YouTube pre-roll exposure effect on purchasing choice

YouTube pre-roll exposure effect on

purchasing choice

A logit regression model

Minh Nhat Vu

1

YouTube pre-roll exposure

effect on purchasing choice

A logit regression model

Minh Nhat Vu – Department of Economics and Business

Master thesis – Track: Marketing Intelligence

Completion date: 06/20/2016

Email:

vu.nhat.minh@student.rug.nl

Student number: S2759101

2

Abstract

3

Table of Contents

I. Introduction ... 4

II. Theoretical framework ... 6

III. Research design ... 10

3.1 Methodology ... 10

3.1.1 The binary logit regression model ... 10

3.1.2 Loyalty measurement ... 12 3.1.3 Quality of fit ... 12 3.2 Data collection ... 13 3.2.1 Data choice ... 13 3.2.2 Descriptive ... 14 3.3 Specification ... 16 IV. Estimation ... 22 4.1 Results ... 22 4.2 Interpretation ... 22 V. Validation ... 24 5.1 Coefficients ... 24 5.2 Hit rate ... 25 5.3 Chi Square ... 27 5.4 Psuedo R Square ... 27 5.5 Multicollinearity ... 28 5.6 Robustness test ... 28 VI. Discussion ... 29

VII. Conclusions and recommendations ... 31

VIII. References ... 33

APPENDIX 1 ... 0

APPENDIX 2 ... 0

APPENDIX 3 ... 0

4

I. Introduction

There is one big similarity between YouTube pre-roll (YT) advertising and Television (TV) advertising and that is viewers’ eternal hatreds towards them (Finan, 2015). This paper will use YT pre-roll and YT interchangeably. Imagine going to watch your favorite show on TV or YouTube and having to wait for 30 seconds or 3 seconds before you can continue. YouTube pre-roll was introduced in December of 2015 and despite its young age, it is not very original. In fact, it is another version of TV advertising as they are both instream ads that are shown before a program a consumer wanted to watch; the differences are: YouTube’s shorter length and the viewers’ ability to skip after 3 seconds.

Even though YouTube and TV advertising have not reached the threshold of perfect balance between annoying and engaging, they work. The elimination of TV advertising had been proven to decrease the aggregate sales of all carbonated soft drinks (CSDs), completely wipe out competitive advantages of brand X and move consumer’s consumption towards fruit juice, bottled water, and milk (Lopez, 2015). An exception to this study would be a meta-analysis done by Lodish (1994) that tested 389 experiments to capture the effects of TV advertising on sales; the result was disappointing with no relation between the amounts of TV advertising and sales impact. Perhaps, such result was due to their unit of measurement, which was advertising weight; as suggested by Lodish in 2001. His renewed study in 2001 showed a small carry over effect of TV advertising campaign from last year on sales of the following year (Lodish, 2001).

5

because it argues that TV and YouTube pre-roll are very similar. If TV advertising has an effect on sales and purchasing choice, then Youtube pre-roll should also has an effect as well.

YouTube pre-roll is an important phenomenon worthy of academic research because of its ability to generate recall, brand awareness, and purchase intent; which are 47 percent, 32 percent, and 44 percent; respectively (Nielsen n.d and Finan, 2015). Additionally, it is also interesting to look into because of its ability to affect consumer behavior (Lopez, 2015) despite their annoyance towards it (Finan, 2015). Another reason YouTube exposure is the focus of this research is its lack of academic attention. This could stem from the scarcity of single source data available for analysis.

It was advised to use household level and daily level when measuring TV advertising effect to avoid losing information (Tellis, 1995). YouTube is similar to TV advertising in nature and YT exposure should vary by HHs and days of the week. Therefore, aggregating YT exposure data by weeks would affect the estimated coefficients (Russell 1988). This emphasizes the importance of using data at daily level to evaluate YT advertising effect in this research paper. Fortunately, this research paper was the first, to the author’s knowledge, to use single source data provided by Gfk that contains daily exposure of YouTube over 90 days. Much of the other works on YouTube advertising had been presented in newspaper articles such as the ones by Anna Richardson quoting from Nielsen (2011) or Baatchi quoting from T-Mobile (2009), controlled experiments (Li, 2015), logical reasoning (Postigo, 2016) and a case study (Bauwel, 2014). However, no academic studies have presented any proven results regarding YouTube from a large data set of real customer activities.

Such research question is also important for managers because their spending on online video advertising increased exponentially as ad view volume increased 47 percent between 2011 and 2012; which implies that users are watching more video ads online more than ever (Li, 2016). However, this number is still a lot lower than TV advertising; YT spending is 12000 dollars (eMarketer, 2016) and TV advertising is 185 billion dollars. It is paramount to let managers know whether it is worth it to spend much money on YouTube pre-roll advertising when its nature is similar to TV ads’. And whether the large amount of money spent on TV is worth it.

6

II. Theoretical framework

TV advertising exposure is the number of times a consumer watches brand X advertising on TV. As brand X belongs to the carbonated soft drink market (CDs), which is quite matured compared to the other markets, any types of advertising would serve to persuade the consumers to buy rather than to inform them of the products. This explained why the elimination of all TV advertising effects on CSDs resulted in a decrease in aggregate CSD sales because consumers would consume other beverages such as fruit juice, bottled water and milk. TV advertising helped brand X capture market shares of the competitors’ by reinforcing loyalty and building brand equity. In other words, when the competitors advertise more, the firms will sell less and vice versa (Erdem and Sun, 2002). Therefore, Lopez concluded that the effect of TV ad for CSDs is important in keeping CSD sales by retaining customer loyalty, market share and demand for that product (Lopez, 2015). This means that without TV advertising, brand X would lose their market share to its competitors as their customers would buy less of brand X and hence, would not opt for brand X as their brand choice; supporting the argument that TV advertising leads to positive effects on consumers buying brand X as their CSD brand.

An older research by Flemming Hansen effectively showed a positive relationship between TV exposure and purchase probability by presenting one unit increase in TV advertising leads to 3.4 percent increase in purchase probability of Kelloggs Cornflakes (Hansen, 2002). One explanation for such change in behavior could be explained by changes in brain activities. One interesting study gave insights into the

Consumer

purchasing choice

Measured by?

Brand loyalty

- Preference of one household for one brand

7

neuroscience of consumer choice suggested that activities at the prefrontal sites during a TV commercial influence consumer choice (Silberstein, 2008). Of the number of consumers included in the study, 18 with high prefrontal activities on the left hemisphere during TV commercial switched to the advertised jam instead of normally choosing the competitor’s jam. Of all the participants, 111 opted for the usual competitor choice but these people did not have high prefrontal activities on the left hemisphere during TV ads exposure. Even though their experiment only yielded positive TV exposure results for 18 participants with high prefrontal activities on the left hemisphere, this outcome still leads to the conclusion that TV advertising has a small effect on consumer choice.

The above conclusions from three studies were challenged by another. One study by Tellis concluded that even though firms spend 250 million dollars on advertising, it still has a weak effect on brand choices in the laundry detergent market (Tellis, 1995). He argued that the evidence of brand loyalty and high advertising motivate brand choice are flawed by incorrect testing or misinterpretation of causality. Additionally, firms spending a lot of money on TV advertising account for trying to meet their competitors advertising to gain more market shares without evidence of increasing the probability of brand choice.

Differences in findings stemmed from the discrepancies in the type of data provided to each researcher. Lopez used data at household (HH) levels aggregated at monthly observations and market areas level (Atlanta, Detroit etc.). The market in his analysis is defined as a combination of monthly and market level (2015). In Tellis’s work, his analysis is at the disaggregate HH daily level at which consumer decisions were made. He also analyzed his data at the aggregate level and found that consumer choices are increasingly sensitive to TV exposure with temporal and cross-sectional aggregation (Tellis, 1995). As my data is at the disaggregate level I would like to test Tellis’s conclusion using the hypothesis of Flemming Hansen. There is a positive relationship between TV exposure and purchase probability of Kelloggs Cornflakes (Hansen, 2002). Hypothesis one is as follow.

H1: TV advertising exposure has a positive effect on purchase probability.

8

This option made YouTube stand out from TV ads because anyone can skip ads after three seconds if they decide they do not want to watch the advertising. According to YouTube themselves, they are able to reach 18-34-year-olds more than any other networks. Google gave an example of Broadway’s Cinderella play enjoying 12% increase in ticket sales after their launch of YouTube pre-roll campaign. They also stated that YouTube influences more purchase decisions than TV advertising. One survey conducted by YouTube yielded a percentage of influence of 78% on beauty products, 61% on smartphones and 73% on automotive vehicles (Google Think Insights, 2014). Despite such large percentages, one still wonders the accuracy of simply asking respondents whether their purchases were influenced by YouTube.

In other words, surveying respondents to get the effect of YouTube on their purchase behavior might not be accurate. Simply asking is not accurate but it is quite hard to grasp the extent of their causal relationship as there had not been any academic articles studying the effects on ads skipping on YouTube. The closest academic study I found was of digital video recorder (DVR) TV ads skipping. This study found no statistical evidence for the skipping effect of DVR on purchase behavior. They attributed the lack of DVR effect to little skipping behavior among the sample group. Even for HHs that had the highest DVR usage, no effect was found. However, this study admitted that TV advertising is not very targeted and suggested that ads skipping might have a positive effect on sales if the messages were targeted more precisely (Bronnenberg, 2010).

YouTube pre-roll is a type of online marketing tool which offers more targeted messages. Unlike other types of YouTube videos, which are posted by YouTubers themselves; YouTube pre-rolls are posted by advertisers and are popular because they enable them to target viewers who have searched for their products and only pay when viewers decided to watch the videos (those who do not skip ad) instead of paying per impressions (those who skip ad) (yinc marketing, 2016). YT pre-roll has the ability to communicate what the viewers would be interested in. This makes it more appealing than non-targeted ads. Even though there had not been any research on targeted YT pre-roll, there were studies on targeted banner ads. One of those resulted in yielding the highest returns and positive effect on purchase probability (Manchanda, 2006). Additionally, YouTube pre-roll advertising increases market share. One campaign from Pepsi claimed to increase channel subscribers from 4000 to over 66000, expanded market share and achieve 50 million views. In line with the TV advertising, increases in market share lead to higher probability of brand choice. Therefore, YouTube pre-roll exposures will lead to higher probability of brand choice. Hence, hypothesis two is:

9

YouTube pre-roll and TV advertising are different from each other in certain ways (Jensen, 2008). A study by Jensen, which surveyed 200 companies in 2008, concluded that offline channel is more preferred but when it comes to financial reasons online is more preferred. YouTube is definitely cheaper than TV which is the first difference between the two. If YouTube were proven to have an effect on purchase probability then, with its cheaper price, it should be more appealing to managers.

The second point of difference between the two types of marketing tools is their audience scope. YouTube is able to target a much younger audience base; those who would not otherwise be exposed to TV advertising at all. The result of a Cadbury Dairy Milk campaign on YouTube saw that 69% of its audiences from 15 to 34-year-olds were reached with YouTube. These people would never have been reached had YT not been used (Richardson, 2011). Young people spend more time online watching YouTube because this gives them more control of their program choice. They are also very tech savvy, which made it easy to seek out programs that interest them on YouTube instead of having to wait for a program to be aired on TV. TV advertising seems to be more popular among the older generations. Adults spend more time watching TV as they get older (Nielsen, 2009). This is because they have more time due to retirement and loss of social contacts. They are also bad with technology and have negative attitudes towards the internet (Teo, 2011). For these reasons, these differences lead to hypothesis three.

H3: YouTube pre-roll advertising has a stronger effect on the younger audience base (1) whereas TV advertising effect is more prominent among the older generations (2).

10

H4: Loyalty has a moderation effect on the relationship between YouTube advertising and consumer purchasing probability.

III. Research design

3.1 Methodology

3.1.1 The binary logit regression model

One popular model to measure advertising effect is the binary logistic regression model or the logit model. A logit model is utilized to explain brand choice probability as a function of purchase explanatory variables (Jones, 1980). Various authors have used this type of model in marketing research. Hlavac and Little used this type of model to represent the probability that a car is purchased by an automobile buyer (1966), Punj and Staelin utilized the model to describe students’ choice of business school (1978), Hansen employed it to estimate the effect of TV exposure and promotion on purchasing choice (2002).

The logit model makes sure that the response function is “curvilinear with asymptotes at both zero and one, which naturally meets the original constraints when the dependent variable is binary (Hansen, 2002). The logit model is also a good model to use in the analysis because advertising response is S-shaped. Meaning many exposures can lead to saturation and few exposures can be lost in noise. The logit is best for capturing an S-shaped response of advertising exposure on sales (Tellis, 1995).

In this particular study, the logit model was used to estimate the probability that a HH purchased brand X given the main explanatory variables are YT exposure, TV exposure and brand loyalty. The dependent variable is binary (0 if brand X is not purchased and 1 if it is). The original sample consisted of 10703 panelists (Household or HHs). The dependent variable is observed at daily level over 90 days from December 30, 2013 to March 29, 2014. It is the recorded times each HH buys brand X daily.

The data provided by Gfk is at disaggregate level. It was suggested by Tellis that disaggregated daily level provides more consistent and less biased estimates. He wrote: "When the focus is the effects of TV advertising exposures the most appropriate data levels is individual household level and daily level” (Tellis, 1995). TV and YT exposures vary by programs and days of the week. If the data were aggregated, the distribution of these variables would be affected, which can affect the estimated coefficients (Russell, 1988).

11 The basic linear model:

𝑌𝑖𝑡 = 𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑌𝑇𝑥𝐿𝑂𝑌𝑖𝑡+ 𝜀𝑖𝑡 (i = 1,… ,N) Where: 𝑌𝑖𝑡 = { 1 𝑖𝑓 ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖 𝑏𝑢𝑦𝑠 𝐶𝐶 𝑖𝑛 𝑎 𝑝𝑎𝑟𝑡𝑖𝑐𝑢𝑙𝑎𝑟 𝑑𝑎𝑦 0 𝑖𝑓 ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖 𝑑𝑜𝑒𝑠𝑛′𝑡𝑏𝑢𝑦 𝐶𝐶 𝑖𝑛 𝑎 𝑝𝑎𝑟𝑡𝑖𝑐𝑢𝑙𝑎𝑟 𝑑𝑎𝑦 𝜀𝑖𝑡 = the unobserved value of a random disturbance term at day t for HH i

𝑌𝑇𝑖𝑡= YouTube pre-roll advertising exposure at time t for HH i

𝑇𝑉𝑖𝑡=TV advertising exposure at time t for HH i

𝐿𝑂𝑌𝑖𝑡=Loyalty of consumer i over the whole sample period for HH i

𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡= moderating effects of loyalty on the relationship of YT and purchasing choice for HH i

P(Yit = 1) = πit was defined as the probability that brand X would be purchased at day t and by definition

P(Yit = 0) = 1 - πit. The probability depends on the intercept, the slope parameter and the independent

variables (Leeflang, 2015). Instead of putting emphasize on the precise value of 𝑌𝑖𝑡, the author focused on

the probability that Yit = 1 given 𝑌𝑇𝑖𝑡, 𝑇𝑉𝑖𝑡, 𝐿𝑂𝑌𝑖𝑡, 𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡 and 𝜀𝑖𝑡.

π𝑖𝑡 = P(Y𝑖𝑡 = 1|𝑌𝑇𝑖𝑡, 𝑇𝑉𝑖𝑡, 𝐿𝑂𝑌𝑖𝑡, 𝑌𝑇𝑥𝐿𝑂𝑌𝑖𝑡, 𝜀𝑖𝑡)

In binomial models, these probabilities are not observed. Only the binary of the final choices of each HHs are observed. In choice modeling, these probabilities are estimated by Maximum Likelihood. High probability predicts that Yit = 1 (will be bought). Basing on the micro-economic theory, choice is based

on utility UB and UN (buy vs. no buy). Customer i will buy if the utility of buying is higher than the utility

of no buy (UB > UN) and vice versa.

The probability of observing Yit = 1 given all the independent variables is equal to the cumulative

distribution function of 𝜀𝑖𝑡 evaluated at 𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡.

The logit model:

𝐹{𝑌𝑖𝑡} =

exp (𝛽0+ 𝛽1𝑌𝑇𝑖𝑡 + 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡)

1 + exp (𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡)

12

3.1.2 Loyalty measurement

Loyalty was not provided in this dataset. However, it was calculated using the share of category requirement (SCR) measurement. SCR is the HH’s shares of category requirements for brand i during time T (range from a month to a year) (Bhattacharya 1996).

𝑆𝐶𝑅𝑖= 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑠 𝑜𝑓 𝐵𝑟𝑎𝑛𝑑_𝑖 𝑇𝑜𝑡𝑎𝑙 𝑐𝑎𝑡𝑒𝑔𝑜𝑟𝑦 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑠 𝑜𝑓 𝑏𝑢𝑦𝑒𝑟𝑠 𝑜𝑓 𝐵𝑟𝑎𝑛𝑑_𝑖 (Jung, 2010) In this case: 𝑆𝐶𝑅𝑖𝑡 = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑠 𝑜𝑓 𝐶𝐶𝑖𝑡 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑠 𝑜𝑓 𝐶𝐶𝑖𝑡+ 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑠 𝑜𝑓 𝐶𝑜𝑚𝑝𝑒𝑡𝑖𝑡𝑜𝑟 𝐵𝑟𝑎𝑛𝑑𝑖𝑡

Basing on that I calculated loyalty (aggregated at HH level) by: count of all brand X purchased during the entire observation period Dec-30 to March-29 for every single HH divided by the count of all brand X plus competitors purchases in the same observation period.

Even though YT and TV vary by days of the week, loyalty is a long process and varies by HHs, not by day. A HH can be loyal to brand X, somewhat loyal or not loyal but cannot be said to be loyal because that HH bought from brand X for one day instead of the competitors’. Therefore, loyalty had to be assessed at the HH level throughout the entire period instead of the daily level. Additionally, SCR is usually calculated in time t ranging from a month to a year (Bhattacharya 1996); therefore, this research also used that data range.

3.1.3 Quality of fit

Measures of quality of fit are important because they guide model specification and ensure that the independent variables explain fully the dependent variable. Guadagni and Little suggested four ways to assess model fit (2008).

U2 for model. A logit model simply predicts probabilities, which must be compared with actual choices. U2 is defined as the “fraction of uncertainty (entropy) empirically explained by the calibrated model relative to the prior distribution of choice probabilities”. This constitutes a null model that represents the maximum uncertainty for the situation at hand. A model under assessment that produces the same uncertainty as the null model explains nothing new and hence, has U2 = 0. A model that explains everything perfectly has a U2 = 1. This model would give a perfect prediction.

U2 is equal to the McFadden’s likelihood ratio index. Which is:

13

𝐿(𝑋) = the log-likelihood of the full model with the explanatory variables (the full model) 𝐿0 = the log likelihood of the null model

Chi-squared tests of model significance. This test helps assess whether adding a parameter or a set of parameters is worthwhile.

Predictive validity. Split the data into two random samples. The first sample is used to train the parameters. This is called the estimation sample. The second sample is called the holdout or validation sample and is used to test the predictive validity. Loyalty is calculated per HHs throughout the whole observation period. Therefore, when the dataset is randomly split into two, the loyalty level per HHs should stay the same.

3.2 Data collection

3.2.1 Data choice

The author worked with panel data collected from actual records of TV and YouTube exposures of brand X campaign for 10703 households over 13 weeks. Actual recollection of consumer viewing was recorded. Both TV and YouTube advertising exposures were recorded via an app on a smartphone, which was situated next to the TV and computer screen to record brand X ad exposures for 13 weeks. To ensure the data quality and that each household watched the ads during air time; Gfk occasionally called the consumers to check if someone in the household was watching them.

TV advertising exposure is the number of times a consumer watches brand X advertising on TV. People with a special DVR device can choose to skip ads but most people without this device are forced to watch the TV ads. The scope of this study only takes into consideration those without the DVR device as I assumed that if people can skip ads, then Gfk would not have been able to record TV ads exposure. On the other hand, with YouTube, everyone has the ability to skip ads without having to download any external program.

The original sample consisted of 10703 panelists (Household or HH). Only those who signed up to be active members of both TV panel and online panel (YouTube and website participants) passive measurement were considered. In other words, those who did not participate as active members were not being measured and hence, might look as if they had zero contacts but, in reality, they might have had more. Therefore, to ensure accurate measurement, it was best to only include those who were active in both panels; which equated to 1304 as the sample size.

14

suggested by Guadagni and Little & Tellis and Weiss to not consider those who bought less than 3 times. Therefore, only those who bought less than 3 times were included in the analysis. This left 117 HHs which was not a large number but Guadagni and Tellis both used 100 HHs in their sample size so I believe this number was appropriate.

Of the 117 HHs left, brand X was bought minimum 4 times and maximum 30 times. 93.2% of all the HHs (109 HHs) were not exposed to YT pre-roll and 6.8% were exposed to YT Preroll. Of those exposed, .9 percent (1 person) was exposed to YT pre-roll 7 times which was the maximum time. Looking at exposure to TV, more people were exposed to TV advertising than YT pre-roll. Only 14 HHs (2%) were not exposed to TV ads at all. One HH was exposed to TV a maximum of 28 times, which was a lot more than YT pre-roll exposure.

For the purpose of prediction and validation, the data were split into two random samples by 70% and 30%. This split was assumed to not affect the data in any way because the dependent variable was recorded at daily level and so were the independent variables. Therefore, it is not time dependent; the brand X purchase is not dependent on the lag effect of one day on the other. The first sample was used to train the parameters. This was called the estimation sample. The second sample was called the holdout or validation sample and was used to quantify the predictive validity. Then the ROC curves for the two samples were created. If the area under the curve for the testing sample is not too much smaller than for the training sample, the model can be said to be valid.

3.2.2 Descriptive

Figure 1 – Advertising exposure by week

As can be seen from the graph, TV exposure gradually decreased from week 1 until week 2 and increased a bit from week 4 to week 5 but then sharply declined to 0 in week 7 and 8. This was a bit strange as it

0 2 4 6 8 0 50 100 150 201401 201402 201403 201404 201405 201406 201407 201408 201409 201410 201411 201412 201413 C o u n t o f H H s Week number

Advertising exposure per week

15

stayed at 0 from week 7 to week 8 then rocketed to 130 times for all HHs only to decrease again in the later weeks. Despite the sudden fluctuations of the TV exposure, the sum of brand X bought seemed not to be affected by it as it gradually increased from week 6 to week 10. Looking at brand X purchase line, one can see two small peaks at week 3 and week 4. In week 10 there was a large peak, which was a coincident with the increase of TV contact from week 8 to 9 if the lag effect were to be taken into consideration. However, just from observation, it did not seem likely because there was zero TV advertising in week 7 and 8 but the amount of brand X bought still increased.

It is interesting to see that YouTube exposure started in week 5 for those who signed up to be in both TV and online panel. Before that they were not exposed to any YouTube advertising. Also, YT pre-roll advertising was less than TV advertising as HHs could be exposed to TV up to 4 times whereas YT a maximum amount of 2 times. Looking at the amount of brand X bought throughout time, it can be seen from the graph that the period from week 1 to week 6, there was a peak in the frequency of HH buying brand X. After YouTube advertising was introduced, the frequencies of brand X bought were steadily increasing. It also started to decrease in a similar pattern as YouTube exposure’s. Also, it was worth it to mention that in 2014, brand X really focused on YouTube as their main channel to communicate their Share a coke campaign which drove massive fan interactions and shares (Hitz, 2014). This could have been accounted for the increase in the brand X bought and TV exposure sudden drop to 0 for two weeks.

Figure 2 - Advertising exposure per day of the month

Figure 2 shows the advertising exposure per day of the month. The scale was at all HHs aggregated exposures on that day. As can be seen in figure 2, YT pre-roll exposure was lower than TV advertising

0 0.5 1 1.5 2 2.5 3 3.5 0 10 20 30 40 50 60 70 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 Co u n t o f al l H H s

Day per month

Advertising exposure

16

exposure. Meaning, less HHs saw YT than TV. However, starting from week 13, with every peak of YT pre-roll, a peak in the brand X purchasing count followed. From week 1 to 12, it seems as if YT pre-roll advertising strategy had not been figured out so the exposures were not very stable. In the same period, TV advertising was experiencing a steep decrease in exposures of HHs from 70 to 15 HHs. This followed by a more stable period but the peaks of TV and brand X purchase did not fit each other, despite there were more TV advertising than YT pre-roll advertising.

Two graphs; one at daily level and one at weekly level were used to look into the results of Tellis (1995). He stated that daily level can provide more information and indeed it provided more information. However, the conclusion of Figure 2 and Figure 1 was somewhat similar.

3.3 Specification

The author was interested in one thing: customer brand X purchasing choice (Yes = purchase/No = no purchase). Therefore, it was logical to choose the variable brand X purchasing choice as the dependent variable for the model.

Dependent variable: Yit = Purit = Choice of brand X purchase for HH i at day t (Yes/No)

The interested independent variables, as described in the conceptual models were (More information can be found in Table 3):

𝑌𝑇𝑖𝑡= YouTube pre-roll advertising exposure at time t for HH i

𝑇𝑉𝑖𝑡= TV advertising exposure at time t for HH i

𝐿𝑂𝑌𝑖𝑡= Loyalty of consumer i over the whole sample period for HH i

𝑌𝑇𝑖𝑡𝑥𝐿𝑂𝑌𝑖𝑡= moderating effects of loyalty on the relationship of TV and purchasing choice for HH i

Full model. In order to test out the data, I decided to create seven models to explain the dependent variable. In the first model, which I called the full model, I included all the variables available in the dataset then looked at the Variables not in the equation table to see which variables could be excluded and which variables would explain the dependent variable well. I expected to have redundancies and correlation between the variables and of course I found an error in the data which led to invalidity of the model. Hence, this model with every variable in the dataset was not recorded.

pre-17

roll advertisement would have an effect on their brand X purchasing choice (the variables can be found in Table 1).

Model 1. There were many variables that fall into the Demographic category shown in Table 1 – Model 1. This model turned out well because there was no error in estimation. However, in assessing whether this was a good model or not; it did not include all the important phenomena in the model therefore it is simple but not complete. Due to this, I decided to take only the significant variable in Variables not in the

Equation table and the Variables in the Equation table – Income. Only income can explain the dependent

well among all the demographic variables; and not too well that, by itself, it can explain the dependent variable.

Model 2. In composing model 2, I first included all the relational characteristic variables. However, the inclusion of the entire set of relational characteristic variables made the model invalid. An error was found in the Iteration history table that leads to the final solution could not be found. I tried expanding the maximum iteration to 200 up to 800 but no final solution was found. I suspected the error was caused by the inclusion of many variables that can explain the dependent variable by themselves. Variables such as Volume total could explain the dependent too well because they were directly related to the dependent variable. Therefore, I decided to run the model again with only variables I argued to have an effect in the conceptual model (as seen in Table 1) and the model had no error in estimation. However, it was too simple and is not complete. Its Cox&Snell R Square is 0 and Nagelkerke R Square is 0.01, which is 0. Therefore, more variables needed to be added to this model.

Model 3. This model was created with all the time variables in the dataset. Similar to the full model and Model 2, Model 3 was invalid so I decided to take out variables that were insignificant in the Variables

not in the equation table and those caused errors in the Iteration history. The only variable left was

Weekday, which made sense because the dependent variable was brand X purchase by different days of the week and was measured on a daily level independent of time. This model was built to explore the data and the relationship between the predictors and the dependent variable. Hence, model 4 was built.

Model 4. Then, I took all the significant variables from Model 2 and 3 and fed those into model 4. As a set of variables, they all yielded significant results and their pseudo R squares were relatively high compared to all other models RC&S = .023 and RN = .053 but this model had yet possessed all the

18

Model 5. Model 5 has loyalty as a moderator for the relationship between YouTube pre-roll and purchasing behavior resulted in an error that questioned the validity of the results. It also led to an insignificant result for both YT pre-roll exposure and the moderator relationship. As this model was invalid in the first place and there was not a significant relationship, I decided not to use this model. Model 6. Hence, model 6 was built with variables from model 4 and interactions suspected to interact logically proven by the conceptual model. This model turned out well without any error. Its pseudo R Squares are significantly higher than all the other models I have created RC&S = .027 and RN = .063.

19 Table 1 - Models and variables

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Demographic Relational characteristics

Timing Significant

variables

Interactions Conceptual model

Variables Age housewives YT exposure Weekday Weekday Weekday Weekday

Young Child TV exposure YT exposure YT exposure YT exposure

Old Child Loyalty TV exposure TV exposure TV exposure

Children 18 years Loyalty Loyalty Loyalty

HH Size Income Income Age housewives

Education YTxLOY TVxAgeHW

Income TVxLOY Income

Size of Town IncomexLOY

District

Table 2 - Correlation table Pur (Yes/No)

Variables Means S.D. 1 2 3 4 5 6 7 8 1 Purchase 0.082686283 0.275426206 1 2 YouTube exposure 0.002043318 0.045159949 .030** 1 3 TV exposure 0.068655497 0.290493714 .016 -.011 1 4 Loyalty 0.90337689 0.18048273 .037** .022* 1 5 Postcode number 4406.510421 2270.94794 .031** -.011 .016 -.034** 1 6 Household Size 2.65372565 1.254407521 -.025* -.002 .021 -.109** -.057** 1 7 Income 8.310993053 4.517374917 -.035** -.040** .013 -.006 -.053** .359** 1 8 Weekday 4.029968669 1.982739102 .130** -.001 -.031** .005 .002 .004 -.003 1 9 Size of town 2.032420651 0.946403193 -.024* .021 .042** -.185** -.021 .064** .076** -.008 **p< .01 and *p<0.05

Table 3 - Information criteria

AIC AICC BIC CAIC

Model 4 1347.805 1347.805 1423.719 1434.719

Model 6 2064.035 2064.035 2257.270 2285.270

Table 4 - VIF scores of Significant variables model (Model 4)

Weekday YT exposure TV exposure Loyalty Income

20

To ensure the variables in Model 4 and 6 were the best variables, I conducted a correlation analysis on all the variables in the analysis to see which other variables had a relationship with Pur. I took note on their Pearson Correlation r values to make sure they were smaller than .50 and larger than -.50 to avoid multicollinearity. Those extensively higher than these numbers were thought to highly correlate with the dependent variable and the other variables in the model, which might lead to confusion of the effects during estimation and multicollinearity.

The correlation in Table 2 showed that all the variables included in Model 4 and 6 fit the criteria stated above besides Age of housewives. Age of housewives was not significant with the dependent variables. However, as stated in the conceptual model, it would be used as a moderator to test how the age of the HHs would affect the relationship between YT pre-roll and TV advertising on brand X purchasing choice, separately. Therefore, this would not be a problem.

I then proceeded to select the model using information criteria such as AIC, AICC, BIC and CAIC. The model with the lowest information criteria was the better model. As can be seen from Table 3, Model 4 was clearly the better model because its AIC was 1347.805, AICC was 1347.805, BIC was 1423.719 and CAIC was 1434.719, which were all smaller than Model 6’s. To make sure there was not any multicollinearity, the variance inflation factor (VIF score) were calculated for the model and all the VIF scores were roughly 1 (Table 4). Therefore, multicollinearity would not be a problem.

The logit model is:

𝐹{𝑌𝑖𝑡} = exp (𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑆𝑈𝑁𝑖𝑡+ 𝛽4𝑀𝑂𝑁𝑖𝑡+ 𝛽4𝑇𝑈𝐸𝑆𝑖𝑡 +𝛽4𝑊𝐸𝐷𝑖𝑡+ 𝛽4𝑇𝐻𝑈𝑅𝑆𝑖𝑡+ 𝛽4𝐹𝑅𝐼𝑖𝑡+ 𝛽4𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡) 1 + exp (𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑆𝑈𝑁𝑖𝑡+ 𝛽4𝑀𝑂𝑁𝑖𝑡+ 𝛽4𝑇𝑈𝐸𝑆𝑖𝑡 +𝛽4𝑊𝐸𝐷𝑖𝑡+ 𝛽4𝑇𝐻𝑈𝑅𝑆𝑖𝑡+ 𝛽4𝐹𝑅𝐼𝑖𝑡+ 𝛽4𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡) 𝑈𝐵𝑖𝑡= 𝛽0+ 𝛽1𝑌𝑇𝑖𝑡+ 𝛽2𝑇𝑉𝑖𝑡+ 𝛽3𝐿𝑂𝑌𝑖𝑡+ 𝛽4𝑆𝑈𝑁𝑖𝑡+ 𝛽4𝑀𝑂𝑁𝑖𝑡+ 𝛽4𝑇𝑈𝐸𝑆𝑖𝑡+ 𝛽4𝑊𝐸𝐷𝑖𝑡+ 𝛽4𝑇𝐻𝑈𝑅𝑆𝑖𝑡 + 𝛽4𝐹𝑅𝐼𝑖𝑡+ 𝛽4𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡+ 𝜀𝑖𝑡

𝑌𝑇𝑖𝑡= YouTube pre-roll advertising exposure at time t for HH i

𝑇𝑉𝑖𝑡= TV advertising exposure at time t for HH i

𝐿𝑂𝑌𝑖𝑡= Loyalty of consumer i over the whole sample period for HH i

𝑆𝑈𝑁𝑖𝑡 = Sunday for HH i 𝑀𝑂𝑁𝑖𝑡 = Monday for HH i 𝑇𝑈𝐸𝑆𝑖𝑡 = Tuesday for HH i 𝑊𝐸𝐷𝑖𝑡 = Tuesday for HH i 𝑇𝐻𝑈𝑅𝑆𝑖𝑡 = Thursday for HH i 𝐹𝑅𝐼𝑖𝑡 = Friday for HH i 𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡 = Income for HH i

21 Table 5 - Description of the independent variables

No. Independent

variables

Description Type of variable

1 YouTube exposure Contact with YouTube Categorical variable assumed ordinal with 3 levels (1 = expose to YT 1 time and 0 = not exposed to YT).

2 TV exposure Contact with TV Categorical variable assumed ordinal with 4 levels (0 = not exposed to TV, 1 = exposed to TV 1 time …4 = exposed to TV 4 times).

3 Loyalty The degree to which a consumer

consistently purchases the same brand within a product class (Bennett, 1995)

Continuous variable. (From 0 to 1 and blank. 0 means no loyalty and 1 means 100 percent loyal. Blank is no information because the HH did not buy from brand X nor the competitors).

4 Weekday The day of the week Categorical (1 = Sunday, 2 = Monday…7 = Saturday)

6 Income How much money does the HH make Categorical variable assumed continuous (From 700 euro to 1100 – 1300 euro)

Table 7 - Classification table - Hit rate of Significant Model (Model 4)

Brand X purchase or not purchase Percentage correct

No purchase Purchase Total

Purchase No purchase 4762 1972 6734 70.7

Purchase 281 326 607 53.7

22

IV. Estimation

4.1 Results

Table 6 - Estimation parameters of model 4

Significant model Step and variables B SE Wald Exp(B) Intercept -2.190** .277 62.583 .112 Main effects YT 1.303** .605 4.634 3.680 TV .240 .133 3.249 1.271 LOY .811** .271 8.960 2.249 Control Income -.028** .010 8.435 .972 Sunday -1.667** .194 73.721 .189 Monday -1.111** .153 52.443 .329 Tuesday -.975** .150 42.371 .377 Wednesday -1.065** .153 48.625 .345 Thursday -.957** .148 41.547 .384 Friday -.306* .126 5.919 .736 Cox&Snell R2 (pseudo R2) .023 Nagelkerke R2 (pseudo R2) .053 McFadden’s (pseudo R2) 0.041 **p<.01 and *p<.05 The logit model becomes: 𝑭{𝒀𝒊𝒕} = exp (−2.190 + 1.303𝑌𝑇𝑖𝑡+ .240𝑇𝑉𝑖𝑡+ .811𝐿𝑂𝑌𝑖𝑡− 1.667𝑆𝑈𝑁𝑖𝑡− 1.111𝑀𝑂𝑁𝑖𝑡− .975𝑇𝑈𝐸𝑆𝑖𝑡− 1.065𝑊𝐸𝐷𝑖𝑡 −.957𝑇𝐻𝑈𝑅𝑆𝑖𝑡− .306𝐹𝑅𝐼𝑖𝑡− .028𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡) 1 + exp (−2.190 + 1.303𝑌𝑇𝑖𝑡+ .240𝑇𝑉𝑖𝑡+ .811𝐿𝑂𝑌𝑖𝑡− 1.667𝑆𝑈𝑁𝑖𝑡− 1.111𝑀𝑂𝑁𝑖𝑡 −.975𝑇𝑈𝐸𝑆𝑖𝑡− 1.065𝑊𝐸𝐷𝑖𝑡− .957𝑇𝐻𝑈𝑅𝑆𝑖𝑡− .306𝐹𝑅𝐼𝑖𝑡− .028𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡) 𝑼𝑩𝒊𝒕= −2.190 + 1.303𝑌𝑇𝑖𝑡+ .240𝑇𝑉𝑖𝑡+ .811𝐿𝑂𝑌𝑖𝑡− 1.667𝑆𝑈𝑁𝑖𝑡− 1.111𝑀𝑂𝑁𝑖𝑡− .975𝑇𝑈𝐸𝑆𝑖𝑡 − 1.065𝑊𝐸𝐷𝑖𝑡− .957𝑇𝐻𝑈𝑅𝑆𝑖𝑡− .306𝐹𝑅𝐼𝑖𝑡− .028𝐼𝑁𝐶𝑂𝑀𝐸𝑖𝑡+ 𝜀𝑖𝑡

4.2 Interpretation

23

of purchasing brand X is 3.680. In sum, a HH is more likely to purchase if that HH is exposed to YouTube pre-roll advertising. This is in line with hypothesis number 2.

H2: YouTube pre-roll advertising exposure leads to high probability of consumer choosing brand X as their brand choice

TV advertising exposure. TV advertising exposure does not significantly predict whether someone would purchase brand X or not. B = .240, Wald = 3.249, p > .05. The odds ratio is smaller than YouTube pre-roll so the effect size is not comparable to each other. Therefore, hypothesis 1 cannot be justified.

Loyalty. Loyalty level of the different HHs significantly predicts HH purchasing probability of brand X. B = .811, Wald = 8.960, p<.01. The odds ratio says that as the loyalty of HHs increases by 1 unit, the change in the odds ratio of purchasing over not purchasing is 2.249. In other words, if loyalty increases, HHs are more likely to purchase brand X. Even though this outcome does not answer hypothesis 4, it is in line with prior argumentation. The more loyal a customer is to a brand, the more likely the person will buy from that brand. If that customer would like to buy a soft drink, and that customer is loyal to brand X, this can mean that customer would buy brand X irrespective of having seen brand X YouTube advertising message or not because loyalty by itself can increase volume purchased (Tellis, 1988). In this case, the result proved that loyalty by itself can significantly predict brand X purchase.

Income. Income level is significant, which means that income significantly predicts whether a HH would purchase brand X or not. B = -.028, Wald = 8.435, p<.01 and the odds ratio can be interpreted as the income increases by 1 unit, the change in the odds ratio is .972. This means if the income level increases; the HHs are less likely to purchase brand X.

Weekday. The day of the week is significant in predicting whether a HH would purchase brand X or not. The baseline was Saturday, therefore, this is the effect of every other day as compared with Saturday. Sunday is significant in predicting brand X purchase BSunday = -1.667, Wald = 73.721, p > .05. The odds

ratio of Sunday can be interpreted as the day of the week changes from Saturday to Sunday, the change in the odds of purchasing brand X is .189. In other words, the odds of purchasing on a Sunday as compared to not buying are 1/.189 = 5.29 times more than a Saturday. Similarly, Monday (BMonday = -1.111, Wald =

-52.443, p <.01), Tuesday (BTuesday = -.975, Wald = 42.371, p <.01), Wednesday (BWednesday = -1.065, Wald

= 48.625, p <.01), Thursday (BThursday = -.957, Wald = 41.547, p <.01), Friday (BFriday = -.306, Wald =

24

2.60 and 1.36 times more than a Saturday, respectively. Therefore, the odds are the highest for Sunday. Sunday is the day with the highest odds that brand X will be purchased.

While it seemed that hypotheses 3 and 4 were not tested because their effects were not reported in Model 4 and Table 6, they were indeed tested. However, the inclusion of these variables led to either invalid model or insignificant results of the main variables. Model 5 attempted to test Hypothesis 4 and Model 6 attempted to test Hypothesis 3.

Looking at Model 6, adding the loyalty as moderator for the relationship between YouTube pre-roll and purchasing behavior resulted in an error that questioned the validity of the results. It also led to an insignificant result for both YT pre-roll exposure and the moderator relationship. As this model was invalid in the first place and there was not a significant relationship, I decided not to use this model. Additionally, because loyalty, by itself could significantly predict purchasing choice and YouTube pre-roll as well, it did not make sense to test the moderator relationship.

Model 6 did not encounter the error Model 5 encountered. However, in comparing Model 4 and Model 6, Model 4 is the better model with lower information criteria. Model 6, nevertheless, provided insights for hypothesis 3. The outcome of Model 6 showed that the age of the housewives (being the person who shops at each HHs) does not significantly moderate TV exposure and purchasing intention. Since the relationship was not significant, it was not reported in the main findings. In conclusion, only hypotheses 1 and 2 were answered fully.

V. Validation

5.1 Coefficients

25

HH with very high income can be very health conscious so they would not buy brand X. Half of Americans drink soda daily and the poor receive 9 percent of their daily calories from it (Daily Mail, 2011). Low income people eat unhealthy food because they could not afford to buy healthy food. One dollar can buy 900 calories of soda when it can only afford 170 calories of orange juice (The Institute of Natural Healing, 2011). This can explain the negative sign but it is out of the scope of this study and the data at hand to test this theory. Last but not least, the days of the week. The baseline is Saturday and all the other days are negative. The interpretation would be that Sunday is the best day to promote brand X because people buy brand X more on the weekend because that is when they go grocery shopping or go out (Jacobe and Jones, 2009).

All the variables included in the model besides TV advertising have a significant p-value. The Wald statistic can determine if the regression coefficients are different significantly from zero or not (H0 = 0). Its cut off value is 3.84. Judging from Table 6, all the parameters besides TV advertising are significantly different from zero as they have Wald value higher than 3.84 and their p value is lower than 0.05.

5.2 Hit rate

The hit rate describes in what percentage of observation the model hit. In other words, how often the model hits its target. However, it seems as if the more observations added to the model, the higher the percentage correct. Therefore, it is easy to get a high hit rate. Also the hit rate can be manipulated by changing the cut off value of the classification table. In this case, I used the mean of the dependent variable which is .0827 as the cut off value for the classification table because the cut off value of SPSS leads to 0% correct in predicting purchasing choices. Perhaps, the amount of purchase is way too low for the hit rate to actually predict choice because the hit rate will predict for the biggest group observed, which is the no purchase group. Note that in the training group, 8.3 percent of the HHs buys whereas 91.7 did not buy. Similarly, in the test group, 7.9 percent of the HHs buy and 92.1 percent of the HHs did not buy.

26

fairs better when the predictors are added to the model. Therefore, the full model results in an overall percentage correct of 69.3% whereas the null model predicts 91.7% correct.

However, that is the percentage correct for the no purchase group. In this article, I am interested in the purchase group even though the purchase group is smaller than the purchase group. Looking at the null model for the purchase group, only 8.26% is correct (607 cases correct/7341 total cases). In the full model in Table 7, of the 607 times the HHs purchases, the full model predicts 326 cases correctly as 281 cases incorrectly which leads to 53.7% correct. In all 30.69% {(1972 + 281)/(4762+326+281+1972)} of the model is correctly classified for the brand X purchase. In conclusion, the null model predicts 8.26% correct whereas the full model predicts 30.69% correct. The null model, in comparison with the full model reduces the prediction by {(1972 + 281)-607}/(7341-607) = 24.44%, which seems like a high number. This indicates a good quality model for the full model.

To explore the importance of each variable to the dependent variable, the predictor importance graph was generated in SPSS modeler. As can be seen in Figure 3, the most important variable is Weekday. The second most important variable is loyalty, then YouTube contacts, income and lastly TV exposure. Therefore, the hit rate is mainly attributable to the weekday and loyalty variable.

Figure 3 - Predictor importance

To see whether YT pre-roll and TV advertising add something to the hit rate, I generated the hit rate with and without the two variables to see how the hit rates would change. Without TV ad exposure and YT pre-roll exposure, the hit rate for purchasing is 30.39%; which means without these variables the hit rate decreased for the no purchase. Meaning they do add value to the hit rate.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 TV ad exposure Income YT pre-roll ad exposure Loyalty Weekday

Least to most important

Pr e d ic to rs

27

5.3 Chi Square

To compare the null model to the full model, we can look at the Chi Square statistics in the Variables not

the in the Equation table. The Overall statistics states that the residual chi-square statistic is 182.653

which is significant at p<.05. This statistic lets the readers know that the coefficients for the Variables not

in the model are significantly different from zero. So any addition of new variables will significantly

affect the predictive power of the model. In other words, forcing more variables in the null model would make a significant contribution to its predictive power. Therefore, the full model is better than the null model.

The overall fit of the model is assessed using the log-likelihood statistic (LL). There is a -2 in front of the log-likelihood number because -2LL has an approximate chi-square distribution and so it enables us to compare values against those that we might expect to get by chance alone and large values of the log-likelihood means poorly fitting statistical model. In this case, the LL of the null model is 4188.506 and of the full model is 4017.070. Since the LL of the full model is smaller than the null model, one can conclude that it is the better fitting model.

The overall fit and the LL lets us know the full model is the better fitting model. The Omnibus Tests of

Model Coefficients table lets us know how much better the model predicts the outcome variables using the

model chi-square statistic. The chi-square measures the difference between the full model with all the variables included and the null model with only the constant. The chi-square measurement = LLnull –

LLfull = 171.435 with significant level p<.05 (Appendix). In sum, this means the full model is better than

the null model.

Many tests were done because the chi-square test statistics are designed to detect differences between a model and the observed data. That means, adding more variables to the model will always improve model fit; which makes it hard to distinguish a realistic improvement in fit between the full model and the full model (Williams, 2016). Due to this, other techniques such as hit rate and LL had to be employed to ensure the correctness of the validation and the parsimoniousness of the model.

5.4 Psuedo R Square

28

5.5 Multicollinearity

Multicollinearity can be a problem because it can cause bias in the effects of the independent variable on the dependent variable. Multicollinearity can be detected through the VIF score. The cut off value of VIF score is 5. As can be seen in Table 4, all the variables meet this criterion because all have a value below 5. Therefore, multicollinearity is not an issue.

5.6 Robustness test

Figure 4 and 5 - Cumulative lift curve for the training sample and the testing sample

29

VI. Discussion

The key conclusion of the analysis and the main emphasis of the research paper is that YouTube pre-roll advertising can predict whether a HH would purchase brand X or not. The odds ratio implies that as YouTube pre-roll advertising increases in 1 unit of calculation, the change in the odds of purchasing brand X would increase by 3.680 points. Even though YouTube pre-roll advertising exposure has a significant effect on brand X purchasing probability and YouTube pre-roll is similar to TV advertising in nature, TV advertising exposure does not yield significant effect.

YouTube pre-roll is new as it was only introduced in 2010 which is 6 years ago and in this dataset HHs were not exposed to YouTube pre-roll as many times as TV advertising. The maximum YT exposure per day measured by Gfk was 2 times whereas TV can be measured up to 4 times per day. However, the descriptive analysis showed that irrespective of such low exposures, YT pre-roll which started only in week 5 has similar frequency as the purchasing behavior ever since it started (at weekly level). Similar pattern was found at daily level, YT pre-roll exposures were low but with every peak of YT pre-roll, a peak in brand X purchasing count followed (Figure 2). This could be attributed to a more targeted nature of online advertising that YouTube pre-roll possesses; or this could be due to the decline in TV exposure strategy.

Suddenly, week 7 and 8 did not present any TV advertising. I did not believe this was a problem in the data because I checked data recorded by Gfk and survey that specifically asked HHs if they had seen TV advertising and they did not. Furthermore, it was not one HH that was exposed to TV advertising; it was 2609 HHs that were not exposed to TV advertising this week. No significant effects of TV advertising could also be attributed to the Share a coke campaign that was so popular on YouTube in 2014. In 2014, The brand X Company really focused on YouTube as its main channel to communicate their Share a coke campaign, which drove massive fan interactions and shares (Hitz, 2014).

30

buying brand X are highest on a Sunday, which is during the weekend. Families usually go shopping on the weekend (Jacobe and Jones, 2009) so this outcome is nothing new but it points the focus YouTube pre-roll advertising promotion to the weekend for managers.

Last but not least, loyalty is the strongest variable to determine purchase behavior as it increases volume purchased of loyal consumers but is not effective in winning new buyers (Tellis, 1988). As this study only looks at the customers who have bought more than three times, needless to say, that means this study only applies to HHs that have bought brand X before. The results of this study show that as loyalty increase by 1 unit, the change in the odds ratio of purchasing over not purchasing is 2.249. So if loyalty increases, HHs are more likely to buy. Therefore, in order to increase purchase probability, companies should target loyal customers because loyalty itself can significantly predicts brand X purchase.

Managers have been spending an exponentially large amount of capital on TV advertising. As mentioned earlier, billions were spent on TV whereas a comparably low amount of thousands were spent on YouTube pre-roll advertising. The results of this study show that while YT pre-roll enjoyed less exposure, its effect is significant in predicting consumer purchasing probability. Additionally, given the low frequency of brand X purchases, it was hard for this model to predict actual choice because all consumers are predicted to belong to the no buy group. Nevertheless, this model is good for sorting people by their propensity to buy brand X. In other words, people who are exposed to YouTube pre-roll advertising, are loyal, have low income and usually buy on the weekend would have a higher probability to purchase brand X. This is an implication for managers to focus on these kinds of people with the above characteristics.

As for academic implications, YouTube pre-roll is important for academic research because of its ability to generate recall, purchase intent and despite the fact that it is annoying; it works. Thus far, lack of academic attention has been given to YouTube pre-roll because it is a new phenomenon and because of the shortage of single source data available on YouTube pre-roll exposure. Next to many newspaper articles and blogs, this study is the first to its knowledge to add information to the, yet rather small library of YouTube pre-roll knowledge base for brand X, which is a product in the CSDs market.

31

This research contributes to the research field of online and offline advertising because YouTube pre-roll is an online advertising tool and TV advertising is an offline advertising tool. This paper directly gives a specific case in which online advertising tools such as YT pre-roll is more prominent than offline advertising tools such as TV advertising. As argued, this could be explained by online advertising tool’s ability to target consumer better than offline tools.

VII. Conclusions and recommendations

The logit regression model provides a good representation of the effect that YouTube pre-roll advertising has on the purchasing of brand X at individual customer level. The training model predicted the behavior of the testing sample of HHs quite well. Significant results were found for YouTube pre-roll advertising effect on the weekend. Loyalty and income also have a significant positive and negative effect on purchasing choice, consecutively. The success of the modeling process was due to the detailed daily records of advertising exposure at household daily level.

The value output of my model is its ability to recommend the latent propensity to buy brand X, which is YouTube pre-roll exposure, low income, loyalty and purchasing occasion on the weekend because these predictors yielded significant effects on the dependent variable of brand X purchasing choice. Therefore, one recommendation for managers is to focus their budgeting more on YouTube pre-roll online advertising instead of TV offline advertising. Another recommendation for managers is to target those who have been exposed to YouTube, are loyal to the company (have bought brand X before), have low income and focus their promotion on the weekend for these people at check-out (for example).

Another value output of my model in the academic world is its accumulation of knowledge for YouTube pre-roll advertising and more proof that disaggregated daily level data is best to explore causal relationship between advertising exposure and purchase probability. Even though the results are positive, they are not perfect. Improvements can be made with a dataset with more brand X purchasing incidents. It is strange that such a popular product has such little purchasing incidents. Maybe future research can look into the value bought because purchase incidents might not be much but the value bought might show a different picture. Future research can also look at the changing health trends of why high income households buy less brand X than low income HHs. More information on competitor advertising choice on YouTube, or display quality can also give a better picture of the HHs and the market.

32

33

VIII. References

Benady, David. "Advertising To The You Tube Generation." Marketing (00253650) (2009): 34-35. Business Source Premier. Web. 29 Feb. 2016.

Bronnenberg, Bart J, Jean-Pierre Dub, and Carl F Mela. "Do Digital Video Recorders Influence Sales?." Journal Of Marketing Research (JMR) 47.6 (2010): 998-1010. Business Source Premier. Web. 5 June 2016.

Burns, W. (March 3, 2015). GEICO Brilliantly Reinvents YouTube Preroll Advertising. Retrieved from Forbes Web site: http://www.forbes.com/sites/willburns/2015/03/03/geico-brilliantly-reinvents-youtube-preroll-advertising/#ba4dd0568b9e

Carter, B (n.d). Why YouTube Pre-Roll Ads Rock & How To Take Advantage of Them. Retrieved from Convince&Convert Web site: http://www.convinceandconvert.com/content-marketing/why-youtube-pre-roll-ads-rock-how-to-take-advantage-of-them/

Court, D (June, 2009). The consumer decision journey. Retrieved from the McKinsey Quaterly Web site: www.mckinsey.com/business-functions/marketing-and-sales/our-insights/the-consumer-decision- journey Daily Mail (September 2011). Sugar nation: Half of Americans drink soda daily - and poor people get 9 per cent of their daily calories from it. Retrieved from the Daily Mail website:

http://www.dailymail.co.uk/news/article-2033153/Sugar-nation-Half-Americans-drink-soda-daily--poor-people-9-cent-daily-calories-it.html

Dutta-Bergman, Mohan J. "Complementarity In Consumption Of News Types Across Traditional And New Media." Journal Of Broadcasting & Electronic Media 48.1 (2004): 41-60. Communication & Mass Media Complete.

Erdem, T., Keane, M. P. and Sun, B. (2008) The impact of advertising on consumer price sensitivity in experience goods markets, Quantitative Marketing Economics, 6, 139–76.

Finan, C. (October, 5, 2015). Why those annoying YouTube pre-roll ads actually work. Retrieved from local solutions Web site: http://blog.cmglocalsolutions.com/why-those-annoying-youtube-pre-roll-ads-actually-work

Google (2014). YouTube insights. Retrieved from Google website:

https://www.thinkwithgoogle.com/research-studies/youtube-insights-stats-data-trends-vol4.html

Hansen, Flemming, Lotte Yssing Hansen, and Lars Grønholdt. "Modelling Purchases As A Function Of Advertising And Promotion." International Journal Of Advertising 21.1 (2002): 115-135. Business Source Premier. Web. 7 Mar. 2016.

Hawkins, Scott A., Stephen J. Hoch, and Joan Meyers-Levy. "Low-Involvement Learning: Repetition And Coherence In Familiarity And Belief." Journal Of Consumer Psychology (Lawrence Erlbaum Associates) 11.1 (2001): 1-11. Business Source Premier. Web. 7 Mar. 2016.

Hitz, L (September, 2014). Simply Summer Social Awards Contestant #3: @CocaCola’s #ShareACoke Campaign. Retrieved on May 19, 2015 from SimplyMeasured website:

34

Hlavac, T. E., Jr., J. D. C. Little. 1966. A geographic model of an Urban automobile market. D. B. Hertz, J. Melese, eds. Proc. Fourth Internat. Conf. Oper. Res., Wiley-Interscience, New York, 302–311.

Jacobe, D & Jones J (October, 2009). Consumers spend more on weekends, payday weeks. Retrieved from Gallip Web site: http://www.gallup.com/poll/123839/consumers-spend-more-weekends-payday-weeks.aspx

Janssens, W. (2008). Marketing Research with SPSS. Harlow, England: Prentice Hall, Financial Times. https://statistics.laerd.com/spss-tutorials/dichotomous-moderator-analysis-using-spss-statistics.php Jensen, Morten Bach. "Planning Of Online And Offline B2B Promotion With Conjoint Analysis." Journal Of Targeting, Measurement & Analysis For Marketing 16.3 (2008): 203-213. Business Source Premier. Web. 3 June 2016.

Leeflang, P. , & Wieringa, J. (2015). Modeling Markets. London: Springer.

Li, Hao, and Hui-Yi Lo. "Do You Recognize Its Brand? The Effectiveness Of Online In-Stream Video Advertisements." Journal Of Advertising 44.3 (2015): 208-218. Business Source Premier. Web. 7 Mar. 2016.

Jones, J. M., F. S. Zufryden. 1980. Adding explanatory variables to a consumer purchase behavior model: An exploratory study. J. Marketing Res. 17(August) 323–334.

Lopez, Rigoberto A., Yizao Liu, and Chen Zhu. "TV Advertising Spillovers And Demand For Private Labels: The Case Of Carbonated Soft Drinks." Applied Economics 47.25-27 (2015): 2563-2576. EconLit. Web. 7 Mar. 2016.

Lunden, I (Jan 20, 2015). 2015 Ad Spend Rises To $187B, Digital Inches Closer To One Third Of It. Retrieved from Web site: http://techcrunch.com/2015/01/20/2015-ad-spend-rises-to-187b-digital-inches-closer-to-one-third-of-it/

Manchanda, Puneet, et al. "The Effect Of Banner Advertising On Internet Purchasing." Journal Of Marketing Research (JMR) 43.1 (2006): 98-108. Business Source Premier. Web. 7 June 2016.

Martz, E (December 3, 2015). What Can You Say When Your P-Value is Greater Than 0.05? Retrieved from The Minitab blog: http://blog.minitab.com/blog/understanding-statistics/what-can-you-say-when-your-p-value-is-greater-than-005

Pauwels, Koen, et al. "Does Online Information Drive Offline Revenues?: Only For Specific Products And Consumer Segments!." Journal Of Retailing 87.1 (2011): 1-17. Business Source Premier. Web. 8 Mar. 2016.

Punj, G. N., R. Staelin. 1978. The choice process for graduate business schools. J. Marketing Res. 15(November) 588–598.

Postigo, Hector. "The Socio-Technical Architecture Of Digital Labor: Converting Play Into Youtube Money." New Media & Society 18.2 (2016): 332-349. SocINDEX. Web. 31 May 2016.

35

Richardson, Anna. "Youtube Can Add Significant Reach For Brands With Its TV-Sized Audience." New Media Age (2011): 26. Business Source Premier. Web. 29 Feb. 2016.

Richard Williams, University of Notre Dame, http://www3.nd.edu/~rwilliam/Last revised January 14, 2016

Risselada, Hans, Peter C. Verhoef, and Tammo H. A. Bijmolt. "Staying Power Of Churn Prediction Models." Journal Of Interactive Marketing 24.3 (2010): 198-208. PsycINFO. Web. 15 June 2016. Sang-Uk, Jung, Thomas S. Gruca, and Lopo L. Rego. "Excess Loyalty In CPG Markets: A

Comprehensive Examination." Journal Of Empirical Generalisations In Marketing Science 13.1 (2010): 1-13. Business Source Premier. Web. 25 May 2016.

Schmidt, Susanne, and Martin Eisend. "Advertising Repetition: A Meta-Analysis On Effective Frequency In Advertising." Journal Of Advertising 44.4 (2015): 415-428. Business Source Premier. Web. 7 Mar. 2016.

Silberstein, Richard B., and Geoffrey E. Nield. "Brain Activity Correlates Of Consumer Brand Choice Shift Associated With Television Advertising." International Journal Of Advertising 27.3 (2008): 359-380. Business Source Premier. Web. 2 June 2016.

Smith, A. (March 6, 2015). Five YouTube pre-roll ads that actually made viewers sit up and take notice. Retrieved on reelseo Web site: http://www.reelseo.com/5-youtube-pre-roll-ads-viewers-noticed/

Steiner, Gary A. "The People Look At Commercials: A Study Of Audience Behavior." Journal Of Business 39.2 (1966): 272. Business Source Premier. Web. 17 Mar. 2016.

Tellis, Gerard J., and Doyle L. Weiss. "Does TV Advertising Really Affect Sales? The Role Of Measures, Models, And Data Aggregation." Journal Of Advertising 24.3 (1995): 1-20. Business Source Premier. Web. 7 Mar. 2016.

Tellis, Gerard J. "Advertising Exposure, Loyalty, And Brand Purchase: A Two-Staged Model Of Choice." Journal Of Marketing Research (JMR) 25.2 (1988): 134-144. Business Source Premier. Web. 7 Mar. 2016.

Teo, T. (2001). Demographic and motivation variables associated with Internet usage activities. Internet Research-Electronic Networking Applications and Policy, 11, 125-137.

The Institute of Natural Healthing (April, 2011). The Economics of Obesity: Why Are Poor People Fat? Retrieved from The INH website: http://www.institutefornaturalhealing.com/2011/04/the-economics-of-obesity-why-are-poor-people-fat/

The Nielsen Company. (2009, April 22). Online engagement deepens as social media and video sites reshape the Internet, Nielsen reports [Press release!,

http:/log.nielsen.com/nielsenwire/wp-content/uploads/2009/04/nielsen-online-global-_pr.pdf