ABSTRACT
Advertisers prefer second-by-second measurements of advertisements over program ratings, but collecting individual viewing data that are accurate to the unit of one second is very difficult and expensive. Under the condition of no additional cost or effort investment, the authors developed a methodology for converting minute-by-minute people-meter data into second-by-second audience ratings. This methodology is based on the successful modeling of television viewers' tuning-in behavior by a uniform distribution and tuning-out behavior during commercials by a beta distribution. The methodology could be applied to measure advertising effectiveness, assess advertising strategies, and aid in future media purchasing and pricing.
MANAGEMENT SLANT
Program-audience may not serve as an adequate indicator of commercial viewership.
Measurement of advertising effectiveness will be more accurate if researchers are able to collect second-by-second data.
Simulations of second-by-second audience numbers can be helpful in measuring advertising audience, assessing advertising strategies, and aiding in future media purchasing and pricing.
Together with the prediction of audience program ratings, the proposed method can predict advertisement audience ratings when advertisers have chosen the time period and program type to broadcast their ads.
INTRODUCTION
Television remains the dominant advertising medium, despite the considerable increase in the number and types of media. As of 2016, television advertising accounted for roughly 40 percent of global advertising spending, and it will remain the most significant channel for advertising in the near future (Yeh and Zhang, 2017). Audience ratings of television programs long have been used as an important reference for measuring advertising effectiveness and for placing and pricing advertisements (Yang, Narayan, and Assael, 2006; Yao, Wang, and Chen, 2017). The serious gap between program and commercial audience ratings due to changes in viewing patterns, however, frequently renders such references spurious (Kent, 2002; Swaminathan and Kent, 2013).
Several studies have identified this problem; an experimental ratings system estimated that advertisements lose a combined 17 percent of program audience over all time slots (Kneale, 1988). Another study found a 21.5 percent loss in commercial audience ratings from the people-meter data of the Netherlands (Van Meurs, 1998). Some studies have shown that zapping—switching to other channels—often occurs during commercial breaks (Danaher, 1995; Rohas-Mendez and Davies, 2005; Siddarth and Chattopadhyay, 1998; Tse and Lee, 2001). Two studies thus concluded that the program audience is not critically important at all for measuring advertisement effectiveness (Kent, 2002; Ross, 1999). Program audience measures may not serve as an adequate indicator of commercial viewership (Schweidel and Kent, 2010).
Advertisers prefer audience ratings for commercials rather than for programs (Schweidel and Kent, 2010). The measurement of advertisement effectiveness is more accurate if second-by-second data can be collected. Although the set-top box used for cable television can supply data in the unit of seconds, the observed unit is a household, not an individual. This is a very serious drawback as far as measuring the effectiveness of a television advertisement is concerned, because individual behavior rather than household behavior is the primary valuable information in most marketing and economic studies (Kent, 2002; Lindorff, 2000).
Another disappointing weakness of set-top-box data is that the boxes count a household as a commercial viewer when the television is on and the corresponding channel is selected; whether anyone is watching, or even present in the room, is unknown (Schweidel and Kent, 2010). People meters installed in collaborative households are set more often to collect information in one-minute intervals because of data-transmission and data-storage requirements. Such storage requirements are increasing and are more expensive, particularly as a growing number of channels emerge. Data-collection companies unlikely will spend significantly greater sums of money in extending people-meter storage to collect second-by-second data (Atkinson, 2008). Collecting such data in one-second intervals using people meters is a waste of resources, because television advertisements occupy only a small portion of broadcast time, and no one, including advertisers, is concerned with the second-by-second fluctuation in viewership of the program itself.
It is thus important to develop a methodology to convert minute-by-minute data into second-by-second data for commercials. To the best of the authors' knowledge, however, almost none of the extant literature to date has resolved this problem adequately. The current study presents the first attempt to fill this gap by developing an innovative methodology of modeling tuning-in (arrival) and tuning-out (departure) viewing behaviors and then simulating second-by-second commercial audience from minute-by-minute data.
HYPOTHESIS DEVELOPMENT
To develop the methodology, the authors proposed two important hypotheses on the statistical modeling of tuning-in (arrival) and tuning-out (departure) viewing behavior during commercial periods. Many casual viewers who are not watching a particular channel may turn on the television and select a channel, or they may switch from other channels to a particular channel randomly. Viewers' arrival time to a certain channel may follow a uniform distribution. This pattern remains unchanged within any particular minute. The authors thus hypothesized the following:
H1: The accurate arrival time follows a uniform distribution over each one-minute interval.
As soon as commercials start, however, because of the commercial-avoidance effect, many individuals leave the room immediately, switch to another channel, or even turn off their television (Tse and Lee, 2001; Wilbur, 2008a; Woltman, Wedel, and Pieters, 2003). This effect gradually subsides toward the end of the commercials. When commercials are over, the program resumes, and the departure behavior returns to the normal program time. The authors thus hypothesized the following:
H2: The departure time follows a decay function over the whole commercial-break period.
The authors tested the hypotheses and validated the use of these distributions with people-meter data obtained from local television stations. With the above modeling, the authors could obtain the second-by-second audience ratings on the basis of the minute-by-minute people-meter data. This methodology can be used elsewhere by others, because the model formulae involved in the hypotheses were derived theoretically and the application and effective use of the methodology are demonstrated by real-life data from different programs and regions. This methodology can be applied to all individuals of collaborative households in the representative sample and generalized to the whole population.
Because this methodology uses only currently available people-meter data as the basis for estimation, it not only requires no additional cost or effort in data collection but also retains most of the advantages of people-meter data. These advantages include the following:
Data are collected in the unit of individuals.
Data coverage is 24 hours a day, seven days a week.
The number of viewers of any specific advertisement therefore can be estimated readily.
One concern is whether individuals push their buttons correctly when using the people meter. A “coincidental survey” answered this question by checking whether the people-meter data are consistent with actual individual behavior (Danaher and Beed, 1993). The results revealed that 92 percent of the panelists pushed their buttons correctly. The people meter thus is believed to record reliably and precisely all channel-switching behavior. The people-meter data used in this study are from Hong Kong TV Broadcasts Ltd. (TVB) and Jiangsu TV in mainland China, and they cover the whole of 2014.
The rest of the article is structured as follows. The authors first introduce the people-meter data in detail. They then illustrate their methodology, including the empirical observation and model development, followed by model calibration and validation. Finally, they present possible applications of the methodology and provide directions for future research.
PEOPLE-METER DATA ANALYSIS
The Hong Kong domestic free tele vision program services are very pervasive and influential, with a penetration rate of approximately 100 percent. Current license holders of domestic free television services are TVB and Asia TV Ltd. (ATV). TVB operates one Chinese JADE channel (TVBJ) and one English WORLD channel, and ATV operates one Chinese HOME Channel (ATVH) and one English WORLD Channel. TVB is a prominent provider of free-to-air television services for Hong Kong's television industry, because its TVBJ maintains a remarkable average of 70 to 80 percent of Hong Kong's television-audience share.
The data used in this study were provided by TVB and cover the whole of 2014. To validate the current study's methodology sufficiently, the authors also used the viewing data of Jiangsu TV in mainland China for 2014. Jiangsu TV, hosted by Jiangsu Province, is one of the most popular television stations in the mainland.
The data are referred to as “people-meter data” because they were collected by people meters installed in 650 collaborating households. The people meter combines the information from a device that records precisely what the television is doing and a remote control that records what each person in a household is doing in his or her viewing. If household member A, for example, begins to watch the television, he should log in by pushing the “A” button on the people meter's remote control. If he leaves the television room during commercials, he should log out by pushing the button again. The people meter thus records precisely when people begin to watch television, when they leave the room, and all channel-switching behavior.
People-meter data therefore integrate the records of people's television-watching behavior and the detailed information of television programs. The output of people-meter data contains the individual viewer ID, which channel he or she is watching, and the specific time the individual started (tuned in) and ended (tuned out) watching a program for each viewer.
After data mining, the authors obtained the number of arrivals and departures in each minute. The people-meter data, however, are accurate only to one minute, because the people meter is calibrated to record any change in set status that lasts for at least one minute. If household member A switches from channel 1 to channel 2 for at least one minute, the change is recorded. If he or she tunes to channel 2 from channel 1 and then back to channel 1 within one minute, this channel switch is not recorded. The released data thus are accurate to one minute, indicating the variation of audience number minute by minute.
At present, many countries and areas, such as Hong Kong and mainland China, still use people meters to measure audience ratings. Television stations, advertisers, and agencies are supplied with the minute-by-minute audience ratings as references to assess the advertisement effectiveness.
The advertisements, however, are always very short, normally lasting just 15 seconds. The minute-level data cannot tell the detailed audience ratings of each advertisement.
That a large part of the audience is lost during commercials has been disclosed by existing studies (Schweidel and Kent, 2010; Wilbur, 2008b), so advertisers and agencies prefer the second-by-second audience ratings of their advertisements, not the minute-by-minute program ratings (Steinberg and Hampp, 2007). This article intends to resolve the problem of inconsistency between supply and demand by translating the obtained minute-by-minute data into the desired second-by-second data.
Television prime time, covering from 19:00 to 23:00, is called “gold time” in China. New dramas are broadcast during this time period, and the vast majority of the advertising revenue is devoted to this time. The authors restricted the current study to this time period, in keeping with most previous studies (Danaher et al., 2011).
Television stations may embed commercials into one episode (in-show commercials) or broadcast commercials between two episodes of one program or between two different programs (between-shows commercials). In Hong Kong, there are three in-show commercial pods in each episode and one between-shows commercial pod. In mainland China, however, authorities have forbidden all in-show commercials for episodic television series since 2012, to improve audience satisfaction.
To make contrast analyses and eliminate particularity, the authors developed and validated their methodology in different advertising situations.
METHODOLOGY
The study's method of translating the minute-by-minute audience ratings into second-by-second audience ratings is easy to understand. The authors chose suitable distributions to model switching behaviors during the television-commercial-watching process. In particular, they fitted the arrival (tuning-in) and departure (tuning-out) behaviors of a channel.
The authors first observed the viewing patterns of individuals from the people-meter data and made two hypotheses about the distribution models of arrival and departure behaviors. They next calibrated and validated these hypotheses in terms of sample data and applied the distribution models to simulate the number of arrivals or departures in each second. Finally, they aggregated the number of simulated arrivals, departures, and viewers who stayed on the channel as the total audience in one second. The central components were model development and validation, which are explained in detail in the following subsections on the basis of TVB data.
Empirical Observation
The sample base was 650 collaborating households, for at least 1,950 individual viewers. The authors chose eight popular dramas broadcast during gold time to analyze: four from TVBJ, and the other four from ATVH. There were 263 episodes and 1,052 commercial pods of the eight dramas. (See Table 1 for summary statistics, e.g., the names, broadcast periods, and total commercial pods of the dramas.)
The authors observed the arrival and departure patterns in each episode of the eight dramas, particularly the commercial pods of each episode. There was little fluctuation in the number of tuning-in occurrences during the whole episode, and the number of tuning-out occurrences varied little during program broadcast time as well. There was considerable variation in the number of tuning-out occurrences when commercials began, however. Another study found, similarly, that in each episode, the first three in-show commercial pods had similar tuning-in and tuning-out patterns, but there was a significantly different tuning-out pattern in the fourth pod (Zigmond, Dorai-Raj, Interian, and Naverniouk, 2009). During the fourth pod, a large number of viewers tuned out, many more than during the first three pods embedded in the episode.
Arrival behavior differs between regular viewers and casual viewers (Lu and Lo, 2009). Regular viewers watch a drama from the very beginning of an episode, because they are familiar with the program schedule. Casual viewers arrive at a channel randomly, because they do not know what program is being shown when they turn to a particular channel.
Casual viewers tune in to a channel for several reasons:
They hear from others that the drama broadcast on this channel is interesting.
They need to rest or experience entertainment after a busy workday or night.
They tune in to this channel to avoid advertisements on other channels.
Casual viewers therefore lead to the variation in the number of viewers tuning in.
If an individual is watching a program on a channel, however, he or she unlikely will switch, except to engage in activities perceived as more important. The number of tuning-out occurrences from a given channel during a program in each minute therefore varies little. When commercials begin, however, the number of viewers tuning out of this channel increases suddenly, because of the commercial-zapping effect. The effect of this phenomenon gradually subsides during the commercial broadcast period. When the commercials end, the normal tuning-out pattern resumes. After an episode of a drama concludes, many more viewers depart in the following pod.
The authors chose a typical episode to show its arrival and departure pattern, an episode from the drama “Beyond the Realm of Conscience,” broadcast on TVBJ in 2014. The authors defined an individual as a viewer in the sample, using the criterion that this individual had watched at least five minutes of this drama. (See Figure 1 for the percentage of the number of viewers tuning in and tuning out of this channel in each minute, during the time period between 20:45 and 21:33 on October 20, 2014.)
The percentage of the number of arrival viewers in the whole episode varied little; there was no specific pattern for tuning in. The percentage of the number of departure occurrences from a given channel in each minute during the program periods also fluctuated little, and it was even more stable than tuning in before and after commercial pods. This observation prompted Hypothesis 1, as stated earlier.
At each pod start time, however, the number of viewers tuning out suddenly became very large; this phenomenon gradually subsided as the advertisements were broadcast. When this pod ended, the normal tuning-out pattern during the program resumed. A decay function, which may be described by a beta, exponential, or gamma distribution over the whole commercial-break period, therefore could be considered. This observation prompted the second hypothesis.
The departure number in the last pod was much larger than that in the first three pods, which is what the authors found from the eight dramas in the data resource (See Figure 1). The authors needed a more detailed observation of the departure pattern in each pod, however, for creating an appropriate statistical model. They calculated the percentage of the number of viewers tuning out in each minute during each pod, averaged over 20 episodes of the same drama (See Figure 2; the data of episode 2 are shown in Appendix A).
There was little variation in the tuning-out patterns during the first three pods, whereas there was a significant difference in the number of tuning-out occurrences between the first three pods and the fourth one (See Figure 2). This is because the first three pods were embedded in an episode of a drama (in-show commercials), and the fourth pod was embedded between the conclusion of an episode of a drama and the start of another (between-shows commercials). Viewers left a channel in the first three pods to avoid advertisements, but the storytelling feature of the drama caused many viewers to stay for the next segment. When one episode was over, however, many more viewers departed from this channel or turned off the television.
To address this observation, the authors classified the departure data in two groups: the first three pods, and the fourth pod. Statistical theory suggests that a decay function, such as a beta distribution, exponential decay, or gamma distribution, could be used for the departure pattern during commercials, with different parameters for the first three pods and the fourth pod, respectively. The authors used another drama to confirm this departure pattern (See Appendix C).
MODEL DEVELOPMENT
Before a discussion of hypothesis testing, it is necessary to explain the details of the underlying model building. The authors use two subsections to introduce the formulae of statistical distribution models for arrival and departure behaviors.
Formulae for Arrival (Tune-In) Behavior
Suppose that the people meter shows that for an individual tuning in to a channel at the t* minute (e.g., the 21:01 minute), the accurate arrival time t (tth second of the t* minute—e.g., the 15th second of 21:01 minute, t = 21:01:15) follows a uniform distribution over that minute—that is, t ~ U(1, 60).
Formulae for Departure (Tune-Out) Behavior
For departure behavior during a commercial break that starts at t0 and ends at t1, which could be more than one minute, a decay function is used for the time of departure, the pattern of which covers the whole period of the commercial break (t0, t1). To be able to use people-meter data, which are measured in the unit of minutes, the authors defined the whole observation time period, T, as the total number of complete minutes covering the commercial break. If, for example, a commercial break is from 21:01:20 to 21:04:30, then T is the duration of the time period from 21:01:00 to 21:04:60. That is, T = 4 minutes = 240 seconds. The commercial start time is t0 = 20 seconds, and the commercial ends at t1 = 210 seconds.
The First Three Commercial Pods
When observing the departure pattern of the first three commercial pods of each episode, the authors found that the departure number abruptly increased even before the commercials started. The explanation for this, provided by the marketing department of TVB, is that a short piece of music (about 10 seconds long) is played at the end of each segment of the drama to indicate the start of the commercials. Many viewers depart from this channel when the music starts.
In this context, the authors extended the time period of the decay distribution for the departure pattern beginning from the start time of the music to the end of a commercial break. If a commercial break, for example, is from 21:01:20 to 21:04:30 and the duration of the music prior to the commercial break is 10 seconds, the decay distribution covers the time period from (t0 − 10) seconds to t1 = 210 seconds. Denote the duration of the short music before the commercial break by tm. The authors termed the time period from (t0 − tm) to t1 as the departure break.
The authors developed a model for departure consisting of a uniform distribution before the music started and after the departure break and a decay distribution during the departure break. The uniform distribution here has a departure rate p1, which is defined as the probability of departure per minute per viewer during the broadcast of the drama. It can be estimated from the departure numbers in normal drama minutes—the minutes with drama broadcast.
The Fourth Commercial Pod
In accordance with the people-meter data, during the one minute prior to the fourth pod, more people departed than in the normal drama program time, but the departures were fewer and steadier than during the commercial time. The experts from TVBJ's marketing department explained that there is always an exit song broadcast for about one minute to show the credit titles and a preview of the next episode of the drama. Because this exit song is an indication of coming advertisements, some viewers start to leave the channel gradually, but many viewers remain on this channel to watch the preview of the next episode, recall the past episode, or wait for the next program.
Viewer departure behavior during the exit song thus can be modeled by uniform distribution, but with a different departure rate as compared with those during the normal drama time. The authors used the departure rate p2 to adjust the viewer departure pattern during the exit song time. They defined p2 as the probability of departure per minute per viewer during the exit-song time, which was estimated from the departure patterns in the exit-song time.
The number of departures from the people-meter data also shows that the departure pattern after the fourth pod and during the remaining part of the last minute of the observation period, (t1, T), followed a uniform distribution but was different from departure patterns during the program time in the prior four segments. This is because another program follows the fourth pod. Some viewers stay to see whether the next program is attractive (the lead-out effect; Cooper, 1993); some viewers tune out. In this case, the authors used another departure rate, p3, to adjust the uniform distribution for the departure behavior in the time interval (t1, T). The probability of departure per viewer per minute after the fourth commercial break, p3, is unknown and was estimated in the model.
Model Calibration and Validation
The data used to calibrate these distributions were the arrival and departure number of viewers in the four commercial pods embedded in the first 20 episodes of a television show. The authors chose to study the drama “Beyond the Realm of Conscience” because it was the most popular drama in 2014 and produced the largest sample of data for analysis. Three decay models can be used to fit the departure data during commercial pods: exponential decay distribution, beta distribution, and gamma distribution. The authors employed maximum likelihood estimation to estimate parameters in the three distributions by calculating each viewer's departing probability and then forming the likelihood function.
The authors used the chi-square test to compare their goodness of fit with the real data (See Table 2). They found that the uniform distribution for arrival behavior fitted well. For departure behavior, the exponential distribution fitted poorly and was unacceptable; the beta and gamma distributions fitted better. Finally, the authors chose the beta distribution to model departure behavior during commercial breaks for its best goodness of fit. The p values of the chi-square test of the uniform distribution for arrival and the beta distribution for departure were more than .3 in the first three pods and the fourth pod. This means that the discrepancy between the real and simulated data was not significant, so the uniform and the beta distributions can be used to model tuning-in and tuning-out behavior in an acceptable and accurate manner.
This successful model validation confirms that the statistical-distribution hypotheses (Hypotheses 1 and 2) for viewers' tuning-in (arrival) and tuning-out (departure) behaviors are acceptable. The authors thus verified the hypotheses and accomplished the central component of their methodology.
An Application Example
Commercial zapping decreases commercial effectiveness considerably through the loss of audience. Advertisers therefore need a more detailed understanding of commercial audience ratings when their advertisements are broadcast. Under the successful model validation, one may employ the distributions to simulate second-by-second audience ratings during a commercial break from minute-by-minute people-meter data. The authors calculated simulated audience numbers by second over a 10-minute interval that included a commercial break, as well as the observed audience numbers by minute (See Figure 3). The authors selected this time interval arbitrarily from the drama, “Beyond the Realm of Conscience.” The time points (marked on the horizontal axis in Figure 3) were as follows:
The drama was shown from 21:27:01 to 21:30:10.
A commercial pod then ran from 21:30:11 to 21:33:56 (enclosed by two vertical dashed lines in Figure 3).
The drama resumed from 21:33:57 to 21:36:60.
The authors found that after the commercials started (i.e., after the first vertical dashed line in Figure 3), the audience number declined rapidly. This simulation provides detailed insights into the variation in the second-by-second viewer number during the commercial pod (solid line in Figure 3). The observed behavior, in contrast, can supply only a ladder-type variation (the dashed line in Figure 3) in the minute-by-minute viewer number. From these findings, one also can understand why the reported audience number directly obtained from people-meter data significantly overestimates the true audience number during commercial breaks.
It is interesting to note from the simulated second-by-second data (See Figure 3) that the advertisements at the beginning of the commercial break provided greater exposure. The audience number continued to decrease while the advertisements were being broadcast, then recovered slowly when the program resumed. (Similar figures can be found in the studies of Swaminathan and Kent, 2013, Figure 5; Zigmond et al., 2014, Figure 1. These figures were drawn from observed second-by-second set-top-box data.)
From the simulated data over the 10-minute period, a clear picture of how viewers are gained and lost in each second and each program emerges. This insight may help advertisers to determine the optimal placement and length of their advertisements when considering the problems of audience ratings, commercial avoidance, and so forth. The use of the minute-by-minute record, however, may lead to serious mistakes.
EMPIRICAL VALIDATION
The core of the proposed methodology was to model tuning behaviors using appropriate statistical distributions and transformation formulae. The authors derived these formulae theoretically in the model-development stage. It was not necessary to test the correctness of the theoretical models themselves as long as the distributions used were appropriate, but it was necessary to demonstrate the application of the methodology with real-life data. In this section, the authors use three different dramas in different regions to validate the practical use of the proposed methodology.
Methodology Validation on TVBJ Drama
To validate the methodology, the authors needed to compare the simulated second-by-second audience ratings with the actual second-by-second ratings. They cooperated with a leading media-research company, whose identity cannot be disclosed for reasons of confidentiality. The company provides creditable and continual investigation on the television audience for Hong Kong and mainland China. For the purpose of the current study, it recalibrated the people meters to record the changes of a one-second interval, which yielded the second-by-second audience ratings covering the three dramas used to validate the methodology.
The authors calculated the simulated number of viewers in each second of the four pods embedded in episode 10 of the TVBJ drama, “Beyond the Realm of Conscience,” compared with the actual second-by-second number of viewers (See Figures 4, 5, 6 and 7). The validation based on this drama showed excellent results. The measure of simulation error was the mean absolute deviation (MAD), which is a well-known criterion used to measure the discrepancy between the simulated data and the actual data. On the basis of the scale of 550 viewers, the MADs of the four pods were 2.3, 1.6, 2.1, and 6.1 viewers per second, respectively. The results are encouraging, indicating the strong performance of the methodology. (See Table 3 for the MADs of each pod in the first 10 episodes of the drama.)
The pattern of audience resumption in the first pod was very different from that of the other three when the commercial was still being broadcast. The manager of the TVBJ marketing department explained that because it was the first pod of an episode, many viewers might tune in for the following part of the episode. The audience number decreased rapidly in the fourth pod, with much more audience tune-out in the between-shows commercials, because the episode was over (See Figure 7).
Methodology Validation on ATVH Drama
To validate their methodology further, the authors applied it to simulate second-by-second audience number during commercials in the drama, “Meter Garden,” broadcast on ATVH. Programs from different channels may constitute different populations, which could have different parameters for the assumed distributions, so the authors recalibrated the models and chose the best fit. The simulated number of viewers and the actual number of viewers of the four pods in episode 10 were calculated (See Figures 8, 9, 10 and 11).
The MADs of pods 1–4 were 2.7, 1.6, 2.0, and 4.9 viewers, on the basis of a scale of 150 viewers. The performance of the methodology applied to the drama broadcast on ATVH was also strong, and the patterns of audience increase and decrease in the commercial pods were similar to those of the TVBJ drama. The authors also showed MADs of each pod in the first 10 episodes to measure the simulation error (See Table 4).
Methodology Validation on Mainland Drama
In mainland China, the most expensive advertisements are between-shows commercials, because the authorities have banned in-show commercials since 2012. After a large number of observations and comparison analyses, the authors found the arrival and departure patterns of the commercial audience of Jiangsu TV similar to those of pod 4 of Hong Kong's TVB. The number of arrivals varied little during the programs and commercials. The number of departures also fluctuated little during programs but increased suddenly when commercials began.
The authors selected a typical between-shows commercial pod to show its viewers' arrival and departure patterns (See Figure 12). This between-shows pod was from the drama, “Divorce Lawyer,” broadcast on Jiangsu TV from August 4, 2014. The drama has 48 episodes and received positive reviews during the broadcast period. The Jiangsu TV station plays two episodes on each weekday and one episode per day on weekends. There thus are 28 between-shows commercial pods related to the drama “Divorce Lawyer.”
The authors calculated the percentage of the number of viewers tuning in and tuning out of this channel in each minute during the time period between 19:34 and 20:59 on August 14, 2014 (See Figure 12). From 19:34 to 20:08, Episode 19 was broadcast (marked Episode 19 between two vertical dashed lines in Figure 12), following which were the between-shows commercials. The departure number increased suddenly at the start time of commercials, then returned to the normal state as the end of commercials was approaching (See Figure 12).
Episode 20 was broadcast after the between-shows pod. In contrast with the departure pattern of the TVB drama (shown in Figure 1), the percentage of departure of the Jiangsu TV drama was higher. One possible reason is that the between-shows pod was longer, so more viewers might have chosen to switch channels during the long, boring commercials.
The authors applied their methodology of modeling the arrival and departure patterns of viewers to the people-meter data of the drama, “Divorce Lawyer,” on Jiangsu TV. The process of model calibration and validation was the same as the authors used with the Hong Kong TVB data. On the strength of reestimated distribution models, the authors simulated the number of viewers in each second during the 28 between-shows commercial pods.
Comparing these results with the actual second-by-second audience number, the authors found that the methodology again performed well (See Table 5). The authors provide a detailed look at the variation of the audience number during the between-shows commercials between episodes 19 and 20 (See Figure 13). The simulated number of viewers (solid line in Figure 13) varied little around the actual number of viewers (dashed line in Figure 13; MAD = 4.9 viewers, Pod 11, scale was 550).
The MADs of the 28 between-shows pods of “Divorce Lawyer” on Jiangsu TV were in the acceptable range, from 4.1 to 7.1. Note that the MADs of pods 6 to 7, 13 to 14, 20 to 21, and 27 to 28 were bigger than the others. A possible reason is that these eight pods were between one episode of “Divorce Lawyer” and one episode of another program, whereas the remaining pods were between two episodes of “Divorce Lawyer.” People's viewing behaviors are much more changeable in the commercial pods between two different programs.
Methodology Validation Using Reach Estimation
To further validate the practical use of the proposed methodology, the authors estimated the reach of each commercial break and compared the reaches obtained using different databases. For a clear validation analysis, the authors defined the pod reach as the total number of people exposed at least once to a given commercial break (i.e., pod; Ali, Turner, Cetintas, and Yang, 2017). The second-based estimate of the pod reach was produced from the simulated second-by-second data. The minute-based estimate was obtained in terms of minute-by-minute data. The actual reach denotes the pod reach calculated from the actual second-by-second data.
The above three dramas are representative because they are from different broadcast stations and regions. The authors thus estimated the pod reach of each cited commercial break in these programs and compared the pod reaches of the minute-based estimate, the second-based estimate, and the actual reach. (See Figure 14 for the comparison results of the pods of the TVBJ drama; See Appendix B for comparison figures of the other two dramas.)
It may be observed that the second-based pod reach was much closer to the actual pod reach than was the minute-based pod reach, which indicates the improvement of the estimation accuracy of the methodology. The minute-based line, in addition, had bigger fluctuations in the reaches of the 40 commercial pods, whereas the second-based estimates from the methodology showed fewer fluctuations. That is, the pod reach from the authors' simulation was more stable in reflecting the actual reach.
Three criteria were used to measure variances of estimation: mean squared error (MSE), mean absolute percentage error (MAPE), and mean estimation error (MEE). MSE, calculated as the average of the squares of the errors between the actual reach and the estimated reach, is always nonnegative, and values closer to zero are better. MAPE, which expresses accuracy as a percentage, is a measure of estimation or prediction accuracy of a method or model in statistics. MEE, defined as the average of the errors between the actual reach and the estimated reach, may be positive or negative, and values closer to zero are better. It can be a good measure of serious over- or underestimations when most errors are negative or positive.
Values of MSEs, MAPEs, and MEEs often are used for comparison purposes. Two or more statistical estimates may be compared with their MSEs, MAPEs, or MEEs as a measure of how well they explain a given set of observations (Bork and Stig, 2015; Mandalinci, 2017; Tofallis, 2015). The authors calculated the MSEs, MAPEs, and MEEs of each program's pod-reach group by minute-based and second-based estimations and compared their values to validate the better performance of their methodology. (See Table 6 for the values of MSEs, MAPEs, and MEEs of the three programs.)
The second-based estimates of pod reach were much more accurate than the minute-based estimates, because the values of MSEs and MAPEs from minute-based estimates were much bigger than those of second-based estimates among the three programs. For MEEs, it is apparent that the second-based group was closer to zero, and the MEEs for the minute-based group had much larger positive values, which highlights again that this methodology performed better than the conventional minute-based method. It also reflects the higher possibility of overcharging the advertising agent when minute-based reach is used. In conclusion, the second-based reach from this methodology was much closer to the actual reach, which is more reasonable and reliable in advertisement pricing.
The authors thus have completed the methodology validation on different dramas from different regions with various populations. Through the mathematical formulae derivation and statistically significant outcomes of the validation analyses, this methodology has been confirmed in practical use.
DISCUSSION
Possible Applications
The authors have started using their methodology for a variety of applications at an advertising agency in mainland China. The second-by-second viewing data are converted from minute-by-minute data for advertisers, who can use them to evaluate the advertisement effectiveness. This may be a useful reference for advertising strategy.
The advertising agency purchased a 10-second between-shows spot during prime time on four television channels. The audience levels measured in minutes of the four channels were the same: 400 viewers. The corresponding advertisement prices were also the same. When the authors referred to the converted second-by-second audience data, however, the gaps among the four channels were very apparent. To illustrate this finding, the authors plotted the viewer numbers during the 10-second advertisement as measured by the scale of minute and second for the four channels (See Figure 15). They plotted the 10-second interval spot by truncating two sides of between-shows commercials.
One can see that the minute-by-minute people-meter data overestimated the real commercial exposures. The audience gap between the minute level (dashed lines) and the second level (solid lines) was especially bigger in channel 4 (average number of viewers per second = 392). Channel 3 retained the audience best; its average number of viewers per second was 399.9 (i.e., 400). The advertising expenses thus were spent well only on channel 3. This sort of analysis can provide a reference for audience measurement and advertisement placement and even can serve as a potential aid in future advertisement purchasing and price determination.
Conclusion and Future Work
Because of commercial-zapping behavior, program-audience ratings are not equal to commercial ratings. Advertisers prefer to observe the audience variation of their advertisements directly rather than to be given general program ratings. Installing people meters in collaborative households is still a popular and persuasive method of collecting television-audience data in many areas in the world.
People-meter data commonly are accurate to one minute. They reveal the changes of viewers' watching behavior in each minute to reduce the cost of data transmission and data storage. Advertisements usually are shorter than one minute, however—perhaps 15 seconds. The minute data cannot indicate the change of audience number in each second in a commercial spot. The current research proposed to develop a methodology to convert the minute-by-minute audience ratings into second-by-second ratings.
According to a careful observation and intensive data mining, people-meter data recorded in one-minute intervals suggest that the number of individuals tuning in to (arriving at) a channel remains stable with few fluctuations during the program and commercial time periods. The number of tuning-out (departure) occurrences before and after commercials also varies little. During commercials, however, the number of departing individuals is markedly different, exhibiting an abrupt increase followed by a decaying trend as the commercial time continues.
Accordingly, the authors modeled arrival behavior as well as departure behavior before and after commercials using a uniform distribution; they used a beta distribution for departure behavior during the commercials. The authors also defined several different departure rates to adjust the different departure situations. They verified the model calibration using chi-square goodness-of-fit tests on the basis of real data from television stations.
In cooperation with a leading media-research company that supplied the second-by-second records covering the validation period, the authors validated their methodology by comparing the simulated second-by-second commercial audience with the actual second-by-second data and estimates of pod reach based on different data bases. They showed the empirical validation in terms of three dramas from different television stations: TVB in Hong Kong, and Jiangsu TV in mainland China. The results are very inspiring. The methodology was accurate enough to simulate the second-by-second viewer number from minute-by-minute people-meter data; the converted data varied little around the actual second-by-second viewing data.
Many feasible applications of this methodology may be useful for advertisers and television stations. The second-by-second audience number can be simulated from the original people-meter data recorded in one-minute intervals (as shown in Figure 3). These simulations can be helpful in measuring advertising audience, assessing advertising strategies, and aiding in future media purchasing and pricing. In this way, the authors have extended the functionality of minute-by-minute people-meter data.
The formulae and models included in the methodology were derived theoretically, and the methodology validations were sufficient, so people can use this approach with confidence for dramas from various regions. Without investing additional resources or effort, advertisers are able to obtain more detailed information about their advertisements from normal people-meter data. This methodology could be used effectively in a wide range of academic research and managerial situations in the future.
The current research has several limitations. First, the statistical distribution assumptions were made mainly in accordance with observed people-meter data, the tuning-in and tuning-out concurrences. The authors have not measured the effect of influenced factors on the switching behavior, which have been confirmed in previous research (Rohas-Mendez and Davies, 2005; Yao et al., 2017).
Second, the authors applied directly the existing beta-distribution model, and they obtained the p value of the chi-square test (.33). That means that if one assumes that the departure time follows a beta distribution, the probability of obtaining the calculated or even larger chi-square statistic is .33. The p value reflects the discrepancy between fitted data and actual data. This significance test often is referred to as the threshold value .05. When the obtained p value is bigger than .05, which means this test is not significant, one cannot reject the proposed hypothesis. Even though the obtained p value in this study was bigger than the threshold value .05, advanced statistical models may be considered to improve the fit accuracy.
In the long run, the authors hope this methodology will be helpful in constructing a forecasting system to predict advertisement-viewing behavior in future commercial time. It has been confirmed in previous studies that two factors are very effective for advertisement-viewing behavior: program broadcast time, and program genre (Schweidel and Kent, 2010; Swaminathan and Kent, 2013). In a certain, specific time period, people may perform certain work (e.g., between 7:00 and 8:00, people are busy preparing to go to the office or school, whereas from 21:30 to 22:30, many people may stay at home and watch television or do some housework). The models for tuning-in and tuning-out behavior thus may change when the program broadcast time changes. Similarly, the program type also may have an influence on viewing behavior; a drama has a higher advertisement audience-retention ratio than a program of public affairs (Schweidel and Kent, 2010).
Given these two factors, researchers may construct different models of fitting commercial-viewing behavior to constitute a forecasting system. A general advertisement-viewing pattern for future programs can be predicted according to the models of the same program-type broadcast in the same time period. As a result, together with the prediction of audience program ratings, the proposed methodology conveniently and quickly can predict advertisement audience ratings when advertisers have chosen the time period and program type to broadcast their advertisements. In addition, future studies may consider extending the model methods and simulation variables to other data bases, as in a previous study (Ephron and Gray, 2001) that modeled the viewers per viewing household from the tuning information of set-meter panels.
ABOUT THE AUTHORS
Lianlian Song is an assistant professor at Nanjing University of Aeronautics and Astronautics, Jiangsu, China. Her research focuses on marketing research and operations management. Her work has appeared in Energy Economics and Journal of Advertising Research.
Peng Zhou is a professor at Nanjing University of Aeronautics and Astronautics. His research areas include business economics and energy economics. His work has been published in Journal of Environmental Economics and Management and Energy Economics.
Geoffrey Tso is an associate professor at City University of Hong Kong. His research areas include business economics and marketing research. His articles have appeared in European Journal of Marketing and Advances in Economics and Business.
Hingpo Lo is an adjunct associate professor in the Department of Civil Engineering at the University of Hong Kong. His research focuses on statistical modeling in marketing research. He has published many research articles in journals such as the European Journal of Marketing and Journal of Advertising Research.
ACKNOWLEDGMENTS
This research was supported financially by the National Natural Science Foundation of China (Grant 71602089), the Natural Science Foundation of Jiangsu Province (Grants BK20160785 and 16GLC001),and the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. 11507817).
APPENDIX A Formulae Derivation in Model Development
GENERAL FORMULAE FOR ARRIVAL BEHAVIOR
Suppose that the people meter shows that an individual tunes into a channel at t* minutes. One assumes that this viewer's accurate arrival time t (tth second of the t* minute) follows a uniform distribution over the one minute—that is, t~U(0, 60). There is no parameter to be estimated. Thus, one can use this model directly. The probability distribution function of arrival time is f(t) = 1/60.
GENERAL FORMULAE FOR DEPARTURE BEHAVIOR
The authors use the beta distribution as an example to illustrate the general formula derivation for departure behavior. Other decay distributions may be considered; the steps of developing the formulae are similar.
During the commercial period, the viewer departure pattern is assumed to follow a beta distribution (from t0 to t1). The probability distribution function is (1) where a and b are unknown parameters in the truncated beta distribution. On the basis of Equation (1), one can find the probability distribution function of departure time in each minute during commercial breaks. Because the people-meter data are in the unit of minutes, viewers' tuning behaviors are complex in the first and last minutes of a commercial break. One needs to derive the specific formula of each minute for viewer-tuning behavior, which is helpful in model calibration and validation and also in the future step of second-by-second data simulation.
THE FIRST THREE PODS
Case 1
In case 1, viewer A departs in the first minute of a departure break (0<t<60), and the commercial break starts in the first minute (t0 + tm < 60). The probability distribution function of departure time during this first minute is as follows: (2) where p1 is defined as the probability of departure per minute per viewer during the broadcast of the drama. It can be estimated from the departure numbers in normal drama minutes (the minutes with drama broadcast). PF is the probability of the portion of the beta distribution from t0 to 60, which may be found as (3)
The factor is used to adjust the probability, because the departure break starts at the t0th second. If a viewer leaves before the t0th second, therefore, his or her departure pattern follows a uniform distribution, leaving behind a beta distribution for the departure pattern in the remaining (60 − t0) seconds. One therefore needs to adjust the probability distribution by a factor .
Case 2
For viewers who departed in each minute in the interval (2, [T − 60]/60), n denotes the nth minute in the time interval (2, [T − 60]/60), n = 2,…, (T/60 − 1). The probability distribution function of the departure time is (4) where PM is the probability of the portion of the beta distribution from 60[n − 1] to 60n, which can be found as (5)
Because the departure pattern in this minute follows the beta distribution for the whole one minute, one needs PM to obtain the truncated beta distribution.
Case 3
For viewers who departed in the last minute (T − 60 < t < T) of the first three pods, the departure pattern is described as follows: (6) where PL is defined in the text as the probability of the portion of the beta distribution from (T − 60) to t1, which can be obtained as follows: (7)
The factor is used to adjust the probability because the advertisement ends at the t1th second. If a viewer departs after the t1th second, therefore, his or her advertisement-viewing duration should be fixed at (t1 − t0) seconds. After the advertisement broadcast period in the next (T − t1) seconds, the departure pattern follows a uniform distribution.
THE FOURTH POD
Case 4
For viewers who departed in the first minute of pod 4 (0 < t < 60), the probability distribution function of the departure pattern is (8) where p2 is defined in the text as the probability of departure per minute per viewer during the exit-song time. This is estimated from the departure patterns in the exit-song time over 20 episodes. PF is the same as defined in the first three pods.
Case 5
For viewers who departed in each minute in the interval (2, [T − 60]/60) in pod 4, n denotes the nth minute in the time interval (2, [T − 60]/60), n = 2, …, (T/60 − 1).
The departure pattern is (9) PM is the same as defined in the first three pods.
Case 6
For viewers who departed in the last one minute (T − 60 < t < T) of pod 4, the departure pattern is shown as follows: (10) where p3 is the probability of departure per viewer per minute after the fourth commercial break, which is unknown and will be estimated in this model. PL is the same as that defined in the first three pods.
Appendix B Viewing Pattern of Tuning Out
The authors plotted the average percentage of the number of departure viewers in each commercial minute over the 20 episodes of another drama (See Figure B1). This drama is called “Meteor Garden” and was broadcast on Asia TV Ltd. (another television station in Hong Kong) between 19:00 and 20:00 each weekday night from April 27, 2014, to May 29, 2014. It can be observed that, except for the values of percentages of departures, the departure patterns are very similar to those for the Hong Kong TV Broadcasts drama (See Figures 2 and B1).
Appendix C
Appendix D Comparison of Estimates of Reach
The comparison results of the estimated pod reaches of the Asia TV Ltd. drama and the mainland drama are presented separately (See Figures D1 and D2).
- Received March 26, 2017.
- Received (in revised form) December 7, 2017.
- Accepted January 17, 2018.
- Copyright© 2019 ARF. All rights reserved.