File Sharing and Film Revenues: Estimates of Sales Displacement at the Box Office

4.1 Contemporaneous Downloads

Our empirical model aims to quantify the sales displacement effect from illegal P2P torrent downloads on box office revenue. The approach is similar to Oberholzer-Gee and Strumpf (2007), with a number of important differences. First, and most obviously, our context is theatrical film revenues rather than music sales. This removes the complicated issue of transforming single downloads (sales) to a proxy for album downloads (sales). Second, the data are observed at daily, rather than weekly, levels. Also, our data are observed at the state/territory level, rather than national level. Third, our identification strategy relies on a downloading cost function which relates to contemporaneous downloads in other Australian states (proxying for number of peers), US DVD release date, and significant structural changes in the Australian internet service industry over our sample period. We discuss identification in more detail below.

Our empirical model is based on the following equation

where Rist and Dist define revenue and downloads of film i in state s on date t, respectively. The logarithmic transformation is applied to revenue and download data since both distributions are bounded below at zero and are right skewed. ^[27] An additional benefit of the log-transformation is that the impact of downloads on revenues can be interpreted as an elasticity. We partition control variables into observable and unobservable vectors xist′ and uist′, respectively. In particular, vector xist′ includes the (time-variant) number of theatres showing film i in state s on date t (THist) which also enters in log-form; the week of the run of film i in state s on date t (WKist); as well as a set of dummy variables for day-of-week effects (DWd, where d indexes day). Additionally, we include state/territory fixed effects ηs and film fixed effects vi. ^[28]

We seek to make causal inference on the parameter φ. However, because popular films at the box office are also likely to be popular on P2P networks, download activity should be treated as endogenous. In terms of eq. [1], this implies that Eu′istδ+ϵist|Dist≠0. To remedy this problem, we employ two-step efficient generalized method of moments (GMM) estimation with robust standard errors (Hansen 1982). ^[29]

Our instruments are chosen based on the assumption of a downloading cost function with “cost-shifters” which would not be expected to be related to demand for theatrical consumption. We propose the following

where CistD is the cost of downloading film i in state s on date t, Pit is the number of peers of film i on date t, DVDit is a binary variable relating to whether film i had been released on DVD (in the US market) at date t, and t represents time (trend) index. The first argument of the cost function relates to the number of peers who are (contemporaneously) “seeding” the film file. Typically, the more seeders in a swarm, the faster the download and the more likely the chance of a successful download. Therefore, we would expect ∂C/∂P<0. Although we don’t observe actual information on the number of peers, we proxy this variable with the (contemporaneous) total number of downloads made across all other Australian states on date t. Regarding the second argument, we expect downloading costs to be inversely related to the US DVD release date under the assumption that once the DVD has been released, higher quality torrent downloads are more likely to exist on P2P networks. This is supported by the second mode observed in Figure 1 that suggests an increase in downloading activity approximately three-month post-theatrical release. Therefore, we expect ∂C/∂DVD<0. The final argument is a simple time trend, which enters the first-stage regression linearly allowing for an exponential fit in levels given the dependent variable (downloads) is measured in log form. As reported in Section 2, the period 2010–2011 saw significant increases in broadband subscriptions, speeds and data allowances in Australia. As a result, downloading costs have fallen significantly over our sample and we conjecture ∂C/∂t<0. We discuss identification issues in more detail in the following section.

4.2 Identification

Identification in our model derives from the use of film fixed effects and inclusion of instruments in the GMM estimation procedure that relates to downloading activity in other states (as a proxy for total number of peers), US DVD release date and a trend variable to account for structural changes in the Australian internet industry.

The primary endogeneity problem we face is one of unobserved film quality that makes popular films at the box office similarly popular on P2P networks. To address this at a global level, we use film fixed effects. Of course, this implicitly treats the endogeneity problem as constant through time, which may not be true. For example, a film may benefit from a national advertising campaign or awards after its release. Our model deals with such temporal (global) demand shocks implicitly by the use of theatres and the assumption that cinema managers respond dynamically to demand shocks by adjustment to the number of daily theatres. The instruments in our GMM estimation, most notably summed downloads in other states, therefore handle a more subtle form of time-varying endogeneity that occurs at a state (local) level, rather than national (global) level. Because we use downloads in other states, the instruments cannot assist in identification when demand shocks occur at the national level. Rather, the type of time-varying demand shocks that our instruments can assist with is state-level effects that may be orthogonal to other states. For example, this could include local newspaper critic reviews or localised word-of-mouth via social networks. We now discuss the instruments we use in more detail.

There are three well-known requirements for a good instrument: (1) that it itself does not belong in the focal relationship, (2) that it is uncorrelated with the error process, and (3) that it is correlated with the endogenous variable. Taken together, the instrument’s effect on the outcome variable occurs only through its association with the endogenous variable. Our first two instruments clearly satisfy the first requirement. By assumption of geographic separation, downloads in other states couldn’t displace sales in a focal state. And US DVD release date certainly has no place as a determinant of Australian box office revenues. As for trend, we note three reasons why we do not believe trend should appear in the focal (revenue) relation: (1) the relatively short time frame of our sample (20 months), (2) no significant increase in cinema infrastructure, and (3) no significant trend effect in sales. ^[30] With the significant changes in internet provision previously discussed, it is therefore appropriate to consider the trend as also satisfying the first requirement outlined above and be excluded from the focal equation.

The second and third requirements deserve further attention. Regarding the first instrument, the use of summed downloads in other Australian states is taken as a proxy for the number of peers (seeders) in the P2P network and, as such, should be inversely related to downloading costs. That is, more downloads in other states suggests more seeders generally on a network and as a result a higher chance of a quicker and successful download. In some respects, the use of the total number of downloads from geographically separated markets as an instrument is similar to the use of (average) price from other markets when modelling demand for differentiated goods (e.g. Hausman, Leonard, and Zona 1994; Nevo 2001). The intuition is that prices are correlated through common marginal cost shocks. Assuming the errors in demand are independent across market, this “cost shifter” approach is valid.

In our context, we similarly require an independence assumption for total downloads in other states to be a valid instrument. Specifically, the unobserved elements of uist′ in eq. [1] should not also be correlated with our instrument. We argue this is true based largely on the panel structure of our data permitting us to use film fixed effects to absorb much of the film level heterogeneity, which is the major source of our endogeneity. ^[31] We do, however, include instruments to assist with further identification owing to potential state-level (local) time-varying effects. We implicitly assume that the remaining (time) variation in the instrument is then more reflective of technical aspects and shocks related to downloading than any systematic variable which relates to unobserved demand. ^[32] For example, a low number of downloads could reflect a poor quality torrent file (e.g. camcorder version) or a problematic tracker (i.e. the server which directs the torrent uploads and downloads not responding). ^[33] As a result, low number of peers may indicate high chances of download failure and associated higher cost (wasted time and data capacity). It is also typically true that more seeders imply a faster download with higher probability of success. Downloaders therefore preference torrents for which they can see a high seeder ratios. At the margin therefore downloaders would prefer to download titles with more active swarms.

Regarding the second instrument, the US DVD release date likely leads to high-quality torrent files becoming available on P2P networks. As a result, downloaders face lower (net) costs of downloading – or, alternatively stated, receive greater (net) benefits. We conjecture the marginal consumer who is sensitive to quality is more likely to download a DVD quality torrent file over, for example, a low-quality cam version.

The third instrument we make use of is related to the changing structural features of the Australian internet landscape. As discussed in Section 2, there were significant increases in data offerings and speeds from Australian ISPs over the sample period. As a direct result, broadband subscriptions, speeds and data downloaded jumped dramatically. In this environment, it would seem that illegal downloading costs would have declined and downloading increased as a result. To capture this we include a simple linear trend as an additional instrument in our model.

4.3 Dynamic Stock of Downloads

The model described in eq. [1] captures the potential sales displacement for film i, in market s, on date t. There are, however, a number of reasons it would seem more appropriate to consider the total number of downloads over some window of time prior to, and including, date t rather than only those downloads observed contemporaneously. For example, illegal downloaders may browse torrent sites for new titles and may initialise a download with the intent of viewing the download at a later time – and, we assume, forego the cinematic alternative. In addition, downloading speeds (at least during the period of analysis in Australia) could be restrictively slow preventing instant playback necessitating files are downloaded in advance of actual consumption. Further, there is also the possibility that once a download has been completed, an individual may share the file with friends which would similarly contribute to potential future sales displacement effects. For these reason, we consider it likely that downloading over some window of time affects future box office sales which would not be captured in our contemporaneous specification. We subsequently refer to this alternate approach as a dynamic stock of downloads in our empirical methodology and consider the modified model:

where DisT defines downloads of film i in state s over a period of T days, where T∈{7,14,21,28}, prior to and including the date of observation. ^[34] Alternatively stated, DisT=∑r=0TDis(t−r). Essentially, the modification simply considers the dynamic stock of downloads over one, two, three or four weeks prior to (and including) date t. Our first instrument related to number of downloaders in other states, as discussed above, Dis′T, is also similarly redefined and now represents total downloads of film i in all other states/territories s′≠s over T, or Dis′T=∑r=0TDis′(t−r). All other variables remain as specified above.

This transformation requires further discussion in relation to role of our instruments. In particular, how to interpret the dynamic stock variable in the cost function. The number of downloads per T is now more reflective of aggregated popularity of the torrent on the P2P networks. The time-invariant heterogeneity will, once again, be picked up though film fixed effect. But the temporal variation will pick up changes in technical aspects of the torrent file. As an example, take the case of a typical film at the box office which is enjoying an average reception and theatre/screens decease familiarly over some number of weeks. Suppose up until week three only a bad (camcorder) torrent was available for this film. As a result, educated downloaders avoided this torrent and only low downloads were observed. Assume, for sake of argument, beyond the third week a better torrent appeared. This would increase numbers of downloads, increased number of peers, and consequently reduce download costs. This temporal variation is entirely related to the changes in the P2P offerings and the instrument is constructed to capture this type of variation. We continue discussion of identification in Section 6.

5.1 Contemporaneous Downloads

Table 3 provides GMM estimation results for the base model where downloads are treated contemporaneously to revenue (i.e. on the same date of observation). The first column reports simple OLS estimates without film fixed effects. The coefficient on the key variable of interest, contemporaneous downloads (Dist), is positive and significantly different from zero at 1%. The other estimated coefficients conform with a priori expectations. Specifically, week-of-run (WKist) and contemporaneous theatres (THist) reveal negative and positive signage, respectively. Again, both are statistically significant. Inclusion of film fixed effects in the second column of Table 3 purges some of the endogeneity between revenue and downloads that exists because of unobservable shared tastes for more popular films. Notably, the estimated coefficient on contemporaneous downloads is reduced but is still significantly positive. The third and fourth columns represent the first- and second-stage GMM estimation defined in eq. [1], respectively. The first stage estimates reveal a strong and significant positive relation with downloads in other states (Dis′t) and the US DVD release date (DVDit), but insignificant relation with trend (t) (although signage is as expected). Using the Kleibergen and Paap (2006) LM test for under-identification, Cragg and Donald (1993) test for weak identification, and the Hansen J test for over-identification (see, for example, Hayashi 2000) we find our instruments reject under- and weak-identification but do not reject over-identification. However, it is well known that this test suffers in large samples.

The second stage results of the GMM estimation reveals a further decrease in the magnitude of the estimated coefficient on contemporaneous downloads when compared to the OLS specification with film fixed effects. The positive relation between downloading activity and revenues is still apparent and significant, but the magnitude of the effect has been substantially reduced. This supports that the excluded variables (instruments) are helping to identify the model by reducing the positive bias induced by the endogeneity of downloads. The other estimated coefficients in the second stage regressions are similar in magnitude and signage to the OLS results with film fixed effects of column 2.

Time-invariant and film-specific variables – such as budget, cast/director appeal (pre-release) advertising, genre, classification rating, etc. – are implicitly captured in our model by the inclusion of film fixed effects. To examine the contributions of some of these variables, we extract individual film fixed effects and consider them against the time-invariant film-specific variables observed in our data set. The correlation between budget and the extracted fixed effects is 0.50 (compared to a correlation of 0.65 with total film revenues); and the correlation between (national) opening week screens and fixed effects is 0.72 (compared to a correlation of 0.82 with total film revenues). In a basic OLS regression with fixed effects as the dependent variable, estimated coefficients of both (log) budget and (log) opening screens are positive and significant at the 1% level of significance with an R2=0.44. Both relationships still hold at 1% when controls are added for the categorical variables of sequel, genre and classification rating with R2=0.64. In comparison, a regression of (log) total revenue on the same set of covariates found similar explanatory evidence in terms of signage and significance with R2=0.81. To the extent that the extracted fixed effects correlate strongly with key variables, which have been shown as important attributes of overall film revenues, these findings validate the use of film fixed effects as serving the model to capture the time-invariant determinants of demand for the films we observe. ^[35]

5.2 Dynamic Stock of Downloads

As discussed in Section 4, we consider it likely that individuals make theatre attendance decisions after a download decision. Under this assumption, it would be more appropriate to consider the number of downloads over some window of time prior to the actual date at which box office is observed. Table 4 provides evidence when this window of time is considered at one, two, three and four weeks prior to (and including) the date of observation. We report only results from the fixed effects and GMM models with all three instruments (i.e. sum of downloads from other states, US DVD release date and trend). In all four models, T= 7,14,21,28, the (log) downloads coefficient from the GMM estimation is less than the OLS fixed effects (as also observed within the contemporaneous results reported in Table 3) providing evidence of a reduction in the inherent bias of this coefficient. However, unlike the contemporaneous case, the relationship between downloads and revenues is now observed to be statistically negative implying a sales displacement effect. The (absolute) increasing magnitude of the download coefficient over the four models, T= 7,14,21,28, is a simple manifestation of the increasing dynamic stock. The estimated coefficient suggests a sales displacement elasticity in the range 0.06–0.31. We discuss the economic interpretation of these results further below.

Regarding the first stage results, the primary instrument relating to download activity in other states continues to be positive (as expected) and highly statistically significant across all four models. Also, the US DVD release date and trend variables are highly significant with positive signage. This is consistent with our a priori intuition related to the cost of downloading being inversely related to the availability of a DVD quality torrent file and also that download costs have fallen due to the changing structural features in the Australian internet service provision industry.

6.1 Instruments and Identification

Correct causal inference depends critically on the model being correctly specified. In modelling a potential sales displacement effect, the inherent difficulty lies in purging the simultaneity between downloads and revenues which manifests in a positive bias on the estimated coefficient in the absence of remedial measures. Fixed effects provide a solution for the time-invariant component of the endogeneity in panel data sets such as ours as the (time-invariant) heterogeneity between films can be accommodated. However, full identification of the effect requires an instrumental variable which also has some temporal variation and therefore removes remaining time-varying unobserved heterogeneous effects. We argue that our instruments are strong and satisfy both economic and statistical requirements. However, it is necessary to restate that the temporal endogeneity we treat with our instruments (specifically, downloads in other states) is related to what we term local (state) effects, rather than global (national) effects. For example, temporal demand shocks that occur due to local media effects and/or local network effects.

Intuitively, identification beyond fixed effects stems from an assumption that temporal changes in downloading activity relate more to technical aspects of file sharing rather than unobserved temporal (global) shocks to demand. This is not to say that there is no unobserved relation between box office revenues and downloads, but simply to say that it is not one that is systematically time varying beyond that accounted for by the inclusion of other control variables (e.g. theatres) and is thus accounted for by film fixed effects.

It is worth further interrogating this critical assumption. As discussed previously, there is similarity in approach to the use of (average) price in geographically separated markets (aka Hausman-type instruments) to instrument price in a focal market when modelling demand for differentiated products. However, many researchers have also noted this approach becomes invalid if unobserved demand shocks are correlated between markets. For example, national advertising/marketing campaigns may generate an unobserved shock to demand and could potentially increase prices. In our context, advertising is one possible source of time-varying heterogeneity which may affect both downloads and revenues at the (global) national level and isn’t observed in our model. However, it is well documented that the large majority of typical advertising spending on a movie occur pre-release (typically in excess of 90%) and therefore would not affect demand temporally beyond what is captured by the film fixed effect.

Similar concerns may also be put forward for other types of post-release promotion of the film. This could, for example, be critics’ reviews or award nominations and/or wins. While we cannot rule out these effects entirely, most films are reviewed prior to (or coinciding with) opening weekend. Additionally, the release delay between the US and Australian markets means many reviews are already available online prior to the Australian release. Award nominations/wins are potentially an issue that could cause an unobserved demand shock to both revenues and downloads. However, the relatively low number of films in our sample which were playing in cinemas when major award nominations and/or wins took place leads us to believe this would not be a significant effect. Moreover, the types of films we consider in our sampling frame (i.e. US wide release films) are often not those that are typically nominated or receive major awards. ^[36]

While we have explained the theoretical justification why our instruments – in particular, the summed downloads of other states – are valid, it is also of value to evaluate the statistical evidence a little more closely. The results of the first stage regressions reveal a particularly high R2 which indicates high explanatory power of the (included and excluded) instruments. This may create concern that the correlation is too high in the sense that the first stage is just recovering the endogenous variable. This would certainly be true if there was near perfect correlation between downloads in the focal market and other states. However, a simple regressions of lnDisT on lnDis′T for T= 1,7,14,28 reveal R2 in the range 0.22–0.34. In terms of simple correlations, the range is 0.47–0.58 between the endogenous (log) download variable and respective (log) downloads in other states. It is apparent that even though there is reasonably strong correlation between the endogenous variable and the instrument – the statistical requirement – it is far from simply recovering itself and the other instruments play an important role. As well, in terms of the statistical requirement that the instrument is uncorrelated with revenues, the condition appears to be well satisfied with simple correlation in the range −0.11–0.08. More formally, it was also observed that the statistical tests rejected under-identification and weak-identification, but not over-identification. However, as noted above, we believe this is in part due to large sample size.

6.2 Opening Week Revenues and Release Delay

One potential criticism of our model is that we are observing daily film revenues over the theatrical life of a film, which would typically be decreasing, whereas the dynamic stock variable may be increasing. Although we include a week-of-run variable in both the first and second stage regressions, this potentially inverse relation may be driving the results. To address this potential issue, we restrict the model described in eq. [1] to only model revenues of films in their opening week of theatrical release. This means the variable WKist is now redundant as all films are observed in their first week of release. Table 5 provides results for the second-stage regressions for the contemporaneous and dynamic models. In all cases the coefficient on the downloads variable is statistically less than zero implying that first-week revenues are subject to a sales displacement effect. As illustrated graphically in Figure 1, the fact that many films are subject to a release delay between the US and Australian markets seems a logical explanation of the decreased first-week sales when there is an opportunity for download prior to the opportunity for legal consumption. ^[37] In our sample of 166 films, the average release gap between the US and Australian markets was 28 days (median of 13 days). One film, Thor, was released in Australia two weeks prior to the US release, a further six films were released in Australian cinemas one week prior to the US release, and 50 films had a simultaneous release with the US release (opening within one day of the US release). The remainder of films had a positive release gap with the greatest gap being Diary of a Wimpy Kid which opened in Australia cinemas six months after its US release.

A simple regression of (log) downloads (observed within the first week) on observation date relative to the US release (controlling for day-of-week, state and film), retrieved a significant positive relation suggesting an increase in the number of first week downloads the more time elapsed since the US release, ceteris paribus. Given that illegal copies of films typically show up after the US release, this evidence is supportive of the release delay between the US and Australian markets providing opportunities for illegal consumption before the film is theatrically released in Australia. When release delays between the US and Australian markets are significant, and extend beyond the US DVD release date, the likelihood of high-quality torrents files appearing on networks increases and increased levels of downloading would displace more sales.

6.3 Weekly Model

The model outlined in Section 4 was estimated with daily data. To examine whether this feature of our data has any bearing on results we also consider a model where revenue and downloads are considered at the week level, rather than daily. We do this both for the contemporaneous and dynamic stock models. The contemporaneous model is analogous to eq. [1] but now the time subscript t represents a week, rather than a day, and is redefined tW. The theatres variable is redefined as THistW(max) so it now represents the maximum number of theatres screening on any day of that week for the film of interest. Also, the day-of-week dummy variables are now redundant. The dynamic stock model is similarly redefined in terms of weeks as TW, where TW now represents the number of weeks included prior to week tW. We consider the dynamic horizons TW=1,2,3,4 implying that downloads are observed for one, two, three and four weeks prior to, and including, week tW.

Results for the weekly model are reported in Table 6 which display second-stage regression results for the contemporaneous and dynamic stock models. In the contemporaneous and TW=1 model, a positive and significant relation between downloads is observed. However, in the TW=2,3,4 models the relationship is negative implying a statistically significant sales displacement relationship. The magnitude of the estimated coefficients suggests a displacement elasticity of 0.15% for a 1% increase in downloads over the four weeks prior to, and also including, the week of observation.

6.4 Forward Looking Dynamic Stock of Downloads

It has been argued that the dynamic stock approach to considering future sales displacement is more realistic than a contemporaneous approach. It might also be possible, however, that an individual forgoes cinema attendance today by making a decision to download the title at some point in the future. Obviously, this argument relies on the title being available for download and that the intention is actually carried out. To test whether downloaders forgo theatrical consumption prior to actually downloading the file, we consider a simple modification of our previously specified dynamic model [3] in which T is now forward looking. We denote this forward looking dynamic stock as TF and define DisTF=∑r=0TDis(t+r) for the horizons TF=7,14,21,28. Similarly our instrument is also redefined as Dis′TF=∑r=0TDis′(t+r) for all s′≠s.

As Table 7 demonstrates, the estimated coefficients on the (log) downloads variable across all specifications (TF=1,7,14,21,28) remain positive implying no displacement effect. Taken with the evidence of the (backward looking) dynamic stock model, this might suggest individuals who partake in downloading (and substitute it for paid cinematic consumption) are time impatient and seek out films on torrent sites early in their theatrical life – often prior to the official release if the film has already been released overseas as discussed above. ^[38] This observation is also consistent with the “movie-maven” subculture among (particularly young and tech savvy) consumers who desire to see a film early and before the masses.

6.5 Sales Displacement Effects

Although we have detected a statistically significant negative impact from file sharing on box office sales, the economic significance of these effects appears relatively small. We demonstrate this by providing some “back-of-the-envelope” numbers concerning the potential sales displacement effects of piracy implied by our results. Table 8 provides estimates of sales displacement or “substitution rates” for the daily, opening week and weekly models as presented in Tables 4–6. ^[39]

If we focus on the daily T= 7,14,21,28 models, we observe sales displacement elasticities of 0.06, 0.16, 0.25 and 0.31, respectively. Given the median daily box office of films in our sample is A$3,593 (Table 2), this translates to a reduction in revenue of between A$2.06 and A$11.01, or between 0.16 and 0.86 people assuming an average ticket price of A$12.87 (Screen Australia reported average ticket price for 2011), for a 1% increase in downloading activity across these time horizons. Given that state/territory-level median number of downloads over one, two, three and four weeks are 680, 1345, 1962 and 2485 (adjusting for Peer Media Technology’s estimate of a 55% market coverage), this suggests that somewhere between 29 and 42 downloads displaces one purchased ticket each day depending upon which model is considered. Put another way, for every 100 downloads somewhere between 2.4 and 3.4 cinema admissions are displaced.

When considering the opening week and weekly models, we find lower levels of displacement for the weekly model but also find substantially higher levels of sales displacement for opening week revenues. Given the range of elasticities from Table 5, and the median opening week daily revenue of A$31,046, we estimate a 1% increase in downloads displaces between 1.65 and 2.01 paid admissions. With the range of the median number of downloads between 371 (T=7) and 725 (T=28), these levels imply that anywhere from 2.2 to 3.8 downloads displaces one paid admission – which appears relatively high – but in terms of economic significance, the overall potential effect is low because of the relatively low levels of downloading actually taking place in the first week. In part this is reflective of the large number of films with simultaneous releases in the United States and Australian market, but it is apparent that the longer a film’s release is delayed between the two markets, the more likely the title is to appear on torrent sites which in turn increases the number of illegal downloads.

In terms of the weekly model’s results of Table 6, and given the median weekly revenue of A$19,142, the estimated displacement for T=2, 3 and 4 (weeks) suggest between 6.0 and 8.3 downloads displaces one paid admission over a weekly time horizon. This finding is lower than the daily model where it took between 29 and 40 downloads over the various windows considered to displace a sale on a given day. In the weekly model, we are observing approximately the same time frame for downloads but now a week for revenue/admissions. However, the calculation is not simply divided by seven as the medians are not related by this scale. As it turns out, the displacement effect is estimated larger in the daily model if the substitution rate is simply multiplied by seven.

To put these “back-of-the-envelope” calculations in perspective, assume that (as some industry reports implicitly suggest) one download displaces one paid admission. Then the total lost revenue of the median film is in the order of A$2.34m (adjusting the observed median downloads of 100,000 – see Table 1 – for 55% estimated market coverage, and assuming ticket price of $12.87), or about 38%. If we compare to the maximum (daily) substitution rate of 3.5, and continuing to assume median downloads of the same number of median downloads and ticket price, this would imply total lost revenues of A$82k (1.3%) for the median film.