Habit formation or word of mouth: What does lagged dependent variable in tourism demand models imply?

Abstract

The effects of habit formation/persistence (HFP) and word of mouth (WOM) each play a critical role in influencing tourists’ decisions regarding whether to visit tourism destinations and therefore tourism policies and tourism management resource allocations. Nevertheless, in previous tourism demand studies, the two effects have been represented by the same time-lagged dependent variable, which makes the variable have an ambiguous meaning and biases the empirical results. The purpose of this study is to solve the ambiguity of a lagged dependent variable in tourism demand. We used economic theories regarding internal habits and external habits to clarify the meanings of HFP and WOM and revised the tourism demand model into a spatial dynamic panel model (SDPM). The empirical results suggested that an SDPM is a more accurate model for modeling tourism demand. The effects of variables in an SDPM are more consistent with theoretical expectations.

Keywords

external habit habit formation habit persistence internal habit negative word-of-mouth effect spatial dynamic panel model

Introduction

The lagged dependent variable, an autoregressive term of tourism, has been commonly included in the modeling of tourism demand. In empirical studies, regardless of whether the estimated coefficients on the lagged dependent variable are positive (Balli et al., 2015; Dogru et al., 2017; Song et al., 2003) or negative (Becken and Carmignani, 2016; Naudé and Saayman, 2005), they are highly significant. Some studies even report the lagged dependent variable as the most important determinant of tourism demand (Song and Witt, 2003; Song et al., 2003, 2010). If such an important variable is omitted from tourism demand analysis, the conclusions about the variables included in the tourism demand model may be incorrect. Even when the lagged dependent variable is included in the model, if it has an ambiguous meaning, conclusions may also be biased by this ambiguity. Nevertheless, the ambiguity is apparent in the literature. The purpose of this study is to solve the ambiguity of a lagged dependent variable in tourism demand.

There are two types of distinguishable explanations for the lagged dependent variable. Some authors maintain that the lagged dependent variable should be interpreted as habit formation (Assadzadeh et al., 2014) or as the strength of habit persistence in travel preferences (Peng et al., 2015). Other authors assert that the lagged dependent variable is to capture the word-of-mouth (WOM) effect on tourism demand. Many others do not distinguish the meaning and use in their papers (Dogru et al., 2017; Garín-Mun, 2006; Song et al., 2010; Witt and Witt, 1995). The same word with different meanings causes ambiguity. Even worse, the two different meanings are in direct contradiction to each other.

The principle underlying habit formation/persistence (HFP) is that consumers’ current utility depends not only on their current consumption but also on a habit stock formed by their own past consumption. Dynan (2000) developed a model to show how habit formation creates a positive link between consumption in two consecutive periods and how habit formation makes consumers gradually adjust their consumption to economic shock. Lally and Gardner (2013) suggested using tangible and randomly scheduled rewards to reduce behavioral complexity and to provide salient cues to promote habit formation and maintain habit persistence.

The principle underlying WOM is that consumers’ current utility does not depend on their own past consumption but rather on other consumers’ past consumption or information. Contrary to habit formation’s positive link between consumption in two consecutive periods, the effects of WOM on consumption can be either positive or negative (Mahajan et al., 1984; Pfeffer et al., 2014). The physical evidence of the negative effects of WOM on demand can be observed from empirical studies. Mahajan et al. (1984: 1389) surmised that “unfavorable information about products tends to carry greater weight with prospective buyers than favorable information.” Chevalier and Mayzlin (2006: 346) echoed that “an incremental negative review is more powerful in decreasing book sales than an incremental positive review is in increasing sales.” The negative WOM can even cause information avoidance and “reduces the (consumer’s) desire to obtain the information” (Golman et al., 2017: 108).

Marketing strategies through WOM are different from those through habit formation and are very different depending on the type of WOM. Marketing strategies for cases with positive WOM include entertainment, useful information, self-concept of relevant things, high-status goods, unique things, common ground, emotional valence, incidental arousal, and accessibility (Berger, 2014). For the cases with negative WOM, the managers should reduce the negative feelings of unsatisfied consumers and “induce in them a sense of forgiveness” (Ghosh, 2017: 164). When WOM is sparse, “targeting of advertising at the best connected areas of the network is generically not an optimal strategy” (Campbell, 2013: 2496).

The definitions of HFP and WOM are different. The effects of these factors on tourism demand are dissimilar. HFP creates a positive link between consumption in the two consecutive periods; however, the link created by WOM could be either positive or negative. Therefore, the management implications under different cases are quite distinct from each other. Thus, the ambiguity raises problems in explaining the meaning of the estimated coefficient and in implementing marketing strategies. Although the ambiguity is obvious, in empirical tourism demand studies, the lagged dependent variable is still riddled with ambiguity due to the indiscernibility between HFP and WOM. Witt and Martin (1987) asserted that ambiguity exists because economic theory does not suggest a clear presence of a lagged dependent variable.

As noted, the recent economic literature has made great progress in solving this ambiguity. HFP has been applied in the study of asset returns and the business cycle (Boldrin et al., 2001), persistence of monetary shocks (Bouakez et al., 2005), endogenous growth (Augeraud-Veron and Bambi, 2015), and aggregate consumption growth (Everaert and Pozzi, 2014). Meanwhile, WOM has been applied in studying stock returns (Tallarini and Zhang, 2005), stock market participation (Brown et al., 2008), asset pricing (Gârleanu and Panageas, 2015), the business cycle (Khorunzhina, 2015), and consumption commitments (Chetty and Szeidl, 2016). Some authors integrate HFP and WOM together (Grishchenko, 2010; Korniotis, 2010; Kraft et al., 2017). In the mentioned literature, HFP depends on consumers’ own history of consumption, while WOM depends on the consumption choices of peers. To clarify, in the abovementioned literature, HFP is replaced by “internal habit,” WOM is replaced by “external habit,” and each has its own symbol instead of integrating the two different meanings into one variable, as has been done in extant tourism literature.

The empirical study of internal and external habits has been carried out by many authors (Ciccarelli and Elhorst, 2017; Grishchenko, 2010; Korniotis, 2010; Verhelst and Van den Poel, 2014). These authors use aggregate data to estimate each coefficient of “internal habit” (as HFP used in tourism literature) and “external habit” (as WOM used in tourism literature). For example, Ciccarelli and Elhorst (2017) used cigarette consumption at the provincial level in Italy, and Korniotis (2010) used aggregate state sales in the United States. They defined internal habit as the consumers’ consumption habits influenced by their own past consumption, represented internal habit by Y_i _{, t−1}, a time-lagged dependent variable, defined external habit as the consumers’ consumption habits influenced by the consumption decisions of other consumers, and represented external habit by $\sum_{j = 1}^{N} ω_{i j} Y_{j, t - 1}$ , a spatially lagged dependent variable. Their methods solved the abovementioned ambiguity problem in tourism demand.

Tourism demand models are crucial for generating accurate forecasts of future tourism demand, which is the foundation for tourism planning and tourism policy formulation. We noticed that the tourism demand models used in the tourism literature suffered from ambiguity, and we believe that the ambiguity problem needs to be solved to prove the accuracy of tourism demand. We observed that HFP and WOM were addressed separate from each other in Korniotis (2010), and thus, we believe that such a methodology can fix the long-existing ambiguity problem. Therefore, we follow Korniotis’ (2010) methodology to test “internal habit” and “external habit” in the tourism demand model.

Theory and literature review

HFP and WOM have been incorporated into economic models, and this integration has contributed some new economic theories. Brown (1952) has used a positive autoregressive component to represent habit formation in a traditional demand model. Some important subsequent contributions of HFP can be seen in Friedman’s (2008) concept of permanent income and the endogeneity of preferences (Becker and Murphy, 1988). Meanwhile, Veblen (1953) develops a theory of conspicuous consumption, the individuals’ consumption activities imitated by their peers. The theory of conspicuous consumption contributes to the theory of interdependence in consumption and the concept of network externalities (Katz and Shapiro, 1985) in which individual preferences are dependent on the consumption activities of others. Preference is driven by the desire for distinction and peer reference (Bourdieu, 1984).

Becker (1996), a Nobel laureate, used an extended utility function to provide rigorous and precise models to determine the specific effects of HFP and WOM on consumption. In the function, the consumer’s utility over goods x is conditional on two stocks, personal consumption capital and social consumption capital:

U = f (x | P, S)

where x is the consumption goods, P is personal capital, and S is social capital. Habit formation occurs through P and peer reference occurs through S. Becker offers an individual theory of consumption that individual choices should be conditioned by personal capital and social capital. Personal capital examines how individuals’ tastes are influenced by their own past behavior; for example, experience in a particular consumption activity can be expected to increase the marginal utility of the time spent in that activity. Social capital examines how the tastes of different individuals are interdependent; for example, “men and women want respect, recognition, prestige, acceptance and power from their family, friends, peers and others” (p. 12), and hence, consumption behavior is driven “with an eye to pleasing peers and others in their social network” (p. 12). In addition, in his model, Becker notes that an increase in social capital can raise or reduce utility, depending on the consumer.

Becker’s frameworks of habit formation involve preferences that account for the formation of personal and social capital and provide rigorous and precise definitions of both types of capital. In the case of personal capital, consumers’ own history influences their current tastes and decisions. In the social capital case, the history of the society or of the social group to which the consumers belong influences their tastes. Economic empirical studies analyzing the effects of personal capital and social capital on consumption have followed the above definitions (Augeraud-Veron and Bambi, 2015; Boldrin et al., 2001; Chetty and Szeidl, 2016; Grishchenko, 2010; Korniotis, 2010; Kraft et al., 2017). In recent studies, personal capital and social capital have usually been termed as “internal habit” and “external habit”, respectively, and each has its own notation, Y_i _{, t−1} and $\sum_{j = 1}^{N} ω_{i j} Y_{j, t - 1}$ , respectively (Grishchenko, 2010; Korniotis, 2010; Kraft et al., 2017).

Methodology

Spatial econometric model

The theories discussed in the second section provide clear definitions of “internal habit” and “external habit,” and the empirical studies following this theory assign “internal habit” and “external habit” each with a distinct variable. These methods separate “internal habit” (habit formation) from “external habit” (WOM) and can solve the ambiguity problem of tourism demand. Therefore, we use this methodology to test “internal habit” and “external habit” in the tourism demand model.

We introduce a method separating “internal habit” (habit formation) from “external habit” (WOM) for estimating the tourism demand model. From the tourists’ perspective, “internal habit” indicates that tourists’ current travel are influenced by their own previous travel, while “external habit” indicates that tourists’ current travel are influenced by other people’s travel. In the model, a time-lagged dependent variable, Y_i _{, t−1}, is used to present the effect of habit formation, and a spatially lagged dependent variable, $\sum_{j = 1}^{N} ω_{i j} Y_{j, t - 1}$ , is used to present the effect of WOM. The model is written as

Y_{i t} = τ Y_{i, t - 1} + ρ \sum_{j = 1}^{N} ω_{i j} Y_{j, t - 1} + X_{i j} λ + ε_{i t}, i = [1, \dots, N], t = [1, \dots, T]

where Y_it is the dependent variable for cross-sectional unit i at time t. The vector X_ij is a vector of the exogenous variables of tourism demand. The spatial weight variable, ω_ij , captures the spatial relationship of unit i and unit j. The spatial autoregressive coefficient, ρ, measures the spatial influence of Y_j _{, t −1} on Y_it .

The model in equation (2) contains both time-lagged and spatially lagged dependent variables; therefore, it is a spatial dynamic panel model (SDPM), featuring spatial and/or time serial dependence between the observations on each spatial unit over time, unobservable spatial and/or time period-specific effects, and endogeneity of the regressors. The commonly used tourism demand model is expressed as equation (2) without the spatial lag variable. Because the later tourism demand model considers time-lagged dependent variables but not the spatial effects of variables, the model is a dynamic panel model (DPM).

The estimation for the SDPM/DPM must address these potentially concurrent problems. In the literature, estimation methods have been developed based on either a bias-corrected maximum likelihood (ML) or a quasi-maximum likelihood (QML) estimator, instrumental variables or generalized method of moments (IV/GMM), or Bayesian Markov chain Monte Carlo (MCMC; Elhorst, 2012). Nevertheless, the parameters estimated using these methods are still considerably biased, for example, ML in Elhorst (2005), IV/GMM in Elhorst (2010), and MCMC in Parent and LeSage (2011).

An alternative estimation method for the SDPM/DPM is system GMM (Blundell and Bond, 1998). System GMM utilizes lagged first differences for the equation in levels and thus has less finite sample bias. Furthermore, the system GMM estimator has been shown to be more efficient. Kukenova and Monteiro (2009) use system GMM, MLE, QMLE, and LSDV to estimate a dynamic spatial panel model and find that the results from system GMM are more efficient than those from maximum likelihood estimator (MLE), quasi-maximum likelihood estimator (QMLE), least squares dummy variable (LSDV). Therefore, due to better efficiency in estimation, the system GMM estimator became one of the most widely used estimators in the empirical analysis of the SDPM/DPM (Bouayad-Agha and Védrine, 2010; Hayakawa, 2015).

However, the system GMM estimators are valid under one strong assumption. Because the system GMM estimates the two equations of differences and levels simultaneously, the consistency of the system GMM estimators relies on there being no first-order serial autocorrelation in the errors of the level equation and the instruments being truly exogenous and, therefore, valid to define the moment conditions. The null hypothesis of no first-order serial correlation can be tested by the Arellano and Bond (1991) test, which tests the hypothesis that there is no second-order serial correlation in the first-differenced residuals and in turn implies that the errors from the level equation are serially uncorrelated. The validity of the instruments can be tested by Sargan statistics. If the null hypotheses are rejected at the usual significance levels, the system GMM estimators are not appropriate for analyzing the data.

The other issue concerning the use of system GMM is that we have to decide whether to use one-step or two-step estimation procedures. The one-step system GMM is estimated based on an initial weight matrix, and no updating of the weight matrix is performed except when calculating the appropriate variance–covariance matrix (VCM). The two-step system GMM uses the VCM estimated by one-step estimation as a weighting matrix to obtain a more consistent estimate of VCM and then reruns the estimator; therefore, the covariance matrix is robust to heteroskedasticity and autocorrelation, and the estimations are more consistent in the presence of panel heteroskedasticity and autocorrelation (Mileva, 2007). The problem with the two-step GMM estimator, however, is that the standard errors it produces are biased downward, specifically in the dynamic panel data setting (Arellano and Bond, 1991; Blundell and Bond, 1998). Judson and Owen (1999) showed in Monte Carlo simulations that when T increases, the one-step GMM estimators perform better than the two-step GMM estimators, and the one-step GMM estimators with robust standard errors are preferred (Arellano and Bond, 1991; Blundell and Bond, 1998). Because we have a long panel, a long time dimension (large T), and few cross-sections (small N), we therefore use the one-step system GMM estimators with robust standard errors in our estimations. In robust one-step system GMM estimators, the Sargan test statistic is inconsistent (Roodman, 2009); therefore, we do not report the Sargan test statistic.

Data source and descriptive statistics

Before estimating equation (2), the factors of the exogenous variables of tourism demand should be determined. The literature on the exogenous variables of tourism demand has been prolific, and the variables used are diverse. The meta-analysis method has been used to systematically justify the importance of diverse exogenous variables from the bulk of research. Lim (1997) performs a meta-analysis review on 100 studies and finds that the most used independent variables are income and transportation costs. The significant effects of these two factors on tourism demand are revealed again in Lim (1999). Crouch (1994) found that income is the most important and statistically significant variable of tourism demand and showed the important influence of the cost of transportation on tourism demand. In a recent article, the importance of population size on modeling tourism demand is also identified (Martins et al., 2017). Some one-off events, such as severe acute respiratory syndrome (SARS) and financial crisis, have significant adverse effects on tourism demand in the affected countries (Song and Li, 2008), as was the case in Taiwan. According to the literature review, income, population, transportation costs, SARS, and financial crisis were selected as exogenous variables for equation (2). The spatial weight was created using queen contiguity-based weights, defining neighbors as any counties that share a border or vertex.

Here, we explain the definitions and sources of the dependent and exogenous variables used to estimate the tourism demand model for Taiwan. We use monthly data of domestic tourism of 19 cities/counties (Table 1). Cities/counties are relevant for the purpose of this analysis, as they generally have a specific institutional framework for regional and tourism policies. The monthly tourist data are collected from the Tourism Statistics Database managed by Tourism Bureau, Taiwan, and have been deseasonalized. The visitor figures are the number of visitors to the scenic spots in Taiwan and are not the volume of business travel. The top 20 popular tourist destinations on the Island are listed in Table 2. The income variable enters the equation as the per capita average monthly salary (in NTD) in real terms in Taiwan. The transportation cost variable, a factor representing travel prices in tourism demand function (Becken, 2011; Ritchie et al., 2010), is measured by the monthly gasoline price index (Gozgor and Ongan, 2017; Poudyal et al., 2013). Data on these two variables, population size, and number of tourists were obtained from the Macro Database maintained by the Directorate-General of Budget, Accounting and Statistics of Executive Yuan, Taiwan. In addition, income, population, tourists, and gasoline index are presented in logarithmic form in the empirical models. Exogenous shock and crisis event, the SARS outbreaks in 2003, and the global financial crisis in 2008–2009 are modeled by dummy variables, which take 1 for the period the shock/event occurred and 0 the otherwise. The two dummies are calculated by the author. The data used are domestic monthly data of 19 cities/counties collected from January 2001 to July 2016. There are 3553 total observations. The descriptive statistics are presented in Table 3.

Table 1.

Descriptive statistics of tourists in the 19 cities/counties.

City/county	Total	Monthly average	Std
Taipei City	457,307,236	2,598,337	1,356,227
New Taipei City	381,233,614	2,166,100	1,028,139
Taoyuan City	72,993,985	414,739	224,043
Hsinchu County	139,298,195	791,467	275,902
Hsinchu City	12,528,413	71,184	96,191
Keelung City	12,582,010	71,489	43,156
Yilan County	83,918,211	476,808	223,906
Miaoli County	31,172,855	177,118	119,763
Taichung City	242,341,582	1,376,941	1,194,987
Changhua County	93,647,101	532086	451,406
Nantou County	176,925,061	1,005,256	441,286
Yunlin County	113,500,008	644,886	593,867
Chiayi County	70,316,439	399,525	864,787
Chiayi City	13,946,670	79,242	43,984
Tainan City	291,077,580	1,653,850	493,588
Kaohsiung City	259,358,132	1,473,626	1,062,580
Pingtung County	101,946,800	579,243	202,134
Hualien County	132,380,517	752,162	319,727
Taitung County	62,110,121	352,898	130,735

Table 2.

The top 20 scenic spots of Taiwan in 2016.

Scenic spots	County	Number of visitors
Dongfong Green Bikeway and Houfeng Bikeway	Taichung City	10,775,803
Daitianfu Temple, Nankunshen	Tainan City	9,773,500
Calligraphy Greenway	Taichung City	8,485,847
Foguangshan	Kaohsiung City	7,747,587
National Chiang Kai-shek Memorial Hall	Taipei City	6,953,671
Lion’s Head Mountain Scenic Area	Hsinchu County	6393095
National Dr. Sun Yat-Sen Memorial Hall	Taipei City	6,246,672
BeiGang ChaoTian Temple	Yunlin County	6,081,000
Mt. Bagua Scenic Area	Changhua County	5,086,839
Cijin Seaside	Kaohsiung City	5,053,000
Lotus Pond	Kaohsiung City	4,822,531
Taichung Park	Taichung City	4,698,811
National Palace Museum	Taipei City	4,665,725
Golden Coast	New Taipei City	4,553,901
Pier-2 Art Center	Kaohsiung City	4,372,019
Rueifan Special Scenic Area	New Taipei City	4,322,675
Daitianfu Temple, Madou	Tainan City	4,073,805
Shihfen Sightseeing Service Center	New Taipei City	3,882,540
The Sunmoonlake Scenic Area	Nantou County	3,727,828

Table 3.

Summary statistics of variables.

Variable	Obs.	Mean	SD	Min	Max
Tourists	3553	825,479	941,564	0	10,300,000
Population	3553	23,000,000	345,815	22,300,000	23,500,000
Income	3553	44330.030	9851.911	37,211	88,285
Gasoline index	3,553	84.220	16.878	56.42	112.15

Empirical findings

To estimate the parameters of tourism demand in the form of the DPM or SDPM using panel data, stationary variables should be used to avoid the problem of spurious correlation. To avoid meaningless results, before performing the empirical DPM or SDPM analysis, we conducted panel unit root tests to examine the stationarity of the variables used in the study. For panel data, the Levin–Lin–Chu test (LLC) and the Im–Pesaran–Shin test (IPS) are proposed in the literature. The LLC relies on the homogeneity assumption in the dynamics of the autoregressive coefficients for all panel units, which indicates that the root must be the same for all series under alternative hypothesis, while the IPS advocates a unit root test corrected for heterogeneous panels and relaxes the restrictive homogeneity assumption of LLC. The alternative hypothesis of the LLC is that all panels are stationary, while the alternative hypothesis of the IPS is that some panels are stationary. Although the IPS is a more flexible panel unit root test, it is common to see both LLC and IPS in the literature as well as the LLC used for a pooled panel unit root test and the IPS for a heterogeneous panel unit root test. Therefore, for the present analysis, we selected both LLC and IPS unit root tests.

The null hypothesis of LLC and IPS is that a unit root exists in the series, that is, the variables are nonstationary. Table 4 summarizes the empirical results of LLC and IPS for a unit root in each of the listed variables. The results show that all variables were significant at the usual testing levels. We obtained overwhelming evidence against the null hypothesis of a unit root and, therefore, concluded that all variables were stationary. The results are in accordance with Chang et al. (2007).

Table 4.

Panel unit root tests.

	IPS	LLC
lntour	−49.986***	−58.965***
lnpop	−20.990***	−21.746***
lnincome	−49.005***	−63.323***
lngas	−29.825***	−30.964***

Note: IPS: Im–Pesaran–Shin test; LLC: Levin–Lin–Chu test.

***Significance at the 1% level.

The one-step system GMM with robust standard errors can be used to estimate the SDPM/DPM if there is no first-order serial correlation in the residuals, which can be tested by the Arellano and Bond (1991) test. The Arellano and Bond test for serial autocorrelation tests the hypothesis that there is no second-order serial correlation in the first-differenced residuals, which in turn implies that the errors from the equation are serially uncorrelated. The optimal lag length 1 is chosen at the minimum AIC from lag 1 period to lag 12 periods. The null hypothesis of the Arellano and Bond test is that there is no serial correlation in the residuals. Table 5 lists that the Arellano–Bond AR(2) test for both SDPM and DPM were insignificant at the usual testing levels, and we did not reject the null hypothesis of no serial autocorrelation. We verified that the assumption of one-step system GMM is satisfied and that the estimated one-step system GMM is feasible.

Table 5.

Dynamic panel estimates.

	DPM	SDPM
laglntourist	0.934*** (0.031)	0.942*** (0.022)
Wlaglntourist		−0.288** (0.109)
lnpop	−0.604 (0.889)	7.957** (3.196)
lnincome	0.687*** (0.155)	0.562*** (0.148)
Gasoline_index	−0.000 (0.000)	−0.000 (0.000)
SARS	−0.125** (0.061)	−0.111* (0.063)
Financial Crisis	0.006 (0.013)	−0.047** (0.023)
constant	3.732 (13.319)	−135.947** (51.924)
Number of obs.	3515	3515
Wald χ ²	6411.02	9438.76
Arellano–Bond AR(1) test	−1.9732**	−1.8552*
Arellano–Bond AR(2) test	1.0137	0.74444

Note: DPM for tourism demand model using time lag variable to present both HFP (internal habit) and word-of-mouth effect (external habit); SDPM for tourism demand model using time lag variable to present HFP (internal habit) and spatial lag variable to present word-of-mouth effect (external habit). Standard errors are represented in parentheses. HFP: habit formation/persistence; DPM: dynamic panel model; SDPM: spatial dynamic panel model.

*Significance at the 10% level; **significance at the 5% level; ***significance at the 1% level, and other notations are defined earlier.

The empirical results from the one-step system GMM estimation of tourism demand in the form of DPM and SDPM are listed in Table 5. According to the estimators, the coefficients of the lagged dependent variables were all significant. The result is that the highly significant positive coefficients of the time-lagged dependent variables in both models are consistent with the findings of other researchers (Balli et al., 2015; Dogru et al., 2017; Song et al., 2003). The estimated coefficient of Wlaglntourist is −0.288, which is significant at the 0.05 level of significance. The result indicates that Taiwan tourism industry suffered negative WOM problems. A 1% increase in tourist arrivals in the neighboring counties suggests that target county is losing 0.288% tourists. According to Golman et al. (2017), the problems are caused by that the tourists to the neighboring counties “reduces (their) desire to obtain the information” (Golman et al., 2017: 108) about the target county. This kind of information avoidance is constructed from the similarity among tourism products in the target county and neighboring counties and imposes a negative externality on others.

Our estimated income elasticity of tourism demand is inelastic, which indicates that domestic tourism in Taiwan is a normal (necessity) good. Our result is in contrast to the belief that tourism is a luxury good (Crouch, 1994). Nevertheless, according to Gunter and Smeral (2016), the income elasticities decreased from period to period. In particular, for 2004–2013, the values of the income elasticities are lower than 1. Furthermore, the values of the income elasticities are lower for the recent research studies conducted in Asian countries (Peng et al., 2015). Therefore, our result is in accordance with the current literature.

The estimated coefficient of the spatially lagged dependent variable in the SDPM was significantly negative, which raised a few issues. First, the effect of external habits on tourism demand in Taiwan is negative, which was not found or mentioned in previous studies. Second, the estimated results of tourism demand models for Taiwan are biased because the spatial lag dependence, an important variable, was not included. Third, the ratio of tourism expenditure to income decreases with income growth because negative WOM have strong effects on decreasing willingness-to-pay for a necessity good (Laamanen, 2013).

The phenomenon of negative external habit (WOM) is less addressed in tourism demand research. The ignorance of negative WOM can lead to inappropriate policies aiming to promote tourism. For example, tourist information is regarded as an important influence on tourists’ choice of destination, and tourist brochures or media materials providing tourism information are recommended for promoting tourism (Molina and Esteban, 2006), but Golman et al. (2017) highlight the existence of information avoidance caused by negative WOM, which reduces tourists’ desire to obtain tourism information. Furthermore, negative WOM is more powerful than positive WOM (Chevalier and Mayzlin, 2006; Mahajan et al., 1984). For cases with negative WOM, common tourism marketing strategies such as entertainment, useful information, self-concept of relevant things, high-status goods, unique things, common ground, emotional valence, incidental arousal, and accessibility are less effective than the strategy to “induce in them a sense of forgiveness” (Ghosh, 2017). Nevertheless, negative WOM cannot be observed in the DPM.

The erroneous conclusions caused by omitting important explanatory variables are listed in Table 5. Some coefficients of the variables in the DPM were contrary to the expectations of the theories; nevertheless, all coefficients of the variables in the SDPM were consistent with the expectations of the theories. The expected effect of population size on tourism demand is positive (Martins et al., 2017), and the expected effect of financial crisis on tourism demand is negative (Song and Li, 2008). The two variables’ signs were contrary to what was predicted by the theories and the empirical studies on the DPM, but these signs were not observed in the SDPM.

Compared with the DPM, our SDPM uses the spatially lagged dependent variable to model external habit (WOM) and uses the time-lagged dependent variable to model internal habit (HFP), instead of using only one time-lagged dependent variable to model the two different habits. The empirical results suggest that the SDPM is a more accurate model for modeling tourism demand than the DPM. The effects of the variables in the SDPM were more consistent with theoretical expectations than those in the DPM. The individual coefficient of external habit in SDPM-recommended policies responds to market conditions better than that in the DPM did.

Conclusion

Both HFP and WOM play critical roles in influencing tourists’ decisions to visit tourism destinations and, therefore, tourism policies and tourism management resource allocations. For promoting HFP-induced tourism, tangible and randomly scheduled rewards are recommended (Lally and Gardner, 2013). For promoting WOM-induced tourism, tourist brochures or media materials providing tourism information are recommended (Molina and Esteban, 2006) when WOM is positive, and offering overt expression of forgiveness and prudent reconciliation are recommended (Ghosh, 2017) when WOM is negative. However, neither effect can be confirmed in the prior tourism demand literature because all effects were integrated into one factor. The integration of different types of effects that induce tourism makes it difficult to plan appropriate tourism management strategies according to the empirical results; furthermore, the integration biases the empirical results of tourism demand analyses.

This article solved the long-existing ambiguity in the tourism demand model. We used economic theories regarding internal habits and external habits to clarify the meanings of HFP and WOM and revised the tourism demand model, SDPM. We used empirical data to estimate the new model and compared the results of the new model with those of prior work. We obtained estimators for HFP and WOM and found that the effect of WOM on Taiwan domestic tourism demand is negative. Specific strategies for overcoming negative WOM were discussed. The observation that there were contrary results in the DPM but not in the SDPM suggested that the SDPM is a more appropriate model than the DPM.

Our SDPM was revised from the tourism demand model, DPM, which was extended to include an additional variable representing WOM to avoid the ambiguity in the DPM. We used the developing methodology, spatial dynamic panel data analysis, to analyze the model empirically. We believe that the new tourism demand model we developed and the spatial econometric method together can improve the accuracy of tourism demand modeling, as listed in Table 5, and contribute to tourism management, as we previously discussed.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Ministry of Science and Technology, Taiwan, under grant number MOST 105-2119-M-142-002.

ORCID iD

Tzu-Ming Liu

References

Arellano

Bond

(1991) Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies 58: 277–297.

Assadzadeh

Mamipour

Najafi

(2014) Estimation of tourism demand function for Australian destinations: a time-varying parameter approach. Anatolia 25: 108–116.

Augeraud-Veron

Bambi

(2015) Endogenous growth with addictive habits. Journal of Mathematical Economics 56: 15–25.

Balli

Tangaroa

(2015) Research note: the impact of marketing expenditure on international tourism demand for the Cook Islands. Tourism Economics 21: 1331–1343.

Becken

(2011) A critical review of tourism and oil. Annals of Tourism Research 38: 359–379.

Becken

Carmignani

(2016) Does tourism lead to peace? Annals of Tourism Research 61: 63–79.

Becker

(1996) Accounting for Tastes. Cambridge, MA: Harvard University Press.

Becker

Murphy

(1988) A theory of rational addiction. Journal of Political Economy 96: 675–700.

Berger

(2014) Word of mouth and interpersonal communication: A review and directions for future research. Journal of Consumer Psychology 24: 586–607.

10.

Blundell

Bond

(1998) Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics 87: 115–143.

11.

Boldrin

Christiano

Fisher

(2001) Habit persistence, asset returns, and the business cycle. American Economic Review 91: 149–166.

12.

Bouakez

Cardia

Ruge-Murcia

(2005) Habit formation and the persistence of monetary shocks. Journal of Monetary Economics 52: 1073–1088.

13.

Bouayad-Agha

Védrine

(2010) Estimation strategies for a spatial dynamic panel using GMM. A new approach to the convergence issue of European regions. Spatial Economic Analysis 5: 205–227.

14.

Bourdieu

(1984) Distinction: A Social Critique of the Judgement of Taste. Cambridge, MA: Harvard university press.

15.

Brown

Ivković

Smith

, et al. (2008) Neighbors matter: causal community effects and stock market participation. The Journal of Finance 63: 1509–1531.

16.

Brown

(1952) Habit persistence and lags in consumer behaviour. Econometrica: Journal of the Econometric Society 20: 355–371.

17.

Campbell

(2013) Word-of-mouth communication and percolation in social networks. The American Economic Review 103: 2466–2498.

18.

Chang

Yang

Liao

, et al. (2007) Hysteresis in unemployment: empirical evidence from Taiwan’s region data based on panel unit root tests. Applied Economics 39: 1335–1340.

19.

Chetty

Szeidl

(2016) Consumption commitments and habit formation. Econometrica 84: 855–890.

20.

Chevalier

Mayzlin

(2006) The effect of word of mouth on sales: online book reviews. Journal of Marketing Research 43: 345–354.

21.

Ciccarelli

Elhorst

(2017) A dynamic spatial econometric diffusion model with common factors: the rise and spread of cigarette consumption in Italy. Regional Science and Urban Economics 72: 131–142.

22.

Crouch

(1994) The study of international tourism demand: a review of findings. Journal of Travel Research 33: 12–23.

23.

Dogru

Sirakaya-Turk

Crouch

(2017) Remodeling international tourism demand: old theory and new evidence. Tourism Management 60: 47–55.

24.

Dynan

(2000) Habit formation in consumer preferences: evidence from panel data. American Economic Review 90: 391–406.

25.

Elhorst

(2005) Unconditional maximum likelihood estimation of linear and log-linear dynamic models for spatial panels. Geographical Analysis 37: 85–106.

26.

Elhorst

(2010) Dynamic panels with endogenous interaction effects when T is small. Regional Science and Urban Economics 40: 272–282.

27.

Elhorst

(2012) Dynamic spatial panels: models, methods, and inferences. Journal of Geographical Systems 14: 5–28.

28.

Everaert

Pozzi

(2014) The predictability of aggregate consumption growth in OECD countries: a panel data analysis. Journal of Applied Econometrics 29: 431–453.

29.

Friedman

(2008) Theory of the Consumption Function. Princeton, NJ: Princeton University Press.

30.

Garín-Mun

(2006) Inbound international tourism to Canary Islands: a dynamic panel data model. Tourism Management 27: 281–291.

31.

Ghosh

(2017) Managing negative reviews: the persuasive role of webcare characteristics. Journal of Internet Commerce 16: 148–173.

32.

Golman

Hagmann

Loewenstein

(2017) Information avoidance. Journal of Economic Literature 55: 96–135.

33.

Gârleanu

Panageas

(2015) Young, old, conservative, and bold: the implications of heterogeneity and finite lives for asset pricing. Journal of Political Economy 123: 670–685.

34.

Gozgor

Ongan

(2017) Economic policy uncertainty and tourism demand: empirical evidence from the USA. International Journal of Tourism Research 19: 99–106.

35.

Grishchenko

(2010) Internal vs. external habit formation: the relative importance for asset pricing. Journal of Economics and Business 62: 176–194.

36.

Gunter

Smeral

(2016) The decline of tourism income elasticities in a global context. Tourism Economics 22: 466–483.

37.

Hayakawa

(2015) The asymptotic properties of the system GMM estimator in dynamic panel data models when both N and T are large. Econometric Theory 31: 647–667.

38.

Judson

Owen

(1999) Estimating dynamic panel data models: a guide for macroeconomists. Economics Letters 65: 9–15.

39.

Katz

Shapiro

(1985) Network externalities, competition, and compatibility. The American Economic Review 75: 424–440.

40.

Khorunzhina

(2015) Real business-cycle model with habits: empirical investigation. Economic Modelling 46: 61–69.

41.

Korniotis

(2010) Estimating panel models with internal and external habit formation. Journal of Business & Economic Statistics 28: 145–158.

42.

Kraft

Munk

Seifried

, et al. (2017) Consumption habits and humps. Economic Theory 64: 305–330.

43.

Kukenova

Monteiro

(2009) Spatial dynamic panel model and system GMM: a Monte Carlo investigation. Available at: http://ideas.repec.org/p/pra/mprapa/11569.html

44.

Laamanen

(2013) Estimating demand for opera using sales system data: the case of Finnish National Opera. Journal of Cultural Economics 37: 417.

45.

Lally

Gardner

(2013) Promoting habit formation. Health Psychology Review 7: S137–S158.

46.

Lim

(1997) Review of international tourism demand models. Annals of Tourism Research 24: 835–849.

47.

Lim

(1999) A meta-analytic review of international tourism demand. Journal of Travel Research 37: 273–284.

48.

Mahajan

Muller

Kerin

(1984) Introduction strategy for new products with positive and negative word-of-mouth. Management Science 30: 1389–1404.

49.

Martins

Gan

Ferreira-Lopes

(2017) An empirical analysis of the influence of macroeconomic determinants on World tourism demand. Tourism Management 61: 248–260.

50.

Mileva

(2007) Using Arellano-Bond Dynamic Panel GMM Estimators in Stata. Economics Department, Fordham University, New York, NY, USA, pp. 1–10.

51.

Molina

Esteban

(2006) Tourism brochures: usefulness and image. Annals of Tourism Research 33: 1036–1056.

52.

Naudé

Saayman

(2005) Determinants of tourist arrivals in Africa: a panel data regression analysis. Tourism Economics 11: 365–391.

53.

Parent

LeSage

(2011) A space–time filter for panel data models containing random effects. Computational Statistics & Data Analysis 55: 475–490.

54.

Peng

Song

Crouch

, et al. (2015) A meta-analysis of international tourism demand elasticities. Journal of Travel Research 54: 611–633.

55.

Pfeffer

Zorbach

Carley

(2014) Understanding online firestorms: negative word-of-mouth dynamics in social media networks. Journal of Marketing Communications 20: 117–128.

56.

Poudyal

Paudel

Tarrant

(2013) A time series analysis of the impact of recession on national park visitation in the United States. Tourism Management 35: 181–189.

57.

Ritchie

Amaya Molinar

Frechtling

(2010) Impacts of the world recession and economic crisis on tourism: North America. Journal of Travel Research 49: 5–15.

58.

Roodman

(2009) A note on the theme of too many instruments. Oxford Bulletin of Economics and Statistics 71: 135–158.

59.

Song

(2008) Tourism demand modelling and forecasting—a review of recent research. Tourism Management 29: 203–220.

60.

Song

Witt

(2003) Tourism forecasting: the general-to-specific approach. Journal of Travel Research 42: 65–74.

61.

Song

Witt

(2003) Modelling and forecasting the demand for Thai tourism. Tourism Economics 9: 363–387.

62.

Song

Witt

, et al. (2010) Tourism demand modelling and forecasting: How should demand be measured? Tourism Economics 16: 63–81.

63.

Tallarini

Thomas

Zhang

(2005) External habit and the cyclicality of expected stock returns. The Journal of Business 78: 1023–1048.

64.

Veblen

(1953) The Theory of the Leisure Class: An Economic Study of Institutions. New York: New American Library.

65.

Verhelst

Van den Poel

(2014) Deep habits in consumption: a spatial panel analysis using scanner data. Empirical Economics 47: 959–976.

66.

Witt

Martin

(1987) Econometric models for forecasting international tourism demand. Journal of Travel Research 25: 23–30.

67.

Witt

(1995) Forecasting tourism demand: a review of empirical research. International Journal of Forecasting 11: 447–475.