The Impact of Time-Shift TV on TV Viewership and on Ad Consumption: Results from Both Natural and Randomized Experiments

Transcription

1 The Impact of Time-Shift TV on TV Viewership and on Ad Consumption: Results from Both Natural and Randomized Experiments Rodrigo Belo 4,a,PedroFerreira 1,2,b,MiguelGodinhodeMatos 3,c, and Filipa Reis 1,3,d 1 Heinz College, Carnegie Mellon University 2 Department of Engineering and Public Policy, Carnegie Mellon University 3 Católica-Lisbon, Portuguese Catholic University 4 Rotterdam School of Management, Erasmus University a rbelo@rsm.nl b pedrof@cmu.edu c miguel.godinhomatos@ucp.pt d frodrigu@cmu.edu August 10, 2016 Abstract We partner with a major telecommunications provider to study the e ect of Time- Shift TV (TSTV) on TV consumption. TSTV automatically records in the cloud programs that were broadcasted live for the past few days and thus allows users to watch the programs they want when they want. In 2012, our industrial partner deployed TSTV to half the TV channels it o ered to premium consumers. Using di erence-in-di erences with Inverse Probability of Treatment Weighting we find that, on average, the introduction of TSTV increased the consumption of daily TV by 11 minutes (p <0.01), from a baseline of 3.4 hours, and did not change the consumption of live TV. Furthermore, we find that the concentration of TV consumption increased after TSTV was available. In 2015, our industrial partner run a randomized experiment in which a random set of households was selected to obtain access to a new set of TV channels broadcasting movies and TV shows. A random subset of these households obtained access to these channels with TSTV and another random subset of them obtained access to these This work has been partially supported by PhD grant SFRH/BD/51568/2011, by Post-Doctoral grant SFRH/BPD/94212/2013 and by the ilab at the Heinz College. We also thank our Industrial Partner for their support. 1

2 channels without TSTV. Using di erence-in-di erences we find that, on average, the former set of households consumed 4.5 more minutes of TV per day (p <0.01), froma baseline of 5.0 hours, and as much live TV as the latter set of households. In addition, the consumption of TV by the former set of households was more concentrated towards the most popular programs. Finally, we show that households do not seem to use TSTV as a new tool to strategically avoid ads. In particular, and in 2015, households given access to the new TV channels with TSTV exit ads in the original TV channels as much as the households given access to these channels without TSTV. Therefore, the concern of advertisers that TSTV may reduce their revenues is unwarranted. However, advertisers do not have data on TV audiences. Instead, content distributors do. Therefore, the latter are in a unique powerful position to deploy auction-based systems to sell TV ad slots taking advantage of the fine-grained information they have on TV viewership, much like websites do today online using cookies. 1 Introduction Watching TV is still the leisure activity that most people devote most time to across the developed world. However, the traditional model of linear TV imposes several restrictions on users. Users interested in specific programs need to accommodate their schedules to the existing programming grids, possibly compromising on convenience. Users with fixed time slots for leisure, and in particular to watch TV, are restricted to watch the content broadcasted during these slots, possibly compromising on the programs they watch. In recent years, technology evolved to address these limitations. Video Cassette Recorders (VCRs), Digital Video Disc (DVD) recorders, Digital Video Recording (DVR) and more recently Time- Shift TV (TSTV) allow users to watch the programs they want after they are originally aired live on TV. TSTV is the most advanced technology in this respect. It automatically records in the cloud programs that were broadcast live for the past few days. These programs can then be easily fetched by users who can watch them at their leisure. This is in contrast with DVR, which requires users to decide ahead of time which programs they would like to watch and to explicitly program their device to record them. Furthermore, DVR does not allow for recording concurrent programs whereas with TSTV all programs that aired in the past few days in the channels covered by the service are available to users. Abreu et al. (2016) s global survey on the availability of non-linear TV 2

3 services shows that TSTV is currently available in most countries in North America, South America, Europe and Asia. In most implementations TSTV allows users to rewind 7 days or less. Nielsen (2015) s total audience report shows that in 2015, 8% of all TV consumption in the US was time-shifted. This statistic is similar in other countries, such as the United Kingdom (BARB, 2011) and France (isuppli Screen Digest, 2011). The introduction of TSTV may trigger several concurrent changes to the way people watch TV. It may impact the amount of TV consumed, the type of programs watched and the time at which people watch TV. Understanding how all these changes might play out is fundamental for content providers, content distributors and advertisers because their interplay may a ect the equilibrium of this supply chain in complex ways. For example, TSTV is likely to attract specific people such as those that already know what they want to watch and thus use TSTV to immediately get to that content. This shifts audiences across channels, which advertisers need to take into account when planning their marketing campaigns. Furthermore, people that use TSTV may do so at the expense of live TV. This may reduce advertising revenues if users skip ads when using TSTV. Reduced advertising revenues may, in turn, hurt the profitability of content distributors who buy content from content providers to attract consumers and sell ad slots to advertisers. Therefore, reduced revenues in the industry may also a ect the production of new content and reduce consumer surplus. The prior literature o ers mixed evidence with respect to ad-skipping (Wilbur, 2008a; Anderson and Gans, 2008; Zigmond et al., 2009). Yet, and given the ubiquity of TV consumption all over the developed world, even small changes in ad consumption at the household level may amount to large sums of money at the industry level. For example, in 2015 in the US, TV ads captured nearly 38% of total ad spending, amounting to billion USD (emarketer, 2016), roughly 0.38% of GDP. In the same year, TV subscriptions reached billion USD, roughly 0.56% of GDP (Statista, 2016b). In this paper, we partner with a major telecommunications provider, here- 3

4 inafter called TELCO, to explore outcomes from both a natural experiment and a randomized experiment aimed at studying the e ect of TSTV on TV consumption. In the end of the summer of 2012, TELCO introduced TSTV to premium consumers on 50 out of the 100 TV channels o ered as part of its basic TV package. These 50 channels correspond to the most popular channels, gathering 90% of total viewership before the introduction of TSTV. The consumption of TSTV increased quickly but also flattened quickly at around 8% of the total TV consumption. Using di erence-in-di erences with inverse probability of treatment weighting we find that after the introduction of TSTV, households watched 11 more minutes of TV per day (p <0.01) from a baseline of 3.4 hours. We also find that they did not change the amount of live TV they watched. The TV consumed in time-shift comes on the top of the prior live TV. This finding provides some evidence that the introduction of TSTV may be, after all, ino ensive to advertisers. However, we also find that the distribution of TV viewership across programs became more concentrated after the introduction of TSTV, which may increase the value of the ad slots associated to the most popular content. This needs to be taken into account by advertisers if TSTV providers move to o ering TSTV without ad-skipping. Increased concentration in the consumption of TV content with TSTV may also a ect the production of content. When the most popular content becomes even more popular it might become risker for small producers to develop new content. In 2015, TELCO ran a randomized experiment in which a random set of households were given access to a new set of 10 TV channels, called the Entertainment Bundle, broadcasting only movies and TV shows. These TV channels were o ered in addition to the original 100 TV channels included in TELCO s basic TV package. Between May 18th and June 30th, a subset of these households were given access to these channels without TSTV and another subset of these households were given access to these channels with TSTV. A third random subset of households were held out from any intervention. This experimental design readily provides for identification. In line with the results obtained using the 4

5 2012 data, we find that households that were given access to the Entertainment Bundle with TSTV watched 4.5 more minutes of TV per day (p <0.01) than households that were given access to this bundle without TSTV, from a baseline of 5.0 hours. Essentially, households substitute viewership time in the general purpose free to air channels original channels for the Entertainment Bundle and more so when the latter is o ered with TSTV. We also find, again, that households did not change the amount of time they devote to watch live TV and that the concentration of TV viewership across the programs o ered as part of the Entertainment Bundle is higher for households that obtained access to this bundle with TSTV. Therefore, we confirm the results obtained with the data from Learning that the introduction of TSTV does not reduce the amount of live TV that households consume may put advertisers at rest. However, we also find that the households that use TSTV are di erent from the ones that do not and advertisers should take these di erences into account in order to plan their ad campaigns ahead of time. Finally, we study how the introduction of TSTV in 2015 a ected the consumption of ads placed by TELCO (for its own products such as the movies available in it s Video on Demand service) between June 1st and June 30th, The channels o ered as part of the Entertainment Bundle did not include ads and therefore our setting resembles a worst case scenario in which all viewership of the Entertainment Bundle is associated with a loss in ad consumption. We find that households that were given access to the Entertainment Bundle with and without TSTV watched less of these ads during the month of June 2015 in proportion to how they substitute viewership time in the original TV channels for viewership time in the Entertainment Bundle. Hence, we do not find that households use TSTV strategically to consume less ads. We confirm this result by showing that the probability of exiting an ad in the original TV channels is the same for these two sets of households and therefore having TSTV does not seem to trigger strategic behavior from consumers with respect to ad avoidance. Our findings have deep implications for content providers, content distributors 5

6 and advertisers. They suggest that these stakeholders should not be concerned with TSTV stealing from live consumption. Instead, they should realize that TSTV changes the concentration of TV viewership, which is likely to increase the value of the ad slots associated to the most popular programs. In parallel, a content industry more concentrated towards these programs may a ect the production of content. In particular, the most popular content providers might charge higher prices to content distributors and, at the same time, it might become riskier for small content providers to develop new content. In addition, TSTV attracts a certain type of households and this information should be taken into account at the time of campaign planning. Finally, we note that advertisers do not have data on TV consumption, live nor time-shifted, but content distributers do. Therefore, the latter are in a unique position to develop business models for advertising akin to those used nowadays on the Internet, in which their fine-grained information about TV audiences might play the role that cookies do in the online world. The remainder of this paper is organized as follows. In section 2 we revise the relevant work and, in section 3, we present the empirical context of this study. In section 4 we present our analysis of the natural experiment of the initial introduction of TSTV by our industry partner. In section 5 we present our analysis of the randomized experiment in which TV channels with and without TSTV were o ered to households. In section 6 we analyze the impact of TSTV on advertising using the data from the randomized experiment. Finally, in section 7, we conclude. 2 Related Work In 2014, American adults spent over 30 hours per week on leisure activities (BLS, 2015). The most popular leisure activity is watching TV followed by socializing and communicating (BLS, 2015). In the first quarter of 2014, on average, they spent 295 minutes per day watching TV plus an additional 35 minutes per day 6

7 using Time-Shift TV (TSTV) (Nielsen, 2015). Roughly half of millennials (18-34 years old) use TSTV (Statista, 2016a). They report that they do so because with TSTV they can watch TV when is convenient, catch missed episodes and pause and rewind, which leads to a more e cient use of the time they devote to watch TV (Statista, 2016a). In short, the introduction of TSTV improves the quality of TV consumption. In particular, TSTV increases both the volume and the variety of content available to users. As a result, TSTV may increase TV consumption at the expense of other leisure activities. In fact, according to Hinz and Eckert (2010), the market for each particular leisure activity may be always unsaturated because leisure activities are highly substitutable for each other. Classical economic theory predicts that consumers are better o with more options (Hotelling, 1929; Dixit and Stiglitz, 1977; Salop, 1979). When choice sets expand consumers can either choose a new option that provides them with more utility or stick to their previous choice obtaining the same utility. Brynjolfsson et al. (2003) showed empirical evidence that consumer welfare increases with more options in online markets because consumers can find options that better match their preferences. However, and also empirically, Boatwright and Nunes (2001) showed that consumers are sometimes unable to internalize the benefits from increased product variety. Consumers may end up buying less if they have to scan too much to find appropriate products. In recent times, recommender systems have been introduced to help consumers navigate the large catalogs offered to them online. At the core, recommender systems distort search costs. They lower the search costs of some products at the expense of others (Hinz and Eckert, 2010). This a ects which products get sold and the resulting market concentration (Hinz and Eckert, 2010). Two e ects may arise with respect to changes in concentration. Increased volume and variety may help users discover products that otherwise they were unable to find. This is likely to reduce market concentration. For instance, Brynjolfsson et al. (2011a) compare the distribution of sales of online and o ine channels using data from a multichannel retailer. They find that even when the products o ered in the two channels are exactly 7

8 the same, the introduction of search engines and recommendation tools in the online channel increases the share niche products. Alternatively, a super-star e ect may arise, whereby users buy even more of the most popular products shifting even more density towards the head of the distribution of sales across programs. This has been observed by Elberse and Oberholzer-Gee (2006) in a study of the distribution of home video products between 2000 and The authors found evidence of both a long-tail e ect - a significant increase in the number of titles selling only a few copies - and of a super -star e ect - among top performers, most sales became more concentrated in fewer titles (Elberse and Oberholzer-Gee, 2006). The introduction of TSTV may change the amount of time devoted to watch TV, both live and total, the types of programs that people watch and the time at which people watch TV. The interplay among these changes may fundamentally change the entertainment industry by readjusting the relative valuation of programs, channels and ultimately advertisement slots, which may require both content providers and content distributors to adapt their business models. These changes stem from two fundamentally di erent responses that may arise from consumers: i) content distributors already optimize programming grids, that is, the higher quality programs air at times when most people are at home and can watch them. In this case, if consumers are relatively homogeneous regarding their fit with programming grids, most people will keep watching live TV at the same time as before, potentially to watch the same programs they used to watch before TSTV became available, and then use TSTV to catch up some programs that might have aired in conflict with their favorite programs or that they might have missed. This type of behavior should not change too much the time at which people watch live TV, and may increase total viewership time through TSTV; ii) users take advantage of the flexibility introduced by TSTV and watch TV at di erent times once TSTV is available. In this case, if users are relatively heterogeneous regarding the content they like and the time at which they prefer 8

9 to watch TV, with TSTV they no longer compromise on one or the other because they can watch the programs that best match their preferences at any time. In this case, most likely most of the TV they used to watch live becomes timeshifted, which may, or may not, change the total amount of time they devote to watch TV. At the outset, it is hard to predict which of the above two scenarios is more prevalent. The heterogeneity across users with respect to their fit with programming grids and to their flexibility to change viewership schedules is likely to determine what happens on average. People may also behave di erently at weekdays and weekends because most people have a more flexible schedule during the latter. Likewise, it is also hard to predict how the introduction of TSTV will a ect the concentration of TV viewership. Some users will take advantage of TSTV to watch more popular programs, in particular, programs that they could not watch before TSTV was available because they aired at the same time in di erent channels. This behavior will shift even more weight towards the head of the distribution of viewership across programs. Other users will use TSTV to find niche programs that best match their preferences flattening the distribution of viewership across programs. Understanding the changes introduced by TSTV raises a number of empirical questions, which we aim at addressing in this paper, such as: how does the introduction of TSTV change live and total viewership time? How does the introduction of TSTV change the types of programs that people watch? How does the introduction of TSTV change the times at which people watch TV? Another stream of research related to our paper is focused on ad skipping behavior. Recording technologies, such as TSTV, allow users to fast forward and thus skip ads. Therefore, advertisers worry that the introduction of TSTV will lead users to strategically substitute live TV for time-shifted TV, which is likely to reduce ad consumption and advertising revenues. However, some empirical studies provide mixed evidence in this respect. For example, Downey (2007) show that with DVR people watch only 59% of the ads that they do with live 9

10 TV. Pearson and Barwise (2007) studied the consumption of ads by 22 households whose TV usage was filmed for a period of 2-3 weeks. The authors found that these households fast-forwarded ads two thirds of the time when watching TSTV. Contrary evidence was provided by Bronnenberg et al. (2010) who analyzed data from a three year field study carried in partnership with five firms in which a sample of fourteen thousand households were solicited to accept TiVo for free. Using propensity score matching (because households included in this field study were not selected at random), the authors found that skipping ads occurred relatively infrequently. Unfortunately, these studies are either observational or field experiments without randomized treatments, which hampers their ability to claim causal e ects. Furthermore, note that even in the absence of recording technologies such as DVR and TSTV, avoiding ads is a common behavior among viewers who can either change channels or start some other activity when commercial breaks come along (Van Meurs, 1998; Krugman et al., 1995; Tse and Lee, 2001). Ultimately, the entertainment industry is supported by advertising revenues. These go through content distributors, who operate a two-sided market: they buy content from producers to attract the consumers attention and sell advertising slots to advertisers (Wilbur, 2008b). The price of ad slots is a function of the characteristics of the audience that is likely to watch them, such as volume and age bracket (Wilbur, 2008b). In the case of TV, the ability to attract specific audiences that generate enough revenue through advertisements is a significant determinant of which content gets to be broadcasted and when (Wilbur, 2008b). The shift towards digital TV, allowing for better measurement of TV audiences can thus help all TV-related stakeholders face the increasingly fierce competition from web-based advertising. The latter has been stealing market in recent times due to the superior targeting e ciency provided by cookies (Evans, 2009). Along these lines, our paper aims at characterizing the consumption of ads as a result of the introduction of TSTV and at studying whether consumers use TSTV strategically to consumer fewer ads by, for example, substituting live TV for 10

11 time-shifted TV. 3 Empirical Context This work was developed in collaboration with a multinational telecommunications provider, hereinafter called TELCO. Our analysis will focus on one region where TELCO is the market leader in Pay-TV services with over one million households. The firm s services include Pay-TV, Video-on-Demand, Digital Video Recording, TSTV, broadband and mobile Internet, fixed and mobile telephony. Subscribers can opt for either the standard or premium service, which di er in the number of TV channels and on the set of complementary features available. Our study focuses on premium subscribers who, in 2012, accounted for roughly 16% of all the households served by TELCO. This statistic grew to around 80% in Premium subscribers have at least one Set-Top-Box (STB) with over one hundred TV channels, a high-speed Internet connection and unlimited fixed telephone. The TV channel o ering from our industrial partner s basic service bundle can be complemented with additional packs of thematic channels such as children, music, sports, documentaries, movies and TV shows, which can be purchased separately for a fixed monthly fee. The Entertainment Bundle is a set of 10 premium channels broadcasting recent and popular movies and TV series. In particular, the TV-series broadcasted in these channels air only a couple of days after their U.S. broadcast. Access to the Entertainment Bundle can be purchased for 13 USD/month. 11

12 4 The Introduction of Time-Shift TV 4.1 Natural Experiment In the end of the summer of 2012 TELCO introduced TSTV for free to premium consumers. This service extended the previous DVR capability and became available for 50 TV channels out of the 100 TV channels included in TELCO s basic TV package. These 50 channels corresponded to the most popular channels, covering 90% of the total TV viewership time. With TSTV, premium consumers were able to rewind one week the content o ered as part of these channels. This service increased both the quantity and the variety of the programs that premium consumers could watch at any point in time. Exploring the e ect of the introduction of TSTV in 2012 allows us to study how the sudden availability of this technology for a significant number of TV channels changes the patterns of TV consumption. This may help anticipate e ects in countries where TSTV is not yet available but might be in the near future. 4.2 Descriptive Statistics Our dataset covers the period between July 15th 2012 and October 15th 2012 and pertains to a random sample of 50,000 standard consumers and 10,000 premium consumers. Figure 1, shows the daily TV viewership for both sets of consumers during this period. On average, households in our sample watched 3.4 hours of TV per day before the introduction of TSTV. This statistic became 3.3 and 3.6 hours per day after the introduction of TSTV for standard and premium consumers, respectively. Figure 2, shows the breakdown of TV viewership between live and TSTV over time. The latter changes quickly after the introduction of TSTV and also flattens quickly at around 8%. Figure 3, shows the breakdown of TV viewership time, live and TSTV, per type of channel. General purpose free to air channels account for most of the viewership, both in live and in TSTV, followed by entertainment and children s. However, the share of viewership de- 12

13 voted to entertainment channels is substantially larger in TSTV whereas that devoted to news is substantially smaller. This could be expected given that news are more time sensitive. 6 Avg. Total Daily TV Viewership (Hours) 4 2 Premium Standard 0 Aug Sep Oct Date Figure 1: Daily TV viewership time (2012) Figure 4 shows the cumulative distribution of TSTV viewership by number of days elapsed since the program was first broadcasted live on TV. Panel (a) shows that on average 80% of the programs watched using TSTV aired in the previous 48 hours. This may be simply a reflection of the fact that older content has a higher search cost with TSTV because the interface on the TV screen requires more clicks to access older content. Panel (b) breaks down the former statistic per type of content. As one could expect, it shows that this statistic 13

14 Percentage of Time Jul 15 Aug 01 Aug 15 Sep 01 Sep 15 Oct 01 Oct 15 Figure 2: Percentage of viewership time in Live and TSTV (2012) (observational dataset, 2012) 1.00 adult % 43.2 % other music international Market share % 34 % lifestyle documentaries sport 11.2 % news % 12 % children entertainment general 0.00 Live TV Viewership Type TSTV Figure 3: Distribution of Live and TSTV viewership time by channel (2012) type 14

15 P( Viewership delay < x ) P( Viewership delay < x ) Viewership delay in days (view time in time shift program air time) Viewership delay in days (view time in time shift program air time) entertainment general news sport (a) Overall view lag (2012) (b) View lag by channel type (2012) Figure 4: TSTV view lag is higher for time-sensitive content, such as sports and news, and lower for entertainment and documentaries. Finally, figure 5 shows the breakdown of live and TSTV viewership per hour of the day at which programs were originally broadcasted. Primetime content, originally aired between 8pm and midnight, is the most watched in both consumption modes. This may simply arise because in order to optimize programming schedules content distributors place the most appealing content at a time when most users are at home. 4.3 Empirical Strategy Measuring Impact on TV Viewership Time We aggregate our data at the week level, which results in a panel with 6 weeks before the introduction of TSTV and 7 weeks after the introduction of TSTV. We use a di erence-in-di erences approach to study the impact of the introduction of TSTV on TV viewership time: Y it = After t + 2 T reated i + 3 T reated i After t + µ it (1) 15

16 2012 Distribution of view time per content types % 7% 37% 25% 20% 9% 5% 13% 38% 18% 15% 11% Morning Noon Afternoon Prime Time Night Dawn Live Viewership Type TSTV Figure 5: Distribution of Live and TSTV viewership time by time block (2012) where i represents a household and t aweek.after t is a dummy variable for the period after the introduction of TSTV, T reated i indicates whether household i was given TSTV (that is, whether it is premium) and µ it is the idiosyncratic error term. Y it indicates an outcome of interest, such as total, live or TSTV viewership time. The error term includes all covariates that might explain the outcome variable. Some of them may be correlated to whether a household is premium. For example, premium households might like to watch more TV. Thus, they register more TV viewership and, exactly because they like to watch more TV, they also have a higher propensity to subscribe the premium service. Such potential correlation between T reated i and µ it renders OLS estimates biased. We address the identification concern above using Inverse Probability of Treatment Weighting (IPTW). This approach can be used to explore the outcomes of natural experiments as a way to minimize selection bias (McCa rey et al., 2004). We weight the observations of standard households by their propensity of being treated. Weights were computed using boosted regression modeling (McCa rey et al., 2004; Guo and Fraser, 2014). This approach ensures that, on average, (weighted) control and treated households are similar on key observed covariates 16

17 that determine whether they are likely to be treated. We use the number of days that households watch TV per month, the number of programs they watch per month, the amount of time they spend watching TV, the hour of the day when they turn on the TV and their tenure with TELCO as proxies for their interest in TV and for their availability to watch TV, which should correlate to their propensity to subscribe the premium service. Table 1, shows how using IPTW improves balancing in our sample. In general, standard consumers watch less TV (hours per day, number of programs watched and number of days watching TV) and joined TELCO later. Therefore, balancing is accomplished by associating larger weights to standard households that watch more TV. Column (4) in Table 1, shows that after weighting the weighted standardized e ect size is smaller than 0.2 for all covariates considered in this analysis, which is the threshold typically considered when using IPTW to achieve good balance (McCa rey et al., 2004). Finally, we are interested in the magnitude of 3, whichmeasureshowthe introduction of TSTV changes the outcome of interest Measuring Impact on TV Viewership Concentration We follow the approach used in Brynjolfsson et al. (2011b) and use the Pareto distribution to study the e ect of the introduction of TSTV on the concentration of TV viewership. The cumulative density function for a random variable in [1, +1] withshapeparameter >0isgivenbyf(x) =1 x. For example, in Brynjolfsson et al. (2011b) s study of sales as a function of rank, the shape parameter represents the decay in sales over rank. A larger shape parameter will place more density towards the head of the distribution. In our setting, we pool together observations of live and total viewership time per program after the introduction of TSTV and we let the dummy variable T otal j indicate whether observation j pertains to live or total viewership. Therefore, each observation in this analysis is a program in a viewership mode. The model we estimate is: log(t ime j )= T otal j + 2 log(rank j )+ 3 log(rank j ) T otal j + j (2) 17

18 where T ime j represents the viewership registered in observation j, Rank j indicates the rank of observation j in the corresponding distribution of viewership and j is the idiosyncratic error term. We expect 2 to be negative, as viewership decreases with rank. We are interested in parameter 3 whose sign indicates how the concentration of TV viewership changes with the introduction of TSTV. Anegative 3 will provide evidence of increased concentration. In this case identifying the e ect of the introduction of TSTV on the concentration of TV viewership relies on assuming that the concentration of live TV did not change concurrently. Section provides some empirical evidence in favor of this assumption. 4.4 Results Impact of TSTV on TV Viewership Time Table 2 shows our di erence-in-di erences results for the time spent watching TV. Column (1) shows that the introduction of TSTV is associated to an increase of hours (11 min) (p <0.01) in TV viewership time per day, which corresponds to a 5.4% increase upon the average time devoted to watch TV before the introduction of TSTV (3.39 hours per day). Column (2) shows that the time spent watching live TV did not change with the introduction of TSTV. Column (3) shows that the consumption of TV using TSTV makes up for all the increase in TV consumption. Therefore, the concern of advertisers that the introduction of TSTV could reduce ad consumption and advertising revenues, by allowing users to skip ads, may be unwarranted Impact of TSTV on the Concentration of TV Viewership Table 3 shows the results obtained from fitting a Pareto distribution to viewership time per program as discussed in section 4.3. As expected, the coe cient on log(rank) isnegative(p < 0.01). The coe cient for the interaction between log(rank) andt otal is also negative (p < 0.01) suggesting that after 18

19 Table 1: Pre-treatment variables before and after weighting. Weights chosen to minimize the mean standardized e ect size (ATT) in the pre-treatment variables (2012) Premium Premium (s.d.) Standard Standard (s.d.) Std. E. Size Weight. Standard Weight. Standard (s.d.) W. Std. E. Size view time total hr sd view time total hr n days weekend n days weekday n programs weekend n programs weekday sd n programs mode first event start hour d view time total hr tv tenure tv tenure:<na> net tenure net tenure:<na> voice tenure voice tenure:<na> Note: Standardized e ect size of a variable is defined as the di erence between the mean of the variable in premium and standard consumers divided by the standard deviation of the variable over the premium consumers. 19

20 Table 2: Viewership time as a function of TSTV (2012) Dependent variable: Total TV Live TV Time-Shift TV (1) (2) (3) TSTV (0.028) (0.028) (0.000) Premium (0.024) (0.024) (0.001) TSTV*Premium (0.033) (0.033) (0.002) Constant (0.020) (0.020) (0.000) Observations 652, , ,799 Akaike Inf. Crit. 4,170, ,128, ,077, Note: p<0.1; p<0.05; p<0.01 Standard errors clustered at the user level. the introduction of TSTV the distribution of total TV viewership time is more concentrated than that of live TV viewership time. Therefore, the introduction of TSTV increased the concentration of TV viewership. This may have significant implications for advertisers because if content distributors move to o ering TSTV without ad-skipping capabilities then the advertising slots associated to the more popular content may become more expensive. This may also have significant implications for content providers and content distributors because the more popular programs may become more expensive when they become even more popular and, in addition, it may also become riskier for smaller producers to create new content. 5 Adding TV Channels and Time-Shift TV 5.1 Randomized Experiment TELCO ran a randomized experiment in May and June 2015 to study the effects of providing households with access to additional TV channels with and 20

21 Table 3: Viewership time as a function of rank (2012) Dependent variable: log(viewership hours) log(rank) Total TV log(rank) * Total TV Constant (0.004) (0.059) (0.005) (0.042) Observations 414,472 R Note: p<0.1; p<0.05; p<0.01 Cluster robust standard errors in parentheses without TSTV. At this time 80% of TELCO s households were premium and the remainder 20% were standard. Most of TELCO s households had already used TSTV by May 2015, and 58% of households in the experimental sample used TSTV at least once between May 1st and May 12, Therefore, exploring the outcomes of this randomized experiment allows for characterizing the e ect of the introduction of TSTV at a later stage in the life cycle of the technology, which may help anticipate the e ect of extending existing TSTV capabilities in countries where TSTV is already available. A random sample of 40,566 premium households that did not subscribe the Entertainment Bundle before was selected for this experiment. A random subset with a third of these households was held out from any intervention during the experiment control group. Another random subset of these households were o ered access to the Entertainment Bundle for free without TSTV that is, the content aired in the channels o ered as part of this bundle could only be watched live. Finally, the remainder of these households were o ered access to the Entertainment Bundle for free with TSTV that is, the content aired in 21

22 the channels o ered as part of this bundle could be consumed both live and using TSTV. The activation of the Entertainment Bundle, with and without TSTV, did not require any intervention of the households in the experiment. They simply received an and a text message notifying them about this temporary o er, which was readily available to use. For technical reasons, the activation of the Entertainment Bundle took one week to complete for households in the sample. The order in which households were activated was random and this week of data was removed from our analysis. The experimental panel that we analyze below covers 6 weeks of data between May 18th and June 30th The design used by TELCO for this randomized experiment is similar to that employed in Godinho de Matos et al. (2015) where the authors studied the substitution between TV and Internet consumption for a sample of pirate households. Anumberofhouseholdsweredroppedfromtheanalysis.Forexample,some households did not register any TV usage during the whole experiment. Other households churned during the experiment and some households had opted-out from marketing campaigns. This attrition in the sample a ected the three experimental groups of households similarly: 11,631 households remained in the control group, 11,752 households remained in the group of households that were given access to the Entertainment Bundle without TSTV and 11,724 remained in the group of households that were given access to the Entertainment Bundle with TSTV. We note that the attrition in the sample was orthogonal to treatment assignment and therefore our findings have causal interpretations. The caveat is that they generalize only to the population of households that watches TV on a regular basis, do not churn and do not opt out from marketing campaigns. This is still the most interesting population of households to study the e ect of TSTV on TV viewership behavior. 5.2 Descriptive Statistics Table 4 shows that TELCO s randomized schedule achieved good balance in key observed covariates across groups of households. The data in this table covers the 22

23 period between May 1st and May 12th 2015, thus before the experiment started. Figure 6 shows the viewership of the channels included in the Entertainment Bundle before and during the experiment. We observe that the households that were o ered these channels started watching them right away. Therefore, learning e ects are unlikely to be at play in our setting. Furthermore, households that obtained these channels with TSTV watched more of them than those that obtained them without TSTV throughout the whole experiment. Table 4: Balance obtained with randomization in pre-treatment covariates (2015) Treatment Group Variable Control Avg Treatment Avg T Test P Value 1 E. Bundle No TSTV Total TV time (min) E. Bundle No TSTV Total CPack time (min) E. Bundle No TSTV TSTV TV time (min) E. Bundle No TSTV Downloads (Mb) 1, , E. Bundle No TSTV VoD expenditure (cents) E. Bundle No TSTV Prior Customer Retention E. Bundle TSTV Total TV time (min) E. Bundle TSTV Total CPack time (min) E. Bundle TSTV TSTV TV time (min) E. Bundle TSTV Downloads (Mb) 1, , E. Bundle TSTV VoD expenditure (cents) E. Bundle TSTV Prior Customer Retention Entertainment Bundle Total Time (min) May 01 May 15 Jun 01 Jun 15 Jul 01 Control Treated No TSTV Treated TSTV Figure 6: Average daily minutes viewed in the Entertainment Bundle (2015) We also observe that some households in the control group watched the TV 23

24 channels o ered as part of the Entertainment Bundle. This arises because some of these households subscribed the Entertainment Bundle organically. Essentially, we have non-compliance with treatment assignment in all of the groups of households used in the experiment: 10.7% of the households in the control group watched channels o ered as part of the Entertainment Bundle at least once for more than 90 consecutive minutes; 50% of the households treated with the Entertainment Bundle without TSTV watched channels o ered as part of this bundle at least once for more than 90 consecutive minutes; and 59.2% of the households tread with the Entertainment Bundle with TSTV watched channels o ered as part of this bundle at least once for more than 90 consecutive minutes. We choose 90 consecutive minutes to evaluate compliance with treatment because this is the average duration of the movies broadcasted in the channels o ered as part of the Entertainment Bundle. In face of these statistics, our OLS results for the e ect of treatment assignment measure the Intention to Treat (ITT), which is a lower bound for the Average Treatment E ect (ATE). Households in our sample watched, on average, 5.0 hours of TV per day before the experiment started. This is an increase of 1.6 hours per day relative to what they used to watch in Figure 7, shows the breakdown of TV viewership time, live and TSTV, per channel type. This figure mimics Figure 3 from General purpose free to air channels account for most of the viewership, both in live and in TSTV, followed by entertainment, news, and children channels. However, the share of viewership time devoted to general purpose free to air and entertainment channels is larger in TSTV whereas that devoted to news is substantially smaller. Figure 8, shows the cumulative distribution of TSTV viewership by number of days elapsed since the program was first broadcasted live on TV. This figure mimics Figure 4 from Figure 8a, shows that on average 80% of the programs watched using TSTV aired in the previous 48 hours. Figure 8b, breaks down the former statistic per type of content. As one could expect, this statistic is larger for time-sensitive content, such as sports and news, and smaller for entertainment and documentaries. Finally, Figure 9, shows the 24

25 breakdown of live and TV viewership per hour of the day at which programs were originally broadcasted. Primetime content, originally aired between 8pm and midnight, is the most watched in both consumption modes and significantly more so in TSTV adult % 52.2 % other music international Market share % 11.1 % 10.7 % 25.4 % 10.2 % lifestyle documentaries sport children news entertainment general 0.00 Live TV Viewership Type TSTV Figure 7: Distribution of Live and TSTV viewership time by channel type (2015) 5.3 Empirical Strategy and Identification Measuring Impact on TV Viewership Time We explore the outcomes of this randomized experiment using an empirical strategy similar to the one we used before to analyze the data from We use di erence-in-di erences to compare households that were given access to the Entertainment Bundle with and without TSTV. Therefore, we run the following regression: Y it = During t + 2 T reatedt ST V i + 3 T reatedt ST V i During t + it (3) where i indicates a household in the experiment that was o ered access to the Entertainment Bundle and t indicates a week. During t indicates the ex- 25

26 P( Viewership delay < x ) P( Viewership delay < x ) Viewership delay in days (view time in time shift program air time) Viewership delay in days (view time in time shift program air time) entertainment general news sport (a) Overall view lag (2015) (b) View lag by channel type (2015) Figure 8: TSTV view lag (2015) 2015 Distribution of view time per content types % 3% 30% 30% 21% 12% 6% 5% 46% 19% 14% 9% Morning Noon Afternoon Prime Time Night Dawn Live Viewership Type TSTV Figure 9: Distribution of Live and TSTV viewership time by time block (2015) 26

27 perimental period and T reatedt ST V i indicates whether household i was given access to the Entertainment Bundle with TSTV. More importantly, in the case of the randomized experiment run by TELCO in 2015, we have identification by design because whether a household was given access to the channels o ered as part of the Entertainment Bundle with TSTV was determined at random. Therefore, we estimate this model using OLS. We are interested in parameter 3, which compares the outcome between households that were given access to the Entertainment Bundle with and without TSTV Measuring Impact on TV Viewership Concentration We also use an empirical strategy similar to one used before to analyze the e ect of TSTV on the concentration of TV viewership. In this case, we run the following specification over programs o ered as part of the Entertainment Bundle : log(t ime k )= T reatedt ST V k + 3 log(rank k )+ 4 log(rank k ) T reatedt ST V k + k (4) where T ime k represents the total viewership time devoted to program k by households in the experiment during the experimental period and Rank k indicates the rank of observation k in the corresponding distribution of viewership. In this case, T reatedt ST V k indicates whether observation k is computed across households that were o ered the Entertainment Bundle with TSTV. As before, we expect 3 to be negative and we are interested in the sign of coe cient 4, which measures how the distribution of viewership time across the programs o ered as part of the Entertainment Bundle is di erent between households that were given access to them with and without TSTV. 27

28 5.4 Results Impact of TSTV on TV Viewership Time Table 5 shows the results obtained using the di erence-in-di erences approach described before. Column (1) shows that households that were given access to the Entertainment Bundle watch more hours per day (4.5 min) of TV (p <0.01) than households that were given access to the Entertainment Bundle without TSTV. This represents an increase of 1.5% upon the average 5.0 hours per day of TV viewership across households included in the experiment before the experiment started. Column (2) shows that households that were o ered access to the Entertainment Bundle with and without TSTV watch the same amount of live TV during the experiment. This result confirms the previous finding from 2012 that the introduction of TSTV does not seem to reduce the amount of time that households devote to watch live TV. Table 5: Viewership time as a function of TSTV (2015) Dependent variable: Total TV Live TV Time-Shift TV (1) (2) (3) Treated TSTV (0.038) (0.036) (0.008) After (0.022) (0.020) (0.004) Treated TSTV * After (0.021) (0.020) (0.005) Constant (0.030) (0.028) (0.006) week Yes Yes Yes Observations 182, , ,507 R Note: p<0.1; p<0.05; p<0.01 Cluster robust standard errors in parentheses. Observations clustered at the household level. Figure 10 shows the e ect of o ering the Entertainment Bundle with TSTV 28

29 relative to o ering it without TSTV on the total consumption of TV per type of channel. Vertical bars represent the 95% confidence interval for the coe cients. The regression results used to construct this figure are included in appendix A. Households that were o ered access to the Entertainment Bundle substitute viewership in the original entertainment channels for viewership in the Entertainment Bundle. This substitution is stronger for households that were o ered access to the Entertainment Bundle with TSTV and this is what leads to the increase in total TV consumption with TSTV Treated TS * After Adult Children Documentaries E. Bundle Entertainment General International Lifestyle Music News Other Sport Effect of Treatment (in hours) Figure 10: Impact of TSTV on TV viewership by channel type We are unable to directly compare the magnitude of the e ect of the introduction of TSTV on TV viewership in 2012 and 2015 because the contexts in which these interventions were carried out are significantly di erent. TELCO introduced TSTV for the first time in At that time, TELCO made TSTV available on 50 TV channels out of the 100 TV channels included in the basic TV package. In 2015, TSTV was added to the set of 10 TV channels o ered as part of the Entertainment Bundle. Not only the number of channels receiving TSTV was di erent in 2012 and 2015 but also the channels that obtained TSTV were of a di erent nature. The 50 TV channels that obtained TSTV in 2012 included general content, entertainment, documentaries, news, sports, chil- 29