Popularity vs. True Quality
What determines the popularity of YouTube videos? Are the most viewed videos really the ones people like the most? What drives people choose certain videos? Answering these questions is particularly important because websites like youtube.com, yahoo.com, google.com or facebook.com make millions of dollars on advertising and these ads should be correctly placed. Did you ever click on a video just because it had a too many views, so you thought that it had to be good? Can we use real data to explain the popularity of Rebecca Black's Friday video?1 The problem is that you click ahead of time and most likely before knowing the true quality of the video (unless is BEP, of course).
Therefore sometimes you need to rely on a potentially misleading title, facebook or twitter links, total number of previous views or ratings from previous viewers.
This is not a concern for YouTube viewers only. Employers face a similar problem as they have to hire before knowing the true quality of the future employee, therefore they have to rely on the name of the institution that granted the degree, previous working experience and recommendation letters. In general, this is a typical problem in microeconomics, where consumption decisions are made before knowing the true quality of the good. YouTube videos are just an excellent example where the data is available.
The Data
The dataset on youtube.gtd contains a sample of 1150 observations, where each observation corresponds to a single YouTube video. These data actually uses an older YouTube format that had more information than the current format. The following is the list of variables:
vweeka: Number of views of the video during week A. (this is our measure of popularity of the videos)
cumview: Total number of views since the videos was uploaded until prior to week A. 1,860,220 in the BEP video above.
rati: Total number of ratings to the video prior to week A. 5,488 in the video above.
come: Total number of text comments to the video prior to week A. Not show in the video above.
favo: Total number of times the video was selected as "Favorite" prior to week A. Not shown in the example. (this is similar to the current "like" option)
star: Number of stars in the rating prior to week A. Goes from 1 to 5; 1 star = poor; 2 stars = nothing special; 3 stars = worth watching; 4 stars = pretty cool; 5 stars = awesome, like the video above.
min: Total time length in minutes. The video above with 4:56 would be 4.93 minutes.
days: Number of days elapsed since the video was uploaded until the beginning of week A. (this is a measure of how old the video is)
good: This is a dummy variable equal to one if star is 3 or greater (worth watching, pretty cool or awesome) and it is zero if it is two star or less (poor or nothing special).
old: This is a dummy variable equal to one if the video is older than 100 days and equal to zero if the video is days is below 100.
The Models and Analysis
a) Get the summary statistics table with all the variables. Discuss three interesting results from this table.
The model we will estimate is the following:
vweeka = b + b cumview + b rati + b come + b favo + b star + b + b days +e 0 1 2 3 4 5 6 7 min
b) Before you estimate this model in Gretl, what signs do you expect the slope coefficients will have? Positive, negative or no effect. Just write the sign and briefly explain your thoughts.
c) Estimate the model in Gretl (or other statistical software). Do the coefficients have the sign you expected?
d) What is the interpretation of the coefficient cumview? Do videos with more prior views more likely to receive more viewers during week A? Can you say that videos that were already popular are more likely to stay popular?
e) How can you explain that the number of ratings affects negatively the number of views in week A?
f) Can you claim that more discussed videos (as measured by come) are also more popular?
g) Under certain restrictions, we can argue that favo and star should be a measure of quality of the videos. What concerns do you have about favo and star as measures of quality of the videos? Do 'higher quality' videos are also more popular?
h) Are longer videos more popular? Don't forget to look at the p-value and the interpretation of the coefficient.
i) Are older videos more popular? Don't forget to look at the p-value and the interpretation of the coefficient.
j) What other variables not included in the model do you think can impact vweeka?
Estimate this slightly different specification:
vweeka = b + b cumview+ b rati + b come + b favo + b star + b + b days +e 0 1 2 3 4 5 6 7 log( ) min
k) What is the interpretation of β1 and β6?
Estimate the model:
vweeka = b + b cumview + b rati + b come + b favo + b star + b + b days +e 0 1 2 3 4 5 6 7 log( ) log( ) min
l) What is the interpretation of β1?
Estimate the following model:
vweeka = b +d old +d good +e 0 0 1 log( )
m) What is the interpretation of δ0 and δ1?
n) We can say that if old = 1 the video is old. If old = 0, the video is young, if good = 1 the video is good, and if good = 0, the video is bad. Following the example in Equation 6.6 in the class notes, create 4 groups. (oldgood, oldbad, youngbad, younggood) Let oldgood be the base group. Are old good videos more popular than young good videos?