The effect of weather on tourism
The direct and delayed effect of three meteorological weather on the sales and
orientation of tourism products and how this is moderated by advertisement
2
The effect of weather on tourism
The direct and delayed effect of three meteorological weather on the sales and
orientation of tourism products and how this is moderated by advertisement
Marco Verkruissen
Department of Marketing
Master thesis Marketing Intelligence
June 2018
Vismarkt 15A
0681044653
M.verkruissen@student.rug.nl
S2545128
3
Abstract
Tourism is a global and still growing industry which contributes to local economies all over the world. An enormous number of factors influences the purchase journey of consumers concerning tourism products. Although there currently is some literature about the effects of weather on these tourism products, almost no studies investigated the effect of domestic weather situations and made a clear distinction between the different meteorological variables. This study combined data from Gfk (Growth From Knowledge) and the KNMI (Koninklijk Nederlands Meteorologisch Instituut) to investigate the effect of weather
situations on the purchase and orientation behavior of consumers on tourism products. Results showed that both temperature and rainfall have a negative and positive effect respectively on purchases of tourism products. Only rainfall showed a similar positive effect in the case of orientation. This implies that unfavorable weather situations in a country are an incentive for consumers to spend holidays abroad. This so called “push” effect has not been found for the delayed effects of weather nor have interaction effects been found between the weather variables and advertisement.
4
Preface
This thesis has been written to complete the master marketing intelligence. Not only does this make and end to my master, but my life as a student as well. It has been some wonderful years in Groningen at the Rijksuniversiteit.
6
5.2 Scientific and managerial implications………41
5.3 Limitations and further research………..43
6. References………..…….….45
7. Appendix………..………50
7
1. Introduction
The tourism sector is one of the largest and fastest growing global industries and is a
significant contributor to national and local economies around the world (Scott & Lemieuxa, 2010). Statistics show that the contribution to economies of global travel and the tourism industry, is 9.6 per cent of the global Gross Domestic Product (GDP) and 7.9 per cent of the world-wide employment (Tourism impact data and forecasts, 2009). Although already being such a massive industry, tourism is most likely still to increase over time due to the increase of income and leisure time (Agnew & Palutikof, 2001). In the Netherlands itself tourism is a growing service as well. In 2016, 81 per cent of all the Dutch people visited a foreign country. The total number of holidays in that respectively year was 17.9 million, spending over 15.6 billion in total (Trendrapport toerisme, recreatie en vrije tijd, 2017).
When considering a holiday abroad, a numerous amount of aspects could influence the consumers purchase decision. Gomez (2005) found that climate is one of those essential factors that facilitates tourism and satisfies tourist’s needs. A numerous amount of activities conducted during a holiday relies on these climate and weather conditions, which can’t be conducted otherwise (Gomez, 2005). With weather and climate satisfying the needs of tourists, the importance is found in the buying behavior of consumers as well. Maddison (2001) for instance, found that both weather and climate are one of the most important aspects for consumers influencing their purchase behavior for tourism products.
Tourism is a high cost product for many consumers, which is why this is given a significant amount of attention unlike inexpensive goods during the consideration of a purchase (Morgan & Pritchard, 2002). This significant amount of attention causes advertisements and marketing to be a crucial element in the tourism industry. Advertisements within this product category are there to persuade and suggest things that consumers haven’t considered before (Morgan & Pritchard, 2002). Regarding this advertising shows its essential role in the sales of tourism products.
8 A contemporary example is the use of brochures of which a high percentage refers to the favorable weather conditions in the tourist country (Gomez, 2005).
Weather and climate playing an essential role for satisfying tourist needs and this being well implemented in the advertisements, it is a major factor in persuading people in purchasing tourism products. Freitas (2003) for example argued that climate and weather are pervasive factors for tourists deciding on a potential holiday destination. This however only shows weather to be an important factor when consumers were already in the buying process. Kozak (2002) showed however that favorable weather conditions can be an incentive for people to spent time in a foreign country. Similarly, it was found in a UK survey that 73% of the respondents considered ‘good weather’ as a main reason for going on a holiday (Goh, 2012). Regarding this it can thus be concluded that weather conditions are not only an important factor for tourism products but can also initiate the purchase of such a product.
All literature described above found that preferable weather and climate conditions in tourist destinations have a positive effect on tourism. The only effect found so far however concerns the ‘pull’ effect, where consumers are drawn to weather conditions in the foreign country. There has been a substantial less amount literature so far on a ‘push’ effect, which describes the phenomenon where domestic climate or weather conditions initiates an incentive to spend time abroad. With a ‘pull’ motivation the consumers are drawn to a location because of its attractions and characteristics whereas with a ‘push’ factor people are motivated by their own internal forces (Mohammad, 2010). Unfavorable weather conditions in the country of
residence could potentially be one of those internal forces.
Although there has been some literature about the ‘push’ effect, most literature focusses on the ‘pull’ effect. Smith (1993) however did argue that one motivation for tourists is to escape the weather in their home country. Unfavorable domestic weather conditions might be just as essential for consumers to purchase tourist products as favorable weather conditions in the tourist country. This might be interesting given that a large amount of relative wealthy countries in Europe, that have more financial resources to spend on tourism, are located on places with unfavorable weather conditions (McCrae, Terracciano, Realo & Allik, 2007).
There is some extensive literature about advertising in the tourist sector, such as the
9 of unfavorable weather conditions in the home country on sales of tourism products might be enforced with proper advertisement.
If it is the case that with unpleasant weather conditions and proper advertising, there are more consumers considering or even purchasing tourism products, this requires a proper adaption of organizations offering tourism products such as flight tickets and accommodations. During those unfavorable weather conditions, the organizations should increase their advertising to persuade those doubting consumers to make the final step in booking a trip. This could increase profits or brand awareness since more consumers are paying attention on these products. On top of that these organizations should increase their advertising in those countries with unfavorable weather conditions to target their customers more effectively.
To investigate if this is the case, the research question is:
To what extend does unfavorable weather conditions in a consumer’s home country, effect consumers to purchase a tourism product or orientate more on a tourism product and how is this moderated by advertisement?
Although there is some literature about weather influencing tourism this is mostly focused on meteorological influences from abroad. This research contributes to existing literature by focusing on the effect of domestic weather variables. In addition, this research makes a distinction between three meteorological variables which’s effects are compared on to another. Currently there is only little literature which makes a similar distinction and those that do used a different set of meteorological variables. If significant results will be found, it might be interesting to investigate the effect of weather on other types of products as well.
The remainder of this research is structured in several sections to get a clear overview. Section 2 provides an analysis of current literature about the effect of weather on tourism products. Following up there is a section about the methodology which gives a detailed
10
2. Theoretical Framework
2.1 Tourism
An essential element in this research is the purchase of and orientation on tourism products representing the dependent variable. Because of this importance in this study a definition of tourism needs to be found. Gomez (2015) defined tourism as the “entirety of the relationships, phenomena and experiences that arise from travelling and overnight stay of people in
locations or areas other than their usual residence” (p. 572). A more detailed definition is “taking a trip to a main destination outside his/her usual environment, for less than a year, for any main purpose (business, leisure or other personal purpose) other than to be employed by a resident entity in the country or place visited” (Yu, Kim, Chen & Schwartz, 2012, p. 446). Both definitions agree on the terms of an overnight stay other than the usual residence. The main purpose could be a variety of things as long as no form employment is present.
2.2 Weather
There are several definitions for weather, even though most are similar. Matzarakis (2006) defines weather as “the present combination of atmospheric elements (physical condition of the atmosphere) at a specific time and location, and the resulting processes in the atmosphere (time scale: days, weeks, months)” (p. 101). Almost similar Gomez (2005) described weather as “the state of the atmosphere in a given place at a given time and can be described for one particular weather station or for a specific area of the earth’s surface” (p. 572). With both definitions the specific time, location and the state of the atmosphere are of essential importance in defining the weather.
11
Table 1: Ranked weather elements in different contexts (Scott & Lemieuxa, 2010, p. 172)
Mieczkowski (1985) originally developed the tourist climate index (TCI) that measured the perceived quality of tourist’s experience using the most relevant weather elements. Although the original TCI consisted of 12 elements, this was eventually reduced to only five as could be observed in table 2 (Scott & McBoyle, 2001). Similarly with Scott & Lemieuxa (2010), the most relevant weather variables consist of precipitation, sunshine and temperature (partly representing the index).
Table 2: Tourist climate index (Scott & McBoyle, 2001, p. 71)
Gomez (2005) stated that the “elements that have the greatest influence on tourism are temperature, number of sun hours, precipitation, wind, humidity, and fog” (p. 576). Once again, apart from some additional elements, sunshine, temperature and rainfall are considered as one of the most crucial ones.
12 2.3 Tourism sales
It is clear that consumers are attracted by the weather conditions of foreign countries, and which subsequently is of major influence on tourism. The consumers own country weather conditions however might be of influence as well. Smith (1993) already argued that tourists make a conscious decision to travel in order to gain a short-term advantage concerning weather conditions compared to those conditions back home. There already exist some
patterns of tourism from northern Europa to the Mediterranean climate in the south during the winter to avoid the cold (Smith, 1993). A similar movement exists in North America, with tourists from the north visiting the much sunnier southern states like Mexico and Florida (Smith, 1993). This effect has been found in the UK as well where a relation between a high amount of rainfall and a higher number of visits to Portugal has been found (Smith, 1990). In a Canadian survey it was found that 23% of the respondents had a primary travel motivation to escape from the weather conditions in their own country (Scott, McBoyle &
Schwartzentruber, 2004). Similarly, Jorgenson & Solvoll (1996) found that 84% of tour charters were destined to places with a relatively high amount of sunshine, which might assume that Norwegian civilians tend to escape their own cold climate.
Agnew & Palutikof (2001) found that climate and weather conditions in the home country have a push factor on tourism, whereas this only applies for unfavorable weather conditions. This effect arises both due to weather conditions of current year and previous year (Agnew & Palutikof, 2001). A similar conclusion was drawn by Scott, Jones & Konopek (2008) who found that a1ºC warmer than average summer in Canada would increase the domestic tourism expenditures by 4%. This phenomenon doesn’t seem to be restricted to explicit continents. In the case of Europe, Becken (2010) stated that “a better climate in the region of residence is related to a higher probability of travelling domestically, whereas poor conditions increase the chance of international travel” (p. 5). Although all these researches indicate there is some evidence for unfavorable weather conditions pushing the tourist to foreign countries, the the different meteorological elements, which together represent weather, aren’t compared one to another.
Palutikof (1999) and Agnew and Palutikof (2001) however, did make a difference between the different weather elements causing tourism. They found that rainfall is a good predictor for outward tourism, whereas sunshine and temperature are good indicators for inward
13 conditions and a duller-than-average July in the year prior to travel appear to encourage more trips abroad (Agnew & Palutikof, 2006). Although these weather variables have different effects, some causing consumers to travel abroad and some causing them to travel
domestically, they seem to be of influence on tourism.
Considering all the literature described above there is some evidence for weather causing consumers to purchase tourism products. Current literature is however barely distinguishing between the different impacts of the different weather elements. So far only the studies of Agnew and Palutikof (2001) made a distinction between different meteorological variables. This leads us to the following hypotheses:
H1: An increasing/decreasing amount of sunshine in the home country decreases/increases the sales of tourism products
H2: An increasing/decreasing temperature in the home country decreases/increases the sales of tourism products
H3: An increasing/decreasing amount of rainfall in the home country increases/decreases the sales of tourism products
Lim, Kim, Choi, Choi & Lee (2010) investigated the effect of weather elements on tourism related website visits. They found significant evidence that on days with rainfall, more people visit the websites on average. On top of that there was a relation between the cloudiness of the day and the amount of website visits. Although the effect was weak, the cloudier the day and thus the less sunshine, the more visits the website had. This shows that certain weather
conditions do not only influence sales as well the consideration or orientation of consumers to book a trip. This is however the only research so far about the effect of weather on
orientation, resulting in that a research confirming these assumptions could be useful. This leads to the following hypotheses.
H4: An increasing/decreasing amount of sunshine in the home country decreases/increases the amount orientation of people for tourism products
H5: An increasing/decreasing temperature in the home country decreases/increases the amount orientation of people for tourism products
14 2.4 Advertisement
Throughout the years, advertisement has been a well investigated concept which has been defined in several ways. There are a couple of definitions that are broadly used in current literature. Defleur & Dennis (1996) defined advertising as “a form of controlled
communication that attempts to persuade consumers, using strategies and appeals, to buy or use a particular product or service” (p. 564). Arens (1996) defined advertising as “the
nonpersonal communication of information, usually paid for and usually persuasive in nature, about products (goods and services) or ideas by identified sponsors through various media” (p. 23). With both definitions communication and persuasion of consumers are important aspects defining advertising, whereas the definition of Arens (1996) introduces a new aspect which is a paid form. More recent definitions seem to have adopted this paid form in their definitions with Armstrong & Kotler (2000) that defined it as “any paid form of nonpersonal presentation and promotion of ideas, goods, or services by an identified sponsor” (p. 446) or Richard & Curran (2002) who defined advertising as a “paid nonpersonal communication from an identified sponsor, vising mass media to persuade or influence an audience” (p. 64).
Within tourism, the effectiveness of advertising has been evaluated by focusing on the extent to which a promotional campaign stimulates people to visit a particular destination (Kim & Hwang, 2005). Kim & Hwang (2005) found evidence for the link between exposure to advertisements and the visit to a destination. The results also showed an indication that the awareness of advertisements and the request of travel information had a strong relation with the visit of a destination (Kim & Hwang, 2005).
A large amount of media channels is used throughout the years to advertise tourism. All the different media channels have their own particular strengths and weaknesses in exposing consumers to different advertisement messages (Kim & Hwang, 2005). As an example, television is more effective for those products that require a visual representation, while radio can only be used to put out a message only using sound (Assael, 1981). With the rise of the internet, it has become one of the most important and effective media channels for consumers to seek information and purchase tourism products (Lim et al., 2010). The different media channels address different cognitive and processes and thus different dimensions of
15 could express itself in consumers browsing on the tourist website, or stated differently,
consumer orientating on tourism products.
Bojanic (1991) has shown that advertisement within tourism isn’t only limited to orientation and sales but could potentially also lead to psychological effects such as destination image formation. This destination image formation may affect a consumer to consider a visit to a foreign country while unfavorable weather conditions in the home country are prevailing. The unfavorable weather conditions may be compared to the image in the consumers mind and serve as an incentive. Regarding this it seems possible that advertising could affect consumers buying behavior and consideration for a holiday in combination with certain weather
conditions. This leads to the following hypotheses.
H7: Advertising has a positive effect on the relation between sunshine and sales
H8: Advertising has a positive effect on the relation between temperature and sales
H9: Advertising has a positive effect on the relation between rainfall and sales
H10: Advertising has a positive effect on the relation between sunshine and orientation
H11: Advertising has a positive effect on the relation between temperature and orientation
H12: Advertising has a positive effect on the relation between rainfall and orientation
2.5 Delayed weather
Weather might not only have a direct impact on consumer’s orientation and sales but could also be delayed due to certain circumstances. People might not be immediately affected by a specific day with unfavorable weather conditions but are affected when these days persist. Bigano, Goria, Hamilton & Tol (2005) found evidence for lagged effects of temperature on the number of consumers booking bed-nights. Especially during hot summers when high temperature persisted this had an immense delayed effect. Similarly, Falk (2015) conducted a study to investigate the impact of weather conditions on overnight stays of domestic and German tourists within Austria. They found that besides current, lagged sunshine and
temperature effects can explain a substantial amount of the variation in domestic and German overnight stays (Falk, 2015).
16 of the tourism market where consumers waited with their purchase to get the most profitable deal. This would assume that there is a delayed effect between weather and sales and/or orientation of tourism products. This leads to the following hypotheses.
H13: An increasing/decreasing amount of sunshine in the home country has a decreasing/increasing lagged effect on the sales of tourism products
H14: An increasing/decreasing amount of temperature in the home country has a decreasing/increasing lagged effect on the sales of tourism products
H15: An increasing/decreasing amount of rainfall in the home country has an increasing/decreasing lagged effect on the sales of tourism products
H16: An increasing/decreasing amount of sunshine in the home country has a
decreasing/increasing lagged effect on the amount of orientation on tourism products
H17: An increasing/decreasing amount of temperature in the home country has a decreasing/increasing lagged effect on the amount of orientation on tourism products
H18: An increasing/decreasing amount of rainfall in the home country has an
increasing/decreasing lagged effect on the amount of orientation on tourism products
2.6 Conceptual model
17
Figure 1: conceptual model
3 Methodology
3.1 Dataset
The dataset used in this research is retrieved from GfK which is a marketing research company originated in Germany. The dataset is on an individual level reflecting a relatively recent period from the 1st of June 2015 till the 31st of September 2016. The dataset consists of 7312 consumers who were active on several websites/applications with different services. These services consisted of accommodations, information/comparison, touroperators/travel agency and flight tickets. Although a logical distinction could be made about these four services they are all considered as tourism products within this research. On top of consumer activity on the website or applications, the data indicates if a consumer applied a search term on one of the four services. Since all these actions are initiated by consumers all these variables are considered as orientation by consumers. The remaining variables consists of affiliates, banners, email, prerolls and retargeting. These variables indicate whether a
18 focal company or a competitor. It is possible that although a consumer orientated on the websites or applications of the focal company, the purchase was made at a competitor.
In order to make this research feasible some demographical information has been added to the dataset. An essential variable in this dataset is the region of residence. The place of residence of the consumers is crucial to determine the corresponding weather at a certain time period that might influence the consumers. Although one might argue that the weather in The Netherlands is relatively the same throughout the country, there could still be some minor differences, which is why the weather data of multiple locations will be obtained. In order to determine the weather on the specific days within the dataset, this needs to be obtained from an external source. This is done via the KNMI (Koninklijk Nederlands Meteorologisch Instituut) which is the national weather forecasting service in The Netherlands. A variety of meteorological variables could be retrieved including the essential ones for this research which are sunshine, rainfall and temperature. To be able to determine the weather, this needs to be matched with the place of residence of the consumers. The place of residence is, apart from Amsterdam, Rotterdam and Den Haag, divided in the sections north, south, west and east. To match the match the correct weather information the weather station which is settled most centrally in the four sections will be chosen to represent the weather in that respectively section. The four chosen weather stations that will represent one of the four sections are Schiphol (west), Lauwersoog (north), Hupsel (west) and Eindhoven (south). Since the consumers whose place of residence either is in Amsterdam, Rotterdam or Den Haag are aggregated, it is not possible to make a distinction between the weather they were exposed to. As a result, those consumers will not be used in this research which is unfortunate but
acceptable considering the size of this group compared to the overall dataset. This will be shown in the descriptive section.
19 effects of income and life stage on tourism. This is sufficient reason to use these two
demographics as control variables.
At first the main relations to be investigated is between weather and the sales and orientation. Regarding this the distinction between the different kinds of touchpoints i.e., websites, applications or search term, isn’t relevant and thus will all be combined to one new variable. This will lead to the assumption that more touchpoints equal more orientation. Secondly there is a distinction made within sales, where consumers either purchased a product at the focal company or a competitor. For the primary investigation this isn’t relevant either and thus no distinction will be made between the two, which resulted in a new combined sales variable. For investigating the moderating effect of advertisement, the different forms of advertisement touchpoints will be combined as well representing one individual advertisement variable. This variable thus shows if and how many times a consumer is exposed to advertisement. Finally, the external acquired weather data requires a modification as well. In the cases of sunshine and rainfall there are some data points stating ‘-1’ in the cases of < 0.05 hour of sunshine and < 0.05 mm of rainfall. Since it is simply not possible to have a negative amount of sunshine or rainfall on a day these values will be modified to ‘0’.
In some particular cases there was no available information for the used variables within this research. These data points which didn’t contain the required information were deleted from the dataset. For instance, there is a substantial number of consumers who didn’t have any available information about the place of residence. These consumers will be removed from the dataset because it can’t be determined in what weather situations they reside. Similarly, a certain amount of the consumers resides in Amsterdam, Rotterdam and Den Haag where no distinction can be made to what weather these consumers are exposed to as well. After deleting the just specified consumers, which contained 1720 data points, 5592 consumers remained in the dataset out of the initial 7312
3.2 Descriptives.
20 categories respectively, which distributions are observable in figure 2. It must be noted that the levels of the two variables aren’t related but are combined in one figure for the sake of convenience.
Figure 2: Distribution of life stage and income levels
Out of 17687 conducted orientations, which is a session where a consumer is browsing for tourism products, 108 purchases (4.8%) were made at the focal company whereas 2138 (95.2%) at a competitor. This might suggest that an extensive amount of people uses the websites/applications of the focal company to only orientate and in many cases make their purchase at a competitor.
21
Figure 3: Distribution of the consumer initiated touchpoints
Regarding the company-initiated touchpoints, which represent the different forms of
22
Figure 4: Distribution of the company initiated touchpoints i.e., advertisement forms
23
Figure 5: Temperature of the different weather stations throughout the observed time period
Figure 6: Sunshine of the different weather stations of the first 100 days of the observed time period -10 -5 0 5 10 15 20 25 30 6/1/2015 8/20/2015 11/8/2015 1/27/2016 4/16/2016 7/5/2016 9/23/2016 DE G R EE S CE LCIUS
Temperature
Schiphol Lauwersoog Hupsel Eindhoven
0 2 4 6 8 10 12 14 16 18 6/1/2015 7/1/2015 7/31/2015 8/30/2015 H O URS O F S UN SH IN E
Sunshine
24
Figure 7: Rainfall of the different weather stations throughout the observed time period
3.3 Model specification
Within the next section the model specifications will be shortly discussed. Since there are multiple methods to investigate the research question considering the nature of the dataset, two different models have been used. Both model types will be shortly discussed with their corresponding mathematical form.
3.3.1 Logistic regression
Logistic regression models have become an integral component of many data analysis where the relation between a response variable and one or more explanatory variables are described (Hosmer & Lemeshow, 2000). What distinguishes a logistic regression of a linear regression is that the outcome of the first is binary or dichotomous whereas with the latter it is
continuous (Hosmer & Lemeshow, 2000). This difference is reflected in the choice of parametric model and the assumptions (Hosmer & Lemeshow, 2000), which makes the logistic regression beneficial since this model isn’t restricted to the assumptions of a linear regression (Hair, Black, Babin, & Anderson, 2010). A potential disadvantage of the logistic regression is that it requires a large sample size which doesn’t seem to raise any concern regarding the sufficient amount of data points used in this analysis. The logistic regression model could be expressed as:
0 50 100 150 200 250 300 350 400 450 500 6/1/2015 8/20/2015 11/8/2015 1/27/2016 4/16/2016 7/5/2016 9/23/2016 R ain fall in mm
Rainfall
25
𝑃(𝑌) = 1
1+exp(−(𝑍)) (1)
In this basic form, Z presents a linear combination of the independent variables used to predict Y. Once all the variables are placed in the model it can be expressed as:
𝑃(𝑌)𝑗𝑡=
1
1+exp(−(𝛽1𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡+𝛽2𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡+𝛽3𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡+𝛽4𝐿𝑖𝑓𝑒𝑠𝑡𝑎𝑔𝑒𝑗+𝛽5𝐺𝑟𝑜𝑠𝑠𝑖𝑛𝑐𝑜𝑚𝑒𝑗))(2)
P(Y) can take the values between 0 and 1, which represents the probability that there is a sale or no sale. Because of this binary variable no distinction is made between a purchase made at the focal company or the competitor. The weather variables present in equation 2 are unique with those of equations 4 and 5. This model however contains two control variables which are Lifestagejt and Grossincomejt. Both of these control variables are factors which levels are
showed in figure 2. The variables represent the life stage and gross income of each consumer and are thus on an individual level.
3.3.2 Linear regression
Linear regression is a widely used technique in marketing research (Leeflang, Wieringa, Bijmolt & Pauwels, 2015). Linear regression models assume that the relation between the independent and dependent variables are linear, which is why this model is easy to interpret (Montgomery & Peck, 1992). As this could be considered as an advantage, this could be considered as a downside as well since the model is limited to this linear relation. Another disadvantage of linear regression, as mentioned before, is that the model should conform to a set of assumptions before being useable. Potential violations of these assumptions will be tested in the results section as well as their demanded remedies. The linear regression is expressed as:
𝑌 = 𝛼 + 𝛽𝑋 + 𝜀(3)
26 𝑆𝑖𝑡 = 𝛼 + 𝛽1𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡+ 𝛽2𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡+ 𝛽3𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡+ 𝛽4𝐴𝐷𝑉𝑖𝑡∗ 𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡 + 𝛽5𝐴𝐷𝑉𝑖𝑡∗ 𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡 + 𝛽6𝐴𝐷𝑉𝑖𝑡∗ 𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡+ 𝛽7𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡−5+ 𝛽8𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡−5+ 𝛽9𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡−5+𝜀𝑖𝑡 (4) 𝑂𝑖𝑡 = 𝛼 + 𝛽1𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡+ 𝛽2𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡+ 𝛽3𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡+ 𝛽4𝐴𝐷𝑉𝑖𝑡∗ 𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡+ 𝛽5𝐴𝐷𝑉𝑖𝑡∗ 𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡 + 𝛽6𝐴𝐷𝑉𝑖𝑡∗ 𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡+ 𝛽7𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡−5+ 𝛽8𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡−5+ 𝛽9𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡−5+ 𝜀𝑖𝑡 (5)
Equation 4 represents the complete linear regression model 2.3 including all the variables. Model 2.2 has an identical equation where only the lagged variables are omitted whereas with model 2 the lagged and the interaction variables are omitted. Same thing applies for equation 5 that represents the complete model 3.3 where model 3.2 and 3 could be derived by
excluding the lagged and interaction variables.
Sit represents the dependent variable sales which is aggregated on a daily level so that the
effect of the corresponding weather on that day can be measured. As a result, the dependent variable states the total sales per day for each separate location and thus no individual level data is considered within this model.
Concerning equation 5, same thing applies for the dependent variable Oit that is aggregated on
a daily level as well. This variable represents the total orientation per day for each location. The orientation in this matter equals the total number of touchpoints that occurred for each location within each day. Due to this aggregation no distinction is made between the different types of touchpoints.
The weather variables Sunshineit, Temperatureit, Rainfallit represent the current weather
situation for each day with the corresponding location. The corresponding lagged effects Sunshineit-5, Temperatureit-5, Rainfallit-5 represents the weather situation of five days earlier at
27 presented below which states all the used variables for all the models used within this
research. 𝛼 = 𝐼𝑛𝑡𝑒𝑟𝑐𝑒𝑝𝑡 𝑃(𝑌)𝑗𝑡= Probabilityofsalesofconsumerjatdayt 𝑆𝑖𝑡 = 𝑆𝑎𝑙𝑒𝑠𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝑂𝑖𝑡 = 𝑂𝑟𝑖𝑒𝑛𝑡𝑎𝑡𝑖𝑜𝑛𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝑆𝑢𝑛𝑠ℎ𝑖𝑛𝑒𝑖𝑡 = 𝐴𝑚𝑜𝑢𝑛𝑡𝑜𝑓𝑠𝑢𝑛ℎ𝑖𝑛𝑒𝑖𝑛0.1ℎ𝑜𝑢𝑟𝑠𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑡 = 𝑇𝑒𝑚𝑝𝑒𝑟𝑎𝑡𝑢𝑟𝑒𝑖𝑛0.1𝑑𝑒𝑔𝑟𝑒𝑒𝑠𝑐𝑒𝑙𝑐𝑖𝑢𝑠𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑡= 𝐴𝑚𝑜𝑢𝑛𝑡𝑜𝑓𝑟𝑎𝑖𝑛𝑓𝑎𝑙𝑙𝑖𝑛𝑚𝑖𝑙𝑙𝑖𝑚𝑖𝑡𝑒𝑟𝑠𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝐴𝐷𝑉𝑖𝑡 = 𝐴𝑑𝑣𝑒𝑟𝑡𝑖𝑠𝑒𝑚𝑒𝑛𝑡𝑎𝑡𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑎𝑡𝑑𝑎𝑦𝑡 𝐿𝑖𝑓𝑒𝑠𝑡𝑎𝑔𝑒𝑗= 𝐿𝑖𝑓𝑒𝑠𝑡𝑎𝑔𝑒𝑜𝑓𝑐𝑜𝑛𝑠𝑢𝑚𝑒𝑟𝑗 𝐺𝑟𝑜𝑠𝑠𝑖𝑛𝑐𝑜𝑚𝑒𝑗= 𝐺𝑟𝑜𝑠𝑠𝑖𝑛𝑐𝑜𝑚𝑒𝑜𝑓𝑐𝑜𝑛𝑠𝑢𝑚𝑒𝑟𝑗 𝜀𝑖𝑡 = 𝑑𝑖𝑠𝑡𝑢𝑟𝑏𝑎𝑛𝑐𝑒𝑡𝑒𝑟𝑚
4. Results
In this section the results of the conducted tests will be discussed. Out of the two main effects to be investigated, the effect of weather on sales could be investigated in multiple ways as stated before. Since there are two methods used to investigate hypothesis 1, 2 and 3 this resulted in two statistical tests. These two models will be compared one to another in a later stage. In addition, some interaction and lagged effects are included in model 2 and 3 which resulted in model 2.2, 2.3, 3.2 and 3.3 respectively. To get a clear overview table 3 shows the different tested relations with the corresponding models and hypotheses.
Before the actual analysis some assumptions are checked since these can potentially affect the results. Subsequently the quality of the models will be validated to ensure the usefulness of the results. The final section will discuss the found results and the accepted or rejected hypothesis. To get a clear overview the results will be presented in several tables and figures.
Relation Statistical test
(method)
Name model Corresponding hypothesis Effect of weather on sales binary logistic
regression
Model 1 H1, H2, H3 Effect of weather on sales multiple linear
regression
28
regression Effect of weather on sales including
moderation
multiple linear regression
Model 2.2 H7, H8, H9 Effect of weather on orientation including
moderation
multiple linear regression
Model 3.2 H10, H11, H12 Effect of weather on sales including
moderation and delayed effects
multiple linear regression
Model 2.3 H13, H14, H15 Effect of weather on orientation including
moderation and delayed effects
multiple linear regression
Model 3.3 H16, H17, H18
Table 3: Overview statistical tests and hypothesis
4.1 Assumptions
Since distinctive statistical tests are used, a different set of assumptions needs to be validated. Starting with the logistic regression a first assumption is that the outcome variable explicitly states values of “0” or “1” since the dependent variable is binary. A second assumption is that the observations need to be independent from each other and the third states that the sample size needs to be sufficient enough (Hosmer & Lemeshow, 2000). All of those assumptions are relatively easy to check and are satisfied after a short investigation. The final assumption to be respected concerns multicollinearity which is present when the independent variables are correlated to each other (Malhotra, 2009). One method for checking for multicollinearity is with variance inflation factors. Considering the threshold of 5 (Leeflang et al., 2015), none of the variables show any present multicollinearity, meaning that the independent variables are not explained by one another. As a result, multicollinearity doesn’t seem to raise any concern for all the models regarding this assumption. For a more detailed insight about the specific variance inflation factors scores see appendix A.
The multiple linear regression has its own set of assumptions which requires to be tested for before conducting any further analysis. These assumptions are apart from multicollinearity, a normal distribution of the residuals, heteroscedasticity and autocorrelation (Leeflang et al., 2015). Testing these assumptions is critical since, violations could cause wrong estimates of the parameters as well as wrong estimates of the variance (Hair et al., 2010).
29 observable that the residuals are not normally distributed. Inspecting the histograms of models 2, 2.2 and 2.3 for instance, the distribution seems to be slightly negatively skewed since the data points are clustered on the left side. After an inspection of the plots it was quite clear that the residuals aren’t fitted on the regression line which indicates non-normality. To confirm this an Anderson-Darling test has been conducted which showed highly significant results (p = < 2.2e16) for all six models, indicating that there is indeed no matter of a normal
distribution. With increasing sample sizes however, it occurs more frequently that this assumption is violated (Leeflang et al., 2015). If the model specification however seems appropriate, it is not extremely important to follow the strict rules for these violations (Leeflang et al., 2015). The used histograms and plots for this assumption can be found in appendix B.
With the presence of heteroscedasticity, the variance of the residuals isn’t equal within all the levels of the independent variables (Malhotra, 2009). To investigate the presence of
heteroscedasticity some plots were developed to get a visual inspection. Although it is hard to determine considering the large number of observations, it seemed that the residuals are relatively equally distributed throughout the range of the X axis in model 2, whereas with the remaining models it seemed clear that this isn’t the case. This would assume that only model 2 isn’t limited by the presence of heteroscedasticity. In addition, the red line was only slightly curved within model 2, whereas this line was curved more in the other models, which
indicated heteroscedasticity as well. To confirm this a Breusch-Pagan test has been
conducted, which showed non-significant results for model 2 (p = 0.1683), model 3.2 (p = 0.0786) and model 3.3 (p = 0.2305). The remaining models showed significant results indicating that heteroscedasticity is present. Plots and results of the Breusch-Pagan tests regarding heteroscedasticity could be observed in appendix C.
The final assumption to be tested is autocorrelation which exists when the residuals have some pattern over time (Leeflang et al., 2015). To test any presence of autocorrelation a Durbin-Watson test have been conducted. For all the models, significant results are found, (p = 1.24e-14) for model 2 and (p = < 2.2e-16) for the remaining models. These results indicate that all the models suffer from autocorrelation.
30 transformations and Box-Cox transformations were conducted without any positive result. After conducting a second Breusch-Pagan test there was still an indication of
heteroscedasticity. This lead to the remedy of using heteroscedasticity and autocorrelation consistent (HAC) estimators, or more specifically the Newey-West procedure, which provides a consistent estimate of the covariance of the model parameters (Newey & West, 1994). The weighting of the covariance is determined by the kernel function and truncation lag or “bandwidth” (Kiefer & Vogelsang, 2002).
With this method the form of heteroscedasticity and autocorrelation doesn’t need to be specified (Croux, Dhaene & Hoorelbeke, 2003). Even when there is no present
heteroscedasticity as in model 2, the use of these robust standard errors is appropriate as well.
4.2 Model quality
In this next section the overall significance and quality of the models will be discussed. Due to the use of two different statistical tests, different methods are required to validate the overall quality of the models.
Model McFadden R2 Cox & Snell R2 Nagelkerke R2
LR-test AIC BIC TDL 1 0.0019 0.0014 0.0027 0.0001 13505 13551 1.294
NULL 13556 13564
Table 4: Quality measures model 1
31 of parameters used in the model. The lower AIC of model 1 indicates it is performing better than the null model. The lower BIC score is comparable with the AIC and shows an
improvement as well (Burnham & Anderson, 2004). The top decile lift (TDL) indicates how much better the model is performing compared to the null model that corresponds to the lift value of 1 (Duncan, 2011). The hit rate has been neglected as a validity measure since it isn’t a valid measure for this dataset. A high hit rate might indicate that the model is performing well in predicting the 0’s, while a vast majority of the dataset contains the value “0” which indicate that no purchase was made. The hit rate thus might indicate it is predicting most cases correctly whereas it is only predicting the 0’s correctly but has a poor performance in
predicting the 1’s.
Table 5: Quality measures models 2, 2.2, 2.3, 3, 3.2 and 3.3
The remaining linear regression models required a different set of quality measures as showed in table 5. One of the most important measures for the fit of the model is the coefficient of determination or R2 (Leeflang et al., 2015). The R2 measures the total variance in the criterion variable explained by the model, which is poorly low with both initial developed models 2 and 3. Regarding these results the models are doing a poor job in explaining the sales and orientation of consumers. This implies that the current variables do not sufficiently explain the variance of sales and orientation and there might be other variables which explain this better, which haven’t been included in these models. Models 2.2, 2.3, 3.2 and 3.3 are built upon from the initial models and have a substantial higher R2. This indicates that the added interaction and lagged effects are better predictors of sales and orientation of consumers. The adjusted R2 penalizes for added variables that do not sufficiently contribute in explaining the dependent variable. With all models this adjusted R2 isn’t much lower compared to the normal
Model R2 Adjusted R2 F-statistic P-value
32 R2 indicating that the added variables don’t improve the model just by chance. The final quality measure is the p-value which shows that all the models are significant on a 0.05 level.
4.3 Main effects
4.3.1 Model 1
Table 7 shows the outcome of the binary logistic regression model. The only two variables that significantly affect sales are sunshine (p = 0.0115) and temperature (p = 0.0005). Whereas sunshine has a positive (B = 0.0015) effect on sales, the effect of temperature is negative (B = -0.0015). This concludes that with more sunshine, people tend to purchase more tourist products which rejects hypotheses 1. Concerning the effect of temperature, which causes people to purchase less is in line with hypotheses 2 and thus accepted. No significant evidence has been found for the effect of rainfall, which causes hypothesis 3 to be rejected as well. Both the control variables gross income (p = 1.38e-09) and life stage (p = 0.0396) show significant results indicating that these affects the sales of tourism products.
Coefficients Estimate Std.error Z-value P-value Odds ratio Intercept -2.2316 0.0812 -27.482 < 2e-16 0.1073565 Sunshine 0.0015 0.0006 2.528 0.0115 1.0015512 Temperature -0.0015 0.0004 -3.485 0.0005 0.9985387 Rainfall 0.0000 0.0005 0.086 0.9312 1.0000414 Gross-income 0.0646 0.0107 6.058 1.38e-09 1.0183707 Lifestage 0.0182 0.0088 2.058 0.0396 1.0667191
Table 6: Outcome model 1
The coefficients of the outcome are hard to interpret since they reflect the change in the log odds. To get better and more interpretable insights these coefficients can be converted to odds ratios. The sunshine’s odds ratio indicates that with every 0.1 hour increase of sunshine, the odds of getting a sale increases times 1.00155. In the case of temperature an increase of 0.1 degrees Celsius decreases the odds of sales times 0.9985387.
Some literature however states that these odds ratios cannot be interpreted that
33 interpretation of the results is the use of marginal effects which is presented in table 7. The sunshine’s marginal effect of 0.000168 indicates that with an increase of 0.1 hour of sunshine the probability of getting a sale increases with 0.0168 percentage points. Temperature has an opposite effect where an increase of 0.1 degrees Celsius, the probability of getting a sale decreases with -0.0158 percentage points.
Marginal effects dF/dx Std.error Z-value P-value Sunshine 1.6763e-04 6.6262e-05 2.5298 0.0114139 Temperature -1.5815e-04 4.5307e-05 -3.4906 0.0004818 Rainfall 4.4775e-06 5.1874e-05 0.0863 0.9312160 Gross-income 6.9848e-03 1.1471e-03 2.0592 0.0394716 Lifestage 1.9687e-03 9.5602e-04 6.0893 1.134e-09
Table 7: Marginal effects model 1
4.3.2 Model 2
Model 2 as stated before investigates the same relation as model 1 but with different results as could be observed in table 8. Similarly, to previous results, the effect of temperature is
significant (p = 0.0003), with an elasticity of -0.3772. A remarkable difference however is that this time not sunshine but rainfall has a significant (p = 0.0052) effect. The size of the effect (B = 0.2608) is like temperature rather large. Regarding the significant effects of temperature and rainfall hypothesis 2 and 3 are accepted, whereas hypothesis 1 is not. This concludes that people tend to purchase more on those days with more rainfall and a lower temperature. The exact effect is that with every mm rainfall more, sales increases with 0.2608 and with every increase of 0.1 degrees Celsius, sales degreases with 0.3772.
Model 2 was able to accept more hypothesis compared to model 1 and found large elasticities which was more in line with current literature. The odds ratios and marginal effects of the logistic regression can’t be simply compared to the elasticities of a linear regression but it seems quite clear that the found effects are larger in model 2. As a result, it seems logical to use this model to build on for additional interaction and lagged effects in a later stage.
Coefficients Estimate Std.error Z-value P-value
Intercept 270.1365 15.3276 17.6242 <2e-16
34
Rainfall 0.2608 0.0933 2.7949 0.0052
Temperature -0.3772 0.1037 -3.6363 0.0003
Table 8: Outcome model 2
4.3.3 Model 3
The outcome of model 3 which is shown in table 9, shows the results of the second main effect which concerns the orientation. Results show that in this case only rainfall has a significant effect (p = 0.0012), which is positive (B = 0.1707) as well. As a conclusion, consumers tend to orientate for tourist products more when rainfall increases, which supports hypotheses 6. Because of a lack of significance hypothesis 4 and 5 are rejected. The effect of rainfall is that with every increase of 1 mm, the number of occurred touchpoints (e.g.,
accommodations website) increases with 0.1707.
Coefficients Estimate Std.error Z-value P-value
Intercept 612.6666 38.8037 16.0208 <2e-16
Sunshine 0.0398 0.2828 0.1407 0.8881
Rainfall 0.1707 0.2073 3.2355 0.0012
Temperature 0.0492 0.2764 0.1780 0.8587
Table 9: Outcome model 3
4.4 Moderation effects
4.4.1 Model 2.2
To investigate whether advertisement has a moderating effect on sales and orientation interaction effects have been added to existing models. The effect of advertisement itself shows significant (p = 2.836e-12) results, where every increase of an advertisement touchpoint (e.g., banners) sales increases with 5.1586. Considering the interaction effects, there are no significant results found on a 0.05 level. The interaction effect of advertisement and sunshine however is significant (p = 0.0626) on a 0.10 level. The original assumption was that the sunshine had a negative impact on sales and this would be enforced with
35 Considering this, hypothesis 7 can’t be accepted nor hypothesis 8 and 9 because of a lack of significance.
Coefficients Estimate Std.error Z-value P-value Intercept 195.5272 17.1980 11.3692 <2e-16 Sunshine 0.2872 0.1367 2.102 0.0356 Rainfall 0.0988 0.1064 0.9286 0.3531 Temperature -0.2018 0.1247 -1.6182 0.1056 Advertisement 5.1586 0.7385 6.9856 2.836e-12 Adv:Sunshine -0.0141 0.0076 -1.8619 0.0626 Adv:Rainfall 0.0027 0.0057 0.4801 0.6311 Adv:Temperature -0.0055 0.0051 -1.0624 0.2880
Table 10: Outcome model 2.2
4.4.2 Model 3.2
The second moderating effect is that of advertisement on the relation between weather and orientation. Once again as observable in table 11, there is only one interaction effect
significant (p = 0.0638) on a 0.10 level which is that of advertisement with rainfall. The effect is positive (B = 0.0191) as expected, which shows that people tend to orientate more with increasing rainfall which is enforced through advertisement. However, since no significant effect on a 0.05 level was found, hypothesis 10, 11 and 12 are all rejected.
36
Table 11: Outcome model 3.2
4.5 Lagged effects
The final assumptions to be tested are the lagged effects of the weather variables. Since the data for this model is aggregated on a daily level the possibilities for the length of the lagged effect is endless. Bigano et al., (2005) found however that temperature has a strong lagged effect on tourism up to five days after which the effect diminishes. The different
meteorological variables might have different ideal lagged effects but within this research it chosen to investigate the lagged effect of five days for all weather variables.
4.5.1 Model 2.3
Table 12 and 13 show the results of the lagged effects on sales and orientation respectively. Concerning sales there only seems to be a lagged effect in the case of sunshine (p = 0.0297). The effect is positive (B = 0.2640) however which isn’t in line with hypothesis 13. Because of this positive effect and the lack of significance of the other delayed weather variables, none of the corresponding hypothesis could be accepted.
These results do give an interesting insight however since the effect of rainfall was more frequently significant compared to sunshine, which is the opposite in the case of delayed effects. It might be possible that especially rainfall has a more direct effect whereas sunshine has more effect on the long term. The effect of the delayed sunshine variable is that with every increase of 0.1 sunshine hours of five days before, current sales increases with 0.2640.
37
Temperature-lag -0.4562 0.1724 -2.6459 0.1581
Table 12: Outcome model 2.3
4.5.2 Model 3.3
Regarding the lagged effects of weather on orientation, only sunshine has a significant (p = 0.0219) lagged effect. Table 13 shows that this effect is even greater (B = 0.5621) than the lagged effect of sunshine on sales (B = 0.2640) as mentioned in previous results. This relatively large effect shows that with every increase of 0.1 hours of sunshine of five days earlier the current sales increases with 0.5621. These findings do not accept any of the corresponding hypothesis however, because of a lack of significance for temperature and rainfall and an opposite expected effect of sunshine.
Coefficients Estimate Std.error Z-value P-value Intercept 485.0302 39.0941 12.4067 <2e-16 Sunshine 0.2157 0.2920 0.7390 0.4599 Rainfall 0.2011 0.2338 0.8601 0.3897 Temperature 0.1445 0.3222 0.4485 0.6538 Advertisement 8.1672 1.2602 6.4807 9.133e-11 Adv:Sunshine -0.0112 0.0139 -0.8060 0.4203 Adv:Rainfall 0.0180 0.0103 1.7369 0.0824 Adv:Temperature -0.0078 0.0100 -0.7802 0.4353 Sunshine-lag 0.5621 0.2453 2.2914 0.0219 Rainfall-lag 0.0349 0.1661 0.2100 0.8336 Temperature-lag 0.0914 0.3268 -0.2796 0.7798
Table 13: Outcome model 3.3
Since there was a substantial number of hypothesis to be tested, table 14 will give a short overview of which of the hypothesis were accepted or rejected.
Hypothesis Accepted / Rejected Hypothesis Accepted / Rejected
38 H2 Accepted H11 Rejected H3 Accepted H12 Rejected H4: Rejected H13 Rejected H5: Rejected H14 Rejected H6: Accepted H15 Rejected H7: Rejected H16 Rejected H8: Rejected H17 Rejected H9: Rejected H18 Rejected
Table 14: Overview accepted and rejected hypothesis
5. Discussion
5.1 Conclusion
This research investigated the influence of weather on sales and orientation of tourism products. Three weather variables that most affect people’s perception of weather was determined by past literature. These three weather variables were sunshine, rainfall and temperature. In addition, a moderating effect with the weather variables has been added to see if this enforces the effect of weather. Since weather might trigger consumers to react to tourism products, this might express itself in a delayed reaction as well. This is why some lagged effects of weather has been examined as well.
39 There could be a numerous amount of reason why models 1 and 2 showed distinctive results. Firstly, the functional form is of relevance which is the representation of the relation between the independent variables and the dependent variables (Leeflang et al., 2015). The logistic regression and linear regression used in this research have distinctive functional forms which could lead to different results. Another possible reason for the different results is the
distinctive level of aggregation of the two models. For model 1 each individual Purchase Id was considered and whether a sale was made within this unique id. The data for model 2 was aggregated on a daily level which presented the total amount of sales for each day. After this modification of the data to be useful for model 2 it is most unlikely that the model would show the exact same results as model 1.
Regarding these differences, model 2 found more significant results which confirms more initially developed hypotheses. On top of that the effects found in the first model were so small that they were redundant, whereas the second model found relative strong effects which is why this model was used for further examination in this research. Coming back to these results, the only effect which wasn’t proven to be significant in model 2 was that of sunshine. This result on first sight seems to be somewhat extraordinary since temperature and rainfall do have a significant impact. It might be the case however that people are more severely affected by temperature and rainfall, because when exposed to these elements, one might argue that the consequences are bigger compared to sunshine. Rainfall and temperature simply requires more drastic measures to avoid getting wet or cold. Sunshine however, has less influence and requires less anticipation. This less severe effect of sunshine might express itself as well in the consumer behavior for tourism products. Due to this more severe effect of rainfall and temperature, people might be more motivated to purchase tourism products. Although this is just a theory it might be a reasonable explanation for the lack of significant findings of sunshine within this research.
When it comes to the main effect of the weather variables on orientation of tourism products, only rainfall showed to have a significant effect, which was positive as expected. This
supports the findings of Lim et al., (2010) who particularly examined the number of
40 either confirm or deny this phenomenon.
A possible explanation however might be that the results are affected by the distribution of the occurred touchpoints within this dataset. A preliminary investigation of the data already showed that accommodations websites was the most visited website. It might be that weather is of less influence on these websites and has a greater effect on for example orientation on flight tickets. The results would be affected in this case since flight tickets touchpoints are barely represented as opposed to accommodations websites in the dataset.
Another possible reason might be that in general the effect of weather is more severe on sales than orientation. People might orientate for tourism products on a regular base and are only slightly affected by weather but are especially convinced for a purchase on those days with unfavorable weather conditions. This could be explained by that both models found at least one significant effect but model 2 resulted in more significant and greater effects.
The moderating effects of advertisement is less supportive to the initially developed assumptions. Regarding sales only a significant interaction effect on a 0.10 level with
sunshine has been found which’s effect was rather small. Looking at the effect on orientation only the interaction with rainfall was significant, once again on a 0.10 level and only with a small elasticity. It is hard to determine why specifically these specific interaction effects were the only significant ones, if considered significant whatsoever.
On the one hand this lack of significant evidence might be caused by the distribution of the different forms of advertisement to which the consumers were exposed within this research. More than 80% of the advertisements that consumers were exposed to was retargeting. A possible explanation thus might be that other forms of advertisement does have a significant interaction effect with weather. Retargeting is a form of advertisement to which consumers are exposed once already browsing on the web, whereas exposure to emails for example could occur at different moments, possibly initiating a chain of events for the consumer. Maybe consumers are triggered to browse for tourism products once they have received an email whereas with retargeting they were already searching for these tourism products.
It might however be concluded that advertisement itself has a positive influence on sales and orientation but does not alter the effect of the weather variables. Consumers might be
triggered to orientate and purchase more due to certain weather conditions and are, because of this higher orientation, more exposed to advertisement. It doesn’t however seem to be the case that advertisement enforces the effect of weather and thus the effect of weather and
41 remarkable that the significant found direct effects of weather in the initially developed
models, become insignificant after the inclusion of interaction effects. Maybe there is no interaction effect but the weather effects require advertisement to be present in order to show significant results. Maybe if no form of advertisement was present at all, no significant weather effects weren’t found either.
The final investigated effects are the delayed effects of weather on sales and orientation. Considering the effects on sales, no significant effect was found for temperature and rainfall, whereas there was for sunshine, which is in line with some of the findings of Falk (2015). Since no significant current effect of sunshine was found in previous models it could be that temperature and rainfall have a direct effect and sunshine has a more delayed effect. The lagged effect of sunshine however was positive which was expected to be negative. Once again it is hard to determine the possible cause of this phenomenon since there is only little literature specifically about lagged domestic sunshine effects on tourism.
A possible explanation for these found effects is due to the size of the delay. It might be that more significant results would be found if the size of the delay would be increased or even decreased. The delayed effect of five days was based on one particular research of Bigano et al., (2005), which isn’t sufficient enough. When different sizes of delayed effects are
investigated more significant evidence could be found.
Another explanation might be that the delayed effects of sunshine were slightly correlated with the current variables of temperature and rainfall. It might that the high amounts of sunshine from five days before didn’t have an effect on sales and orientations of the current day but were actually influenced by the current effect of temperature and rainfall. Maybe there is some sort of pattern within weather were favorable weather situations are followed by unfavorable weather. As a consequence, it might have seemed that sunshine had a delayed effect but this effect is actually captured in the current effect of temperature and rainfall.
5.2 Scientific and managerial implications
42 There is currently only a limited amount of literature which investigates the relation of
domestic weather which causes a so called “push” effect for tourism. Those researches that did, didn’t always make a clear distinction between the different meteorological variables or chose a different set of weather variables. On top of that it can be concluded that there is almost no literature which investigates potential interaction or lagged effects of weather. This is where this research contributes to current literature. Regarding the interaction effects, barely some significant evidence was found. Other forms of advertisements could result in distinctive results however. Different effects that could interact with weather could be thought of as well. There is an endless amount of possibilities to consider for interaction effects which hasn’t been done before.
This research has shown that there is a clear distinction between the effect of the different meteorological variables regarding significance and effect size. In addition, it has been found that temperature and rainfall have a more direct impact whereas sunshine has a more severe effect ones delayed. This as well might be a good starting point to further investigate the delayed effects of weather, with possibly different sizes of this delay.
With the findings in this research in mind it might be interesting to converse consumers on those days with high rainfall or a low temperature i.e., unfavorable weather conditions. As it seems that people orientate and mostly purchase more during these particular days, managers of tourism organizations could seduce potential customers with discounts or other methods on those specific days. Due to the high costs, people only go on holiday a certain amount of times per year or maybe even less (Morgan & Pritchard, 2002). Once they purchased a
tourism product at a certain organization that customer won’t consider a similar purchase for a relatively long time. This makes it essential for tourism organizations to react on the right moment which, as the results show, are on those days with unfavorable weather conditions i.e., high rainfall and low temperature.
Besides the importance of targeting at the right time, weather can also be helpful for determining which customers to target. Tourist organizations should focus on those
43 5.3 Limitations and further research
Every research could potentially be improved including this particular research. The next section will be about the shortcomings and potential improvements for further research.
Firstly, the demographic information about the consumers in the dataset could have been more detailed in some cases. For this research it was essential to determine the place of residence as detailed as possible to match the weather situation during the time period of the dataset. Unfortunately, the place of residence was only divided in four sections. This caused it to be necessary to work with “general” weather variables which in some cases wouldn’t match with the actual weather conditions in that certain place. Especially with the weather variable rainfall it is not uncommon that precipitation occurs short and locally. In further research it might be beneficial to use weather data which is more detailed, to examine the effect of weather even better.
A second limitation concerns the data about the sales and orientation of consumers. Out of the 17687 sessions that were conducted by consumers there were only 2246 sales, out of which only 108 purchases were made at the focal company. It is well known that a large portion of the consumers do not converse, but more significant effects might have been found if the data was larger including more sales. Especially with the low amount of focal purchases, it is hard to determine a difference between these sales and those made at the competitors, which is why this was neglected in this research. Further research should use a larger and richer dataset to obtain even more significant results.
Considering the moderation effect of advertising, the distribution of advertisements to which consumers were exposed to should be more equal. Even though the data contains five
different categories of advertisements, more than eighty percent is exposed to one particular form of advertising which is retargeting. Since there haven’t been found significant
interaction effects with advertising this might be a consequence of retargeting as stated in the discussion. Further research should have a more distributed occurrence of the advertisement categories.
