On Effects of Personality Traits on Travelers’ Heterogeneous Preferences: Insights from a Case Study in Urumqi, China

Abstract

Personality is a psychological concept which reflects people’s characteristic patterns of thoughts, feelings, and behavior. In this sense, it is straightforward to state that people’s behavior is fundamentally affected by their personalities. However, the concept of personality is rarely considered in most travel behavior studies. Given this fact, this paper aims to investigate the effect of personality traits on travelers’ heterogeneous preferences in the context of air itinerary choice. To this end, travelers’ stated choices for air itinerary and information reflecting travelers’ personality traits were collected. After defining specific personality traits based on the collected data, a hybrid choice model with interaction effects between travelers’ personality traits/socio-demographic characteristics and alternative-specific attributes were established and estimated and next the contributions of personality traits and socio-demographic characteristics to travelers’ choice behavior were compared and analyzed. The results confirmed the effects of personality traits on travelers’ heterogeneous preferences. However, the results also revealed that the magnitudes of such effects are not as great as the effects of socio-demographic characteristics.

1. Introduction

If one looks into the literature of choice modeling, he/she could find numerous studies about identifying determinants in terms of decision makers’ behavior in certain choice context, and he/she also could find many studies about how psychological effects, such as loss aversion effect [1,2] and regret avoidance effect [3], affect decision makers’ choice behavior. However, one may get the impression that choice modeling community seems not interested in answering the question why on earth a decision maker would show a certain behavior or why those psychological effects exist. Regarding such questions, borrowed from the findings from psychology, it is reasonable to state that the answer may lie in decision makers’ personalities, which distinguish one from others and make him/her a unique person.

According to Coon & Mitterer, personality is “a person’s unique long-term pattern of thinking, emotions, and behavior” [4]. Two main points could be drawn from this statement. First, everyone has his/her unique personality. As we stated above, personality is used to distinguish one from others. Second, personality is stable and would not be changed frequently. To sum up, personality refers to “the consistency in who you are, have been, and will become”. The representation of someone’s personality is normally based on the concept of trait. According to Allport, personality traits are organized mental structures describing quality that person shows in most situations [5]. In the literature, someone’s personality is normally represented by several traits, and the best accepted and most commonly used model of personality trait must be the Big 5, short for “the big five personality traits”. According to this model, people’s personalities could be generally represented by five traits: extraversion, neuroticism, agreeableness, conscientiousness and openness. Extraversion is a tendency to seek stimulation and to enjoy the company of other people; neuroticism is the tendency to experience negative emotions, such as anger, anxiety, or depression; agreeableness is a tendency to be compassionate toward others; conscientiousness is a tendency to show self-discipline, to be dutiful, and to strive for achievement and competence; openness is a tendency to enjoy new intellectual experiences and new ideas [6]. Although there are a number of studies agree with the big five personality traits, the items that represent these traits are various. For instance, the International Personality Item Pool collected over 3000 items of personality traits that have been used in the literature [7].

In the field of travel behavior analysis, the concept of personality has also been largely ignored for a long time. On the one hand, travelers’ personality is a psychological factor that cannot be observed directly by the researchers, which causes many difficulties to model in the early days of travel behavior analysis. On the other hand, personality, by its definition, is constructed and applied to distinguish one person from others whose relationship with travel behavior is subtle and not noticeable. In fact, when transportation community started to take travelers’ psychological factors into consideration, travelers’ attitude is their first choice [8,9] as attitude is a disposition to respond favorably or unfavorably to a given external target [10]. In this way, travelers’ attitude is naturally with the ability to reflect travelers’ preferences and easily to be thought as important for travelers’ choice behavior.

Nevertheless, with the development of modeling techniques and research on travel behavior, we believe it is time to put travelers’ personality on the table and study its potential role in travel behavior. Therefore, the aim of this paper is to emphasize again and also remind the transportation community of the importance of travelers’ personality when modeling travel behavior. The effect of personality on choice behavior may vary, while the attention of this paper is paid to its effect on heterogeneous preferences in the context of air itinerary choice. In this sense, the contribution of this study is twofold. First, this study confirms the significant effects of travelers’ personalities on their preferences. Second, this study compares such effects with the effects of travelers’ socio-demographic attributes and proves that the effects of travelers’ personality is more subtle.

The remainder of the paper is structured as follows. Section 2 gives a brief literature review about studies of personality in transportation. Section 3 shows the details of data collection. Section 4 presents the implementation of exploratory factor analysis for the scales of personality traits. Section 5 introduces the model specifications. Section 6 gives the model estimation results and corresponding analysis. Section 7 discusses the conclusions and summaries the paper.

2. Literature Review

Now let us look into the studies of personality in transportation. Instead of exploring the role of travelers’ personality on their travel behavior, researchers are more interested in the effect of personality related to traffic safety. Poó & Ledesma investigated the relationship between drivers’ personality traits and driving styles and the various effects of personality traits in different socio-demographic groups [11]. Constantinou et al. and Lucidi et al. focused on the same topic but in particular young and elderly samples, respectively [12,13]. Wood et al. investigated the variation of effects of drivers’ personality traits on automobile fatality rates in United States [14]. It found that the fatality rates is higher when drivers described themselves as being depressed, moody, and quarrelsome. What is interesting is that it found that some relationships between personality traits and fatality rate are less intuitive. Meanwhile, there are studies putting their efforts on how drivers’ personality traits would influence their driving behavior. Mallia et al. found that drivers’ personality traits were associated to aberrant driving behavior both directly and indirectly (through their attitudes), which was also confirmed by Monteiro et al. [13,15]. In addition, the effect of personality on driving behavior or traffic safety is culture-dependent as personality is also culture-dependent [4]. Therefore, some researchers focus on the effect of personality in different culture context. For instance, Nordfjærn et al. investigated the effects of personality traits related on traffic safety and compared such effects between Turkish and Iranian samples [16]; Shen et al. presented a Chinese case study about the effects of personality traits on prosocial and aggressive driving behavior [17].

In terms of the field of travel behavior analysis, it is easy to think of that travelers’ personality is related to their acceptance of new mobility types or travel-related policies. Kim et al. investigated travelers’ acceptability of sustainable transport policies, which was the carbon taxation [18]. It found that personality traits were indirectly related to the acceptability through both attitude and perception and not all personality traits were related to the acceptability. Lopez-Carreiro et al. investigated travelers’ willingness to adopt MaaS ending up with the conclusion that it was directly influenced by travelers’ attitudes toward MaaS which were further influenced by fundamental personality traits [19]. There are other studies paying attention to effects of travelers’ personality traits on public transport usage [20,21]. Meanwhile, it is easy to think of the potential effects of travelers’ personality traits on their travel satisfaction and well-being. Gao et al. explored the influence of travel satisfaction on extent subjective wellbeing or life satisfaction where travelers’ personality traits played the key roles [22]. It found that both personality traits and socio-demographic characteristics could influence the relationship between travel satisfaction and well-being, and personality traits were to some degree influenced by travelers’ socio-demographic characteristics. The same authors also presented a study about the effect of personality traits on daily trips [23]. Munoz & Laniado investigated air passengers’ demands for international round-trip flights by assuming passengers’ airline choice behavior partially depends on their satisfaction which in turn was influenced by personality traits [24]. In terms of the relationship between personality traits and travel mode choice behavior, Johansson et al. could be the one of the first studies, who studied the influence of personality traits on travelers’ commuting mode choice [25]. Yazdanpanah & Hosseinlou is another study who investigated the effect of passengers’ personality traits in the context of airport public transport access mode choice [20]. After the outbreak of COVID-19 pandemic, Mussone & Changizi investigated the relationship between personality traits (and other subjective factors) and transport mode choice during the first lock-down in Milan, Italy [26]. Finally, another interesting research line about personality in travel behavior analysis is how personality influences the travelers perceive external information which further influence travelers’ behavior [27,28,29].

All in all, studies about effects of travelers’ personalities on travel choice behavior is relative scarce in the literature. Although the modeling techniques are well-developed, it seems that transportation community still not widely realize the importance and the role of personality in choice behavior analysis. Inspired by this, this study tries to remind the transportation community of the effects of travelers’ personality in travel choice behavior using a stated preference case study in the context of air itinerary choice.

3. Data Collection

The data used in this paper is actually a sub set of a bigger data set which serves for a project aiming to uncover the potential effects of decision makers’ personalities on both decision mechanism and choice preferences. In order to collect such data, a questionnaire based on s stated choice experiment was applied to capture respondents’ hypothetical choice behavior in the context of air itinerary choice. Meanwhile, respondents’ self-reported personality information and their socio-demographic characteristics were also collected. The followings describe the details of the questionnaire.

3.1. Questionnaire

In terms of the stated choice experiment, air trips from Urumqi to Beijing were considered. Three hypothetical and unlabeled air itineraries consisted the choice set, which were described by several attributes: departure time, arrival time, flight time, connection, departure punctuality, arrival punctuality, ticket fare and airline meal. Most of these attributes were repeatedly confirmed by other studies as significant determinants that would influence travelers’ air itinerary choice behavior [30,31,32,33,34]. One thing should be noted is that in this study we differentiated the effects of departure punctuality and arrival punctuality which are normally combined into one in previous studies. The descriptions of these attributes and their corresponding levels considered in the experiment are presented in

Table 1. Alternative-specific attributes and corresponding levels.

Attributes	Description	Levels
Departure time	The time moment that the airplane take off in the airport	6:00; 12:00; 18:00; 24:00
Arrival time	The time moment that the airplane arrives at the destination airport	Calculated based on departure time and flight time
Flight time	Travel time from origin to destination airports, including the connection	3.5 h; 4 h
Connection	Whether the itinerary contains an intermediate stop	No connection; Once, 2 h
Departure punctuality	Average punctuality about departure for an air itinerary (%)	60%; 70%; 80%; 90%
Arrival punctuality	Average punctuality about departure for an air itinerary (%)	60%; 70%; 80%; 90%
Ticket fare	Price of a ticket for an air itinerary (RMB)	¥1000; ¥1500; ¥2000; ¥2500
Airline meal	Whether the airline provides meals on the flight	Yes; No

.

According to the considered alternative-specific attributes and corresponding levels, an orthogonal design was implemented using R package “support.CEs.” [35]. Finally, 16 profiles (i.e., choice sets) were generated, which were further grouped into 2 blocks. Therefore, each respondent was requested to complete 8 choice tasks. Figure 1 shows an example of the stated choice tasks. Different from the traditional choice tasks used in most previous studies where respondents could only choose the best alternative, in this study respondents were allowed to choose multiple alternatives. For instance, the option “I prefer itineraries A or B, they are almost the same” indicates that the utilities of itinerary A and itinerary B are almost the same and their difference is minor and hardly to be identified by respondents. In this way, we can measure the effect of limited conceptions of respondents on choice behavior [36,37]. However, this is already beyond the scope of the current study and will be investigated in our further analyses.

In terms of respondents’ personality traits, the big five inventory from John & Srivastava was adopted [38]. More specific, all reverse-scored items were removed to avoid respondents’ potential misunderstanding.

Table 2. Constructed Scales for Personality Traits.

Personality Traits	Items	Statements
Agreeableness	item24	Likes to cooperate with others
	item5	Is helpful and unselfish with others
	item12	Has a forgiving nature
	item15	Is generally trusting
	item21	Is considerate and kind to almost everyone
Conscientiousness	Item23	Makes plans and follows through with them
	Item2	Does a thorough job
	Item8	Is a reliable worker
	Item18	Perseveres until the task is finished
	Item26	Does things efficiently
Extraversion	Item1	Is talkative
	Item7	Is full of energy
	Item11	Generates a lot of enthusiasm
	Item17	Has an assertive personality
	Item22	Is outgoing, sociable
Neuroticism	Item13	Worries a lot
	Item19	Can be moody
	Item9	Can be tense
	Item3	Is depressed, blue
	Item25	Gets nervous easily
Openness	Item10	Is ingenious, a deep thinker
	Item14	Has an active imagination
	Item16	In inventive
	Item4	Is original, comes up with new ideas
	Item20	Values artistic, aesthetic experiences
	Item6	Is curious about many different things
	Item27	Is sophisticated in art, music, or literature

presents the details of the items of each personality trait that were used in this study. Respondents were requested to read these statements of items and to report the extent to which they agree or disagree with these statement using a 5-point Likert scale (from disagree strongly to agree strongly). In terms of respondents’ socio-demographic characteristics, information about gender, education level, household location, monthly expenditure, occupation, and age were collected.

3.2. Data Description

The survey was administered in September and November 2019 in Urumqi, China. Some questionnaires were distributed online while some were collected in person. Finally, 409 respondents completed and provided valid questionnaires. Since each respondent was requested to complete 8 stated choice tasks, in total 3272 valid observations were collected. However, since we allowed the respondents to choose multiple alternatives, those indicating that two or all itineraries are almost the same were removed. At last, 2342 observations from 385 respondents remain, in which 864 indicate Itinerary A was chosen (36.89%), 808 indicate itinerary B was chosen (34.50%) and 670 indicate Itinerary C was chosen (28.61%).

In terms of respondents’ personality traits,

Table 3. Descriptive statistics of respondents’ scores in terms of personality items.

	Mean	Median	Mode	Standard Deviation	Minimum	Maximum
item1	3.23	3	3	1.05	1	5
item2	3.39	3	3	1.01	1	5
item3	2.62	3	3	1.09	1	5
item4	3.33	3	3	0.97	1	5
item5	3.76	4	4	0.96	1	5
item6	3.57	4	3	1.01	1	5
item7	3.55	4	3	0.99	1	5
item8	3.82	4	4	0.97	1	5
item9	2.79	3	3	1.13	1	5
item10	3.48	3	3	1.05	1	5
item11	3.46	3	3	1.02	1	5
item12	3.74	4	4	1.06	1	5
item13	3.28	3	3	1.20	1	5
item14	3.48	3	3	1.03	1	5
item15	3.46	3	3	1.05	1	5
item16	3.30	3	3	1.04	1	5
item17	2.81	3	3	1.16	1	5
item18	3.38	3	3	1.11	1	5
item19	2.61	3	3	1.24	1	5
item20	3.14	3	3	1.19	1	5
item21	3.63	4	4	1.04	1	5
item22	3.47	3	3	1.06	1	5
item23	3.35	3	3	1.08	1	5
item24	3.41	3	4	1.08	1	5
item25	3.14	3	3	1.15	1	5
item26	3.40	3	3	0.97	1	5
item27	3.36	3	3	1.20	1	5

shows the descriptive statistics of the scales. From Table 3 we can see that the self-reported scores with respect to all scales are skewed to the high rates: the means are around 3 and the median/mode are either 3 or 4. In terms of respondents’ socio-demographic characteristics,

Table 4. Descriptive statistics of respondents’ socio-demographic characteristics.

Socio-Demographic Characteristic	Level	Percentage
Gender	Male	75.1%
	Female	24.9%
Education	Bachelor	37.9%
	Master/PhD	62.1%
Household location	Rural	53.3%
	Urban	46.7%
Monthly expenditure	0~1000	22.5%
	10,001~15,000	0.5%
	1001~2000	46.2%
	15,001~20,000	0.2%
	2001~3000	19.6%
	3001~4000	3.9%
	4001~5000	2.9%
	5001~6000	3.0%
	6001~7000	0.2%
	8001~9000	0.8%
	above 20,000	0.2%
Occupation	Government Staff	2.4%
	University Student	93.0%
	Company Staff	3.4%
	Others	1.2%
Age	16	0.2%
	17	0.8%
	18	9.3%
	19	10.7%
	20	6.6%
	21	3.9%
	22	7.4%
	23	14.2%
	24	13.6%
	25	10.3%
	26	4.9%
	27	3.2%
	28	3.4%
	29	3.7%
	30	3.4%
	31	1.2%
	32	1.0%
	33	0.5%
	34	0.2%
	35	0.3%
	37	0.2%
	38	0.3%
	39	0.2%
	40	0.5%

shows the descriptive statistics of each attribute. From Table 4, several general images of the respondents could be pictured. First, over three-fourths (75.1%) of the respondents are male. Second, all respondents have at least a bachelor degree, of which around two thirds (62.1%) have a master or PhD degree. Third, the household location is almost even distributed: 53.3% of the respondents live in rural region and 46.7% live in urban region. Fourth, most of the respondents are students (93.0%) as the questionnaires were largely distributed in universities in Urumqi. Since most of the respondents are university students, most are aged below 30 (95.6%).

4. Exploratory Factor Analysis

The scales of the big five personality traits used in this study were directly adopted from John & Srivastava (1999), therefore the reliability seems guaranteed [38]. However, there still may be some potential bias as they were developed over 20 years and for western countries. In this sense, it is necessary to implement exploratory factor analysis to test the reliability and also to define the valid personality traits for following analysis. First of all, the Bartlett test of sphericity and KMO measure of sampling adequacy test were adopted to examine whether the scales are appropriate for factor analysis. The results are presented in

Table 5. KMO and Bartlett’s test.

KMO Measuring of Sampling Adequacy	Overall MSA	0.857
Bartlett test of sphericity	${\chi }^2$	3737
	Degree of freedom	351
	p-value	<0.001

, which shows that the overall MSA value is 0.857 for KMO test and the p value is lower than 0.001 for Bartlett test of sphericity. According to Cerny & Kaiser and Hair et al., a MSA value larger than 0.8 and a p value lower than 0.05 could be seen as meritorious and appropriate for factor analysis of the scales [39,40]. Therefore, it could be concluded that the factor analysis can be applied for the scales.

As one aim of factor analysis is to test whether the scales from John & Srivastava are still useful and appropriate in the current case, therefore the number of personality traits in factor analysis was fixed to five [38].

Table 6. Rotated factor matrix in exploratory factor analysis.

	Factor 1	Factor 2	Factor 3	Factor 4	Factor 5
item12	0.763
item8	0.547
item7	0.542
item14	0.530
item15	0.528
item11	0.507
item5
item10
item23		0.725
item2		0.650
item26		0.625
item18		0.579
item24		0.545
item21
item1			0.751
item4			0.625
item22			0.615
item16			0.523
item6
item3				0.744
item9				0.721
item19				0.672
item25				0.636
item13				0.584
item17				0.506
item20					0.823
item27					0.771

Note: Items that are selected for following analysis are marked in bold.

presents the factor loadings based on “varimax” ration method, in which only values larger than 0.5 are shown. Comparing to Table 2, it is surprising to find that only two factors (i.e., Factor 2 and Factor 4) are generally in line with the scales from John & Srivastava [38]. Specifically, Factor 2 is basically consistent with the scales of the trait “conscientiousness” and Factor 4 is basically consistent with the scales of the trait “neuroticism”. In addition, the Cronbach’s alpha was calculated for the two factors in order to examine the reliability of the scales, which a value larger than 0.7 is normally believed as acceptable [40,41]. The alpha values for the trait “conscientiousness” (i.e., Factor 2) is 0.71 and for the trait “neuroticism” (i.e., Factor 4) is 0.73, both of which confirm the reliability of the scales in terms of the corresponding traits. Therefore, based on the above analyses, these two traits were derived and used in the following analyses.

5. Models

Since respondents’ personality traits are unobserved, a hybrid choice model which integrated a discrete choice model and a latent variable model is adopted [42,43]. After the two decades of development, hybrid choice model has been widely accepted to capture the effects of latent determinants on travel choice behavior [44]. Besides, since the choice tasks in the stated choice experiment are unlabeled, it seems unreasonable to introduce respondents’ personality traits as explanatory variables to measure the main effects of the utility function. Therefore, in this study they are introduced into the model to measure the interaction effects with alternative-specific attributes instead. In addition, we also consider the interactions of alternative-specific attributes with socio-demographic attributes in order to compare with the interactions with personality traits. The following equations describe the utility functions of the three air itineraries:

$$U_{ntj}=V_{ntj}+{\epsilon }_{ntj}$$

(1)

$$V_{ntj}={\beta }_j+{\sum }_k{\beta }_k·x_{ntjk}+{\sum }_k{\sum }_z{\alpha }_{kz}·x_{ntjk}·{PT}_{nz}+{\sum }_k{\sum }_l{\gamma }_{kl}·x_{ntjk}·{SO}_{nl}$$

(2)

where $U_{ntj}$ denotes the random utility of alternative $j$ for respondent $n$ in the choice task $t$, which consists of a representative utility $V_{ntj}$ and a pure error utility ${\epsilon }_{ntj}$; $x_{ntjk}$ denotes the alternative-specific $k$ of alternative $j$ in the choice task $t$ for respondent $n$; ${PT}_{nz}$ denotes the personality trait $z$ for respondent $n$; ${SO}_{nl}$ denotes the socio-demographic attribute $l$ for respondent $n$; ${\beta }_k$, ${\alpha }_{kz}$ and ${\gamma }_{kl}$ are associated parameters to be estimated; ${\beta }_j$ is the alternative-specific constant for alternative $j$. For unlabeled choice tasks, one may argue that the alternative-specific constants should be the same for all alternatives and no need to be estimated. This argument might be true but Hensher et al. also argues that there may be other bias caused by the experiment or the choice task itself (e.g., left-oriented bias) and therefore suggests estimating the alternative-specific constants first [45]. In this study we follow this suggestion and estimate the alternative-specific constants.

In addition, since each respondent was requested to complete 8 choice tasks (although some may be removed if the respondent choose multiple air itineraries), panel effect, which captures the potential correlation among sequential choice tasks completed by a single respondent, should be taken into consideration. Therefore, three random terms (${\eta }_{nj}$) are introduced in the utility function and Equation (1) should be modified as follows:

$$U_{ntj}=V_{ntj}+{\phi }_{nj}+{\epsilon }_{ntj}$$

(3)

Normally, the random term ${\eta }_{nj}$ is assumed to follow a normal distribution with mean of zero and standard deviation ${\sigma }_j$.

In terms of the latent variable model part, respondents’ personality traits are represented by some specific scales which are shown in Table 2. The measurement relationship between personality traits and scales could be described as follows:

$$I_{nzi}={\varrho }_{i0}+{\rho }_{iz}·{PT}_{nz}+{\varpi }_{ni},{\varpi }_{ni}~N\left(0,{\sigma }_{{\varpi }_i}\right)$$

(4)

where $I_{nzi}$ denotes the $i$th scale of personality trait $z$ for respondent $n$ and is made up of a constant ${\varrho }_{i0}$, personality trait $z$ and an normally distributed error term ${\varpi }_{ni}$ with mean of zero and standard deviation ${\sigma }_{{\varpi }_i}$; ${\rho }_{iz}$ is the scale $i$’s measurement parameter on the personality trait $z$. Further, the personality traits are structured by a constant and a normally distributed error term. Therefore, the structural relationship could be presented as follows:

$${PT}_{nz}={\phi }_{z0}+{\zeta }_{nz},{\zeta }_{nz}~N\left(0,{\sigma }_{{\zeta }_z}\right)$$

(5)

where ${\phi }_{z0}$ is the constant and ${\zeta }_{nz}$ is the normally distributed error term with mean of zero and standard deviation ${\sigma }_{{\zeta }_z}$. One thing should be noted is that in the structural relationship there are no other explanatory variables such as respondents’ socio-demographic attributes. The reasons are as follows. First, respondents’ socio-demographic attributes have been introduced into the discrete choice model part to measure the interaction effects. Introducing these attributes again in the latent variable model part may cause unexpected correlation or bias. Second, the aim of this study is not to explore the construction of respondents’ personality traits but to explore the effect of personality traits on respondents’ heterogeneous preferences.

Combing the above equations, a hybrid choice model is established. Maximum likelihood estimation could be used for model estimation. The log-likelihood function is given as follows:

$$LL=\sum _n\mathit{ln}\left(\underset{\mathbf{\zeta },\mathbf{\phi }}{\overset{}{\iint }}\left(\begin{array}{c}{\displaystyle \prod _t}{\displaystyle \prod _j}{P_{ntj}\left(y_{ntj}=1|\mathbf{\phi }\right)}^{y_{ntj}}\times \\ {\displaystyle \prod _i}\frac{1}{{\sigma }_{{\varpi }_i}}·\varphi \left(\frac{I_{ni}-{\varrho }_{i0}-{\sum }_z{\rho }_{iz}·{PT}_{nz}}{{\sigma }_{{\varpi }_i}}\right)\times \\ {\displaystyle \prod _z}\frac{1}{{\sigma }_{{\zeta }_z}}·\varphi \left(\frac{{\zeta }_{nz}}{{\sigma }_{{\zeta }_z}}\right)\end{array}\right)d\mathbf{\zeta }d\mathbf{\phi }\right)$$

(6)

where $LL$ denotes the log-likelihood; $y_{ntj}$ is a dummy variable, which equals 1 if the respondent $n$ chooses alternative $j$ in choice task $t$ otherwise 0; $P_{ntj}\left(y_{ntj}=1|\mathbf{\phi }\right)$ denotes the respondent n’s choice probability of alternative $j$ in choice task $t$ conditioning on the error terms $\mathbf{\phi }$, which is derived from Equations (2) and (3) based on an assumption that the pure error utility ${\epsilon }_{ntj}$ follows IID standard Gumbel distribution; $\mathbf{\phi }$ is a vector of the random terms that capturing panel effect; $\varphi \left(·\right)$ represents the probability density function of standard normal distribution; $\mathbf{\zeta }$ is a vector of the error terms of all personality traits. Since the log-likelihood function contains integration, simulation is needed to mimic the distributions of random error terms [46].

6. Results Analysis

6.1. Model Estimation

The models were estimated using the R package “maxLik” (version 1.5-2) [47]. All socio-demographic attributes (except age) and the alternative-specific attribute departure time, flight time, connection were regarded as categorical and effects coded [45]. The distributions of the random terms were simulated by 100 scrambled Halton draws based on different prime numbers [48,49]. In addition, for the sake of model identification, the alternative-specific constant associated to air itinerary A was fixed as 0; one of the measurement parameters associated to the corresponding personality trait in Equation (5) was fixed to 1, i.e., ${\rho }_{1z}=1,\forall z$; the structural constants in Equation (6) were fixed to 0, i.e., ${\phi }_{z0}=0,\forall z$. Besides, the standard deviations of random terms that capture panel effect (i.e., ${\phi }_{nj}$) were also assumed identical as the choice tasks are unlabeled.

A pilot estimation was conducted. Insignificant main or interaction effects were removed (i.e., the associated parameters were fixed to 0). The final results are presented in

Table 7. Estimation results (discrete choice model part).

	Estimate	Std. Error	p-Value
alternative-specific constant
air itinerary A	0.0000
air itinerary B	−0.1837	0.0676	0.0066 ***
air itinerary C	−0.3330	0.0737	0.0000 ***
alternative-specific attribute
departure time
8:00	0.2357
13:00	0.4401	0.0571	0.0000 ***
18:00	0.0756	0.0623	0.2249
23:00	−0.7514	0.0643	0.0000 ***
flight time
3.5 h	0.1475
4 h	−0.1475	0.0449	0.0010 ***
connection
no connection	0.3500
once, 2 h	−0.3500	0.0269	0.0000 ***
departure punctuality	0.9191	0.2723	0.0007 ***
arrival punctuality	2.3012	0.2690	0.0000 ***
ticket fare	−0.2340	0.0316	0.0000 ***
interaction effect
departure time × gender
8:00 × male	−0.1857
13:00 × male	−0.0115
18:00 × male	−0.0065
23:00 × male	0.2037
8:00 × female	0.1857
13:00 × female	0.0115	0.0544	0.8329
18:00 × female	0.0065	0.0606	0.9145
23:00 × female	−0.2037	0.0581	0.0005 ***
departure time × household location
8:00 × rural	−0.0204
13:00 × rural	0.0741
18:00 × rural	−0.1046
23:00 × rural	0.0509
8:00 × urban	0.0204
13:00 × urban	−0.0741	0.0493	0.1329
18:00 × urban	0.1046	0.0558	0.0609 *
23:00 × urban	−0.0509	0.0504	0.3133
departure time × NEU
8:00 × NEU	−0.3045
13:00 × NEU	0.1575	0.1022	0.1234
18:00 × NEU	−0.1754	0.1129	0.1204
23:00 × NEU	0.3224	0.1049	0.0021 ***
flight time × education
3.5 h × bachelor	0.0929
3.5 h × master/PhD	−0.0929
4 h × bachelor	−0.0929
4 h × master/PhD	0.0929	0.0423	0.0282 **
connection × household location
no connection × rural	−0.0791
no connection × urban	0.0791
once, 2 h × rural	0.0791
once, 2 h × urban	−0.0791	0.0241	0.0010 ***
connection × NEU
no connection × NEU	−0.0927
once, 2 h × NEU	0.0927	0.0518	0.0739 *
arrival punctuality × education
arrival punctuality × bachelor	0.8618
arrival punctuality × master/PhD	−0.8618	0.2616	0.0010 ***
ticket fare × age	0.0049	0.0013	0.0001 ***
ticket fare × CON	−0.0555	0.0148	0.0002 ***
ticket fare × NEU	0.0288	0.0131	0.0278 **
panel effect	0.2890	0.0585	0.0000 ***
number of scrambled Halton draw	100
sample size	2342
initial log-likelihood	−2572.950
final log-likelihood	−2012.686
rho-squared	0.2178
adjusted rho-squared	0.2065

*** p-value < 0.01; ** p-value < 0.05; * p-value < 0.1. “CON” indicates the trait conscientiousness; “NEU” indicates the trait neuroticism. All variables are effects coded.

for the discrete choice model part and in

Table 8. Measurement and structural relationship for latent variable model part.

		Factor	Indicator	Estimate	Std. Error	p-Value
Measurement relationship	Measurement parameter	Conscientiousness	item2	1.2394	0.1891	0.0000 ***
			item18	1.1647	0.1831	0.0000 ***
			item23	1.4581	0.2272	0.0000 ***
			item26	1.0000
		Neuroticism	item3	1.2978	0.1896	0.0000 ***
			item9	1.3576	0.2019	0.0000 ***
			item13	0.8251	0.1655	0.0000 ***
			item19	1.1888	0.2098	0.0000 ***
			item25	1.0000
	scale constant	Conscientiousness	item2	3.3710	0.0588	0.0000 ***
			item18	3.3644	0.0658	0.0000 ***
			item23	3.3236	0.0595	0.0000 ***
			item26	3.3772	0.0541	0.0000 ***
		Neuroticism	item3	2.6292	0.0608	0.0000 ***
			item9	2.7806	0.0619	0.0000 ***
			item13	3.2805	0.0673	0.0000 ***
			item19	2.6071	0.0724	0.0000 ***
			item25	3.1657	0.0648	0.0000 ***
	scale std. deviation	Conscientiousness	item2	0.7658	0.0402	0.0000 ***
			item18	0.9185	0.0426	0.0000 ***
			item23	0.7779	0.0398	0.0000 ***
			item26	0.8105	0.0356	0.0000 ***
		Neuroticism	item3	0.7672	0.0409	0.0000 ***
			item9	0.7782	0.0420	0.0000 ***
			item13	1.0948	0.0587	0.0000 ***
			item19	1.0105	0.0490	0.0000 ***
			item25	0.9858	0.0509	0.0000 ***
Structural relationship	factor constant	Conscientiousness		0.0000
		Neuroticism		0.0000
	factor std. deviation	Conscientiousness		0.5295	0.0599	0.0000 ***
		Neuroticism		0.6066	0.0766	0.0000 ***

*** p-value < 0.01; ** p-value < 0.05; * p-value < 0.1.

for the latent variable model part. The convergent log-likelihood for the completed hybrid choice model is −6947.828, while the convergent log-likelihood for the ones without interaction effects of respondents’ socio-demographic characteristics and personality traits are −6978.404 and −6964.696, respectively. Comparing these three log-likelihood values using likelihood ratio test (Hensher et al., 2015 [45]), it is easy to conclude that although the interaction effects of respondents’ socio-demographic characteristics or personality traits are both significant, the magnitudes of personality traits seem not as great as socio-demographic characteristics. This conclusion could also confirm the above statement in the Section 1 that the effect of personality is subtle and not noticeable in choice behavior.

In terms of the discrete choice model part, the adjusted rho-squared is 0.2065, which is satisfactory [50]. The alternative-specific constants for air itineraries B and C are also significant and negative, which imply that all else being equal, respondents are more likely to choose air itinerary A (the left one), followed by air itinerary B (the middle one) and air itinerary C (the right one) in sequence. It seems that even if in unlabeled choice tasks, bias still exists and these constants do need to be estimated. As for the panel effect, the results in Table 7 also confirm its significance. Therefore, we can conclude that the choices made by a single respondent do have correlations.

Table 8 presents the measurement and structural relationship for latent variable model part of the hybrid choice model. All parameters, including measurement parameters, scale constants and standard deviations, and personality trait standard deviations are significant at a 99% probability level, which indicates good structures of personality traits and good reliability of the corresponding scales for each trait.

6.2. Analysis of Alternative-Specific Attributes

In terms of alternative-specific attributes, eight attributes were presented to respondents in each choice task, they are: departure time, arrival time, flight time, connection, departure punctuality, arrival punctuality, ticket fare and airline meal. Details can be found in Table 1. Except the attribute arrival time, which was calculated based on departure time and flight time, all attributes were introduced into the model and estimated. The reason that arrival time was not included is to avoid the strong correlation with departure time and flight time. The results in Table 7 show that all considered attributes are significant except the attribute airline meal. The insignificance of airline meal indicates that in the given choice context, which is a trip from Urumqi to Beijing, airline meal seems not important to the respondents although the flight time is already relatively long (i.e., 3.5 or 4 h).

As for the attribute departure time, results show that respondents prefer the air itineraries departing at 13:00 (the day time) the most, followed by those departing at 8:00 (the morning), 18:00 (the early evening) and 23:00 (the evening) in sequence. This finding is in line with previous studies [34]. As for the attribute flight time, the results show that respondents prefer the air itineraries with shorter flight time, which is straightforward and intuitive. As for the attribute connection, the results show that respondents prefer the air itineraries with no connection. This conclusion is also intuitive but if we compare the coefficients associated with the attribute flight time, we may find that respondents seems more likely to avoid air itineraries with (2-h) connection than those with longer (4-h) flight time. As for the attributes departure punctuality and arrival punctuality, the results show that their influences to respondents’ air itinerary choice behavior are different. Specifically, respondents care more about air itineraries’ arrival punctuality than the departure punctuality. This conclusion indicates to the airlines that policies aiming to improving arrival punctuality can lead to more significant benefits. As for the attribute ticket fare, the results show that respondents prefer air itineraries with lower ticket fare.

6.3. Analysis of Personality Traits and Socio-Demographic Attributes

In terms of the interaction effects of personality traits with alternative-specific attributes, the results in Table 7 show that both considered personality traits “conscientiousness” and “neuroticism” are significant to some degree. Specifically, they both have significant interaction effects with air itineraries’ ticket fare. Respondents with higher score in the trait conscientiousness are more sensitive to ticket fare while respondents with higher score in the trait neuroticism are less sensitive to ticket fare. According to the definition of these two traits, we may interpret that respondents with higher score in the trait conscientiousness shows high self-discipline and may be more likely to control things by themselves and unwilling to spend more to make things right while respondents with higher score in the trait neuroticism are easier to get panic and nervous and therefore may be willing to spend extra cost to make things right. One may argue that such interpretation is weak and debatable. We acknowledge this argument and also believe that this could confirm the statement in Wood et al. (2020) that the effect of personality traits is less intuitive [14]. In addition, the trait neuroticism also has significant interaction effects with the attributes departure time and connection. In detail, respondents with higher score in the trait neuroticism are more likely to choose an air itinerary which departs in the evening or has connection.

In terms of the interaction effects of socio-demographic attributes with alternative-specific attributes, the results in Table 7 show that respondents’ gender, household location, education level, and age are significant. As for gender, females are less likely to choose an air itinerary which departs at 23:00 which is imaginable as females may feel unsafe in the evening especially that airports are normally located outside of the city center. As for household location, respondents who live in urban region prefer to air itineraries which departs at 18:00 or have connection. This may be caused by that people living in urban region are more familiar with air trip and more likely to enjoy the trip or night life. As for education level, results show that respondents with a master or PhD degree are less sensitive to flight time and arrival punctuality. The reason behind could be multiple. It may be because that respondents with a master or PhD degree are better at time management or they are just more tolerant. As for age, results show that relatively older respondents are less sensitive to ticket fare, although this interaction effect is minor comparing to the main effect of ticket fare. This may be caused by that most of the respondents in this study are students but the older have already graduated and they can make money by themselves.

Although we cannot compare with the interaction of personality traits and socio-demographic attributes directly, it is still easy to find that there are more alternative-specific attributes that significantly interacted with socio-demographic attributes than with personality traits and interaction effects with socio-demographic attributes are more interpretable.

7. Conclusions, Limitations and Paths for Future Research

7.1. Conclusions

Although people’s behavior could have significant differences across their socio-demographic characteristics, it is believed that their behavior is actually rooted in fundamental psychological factors. Just like the definition of personality, it is the personality rather than the age or gender that makes an individual unique in the world. From this point of view, it is important and essential to study the effect of personality in people’s behavior.

This study tries to remind the transportation community of not ignoring travelers’ personality while modeling travel behavior. A stated choice case in the context of air itinerary from Urumqi to Beijing was presented to confirm the above arguments and also compare the contributions of personality traits and socio-demographic attributes to travelers’ heterogeneous preferences. In detail, a survey including a stated choice experiment and questions about personality traits was carried out, based on which data were collected. Next, exploratory factor analysis was implemented to test the validity of the big five inventory for personality traits and also to define valid personality traits that can be used for modeling. Finally, a hybrid choice model integrating a discrete choice model and a latent variable model was established and estimated. The estimated results reveal satisfaction goodness of fit and provide fruitful insights about the effects of personality traits.

From this study, some general conclusions in terms of personality can be made. First, the results of exploratory factor analysis tell that the well-developed big-five inventory in the literature may be not always appropriate in specific cases. The reasons may lie in the differences of the respondents’ culture background, and may also simply in that the inventory is behind the times. In addition, translation from one language to another language, such as from English to Chinese, can also cause inconsistency. This conclusion reminds the researchers that whether the construct scales of personality traits are from classic references or not, it is always necessary to implement factor analysis to examine the validity and reliability of the scales based on their collected data set. Second, the results from the hybrid choice model confirm the significant effects of personality traits on travelers’ heterogeneous preferences. However, comparing to the effects of travelers’ socio-demographic characteristics, such effects seems minor and sometimes more difficult to interpret. One thing should be emphasized is that such uninterpretable feature does not mean the effects of personality traits make no sense. Instead, it is actually a reflection of the features of personality that it is a fundamental psychological factor whose relationship with behavior is subtle and indirect.

Given the specific meanings of the significant personality traits, there are some practical implications can be drawn from the conclusions. On the one hand, the trait “conscientiousness” indicates travelers’ tendency to show self-discipline, to be dutiful, and to strive for achievement and competence. Therefore, airlines can provide customizable or integrated travel service packages which including all facets that a traveler need to decide. With high or low degrees of conscientiousness, travelers may be more willing to choose customizable or integrated travel services packages. On the other hand, the trait “neuroticism” indicates travelers’ tendency to experience negative emotions. In this sense, it is essential for airlines to improve the level of services (such as shorten the times of intermediate stop according to the findings of this study) as travelers with higher degrees of neuroticism is more likely to enlarge even tiny negative experiences and express such emotion in their social networks.

7.2. Limitations and Future Works

One limitation of this study is that we use the scales of personality traits from previous study. Our original consideration is that these scales are widely used and the reliability and validity are ready confirmed repeatedly. However, it turns out this is not always true, which disables further investigation on effects of personality on choice preferences in this study. Another limitation goes to the sample used in this study. Most of the questionnaires were collected in Xinjiang University, which actually hurts the generality of the findings and conclusions in this study, although it is not a key issue and still acceptable,

In addition, there are still some things could be done in the future. First, the roles of personality in people’s behavior may be multiple. On the one hand, as this study indicated, people’s personality could influence their choice preferences. On the other hand, people’s personality could also influence their decision-making mechanism. For instance, people who are easier to be blue may be more likely to adopt regret-avoidance mechanism rather than utility-purchasing mechanism; people who are easier to be happy may be more likely to ignore minor difference between alternatives. This issue is actually our next work as the survey was designed to model such effects of personality traits. Second, a same person may behave totally different in different choice context, which indicates that the effect of personality traits may also vary across choice context. For instance, in the context of choosing among bus, taxi, metro and car, the personality trait openness may not show its effects. However, once a new mobility service, such as MaaS, comes into the market, this trait may have great contributions to the behavior of choosing MaaS as it reflects the tendency to enjoy new intelligent experiences and new ideas. On the one hand, applications in terms of different choice context should be encouraged. On the other hand, what is the general conclusions of the effect of personality in different choice context is actually a more interesting question needs to be answered. Third, previous studies confirm the significant effect of travelers’ attitudes and perceived images on their choice behavior, and this study argue that travelers’ choice behavior is fundamentally influenced by their personalities. However, there is still a lack of comparison among these factors. Questions such as which factor has the greater power to influence choice behavior or is there any relationship between personality and attitude should be answered in the future.