We find that file sharing has no statistically significant effect on purchases of the average album in our sample. Moreover, the estimates are of rather modest size when compared to the drastic reduction in sales in the music industry. At most, file sharing can explain a tiny fraction of this decline. This result is plausible given that movies, software, and video games are actively downloaded, and yet these industries have continued to grow since the advent of file sharing. While a full explanation for the recent decline in record sales are beyond the scope of this analysis, several plausible candidates exist. These alternative factors include poor macroeconomic conditions, a reduction in the number of album releases, growing competition from other forms of entertainment such as video games and DVDs (video game graphics have improved and the price of DVD players or movies have sharply fallen), a reduction in music variety stemming from the large consolidation in radio along with the rise of independent promoter fees to gain airplay, and possibly a consumer backlash against record industry tactics.26 It is also important to note that a similar drop in record sales occurred in the late 1970s and early 1980s, and that record sales in the 1990s may have been abnormally high as individuals replaced older formats with CDs (Liebowitz, 2003).

[:24]

Like most other authors they do a panel data regression of albums sales over time against instrumented estimates of downloading (in an attempt to eliminate the default correlation between downloading and sales that will otherwise bias estimates). They also first difference after finding stationarity in their sales data.

This study has sparked a great deal of debate and has detailed rebuttal from among others Liebowitz. The main shortcoming is to estimating the interaction of illegal downloading and legal sales by looking at contemporaneous changes, i.e. what effect downloading in week X has on sales in week X. However it seems clear that substitution can occur across time not simply contemporaneously.

Futhermore as Blackburn argues persuasively, pooling across albums is a major problem (a problem not of course confined to this paper but present in all of the other papers discussed). As Blackburn demonstrates pooling may lead to a 0 or even +ve effect across all albums (because positive effects on some albums cancel out negative effects on others). However this does not mean that overall effect on industry is negative since the per-album effect may be correlated with per-album sales.

Summarizing, if music downloading reduces the probability of buying by 30%; if 15% of the population download music; if downloaders are twice as likely to buy music than non-downloaders; if — conditional on buying — downloaders and non-downloaders buy the same quantity of units, under all these assumptions sales in 2002 would have been 7.8% (0.3*[(0.15*2)/115)]) higher than the level they experienced.

[zentner_2003:23]

Zentner, who is a PhD student at Chicago, produced one of the first papers in this area but there are reasons to have reservations about Zentner’s analysis and his conclusions. For example in his calculation of loss he assumes that downloaders are likely to buy twice as much music on average as non-downloaders. But this entirely ignores the frequently cited possibility that downloaders large propensity to purchase is positively correlated with downloading (the so-called sampling effect). If this is so you cannot simple multiply estimated reduction in purchases due to (illegal) downloading with their likelihood to purchase uncorrected for downloading (the two variables may be correlated in opposite directions which means the uncorrected estimates overestimate loss). To illustrate suppose that normally all individuals would purchase the same amount of music but downloaders hear more and therefore would purchase twice as much. Then in this framework downloading actually increases sales by (1-0.3)*200 – 100 = 40%.

Another issue is his method for dealing with unobserved heterogeneity in taste for music by taking an instrumental variables approach. His chosen instruments are measures of internet sophistication of broadband access. However both these sets of variables while clearly correlated with downloading (as is intended) are also likely to be correlated with unobserved heterogeneity (through indirect variables such as living in a city).

Pietz and Waldbroeck

We have analyzed the RIAA`s claim that music downloads are causing a substantial decrease in music sales. Our macro data confirm their fear: we find that music downloading could have caused a 20% reduction in music sales worldwide between 1998-2002. While this is only a crude estimate, we believe that it is a good reference value that other studies, especially microeconometric ones, could use to assess the exact substitution that has taken place between CDs and MP3s. Our analysis also reveals that other factors than music downloads on file-sharing networks are likely to be responsible for the decline in music sales in 2003.

[:78]

Moreover in several of the regressions including the most complete one (i.e. the one with all explanatory variables included) downloads are not statistically significant (at the 10% level).

We argue that successfully measuring the possible sales-displacing effect of unpaid music downloading requires data on the quantities of purchases and downloads made by individuals, leading us to conduct original surveys. Using a variety of empirical approaches, we document that downloading displaces sales among a convenience sample of college students. The estimate we consider most conservative indicate that an additiona l download reduces sales by between 0.1 and 0.2 units. As a result, for the individuals in our sample, downloading reduced expenditure by about 10 percent but possibly much more. Supporting incomplete sales displacement is our finding that downloaded music is valued much less than purchased music.

While downloading reduces expenditure (on hit albums, 1999-2003) by $25 per capita in the sub-sample for which we perform a direct welfare analysis of downloading, it raises sample consumers` welfare associated with these albums by $70 per capita. Some of the benefit to consumers are transfers from sellers, but most of the benefit ($45 per capita) comes from reductions in deadweight loss.

Two facts bear emphasis again. First, our sample is not representative, so our results should not be generalized. Second, our evaluation of welfare takes supply as given. It is entirely possible that downloading has important effects on the quantity and types of music recorded and marketed in the first place. This is an important area for further research.

[:28-29]

Due to its survey design and acquisition thereby of individual level quantity and (self-reported) valuation data able to address the more fundamental question of social welfare. Point out that survey based studies are the only real way to identify as they allow us to monitor individual changes and avoid the issues that pooling brings (for example a negative relation between donwloading and sales could simply indicate that downloaders and buyers have different tastes p. 8).

But the really important aspect of this paper is that is among the first to obtain the data necessary to estimate the welfare impact of downloading. They find a consumer welfare benefit of $70 per person of which around $45 is social welfare benefit. This is a very large amount and over twice the loss estimated in terms of reduced sales.

In particular, the point estimates imply that the median ‘new’ artist, whose weekly sales are 2,163 albums, would see a decrease in weekly sales of 101 albums per week were files shared to be reduced by 10%. A similar calculation can be made for an artist of maximum popularity. At the median level of sales for these artist, the estimate implies an increase in sales of 490 albums per week if file sharing were to be reduced by 10%. This stark contrast between the magnitudes of the effects for artists of varying levels of popularity highlights the importance of this heterogeneity in estimating the aggregate effects of file sharing.

…

A similar calculation can be made for estimating the total effect of file sharing on sales. To estimate the aggregate effect of a 30% reduction in file sharing across the board,33 I simply subtract out the effect of the deleted files from the second stage estimation in Table 4 and then aggregate up to market level numbers using the appropriate weights. The estimated effect of such an across-the- board reduction in file sharing is to increase aggregate sales by 15%. Again, while these calculations were useful for placing the analysis inside the framework of the previous literature, they do not take into account competition effects across albums, and so the effects of file sharing will be overstated in these estimates.

[:21]

Introduces a multinomial logit model to allow for competition effects (i.e. fact that buying one album reduces likelihood of buying another as budget is fixed).

Doing this reveals that as a result of the lawsuit strategy followed by the RIAA against users of file sharing networks, album sales increased by 2.9% over the 23 weeks in the data sample after the strategy was announced. During this period, actual record sales in the U.S. were an average of 11,470,652 albums per week, based on national level data reported by Billboard magazine (2003) each week, and thus would have been 11,147,378 per week in the absence of the reduction in file sharing caused by the lawsuit strategy. Again using a baseline of $5 markup per CD, this translates to an increase in industry profits of $1,616,370 per week, or $37 million over the 23 week period after the lawsuit strategy was announced to the public. Note that, as expected, this increase in profits is much smaller than the number obtained using the simple reduced form estimates (approximately $160 million), which fail to account for the effect of competition among albums.

[:30]

p.32 provides a very interesting table that lists the estimated effect of 30% less file-sharing on artists depending on their position in the popularity distribution. By percentile (with 1% being lowest selling, 100% the highest selling) we have the break even point at the 75th percentile: that is the bottom 3/4 of artists gain from file-sharing while the top 1/4 lose.

Percentile | Actual Sales | Sales with 30% less file-sharing
1%                73             70
5%                170            166
10%               281            277
25%               757            745
50%               2852           2851
75%               10110          9831
90%               26531          26934
95%               45255          47357
99%               133983         165054

Note: reason why aggregate estimates (i.e. without accounting for popularity) can yield no effect while actual effect is negative is that popularity is interacted with download quantity ie. phi * P_{i} * q_{i}^{t}. Since download quantity is correlated with popularity this means setting phi to 0 is not innocent.