Book prizes and sales

I’ve always been wondering whether different kinds of awards increase the sales of certain goods. Say a literary award for books or an award for a wine. Turns out there is a rather extensive literature on the topic and the answer appears to be yes.

Mostly difference-in-differences models have been used to answer this question, but recently Ponzo and Scoppa (2013) came out with a working paper in which they use a technique called regression discontinuity design (RDD), which I find quite interesting.

The authors study the effect of a presitigious literary prize called the Strega Prize on Italian book sales. The prize appears to be a very serious one. It is awarded once a year to one book based on the votes of 400 experts.

All models estimated by the authors indicate that the Strega Prize indeed has a highly significant effect on sales after it is awarded. There are a bunch of robustness checks in the paper as well.

What is really interesting to me, however, is the regression discontinuity design they use. Basically what this entails is the following: there is a certain threshold above which the independent variable (the book sales) received a treatment (the prize), below which they didn’t.

Think of it this way: suppose for a second that every book that gets at least 50 votes is awarded the prize. Assume that Book A received 45 votes, Book B 55 votes, Book C 295 votes and Book D 5 votes. Then Book B and C received the treatment. But how can we really tell whether their (potentially) higher sales are due to the prize or due to the fact that they are inherently better books that just simply sell better. And since they’re better, they received more votes as well.

Of course, it’s clear to see how one could use a diff-in-diff analysis here. Roughly speaking, if Book C sold twice as well as Book A before the prize, and after being awarded the prize it sold three times as well, then that extra factor of sales can be attributed to the prize.

An RDD model is somewhat different. The main argument is that while Book A did not receive the prize, Book B did. However, the difference in votes between A and B is negligible. In an RDD model we assume that books just above the threshold (like Book B) would perform the same as books just below the threshold (like Book A), at least without treatment. So if there are any differences between the sales of Book A and Book B after the award, it is due to the prize.

Now, in reality only the book with the most votes gets the prize, so the authors transformed all the votes by substracting the number of votes the second book received plus 1 from each book’s votes. This way, only the winner had positive votes, the second book had -1 votes and the other books had various negative votes. So the threshold for winning the prize was normalized to 0 for all years. The model to be estimated then looks like this:

\ln{Sales_{it}} = \beta_0 + \beta_1 {Prize}_{it} + f({Votes}_{it}) + {Year}_{it} + \theta X_{it} + \epsilon_{it}

where Prize is a dummy, X is a vector of control variables and

f({Votes}_{it}) = \gamma_1 {Votes}_{it} + \dots + \gamma_k {Votes}_{it}^k.

The authors estimate models with k = 1, 2 and 3. In this case thus, the intrinsic quality of the books (as measured by the number of votes) is captured by the gamma coefficients in f(Votes). Therefore, the coefficient of the Strega Prize, beta_1, measures the effect of the prize and it is unaffected by quality differences among books.

This specification can perfectly capture a discontinuity or jump that can be expected in sales due to the prize. Just consider a book that got -1 votes (a book in the second place) and a book that got 1 vote in transformed votes (i.e. a book that won with only 2 votes ahead of the runner-up). Normally, 2 measly votes should not matter but this is where the model can produce discontinuity thanks to the Prize dummy variable. Using coefficients from the simplest model the authors estimated we have that in our theoretical example

\ln{Sales}_1 = 1.9746 \times 0 + 0.0143(-1) = -0.0143

\ln{Sales}_2 = 1.9746 \times 1 + 0.0143(+1) = 1.9889.

While based on the intrinsic quality of the books, one would expect a 2 * 0.0143 = 0.0286 difference in the logs of sales, the model predicts a huge difference of roughly 2.

Here is a picture of this discontinuity at the threshold:

discontinuity_rdd

This picture shows the model where f(Votes) was specified to be a third-order polynomial, as the shape of the fitted curve attests.

As mentioned above, the authors carry out a lot of robustness checks and they also use a diff-in-diff model to analyze the effect of the Strega Prize. They arrive at the conclusion that the Prize significantly increases sales, it has a huge effect.

Perhaps this is not so surprising given the prestige of the Prize, which is probably quite well-known and well-communicated in Italy. The main reason I found this paper interesting is because of the methodology used.

On the other hand, it would be interesting to see whether not so well-known, potentially bogus prizes also increase sales. For instance, if you check out a grocery store’s wine section, you will find multiple wines with stickers attesting that they won all kinds of awards. Surprisingly, this goes even for cheaper wines. I am no expert, but clearly many of those awards are sort of bogus. Also, most consumers have no idea about which wine award is prestigious. So in a situation when information is so imperfect that consumers can’t tell a “phony” award from a prestigious one, how much do awards matter? Something perhaps for future work.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s