While our results demonstrate that the 2014 website blocks caused an increase in visits to
legal subscription sites, one might ask whether all ofthese additional visits came from users who
were already subscribed or if the blocks caused some non-subscribers to begin paying for legal
subscriptions. We do not have e-commerce data on actual signups or subscriptions, but we can
assume that an individual who made no visits to legal subscription sites in the 3 months before
the blocks was not a paying subscriber. We first limit our sample to only individuals who made
no visits to subscription sites in the pre-period. We then define a binary variable
NewSubscriber, equal to 1 if the user made a number of visits to legal subscription sites above
some threshold in the post period, and equal to zero otherwise. Using this variable we estimate

the following cross-sectional model on the post period:
NewSubscriber; = ßg + BıTreatmentlIntensity; + e; (3)

Equation (3) measures whether treatment intensity — pre-blocked usage of subsequently blocked
sites — is associated with a higher likelihood of becoming a new subscriber in the post period.
We estimate this model via a logistic regression, the results of which are shown in table 5. Mak-
ing a single visit to subscription sites in the post period may indicate exploration of the site with-
out actually signing up, which is why we vary the threshold number of visits necessary in the

post period required to indicate becoming a new subscriber (columns 1 through 3).

29

Table 6: Impact of 53 Site Block in November 2014
on New Legal Subscriptions Estimated Via Logistic Regression

 

 

a) 2) 3)
Dependent variable: >1 post-period legal >2 post-period legal >3 post-period legal
subscription visits __ subscription visits subscription visits

Treatment intensity 0.00909** 0,0111** 0.0102*
(0.00341) (0.00385) (0.00444)

Constant -1.096*** -1.918*** -2.378+*+*
(0.0313) (0.0404) (0.0484)

N 5759 5759 5759

Log-Likelihood -3262.513 -2233.499 -1697.341

 

Notes: Standard errors are shown in parentheses. + p<0.10 * p<0.05 ** n<0.01 *#* n<0,001. Model is estimated
on a subset ofusers who made zero pre-period visits to paid subscription sites.

The coefficient on treatment intensity in each column indicates a positive and statistically
significant relationship between pre-period usage of blocked sites and post-period likelihood of
becoming a new subscriber. For each additional visit to blocked sites in the pre-period, an indi-
vidual’s probability of becoming a new subscriber in the post period increases by about 1% over
the baseline probability. The similarity ofthese coefficient across all three columns indicates that
the threshold number of visits required to indicate an actual subscription does not materially im-
pact our estimates. We acknowledge that without a fixed effects model, one might suggest that
individuals with higher treatment intensity are more likely to use paid subscription sites as well.
We have partly controlled for this problem by only looking at individuals who, in spite of their
varying levels of treatment intensity, were non-subscribers during the pre-period. Thus we are
only looking at individuals with similarly low propensity to subscribe. We have shown that us-
ers more affected by the blocks were more likely to become new paid subscribers in the post pe-
riod than users less affected by the blocks, which suggests a causal interpretation. In Table A5 in
Appendix D, we also estimate this model for the balanced panel and find coefficients as high as

0.018. We discuss the economie significance of this in section 6.2.

30

We next ask whether our findings change when fewer sites are blocked. To do so, we exam-

ine the blocking of 19 sites in November 2013 and also a single site in May 2012.

3.2 November 2013 Blocking of 19 Major Piracy Sites

Recall that because of privacy concerns PanelTrack would only release monthly data ag-
gregated into consumer groups for 2012 and 2013. While using aggregate grouped data is clearly
inferior to using individual data, our analysis is able to recover the impact of website blocks on
legal and illegal media usage with the appropriate inferential statistics. We rely on the estimator
and inference suggested by Donald and Lang (2007), who discuss methods for correcting for
common group errors when the treatment is assigned at the group level such as in our data. The
estimator they recommend as most efficient in many circumstances is the “between-group esti-
mator”, which is in fact a regression of grouped means against group average outcomes. The ap-
proach relies on the fact that each data point is based on the aggregated behavior of a sufficiently
large underlying group and thus is measured with greater precision than if each observation were
generated by one individual, which is precisely the data for the website blocks in 2012 and 2013
we have at our disposal. Inference for our estimators is given by a t,._, distribution, where G in-

dicates the number of groups.?!

Specifically, we estimate the following model:
InVisits;. = yo + yımonth, + y3TreatmentIntensity; - month, + u it St (4)

where all terms are the same as in (1) except that the j subscript now denotes a group of consum-

ers (as opposed to i indexing the individual). Because our data are aggregated across groups, the

 

* Note this distribution is more conservative than using a t-distribution with Tu +; +. +6-2 degrees of freedom that
would be used were we to estimate pooled OLS. This point is made in lecture notes by Jeffrey Wooldridge (2007).

31

resulting visits are large enough that we can estimate OLS. However, we log these data because
visits are right skewed and because we expect trends across the groups to be comparable on a rel-
ative (percent) basis. y, is the coefficient of interest, and as a test ofthe parallel trends assump-
tion, we ask if yz is 0 for all months before the blocks. We plot all y, coefficients below for the

various outcome variables.

Figure 2: Effect of November 2013 Blocks on Outcomes

Unblocked torrent sites visits Cyberlocker sites visits

Partiat 19 Piracy Partial 19 Piracy
Treatment Sites Blocked | Treatment Sites Blocked
I

.01

02

. ‚025

0
|
+
|
—
|
|
!
u

.015

-01
o

005  .01
—

Coefficient of Treatment Intensity
Coefficient of Treatment Intensity

0

’

|

|

|

|

|

|

|

|

|

|

F

|

t .
-.03 -.02

 

 

r I. " T ——_ m
Aug. '13 Sep. '13 Oct. '13 Nov. '13 Dee. '13 Jan. '14 Feb. 14 Aug. '13 Sep. '13 Oct. "13 Nov. '13 Dec. '13 Jan. "14 Feb. 14
VPN sites visits Legal subscription sites visits
- Partial 19 Piracy 8 Partiat 19 Pıracy
E Treatment Sites Blorked 2 Treatment |Sites Blocked
2
c c
£ | 28
E ; ES
3° | 004 5
Fe F
5 ss? “-—— — —t 7 na ee
E ® =
3 35
E ® | 5°
9o4+—-—-—- —- - -— IN — per +8
oO On
— ; n——-—- = : Mm
Aug. '13 Sep. '13 Oct. '13 Nov. '13 Dec. '13 Jan, '14 Feb. "14 Aug. 13 Sep. 13 Oct. '13 Nov. '13 Dec. "13 Jan. '14 Feb. '14

In the post period, it appears as if visits to unblocked cyberlocker piracy sites decreased
as aresult ofthe blocks while visits to legal subscription sites increased — both ofthese results
are consistent with our results from 2014. It also appears as if VPN visits and visits to unblocked

torrent sites increased as a result ofthe November 2013 blocks.

32

While the parallel trends assumption holds for cyberlocker visits, for legal subscription,
and (almost) for VPN site visits, it fails for unblocked torrent site visits. Heavier users ofthe
blocked sites appeared to increase their usage of unblocked torrent sites more in the pre-period
than lighter users. We are not certain why this is the case. It is possible that because the court
case which ordered the blocks occurred during October 2013, some users of pirate sites may
have had advance knowledge of the blocks and started to rely more other sites. Alternately,
some of the unblocked torrent sites may in fact be proxy or mirror sites for the blocked sites. Ei-
ther way, any results for the effect of the 2013 wave of blocks for visits to unblocked torrent sites
must be taken with caution as they may not be causal, but we can say that we find no evidence of

a decrease in usage of unblocked torrent sites caused by the 2013 blocks.

To measure the overall effect of the November 2013 blocks on the outcome variables and

to determine statistical significance, we estimate the following model:

InVisits;, = yo + yıPost, + y2TreatmentIntensity; - Post, + y3PartialTreatment,

(5)

+ y4Partial, - TreatmentiIntensity; + u it St

Model (5) is similar to (2) except that j indexes each group and the outcome variable is logged

visits due to our ability to estimate using OLS.

33

Table 7: Estimated Impact of 19 Site Block in November 2013 on User Site Visits

 

 

 

 

 

(1) (2) 3) (4)
Dependent variable: Unblocked tor- Cyberlockers VPN sites Legal subscrip-
. . rent sites tion sites
Post Treatment -0.149 -0.527** -0.0540 -0.0339
(0.0934) (0.121) (0.418) (0.0974)
Post X treatment intensity 0.00689+ -0.0141** 0.0454 ** 0.0134 **
(0.00324) (0.00305) (0.0140) (0.00359)
Partial Treatment 0.0295 -0.214 0.456+ -0.00699
(0.0697) (0.122) (0.225) (0.0691)
Aa X treatment inten- 0.00881** -0.00339 0.0151 -0.00843*
(0.00186) (0.00343) (0.0109) (0.00364)
Constant 13.74 +** 13.55 *+** 9,92 7r + * 13.78* ++
(0.0318) (0.0472) (0.147) (0.0380)
User Group FE? Y Y Y y
N 70 70 70 70
User Groups 10 10 10 10
Adjusted R2 .125 . .633 221 255

Notes: Standard errors are shown in parentheses and elustered by user group. + p<0.10 * p<0.05 ** p<0,01 *** p<0.001
In Table 7, we see from the Post Treatment dummy that all ofthe outcome variables were de-
creasing over time, though the decreases are relatively small for VPN sites and legal subscription
sites. The coefficients of interest are those on the post x treatment intensity interaction term.
Here, we see an increase usage of unblocked torrent sites (significant at alpha = 0.1). Because we
know from the time plot in Figure 2 that this may be an extension of pre-existing trends we can-
not make a strong claim as to whether these blocks increased usage of unblocked torrent sites,
but there is no evidence of a decrease. We do observe a causal decrease in visits to unblocked
cyberlocker sites, a causal increase in visits to VPN sites, and a causal increase in visits to legal
subscription sites. Thus, our results indicate that, like the blocking of 53 sites in November
2014, the blocking of 19 sites did drive some users to paid legal streaming sites and reduced at

least some forms of piracy (cyberlockers). An individual who made one more visit per month to

34

blocked sites during the pre-period increased her monthly visits to legal subscription sites 1.34%

more than she would have if not for the blocks.

We followed Donald and Lang (2007) in computing p-values when outcome variables are
aggregate data from large groups, and believe this to sufficiently correct for any downward bias
in standard errors. However, because our number of clusters is small, we also impute even more
conservative p-values using the wild cluster bootstrap approach (Cameron et. al. 2008). These p-
values on the coefficients of interest can be found in Table A6 in Appendix D (there the coeffi-

cient for visits to paid legal subscription has a p-value of 0.08).

Both the 2013 and 2014 waves of blocks involved the blocking of a number of major piracy
sites. Next, we ask whether the blocking of one major piracy site — an experiment more akin to
those in Poort et. al. (2014) and Aguiar et. al. (2018) — demonstrates similar outcomes or pro-

duces a different set ofresults.

3.3 May 2012 Blocking of The Pirate Bay

The data we obtained from PanelTrack to study the blocking of The Pirate Bay in 2012 are
similar to the data from 2013: we observe outcomes by consumer group by month, generated from
balanced panel of consumers observed in all months. Again, PanelTrack sorted consumers into

groups based on pre-block usage of the blocked site, in this case, The Pirate Bay.

We estimate model (4) for each of the outcome variables and plot the coefficients of interest

for the models in Figure 3.

35

Figure 3: Effect of May 2012 Pirate Bay Block on Outcomes

Unblocked torrent sites visits

4 Partial The Pirate Bay
Treaiment Blocked

‚004

.002

Coefficient of Treatment Intensity
0
7
+
|
1

-002

= A VE VEEHESEEEEE VESHEEREEHEEENE HERSRERREEEEE: KERSESHEEHEEE RER
Feb. '12 Mar. 12 Apr. 12 May "12 Jun. 12 Jul. "12 Aug. '12
VPN sites visits

Partial The Pirate Bay
Treatment Blocked

‚03

.02

0
*
I
|
*

Cosfficient of Treatment Intensity
01

 

-01

I.
Feb. '12 Mar. 12 Apr. "12 May '12 Jun. 12 Jul. 12 Aug. 12

36

‚006

.004

Coefficient of Treatment intensity
-.002

-.004

01

‚005

Coefficient of Treatment Intensity

-.005

.002

0

0

Cyberlocker sites visits

Partiat

The Pirate Bay
Treatment Iocked

1 en

Feb. '12 Mar. '12 Apr. '12 May 12 Jun. 12 Jul. "12 Aug. '12

Legal subscription sites visits

Partial The Pirate Bay
Treatment Blocked

mu

Feb. '12 Mar. ‘12 Apr. '12 May ''12 Jun. "12 Jul. "12 Aug. '12

 

i

The 2012 blocking ofthe Pirate Bay appears to have caused an increase in visits to un-
blocked torrent sites as well as visits to VPN sites. We observe no clear effect on visits to cyber-
lockers and while we may see an increase in usage of legal subscription sites in the month after
the blocks, it disappears by the second and third months after the blocks. The parallel trends as-
sumption appears to hold for visits to unblocked torrent sites, VPN sites, and (nearly) for un-
blocked cyberlockers. However, the parallel trends assumption fails for visits to legal subscrip-
tion sites as treatment intensity appears positively correlated with changes in visits to subscrip-
tion sites during the pre-period. There is a compelling explanation for this fact: one ofthe paid

subscription sites, Netflix, was introduced to the UK in January 2012, and it quickly became

popular due to the fame ofthe brand. During the initial adoption period, we argue that people

who were pirating a lot of content (relative to people who were pirating little) are more likely to

have an initial interest in Netflix and therefore subscribe. This would explain the elevated y, co-

efficients in March and April. The direction of the pre-existing trend would actually suggest that

the correlation should have increased in the post period, and instead we see it remain flat or de-

crease other than in June 2012. Thus we conclude that blocking The Pirate Bay caused no last-

ing increase in paid legal consumption and at most a temporary one month increase.

Next, we estimate (5) for each ofthe outcome variables and present the results below. ?

Table 8: Estimated Impact of The Pirate Bay Block in May 2012 on User Site Visits

 

 

 

 

) 2 3) a
Dependent variable: Unblockea Br Cyberlockers VPN sites Legal subserip-
In rent sites . tion sites
Post -0.207** -0.388+ -0.962+ -0.576+
(0.0434) (0.173) (0.508) (0.292)
Post X treatment intensity 0.0022 1*** 0.000769 0.0106** 0.00143
(0.000363) (0.000935) (0.00261) (0.00148)
Partial treatment -0.312** -0.340+ -1.085* -0.707+
(0.0811) (0.160) (0.392) (0.322)
Partial X treatment inten- 0.000720 -0.00159+ 0.0108** 0.00119
sity (0.000436) (0.000835) (0.00233) (0.00166)
Constant 14.67*** 13.85*** 9.897 ** 11.90***
(0.0216) (0.0785) (0.217) (0.141)
User Group FE? Y Non Y Yo
N 70 70 70 70
User Groups 10 10 10 10
Adjusted R2 0.274 0.288 0.050 0.203

 

 

Notes: Standard errors are shown in parentheses and clustered by user group. + p<0.10 * p<0.05 ** p<0.01 *** p<0.001

?2 As with the 2013 blocks, we present in Table A7 in Appendix D the same estimates but with standard errors esti-
mated using the wild cluster bootstrap approach. The increase in visits to unblocked torrent sites remains significant

with a p-value of 0.036.

In Table 8 we observe a statistically significant increase in usage of other unblocked torrent sites
such that a person visiting The Pirate Bay one more time during the pre-period increased her us-
age of other torrent sites 0.22% more than she would have after the blocks if she had not been
using The Pirate Bay. We observe a statistically significant increase in usage of VPN sites, indi-
cating that, after the block, some heavy users ofthe Pirate Bay turned to using a VPN to circum-
vent the blocking of the site. However, the economic significance of this may be small as the
constant term here is small — visits to VPN sites in the data are relatively low. Finally, the coef-
ficient for paid legal streaming sites is positive but small and statistically insignificant. Typi-
cally, this might lead to an inconclusive interpretation — is the increase positive or 0? From Fig-
ure 3 we know that any increase in the post period is driven entirely by the first month, after
which it disappears. And we also know that even this first month effect may be the result ofa
pre-existing trend. Thus, like Aguiar et. al. (2018) we find no lasting causal effect ofthe May

2012 blocking of The Pirate Bay on legitimate consumption.

3.4 Summary of Empirical Results

In summary, we found that the 2014 blocking of 53 major piracy sites not only decreased
visits to the blocked sites but also caused a decrease in usage of other unblocked piracy sites. We
observe that it causally increased usage of paid legal streaming sites and may have been associated
with an increase in new paid subscriptions. Together, these results imply that supply-side antipi-
racy enforcement can be effective in turning users of illegal piracy channels toward paid legal
consumption. In November 2013 when 19 major piracy sites were blocked, we do not observe a
causal decrease on visits to unblocked torrent sites but we do observe a causal decrease in visits to
unblocked cyberlocker sites. We also observe a statistically significant increase in usage of paid

legal streaming sites. Finally, consistent with the literature, we found that the May 2012 blocking

38