo, and Tam 2020; Shin, Perdue, and Pandelaere 2020). Research also identified effects on tourist expectations and purchase intentions from posted UGC characteristics (e.g., ratings, valence, lengths, readability, and language style) and from sharer characteristics (e.g., travel and posting frequencies) (e.g., Liu, Wu, and Li 2019; Kaosiri et al. 2019).

The current research focuses on UGC effect examinations, as most existing work evaluate UGC effects on purchase intention/preference rather than more concrete outcomes such as actual arrivals. Additionally, while most studies explored the effects of UGC evaluations on tourist attitudes/behaviors, effects of the fundamental forces shaping the distribution and influence of these UGC evaluations—individual UGC sharing/using behavioral features—are still largely underexamined yet valuable to understand and manage the root of UGC influences (Q2–3).

Moreover, UGC spreading features and effects are noted by studies as varying by region. It was found that US and Spanish tourists predominantly view RS while Swiss and UK tourists focus on SNS; German and Nordic tourists less are likely to use or be influenced by social media versus Chinese tourists, who actively adopt UGC (Gretzel, Kang, and Lee 2008; Jacobsen and Munar 2012; Wilson, Murphy, and Fierro 2012). Kietzmann and Canhoto (2013) accordingly proposed the need to identify the context-based variations of UGC effects.

Overall, the pressing research demand has shifted focus from generally depicting UGC communicative patterns (“what”) and effects on tourist attitudes/behaviors (“to what extent”), to causal mechanisms (“how”) and underlying psychological/sociocultural rationale (“why”) of UGC distinctly shaping tourist attitudes/behaviors (Hudson and Thal 2013). As UGC patterns and extent of effects are ever-evolving, identifying the causal mechanisms of UGC effects (“how”) is a vital yet unexplored area, it being the premises for predicting long-term UGC patterns/effects and strategic UGC marketing (Q1).

This study addresses the aforementioned gaps revealed in existing literature, which can be summarized into the demand for a holistic evaluation of UGC effects that accounts for (1) the causal mechanisms of UGC effects (Q1), particularly as driven by the fundamental forces of individual UGC sharing/using features (Q2–4); (2) simultaneous UGC effects via multiple social media channels (Q5); (3) the underrepresented passive/inactive UGC contributors/users with UGC effect expanding promises; (4) the region-dependent nature of UGC effects (Q6); and (5) longitudinal evaluation of UGC effects considering Internet’s longer-term UGC storage and consequently more long-lasting effects than traditional word-of-mouth impacts (Q7).

Methodology-wise, UGC effects in tourism and hospitality have mostly been analyzed with correlation explorations (Lu and Stepchenkova 2015). These include advanced big data analyses that harvest vast social media comments/ratings to reveal correlations between social media sharer/user features, UGC content, and actual customer behaviors to provide competitively accurate behavioral predictions (Li et al. 2018). Despite their leading predictive power and generalizability, these cutting-edge methods are still limited in ability to draw causal effects from UGC features, given the absence of control group in real-life settings (Zeng and Gerritsen 2014). Revealing the causal mechanisms of UGC effects is nevertheless warranted to develop effective UGC marketing and management interventions.

Experimental designs have been adopted to identify such causal mechanisms, by manipulating UGC features (e.g., valence, order, and contributor expertise) (e.g., Ladhari and Michaud 2015). Both influential UGC features to individual choices and potential moderators of UGC effects were detected (e.g., level of involvement and trust/knowledge about the commented businesses/destination). While experimental design is the most restrictive form of causality analyses, its results are hardly generalizable from simplified and restrictively controlled lab settings to complex and dynamic real-life settings. Furthermore, actual purchases/arrivals, the outcome of greatest industrial interest, cannot be observed in regular experimental settings. An exception regarding those limitations is field experiment, which however is costly and hardly controllable in studying UGC effects. There is thus a pressing need for a methodology that can cost-effectively and restrictively identify causal effects as well as replicate real-life settings, where agent-based modeling (ABM) is proposed to fulfill these methodological demands.

Agent-Based Modeling

Agent-based modeling (ABM) is a class of computational models that can simulate a complex macro-level system (e.g., global tourist–destination system) based on formally assumed simple behavioral rules of individual agents (e.g., tourists) and learning algorithms (Nicholls, Amelung, and Student 2017). By simulating micro-level agents’ behavioral processes (e.g., individual tourists’ UGC using/sharing behaviors) and interactions with each other (e.g., UGC communication between tourists), it allows the detection of macro-level pattern variations (e.g., global visit distribution variations) caused by individual agents’ behavioral changes, which is hardly observable using traditional analytical models. It demonstrates advantages in unveiling the hidden causal mechanisms driving the macro-level evolvement (Bonabeau 2002).

International tourism is a highly complex system, in that it contains autonomous and heterogeneous tourists, diverse destinations, and various travel decision-aid sources, as well as potential interactions/evolvements of these components. When resolved with traditional analytical models, the multifaceted realities are largely simplified/abstracted to build theories with generalizability at the expense of accuracy (Bonabeau 2002). The unrealistic assumptions (e.g., homogeneity, linearity, and equilibrium) often fail to gauge the complex travel patterns (Nicholls, Amelung, and Student 2017). ABM instead allows tourists to be heterogeneous in behavioral patterns, make boundedly rational decisions based on imperfect information collection/interpretation, perform evaluations based on interactions with each other and the environment, and adapt based on travel experiences and environmental changes (Kiesling et al. 2012). These are possible given the ability of ABM in simultaneously estimating a large number of parameters, nonlinear relationships, and multiway interactions, which is impossible given the limited computing power of traditional models (Bonabeau 2002). ABM is thereby a well-suited tool for identifying causal mechanisms of UGC effects on international arrival distribution with both restrictiveness and realism.

ABM can be employed with restrictive validations against empirical data and maximally replicating reality, to produce an accurate prediction of future system patterns (Bruch and Atwell 2015). Yet when empirical data is unavailable, ABM functions as an explanatory tool. It can experiment with possible scenarios simulated and compare outcomes to identify reasonable explanations and propose theoretical advances, without having to be anchored to existing empirical evidence (Willer, Kuwabara, and Macy 2009).

The purpose of this study is two-fold: it is explanatory with the mission of identifying the most likely causal mechanisms underlying the UGC effects on international arrival distribution (Q1), and also approximately predict the comparative variations of international arrivals between global regions (the distributional variation) corresponding to UGC behavioral changes (Q2–7) in both the short and long term. There is no intention to predict the precise arrival number of each country, given the inaccessibility of comprehensive statistics on global UGC distribution, and the impossibility to account for all potential influential factors to international tourism.

ABM Applications in Destination Choices

Tourism scholars have shown the growing interest in employing ABM to simulate tourist–destination dynamics (Nicholls, Amelung, and Student 2017). They primarily explore how a contextual variation (e.g., sociodemographic/economic/policy) may shape the future tourism demand (e.g., Vinogradov, Leick, and Kivedal 2020), or develop a model forecasting future visitations (e.g., Johnson and Sieber 2011). These studies have generally modeled a narrow set of homogeneous destinations within a smaller geographical scope (e.g., city/province/country), at which level richer individual/destination data become accessible for validations and hence allows for more accurate quantitative predictions. Scholars also recognized that even after modeling within a smaller geographical scope the accuracy of quantitative predictions is still limited, pending on the number of parameters to be validated (Batty 2008).

A parallel line of ABM applications has focused on developing theoretical ABM models not grounded in empirical data (e.g., Alvarez and Brida 2019). The more relevant work to the current study, involving Boavida-Portugal, Ferreira, and Rocha (2017) that pioneer the attempt to incorporate social network influences in developing a theoretical ABM model of destination choices, and Zhai, Zhong, and Luo (2019) focusing on how a destination’s SNS posts could influence the collective tourist emotion.

Potential advancements of these existing attempts are spotted. Deterministic modeling of tourist choices is still dominant in existing attempts, where a destination is selected once a tourist’s destination evaluative threshold is exceeded, or an array of destination selective criteria are met (e.g., price level/distance/weather/capacity). A probabilistic perspective is nevertheless demanded to reflect the randomness in tourist choices based on contextual influences (e.g., companions’ interest), limited information access and processing capability, or decision-making heuristics (i.e., speed-accuracy tradeoff) (Van Middelkoop, Borgers, and Timmermans 2003). Also, a destination attribute typology that is more comprehensive (including infrastructure, sustainability practices, and safety measures) and generalizable to different destinations is demanded for cross-destination validations and benchmarking. Furthermore, the weighted importance of different attributes in destination selection and satisfaction determination should be incorporated. Additional important determinants of social influences beyond similarity with social connections should also be considered, such as expertise of connections or between-channel information accessibility differences. Lastly, novelty has been popularly adopted as a destination selection criterion, but the presence of loyalty that drives repeated visits is largely ignored. The current study adds to the existing ABM applications in destination choice studies by addressing these gaps.

Overview

Entities, state variables and scales

One type of entities is international travel destinations, where 134 patches represent 134 countries of 14 global regions. The countries are selected following Travel and Tourism Competitiveness Report 2014 (World Economic Forum 2015) and characterized by destination attribute-performance values retrieved from the report, the real-time updated overall RS rating by visitors, as well as the current and accumulative numbers of arrivals. The report has the earliest evaluation available on global destinations that contains a comprehensive and standardized list of attributes for cross-country and over-year performance comparisons. It is also competitively objective and accurate in reflecting actual destination performance levels, as compared to subjective destination ratings by certain tourist segments. The report includes each country’s performance levels (using a 7-point scale) on 12 crucial attributes to destination competitiveness and tourist satisfaction. The destination attribute values are fixed at the 2014 level ¹ and all other parameter initializations correspondingly adopt the data of the same year, so that international arrivals of following years can be used for validations. The destination attribute settings (attri.num and attri.perf) are detailed in Table 1, where the rationale for setting all model parameters is specified.

OPEN IN VIEWER

Table 1.

Setting up the Fixed and Calibrated Parameters.

Calibrated/Sensitive Parameters	Description	Calibration Restriction	Calibrated Initial Value
WRS WSNS	The cross-tourist baseline weighting of RS and SNS posts in determining destination choices; individuals may update these overtimes based on their perceived value of two UGC sources (WRS.i(t) and WSNS.i(t))	Within the range of [0.01, 0.5]; either source is assumed as weighed more heavily in destination evaluations given no consensus in literature (e.g., Munar and Jacobsen 2014; Jacobsen and Munar 2012)	0.185 0.269
prob.RS.share.mean prob.SNS.share.mean prob.RS.share.sd prob.SNS.share.sd	The cross-tourist normal distribution of probability to share travel memories on RS/SNS	Given the huge global international tourist population, prob.RS/SNS.share/use should follow a normal distribution (with means and SDs calibrated within [0.01, 1] and [0, 0.2], respectively	0.165 0.708 0.17 0.142
prob.RS.use.mean prob.SNS.use.mean prob.RS.use.sd prob.SNS.use.sd	The cross-tourist normal distribution of probability to use RS/SNS information about destinations		0.1670.1980.0450.17
attri.prefer.mean attri.prefer.sd	The cross-tourist normal distribution of preference level for each destination attribute ( a t t r i . p r e f e r h i n )	With the assumed normal distribution of destination preference levels across the global tourist population, the mean was calibrated within [1, 7] to match the scale of destination attribute performances, and SD was calibrated within [0, 2]	3.060.85
percent.weak.links	Percentage of SNS connections with relatively weaker influences on an individual’s destination evaluations	Calibrated within [0.5, 1], given the empirical evidence in Luo and Zhang (2015) that the majority of SNS connections (e.g., 69% in Chinese samples) are weak links	0.768
relative.link.strength	The relative strength of weak SNS links to strong links, with the strength of strong links as the reference (=1)	Calibrated within [0.1, 0.5], an empirically established range (Luo and Zhang 2015; Reichelt, Sievert, and Jacob 2014)	0.293
expertise.ratio closeness.ratio similarity.ratio	The extent of influence from a tourist’s perceived expertise, closeness, and similarity of SNS connections on the perceived credibility of their SNS shares	Calibrated within [0.01, 1]	0.1710.4170.411
prob.repeat	Probability for a tourist to visit an international country repeatedly	Calibrated within [0, 1]; prob.repeat > .5 in existing empirical studies (e.g., Aliman et al. 2016; Alegre and Cladera 2006), yet their data are from popular global destinations, while here less-popular countries are also incorporated	0.562
prob.multidest	Probability for a tourist to visit another country within the same region during the trip	Calibrated within [0, 1]	0.575
R	Coefficient in tourist satisfaction estimation	Calibrated within [0, 1]	0.667
F	The relative weight of most recent UGC information to all acquired UGC information	Calibrated within [0, 1], no unanimous support in literature for recency effect or primacy effect (Qiu and Wang 2011)	0.619
k1	The coefficient for the linear term of satisfaction to update the probability of UGC sharing	Calibrated within [0, 1]	0.243
k2	The coefficient for the quadratic term of satisfaction to update the probability of UGC sharing	Calibrated within [0, 1]	0.62
B	Coefficient determining the WRS/SNS updates	Calibrated within [0, 1]	0.55
Fixed/less-sensitive Parameters	Description	Initial Value
attri.num	Number of destination attributes	12, according to Global Destination Competitiveness Report 2015-2019. They are (1) natural resources; (2) cultural resources and business travel; (3) air transport, ground, and port infrastructure; (4) tourist service infrastructure (e.g., accommodations, car rental, air travel); (5) international openness; (6) price competitiveness; (7) environmental sustainability; (8) tourism business environment and prioritization of travel/tourism; (9) safety and security; (10) health and hygiene; (11) human resources/labor market; and (12) technology/ICT readiness. Other potential determinants (e.g., geographical and cultural distances) are only accounted for within individual-specific intercepts given the lack of consensus on the inhibitory vs. attractive effects from distances on choosing foreign destinations (McKercher, Chan, and Lam 2008; Nicolau and Mas 2006).
attri.perf	The actual performance levels of attributes for a destination	The within-region actual performance level of each attribute (mean and SD across countries) for 14 global regions in 2014 were retrieved from 2015 Global Destination Competitiveness Report (World Economic Forum 2015)
n1-3	Numbers of basic, performance, and excitement types of destination attributes for each tourist	n1 = n2 = n3 = 4, to eliminate the confounding effect from uneven attribute distribution across three attribute types
prob.travel	The relative international travel probability of tourists from fourteen global regions	The retrieved average number of international trips taken by people living in each region in 2014 (UNWTO 2015), the relative international travel probability is calculated for each region using Northern Europe with the highest international travel frequency as the reference probability level of 1.
pR	Rewiring probability to develop SNS network	0.26; obtained by averaging the estimates of 12 popular SNS of different types over a five-year period (2005–2010) (Kang et al. 2015)
clust.coeff	Clustering coefficient	The SNS connections are rewired until clust.coeff falling in [0.04, 0.17], based on estimates of 12 most popular SNS sites by Kang et al. (2015)
e	The context-invariant uniqueness of each tourist	A fixed random number for each tourist following standard normal distribution (mean = 0, SD = 1)
ε	Random error in dest.eval that represents alternative factors not accounted for (e.g., marketing information)	A random number picked from [–1, 1] uniform distribution at each time step

Tourists are agents in the model (n=100, representing 100% of global international tourist population). They are randomly assigned to 134 countries as their countries of residence, with cross-region proportions following 2014 global population distribution (World Bank 2015). Each tourist is characterized by his or her country of residence, probability to travel internationally, preference about destination attributes, importance/three-factor designation of attributes, memory of previous trips (i.e., destinations and historic satisfaction levels), RS/SNS sharing/using patterns (i.e., probability to share/use), weighting of RS/SNS in decision making, number and strengths of SNS connections, and real-time evaluation updates from connections. The model runs for 100 time-steps, with each time-step equivalent to 6 months in length.

Process overview and scheduling

As demonstrated in Figure 1, at each time-step, a tourist (i) travels internationally at a probability (prob.travel_i). He or she searches for information both internally (from memory of previous experiences and accumulative SNS influences) and externally from RS and other sources for destination evaluation (Hwang et al. 2006). In evaluating destinations, the tourist uses SNS-shared destination evaluations at the probability prob.SNS.use_i and searches for RS destination ratings at prob.RS.use_i. For destinations with the highest evaluation scores, tourist only consider a previously visited destination at a probability of prob.repeat. Then the tourist is more likely to visit the destination (h) with the highest overall evaluation ( d e s t . e v a l h i ( t ) ) at a time-step (t), yet only at a probability proportional to the destination’s evaluation score (P_hi(t)). After a visit, he or she evaluates the satisfaction with the destination (dest.sat_hi(t)) and probabilistically chooses to share the dis/satisfaction level in terms of UGC via the two channels (at prob.RS.share_hi(t) and prob.SNS.share_hi(t)). At the probability prob.multidest this tourist chooses to visit multiple destination on a trip and accordingly further evaluate another destination within the same region. At each time-step, each tourist updates the memory of visited destinations and associated satisfaction levels, as well as received SNS shares from connections; similarly, each destination updates its current and accumulative visit numbers as well as overall RS evaluation.

OPEN IN VIEWER

Figure 1.

Diagram of user-generated content (UGC) mechanisms.

Design Concepts

Theoretical and empirical background

The rationale underlying the overall tourist behavioral process is grounded in innovation diffusion theory (Rogers 2010) that explains how communication of an idea or product gains momentum and spreads through a social system over time, possibly via different communication channels (e.g., mass communication or word-of-mouth). The adapted diffusion process by Thiriot and Kant (2008) is adopted given its superiority in representing the UGC spreading process, wherein the individuals’ innovation diffusion follows the steps of innovation awareness, information searching, adoption decision making, and word spreading.

The assumptions of each submodel and theoretical/empirical support are described as follows, in the order illustrated in Figure 1. More estimation details about each submodel can be found in Table 2.

OPEN IN VIEWER

Table 2.

Submodel Estimating Equations.

Submodel	Equation	Description
Receiving SNS evaluations	Wij = e x p e r t i s e . r a t i o * t r a v . f r e q j ∑ n = 1 c o n n e c t i o n . n u m i t r a v . f r e q n + c l o s e n e s s . r a t i o * l i n k . s t r e n g t h i j + s i m i l a r i t y . r a t i o * ∑ k = 1 a t t r i . n u m | a t t r i . p r e f e r i k − a t t r i . p r e f e r j k | ∑ j = 1 c o n n e c t i o n . n u m i ∑ k = 1 a t t r i . n u m | a t t r i . p r e f e r i k − a t t r i . p r e f e r j k |  (1)   l i n k . s t r e n g t h i j = { 1 , i f i j l i n k i s s t r o n g r e l a t i v e . l i n k . s t r e n g t h , i f i j l i n k i s w e a k  (2)	• expertise.ratio, closeness.ratio, and similarity.ratio denote the extent of influence from tourist i’s perceived expertise, closeness, and similarity of his or her SNS connection j on his or her value weighting of SNS shares by j.• Perceived expertise is represented by the relative travel frequency of connection j ( t r a v . f r e q j ) to tourist i’s other connections, calculated as dividing t r a v . f r e q j by the sum of travel frequencies across tourist i’s all SNS connections. The total number of connections (connection.numi) is counted and stored with each tourist once the small-world SNS network is set up.• Perceived closeness indicates to what extent is tourist i bonding with connection j and fully trusting j in their honestly passing on destination information, represented by link strength between i and j ( l i n k . s t r e n g t h i j ) and family and friends are generally assigned strong links. The strength of strong links is set to 1, so that the relative strength of weak links to strong ones (relative.link.strength) represents the strength of weak links.• Perceived similarity of connection j by tourist i is calculated based on the sum of tourist i’s attribute preference differences with all SNS connections.• All of these variables are kept in the same scale of [0,1] to avoid scale difference bias.
	S N S . m e a n h i ( t ) = (1 – f) * S N S . m e a n h i ( t − 1 ) + f * ∑ j = 1 S N S . s h a r e r . n u m h i ( t ) ( W i j * d e s t . s a t h j ( t ) ) S N S . s h a r e r . n u m h i ( t )  (3)	• Tourist i’s received updated SNS evaluations about destination h at time t is thus calculated with the weighted (Wij in equation 1) average of destination dis/satisfaction ratings by those SNS connections who just visited and shared SNS about destination h ( d e s t . s a t h j ( t ) )• Given how many connections have just visited destination h and the probability for each connection j to share dis/satisfaction rating ( d e s t . s a t h j ( t ) ) with tourist i ( p r o b . S N S . s h a r e j ( t ) ), the number of sharers ( S N S . s h a r e r . n u m h i ( t ) ) is counted at each time step for equation 3 calculation, wherein the coefficient f, as the relative weight of the most recent UGC to previously acquired UGC information, accounts for differentiated effects from UGC posts by time order (Qiu and Wang 2011).
Retrieving RS evaluations	R S . m e a n h ( t ) = (1 – f) * R S . m e a n h ( t − 1 ) + f * ∑ l = 1 R S . s h a r e r . n u m h ( t ) ( M l ∑ l = 1 R S . s h a r e r . n u m h ( t ) M l * d e s t . s a t h l ( t ) ) R S . s h a r e r . n u m h ( t )  (4)	• The average RS rating on destination h at time t ( R S . m e a n h ( t ) ) is calculated by equation 4, where R S . s h a r e r . n u m h ( t ) as the number of reviews is determined by who (l) visited destination h at time t and tourists’ probability to share destination dis/satisfaction ( d e s t . s a t h l ( t ) ) via RS ( p r o b . R S . s h a r e l ( t ) ).• The updated RS rating tourists receive at time t is calculated with the weighted average of destination dis/satisfaction ratings by tourist l about destination h ( d e s t . s a t h l ( t ) ) , where the weightings( M l ) indicate a sharer’s relative expertise on RS and are represented by the percentage of total RS shares he or she contributed (Lee, Law, and Murphy 2011).• The post-order effect differentiation is also considered by incorporating coefficient f
Destination Evaluation	d e s t . e v a l h i ( t ) = W R S . i ( t ) * R S . m e a n h ( t ) + W S N S . i ( t ) * S N S . m e a n h i ( t ) + ( − W R S . i ( t ) − W S N S . i ( t ) ) * d e s t . e x p h i ( t ) + e + ε  (5a) d e s t . e v a l h i ( t ) = W R S . i ( t ) * ( 1 + W S N S . i ( t ) * W R S . i ( t ) 1 − W S N S . i ( t ) ) * R S . m e a n h ( t ) + ( 1 − W R S . i ( t ) − W S N S . i ( t ) ) * ( 1 + W S N S . i ( t ) * 1 − W R S . i ( t ) − W S N S . i ( t ) 1 − W S N S . i ( t ) ) * d e s t . e x p h i ( t )  (5b)	• Each destination selection is evaluated following equation 5a, where WRS.i(t) and WSNS.i(t) denote the extent of influences from two UGC sources on tourist i at time t, making the weight for previous experiences (1 – WRS.i(t) – WSNS.i(t)).•  d e s t . e x p h i ( t ) R S . m e a n h ( t ) (equation 4), S N S . m e a n h i ( t ) (equation 3) are set in the range [1,7] to echo actual attribute performance levels.• Individual tourist uniqueness is represented by context/time-invariant intercept e • External information sources such as marketing promotions and offline WoM are only accounted for with individual random error ( ε ).• Contextual variations are incorporated by temporarily adapting equation 5a to 5b. For instance, when tourist i has not received any SNS shares about destination h ( S N S . m e a n h i ( t ) = 0 ), or has chosen not to use SNS information at time t, equation 5a is temporarily transformed by reallocating WSNS.i(t) to other sources proportional to their relative weightings. The absence of RS influence can be transformed similarly.• Also, WRS.i(t) and WSNS.i(t) are updated by tourists based on the degree trip satisfaction (dest.sathi(t)) validates/invalidates aggregated destination UGC ratings ( R S . m e a n h ( t ) and S N S . m e a n h i ( t ) ) (equation 11a-b). For instance, if the valences of dest.sathi(t) and R S . m e a n h ( t ) ( S N S . m e a n h i ( t ) ) both are the same, then the more similar they are in magnitudes, the greater WRS (WSNS) will increase (with extent determined by coefficient b and the fraction portion ensuring its range of (0,1)).
Destination Selection	P h i ( t ) = e d e s t . e v a l h i ( t ) 1 + e d e s t . e v a l h i ( t )  (6)	• Tourist i chooses a destination h with the highest evaluation at a probability Phi(t), estimated with a well-established binary logistic equation (Stemerding, Oppewal, and Timmermans 1999).• At (1 – Phi(t)) when destination h is not selected, the tourist would visit the destination h2 with 2nd highest evaluation at Ph2i(t), so on and so forth.
After-visit Satisfaction Formation	dest.sathi(t) = sat.basichi(t) + sat.performancehi(t) + sat.excitementhi(t) (7) s a t . b a s i c h i ( t ) = ∑ n = 0 n 1 [ − ( a t t r i . p e r f h n − a t t r i . p r e f e r h i n ) 2 + r * ( a t t r i . p e r f h n − a t t r i . p r e f e r h i n ) ]  (8) s a t . p e r f o r m a n c e h i ( i ) = ∑ n = 0 n 2 [ r * ( a t t r i . p e r f h n − a t t r i . p r e f e r h i n ) ]  (9) s a t . e x c i t e m e n t h i ( t ) = ∑ n = 0 n 3 [ ( a t t r i . p e r f h n − a t t r i . p r e f e r h i n ) 2 + r * ( a t t r i . p e r f h n − a t t r i . p r e f e r h i n ) ]  (10)	• Based on three-factor theory that captures the asymmetric impacts of destination attributes (basic, performance, and excitement attributes), equations 7–10 are used to calculate dest.sathi(t), with the coefficient r, attribute preference levels ( a t t r i . p r e f e r h i n ), actual attribute delivery levels ( a t t r i . p e r f h n ) and number of basic / performance / excitement attributes (n1–3).
After-visit UGC reliance updates	{ W R S . i ( t ) = W R S . i ( t − 1 ) − b * Z ( | d e s t . s a t h i ( t − 1 ) − R S . m e a n h ( t − 1 ) | ) Z ( | d e s t . s a t h i ( t − 1 ) − R S . m e a n h ( t − 1 ) | ) + 1 , w h e n d e s t . s a t h i ( t − 1 ) * R S . m e a n h ( t − 1 ) < 0 W R S . i ( t ) = W R S . i ( t − 1 ) + b * ( 1 − Z ( | d e s t . s a t h i ( t − 1 ) − R S . m e a n h ( t − 1 ) | ) Z ( | d e s t . s a t h i ( t − 1 ) − R S . m e a n h ( t − 1 ) | ) + 1 ) , w h e n d e s t . s a t h i ( t − 1 ) * R S . m e a n h ( t − 1 ) > 0  (11a) { W S N S . i ( t ) = W S N S . i ( t − 1 ) − b * Z ( | d e s t . s a t h i ( t − 1 ) − S N S . m e a n h ( t − 1 ) | ) Z ( | d e s t . s a t h i ( t − 1 ) − S N S . m e a n h ( t − 1 ) | ) + 1 , w h e n d e s t . s a t h i ( t − 1 ) * S N S . m e a n h i ( t − 1 ) < 0 W S N S . i ( t ) = W S N S . i ( t − 1 ) + b * ( 1 − Z ( | d e s t . s a t h i ( t − 1 ) − S N S . m e a n h ( t − 1 ) | ) Z ( | d e s t . s a t h i ( t − 1 ) − S N S . m e a n h ( t − 1 ) | ) + 1 ) , w h e n d e s t . s a t h i ( t − 1 ) * S N S . m e a n h i ( t − 1 ) > 0  (11b)	• WRS.i(t) and WSNS.i(t) are updated by tourists based on the degree trip satisfaction (dest.sathi(t)) validates/invalidates aggregated destination UGC ratings ( R S . m e a n h ( t ) and S N S . m e a n h i ( t ) ) (equation 11a-b). For instance, if the valences of dest.sathi(t) • and R S . m e a n h ( t ) ( S N S . m e a n h i ( t ) ) both are the same, then the more similar they are in magnitude, the greater WRS (WSNS) will increase (with the extent determined by coefficient b and the fraction portion ensuring its range of (0, 1)).
After-visit travel memory updates	d e s t . e x p h i ( t ) = ( 1 − f ) * ( d e s t . e x p h i ( t − 1 ) ) + f * ( d e s t . s a t h i ( t ) )  (12)	• dest.exphi(t) employed in selecting destinations (equation 5a-b) is updated after each visit with weighted most-recent destination satisfaction (dest.sathi(t)).• f is the parameter depicting the weighting of satisfaction with the most recent visit in overall satisfaction evaluation about a destination (recent plus previous visits)
After-visit UGC shares	p r o b . R S . s h a r e h i ( t ) = p r o b . R S . s h a r e i − k 1 * Z d e s t . s a t h i ( t ) + k 2 * Z d e s t . s a t h i ( t ) 2  (13) p r o b . S N S . s h a r e h i ( t ) = p r o b . S N S . s h a r e i − k 1 * Z d e s t . s a t h i ( t ) + k 2 * Z d e s t . s a t h i ( t ) 2  (14)	• The baseline p r o b . R S . s h a r e i a n d p r o b . S N S . s h a r e i are updated into p r o b . R S . s h a r e h i ( t ) and p r o b . S N S . s h a r e h i ( t ) , respectively, corresponding to the updated trip satisfaction (dest.sathi(t) in equation 7), where more extreme dis/satisfaction increase UGC sharing probabilities. k1 and k2 are the coefficients for the linear and quadratic terms of tourist satisfaction, respectively; and Z d e s t . s a t h i ( t ) indicates standard score (Z) of dest.sathi(t).

Submodel 1: receiving SNS evaluations

The SNS connection structure of tourists is set up as a small-world network (Watts and Strogatz 1998). It is recognized as in competitive capacity to imitate the SNS structure (Kang et al. 2015), by capturing two key features of the structure, that of clustering, and only a minority of individuals (“influencers”) possessing extensive links to and imposing significant influences on others (strong links) (Luo and Zhong 2015). The key parameters depicting the SNS structure (p_R and clust.coeff) are set up following the estimated small-world network structure by Kang et al. (2015) from 12 popular SNS sites.

Whenever a tourist share his or her destination dis/satisfaction via SNS, his or her connections update their received overall SNS evaluation of the destination ( S N S . m e a n h i ( t ) ) after weighting the value of this share (W_ij), based on perceived expertise, closeness/trustworthiness, and similarity in preference with the sharer, according to Reichelt, Sievert, and Jacob (2014).

Submodel 2: retrieving RS evaluations

Considering RS as providing more centralized information than SNS, the overall RS evaluation of a destination ( R S . m e a n h ( t ) ) is assumed as accessible to every tourist (Parikh et al. 2014) and calculated as the weighted average of all RS shares on a destination.

Submodel 3: destination evaluation

The model focuses on the effects of UGC than other external information sources on destination selection. Therefore, along with the overall RS evaluation attached to destinations ( R S . m e a n h ( t ) ) and individually stored SNS evaluation of destinations ( S N S . m e a n h i ( t ) ) , only the internal information source of recalled destination experience ( d e s t . e x p h i ( t ) ) is explicitly modeled to determine individual destination evaluation ( d e s t . e v a l h i ( t ) ). It is empirically supported as possessing the most travel-decision explanatory power (Jacobsen and Munar 2012). Other information sources are captured with the error term.

Submodel 4: destination selection

The binary logistic equation used to estimate P_hi(t) at which tourists choose a destination is well-established in the choice modeling literature (Stemerding, Oppewal, and Timmermans 1999). At (1 – P_hi(t)) when destination h is not selected, the tourist would visit destination h₁ with the next highest evaluation at P_h1i(t), so on and so forth.

Submodel 5: after-visit destination satisfaction formation

The tourist satisfaction with a destination (dest.sat_hi(t)) is modeled based on the three-factor theory of tourist satisfaction (attributes categorized into basic, performance, and excitement factors) to capture asymmetric impacts of destination attributes (Matzler and Sauerwein 2002). The impacts of performance attributes are symmetric, where the variation of dest.sat_hi(t) is proportional to the discrepancy between tourist attribute preference and actual destination delivery. The impacts of basic and excitement attributes are nevertheless asymmetric: when the delivery level of a basic attribute is higher (lower) than tourist preference, dest.sat_hi(t) slightly increases (sharply decreases); the excitement attributes are the opposite, that the better-than-preference (worse-than-preference) delivery creates significant increases (slight decreases) of dest.sat_hi(t). These impacts are empirically established as best captured by a negative quadratic equation, a linear equation, and a positive quadratic equation, respectively (e.g., Finn 2012).

Submodel 6: after-visit UGC reliance updates

Considering that trip satisfaction significantly influences perceived value and consequent use of UGC (Elliot, Li, and Choi 2013), the weightings of UGC via RS (W_RS.i(t)) and SNS (W_SNS.i(t)) in destination evaluation are constantly updated based on the degree the trip satisfaction (dest.sat_hi(t)) confirms/disconfirms UGC destination ratings ( R S . m e a n h ( t ) and S N S . m e a n h i ( t ) ).

Submodel 7: after-visit travel memory updates

After each visit, a tourist updates the remembered experiences of the visited destination (dest.exp_hi(t)) by averaging with weighted satisfaction of this recent visit (dest.sat_hi(t)). If not as the first visit, the destination experience in memory should reflect a recency effect, where the dis/satisfaction with the recent trip is better remembered and weighed more in destination evaluation (Filieri, Alguezaui, and McLeay 2015).

Submodel 8: after-visit UGC shares

After a trip, each tourist shares destination dis/satisfaction at a probability via RS ( p r o b . R S . s h a r e h i ( t ) ) or/and SNS ( p r o b . S N S . s h a r e h i ( t ) ). Such UGC sharing probabilities are also updated after each trip according to the customer satisfaction and word-of-mouth model (CS-WOM) (Anderson 1998), in order to capture how UGC sharing probability evolves in an asymmetric U-shape with tourist satisfaction variations (Ranaweera and Jayawardhena 2014).

Details

UGC Behavioral Scenarios

Three UGC behavioral features potentially crucial to UGC distribution yet less explored in existing literature are manipulated and assessed for effects in the experiments, containing two using features (prob.RS/SNS.use and W_RS/SNS), and one sharing feature (prob.RS/SNS.share). The scenarios are built by setting each of three behavioral features at its extreme values (within respective acceptable ranges), then the between-scenario regional arrivals are compared to identify independent and joint effects of these features. Population average of prob.RS/SNS.use (prob.RS/SNS.share) is set to extreme values of low (10%) versus high (90%) within the possible range of (0, 1) to develop the scenarios of LruLsu, LruHsu, HruLsu, and HruHsu ² (LrsLss, LrsHss, HrsLss, and HrsHss ² ), where potential RS-SNS interaction effects (Q5) can also be examined. Similarly, the initial global W_RS (W_SNS) is set to low (0.05) versus high (0.45) extreme values in the acceptable range (0, 0.5) to construct four scenarios—LrwLsw, LrwHsw, HrwLsw, and HrwHsw ² .

Contextual and Interaction Effects

Greater tourist trust in UGC can be derived from reliable and quality offerings from destinations, implying a stronger UGC effect for better-performance destinations (Narangajavana et al. 2017). Global regions are accordingly categorized to investigate interaction effects between UGC behavioral features and regional performance indicators (Q6). The 14 global regions are categorized into four types by the medians of two performance indicators of within-region performance average (P_region) and within-region performance variability (V_region) using 2014 global destination competitiveness data (World Economic Forum 2015). As UGC behavioral effects may also vary by time (Hays, Page, and Buhalis 2013), potential interaction effects between UGC features and time-steps (10, 20, . . . , 100) are also identified (Q7).

As the perceived quality of shared UGC largely influences the importance tourists assign to received UGC in destination decision making (Yen and Tang 2019), expected interaction effects between UGC sharing and using features are thereby checked (Q4), such as between prob.RS(SNS).use and prob.RS(SNS).share. The model also accounts for region/time-based variations of these interaction effects.

Results

To examine context-dependent independent and interactive UGC behavioral effects on global visit distribution, nonparametric rank-based analysis of variance–type statistics (ATS) is adopted. As the average country arrivals for 14 regions, the simulation outputs and DV are nonnormal and violate the homogeneity of variances. ATS thus fits in testing marginal distribution equality (Brunner et al. 2002). The R package nparLD (Noguchi et al. 2012) is chosen for analyses, as it produces robust ATS tests for factorial experiments with longitudinal measures, with confidence intervals also accounted for. DV across time steps are converted into ranks, where data are stratified following between-sample factors of 64 UGC behavioral feature statuses and four performance-differentiated region types, with one repeated/within-sample factor (10/20/ . . . /100 time steps). ATS is then employed to test the mean rank differences attributed to between- and within-sample factors. Relative treatment effects (RTEs) of each factor level (with confidence intervals) and ATS statistics are generated to determine effect significance, revealing the probability for a randomly chosen observation with a smaller value than a random observation falling in the tested factor level (Lemaire, Viblanc, and Jozet-Alves 2019).

To identify how the UGC using/sharing features for each social media channel independently contribute to international arrival distribution (Q2–3), the f2.ld.f1 model of nparLD package is adopted (Noguchi et al. 2012), with each feature and region classifier (P_region/V_region) as between-sample factors and time controlled as within-sample factor. Results in Table 3 found most independent UGC behavioral effects dependent on regional performance averages (P_region) (Q6), with the exception of prob.SNS.use (prob.SNS.use × P_region: ATS = .414, p > .1). All of them do not vary by within-region performance variances (V_region) at p = .05 level. These suggest further exploration of UGC behavioral effects separately in regions varied by P_region.

OPEN IN VIEWER

Table 3.

ATS Statistics of UGC Behavioral Effects.

	UGC Behavioral Feature × Pregion × Time				UGC Behavioral Feature × Vregion × Time
	Independent	× Time	× Pregion	× Pregion × Time	Independent	× Time	× Vregion	× Vregion × Time
prob.RS.use	180.781***	455.857***	311.164***	104.646***	123.823***	362.826***	2.024	3.916*
WRS	94.972***	177.069***	113.716***	39.789***	66.628***	142.705***	0.739	0.372
prob.RS.share	4.748*	24.25***	86.106***	30.302***	3.351	19.634***	0.646	3.516*
prob.RS.use × WRS	108.862***	240.263***	150.788***	50.7***	73.133***	189.235***	1.034	1.573
prob.RS.share × WRS	33.593***	67.896***	67.666***	25.784***	23.336***	54.487***	0.547	1.502
prob.RS.use × prob.RS.share	63.143***	164.838***	141.073***	46.847***	42.731***	130.7***	1.229	4.04**
prob.SNS.use	3.896*	9.492***	.414	2.936*	2.776	7.71***	2.494	1.991
WSNS	6.082*	18.411***	9.218**	5.251**	4.329*	14.944***	0.464	1.59
prob.SNS.share	1.276	5.701**	5.054*	1.852	.908	4.631*	0.251	1.14
prob.SNS.use × WSNS	4.123**	11.606***	4.211**	3.58**	2.932*	9.414***	0.257	0.839
prob.SNS.share × WSNS	2.454	8.77***	5.492***	2.548*	1.745	7.119***	0.269	1.305
prob.SNS.use × prob.SNS.share	2.106	6.165***	3.623*	2.747**	1.497	5.004***	0.402	1.281
prob.RS.use × prob.SNS.use	61.9***	155.71***	104.812***	36.81***	42.378***	123.858***	0.903	2.106
WRS × WSNS	34.839***	66.2***	41.126**	14.542***	24.409***	53.291***	0.43	0.333
prob.RS.share × prob.SNS.share	2.014	10.133***	31.308***	11.417***	1.42	8.2***	0.3	1.59

Note: prob.RS/SNS.use = probability of using UGC via RS/SNS for destination evaluations; W_RS/SNS = weighting of UGC via RS/SNS in destination evaluations; prob.RS/SNS.share = probability of sharing UGC via RS/SNS; P_region = average country performance level within a region; V_region = performance variance within a region.

*

p ≤ .05, **p ≤ .01, ***p ≤ .001.

Q2–3 are then examined in higher- and lower-P_region regions separately, using the f1.ld.f1 model of nparLD package to estimate the between-sample factor of each UGC behavioral feature and control the within-sample time factor. Results show that among higher-P_region regions, the higher prob.RS.use, W_RS, and prob.RS.share facilitate more international arrivals (prob.RS.use: RTE_Hru – RTE_Lru = 0.03, ATS = 39.594, p < .001; W_RS: RTE_Hrw – RTE_Lrw = 0.014, ATS = 9.162, p < .01; prob.RS.share: RTE_Hrs – RTE_Lrs = 0.026, ATS = 30.857, p < .001), whereas UGC behavioral effects via SNS are insignificant (prob.SNS.use: RTE_Hsu – RTE_Lsu = −0.002, ATS = 0.126, p > .1; W_SNS: RTE_Hsw – RTE_Lsw = 0.004, ATS = 0.975, p > .1; prob.SNS.share: RTE_Hss – RTE_Lss = 0.006, ATS = 1.183, p > .1). The opposite was found for lower-P_region regions, for UGC behavioral effects via RS. The higher prob.RS.use, W_RS, and prob.RS.share all lead to fewer international arrivals (prob.RS.use: RTE_Hru – RTE_Lru = −0.114, ATS = 774.85, p < .001; W_RS: RTE_Hrw – RTE_Lrw= −0.076, ATS = 375.722, p < .001; prob.RS.share: RTE_Hrs – RTE_Lrs= −0.042, ATS = 106.387, p < .001), where negative UGC behavioral effects via SNS are further uncovered (prob.SNS.use: RTE_Hss – RTE_Lss = −0.01, ATS = 4.257, p < .05; W_SNS: RTE_Hsw – RTE_Lsw = −0.02, ATS = 20.342, p < .001; prob.SNS.share: RTE_Hss – RTE_Lss = −0.012, ATS = 7.62, p < .01). It is evident that UGC spread through neither social media channels facilitate arrivals to low-average-performance regions, per Narangajavana et al. (2017). Yet increased UGC spreading via SNS shows less-negative effects than via RS. Such region-dependent favorability of RS- versus SNS-based UGC in promoting international tourism is not yet explored in the tourism literature.

Interaction effects between UGC behavioral features (Q4) are also revealed for each social media channel across region types, using the f2.ld.f1 model, where the between-feature interaction terms and P_region/V_region are between-sample factors and the within-sample time factor is controlled. All between-feature interaction effects are found dependent on P_region (p’s < .05) rather than V_region (Q6) (Table 3). Accordingly, examining the interaction effects separately among higher- and lower-P_region regions reveal that the mutual-intensifying effect between prob.RS.use and prob.RS.share exists among higher-P_region regions (prob.RS.use × prob.RS.share: ATS = 5.84, p < .05), while the mutual-intensifying effects between prob.RS.use and W_RS, between prob.SNS.use and W_SNS, as well as between prob.SNS.use and prob.SNS.share exist among lower-P_region regions (prob.RS.use × W_RS: ATS = 90.103, p < .001; prob.SNS.use × W_SNS: ATS = 6.461, p < .05; prob.SNS.use × prob.SNS.share: ATS = 8.059, p < .01). These interaction effects confirm that the three features should be evaluated all together for each region type to accurately understand their joint effects and inform the region-based optimal UGC distributive status.

Correspondingly, the pairwise comparisons of joint UGC behavioral effects are further conducted to identify which combination of UGC behavioral features generates the most arrivals to each region, thus providing DMOs with a promising direction for UGC marketing. Function mctp (multiple contrast tests) of R package nparcomp (Konietschke et al. 2015) is adopted for simultaneous comparisons between all pairs of UGC behavioral statuses. This is achieved by controlling the familywise error rate and using simultaneous confidence intervals (SCI) plus multiplicity adjusted p-values to make competitively accurate comparisons. The approach imposes no restrictive distributional assumptions on samples and is robust to unbalanced factor designs due to unequal numbers of regions in different types. The RTEs and SCIs of all compared effects are listed in Table 4. Using Tukey contrast matrix and Fisher-approximation, simultaneous all-pairs comparisons are made for UGC effects via RS, at each level of prob.RS.use × W_RS × prob.RS.share at the p = .05 level. Results show that for higher-P_region regions, the most arrivals are generated when prob.RS.use and prob.RS.share are high while the least occurs when prob.RS.share is low (Figure 2a). It is in line with Tsao et al. (2015) that the quantity of customer reviews can potentially foster customer trust in reviews and enhance purchasing intention. With more RS shares is thus a prerequisite for motivated RS usage to enhance arrivals at higher-average-performance regions. The same pairwise comparisons for UGC effects via SNS (prob.SNS.use × W_SNS × prob.SNS.share) show no significant differences between any UGC statuses (Figure 2b).

OPEN IN VIEWER

Table 4.

The Pairwise Comparison Results of Interaction Effects between UGC Behavioral Features.

	HpHv (RTE[SCI])	HpLv (RTE[SCI])	LpHv (RTE[SCI])	LpLv (RTE[SCI])
UGC via RS
Hru.Hrw.Hrs	0.601 [0.591, 0.61]	0.621 [0.609, 0.633]	0.3 [0.292, 0.308]	0.296 [0.29, 0.303]
Hru.Hrw.Lrs	0.573 [0.564, 0.582]	0.59 [0.578, 0.601]	0.324 [0.317, 0.331]	0.35 [0.341, 0.359]
Hru.Lrw.Hrs	0.595 [0.586, 0.603]	0.645 [0.634, 0.655]	0.386 [0.379, 0.393]	0.374 [0.368, 0.380]
Hru.Lrw.Lrs	0.576 [0.568, 0.584]	0.598 [0.588, 0.607]	0.441 [0.433, 0.449]	0.435 [0.429, 0.442]
Lru.Hrw.Hrs	0.583 [0.574, 0.591]	0.62 [0.61, 0.63]	0.426 [0.418, 0.434]	0.424 [0.418, 0.431]
Lru.Hrw.Lrs	0.574 [0.566, 0.582]	0.596 [0.587, 0.606]	0.459 [0.451, 0.467]	0.404 [0.397, 0.411]
Lru.Lrw.Hrs	0.572 [0.564, 0.58]	0.602 [0.592, 0.611]	0.475 [0.467, 0.483]	0.461 [0.455, 0.468]
Lru.Lrw.Lrs	0.562 [0.555, 0.57]	0.591 [0.581, 0.6]	0.484 [0.476, 0.492]	0.462 [0.457, 0.468]
UGC via SNS
Hsu.Hsw.Hss	0.58 [0.57, 0.588]	0.616 [0.605, 0.626]	0.385 [0.377, 0.393]	0.367 [0.361, 0.374]
Hsu.Hsw.Lss	0.572 [0.564, 0.581]	0.599 [0.589, 0.609]	0.406 [0.398, 0.414]	0.405 [0.399, 0.412]
Hsu.Lsw.Hss	0.573 [0.565, 0.582]	0.609 [0.599, 0.620]	0.417 [0.409, 0.425]	0.408 [0.401, 0.415]
Hsu.Lsw.Lss	0.572 [0.563, 0.58]	0.604 [0.594, 0.614]	0.433 [0.425, 0.441]	0.414 [0.407, 0.421]
Lsu.Hsw.Hss	0.58 [0.571, 0.588]	0.616 [0.605, 0.626]	0.409 [0.401, 0.417]	0.408 [0.401, 0.415]
Lsu.Hsw.Lss	0.587 [0.579, 0.596]	0.595 [0.585, 0.605]	0.407 [0.399, 0.415]	0.408 [0.401, 0.415]
Lsu.Lsw.Hss	0.574 [0.566, 0.582]	0.609 [0.599, 0.619]	0.412 [0.404, 0.420]	0.416 [0.409, 0.423]
Lsu.Lsw.Lss	0.587 [0.579, 0.595]	0.604 [0.594, 0.614]	0.41 [0.402, 0.418]	0.419 [0.413, 0.426]

Note: RTE = relative treatment effects; SCI = simultaneous confidence intervals; Hp/Lp = higher/lower-P_region; Hv/Lv = higher/lower-V_region; Hsu/Lsu = high/low prob.SNS.use; Hsw/Lsw = high/low W_SNS; Hss/Lss = high/low prob.SNS.share.

OPEN IN VIEWER

Figure 2.

Pairwise comparisons of arrival influences from user-generated content (UGC) behavioral features via (a) RS and (b) SNS.

For lower-P_region regions, however, low prob.RS.use and W_RS, namely, the minimal reliance on RS for destination evaluations, lead to the most arrivals, while heavy RS reliance given high prob.RS.use and W_RS results in the fewest arrivals (Figure 2a). With minimal reliance on SNS, no scenarios are significantly superior to others, yet heavy SNS reliance with high prob.SNS.use and W_SNS results in the fewest arrivals (Figure 2b). All of these statistically significant between-pair differences (at p = .001 level) elucidate region-dependent UGC optimal/worst distributive status in attracting international arrivals. They also confirm that after between-feature interaction effects being considered, increased UGC distribution via RS is still preferable for higher-average-performance destinations while any form of UGC popularity only harms lower-average-performance destinations.

The between-channel (RS-SNS) interaction effects are also identified (Q5), with f2.ld.f1 analyzing the two between-sample factors of between-channel interaction term and P_region/V_region, and controlling for time. prob.RS.use × prob.SNS.use (× P_region: ATS = 104.812, p < .001) interaction effects are statistically significant and P_region dependent (Q6) (Table 3). Specifically, for Higher-P_region Higher-V_region regions, increased prob.SNS.use weakens the visit-facilitating effect of prob.RS.use (ATS = 3.262, p < .05); for Lower-P_region Higher-V_region regions, however, increased prob.SNS.use ameliorates the visit-inhibiting effect of prob.RS.use (ATS = 4.622, p < .05).

Lastly, while all these ATS statistics have within-sample over-time variance controlled, the revealed statistically significant UGC behavioral effects (i.e., independent, between-feature, and between-channel) are more or less strengthened with time elapse (× Time: p’s < .05) (Q7) (Figure 3).

OPEN IN VIEWER

Figure 3.

Demonstrative graph of user-generated content (UGC) behavioral effects with time elapses.

Discussions

The proposed three UGC behavioral features are all found as crucial in determining international arrival distribution (Q2–3). The revealed interaction effects between features, region types, and social media channels (Q4–6) can be largely explained by the nature of UGC distribution via RS versus SNS. As a relatively centralized information source, RS allows every sharer’s destination experiences to more or less influence all other RS users’ destination evaluations (Luo and Zhong 2015). With the increasing number of tourists posting experiences on RS (enhanced prob.RS.share), each destination attribute is more likely to receive a fair overall rating that approximates the actual performance. This increase is more favorable to overall higher-performance than lower-performance destinations (Q6). Such between-region discrepancies can be broadened when more tourists use RS for destination information (enhanced prob.RS.use) or designate greater weight to RS in destination selections (enhanced W_RS) (Q4). High prob.RS.use supplemented with high prob.RS.share is particularly beneficial in generating the most arrivals to higher-average-performance destinations.

SNS is a primarily localized information source where UGC distribution and influence are limited to SNS connections sharing similar interests or social relationships (Kim and Jang 2019). Most SNS influences thus occur within homogeneous SNS user groups, with limited impacts on the broader public’s destination choices (Gursoy, Del Chiappa, and Zhang 2017). Only a minority of influencers can significantly influence those beyond their direct SNS connections (Yoo, Gretzel, and Zach 2011). The SNS-based destination evaluations thus tend to be biased by this minority. Regions with lower-average-performance levels but with strength in some attributes of particular importance to influencers (more occurrences in Lower-P_region Higher-V_region regions) can accordingly enjoy some level of positive SNS exposure. Despite the overall negative RS and SNS ratings, such positive SNS exposure on some attributes can ameliorate the overall negative UGC effects to some degree. This can explain insignificant effects from SNS-based UGC distribution features on arrivals to Higher-P_region regions, considering their supposedly positive SNS outcomes weakened by the distracted arrivals to Lower-P_region Higher-V_region regions. It also explains why SNS-based UGC distributions show less-negative effects on Lower-P_region Higher-V_region regions, and why increased SNS spreading ameliorates the negative influence of greater RS spreading on Lower-P_region regions but attenuates its positive influence on Higher-P_region regions (Q5).

Conclusions and Implications

This study develops and validates an ABM model to reveal how UGC distribution via RS and SNS can influence global visit distribution (Q1). Simulated experiments further identify to what extent would the UGC behavioral features—that of probability to share UGC via RS (SNS), probability to use UGC via RS (SNS), and weighting of UGC from RS (SNS) in destination evaluations—independently and jointly form the international travel distribution. Study findings confirm both independent and interaction effects of these features (Q2–4), and the need to assess all features simultaneously. For instance, regions leading in attribute performance maximally benefit from RS if most tourists are heavy sharers and users of travel-related RS.

Region-based analyses (Q6) reveal that the greater UGC distribution via RS favors regions superior in attribute performance, while the more UGC spread via SNS aid lower-average-performance regions. The identified between-channel interaction effects of UGC using probability (Q5) further support how increased posts of SNS-based UGC are more friendly for lower-performance regions due to buffering of the negative RS effects, where weakened positive RS effects are seen on higher-performance regions. The revealed time-based UGC effects (Q7) suggest that while DMOs develop UGC management strategies based on most current UGC patterns, intensifying the effects of UGC behavioral features over time should be considered.

Theoretical and Methodological Implications

This study helps identify potential mechanisms behind UGC influences on macro-level destination visit distribution, which has remained a black box in the literature (Chen and Law 2016). It also contributes with a holistic examination of UGC effects, via (1) integrating various theories/propositions to identify the bottom-up UGC influence mechanism (e.g., innovation diffusion theory (Bass 1969), three-factor theory (Matzler and Sauerwein 2002), CS-WOM model (Anderson 1998), and the eWOM credibility framework (Reichelt, Sievert, and Jacob 2014)), (2) accounting for effect variations by context of distribution channel (RS/SNS) and region performance level, (3) adopting a comprehensive system of destination attributes that allows global cross-destination benchmarking, (4) considering both active and passive/inactive UGC users in UGC effect evaluation, and (5) is the first attempt to systematically explore UGC effects on destination visits in global scope and longitudinal lens.

The developed ABM model further strives to fully account for the intricacies and nuances of UGC distribution and influences by maximally embracing (1) between-individual heterogeneity in preferences/behaviors, (2) individual probabilistic decision-making across information searching, destination visit, and UGC sharing, (3) realistic individual decision rules, for example, the weighted evaluation of destination attributes, (4) the effects of unconscious SNS exposure as compared to the intentional RS search, and (5) possibility of repeated and multidestination visits in destination choices. Consequently, a competitive predictive accuracy can be achieved within the computational power limit.

The study also supplements gaps in social media literature by addressing the differentiated and interaction effects of social media channels on tourist choices, particularly after controlling for longitudinal and regional variances. The accuracy of SNS effect evaluations is also improved by incorporating powerful influences from constant passive exposure to SNS.

Furthermore, the integration of multiple theories into the ABM model promises theoretical advances beyond the expanded explanatory power for UGC mechanisms. Innovation diffusion theory (IDT) (Bass 1969), as the foundation of the developed model, is proven by the validated model of its applicability to streamlining the international travel process under UGC influences. This study also strengthens its external validity by filling two gaps in its existing applications for accurately capturing the real-world innovation diffusion process (Kiesling et al. 2012). First, more sophisticated rules are added in determining the extent of social influences on innovation adoptions, such as making innovation-adoption decisions (i.e., choosing a new destination) based on the feedback from adopters (i.e., previous visitors) rather than merely based on the number of adopters; second, by allowing an innovation (i.e., a new destination) to be evaluated by multiple attributes and jointly with its competitors (i.e., other destinations). The explanatory power of the original theory is also expanded after adapting to the unique tourism settings, where the repeated adoptions after the initial adoption (i.e., repeated visits) and adopting multiple innovations at once (i.e., multidestination trip) are also investigated.

This study also introduces three-factor theory of tourist satisfaction (Matzler and Sauerwein 2002), which can be an add-value toolkit for attribute-level UGC influence analyses and guiding the strategic UGC management. Recognizing the differential impacts from various destination attributes on tourist satisfaction and UGC responses, the integrated model can potentially help identify those attributes producing the most UGC influences on a target market, which should be accordingly prioritized for management. The current model only randomly categorizes attributes as basic/performance/excitement factors for each tourist, yet the proposed function can be realized when modeling the UGC dynamics of a specific tourist segment with known attribute preference, or even better when the distribution of tourist attribute preference across global regions/countries becomes available.

The CS-WOM model (Anderson 1998) adds realism to the UGC-influence model by capturing the U-shape relationship between tourist satisfaction level and UGC sharing probability. Moreover, its integration allows the assessment (merely via simulations) of various possible tourist-satisfaction management strategies, for example, types of add-value services to be offered or ways to recover from service failure, to select the most effective strategy fostering positive UGC shares and ultimately increasing the overall UGC rating and international visits.

The supported eWOM credibility framework (Reichelt, Sievert, and Jacob 2014) by the validated model highlights the need to incorporate the non-utilitarian criteria in tourist choice modeling. It shows that in SNS-based UGC evaluation, tourists may not rely much on the expert recommendations that better capture the objectively superior-performing destinations. Rather, they value the most destinations recommended by connections close in relationship or similar in interest/value, perhaps because of the gained sense of social belongingness and approval. Whether such tendency stands in RS credibility evaluation is valuable for further explorations. While the current model only employs travel expertise to approximate the credibility weighting of RS shares, future models may add similarity into calculating the proxy and examine in-depth the between-channel UGC-evaluative in/consistency.

The validated model also supports the reliability of certain travel and UGC behavioral patterns of tourists, by demonstrating consistent findings from the current global scope model with existing local/regional empirical evidence. For instance, tourists primarily share UGC with SNS but plan trips using UGC from RS; a considerable portion of tourists are repeat visitors; and the more critical determinant of destination choices than UGC is previous destination experiences, particularly the recalled most recent experiences. These confirmed patterns set a solid reference point for future global scope modeling of tourist behaviors.

Methodologically, it demonstrates ABM as a valuable tool for evaluating UGC/social media effects, as its simulated experiments can not only explore the causal effects of individual UGC/social media behaviors on macro-level visit distribution, but also account for complex contextual and temporal factors to enhance its explanatory power and real-life generalizability. The model is built on assumptions and parameters generated from a series of analytical models to maximally strengthen its restrictiveness (Rand and Rust 2011). The rigorous procedure this study demonstrates for ABM development, validation, and reporting, as well as the simulated experiments design and analyses, can be adapted by tourism scholars for broader applications. It offers promising directions for future UGC/social media studies to simultaneously achieve restrictiveness and realism.

Practical Implications

This study sheds light on the potential for DMOs to optimally capitalize on UGC in generating arrivals, by intentionally promoting the UGC behavioral patterns among targeted tourist markets that match with their destinations’ performance levels. For instance, promoting more RS-based UGC adoptions and shares, for example, using deals or other incentives to solicit more RS shares and adoptions, is optimal for both regions with countries homogeneously high in performance (i.e., Western and Northern Europe/Oceania) and regions high in both country performance average and variance (i.e., East Asia/East Europe/Southern Europe/North America including Mexico). For the latter regions, DMOs should also caution about the conflicting effects between RS and SNS user portion increases and consider distracting market from SNS communication to reduce SNS influences.

Yet for regions lower in average country performance, UGC distribution may not be as effective or even harmful as a promotion approach. These include regions with countries homogeneously lower in performance (i.e., Sub-Saharan Africa/North Africa/South Asia/South and Central America/the Caribbean), and regions with low country performance average but high country/attribute variance (i.e., Southeast Asia/Middle East). With UGC’s inevitable popularity, encouraging market reliance on SNS- than RS-based UGC is less harmful for these regions, particularly the low-average high-variance regions, given SNS ameliorating the negative RS effects. As such, they can provide incentives for UGC shares/uses via SNS or strategies to improve perceived credibility of SNS-based UGC (e.g., by providing DMO feedback to SNS posts) (Sparks, So, and Bradley 2016).

By simply manipulating parameter values and simulating, global DMOs can utilize the validated ABM model (with the codes available on reasonable requests) to generate rough extent forecasts and visual demonstration of global regions/countries’ trends in UGC rating and international visit based on different contexts. For example, by changing the value(s) of prob.RS/SNS.share/use to reflect the future possible UGC using/sharing patterns of targeted markets, a national DMO can predict the approximate extent to which these potential UGC trends could facilitate/harm the country’s short-/long-term arrivals. Experimenting with different values of prob.RS/SNS.share/use is also helpful for DMOs to identify which market segment (i.e., with what frequency of RS/SNS sharing/using) best fits the country’s attribute-performance levels and can produce optimal UGC and visit outcomes.

Besides manipulating UGC behavioral features, DMOs can also manipulate other parameters to inform context-based UGC marketing strategies. For instance, when data become available, tourists’ three-factor designation of destination attributes can be updated (e.g., enhancing the portion of tourists who designate safety/security as a performance than basic factor to reflect its growing importance to tourists in light of COVID-19 influences), and so can the destinations’ attribute-performance levels (e.g., updated performance scores to reflect the destinations’ progress recovering from COVID-19 influences). DMOs can then conduct before-update and after-update simulations to generate the visit trend lines across various UGC distribution statuses and visually identify any significant pattern changes due to the updates. This helps them understand to what extent the market preference or destination performance may influence UGC impacts and thus how to adapt UGC marketing strategies accordingly. The ABM model can further serve as a prototype that can be adapted to guide UGC marketing of destinations with micro-level geographical scopes, for example, how UGC distribution could affect arrivals to different states/provinces/counties/cities.

To summarize, this model provides destinations with a UGC effect assessment tool to facilitate its co-productive than co-destructive role (Lund, Scarles, and Cohen 2020). DMOs can achieve this end with this model by merely experimenting with parameter value variations, simulating, and compare outcome differences to identify effects, yet if their purpose is to produce accurate quantitative forecasts of future trends in visits and UGC, the model needs to be thoroughly adapted to the specific context with additional validations and calibrations before implementation. More information about how to employ ABM in real-life practices can be found in Netlogo official website(https://ccl.northwestern.edu/netlogo/docs/), which is nevertheless beyond the scope of the current study.