THE TIME-DEPENDENT EFFECT OF ONLINE
INFORMATION SEARCH AND V
ALENCE OF
WORD-OF-MOUTH ON BOX OFFICE REVENUE
Evert de Haan (s1808680)
July 19, 2011
By: Evert de Haan 2 Abstract
In this empirical paper we have investigated how the online search volume (the amount of people that search online for information) of a particular movie is related to the box office revenue of that movie. Furthermore we have investigated the time dependent effect of the valence of word-of-mouth on box office revenue. We have conducted this with the use of a dynamic panel data model with instrumental variables to capture the endogeneity of some of the explanatory variables. We have found that in the first three days of a movie’s release the effect of the online search volume is insignificant, but this effect turns positive and significant after these three days. This may be due to the fact that people are first weighing out what they will do with the information which they have found and are later more anticipating towards this information. What we also have found is that the effect of the valence of word-of-mouth is in the first stage of the movie’s release positive, but turns and stays insignificant after ten days, which could explain why there have been mixed findings in the literature up to now.
By: Evert de Haan 3 1. Introduction
In this empirical paper we will inspect how the online search volume, in terms of the amount of people that search online for information, of newly released movies will provide information on the (future) demand for that movie (measured in terms of box office revenue). While we already know quite a lot about the effect of consumer
generated information (e.g. consumer reviews) on revenue, we do not know much about how the amount of people that search for information influences revenue. Since the amount of people that search for information may provide more insight into how many consumers are actually interested in a certain product, this may also be a good indicator for the demand of that product. In practice we for instance have seen that highly
anticipated products, such as mobile phones and game consoles, have generated much search for information during the Christmas season (Wray 2009). If this additional search for information is also an indication for additional sales for these products is however not investigated up to now as far as we know.
Although the relationship between online consumer reviews and movie revenue has already received much attention (see Hadida (2009) for an overview of empirical studies in the motion picture industry), this paper will provide some new insights. First of all we have included the not yet researched variable of ‘online search volume’ in our study. The online search volume captures the amount of people that have searched online for
By: Evert de Haan 4 online reviews and other product information. The online search volume can therefore say something about the amount of people that is reached by the online word-of-mouth. Including this variable can therefore further increase our understanding of how online information can determine the demand for a product.
A second new insight that this paper provides is that we are going to investigate why in some studies the valence of the word-of-mouth of a movie has been found to be significantly related to the movie’s revenue, while other studies have found no significant effect. We propose that the effect of the grade of the movie is moderated by the amount of time that the movie has already been in theaters; when the movie is just released in theaters the grade is expected to have a positive effect on revenue because the grade is given by ‘early adopters’ who’s opinion is highly appreciated by the ‘followers’. The early adopters can be considered to be fans of the movie’s franchise, its director, or the actor(s) playing in the movie. Because of this they may provide a (positive) bias in the grade which is not entirely corrected for by the followers (people who will go see the movie in a later stage of its release), which is in line with the results found by Li, and Hitt (2008) for the online purchase behavior of books. When the movie is longer in theaters the effect of the grade diminishes since all the innovators and early adopters have by then already given their opinion and the opinion of the later adopters is expected to be less valuable for people when making decisions in terms of which movie to see.
By: Evert de Haan 5 only interesting for the motion picture industry and research that will be conducted in this field, but this information could be applicable in other product categories as well.
To give an answer to our research question we have collected daily revenue information from a total of 165 movies that were released between September 2008 and September 2009 in the domestic (US and Canadian) market, together with other variables of interest, including word-of-mouth related variables (amount of reviews and the daily grade) and the new search variable. In the first part of this paper we provide the
theoretical foundations of our research, and after this we will discuss the data collection efforts and summarize the data which we have collected, followed by the models that we have used to test our hypotheses and the conclusions that we have drawn.
2. Literature review
By: Evert de Haan 6 (2010) also have shown that 68% of the US and Canadian population of 2 years and older went to the movies at least onece in 2010, and a total of 1.3 billion tickets where sold in this market in the same period (in average 6 tickets per year per moviegoer), which is almost four times as high as the amount of theme park tickets that were sold in 2010 in the same region. A second reason that makes the motion picture industry an interesting and relevant industry to do research in is that rich data such as the (daily) box office revenue and online consumer and professional reviews are freely available for many different movies (Eliashberg, Elberse and Leenders 2006). Also new movies are released on a regular basis; according to MPAA (2010) there were 558 movies released in 2009 in theaters in the US and Canadian market, thus there are sufficient unique observations to draw conclusions on a wider basis than for only one or a few product introductions.
Previous studies have already investigated what the effect of online word-of-mouth is on the revenue of a movie. Many studies (e.g. Duan, Gu, and Whinston 2008,
By: Evert de Haan 7 furthermore been shown that the reason why the revenue of different movies are highly unequally distributed is because of social influences (i.e. the word-of-mouth that is generated by the movie) (Delre et al. 2010). This means that word-of-mouth can actually make or break a movie, which is also the case for other experience goods such as books and cd’s, as Delre et al. (2010) have shown. Also other studies (e.g. Dye 2000, Sornette et al. 2004, Sorensen 2004, and Kohli and Sah 2003) support the view that experience goods are mainly driven by buzz. This makes the movie industry ideal for studying the impact of word-of-mouth.
We however propose that word-of-mouth is only one side of the coin in
understanding social behavior and social influences of consumers. Word-of-mouth is an active process where a person shares his or her opinion about a certain product with other people (Westbrook 1987). The amount of word-of-mouth that is generated can therefore be an indicator for how many people will be reached and influenced by this form of communication, and as a consequence can be a good indicator for (future) success (in terms of revenue and growth) of a product, which also has been found in the movie industry in many studies (Hadida 2009). The amount of word-of-mouth does however not provide insight in how many people are actually interested in receiving these kinds of information. Consumers (and also professional reviewers and traditional forms of
By: Evert de Haan 8 surrounding the movie may not reach that many people. Because of this we will in
contrast to previous studies also take the demand side of information into account. This will be done by including a variable that will capture the amount of people that search online for information on a daily level for each individual movie. This can also be seen in line with the two dimensions of influences by Bearden, Netemeyer and Teel (1989); the normative influences are about following the norms that are set by other consumers (captured by the word-of-mouth volume and valence), while the informative influences is focused on colleting the information that you want to have to make the right decision (captured by the search volume).
Since the online search volume says only something about voluntary search for information and thus the interest that consumers have in a product, this can be a very good indicator for (future) demand of a product. What is expected is that when the movie has just been released people will search for information about the movie because the movie receives a lot of attention (in the news and in reviews). Although this initial search may be a good indicator for the impact of the movie and the buzz that surrounds it, it may not say a lot about the amount of people that are actually interested in going to see the movie (or purchase a product, as Dye (2000) has pointed out). Over time, when the movie receives less attention in the media and the buzz settles down, the online search is
By: Evert de Haan 9 reasoning that consumers first have to let the buzz sink in before they will take action and search for additional information and to decide to go see the movie.
Furthermore movies that still have a wide release after a longer period of time, which is quite rare in the motion picture industry (Radas and Shugan 1998; Delre et al. 2010), have had a better opportunity to generate online information. Due to the longer period of the movie’s release the information that consumers find will be more objective and balanced (since more people have been able to give their opinion), the change that people will use this information to decide to go see the movie is also expected to increase. When the buzz for the movie diminishes, the information that is still spread after a period of time is expected to be more accurate and less hyped, and therefore this can be more helpful for people that search for information. On top of this movies that still attracts visitors after a longer period of time send out a positive quality signal (De Vany and Lee 2001), which is likely to increase the chance that people who search for
information will also go visit the movie. This leads us to the following hypotheses:
H1a: The online search volume of a movie is positively related to the box office revenue. H1b: The relationship between online search volume and box office revenue becomes
stronger (more positive) the longer the movie has been in release.
There have been some mixed findings when it comes to the effect of the valence of word-of-mouth on the revenue of the product. Chevalier and Mayzlin (2006) have for instance investigated how consumer ratings of books at Amazon.com and at
By: Evert de Haan 10 one-star reviews (the most negative reviews) has a stronger impact than the fraction of five-star reviews (the most positive reviews) on sales. A reason for this could be that reviews on Amazon.com and at Barnesandnoble.com tend to be very positive (according to Chevalier and Mayzlin (2006)), and therefore one negative review may have a stronger impact on the likelihood to buy than one of the many positive reviews.
Chen, Wu, and Yoon (2004) conducted a similar study, where they also used data from Amazon.com and inspected how consumer ratings influence the ranking of the books. Opposed to Chevalier and Mayzlin (2006) they found no significant effect. No explanation could be found in the literature on why these different results where found, but one reason could be that Chen et al. (2004) used cross-sectional data and Chevalier and Mayzlin (2006) have used panel data. The conclusion that can therefore be drawn is that there is no significant correlation between the sales of a book and the rating of that same book, but when looking over time, a change in rating for one book does positively affect sales of that same book. Thus when inspecting the impact of valence of word-of-mouth on sales it is more appropriate to use panel data instead of cross-sectional data, especially when the topic of interest is how valence of word-of-mouth affects sales.
By: Evert de Haan 11 movies and low valence movies), and not about the changes over time within one movie (i.e. the effect of a change in valence for one movie). On the other hand, also some studies that do use panel data (e.g. Duan et al. (2008) and Duan et al. (2009)) have found no significant effect of word-of-mouth valence on box office revenue.
What we expect is that there is a positive effect of the grade on the movie’s
revenue, but that this effect decreases over time when the movie is longer in release. The reason for this is that people that are most enthusiastic about a movie are in general going to see the movie as soon as it comes out and these people are believed to be opinion leaders. The opinion of these people is generally highly valued, while the opinion of the followers (people that will see the movie after it is already in release for some time) is expected to have a lesser impact. Furthermore the first impression of a product may make or break it, and this first impression is given by the persons that see the movie during its opening days or weeks. This results in the following hypotheses:
H2a: The valence of word-of-mouth is positively related to a movie’s box office revenue. H2b: The effect of the valence of word-of-mouth on box office revenue decreases the
longer the movie is in theaters.
3. Data
By: Evert de Haan 12 see how the effects are on a daily level, especially since the data for the other variables (box office revenue, grade and the amount of reviews) are also available on a daily level. A second and more serious downside is that the data from Google Trends are not the absolute amount of people that are searching for a specific movie, but the data are
standardized per search term, making the comparison between different movies harder. A third reason why these data are not appropriate for this research is that no distinction can be made for a search term that has multiple meanings; it is for instance not known if a person who is searching for ‘Titanic’ is searching for information on the ship or the movie with the same name from 1997.
By: Evert de Haan 13 The daily box office revenue and the amount of theaters that show the movie on a specific day was collected from the website Box Office Mojo.com
(http://www.boxofficemojo.com). Furthermore we also collected descriptive variables from Box Office Mojo.com for the movies in our dataset, such as the genre, budget and if the movie is a sequel. For the consumer reviews we have collected data from the Internet Movie DataBase (http://www.imdb.com). We collected all reviews from the Internet Movie DataBase for all movies in our dataset and coded the date on which the review was posted and the grade given per review (which is on a 1-10 scale).
Table 1: Summary statistics of the movies in our dataset (n=165)
Variable Summary Statistics
Genre Action (11.5%), Animation (6.1%), Comedy (29.1%), Drama (21.2%), Horror/Thriller (15.2%), Other (17.0%)
Sequel 9.1% of the movies
3D 5.5% of the movies
MPAA-rating1 G (3.0%) PG (18.2%), PG-13 (41.2%), R (37.6%) Major star Oscar won (26.8%), Oscar nominated (24.2%) Major director Oscar won (5.5%), Oscar nominated (3.6%)
Mean SD Minimum Maximum
Budget (n=102) $58,538,614 $54,988,335 $500,000
(Fireproof)
$250,000,000
(Harry Potter and the Half-Blood Prince)
Total search volume (over 100 days) 413,411.29 596,860.27 3,600 (Management) 3,454,100 (Transformers: Revenge of the Fallen)
Total domestic revenue $55,246,015 $67,182,932 $15,078
(The Education of Charlie Banks) $402,111,870 (Transformers: Revenge of the Fallen) Amount of reviews 115.85 169.08 2 (Management) 967 (Star Trek)
Average review grade 6.42 1.31 2.54
(Delgo)
9.96
(The Education of Charlie Banks)
Run time (minutes) 108.55 20.15 63
(Cheri)
168
(The Curious Case of Benjamin Button)
1
By: Evert de Haan 14 Table 1 shows the descriptive statistics of the different variables, where ‘major star’ (or ‘star power’) means that one of the main actors/actresses (as listed on Box Office Mojo) in the movie has won an Academy Award (Oscar) for best actor or best actress before the movie’s release, or if one of the main actors/actresses has been nominated for this award (but did not win) before the movie’s release. For ‘major director’ this is the same, but then for the Best Director Academy Award. This is comparable to how Basuroy, Chatterjee and Ravid (2003) have defined the star power of a movie in their study. The Internet Movie DataBase (http://www.imdb.com) has been used as a data source for these awards. The variable ‘3D’ indicates that the movie has (also) had a 3D-release. As can be seen we have a wide spread of different movies in terms of budget, genre, MPAA-rating, average grade, amount of reviews and the amount of search for information.
4. Research design
By: Evert de Haan 15 but on the other hand not with the error term (Ebbes 2004). Because of this we have chosen to use lagged variables as instruments, despite the critiques on this technique.
The pooled version, which assumes that all parameters will be the same for all movies, of our model is as followed:
ln(REVENUEit) = β0 + β1ln(REVENUEit-1) + β2ln(REVIEWSit) + β3REVIEWDUMit + β4ln(SEARCHit) + β5ln(GRADEit) + β6ln(SCREENSit) + β7ln(DAYit) + β8WEEKENDit + β9(ln(DAYit)*ln(GRADEit)) +
β10(ln(DAYit)*ln(SEARCHit)) + uit (1)
Where β0 is the constant term which is assumed to be equal for all movies, the other β’s are the elasticity coefficients (for example β2 is the short-term (direct)
review-elasticity, indicating that one percent more reviews for movie i on the day t will result in β2 percent more revenue for movie i on the day t), and uit is the error term for movie i on day t. REVENUEit is the revenue of movie i on day t, REVIEWS indicates the amount of reviews, SEARCH is the search volume, GRADE is the grade of the movie, SCREENS is the amount of theaters screening the movie, DAY is the number of days movie has been in release (set to 1 on the first day of the movie’s release) and WEEKEND indicates if the day is a day of the weekend. Furthermore we have included two interaction terms that are needed to answer our hypotheses. The interpretation of these interaction terms will follow the procedure of Irwin and McClelland (2001).
By: Evert de Haan 16 been in release, the value of REVIEWSit is on some days zero and therefore no logs can be taken. To resolve this problem we will use ln(REVIEWSit+1) as a proxy for
ln(REVIEWSit) and we have included a dummy variable (REVIEWDUMit) which is set to one in the case REVIEWSit is equal to zero. The parameter for this dummy variable (β3) is an indication of how well ln(REVIEWSit+1) is fitting ln(REVIEWSit). Ideally the parameter is non-significant, in the case it is significant it will correct for the arbitrary +1 (which could also have been set to +0.1 or +100 for example).
Also on days when no review has been posted, no daily grade is available.
Furthermore on days when there are only a few reviews available, the average grade per day can become highly fluctuating (i.e. increasing variance over time, a form of
heteroscedasticity). One solution for this could be to take the average grade up to day t. A downside of this is that when the amount of reviews increases over time, the average movie grade will become more stable (i.e. less variance) over time. To resolve these problems it has been chosen to use the exponentially weighted moving average (EWMA), for which the following formula has been used:
GRADEir = α · REVIEWGRADEir + (1 - α) · GRADEir-1 (2)
By: Evert de Haan 17 value has been chosen because this makes the variance of the grade over time stable, and it also makes the last ten reviews (the first page someone sees when looking at the reviews of imdb.com) have a total weight of about 90%.
Figure 1 as an example shows the grade of the movie “Harry Potter and the Half-Blood Prince” using three different weighting methods. As can be seen the variance of the unweighted average grade up to day t is very low and becomes even lower over time; after two weeks the line is almost flat. On the other hand when we look at the grade given on day t (without taking past values into account) the variance is increasing over time, this is because the amount of reviews goes down over time (thus one low or one high grade has a large impact on the daily average grade). The EWMA grade can be seen as a compromise between these two extremer methods and as can be seen the variance is more constant over time, thus supporting the idea of using the EWMA grade instead of the other two grades.
Figure 1: The effect of the different weighting methods for the grade of the movie ‘Harry Potter and the Half-Blood Prince’
By: Evert de Haan 18 of reviews on all previous days will have an effect on current revenue. The direct effect of reviews on revenue is equal to β2, the effect of the previous day’s amount of reviews is equal to β1* β2 and for the day before that it is(β1^2)* β2. The long-term (direct plus all lagged) elasticity of reviews can therefore be calculated as follows (Bond 2002):
(3)
The advantages of a dynamic panel data model is next to the fact that we are able to distinguish between short- and long-term effects, that we also will capture serially
correlated shocks, which can be crucial for recovering consistent parameter estimates (Bond 2002).
Next to the pooled model we will also be estimating a fixed effects version of the model which looks as followed:
ln(REVENUEit) = β0i + β1ln(REVENUEit-1) + β2ln(REVIEWSit) + β3REVIEWDUMit + β4ln(SEARCHit) + β5ln(GRADEit) + β6ln(SCREENSit) + β7ln(DAYit) + β8WEEKENDit + β9(ln(DAYit)*ln(GRADEit)) +
β10(ln(DAYit)*ln(SEARCHit)) + uit (4)
By: Evert de Haan 19 movie’s base-revenue. Next to these two models we will also be estimating the following random effects model:
ln(REVENUEit) = β1ln(REVENUEit-1) + β2ln(REVIEWSit) + β3REVIEWDUMit + β4ln(SEARCHit) + β5ln(GRADEit) + β6ln(SCREENSit) + β7ln(DAYit) + β8WEEKENDit + β9(ln(DAYit)*ln(GRADEit)) + β10(ln(DAYit)*ln(SEARCHit)) + Φi +
uit (5)
As can be seen the movie specific intercepts are now left out and changed into Φi. These are movie-specific random effects (Hsiao 1986). An assumption is that the data that is being analyzed consist out of a hierarchy of different populations for which the differences relate to that hierarchy. An advantage over the fixed effects model is that the random effects model is more efficient.
Since we are using panel data, when the dependent variable has a unit root this may result in overestimated t-ratios (Granger and Newbold 1974). Because of this we will apply the Harris-Tzavalis unit-root test (Harris and Tzavalis 1999), which is a unit root test designed for panel data where the null hypothesis is that there is a unit root.
In line with Khanna, Posnett, and Sandler (1995) we will first compare the fixed effects model with the pooled model with the use of a F-test, where a significant F-test indicates that the fixed effects model is preferred over the pooled model. After that we will compare the fixed effects model with the random effects model by applying the Hausman specification test (Hausman 1978), which is calculated as followed:
By: Evert de Haan 20 The null hypothesis in this test indicates that the random effects estimator is
consistent and efficient. The alternative hypothesis indicates that the random effects estimator is inconsistent, while the fixed effects estimator is consistent. Thus when the Hausman specification test is significant, the fixed effects model is preferred, otherwise the random effects model is preferred. After we have found the most appropriate model we will furthermore inspect the lag structure of the model in terms of the REVENUE, GRADE, REVIEWS and SEARCH variables. We will be doing this in a stepwise
approach; we will include one additional lag term at a time (i.e. first include t-1, then t-2, then t-3) and we will stop at the lag where the new variable is still significant.
In the next part we will estimate the models, select the appropriate model and test our hypotheses.
5. Results
In table 2 we show the correlation between the variables on the aggregated movie level, with the p-values between brackets. As can be seen in the amount of reviews is highly correlated with the search volume, thus movies that have many reviews also in general have many people that search for information. Furthermore it can be seen that these two variables are also positively correlated with the movie’s budget and revenue.
Table 2: Correlations on the aggregated movie level (n=165)
1. 2. 3. 4. 5. 1. Amount of reviews 1.000 2. Search volume 0.926 (<0.001) 1.000 3. Average grade -0.068 (0.383) -0.045 (0.564) 1.000 4. Budget (n=102) 0.563 (<0.001) 0.544 (<0.001) -0.189 (0.057) 1.000
By: Evert de Haan 21 The average grade is at the 10 percent significance level negatively correlated with the movie’s budget, indicating that higher budget movies are on average graded lower.
In table 3 it can be seen that on the daily observation level, the amount of reviews is still significantly correlated with the search volume, but to a lesser extend than with the aggregated level data. This reduces the potential problems of multicollinearity in our models. Furthermore we can see that the amount of revenue is higher correlated with the search volume than with the amount of reviews . Screens (the amount of theaters
screening the movie at time t) is also significantly correlated with both variables, and screens is also highly correlated with the movie’s daily revenue. We can also see that the search volume is negatively correlated with the grade, and so is the amount of screens. It could be the case that negative word-of-mouth generates a part of the online search volume, and that that the audience of widely released movies is more critical than the audience of smaller (independent) movies. Finally we see that the amount of reviews, the amount of revenue and the amount of screens all decrease over time, which is in line with what could be expected.
Table 3: Correlations on the daily observations level (n=10,993)
By: Evert de Haan 22 Before estimating the models we have performed the Harris-Tzavalis unit-root test for the dependent variable including a time trend and panel specific means. The test was highly significant (rho=0.743, p<0.001), indicating that revenue is trend-stationary (i.e. has no unit root).
After this we have estimated the three models, for which the results can be seen in Table 4 with the standard errors between parentheses and the p-values between square brackets. What we can first of all see is that the parameters of the pooled model and the random effects model are quite similar to each other, but the parameters of the fixed effects model do differ quite substantial from the other two models. A reason behind this is that both the pooled model and the random effects model are inconsistent if movie-specific effects are present (Cameron and Trivedi 2005).
By: Evert de Haan 23 Table 4: Parameter estimates
Variable Pooled Model Panel-data models
Fixed Effects Random effects
ln(REVENUEit-1) 0.729 (0.007) [<0.001] 0.361 (0.009) [<0.001] 0.726 (0.007) [<0.001] ln(REVIEWSit) 0.079 (0.016) [<0.001] 0.071 (0.017) [<0.001] 0.083 (0.017) [<0.001] REVIEWDUMit 0.034 (0.019) [0.078] 0.054 (0.019) [0.004] 0.035 (0.020) [0.077] ln(SEARCHit) -0.028 (0.022) [0.192] -0.050 (0.024) [0.034] -0.024 (0.021) [0.247] ln(GRADEit) 0.697 (0.084) [<0.001] 0.319 (0.099) [0.001] 0.780 (0.094) [<0.001] ln(SCREENSit) 0.257 (0.007) [<0.001] 0.524 (0.009) [<0.001] 0.264 (0.008) [<0.001] ln(DAYit) 0.107 (0.045) [0.018] -0.826 (0.054) [<0.001] 0.156 (0.050) [0.002] WEEKENDit 0.254 (0.013) [<0.001] 0.552 (0.013) [<0.001] 0.252 (0.013) [<0.001] ln(DAYit)*ln(GRADEit) -0.129 (0.024) [<0.001] -0.087 (0.028) [0.002] -0.151 (0.027) [<0.001] ln(DAYit)*ln(SEARCHit) 0.024 (0.005) [0.001] 0.050 (0.006) [<0.001] 0.025 (0.006) [<0.001]
Now that we know which model is preferred we will test the lag structure of this model. We have done this by first including one additional lag for the REVENUE, GRADE, REVIEWS, and SEARCH variables. The first additional lag for REVIEWS (t=-1.35, p=0.178) and for GRADE (t=-0.280, p=0.778) where not significant. We also have tested other lags of the REVIEWS and GRADE variables (up to five lags), but none of these where significant. So for both the REVIEWS and GRADE variable no lags will be included in the final model. This does however not mean that REVIEWS and GRADE have no lagged effect, but this effect is already captured by the lagged dependent variables.
By: Evert de Haan 24 Table 5: Parameter estimates
Variable Fixed Effects
(with additional lags) Coefficient Std. error p-value
ln(REVENUEit-1) 0.610 0.012 <0.001 ln(REVENUEit-2) -0.312 0.010 <0.001 ln(REVIEWSit) 0.079 0.016 <0.001 REVIEWDUMit 0.054 0.018 0.003 ln(SEARCHit) 0.240 0.035 <0.001 ln(SEARCHit-1) -0.144 0.034 <0.001 ln(SEARCHit-2) -0.093 0.025 <0.001 ln(GRADEit) 0.342 0.106 0.001 ln(SCREENSit) 0.568 0.009 <0.001 ln(DAYit) -0.790 0.080 <0.001 WEEKENDit 0.347 0.014 <0.001 ln(DAYit)*ln(GRADEit) -0.099 0.029 0.001 ln(DAYit)*ln(SEARCHit) 0.028 0.006 <0.001
With H1a we expected that the online search volume of a movie is positively related to the box office revenue of that same movie. We can see in Table 5 that the direct effect is indeed significant and positive, but this is partly compensated with two negative lagged effects. Due to these lagged effects the total (long term) effect becomes
By: Evert de Haan 25 Figure 2: Long term search elasticity over time
As can be seen in Figure 3, the long-term grade elasticity is for the first nine days of the movie’s release positive and significant, after which it becomes and stays
By: Evert de Haan 26 Figure 3: Long-term grade elasticity over time
What we furthermore have found is that the amount of reviews for a movies is significantly contributing to the box office revenue, that the revenue is decreasing over time, that the amount of theaters that are showing the movie is positively related to the revenue, and finally that revenue is higher during weekends. All of this is in line with previous studies.
6. Recommendations and conclusion
By: Evert de Haan 27 We have also increased the understanding in terms of the valence of word-of-mouth. What may be expected from theory, but for which contradicting results are available from past empirical studies, is that the valence of word-of-mouth is indeed positively related to the box office revenue, but that this effect is only significant in the early stage of the movie’s release. After nine days this effect becomes insignificant, which can explain why many authors have found a non-significant effect for the valence of word-of-mouth on box office revenue.
This also strengthens the idea to make a good first impression when it comes to releasing a new product; what people think of the product during the introduction phase is of crucial importance for the future success of a product. If there is negative word-of-mouth in the introduction stage of a product, then that may be something that cannot be recouped, even if the word-of-mouth becomes positive in a later stage of the product’s life-cycle. For managers it is therefore important that the first impression of a new product is positive, much more important than the impression after a product has already been on the market for a while. Furthermore for managers generating buzz for a new product may still be important, but the initial size of the buzz may not be a good indicator for future success. Providing consumers with good and valuable information is not only important in the introduction stage, but becomes even more important after a while.
By: Evert de Haan 29 References
Basuroy, Suman, Subimal Chatterjee, and S. Abraham Ravid (2003), “How Critical Are Critical Reviews? The Box Office Effects of Film Critics, Star Power, and Budgets,” Journal of Marketing, 67 (4), 103-117.
Basuroy, Suman, Kalpesh Kaushik Desai, and Debabrata Talukdar (2006), “An Empirical Investigation of Signaling in the Motion Picture Industry,” Journal of Marketing
Research, 43 (May), 287–295.
Bearden, William O., Richard G. Netemeyer, and Jesse E. Teel (1989), “Measurement of Consumer Susceptibility to Interpersonal Influence,” Journal of Consumer
Research, 15 (4), 473-481.
Bond, Stephen R. (2002), “Dynamic Panel Data Models: A Guide to Micro Data Methods and Practice,” Portuguese Economic Journal, 1, 141-162.
Cameron, A. Colin, and Pravin K. Trivedi (2005), “Microeconometrics Methods and Applications,” Cambridge University Press, New-York (NY, USA).
Chen, Pei-Yu, Shin-yi Wu and Jungsun Yoon (2004), “The Impact of Online
Recommendations and Consumer Feedback on Sales,” Twenty-Fifth International Conference on Information Systems, 711–724.
Chevalier, Judy, and Dina Mayzlin (2006), “The Effect of Word of Mouth on Sales: Online Book Reviews,” Journal of Marketing Research, 38 (August), 345–354. De Vany, Arthur, and Casey Lee (2001) “Quality Signals in Information Cascades and
the Dynamics of the Distribution of Motion Picture Box Office Revenues,”
Journal of Economic Dynamics and Control, 25 (3/4), 593-614.
Delre, Sebastiano A., Wander Jager, Marco A. Janssen, and Tammo H. A. Bijmolt (2010), “Simulating the Motion Picture Market: why do the hits take it all?,” Center for Research in Organization and Management Working Paper No. 10-001.
Duan, Wenjing, Bin Gu, and Andrew B. Whinston (2008), “Do Online Reviews Matter? An Empirical Investigation of Panel Data,” Decision Support Systems, 45 (4), 1007-1016.
Duan, Wenjing, Bin Gu, and Andrew B. Whinston (2009), “The Dynamics of Online Word-of-Mouth and Product Sales - An Empirical Investigation of the Movie Industry,” Journal of Retailing, 84 (2), 233-242.
By: Evert de Haan 30 Ebbes, Peter (2004), “Latent Instrumental Variables - A New Approach to Solve for
Endogeneity,” Labyrinth Publications, Ridderkerk (The Netherlands). Elberse, Anita and Jehoshua Eliashberg (2003), “Demand and Supply Dynamics of
Sequentially Released Product in International Markets: The Case of Motion Pictures,” Marketing Science, 22 (3), 329–354.
Eliashberg, Jehoshua, Anita Elberse, and Mark A.A.M. Leenders (2006), “The Motion Picture Industry: Critical Issues in Practice, Current Research, and New Research Directions,” Marketing Science, 25 (6), 638-661.
Goettler, Ronald L. and Phillip Leslie (2005), “Cofinancing to Manage Risk in the Motion Picture Industry,” Journal of Economics and Management Strategy, 14 (2), 231–261.
Granger, Clive William John, and Paul Newbold, (1974), "Spurious Regressions in Econometrics". Journal of Econometrics, 2, 111-120.
Hadida, Allègre L. (2009), “Motion Picture Performance: A Review and Research Agenda,” International Journal of Management Reviews, 11 (3), 297-335. Harris, Richard D. F., and Elias Tzavalis (1999), “Inference For Unit Roots in Dynamic
Panels Where the Time Dimension is Fixed,” Journal of Econometrics, 91, 201-226
Hausman, Jerry A. (1978), "Specification Tests in Econometrics", Econometrica, 46 (6), 1251-1271.
Hennig-Thureau, Thorsten, Mark B. Houston, and Gianfranco Walsh (2006), “The Differing Roles of Success Drivers Across Sequential Channels: An Application to the Motion Picture Industry,” Journal of the Academy of Marketing Science, 34 (4), 559-575
Holbrook, Morris B. (2005), “The Role of Ordinary Evaluations in the Market for Popular Culture: Do Consumers Have ‘Good Taste’?,” Marketing Letters, 16 (2), 75-86.
By: Evert de Haan 31 Irwin, Julie R., and Gary H. McClelland (2001), “Misleading Heuristics and Moderated
Multiple Regression Models,” Journal of Marketing Research, 38 (February), 100-109.
Khanna, Jyoti, John Posnett, and Todd Sandler (1995), “Charity Donations in the UK: New Evidence Based on Panel Data,” Journal of Public Economics, 56, 257-272. Kohli, R. and Sah, R. (2003), “Market Shares: Some Power Law Results and
Observations,” Harris School, Working Paper, series 04.1,
http://www1.gsb.columbia.edu/mygsb/faculty/research/pubfiles/rk35/Kohli%5Fm arket%5Fshares%2Epdf
Larceneux, Fabrice (2007), “Buzz and recommendations on the Internet. What impacts on box-office success?,” Applications en Marketing (English Edition), 22 (3), 43-62.
Li, Xinxin, and Lorin M. Hitt (2008), “Self-Selection and Information Role of Online Product Reviews,” Information Systems Research, 19 (4), 456-474.
Liu, Yong (2006), “Word of Mouth for Movies: Its Dynamics and Impact on Box Office Revenue,” Journal of Marketing, 70 (3), 74-89.
Miguel, Edward, Shanker Satyanath, and Ernest Sergenti (2004), “Economic Shocks and Civil Conflict: An Instrumental Variables Approach,” Journal of Political
Economy, 112 (4), 725-753.
MPAA (2010), Theatrical Market Statistics 2010. Washington D.C.: MPAA. Pokorny, Michael and John Sedgwick (2001) “Stardom and the Profitability of Film Making: Warner Bros. in the 1930s,” Journal of Cultural Economics, 25 (3), 157–184. Radas, Sonja, and Steven M. Shugan (1998), “Seasonal Marketing and Timing New
Product Introductions,” Journal of Marketing Research, 35 (3), 296-315. Sornette, D. Gilbert, T. Helmstetter, A. and Ageon, Y. (2004) Endogenous versus
Exogenous Shocks in Complex Networks: an Empirical Test Using Book Sales Rankings, Physical Review Letters, 93, 228701-4
Sood, Sanjay, and Xavier Dreze (2006), “Brand Extensions of Experiential Goods: Movie Sequel Evaluations,” Journal of Consumer Research, 33 (December), 352–360. Sorensen, A. T. (2004), “Bestseller Lists and Product Variety: The Case of Book Sales,”
working paper, http://www.stanford.edu/~asorense/papers/bestsellers.pdf Westbrook, Robert A. (1987), “Product/Consumption-Based Affective Responses and
By: Evert de Haan 32 White, Halbert (2001), “Asymptotic Theory for Econometricians,” Academic Press, New
York (NY, US).
Wray, Richard (2009), “IPhones and games consoles drive online search traffic,” link: http://www.guardian.co.uk/money/2009/dec/15/online-search-christmas-presents (retrieved 4th of July 2011).
