TV Channel Search and Commercial Breaks - GitHub Pages · The data patterns support the idea that viewers search for ... timing of breaks affects TV channels ... Television is still

SONG YAO, WENBO WANG, and YUXIN CHEN*

This study investigates time lapses that interrupt product consumption.Preeminent examples are commercial breaks during television or radio pro-gramming. The authors suggest that breaks facilitate consumers’ search foralternatives. Specifically, when there is so much uncertainty that consumersare unclear about utility levels of different products, they engage in costly searchto resolve the uncertainty. For TV programming, breaks lower the opportunitycost of search, allowing the consumer to sample alternative channels withoutfurther interrupting the viewing experience on the current channel. Using datafrom the Chinese TVmarket, the authors estimate a sequential searchmodel tostudy consumer TV channel choice behavior. The data contain a quasi-naturalexperiment due to aChinese government policy change on commercial breaks.The natural experiment creates exogenous variations in the data that enable theempirical identification of heterogeneous consumer preference and search cost.The data patterns support the idea that viewers search for alternatives duringcommercial breaks. Drawing on the estimates, the authors investigate how thetiming of breaks affects TV channels’ viewership, offering insights about how tostrategically adjust the timing of breaks.

Keywords: advertising, television, consumer search, natural experiment,demand estimation

Online Supplement: http://dx.doi.org/10.1509/jmr.15.0121

TV Channel Search and Commercial Breaks

Television is still the dominant medium for advertising. As of2013, TV commercials account for 40% of global advertisingspending and will remain one of the most significant advertisingchannels in the foreseeable future (Yeh and Zhang 2013). Thisarticle investigates how TV viewers make channel-switching deci-sions when they face uncertainty about the programming of regularshows and commercials on alternative channels. Furthermore, we

explore strategic decisions of TV channels about the timingof commercial breaks in response to viewers’ switching activity.

Television programming changes over time. As a result, themost preferred channel for a viewer varies as time passes.However, at any given point, the viewer has uncertainty aboutthe programming of alternative channels other than the one heor she iswatching. To determinewhether any alternative channelsare preferable to the current one, the viewer must search thealternatives to resolve the uncertainty. Such searches are costlyto the viewer due to the time and cognitive efforts spent duringthe search.1 Furthermore, they disrupt the viewing experienceon the current channel. Accordingly, viewers may refrain fromsearching during the programming of regular shows. Conven-tionalwisdomholds that viewers dislike commercials (e.g., Elliot2004). However, during a commercial break, channel searchdoes not further disrupt one’s regular show-viewing experience.Thus, the commercial breakmay be a natural opportunity for theviewer to search and switch to alternatives. Correspondingly,the timing of commercial breaks becomes a crucial strategic

*SongYao is Associate Professor ofMarketing, Carlson School ofManagement,University of Minnesota (e-mail: [email protected]). Wenbo Wang is AssistantProfessor of Marketing, Hong Kong University of Science and Technology(e-mail: [email protected]). Yuxin Chen is Distinguished Global Professor ofBusiness, New York University Shanghai, with affiliation at the Stern School ofBusiness, New York University (e-mail: [email protected]). This manuscript waspreviously circulated under the working title “The Value of Sampling.” The authorswould like to thank seminar participants at Hong Kong University of Science andTechnology, Shanghai Jiaotong University, Stanford University, University ofHouston, University of Southern California, Washington University at St. Louis,and Marketing Science Conference 2014, as well as Anja Lambrecht, Tat Chan,AnthonyDukes,WesHartmann, JimLattin, DinaMayzlin, HarikeshNair, NavdeepSahni, and Stephan Seiler, for their feedback. The research is partially funded by theHong Kong Research Grants Council Early Career Scheme (#699213) and theMcManus Research Grant at the Kellogg School of Management, NorthwesternUniversity. The authors are listed in reverse alphabetical order. Coeditor: RajdeepGrewal; Associate Editor: Avi Goldfarb.

1In our setting, we specifically consider viewers switching channels toresolve the uncertainty. Alternatively, it is possible to explore channels usingmagazines such as TV Guide or the “picture-in-picture” technology equippedon some TV sets, which also involves the cost of time and effort.

© 2017, American Marketing Association Journal of Marketing ResearchISSN: 0022-2437 (print) Vol. LIV (October 2017), 671–686

1547-7193 (electronic) DOI: 10.1509/jmr.15.0121671

decision for TV channels because viewers’ channel-switchingactivity bears crucial implications for advertising revenues.

To investigate the TV-viewing behavior of consumers whenthey face uncertainty, we calibrate an empirical model underthe framework of the classical economic search model. We usedetailed rating, program-scheduling, and individual channel-switching data from the Chinese TV market, taking into ac-count the effect of commercial breaks on viewing experience. Ourdata cover the period from December 11, 2011, to January 19,2012.One unique feature of our data is an exogenous policy shockthat dramatically changed the distribution of commercial breaksin TV channels’ programming. On January 1, 2012, the Chinesegovernment abruptly changed its regulation of commercial breaksfor all episodic TV series. Before January 1, 2012, TV channelscould broadcast commercials (1) between two different shows, (2)between two episodes of the same show (both cases are labeled“between-show” henceforth), and (3) during a show or an episode(labeled “in-show” henceforth). Episodic TV series between 7 and9 P.M. could have at most one commercial break per episode, andthe break was not to exceed one minute. Starting on January 1,2012, however, the authorities banned all in-show commercialbreaks for episodic TV series nationwide with the intention ofimproving consumers’ viewing experience. This dramatic regu-latory change was announced on November 25, 2011, less than40 days before it became effective (State Agency of Radio, Filmand Television of the People’s Republic of China 2011). Thisabrupt announcement left TV networks little time to strategicallyadjust their programming in response to this new rule, especiallywhen commercial slots inChinawere normally sold severalmonthsin advance of the broadcasting. Networks had to fully refund ad-vertisers for their in-show commercial slot purchases ormove somein-show commercials into between-show slots in conjunction witha partial refund. Consequently, for the period of our observationwindow, this regulatory change created a quasi-natural experimentthat exogenously changed the distribution of commercial breaks inprogramming (both amounts and types) and allowed us to observeTV-viewing behavior both before and after the change. Moreimportantly, the exogenous data variations underpin our empiricalidentification strategy as we discuss in the section “Identification:The Separation of Utility and Search Costs.”

Using our model, we are able to show that the effects ofcommercials vary across channels. Suppose that a viewer iswatching a channel that turns out to be of low utility to him or her.During commercial breaks, the viewer is likely to search alter-native channels with higher (expected) utility levels because themarginal gain of the search is high. When the commercials areremoved altogether, the viewer is likely to refrain from searchingduring regular shows because the search leads to disruption of thecurrent show that would not otherwise exist. Thus, the viewerbecomesmore likely to stickwith the current channel.2 In contrast,if the viewer is watching a preferred channel, the marginal gain ofsearching alternative channels is low even during commercialbreaks. Thus, the viewer is less likely to search the alternativesanyway.Consequently, the removal of commercials has less impacton the viewing behavior of a viewer who is already watching apreferred channel.

Drawing on such insights, we explore the implications of thetiming of commercial breaks across TV channels. A TV channel

may have the incentive to either coordinate (synchronize) itscommercial breaks with other channels or differentiate the timingfrom competing channels, depending on how viewers make theirdecisions. In our empirical application, we find that low-ratedchannels should try to synchronize their breaks with high-ratedones.Doing so lowers the expected return of searching alternativesduring breaks, because other channels will also be airing com-mercials. Thus, it discourages viewers from switching channelsduring breaks. In contrast, a high-rated channel should try todifferentiate the timing of its breaks from competing channels,especially those with lower ratings. With the differentiation, thehigh-rated channel may poach viewers from competing channelsthat are on commercial breaks, because the high-rated channel isnot broadcasting commercials at the same time. Meanwhile, thehigh-rated channel loses fewer viewers to competitors during itsown breaks (vs. a low-rated channel).

The contributions of this article are threefold. First, we advancethe empirical literature on consumer TV-viewing behavior. Werelax the assumption that viewers have full information aboutprogramming. Indeed, we propose that viewers’ switching de-cisions inherently depend on their uncertainty about the pro-gramming of regular shows and commercials of alternativechannels. The seminal work of Lehmann (1971) has led to agrowing body of studies exploring consumer TV show choicesand switching decisions, such as Goettler and Shachar (2001),Rust and Alpert (1984), Shachar and Emerson (2000), Wilbur(2008), and Yang et al. (2010). One common assumption in theliterature is that viewers have little uncertainty about alternativeoptions; a few exceptions are Moshkin and Shachar (2002),Byzalov and Shachar (2004), and Deng and Mela (2017), whichassume that viewers have uncertainty about shows beforewatching. In particular, the first two articles explore the in-formational role of TV commercials. They show that promotionalads for upcoming shows by the networks (“Tune in”) reduce theuncertainty and increase the likelihood of better matches betweenviewers and shows. In our model, we explicitly consider viewers’channel choices under the framework of a classical sequentialsearch model. The consumer must search to know exactly analternative channel’s programming. The observed TV ratings andchannel-switching activities are the outcome of a unified frame-work of viewers’ optimal search and utility maximization.

Second, the identification of empirical search models is oftenproblematic because consumer preference and search cost areconfounded in field data (e.g., Sorensen 2000). The growingempirical literature on search models has paid considerable at-tention to addressing this identification concern (e.g., Chen andYao 2016; De Los Santos, Hortacsu, andWildenbees 2012; Hongand Shum 2006; Honka 2014; Hortacsu and Syverson 2004;Koulayev 2014; Pinna and Seiler 2017). In our data, as an ex-ogenous shock, the regulation changed the distribution of pref-erence independent of search cost, providing us with a convincingidentification approach.

Third, with the second contribution of identifying the searchmodel, we are able to advance the research on the timing of breaksthat interrupt product consumption. The timing of commercialbreaks is an important strategic decision for TVchannels and radiostations (e.g., Sweeting 2006, 2009; Wilbur, Xu, and Kempe2013). While our empirical context is the TV industry, our re-search also sheds light on other scenarios in which breaks havelower utility levels than the consumption utility of the focalproduct. For example, consumers may experience such breaks inthe context of video games after they finish playing one game but

2Under full information, where the viewer knows the details of all showsacross all channels, there would be no need to search. The optimal choice forthe viewer is to watch the most preferred channel of that moment.

672 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

have to wait for its sequel to be launched. On one hand, bysynchronizing breakswith competingfirms, a firm can prevent itsown consumers from leavingduring its breaks.On the other hand,by differentiating the timing of breaks, a firm can potentiallypoach consumers from competitors while the competitors are onbreaks. The trade-off depends on the characteristics of consumersin a specific market. With our empirical model, we can char-acterize consumer TV-viewing activities and thus offer mana-gerial prescriptions for TV channels pertaining to their timing ofcommercial breaks.

The article is organized as follows.We first discuss the data thatunderpin our study and providemodel-free evidence about viewersearch activity.Next,we detail the sequential searchmodel used todescribe viewing behavior and discuss the estimation strategy andidentification. Then, we present the results and explore policyimplications. We conclude with a discussion of main findings.

DATA

Thedatawere provided by a leadingmedia research company,whose identity we cannot disclose for reasons of confidentiality.The company compiles data on the world’s largest TV-viewingaudience in mainland China and Hong Kong. Using set-topmeters, the company collects and constructs TV ratings data thatrepresent the viewing activities of about 370 million householdsin mainland China and 2.4 million households in Hong Kong.(Throughout this article, we use “viewer” and “household” in-terchangeably; we do not explicitly consider group decisionswithin a household, as discussed in Yang et al. [2010]).

One unique feature of the data is that they cover the period of aquasi-natural experiment. On January 1, 2012, the Chinesegovernment banned all in-show commercials for episodic TVseries. This swift policy change left the networks with little timeto strategically change their programming in the short run. Thisis because, according to Chinese government regulation, anyprogramming change by TV networks needs a prolonged reviewprocess by the government agency, which takes more than50 days for just the initial round. Any appeal to an initial denialtakes another 30 days (State Agency of Radio, Film andTelevision of the People’s Republic of China 2010). Because ofthis long review process, and because neither TV networks noradvertisers were aware of the new policy before its announce-ment on November 25, 2011, commercial slots had been soldseveral months before the regulatory change announcement. Weare confident that neither of these parties anticipated the changebecause if they had, they would have not sold or bought thein-show commercial slots.3 According to our discussion with

multiple networks and advertisers, networks had to fully refundadvertisers for their in-show slot purchases or move the com-mercials to between-show slots and issue a partial refund. Thedistribution of commercials was changed by the plausible ex-ogenous shock of the new policy in both amounts and types.Consequently, we observe TV-viewing behavior under differentdistributions of commercials; within at least a short time win-dow, there are minimal changes in TV programming of regularcontent, due to the long administrative review process.

We focus on a short period of prime-time data from the BeijingTV market, eight days before and eight days after the policychange. Specifically, the data are from Monday to Thursdayduring theweeks ofDecember 11 andDecember 18, 2011, andJanuary 8 and January 15, 2012. In the data, we observe thefollowing components that are crucial for our analyses of con-sumer viewing behavior:

• Rating data from the top 29 channels of one-minute intervalsduring prime time, from7:30 to 10 P.M.: These 29 channels accountfor 80% of the Beijing TV market share. Following the industrystandard, the rating of a channel for a given period is defined as thepercentage of viewers who have tuned to that channel during thatperiod, out of all viewers who own TV sets. In other words, therating data reflect the market shares of the channels during eachone-minute interval, including the share of people who are notwatching TV at that moment.

• Individual-level, set-top box TV-viewing data of 1,022 viewers:These viewers are a representative, random sample from the panelused for calculating the channel ratings. For each viewer, weobserve which channel was being watched (including no channel,if TVwas not on) on a second-by-second level from 7:30 to 10 P.M.

• Programming data from each channel: For each one-minute in-terval, we observe whether the interval is broadcasting in-showcommercials, between-show commercials, or regular shows,whose genres include episodic TV, sports, medical and health,news, and so on. (In the data, a minute is defined as a commercialbreak if it contains at least 30 seconds of commercials.) We latercontrol for these genre fixed effects of regular shows in ourempirical estimation.

TV Market Before and After the Ban

The ban on in-show commercials had a profound effect onconsumer viewing behavior and, hence, TV ratings. First, theregulation inevitably affected the amounts and types (in-showvs.between-show) of commercials during the total broadcastingtime. Table 1 documents the changes. For each channel, wecalculate the percentage of commercial break minutes during agiven hour (i.e., 60minutes).We report the distribution of the in-show and between-show commercial percentages across chan-nels, episodic versus nonepisodic shows, and before versus afterthe commercial ban. In Appendix B, we also report the sameset of statistics after further dividing channels into three tiersaccording to their October/November 2011 median ratings (i.e.,high-rated, median-rated, and low-rated channels). The distri-butions across rating tiers are similar to each other and to those inTable 1.

While the large standard deviations of the statistics prevent usfrom reaching statistically meaningful conclusions, we may stillobserve some patterns. The percentages of in-show commercialsdropped after the ban; in-show commercials for episodic TVseries dropped to 0. Between-show commercials increasedslightly. This is consistent with our discussion with channelmanagers and advertisers, which revealed that some in-showcommercials soldwere shifted to between-show slots.We further

3Another piece of evidence that networks and advertisers were unaware ofthe regulatory change beforehand comes from online keyword search vol-ume. In Appendix A, we depict the online search volume for the string “TVshow commercials” (in Chinese) using Google Trends Index from January 1,2011, to December 31, 2012. In November 2011, upon the announcement ofthe new policy, the search index reached its highest level (Google alwaysnormalizes the highest search volume to 100). However, before the an-nouncement, the index was consistently at a much lower level for months.Note that the index does not directly measure the knowledge networks andadvertisers had about the new policy; however, if the networks and ad-vertisers had any knowledge in advance, we would expect at least someinformation leakage to have led to increases in online search of relevantkeywords days orweeks before the government announcement. Accordingly,the consistently stable and low search volume before November 2011supports the notion that networks and advertisers were not aware of the policychange in advance. Baidu Search Index, a similar search index by the searchengine Baidu, shows a similar pattern as Appendix A.

TV Channel Search and Commercial Breaks 673

evaluate the variation in the amount of commercials acrosschannels within the same hour. The purpose of this calculationis to see whether channels differed considerably in amount ofcommercials. More precisely, for each type of commercials be-fore and after the ban (episodic/nonepisodic in-show/between-show), we regress the amount of commercials on the day-hourfixed effects. We then compute the standard deviations of theresiduals of these regressions. The standard deviations of theresiduals, reported in Table 1, can be viewed as the magnitudeof variation in amount of commercials across channels withinthe same hour. In comparison with the overall variation acrossboth channels and day-hours, we can see that the across-channelvariation was much smaller, on average less than one-third ofthe overall variation. This implies that amount of commer-cials varied little across channels, and most of the fluctuationcame from the across-time variation.

Table 2 further shows the regular show programming (ex-cluding commercials) ratings of one-minute intervals before andafter the regulatory change. After the commercial ban, the ratingswere not significantly improved. In fact, the average rating acrosschannels and intervals dropped slightly after the ban (.77 vs. .65;see thefirst rowofTable 2).EpisodicTVseries’ ratings on averageincreased, while nonepisodic TV shows’ dropped, potentially dueto the change in amount of commercials across the two types ofshows.

One question of interest is whether ratings within a channelwere relatively stable over time (i.e., some channels consistently

had high ratings while others consistently had low ratings). Toanswer this question,wefirst compute the standard deviations ofone-minute ratings across both channels and one-minute in-tervals before and after the commercial ban. The results arereported in parentheses in the first row of Table 2. We thencompute across-time rating variation within a channel. Spe-cifically, we calculate the standard deviations of residuals fromregressions of ratings onto channelfixed effects, before and afterthe ban. Note that using log-ratings as the dependent variableleads to similar results. With channel fixed effects controlled,these standard deviations provide us with the assessment ofthe average rating fluctuation over time within each channel.The results are reported in the second row of Table 2. From theresults, we can see that before the ban, the majority of ratingvariation came from the difference across channels. Before theban, the overall standard deviation was quite high, reaching1.48. The across-time variation within a given channel was onlyabout one-quarter of that level (.39/1.48 = .26). After the ban, theoverall rating variation became smaller (.87). The across-timevariation (.37) still accounted for less than half of the variation.In other words, the ratings of each channel were relatively stableover time comparedwith the variation across channels. Thus,wenext consider the ratings within each rating tier of channels. Wefirst collect the median rating of each channel during Octoberand November 2011 (i.e., before the data window we use forestimation). We rank the channels from 1 to 29 according tothe median ratings and then categorize them into three tiers:

Table 1DESCRIPTIVE STATISTICS: PERCENTAGES OF COMMERCIAL TIME OUT OF TOTAL BROADCASTING TIME ACROSS CHANNELS AND

DAY-HOURS

M SDSD

Across Channels5th

Percentile95th

Percentile MaxAverageFrequency

Episodic Shows, In-Show CommercialsBefore the ban 1.41 2.15 .98 0 5.67 10.67 1After the ban 0 0 0 0 0 0 0

Episodic Shows, Between-Show CommercialsBefore the ban 4.26 7.79 2.12 0 16.67 21.67 2After the ban 4.95 8.24 2.27 0 23.33 35.00 2

Nonepisodic Shows, In-Show CommercialsBefore the ban 4.25 5.97 1.61 0 18.33 23.33 2After the ban 3.01 4.58 1.30 0 13.33 28.33 2

Nonepisodic Shows, Between-Show CommercialsBefore the ban 3.41 5.36 1.31 0 13.33 26.67 2After the ban 3.03 4.56 1.15 0 11.67 21.67 2

Table 2DESCRIPTIVE STATISTICS: RATINGS OF REGULAR SHOW DURING ONE-MINUTE INTERVALS (EXCLUDING COMMERCIAL BREAKS)

Average Before the Ban (SD) Average After the Ban (SD)

One-minute ratings across channels and intervals .77 (1.48) .65 (.87)One-minute ratings across intervals (SD, within channel across-time variation) (.39) (.37)

By Show Type Across Channels and IntervalsEpisodic TV series one-minute ratings 1.06 (2.28) 1.24 (1.19)Nonepisodic shows one-minute ratings .53 (.50) .46 (.54)

By Channel Rating Ranking Across Channels and IntervalsChannels 1–10 one-minute ratings 1.48 (2.25) 1.03 (1.22)Channels 11–20 one-minute ratings .45 (.42) .45 (.36)Channels 21–29 one-minute ratings .28 (.29) .41 (.26)


high-rated, median-rated, and low-rated channels (indexed aschannels 1–10, channels 11–20, and channels 21–29, re-spectively). As shown in Table 2, there was a drop in the ratingsof high-rated channels (channels 1–10) after the ban. In com-parison, low-rated channels (channels 21–29) witnessed anincrease in their ratings after January 1, 2012. There may bealternative factors contributing to this observed pattern,4 andwith the large standard deviations, we cannot obtain conclu-sive insights without a formal model. However, one potentialexplanation is that this pattern is consistent with the consumersearch conjecture we proposed earlier. Low-rated channelsmight still attract a reasonable amount of viewers becausesometimes they still broadcast high-quality shows. Before theban, peopleweremore likely to switch to higher-ranked channelsduring commercial breaks, especially if the utility of watchinglow-rated channels dropped at the time. After the ban, however,people weremore likely to staywith their original channels. Thiswould result in the average ratings increase for low-rankingchannels and the decrease for high-ranking channels.

Viewers’ searching behavior also changed after the regulation.In the individual-level data, we observe a viewer’s second-by-second activities. We first define “searching a channel” as stayingat a given channel for at least five seconds to explore the pro-gramming at that channel. (We considered alternative intervals forthe definition, including three, seven, and ten seconds; the insightsare unchanged.) We also define “channel chosen” during a one-minute interval as the channel that is watched the longest.5 Wecalibrate the average number of channels searched during a one-minute interval across viewers and intervals. Table 3 shows thenumber of searches before and after the regulatory change. Wefind that the average number of searches across viewers andintervals decreased after the ban. The ban had a bigger impact forintervals with episodic TV shows, which is not surprising becausethe regulation only applies to such shows. While there may bealternative explanations for the decrease in the number of searchesand the changes are statistically insignificant, the average effectsare again consistent with our conjecture: with fewer in-showcommercial breaks for TV shows, viewers on average searchless and thus are more likely to stay with their original channels.Next, we provide some additional evidence for consumer search.

Evidence of Consumer Search

In this section,we further consider evidence fromdata to showthat viewers search for alternative channels during commercialbreaks.

Evidence from aggregate rating data. The first piece ofevidence is predicated on the notion that the viewer switches toalternative channels during a commercial break. After the break,the viewer should switch back to the original channel if it has ahigher utility level than the alternative channels. If, on average,

many viewers do not return to the original channel, we can inferthat some alternative channels have higher utility. However, ifthe viewer has full knowledge about the higher utility of thealternative channel, as a rational agent, he or she should havewatched that channel even before the commercial break. Em-pirically, if we observe that postcommercial ratings of channelsare on average lower than their precommercial levels, it isconsistent with our conjecture about viewers’ uncertainty of al-ternatives and searching and switching to better alternatives (notethat this is only a necessary condition for “search”). Accordingly,we consider a linear regression of one-minute ratings (in loga-rithm) on commercial dummy and lagged commercial dummy(lagged by one-minute interval), after controlling for channel,hour,weekday,week, and showfixed effects. Table 4 presents theresults (showing only coefficients regarding commercials). Wecan see that the commercial variable has a significant impact onratings. More importantly, the lagged commercial variable alsohas a significant negative effect on ratings, which is consistentwith our conjecture.

An alternative explanation for the observed pattern, however, isviewer inertia. After the commercial break ends at the originalchannel, viewers who do not return might be lazy or forget toswitch back. As a robustness check, we create a dummy ofprecommercial rating. For a given channel during a given period,the dummy takes the value 1 if the channel’s precommercial ratingis lower than the then-current median rating across all channels.We then reestimate the linear regression model given earlier butfurther interact the lagged commercial dummy with the pre-commercial rating dummy. Intuitively, if inertia is the reason thatratings do not return to the precommercial levels, the laggedcommercial should have similar impact across high-rated channelsand low-rated ones. In contrast, if viewer search is themain reason,the channels with a lower precommercial rating should suffer agreater reduction in their rating after the commercial ends. This isbecause viewers who were watching channels with lower pre-commercial ratings will be more likely to search, find betterchannels, and not switch back to the originals. Table 4 againpresents the results. Consistent with the explanation of viewersearch, the interaction term has a significant negative coefficient,implying a greater impact on channels with low precommercialratings. In fact, the effect of lagged commercial dummy hasbecome insignificant. In other words, although people mightstill switch away during commercials, they switch back after thecommercial ends if the original channel had a high rating beforethe commercial break.

Evidence from disaggregate data.We have access to 1,022viewers’ TV-watching data up to the second-by-second level.We consider somemodel-free evidence in addition to evidencediscussed earlier for viewers’ searching activities, using theseindividual-level data.

Before the ban, by government regulation, episodic TVshows during prime time could only have one in-showcommercial break per episode, and that break could not belonger than oneminute (National Bureau of Radio,Movie, andTelevision 2009). Accordingly, for episodic TV shows beforethe ban, at each in-show commercial break, we track everyviewer’s channel-switching activity for five seconds before,one minute during, and thirty seconds after the commercialbreak—in total, 95 seconds. We divide the 95 seconds into19five-second intervals. Thefirst interval starts right before thein-show commercial break. The commercial break starts onthe secondfive-second slot and ends at the thirteenth.We index

4For example, one alternative possibility is that right after January 1, low-rated channels introduced some hit shows that were more popular than theshows on high-rated channels. This scenario, however, is very unlikelybecause it requires the average show popularity across all low-rated channelsto have been higher than the average across all high-rated channels after butnot before January 1, 2012.

5Under this definition, the channel watched the longest may not be the onewatched last during an interval. In such cases, we choose to drop thosechannels after the “watched/chosen” channel. Conceptually, this implies thatthe viewer engages in a new search process during this interval after he/shehas finished one search process and decided on a channel. We consider onlythe first round of the search process in the model and estimation.


the initial channel watched by each individual before thecommercial break as “1,” the second channel watched for atleastfive seconds as “2,” the third channel as “3,” and so on. Byconstruction, at the first interval all individuals are on channel“1,” that is, their initial channels. Depending on a viewer’sinitial channel’s ranking at the time of the first interval (i.e.,precommercial rating), we divide individuals into three groups:low-ranked, median-ranked, and high-ranked initial channels.(In the Web Appendix, we further show a set of 29 graphsrepresenting a group of viewers whose initial channels areranked 1–29. The insights of the 29 graphs are similar tothose of the three graphs in Figure 1.)

In the second interval, when the commercial starts, someviewers switch to other channels (channel “2”). People con-tinue switching channels in the following intervals. We ploteach viewer’s activities in Figure 1, where the three subfigurescorrespond to low-, median-, and high-ranked initial channels.The horizontal axis stands for time intervals, and the verticalaxis represents searched channel indices. Correspondingly,each dot in the graph is the combination of a five-second timeslot and a channel index. To understand the figure, imaginethat a viewer stays at the same channel throughout the 95seconds. In this case, we should see this viewer appearing onchannel “1” for all time intervals. If the viewer starts on aninitial channel and then switches to a second channel and staysthere, we should see this viewer appearing on channel “1” ininterval 1 and then on channel “2” for the remaining time slots.

Because there may be many viewers on each spot in the graph,to avoid overplotting and clearly show the patterns in the data,we allow the points to jitter. Accordingly, if there are moreindividuals at a given spot in the graph, that spot will show ahigher density of dots. To make the densities even moretransparent, for each group (low-, median-, and high-rankedinitial channels), we also depict the percentage of individualson each channel at a given time interval. For example, in thesecond interval of the first graph, channels “1” and “2” show“54.8” and “45.2,” respectively. This means that when thecommercial starts on the second interval, 54.8% of individualsstay at their initial channels and 45.2% switch to their secondchannels.

From these figures, we can see that when the commercialbreaks start at the second interval, some viewers stay at channel“1,”whilemany switch to channel “2.”At the third five-secondinterval, even fewer people stay at their original channels, andsome start to explore their third channels. The length of an in-show commercial break is one minute, per the regulation.Rational and fully informed viewers who were watching theirmost preferred shows before the commercial break shouldreturn to their original channel “1” when the break ends, afterabout one minute (the 13th five-second interval). However, asdemonstrated by the dense plots and high percentages onchannels “2” and “3” across all three graphs, the majority ofviewers do not return to channel “1” and instead stay at one ofthe alternatives they have searched during the commercial

Table 4EFFECT OF COMMERCIALS AND LAGGED COMMERCIALS ON LOG RATINGS

Log Rating

Estimates (SE) WithoutLagged Commercials

Estimates (SE) withLagged Commercials

Estimates (SE) with Precommercialsand Lagged Rating Interaction

Constant .610 (.048) .612 (.048) .615 (.048)Commercial dummy −.201 (.046) −.103 (.043) −.106 (.047)Lagged commercial dummy — −.130 (.013) −.042 (.034)

Lagged commercial dummy interacted — — −.151 (.042)with precommercial rating

Channel fixed effects Yes Yes YesHour fixed effects Yes Yes YesWeekday fixed effects Yes Yes YesWeek fixed effects Yes Yes YesShow fixed effects Yes Yes YesNumber of observations 69,600 69,136 68,672Adjusted R2 .68 .69 .69

Notes: Boldface indicates that the estimate is significant at the 95% confidence level.

Table 3DESCRIPTIVE STATISTICS: SEARCH ACTIVITIES DURING ONE-MINUTE INTERVALS

Average Before the Ban (SD) Average After the Ban (SD)

Number of searches in one minute across viewers and intervals 3.53 (1.94) 2.70 (1.41)

By Show TypeViewers who were watching episodic shows 2.89 (1.94) 1.98 (.90)Viewers who were watching nonepisodic shows 4.33 (2.46) 4.30 (2.23)

By Rating RankingViewers who were watching Channels 1–10 3.12 (1.99) 3.01 (1.97)Viewers who were watching Channels 11–20 3.59 (2.32) 3.51 (2.21)Viewers who were watching Channels 21–29 3.84 (2.13) 2.97 (2.03)


breaks. The high-ranked initial channel group has a slightlyhigher percentage of viewers switching back to their initialchannels. In contrast, the low-ranked initial channel group hasmore people staying at new channels. The pattern in the figure

is again consistent with the findings in the previous analyses. Itimplies that people take the opportunity of commercial breaksto explore alternatives, and the exploration may lead to optionsmore preferred than their previous choices. Consequently, we

Figure 1CONSUMER SWITCHING ACTIVITIES FOR EPISODIC TV SHOWS BEFORE THE COMMERCIAL BAN

0.01 0.02

5.31 6.3919.8854.8 3.78 2.72 2.02 4.04 6.4 12.11 26.42 29.98 34.7 37.88 41.25 41.23 40.73 40.97

39.3740.3338.9839.639.8640.0439.8739.8331.1221.9717.2511.3113.5517.6224.7726.3149.3745.2

30.75 44.11 37.41 33.04 27.3 25.48 30.13 32.74 31.86 23.86 22.64 19.52 17.63 15.36 15.92 15.28 15.86

3.373.223.343.284.174.946.35

0.40.390.510.470.430.731.051.515.5710.921.1418.776.51 13.11

2.39 6.32 7.77 4.21 2.64 0.88 0.13 0.11 0.07 0.03 0.04 0.02 0.04 0.02

0.82 1.15 0.6 0.27 0.05

8.2318.4125.0924.26 24.93 30.05 30.53 31.15

00

2

100

4

6

8

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Five-Second Intervals, Initial Channel Low-Ranked

Five-Second Intervals, Initial Channel Median-Ranked

Five-Second Intervals, Initial Channel High-Ranked

1.87 4.99 6.08 3.44 2.04 0.71 0.12 0.07 0.04 0.03 0.02 0.02 0.02 0.02

0.330.340.340.340.470.60.91.264.659.5213.1718.616.611.264.99

12.73 27.03 30.59 35.25 37.74 41.04 41.24 41.29 41.59

44.93 49.54 26.95 25.64 18.39 13.8 11.83 17.65 22.59 31.42 39.29 39.92 40.12 40.05 39.64 39.57 39.44 39.32

15.4115.5215.5815.5517.4819.132224.0432.1732.5230.0826.1928.0333.1537.614430.06

23.45 24.29 28.9 30.17 30.83 28.58 24.92 17.42 7.99 6.3 4.72 4.17 3.31 3.17 3.3 3.27

0.390.420.410.440.530.721.091.485.3610.7114.7120.4418.412.975.89

2.35 6 7.43 4.22 2.47 0.84 0.15 0.09 0.05

0.8 1.05 0.48 0.25 0.07 0.01 0 0 0 0 0

15.12

28.66

0.64 0.86 0.36 0.19 0.04 0 0 0 0 0 0 0 0

0.04 0.03 0.02 0.03 0.02

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0

2

4

6

8

0

2

4

6

8

100 55.07 20.41 5.6 6.56 4.24 2.8 2.23 4.27 6.53

20.55 22.22 27.4 29.29 30.23 27.13 23.46 16.21 7.12 5.66 4.34 3.64 2.93 2.91 2.94 2.97

14.6514.7214.6814.6416.5817.920.7623.0831.3432.8231.4427.8433.71 29.2637.342.727.53

42.88 49.83 28.82 27.7 20.68 15.73 13.61 19.47 24.34 32.73 39.74 40.24 40.04 39.63 38.92 38.85 39.06 38.9

43.1342.9243.243.1639.6437.0832.3728.6714.327.634.982.83.55.087.796.9322.6457.12100

Sea

rch

ed C

han

nel

Ind

exS

earc

hed

Ch

ann

el In

dex

Sea

rch

ed C

han

nel

Ind

ex


see that many people do not return to their original channel butinstead stay at an alternative channel. In conclusion, we findthe data patterns discussed in this section are consistent withthe model of consumer searching during commercial breaksfor better alternatives under uncertainty. In the next section, weformally introduce the sequential search model used to de-scribe TV-viewing behavior.

MODEL

Utility

During period t (defined as a minute), there are J + 1 al-ternative options available to consumer i, watching one of theJ TV channels (j = 1, 2, ..., J) or choosing the outside optionof not watching TV (j = 0).

The utility of viewer i for watching channel j during periodt is

uijt = g 0jtii + njt + bInShowAdi InShowAdjt

+ bBtwShowAdi BtwShowAdjt

+hbContinuei Iijt + bNoStarti

�1 − Iijt

�iSameShowjt + eijt,

(1)

where g jt is a vector of dummy terms, including fixed effectsof show genre, hour, weekday, and week6; ii is the vector ofthe coefficients for g jt; njt ~ Nðnj,s2

nÞ is a channel-time-specific intercept term, which follows a normal distribution,with themean as nj and standard deviation assn; and njt can beviewed as the channel’s “quality” at minute t that is commonacross individual viewers. Essentially, the mean nj can be seenas a channel fixed effect term that measures the average “quality”level of the channel, which is common across viewers andtime. Each period, the realized “quality”may deviate from themean nj, and sn captures the average magnitude of the de-viation. Note that these intercept terms (njt) are measuredagainst the baseline of “not watching TV,”which is normalizedas n0t = 0. InShowAdjt is the in-show commercial dummy forperiod t, which takes the value 1 if minute t is an in-showcommercial break. Similarly, BtwShowAdjt is the dummy forbetween-show commercials. Commercials affect one’s viewingexperience, and bInShowAdi and bBtwShowAdi account for such ef-fects. People often demonstrate strong state dependency in TV-viewing behavior (Byzalov and Shachar 2004), especially if theprogramming is a continuation of the same show. We thusdefine SameShowjt as a dummy variable taking the value 1 ifchannel j is broadcasting the same show during period t asthe previous period. And if period t is a commercial break,SameShowjt takes the value 1 if the channel continues the sameprecommercial show when the break ends. We further in-troduce an indicator Iijt, which takes the value 1 if the viewerwas watching channel j in the previous period or before thecommercial break if the then-current period is a commercialbreak. Accordingly, under such a specification, the coefficientbContinuei captures consumers’ preference for continuing towatch

the same show, if any. In comparison, if the consumer did notwatch channel j in period t − 1, bNoStarti measures the “missing-the-start-of-the-show” effect; that is, the consumer may dislikestarting viewing in the middle of a show. Finally, eijt representsidiosyncratic preference shocks and follows an i.i.d. standardnormal distribution.

There is also the outside option of not watching TV. Foridentification purposes, we normalize the mean utility level ofthe outside option to 0, and ei0t ~ Nð0, 1Þ:

ui0t = ei0t:(2)

Uncertainty and Search Cost

Let the preference shocks fei$tg be i.i.d., following a standardnormal distribution. We assume that the viewer always knowsei0t for the outside option, no matter whether he/she chose theoutside option in the previous period. Furthermore, if theconsumer starts period t with channel j, it is reasonable toassume that the viewer knows the exact level of eijt, njt, and allprogramming attributes for channel j, including InShowAdjt,BtwShowAdjt, SameShowjt, and the fixed effects of genre,hour, weekday, and week.

For channel k „ j that the viewer is not watching at thebeginning of period t, we assume that the viewer knows thegenre, hour, weekday, and week fixed effects.7 However,before search, the viewer is uncertain about eikt, nkt, andother attributes of the programming, including InShowAdkt,BtwShowAdkt, and SameShowkt. Before exploring channel k,the consumer only knows the distributions of these com-ponents. We assume that the consumer knows eikt ~ Nð0, 1Þ,nkt ~ Nðnk,s2

nÞ, and the joint distribution of programmingattributes of InShowAdkt, BtwShowAdkt, and SameShowkt.We use the observed tier-minute-specific (high-, median-, andlow-rated channels) empirical distribution of the attributes asthe joint distribution known to the viewer.8 Because of therestrictive regulation on the amount of commercials and theprolonged review process for any schedule changes, the dis-tribution is quite stable over time, so such an assumption isreasonable. In a context where the consumer does not know theattribute distributions, this model cannot be applied, and wecall for further research on the topic of consumer search andlearning the distribution during the search.

After searching the channel in period t, the viewer learns theexact levels of eikt and nkt and the programming attributes ofInShowAdkt, BtwShowAdkt, and SameShowkt for the durationof period t. To search a channel during a given period, however,

6Because the policy change can only be considered a quasi-natural ex-periment, we cannot exhaustively rule out other events that happened at thesame time and also affected viewing behavior. Controlling for the time-specific fixed effects (week) is crucial to mitigate such a concern.We also runthe same model but further control day fixed effects, and the results aresimilar. The identification assumption here is that the other factors affectingviewing behavior have the same effect across channels. Thus, they may becaptured by the time-specific fixed effects.

7One implicit assumption here is that the viewer knows the show genre ofchannel k at period t. We examine the schedules of the 29 channels. Thegenres of each hour during prime time are quite stable over time. Also, theschedule of shows is publicized well in advance, and any schedule changetakes more than 50 days to be reviewed by a government agency before goinginto effect. Thus, we consider this assumption tenable.

8More precisely, for a given tier (high-, median-, or low-rated according toOctober/November 2011 median ratings), we pool channels of the same tier.Then, for a given minute (e.g., 8:00−8:01 P.M.) before or after the ban, wecompute the proportions of channels during that minute that were broad-casting (1) an in-show commercial break, (2) a between-show commercialbreak, and (3) the same show as during the last period, out of all observationsacross channels and days. Ideally, we should evaluate the distributions aschannel-specific. However, because we only have a short time window, theobservations for one channel are too sparse to construct the distribution. Thisis a limitation of the data and, with a larger data set, one should use theempirical distributions at a more granular level.


is costly. There is a search cost Costi for each channel searched,which can be interpreted as the cognitive cost incurred due totime and effort spent on evaluating the channel.

Viewer Decisions

The decisions of a viewer include (1) whether and how tosearch alternative channels, and (2) after the search stops,given the channels searched and the outside option, whichoption to choose during period t. We assume that the viewer isfully rational when making the search and choice decisions.This assumption is necessary for model tractability.

The optimal rule for the viewer’s second decision is straight-forward: the consumer should choose the option that has thehighest utility. We therefore focus our discussion on the firstdecision.

Denote the consumer’s belief about the utility distribution ofan unsearched option k as FðuiktÞ, which depends on the dis-tributions of preference shocks fei$tg and {n$t} and the pro-gramming attributes. Because we assume that the viewer knowsthe distributions of fei$tgand {n$t} and the programming at-tributes of InShowAdkt, BtwShowAdkt, and SameShowkt (see“Uncertainty andSearchCosts”), FðuiktÞ is known to the viewer.Note that Fð$Þ is nonstandard and does not have a closed form.Accordingly, we use a simulation approach in estimation later(Step 3 in the “Estimation” section).

Let u*i be the highest utility among the then-current optionsthat have already been searched. The expected marginal gainfor searching an additional option k is (Weitzman 1979)

ð‘u*i

�uikt − u*i

�dFðuiktÞ:

The optimal decision rule for the viewer is to continuesearching as long as the expected marginal gain is greater thanthe search cost, that is,

ð‘u*i

�uikt − u*i

�dFðuiktÞ ‡ Costi:(3)

Furthermore, if multiple candidate channels have positive netreturns, the consumer should search the one with the highestlevel.

For the ease of exposition, we next introduce the concept of“reservation utility,” zikt. Let channel k be an unsearchedoption. If the reservation utility zikt = u*i , the viewer is in-different between searching k or not. That is, when zikt = u*i ,Equation 3 holds with equality.

According to the classical search literature (e.g., Weitzman1979), the optimal search strategy described above can beequivalently expressed using the reservation utility:

1. The consumer continues the search if any unsearched optionhas a reservation utility greater than the then-current maxi-mum u*i ;

2. If the search continues, the consumer should search the optionwith the highest reservation utility.

Heterogeneity

Denote the model parameters as Q0i, fnjg"j, and sn, where

Qi =hii0,bInShowAdi , bBtwShowAdi , bContinuei , bNoStarti , Costi

i0:

In other words, viewers have common fnjg"j and sn, but Qivaries across individuals.

Further define

Qi = Q + Sis, and(4)

Q =hi0, bInShowAd, bBtwShowAd, bContinue, bNoStart, Cost

i0,(5)

whereQ is thevector of themeanparameters ofQi;Si is anm×mdiagonal matrix that captures unobserved heterogeneity (m is thedimension of Q) and contains diagonal elements that followindependent standard normal distributions; and s is an m-vectorthatmeasures the relativemagnitude of unobserved heterogeneity.Together, Sis accounts for the heterogeneity distribution acrossviewers in the market. The model coefficients to be estimated are

W =hQ0,s0,

�nj�"j,sn

i0:(6)

ESTIMATION AND IDENTIFICATION

Estimation

To reiterate, for unsearched channels, we assume that theconsumer knows the fixed effects, the distribution of (ei$t),fnjg"j, and sn, and the distribution of TV programming(InShowAdkt, BtwShowAdkt, and SameShowkt). This as-sumption is consistent with the Chinese TV market, wherein(1) program schedules are fairly stable and well publicized inadvance, and (2) the frequency, duration, and scheduling ofcommercials are stable and strictly regulated by the government.

The estimation is implemented subjecting to the followingtwo criteria:

1. At the aggregate level, minimize the difference between ob-served ratings and simulated ratings according to the optimalsearch model detailed in the previous section.

2. At the disaggregate level, minimize the difference betweenobserved activities and simulated activities of search andchannel switching according to the optimal search model.

To be specific, we simulate channel ratings of period t as thefollowing:

1. Draw R = 3; 000 individual pseudoviewers and allocate themto the channels and outside option according to the ratings atthe beginning of each period observed in the data (i.e., marketshares of channels at the beginning of a given period).

2. For a given individual, draw the heterogeneity components Sifrom independent standard normal distributions.

3. Conditioned on a set of parametersW, the draws ofSi, and channelattributes, evaluate the reservation utilities zi$t of all unsearchedoptions by solving for zi$t with Equation 3 set to equality. Becausewe assume that viewers know only the distribution of pro-gramming attributes and nkt ~ Nðnk,s2

nÞ, we need to use simu-lation to evaluate the integral in Equation 3. To do so,wemake 100draws of attributes from their observed empirical distributions andnkt from Nðnk,s2

nÞ. For each set of draws, we solve for zi$t. Wethen take the average across the 100 sets of zi$t.

4. For each channel, draw the channel intercept shock njt fromNðnj,s2

nÞ.5. Determine the individual’s utility level of the initial option at thebeginning of period t. If the individual had the outside option inthe previous period, draw ei0t from standard normal distributionand use it as his/her then-current maximum utility u*. If theindividual was watching TV channel j in the previous period,calculate the mean utility level using channel j’s njt, attributeslevel in period t, the viewer’s heterogeneity draws Si, and a setof parameters ½Q0,s0�0. Further draw the preference shock eijtfrom standard normal distribution. Note that both channel j and


the outside option do not need the search. Accordingly, thegreater of uijt and ei0t is the consumer’s initial u* of period t.

6. Among all unsearched options, if the maximum reservationutility is greater than the then-current u*, the consumer searchesthat option. Draw ei$t for the just-searched option from standardnormal distribution, together with the draw of n$t from Step 4,evaluate the overall utility, and update u*.

7. Repeat Step 6 until none of the remaining reservation utilities isgreater than u*. Among the searched options, the optionwith themaximum utility u* is the final choice of the viewer in period t.

8. Because researchers do not know the realization of n$t and e:even after the viewer’s search, we repeat Steps 4–7 100 timesto integrate out the uncertainty of n$t and e:.

9. Iterate through all pseudoviewers to determine their choices inperiod t. Note that we can recycle the same random draws of njt(Step 4) for all viewers because njt is channel-specific andcommon across individuals for the same period.

From these steps, we can obtain the following:

1. at the aggregate level, the simulated ratings of option j (j = 0, 1,2, ..., J) in period t, that is, the aggregated shares of the optionschosen by the R = 3,000 individuals;

2. at the disaggregate level, among those whose initial choice is j(j = 0, 1, 2, ..., J) at the beginning of period t, the percentage ofviewers who choose to switch during period t; and

3. at the disaggregate level, the simulated average number ofsearches among those viewers who have the same initialchoice j (j = 0, 1, 2, ..., J) at the beginning of period t.

We use a minimum distance estimator to estimate the pa-rameters. Define vectors Gr, Gs, and Gn as

(7) Gr =�rjt − rjtðWÞ�"j,t

Gs =�sjt − sjtðWÞ�"j,t, and

Gn =�njt − njtðWÞ�"j,t

(8) GðWÞ =24 Gr

Gs

Gn

35,

whereW are parameters defined in Equation 6; rjt and rjt are theobserved and simulated ratings of channel j in periodt, respectively; sjt and sjt are the observed and simulatedswitching percentages of viewers whose initial choice is j at thebeginning of period t, respectively; and njt and njt are theobserved and simulated average numbers of searches forviewers whose initial choice is j at the beginning of period t,respectively. We construct the estimator to minimize the dis-tance between the observed and simulatedmeasures of interest:

W* = argminW

G0WG,(9)

where W is the sample weighting matrix.9 The 95% confidenceintervals of parameters are obtained using bootstrapping.

Identification: The Separation of Utility and Search Cost

Empirical search models face the challenge of identificationbecause it is difficult to separate preference and search cost

using field data (e.g., Sorensen 2000). To give a heuristicexample, suppose we observe that the consumer did not searchin the data. Even with the normalization of the outside optionand a known distribution for the preference shocks, it is unclearwhether the “not searching” stems from the consumer’s highsearch cost or a low expectation about the alternatives. For-mally, it is possible to vary both search cost Costi and pref-erence uikt such that the equality in Equation 3 always holdstrue, rendering the model unidentified. As a result, the ob-served ratings (from aggregate data) and switching patterns(from disaggregate data) alone are not sufficient to identifypreference and search cost.

We next discuss the identification of the searchmodel in oursetting. We are particularly interested in what assumptions anddata features are crucial for the identification ofmodel parameters.

To separate utility and search cost, we need exogenousvariations in the data that affect either utility (left-hand side ofEquation 3) or search cost (right-hand side of Equation 3), but notboth. One example of such exogenous variations is instrument orexclusion restriction variables. For example, it is reasonable toexpect time-constrained consumers to incur a higher search cost.Accordingly, when Pinna and Seiler (2017) study consumers’price search activities during grocery shopping trips, the authorsuse consumers’ walking speeds to instrument their search costs,because more time-constrained consumers tend to walk faster onaverage. Similarly, in Chen and Yao (2016), in the context ofconsumers’ hotel searches and booking, the authors use days-till-check-in-date as an exclusion variable to measure one’s timeconstraint. In both cases, the exclusion restrictions affect one’ssearch cost but not the preference.

In our current setting, however, we do not have such ex-clusion restrictions in the data. Fortunately, the government’sregulation on commercial breaks inevitably affects the utilitylevel of each channel as in-show commercials are suddenly andexogenously removed. In comparison, such a policy has littleeffect on consumers’ search cost, which is attributed to the timeand effort spent on evaluating a channel. Consequently, thepolicy change acts as an exogenous shock that helps to sepa-rate utility from search cost, making identification feasible.In particular, consider the following two equations basedon Equation 3, with Fbefore and Fafter standing for the utilitydistributions before and after the ban, respectively, and withthe reservation utility zijt plugged in:

ð‘zjt

�uijt − zijt

�dFbefore

�uijt

�= Costi, and;(10)

ð‘zjt

�uijt − zijt

�dFafter

�uijt

�= Costi:(11)

These two equations determine the search activities andconsequently the ratings of channels before and after the ban.Note that search cost Costi (right-hand side of Equations 10and 11) stays stable, while Fbefore and Fafter differ. Accordingly,the observed changes before and after the commercial ban inchannel ratings, switch patterns, and the numbers of searcheswill be attributed to the utility changes.

In our data,we observe ratings, switch patterns, and numbersof searches across channels and time and, most important,before and after the policy change. We have also made thestandard assumptions as in classical discrete-choice mod-els, including (1) the outside option, not watching TV, is

9Weuse two iterations of the estimator to obtain theweightingmatrixW.We startwith an identitymatrixas theweightingmatrixanduseEquation9 toobtain the “first-iteration” estimates W.With these estimates, we can calculate the estimated variancematrix of GðWÞ in Equation 8. The inverse of this variance matrix is then used inthe second iteration as the weighting matrix to reestimate the coefficients.


normalized to have zero mean utility; (2) preference shocksfollow a known distribution (standard normal distribution);(3) the heterogeneity in preference follows a normal dis-tribution; and (4) viewers know the distribution of pro-gramming attributes. Accordingly, because search cost andpreference can be separated as discussed, if we considersearch cost Costi as if it is an additional component in one’sutility function, the mean and heterogeneity of the co-efficients will be identified, in a fashion similar to classicaldiscrete-choice models with both aggregate and micro data(e.g., Berry, Levinsohn, and Pakes 2004).

RESULTS

Parameter Estimates

Table 5 reports the parameter estimates. Commercial breaksseverely reduce consumer utility levels. In particular, the averagemagnitude of the disutility of oneminute of in-show commercial(−4.56) is greater than the utility of watching one minute ofepisodic show (4.36), the show genre with the highest utilitylevel. Between-showcommercials cause a slightly higher drop inutility than in-show commercials, though the difference is in-significant. If the current channel continues airing the sameshow, it increases the utility level significantly, by 3.32. If theviewermisses the beginning of a show, it decreases his/her utilitylevel by 2.11. These coefficients together imply that viewersprefer to watch a show without disruption and in its entirety.

In terms of search cost, the average is relatively low (1.12).To put this value in perspective, one minute of in-show andbetween-show commercials decreases one’s utility level by4.56 and 5.30, respectively. For an average viewer, supposethat (1) the current channel starts a commercial break (in-showor between-show), (2) the current channel continues airing thesame show after the break, and (3) the viewer does not expectto miss the beginning of a show of the same genre on an

alternative channel. Other things being equal, the low searchcost implies that the viewer may start searching the alterna-tive if he/she expects it is not on a commercial break (i.e.,3:32 + 1:12 < 4:56, and 3:32 + 1:12 < 5:30).

Model Fit

Next, we discuss several tests to examine the fit of themodel. First, we consider the model’s ability to predict channelratings. We have 16 days of ratings data. There are 150 one-minute intervals per day for 29 channels and the outside option.In total,we have72,000observations of ratings (30 × 150 × 16).Out of these observations, we randomly reserve 15 intervals perday as a holdout sample, which contains 7,200 observations.Using the remaining 64,800 observations, we reestimate themodel. For the holdout sample, we then calculate the out-of-sample mean absolute percentage error (MAPE) of channelrating predictions. The MAPE is .12. In comparison, we alsoestimated a classical discrete-choice demand (logit) model withaggregate market share data in the spirit of Berry (1994). In thediscrete-choice demand model, we control for the same set ofcovariates as in the search model. The MAPE deteriorates to.19.

Then, we measure the model’s ability to predict channel-switching patterns. At the disaggregate level, for each group ofviewers who were watching channel j (j = 0, 1, ..., 29) in theprevious period, we predict the percentages of viewers whowill switch to alternative options. The out-of-sample MAPE is.20. We further consider a discrete-choice model (logit) whereeach viewer has full information about alternative channels.We estimate the logit model and control for the same set ofcovariates as in the search model. We then use the model topredict the percentages of switching viewers in the disag-gregate sample. We find that the out-of-sample MAPE of thisdiscrete-choice model is .34, much worse than in the searchmodel.

Finally, we consider the prediction of average numbers ofsearches conditional on viewers’ previous channels. Usingthe disaggregate data, the out-of-sample MAPE on thepredicted numbers of searches is .19. Since the full-information, discrete-choice model by definition assumesthat the consumers consider all available products, there is nomeaningful prediction on the numbers of searches. Overall,our model has a decent fit and is more accurate than discrete-choice models, which assume viewers have full informationabout alternative channels.

POLICY IMPLICATIONS: THE TIMING OFCOMMERCIAL BREAKS

The Lack of Commercial Timing Coordination

The timing of commercial breaks is always an importantstrategic decision of TV channels. The channels may chooseeither to synchronize or to differentiate the timing of theircommercial breaks. Sweeting (2006, 2009) uses commercialbreak data from radio stations to investigate the timing de-cisions. The author shows that the equilibrium may depend onthe characteristics of a specific market, especially how viewersswitch channels during commercial breaks. As documented byour analyses and extant literature (e.g.,Wilbur, Xu, andKempe2013), many viewers do not return to their original channelsafter the commercial breaks end, causing the original channelsto lose viewership and damaging advertising revenue poten-tials of channels. Intuitively, if all channels have commercials

Table 5ESTIMATES FOR THE SEARCH MODEL

Estimate Heterogeneity

UtilityIn-show commercial (minute) −4.56 (−5.19, −3.55) .99 (.17, 1.99)Between-show commercial

(minute)−5.30 (−5.99, −4.14) .97 (.14, 1.88)

Current channel same show 3.32 (2.25, 4.79) 1.04 (.24, 1.93)Alternative channel same

show−2.11 (−3.19, −.16) 1.37 (.49, 3.16)

Episodic TV series 4.36 (1.03, 7.04) 2.97 (.79, 3.99)Sports events 3.19 (2.10, 5.66) 2.22 (.51, 4.14)Medical and health 1.97 (.96, 4.85) 4.97 (2.71, 7.31)News −1.82 (−4.17, −.91) 3.24 (1.08, 5.17)Other types of shows −.51 (−1.23, .10) 4.11 (2.59, 5.15)Channel-time “quality” (njt)

Channel fixed effects,mean (nj)

Yes —

Standard deviation (sn) 1.19 (.61, 2.16) —

Weekday fixed effects Yes YesWeek fixed effects Yes YesHour fixed effects Yes Yes

Search cost 1.12 (.17, 2.09) 3.62 (2.47, 5.00)

Notes: Boldface indicates that the estimate is significant at the 95%confidence level. Values in parentheses denote 95% confidence intervals.


at the same time, viewers have less incentive to switchchannels during the breaks because they will also find com-mercials on alternative channels. However, such a coordi-nation is delicate. A channel may have an incentive to deviatefrom the coordinated timing because it may gain more viewersby not broadcasting commercials when all its competitors areairing commercials.

Pertaining to the observed timing of commercial breaks inour data, TV channels had little coordination, even thoughthere is no explicit regulation against such a practice acrosschannels (see Chinese State Council [1997], the guideline forall regulations pertaining to broadcasting and television). Toillustrate this, we focus on in-show commercials before theregulation change.10 We introduce a time-channel-specificdummy variable, which reflects whether a channel wasbroadcasting in-show commercials during a one-minute in-terval. The dummy variable takes the value 1 if a given channelwas showing in-show commercials during a specific intervaland 0 otherwise. For every pair of channels, we then calculatethe correlation of the dummy variables across time. If thechannels usually had the same timing for commercial breaks(i.e., starting commercials during the same minute), the cor-relation coefficients of the dummy variables should be closedto 1. Figure 2 depicts the histograms of these correlation co-efficients. The majority of the correlation coefficients are verydistant from 1. The mean correlation is only .04 and statisticallyinsignificant, and the median is in fact negative.

It is possible that at a given time, a channel’s decision oncoordination depends on how its rating compares to its com-petitors’. To check this possibility, for each in-show commercialbreak of a given channel, we record its precommercial ratingand its competing 28 channels’ ratings before the focal com-mercial break. We divide the competitors into three break-specific groups: (1) similar-rated, a channel whose rating waswithin – .07 of the focal channel’s (i.e., – 5% SD of ratingsbefore the ban; we also considered – 10% SD, and the in-sights remain the same); (2) higher-rated, a channel whoserating was at least .07 higher than the focal channel’s; and (3)lower-rated, a channel whose rating was at least .07 lowerthan the focal channel’s. Note that the grouping is not fixedover time because of the variation in ratings across breaks.Next, for the focal channel and each of its competitors in agiven tier, we repeat the same correlation calculation for thedummy variables across time. Figure 3 shows three histo-grams of the correlation coefficients by pooling all channelswithin the same tier, as just defined. Accordingly, the threehistograms correspond to a focal channel’s timing correlationwith competing channels that have lower, similar, or higherratings, respectively. As we can see, the patterns are essentiallythe same as in Figure 2, which shows little coordination intiming across groups. In Appendices C–D, we further showanother set of histograms focusing on adjacent minutes insteadof the same minute (i.e., one minute before or after the focalchannel’s commercial break). The insights remain the same. Inconclusion,wefind little evidence that channelswere coordinatingtheir commercial breaks at the time of our data window.

The Effect of Commercial Timing Coordination on Ratings

Having established that channels had little coordination intheir commercial timing, we are interested in how a full co-ordination of timingmight affect viewers’ behavior and ratingsacross channels for both regular shows and commercial breaks.We hereby consider the following policy simulation on in-show commercial timing coordination. Drawing on observedrating data before the ban, we define a daily top channel as theone with the highest average rating during a given day. Whenthe remaining nontop channels of that day broadcast showsthat overlapped in time with shows of the top channel, weadjust the timing of the nontop channels’ in-show breakssuch that their commercials are all synchronized with the topchannel’s. (Note that it is nearly impossible for the channels tohave no overlapping shows in their programming, andwe haveno such observations in our data.) More precisely, in the data:

1. During the overlapped portion, both the top channel and anontop channel had breaks, and the numbers of breaks were thesame. In this case, we align the nontop channel’s breaks suchthat they start at the same time as the top channel’s breaks.

2. During the overlapped portion, both channels had breaks, butthe numbers of breaks were different. In this case, we firstcombine or divide the nontop channel’s breaks so that (a) itsbreaks all have the same length and (b) the number of breaksbecomes the same as the top channel (i.e., the length of eachbreak may change from its original observed level). Next, wealign these new breaks of the nontop channel such that theystart at the same time as those of the top channel.

3. During the overlapped portion, the nontop channel had breaks,but the top channel had none. In this case, we relocate thenontop channel’s breaks randomly to its nonoverlapped por-tion of programming as in-show breaks.

4. During the overlapped portion, the nontop channel had nobreaks, but the top channel had breaks. In this case, we addnew in-show breaks to the nontop channel and align their startsto the top channel’s breaks.

Next, on the basis of our model and estimates, we simulatetwo sets of ratings of all channels for the duration of the topchannel’s regular shows, including the ratings of the com-mercial breaks. The first set of ratings is simulated accordingto the observed commercial timing without the adjustmentspreviously mentioned. The second set is based on the timingafter the adjustments. Note that as a consequence of thoseadjustments, for the second set of simulated ratings, all

Figure 2COORDINATION OF COMMERCIAL BREAKS

0

1

2

3

4

–.5 .0 .5 1.0Before the Commercial Ban: Correlation of Commercials Across Channels

Den

sity

10After the regulation change, episodic programs no longer had in-showcommercial breaks. Also, the insights are similar to those detailed later whenwe analyze using in-show commercials of nonepisodic programs after theregulation change. For between-show commercials, the coordination is in-herently difficult because the shows often end at different time slots.


channels have the same in-show commercial starting time andfrequency. We also assume that viewers know the updateddistributions of in-show commercials. Next, we inspect theeffect of timing coordination on ratings.

Effect on average ratings.We first compare the two sets ofaverage ratings for the whole data period, including bothregular shows and commercials. We find that with synchro-nized commercial timing, the average ratings of high-ratedchannels drop, while the average ratings of low-rated channelsrise. To be specific, we again use the October/November 2011median rating of each channel to divide them into high-rated(10 channels), median-rated (10), and low-rated (9) groups.We then compare the average ratings between the two sim-ulations (i.e., differentiated timing vs. synchronized timing).When the break timing is changed from differentiated tosynchronized, the average rating of high-rated channelsacross channels and intervals drops from 1.46 to 1.42, a2.7% decrease. In particular, for the average rating of eachhigh-rated channel across time, all have lower values in the

synchronized condition, and seven channels’ ratings dropsignificantly at the 95% confidence level. For low-ratedchannels, in contrast, the average across channels and timeincreases by 6.4%, from .29 to .31. The average ratingsacross time for all nine channels become higher, and foreight of the nine, the changes are significant at the 95%level. For the median-rated channels, the changes are mixed.Four of the ten channels have an increase in average ratings,but none is significant; the remaining six have decreasedratings, and one of them drops significantly at the 95% level.Overall, the average rating for median-rated channels acrosschannels and time becomes slightly lower (−.008), but thechange is statistically insignificant.

To understand the results from this simulation, note thatviewers are more likely to search other channels duringthe commercial breaks than during a show, as shown in thediscussion of data patterns in the “Data” section (and thesimulation). Intuitively, for viewers of low-rated channels,compared with viewers of high-rated channels, such searcheshave a higher likelihood of finding alternatives with higherutility levels. So, compared with high-rated channel viewers,who are more likely to return to their original channels after thesearch or after the break, viewers of low-rated channels are lesslikely to do so (as we showed earlier in the “Evidence ofConsumer Search” section, where low-rated channels take agreater hit in ratings after a commercial break than high-ratedchannels). But when the commercial breaks are coordinated,viewers are less likely to search overall, because the alternativechannels are showing commercials at the same time. As a result,coordination enhances the average rating of low-rated chan-nels. At the same time, the lower likelihood of search impliesthat high-rated channels gain fewer viewers who might haveswitched from low-rated channels, in comparison with thesituation where commercial timing is differentiated. Accord-ingly, the average rating of high-rated channels deteriorates dueto the coordination. Finally, for median-rated channels, ratingsface both the upward and downward forces when commercialtiming is coordinated, and the net result is uncertain. From amanagerial perspective, this simulation implies that low-ratedchannels should try to synchronize their commercial breakswith high-rated channels, and high-rated channels should try todifferentiate their breaks from competing channels. Further-more, while it is beyond the scope of the current study, the effectof channel mergers on the timing of commercial breaks isanother area with rich opportunity for exploration.

Effect on commercial ratings.Wenext focus on the averageratings of commercial breaks.We find that under synchronizedcommercial timing, all three tiers of channels (high-rated,median-rated, and low-rated) achieve higher average ratingsfor their commercials. We use a similar approach as in theprevious section to divide channels into three tiers. When thecommercial timing is synchronized, the average rating ofcommercial breaks on high-rated channels increases from 1.07to 1.16, an 8.4% increase. For all of the ten high-rated channels,the commercial ratings increase under the synchronized timing,and seven channels have significant increases at the 95% con-fidence level. For the low-rated channels, the average commercialratings across channels and time increases by 12.5%, from .24 to.27. Each low-rated channel sees an increase in rating that issignificant at the 95% confidence level. For the median-ratedchannels, the situation is similar: the average commercialrating for median-rated channels increases from .35 to .38

Figure 3COORDINATIONOFCOMMERCIAL BREAKSACROSSDIFFERENT

RATING GROUPS

0

1

2

3

.0 .5 1.0Correlation of Commercials with Lower-Rated Channels

Den

sity

0

1

2

3

4

.0 .5Correlation of Commercials with Similar-Rated Channels

Den

sity

0

1

2

3

.0 .5 1.0Correlation of Commercials with Higher-Rated Channels

Den

sity


under synchronized commercial timing. All median-ratedchannels achieve higher commercial ratings, and for nineout of the ten channels, the improvement is significant com-pared with the nonsynchronized case.

At first sight, this may seem surprising because the earlierresults pertaining to overall average ratings show that undersynchronized commercial timing, low-rated channels have higheraverage ratings, but the ratings of high-rated channels drop.However, note that when the commercial timing is synchronized,viewers search less during commercial breaks for all channels. Inother words, independent of the rating tier of a channel, theviewer is more likely to sit through the commercial, rather thansearching alternatives, when commercials are synchronized.

In terms of the ad revenues for the different channel tiers,however, the implication is more complicated. The industrystandard to calculate the price paid by advertisers is based on theaverage rating of the show duringwhich the commercial is aired,instead of on the rating of the commercial break. Under dif-ferentiated commercial timing, the ratings of the regular showand the commercial are highly correlated. Under synchronizedcommercial timing, in contrast, the correlation is diminished oreven reversed, as shown in the exercise in the previous section.As a result, the low-rated channels benefit from the synchronizedtiming of commercials, but high-rated channels are at a loss. Inshort, we again reach the same insight: high-rated channelsshould try to differentiate the timing of commercials from theircompetitors, while low-rated ones should do the reverse.

CONCLUSION

It is crucial for TV channels to understand how commercialbreaks affect their viewership and how they can strategicallymanage the timing of breaks. Our article proposes a sequentialsearch model and uses it to measure consumers’ choices of TVchannels. The model provides a more realistic characterizationof TV-viewing behavior. We advance the TV-viewing choiceliterature by allowing uncertainty in consumers’ knowledgeabout upcoming programming. In our model, a consumer needsto search alternative channels to resolve the uncertainty inprogramming. The search is costly. We show how commercialbreaks affect the trade-off between search cost and expectedreturns and ultimately influence how viewers behave.

We apply the model to a unique data set of the Chinese TVmarket. In our data, the market experienced an exogenouspolicy change whereby the government banned all in-showcommercial breaks for episodic programs. The identification ofclassical empirical searchmodels is difficult because consumerpreference and search cost are often confounded in field data.The regulation change in our data functions as an exogenousshock that affects the preference distribution independent ofsearch cost, which makes identification feasible. Drawingon the model estimates, we also present managerial insightsfor TV channels regarding the timing of breaks. Low-ratedchannels will benefit from synchronizing their breaks withhigh-rated competitors to prevent their own viewers fromsearching and leaving after the search during commercialbreaks. High-rated channels should try to differentiate theirbreaks from competitors. When a high-rated channel differ-entiates its timing from competitors, it can capture churningviewers from competing channels during their breaks.Meanwhile, when the high-rated channel has commercialbreaks itself, its viewers are on average less likely to search orswitch because of lower expected net return of the search.

Breaks between consumptions appear in other contexts, forexample, commercial breaks in radio shows and time lapsesfor video game sequels. Accordingly, the results of strategictiming decisions of TV commercials may be applied morebroadly.

Television remains the preferred advertising medium formany firms. However, with the TV industry facing increas-ingly intense competition from new media channels and theadvance of technology that enables consumers to skip com-mercials, it is imperative that the TV industry, advertisers, andpolicy makers obtain better insights about how viewers re-spond to commercial breaks and hence strategically determinethe timing of breaks. We hope that our study can enrich theunderstanding of viewer behavior and encourage future studiesin this area.

REFERENCES

Berry, Steven, James Levinsohn, and Ariel Pakes (2004), “Dif-ferentiated Products Demand Systems from a Combination ofMicro and Macro Data: The New Car Market,” Journal of Po-litical Economy, 112 (1), 68–105.

Berry, Steven T. (1994), “Estimating Discrete-Choice Modelsof Product Differentiation,” RAND Journal of Economics, 25 (2),242–62.

Byzalov, Dmitri, and Ron Shachar (2004), “The Risk ReductionRole of Advertising,” Quantitative Marketing and Economics,2 (4), 283–320.

Chen, Yuxin, and Song Yao (2016), “Sequential Search with Re-finement: Model and Application with Click-Stream Data,”Management Science (published electronically September 28),DOI: 10.1287/mnsc.2016.2557.

Chinese State Council (1997), “Decree No. 228: Regulations onBroadcasting and Television Administration,” http://www.pkulaw.cn/fulltext_form.aspx?Gid=18601(in Chinese).

De Los Santos, Babur I., Ali Hortacsu, and Matthijs R. Wildenbees(2012), “Testing Models of Consumer Search Using Data onWeb Browsing and Purchasing Behavior,” American EconomicReview, 102 (6), 2955–80.

Deng, Yiting, and Carl F. Mela (2017), “TV Viewing and Adver-tising Targeting,” Journal of Marketing Research, forthcoming,DOI: 10.1509/jmr.15.0421.

Elliot, Stuart (2004), “A Survey of Consumer Attitudes Reveals theDepth of the Challenge That the Agencies Face,” The New York

Appendix AGOOGLE TRENDS INDEX OF “TV SHOW COMMERCIALS”

0

20

40

60

80

100

120

Jan-

11Feb

-11

Mar

-11

Apr-1

1M

ay-1

1Ju

n-11

Jul-1

1

May

-12

Jul-1

2

Aug-1

1Sep

-11

Oct-11

Nov-1

1Dec

-11

Jan-

12Feb

-12

Mar

-12

Apr-1

2Ju

n-12

Aug-1

2Sep

-12

Oct-12

Nov-1

2Dec

-12


Times (April 14), http://www.nytimes.com/2004/04/14/business/media-business-advertising-survey-consumer-attitudes-reveals-depth-challenge.html.

Goettler, Ronald L., and Ron Shachar (2001), “Spatial Competitionin the Network Television Industry,” RAND Journal of Eco-nomics, 32 (4), 624–56.

Hong, Han, and Matthew Shum (2006), “Using Price Distributionsto Estimate Search Costs,” RAND Journal of Economics, 37 (2),257–76.

Honka, Elisabeth (2014), “Quantifying Search and Switching Costsin the U.S. Auto Insurance Industry,” RAND Journal of Eco-nomics, 45 (4), 847–84.

Hortacsu, Ali, and Chad Syverson (2004), “Product Differentiation,Search Costs, and Competition in the Mutual Fund Industry: ACase Study of S&P 500 Index Funds,” Quarterly Journal ofEconomics, 119 (2), 403–56.

Koulayev, Sergei (2014), “Search for Differentiated Products: Iden-tification and Estimation,” RAND Journal of Economics, 45 (3),553–75.

Lehmann, Donald R. (1971), “Television Show Preference: Ap-plication of a Choice Model,” Journal of Marketing Research,8 (1), 47–55.

Moshkin, Nickolay V., and Ron Shachar (2002), “The AsymmetricInformation Model of State Dependence,” Marketing Science,21 (4), 435–54.

National Bureau of Radio, Movie, and Television (2009), “Regu-lation 61: Measures for the Broadcasting of Radio and TelevisionAdvertisements,” http://www.sarft.gov.cn/art/2009/9/10/art_1583_26310.html(in Chinese).

Pinna, Fabio, and Stephan Seiler (2017), “Estimating Search Ben-efits from Path-Tracking Data: Measurement and Determinants,”Marketing Science, 36 (4), 565–89.

Rust, Roland T., and Mark I. Alpert (1984), “An Audience FlowModel of Television Viewing Choice,”Marketing Science, 3 (2),113–24.

Shachar, Ron, and John W. Emerson (2000), “Cast Demographics,Unobserved Segments, and Heterogeneous Switching Costs in aTelevision Viewing Choice Model,” Journal of Marketing Re-search, 37 (2), 173–86.

Sorensen, Alan T. (2000), “Equilibrium Price Dispersion in RetailMarkets for Prescription Drugs,” Journal of Political Economy,108 (4), 833–50.

State Agency of Radio, Film and Television of the People’s Re-public of China (2010), “Regulation 63: TVContentManagement

Appendix BDESCRIPTIVE STATISTICS: PERCENTAGES OF COMMERCIAL TIME AMONG TOTAL BROADCASTING TIME ACROSS CHANNELS

AND HOURS

M SD 5th Percentile 95th Percentile Max

In-Show Commercial Percentage

Before the Ban: Episodic ShowsChannels 1–10 1.31 2.67 0 6.67 16.67Channels 11–20 1.34 2.09 0 6.67 11.67Channels 21–29 1.60 4.37 0 8.33 20.00

After the Ban: Episodic ShowsChannels 1–10 0 0 0 0 0Channels 11–20 0 0 0 0 0Channels 21–29 0 0 0 0 0

Before the Ban: Nonepisodic ShowsChannels 1–10 3.59 5.11 0 16.67 18.33Channels 11–20 4.34 6.13 0 18.33 21.67Channels 21–29 4.86 6.58 0 18.33 23.33

After the Ban: Nonepisodic ShowsChannels 1–10 2.81 4.08 0 11.67 16.67Channels 11–20 2.43 3.67 0 10.00 15.00Channels 21–29 4.13 5.74 0 15.00 28.33

Between-Show Commercial Percentage

Before the Ban: Episodic ShowsChannels 1–10 4.32 7.65 0 16.67 23.33Channels 11–20 6.67 9.38 0 21.67 36.67Channels 21–29 1.51 4.41 0 13.33 18.33

After the Ban: Episodic ShowsChannels 1–10 5.03 8.48 0 21.67 25.00Channels 11–20 6.81 9.14 0 23.33 35.00Channels 21–29 2.38 5.92 0 20.00 25.00

Before the Ban: Nonepisodic ShowsChannels 1–10 2.51 4.91 0 11.67 26.67Channels 11–20 4.34 6.13 0 18.33 21.67Channels 21–29 4.49 5.18 0 13.33 18.33

After the Ban: Nonepisodic ShowsChannels 1–10 2.16 4.26 0 11.67 21.67Channels 11–20 2.97 4.69 0 11.67 20.00Channels 21–29 4.06 4.54 0 13.33 21.67


Regulations,” (May 14), http://www.gov.cn/flfg/2010-05/20/content_1609751.htm (in Chinese).

State Agency of Radio, Film and Television of the People’s Re-public of China (2011), “Regulation 66: Supplementary Pro-visions on the Measures for the Administration of BroadcastingAdvertisements,” (November 25), http://www.gov.cn/flfg/2011-11/28/content_2005138.htm (in Chinese).

Sweeting, Andrew (2006), “Coordination, Differentiation, and theTiming of Radio Commercials,” Journal of Economics & Man-agement Strategy, 15 (4), 909–42.

Sweeting, Andrew (2009), “The Strategic Timing Incentives of Com-mercial Radio Stations: An Empirical Analysis Using MultipleEquilibria,” RAND Journal of Economics, 40 (4), 710–42.

Weitzman, Martin L. (1979), “Optimal Search for the Best Alter-native,” Econometrica, 47 (3), 641–54.

Wilbur,KennethC. (2008),“ATwo-Sided,EmpiricalModelofTelevisionAdvertising andViewingMarkets,”Marketing Science, 27 (3), 356–78.

Wilbur, Kenneth C., Linli Xu, and David Kempe (2013), “Cor-recting Audience Externalities in Television Advertising,” Mar-keting Science, 32 (6), 892–912.

Yang, Sha, Yi Zhao, Tulin Erdem, and Ying Zhao (2010), “Mod-eling the Intrahousehold Behavioral Interaction,” Journal of Mar-keting Research, 47 (3), 470–84.

Yeh, Johnson, and Ming Zhang (2013), “Taking the Pulse ofChina’s Ad Spending,” McKinsey Quarterly (June), http://www.mckinsey.com/industries/media-and-entertainment/our-insights/taking-the-pulse-of-chinas-ad-spending.

Appendix CCOORDINATION OF BREAKS ACROSS DIFFERENT RATING

GROUPS: BEFORE THE FOCAL BREAK

0

1

2

3

.0 .5Correlation of Commercials with Lower-Rated Channels: Before Focal Channel's Break

Den

sity

0

1

2

3

–.3 .0 .3 .6 .9Correlation of Commercials with Simiar-Rated Channels: Before Focal Channel's Break

Den

sity

0

1

2

–.3 .0 .3 .6 .9Correlation of Commercials with Higher-Rated Channels: Before Focal Channel's Break

Den

sity

Notes: In this figure, we use a similar approach as used in the section onpolicy implications to create correlation coefficients of commercial timing.However, instead of focusing on the commercial breaks in the same minuteacross channels, we focus on the one minute before the focal channel starts itscommercial break

Appendix DCOORDINATION OF BREAKS ACROSS DIFFERENT RATING

GROUPS: AFTER THE FOCAL BREAK

0

1

2

3

–.3 .0 .3 .6 .9Correlation of Commercials with Lower-Rated Channels: : After Focal Channel's Break

Den

sity

0

1

2

3

–.3 .0 .3 .6 .9Correlation of Commercials with Similar-Rated Channels: After Focal Channel's Break

Den

sity

0

1

2

3

–.3 .0 .3 .6 .9Correlation of Commercials with Higher-Rated Channels: After Focal Channel's Break

Den

sity

Notes: In this figure, we use a similar approach as used in the section onpolicy implications to create correlation coefficients of commercial timing.However, instead of focusing on the commercial breaks in the same minuteacross channels, we focus on the one minute after the focal channel starts itscommercial break


Copyright of Journal of Marketing Research (JMR) is the property of American MarketingAssociation and its content may not be copied or emailed to multiple sites or posted to alistserv without the copyright holder's express written permission. However, users may print,download, or email articles for individual use.

Documents

TV Channel Search and Commercial Breaks - GitHub Pages · The data patterns support the idea that viewers search for ... timing of breaks affects TV channels ... Television is still