Multiple Relationships between Streaming and Linear TV: Examining Media Substitution Theory Using Big Data

ABSTRACT With the prevalence of connected TV, streaming is replacing linear TV while expanding its functionality. To structurally explain this replacement based on functional similarity, we applied two-way fixed-effects regression to log data from 197,273 smart TVs in Japan from July 2019 to June 2022. Results showed that professional videos on demand primarily substituted linear TV’s recorded viewing, dramas, and movies, whereas substitution by YouTube was broader. Catch-up streaming substituted recorded viewing while complementing viewing without recording. These multiple relationships support the applicability of media substitution theory and foreshadow that functional expansion of streaming would further contribute to linear TV’s replacement.


Introduction
Every medium worldwide has been heavily affected by the Internet, including radio, phones, newspapers, books, and even academic journals.Television (TV) is no exception; TVs connected to the Internet, namely connected TVs (CTVs), have become increasingly prevalent.A 2022 survey showed that 87% of US households have at least one CTV (Leichtman Research Group, 2022).CTVs enable users to use online video streaming services alongside linear TV. 1 According to Nielsen (2022), streaming accounted for 36.9% of TV viewing time in the US in September 2022.
Thus, given this shift, media industry players must now grapple with an unprecedented expansion of domains and players, as shown by the following two trends.
The first is the expansion of streaming formats and content genres.Streaming pioneers (e.g., YouTube, Netflix) have long attracted viewers using Internet-based on-demand formats.However, new streaming services have emerged that incorporate linear formats.Examples include virtual multichannel video programming distributors (vMVPDs), free ad-supported streaming TV (FAST), and comprehensive subscription video-on-demand services such as Paramount Plus.These services emphasize live news and sports, at which linear TV has excelled (Comcast Advertising, 2022;FCC, 2020).This trend implies that streaming is expanding its functionality into areas closer to linear TV, and its potential impact is unclear.
The second trend is that streaming services by broadcasters are becoming less limited to catch-up services as complements of linear TV.A typical example is BBC iPlayer, which aims to pivot from a catch-up service into a go-to destination for audiences (Ofcom, 2019).Examining this strategy shift requires understanding the relationship between linear TV and catch-up streaming services, a more traditional format for broadcasters.
To comprehend this linear TV's replacement with streaming, this study aimed to structurally explain the relationship between streaming and linear TV, by focusing on various formats and content genres.Consequently, we supported the media substitution theory, which argues the effect of functional similarity on time allocation (Kaye & Johnson, 2003;Lin, 1994Lin, , 1999)).
Our structural explanation involves two aspects.One is demonstrating multiple internal relationships that constitute the whole relationship between streaming and linear TV.We achieve this through a longitudinal analysis of temporal relationships using big data with sufficient granularities from smart TVs in Japan, where linear TV was highly impacted by various streaming services.The other aspect is explaining each relationship through functional similarity and complementary effect.The foundations for our hypotheses correspond to this explanation.

TV Landscape in Japan
TV in Japan has been majorly influenced by five commercial broadcasters, called "key stations," and their networks (Nishioka, 2017).However, with CTV's prevalence, global streaming services such as YouTube, Netflix, and Amazon Prime Video have gained popularity in Japan.According to a 2021 public survey, 95.9% of the households in Japan have TVs, and 55.8% have CTVs (Ministry of Internal Affairs and Communications, 2022).
Based on the criteria in the Methods section, this study considered these three global services and four Japanese services: TVer, Hulu, U-NEXT, and ABEMA.TVer is a free catch-up streaming service that offers on-demand viewing within seven days after broadcast.Key stations' networks provide most of the programs on TVer.Hulu and U-NEXT are subscription video on demand services in Japan.Note that although Hulu in Japan is owned by a key station, its content is not limited to the owner's network, unlike TVer.ABEMA is a hybrid streaming service that incorporates linear and ondemand, free and subscription options, such as FAST in the US.

Media Substitution Theory
Our study employs media substitution theory as the theoretical approach, one of the two interrelated approaches to investigate media competition.This theory examines media competition based on resource allocation and argues that new media could displace traditional media.Since Lazarsfeld (1940) examined the displacement of print media by radio, scholars have investigated subsequent displacement by TV (Belson, 1961), the Internet (Kayany & Yelsma, 2000;Lee & Kuo, 2002), and streaming (Cha, 2013a;Lee & Lee, 2015).Time allocation is a primary indicator in media substitution theory.Westlund and Färdigh (2015) defined displacement as the reduction in the use of one medium due to the use of another, contrasting this with its converse complementarity.Moreover, Ha and Fang (2012) distinguished between permanent replacement (displacement) and temporary replacement (substitution), proposing that substitution potentially triggers displacement.
The second approach describes media competition through functional similarities among media.Media function has been examined through surveys by employing the uses and gratifications theory (UGT) (Katz et al., 1974) and the theory of niche (Dimmick & Rothenbuhler, 1984) as frameworks.UGT investigates the audience's motive and gratification to media corresponding to the function each medium offers to the audience.Niche theory utilizes UGT and the analogy of ecology to describe spatial competition, investigating which media combinations have functional similarity and superiority.
Functional displacement is a key concept in media substitution theory, which bridges the two approaches by referring to fulfilling the same function via a different medium (Kayany & Yelsma, 2000).This concept assumes that viewers evaluate the satisfaction offered by each medium and their own needs (Katz et al., 1973;Lin, 1994) and then shift their time to a new, functionally similar media (Atkin et al., 1998).Thus, this concept suggests that functional similarity predicts and explains the competition for resource allocation.
Therefore, studies on resource allocation have attempted to explain the difference in the impact of new media by the difference in functional similarities (Kraut et al., 2006;Lee & Kuo, 2002;Liebowitz & Zentner, 2012).Nevertheless, to our knowledge, studies on the recent shift in time allocation from linear TV to streaming are insufficient.

Empirical Studies of Temporal Relationships Among Media
This section reviews methodological development and findings from previous media substitution research that informed our approach.
Scholars have long noted the need for a longitudinal study design to estimate temporal relationships by controlling unobserved confounders (Lee & Kuo, 2002;Mutz et al., 1993).Kraut et al. (2006) showed that longitudinal and cross-sectional analyses yield different results regarding the relationship between the Internet and TV.In more recent methodological developments, scholars have analyzed passive log data, enabling large sample sizes (Belo et al., 2019;Fudurić et al., 2020;Kim et al., 2020).Although generalizing results should be done cautiously because of the potential for systematic biases stemming from nonrandom sampling, large samples reduce random variance, enabling granular analysis.
Studies have also examined temporal relationships in various media combinations.Substitutional relationships have been found between Internet use and TV viewing (Kraut et al., 2006;Lee & Kuo, 2002;Liebowitz & Zentner, 2012, 2016).Jang and Park (2016) found substitutions among paper, TV, and computer media.They also found complementarity between telephone and computer media in total usage and substitution in information searching.Kim et al. (2020) found complementarity between mobile and PC device use and among mobile device usage by genre.Complementarity also exists between different formats offered by the same content providers.Belo et al. (2019) found complementarity between live and time-shifted viewing of linear TV and noted a catch-up effect.Xu et al. (2014) found it between Fox News' mobile app and website.
Regarding streaming and linear TV, the focus of this study, Silva and Lima (2022) found that the popularity of Netflix was associated with reduced cable TV subscribers.Fudurić et al. (2020) found that viewing live news and sports on linear TV reduced cord shaving and attributed it to streaming's inadequate coverage of these genres.Cha (2013a) and Lee and Lee (2015) focused on temporal relationships, like our study.Cha (2013a) found a substitutional relationship between user-generated content (UGC) and linear TV and suggested that the TV network's streaming video could complement linear TV. Lee and Lee (2015) found streaming's substitution effect on cable TV, attributing it to content overlap.
These prior studies showed varying degrees of substitution and complementarity in the relationship between streaming and linear TV.They call for further research distinguishing content genres and streaming services (Fudurić et al., 2020;Lee & Lee, 2015).Thus, this study was designed to illuminate this complexity using large, granular, and longitudinal data.

Hypotheses
This study aimed to structurally explain the temporal relationship between streaming and linear TV.Accordingly, we considered various formats and content genres, and formulated hypotheses based on the presumed functional similarities and complementary effects among them.Linear TV was decomposed into live and recorded viewing2 and further into program genres for live viewing.
For streaming, we grouped seven services into four categories based on formats and contents.The first category, professional video-on-demand services (PVODs), includes Netflix, Amazon Prime Video, Hulu, and U-NEXT.These services provide professional content in an on-demand format.The second category is the UGC video-on-demand service, represented by YouTube.The third category is hybrid streaming, represented by ABEMA, combining live and on-demand formats.CyberAgent ( 2022) reports that viewing time on linear format comprised 59% of ABEMA's total viewing time in March 2022.Finally, the fourth category is catch-up streaming, represented by TVer, offering on-demand access to recently aired linear TV content for a limited period.
Our hypotheses comprise five on substitution relationships and one on complementary relationships.Hypotheses on substitution relationships were derived from three perspectives on functional similarity.The first perspective concerns viewing attitudes.Classification of viewing attitudes into ritualistic and instrumental is well-known categorization based on UGT (Rubin, 1983(Rubin, , 1984) ) and has been validated for CTV (Tefertiller & Sheehan, 2019).Ritualistic viewing is passive, whereas instrumental viewing is active and content-focused (Rubin & Perse, 1987).Studies comparing linear TV viewing and the use of the Internet have characterized the former as more ritualistic compared to the latter (Cooper & Tang, 2009;Ferguson & Perse, 2000;Metzger & Flanagin, 2002).However, recorded viewing of linear TV has been characterized as active and selective, namely instrumental (Henke & Donohue, 1989;Levy, 1987).Furthermore, streaming's on-demand format encourages the instrumental viewing of entertainment content (Tefertiller & Sheehan, 2019), as demonstrated by the highly content-focused viewing called "binge-watching" (Jenner, 2016).
However, studies have suggested that YouTube and ABEMA are not limited to instrumental viewing.YouTube caters to broader purposes, encompassing information, co-viewing and social interaction (Haridakis & Hanson, 2009;Kim et al., 2016).ABEMA offers both linear and on-demand formats, and linear viewers watch a broader range of news genres than ondemand viewers (Takano et al., 2021).
These studies highlight the functional similarity between PVODs and recorded viewing of linear TV for instrumental viewing, whereas YouTube and ABEMA encompass functionality for ritualistic viewing, prevalent in live viewing.Based on this discussion, we present the first set of hypotheses.
H1a: YouTube and ABEMA have a larger substitute effect on live viewing compared with PVODs.
H1b: PVODs have a larger substitute effect on recorded viewing compared with YouTube and ABEMA.
The second perspective concerns shared capability.Recorded viewing and catch-up streaming both enable time-shifted viewing of linear TV.Therefore, we hypothesized that catch-up streaming could substitute traditional recorded viewing as a new form of time-shifted viewing, as follows.

H2:
TVer has a substitute effect on recorded viewing.
The third perspective concerns content genres.Preferences for content genres are based on video viewing motives (Cha, 2013b), suggesting substitutability between media offering similar genres (Belo et al., 2019;Fudurić et al., 2020;Lee & Lee, 2015).Despite the functional expansion of streaming, we argue that PVODs in Japan still primarily focus on scripted entertainment, such as dramas, movies, and anime.Among popular programs on Japanese Hulu, four of the top five are scripted entertainment (Hulu, 2022), while U-NEXT's ranking is dedicated to scripted entertainment (U-NEXT, 2022).Although official data regarding popular content on Amazon Prime Video is unavailable, we contend that news and sports consumption is limited owing to the scarcity of dedicated channels in Japan.This focus on scripted entertainment also aligns with Netflix's (2022) global strategy, which emphasizes scripted entertainment over news and sports.Conversely, YouTube's UGC covers various genres, and ABEMA offers live news and sports channels alongside scripted entertainment.Based on these content characteristics, we propose the following set of hypotheses.
H3a: YouTube and ABEMA have a larger substitute effect on news and sports compared with PVODs.
H3b: PVODs have a larger substitute effect on scripted entertainment compared with YouTube and ABEMA.
The last hypothesis explores the complementary aspects of catch-up streaming while maintaining focus on the content genre.Cha (2013a) found that viewers of TV networks' streaming videos have a favorable attitude toward the networks and suggested that they watch missed programs on these sites.Additionally, Belo et al. (2019) revealed the complementary effect of time-shifted viewing on live viewing, which they attribute to encouraging viewers to catch up with the live broadcast schedule.Therefore, we hypothesized that catch-up streaming only complements series programs since non-series programs do not require contextual knowledge of previous episodes.

H4:
TVer has complementary effects only on series programs.

Data
We analyzed TV data from 197,273 smart TV devices in Japan from July 2019 to June 2022. 3These data covered smart TVs from one major Japanese TV manufacturer and were limited to devices with available viewing data for all quarters.All data used in this study were provided by INTAGE Inc., a leading marketing research company in Japan.
Our TV viewing log data comprised each TV device's exposure status every 15 seconds.These data enabled us to distinguish TV viewing into categories, such as live and recorded viewing of linear TV, each streaming service, and others.Furthermore, as the channels for live viewing were logged, we could distinguish the program genres for live viewing, combined with electronic program guide data. 4As a limitation, viewers' attribute information was unavailable since our data were device-based.

Measures and Descriptive Statistics
In media substitution theory, time allocation is a key indicator of substitution and complementarity among media.Therefore, each device's average daily viewing hours were calculated by categories and quarters.We aggregated every 15 seconds of data into 12 quarters based on the quarterly organized programming in Japan and the need to balance the number of observations with the stability of each observation.
Tables 1 and 2 show descriptive statistics for average daily viewing hours by our media categories.Linear TV was disaggregated into live and recorded viewing and further into program genres for live viewing.The genres in our study align with our hypotheses.H3a focuses on soft and hard news and sports, while scripted entertainment in H3b refers to dramas, anime/tokusatsu, and movies.Series programs in H4 correspond to dramas and anime/tokusatsu.Note that models by genre exclude those with less than a 1% share in live viewing.For streaming, we categorized services according to our hypotheses.PVODs consist of Netflix, Amazon Prime Video, Hulu, and U-NEXT.Services with less than a 1% share of streaming viewing and non-streaming services were included in the "other services" category in Table 1 and excluded from all models.
The limited coverage of our data required the use of additional procedures.Our recorded viewing data included only viewing through USB-connected Hard Disk Drives (HDDs).We could not include recorded viewing through HDMI-connected recorders in the recorded viewing category because we could not distinguish HDMI-connected recorders from other HDMI-connected devices.Therefore, we extracted the HDD group of 42,006 devices with available HDD data for all quarters as a subset of the ALL group of 197,273 devices to estimate a part of the later statistical model.We presumed that we could cover most of the HDD group's recorded viewing, as it would be unusual if both an HDD and recorder were connected to a single TV device and used for recorded viewing.Table 1 shows the differences between the ALL group and the HDD group.As expected, the HDD group spent more time on recorded viewing than did the ALL group (ALL: Mean = 0.229; HDD: Mean = 0.957; difference = 0.728).Note that the HDD group spent less time on the others category than did the ALL group (ALL: Mean = 1.577;HDD: Mean = 0.725; difference = −0.852),as we anticipated.
Before we examine the structural relationship between streaming and linear TV, a brief overview is presented.Table 3 shows quarterly descriptive statistics that suggest a time allocation shift from live viewing of linear TV to streaming.We focused on the composition percentages to avoid irregular changes in total TV viewing time associated with the COVID-19 pandemic.Streaming increased from 19.5% to 28.2%, and live viewing decreased from 57.7% to 52.0%.

Model
A two-way fixed-effects regression model was used in this study to estimate the multiple effects of streaming on linear TV.As our data were device-based, unavailability of viewers' socioeconomic attributes could have confounded streaming and linear TV.However, two-way fixed-effect regression controlled these unobserved factors as fixed effects.5 Therefore, this model estimated the effects of changes in streaming on changes in linear TV viewing.In fact, longitudinal studies have used two-way fixed-effect regressions to understand the relationship between new and traditional media (Liebowitz & Zentner, 2012, 2016;Silva & Lima, 2022).Our model was specified as follows.
This model was estimated for each linear TV category c, representing live viewing, recorded viewing, and program genre.To compare the regression coefficients b cj as the effect magnitude across different linear TV categories, Y cit was divided by its mean � Y c .As the error term ε cit was clustered by device, clusterrobust standard errors by device were estimated (Cameron & Miller, 2015).We used the "feols" function in fixest, the successor to the R package FENmlm (Bergé, 2018).After de-meaning to exclude fixed effect, no variance inflation factor exceeded 1.01 for any explanatory variable X ijt across all regressions.
The following 95% confidence interval was employed in the two-sided significance tests of the differences in the magnitude of regression coefficients between the streaming services related to hypotheses H1a, H1b, H3a, and H3b (Shrout & Yip-Bannicq, 2017).Given our large sample size, we approximated the t-distribution to the standard normal distribution.

Y cit
Average daily hours for linear TV category c for device i in quarter t X ijt Average daily hours for streaming service j for device i in quarter t b cj Regression coefficient for the effect of change in X ijt on change in Y cit μ ci Device i's fixed effect for category c γ ct Quarter t's fixed effect for category c ε cit Error term

Hypothesis Testing for Live and Recorded Viewing
In this section, we tested the hypotheses regarding functional similarities in viewing attitude (H1a and H1b) and shared capability (H2).Table 4 presents the estimation model results for live and recorded viewing, while Table 5 compares the magnitudes of coefficients between streaming services related to our hypotheses.All regression coefficients were significantly negative, except for TVer in live viewing, and the magnitudes varied among services.These findings support H1b and H2, and partially support H1a. H1a states that YouTube and ABEMA have a larger substitution effect on live viewing than PVODs.This was supported when comparing YouTube and PVODs.As shown in Columns 1 in Tables 4 and 5, YouTube had a significantly larger negative impact on live viewing than PVODs.Although ABEMA also negatively affected live viewing, it did not significantly differ from what PVODs did.
By contrast, H1b posits that PVODs have a larger substitution effect on recorded viewing than YouTube and ABEMA.This was supported by Columns 3 in Tables 4 and 5. PVODs exhibited a significantly larger negative effect on recorded viewing than YouTube and ABEMA.
Finally, H2 asserts that TVer has a substitution effect on recorded viewing.This is corroborated by Column 3 in Table 4, which shows a significantly negative coefficient for TVer in recorded viewing.However, the coefficient for live viewing within the same HDD group was insignificant, as in Column 2 in Table 4. 6

Hypothesis Testing for Each Program Genre
In this section, we tested the hypotheses regarding functional similarities in content genres (H3a and H3b) and the complementary effect of catch-up viewing (H4).Table 6 presents the estimation results for live viewing by program genres, while Figure 1 compares the coefficients' magnitudes between services to test our hypotheses.The results partially support H3a and H3b but not H4.
H3a posits that YouTube and ABMEA have a larger substitution effect on news and sports than PVODs.Columns (1)-(3) in Table 6 and Figure 1 support H3a when comparing YouTube and PVODs, but not ABEMA and 6 The positive effect of TVer on live viewing was statistically significant in the ALL group but not significant in the HDD group [columns ( 1) and (2) in     Note.A confidence interval not containing zero indicates statistically significant differences in regression coefficients.Confidence intervals are wider for streaming services and linear TV categories with lower viewing hours, and for device groups with smaller sample sizes.CI = confidence interval; LL = lower limit; UL = upper limit.
PVODs.YouTube had a significantly larger substitution effect on news and sports than PVODs.Furthermore, YouTube demonstrated a large substitution effect across various genres, including the most popular genres, news, and variety.By contrast, PVODs' negative effect on news and variety was less than half of dramas and movies.This difference in the magnitude of the regression coefficients contributed to the overall effect on live viewing examined in H1a.The direction of differences in regression coefficients between ABEMA and PVODs for news aligns with H3b but was not statistically significant.
H3b posits that PVODs have a larger substitute effect on scripted entertainment than YouTube and ABEMA.Columns ( 4)-( 6) in Table 6 and Figure 1 support H3b when comparing YouTube and PVODs, except for anime/tokusatsu, but were not significant when comparing ABEMA and PVODs.PVODs' substitution effect on dramas and movies was larger than YouTube's but smaller on anime/tokusatsu because PVODs' substitution magnitudes on anime/tokusatsu are about half of that on dramas and movies.The direction of differences in regression coefficients between ABEMA and PVODs aligns with H3b but was not statistically significant.Finally, H4 states that TVer would have complementary effects only on series programs.Table 6 shows that TVer had positive effects on dramas and anime/tokusatsu, namely series programs.However, TVer had positive effects on many other genres beyond dramas and anime/tokusatsu.The largest positive effect was on hobby/education, and variety exhibited the third largest effect.Thus, H4 was not supported.

Discussion
To examine media substitution theory, we analyzed the relationship between streaming and linear TV on smart TVs.Overall, our results highlight the applicability of media substitution theory.Except for TVer, all streaming services had substitution effects on live and recorded viewing of linear TV.This finding is similar to that of Belo et al. (2019) for TV devices7 and differs from that of Kim et al. (2020) for mobile devices.There are two possible reasons for the different results between TVs and mobile devices.The first is the ubiquity of mobile devices (Kim et al., 2020;Okazaki & Mendez, 2013).While mobile devices are available in various locations, TVs are available only at fixed locations.This limitation caps the total consumption time and facilitates substitutional relationships.That is, the media substitution theory is premised on resource limitation.The second reason is functional similarity.Kim et al. (2020) included non-video activities in their analysis, whereas Belo et al. (2019) and this study analyzed only video viewing.

Substitutional Relationships Resulting from Functional Similarities
We uncovered substitutional relationships resulting from functional similarities in viewing attitudes, shared capability, and content genres.The results suggest the functional displacement of linear TV with streaming, generally supporting media substitution theory.YouTube and PVODs results are mainly consistent hypotheses on viewing attitudes and content genres, whereas TVer results align with the hypothesis on shared capability.This consistency foreshadows the impact of the ongoing functional expansion of streaming and provides managerial implications.Unclear results for ABEMA and the deviation from the hypothesis for anime/tokusatsu are discussed later as limitations of this study.
PVODs had larger substitution effects on recorded viewing, as well as live viewing of dramas and movies, than YouTube did.This result agrees with Belo et al. (2019) and Fudurić et al. (2020), who found that content overlaps enhance substitutional relationships, and common active viewing attitudes toward content, namely instrumental viewing (Henke & Donohue, 1989;Levy, 1987;Tefertiller & Sheehan, 2019).Conversely, YouTube had broadly large substitution effects on both live and recorded viewing and all program genres, including news and sports.This result extends that of Cha (2013a), who found a substitutional effect of UGC streaming on linear TV.More important is the consistency with YouTube's broad function (Haridakis & Hanson, 2009;Kim et al., 2016).TVer showed a substitution effect on recorded viewing, indicating that catch-up streaming substitutes recorded viewing as a superior form of time-shifted viewing.
The lesser degree of substitution observed for live viewing and news by PVODs, when compared to recorded viewing and other genres, somewhat upholds the utilities of linear format (Horst et al., 2018;Webster, 2009) and TV as information media (Leiner & Neuendorf, 2022).However, media substitution theory implies that streaming could functionally displace linear TV even in live viewing, news, and sports, through its functional expansion.Therefore, linear streaming services could substitute linear TV more broadly and rapidly than could on-demand streaming.Furthermore, YouTube's broad substitutional effect suggests that even on-demand streaming can function similarly to linear TV."Linear" is a format, not a function.For every player in the media industry, the critical factor for success on CTV should be comprehending the essential functions that the traditional linear TV has provided.They should then endeavor to effectively fulfill such functions via the Internet.

Complementary Effects of Catch-Up Streaming
TVer had complementary effects on live viewing beyond series programs for which catch-up effects were expected.This result extends that of Belo et al. (2019), who found complementarity between live and time-shifted viewing in the entertainment genre and mentioned catch-up effects.We believe that two types of complementary effects caused this result.The first is a shortterm spillover effect.TV viewers tend to watch the same channel when live viewing to avoid switching costs (Danaher & Dagger, 2012;Rust & Alpert, 1984).Therefore, even if TVer had direct complementary effects only on series programs, these effects would spill over to subsequent programming.The second is a long-term effect, namely the development of viewing habits and affinity to a channel.These hypothesized mechanisms of complementarities need to be tested via further research.

Limitations
This study had three main limitations.First, the coverage was limited to smart TVs from one manufacturer in Japan.Additionally, we could not include streaming devices such as Fire TV Stick.Therefore, it may not be possible to extrapolate our results to CTV as a whole or to other countries.
Second, there were limited granular categories for streaming and recorded viewing.Streaming could not be decomposed more granularly than at the services level, and recorded viewing could not be decomposed further, complicating the investigation into why the results for ABEMA and anime/tokusatsu did not fully align with our hypotheses.Furthermore, we could not determine whether the complementarity between TVer and live viewing existed between only the same genres or even between different program genres.
Third, there was a difficulty in demonstrating causality.Although twoway fixed-effects regression extracts longitudinal relationships, they differ from rigorous causal relationships.To investigate causality, it is necessary to use a dynamic model that incorporates the temporal order of variables (Paul, 2009) or an experimental study (Belo et al., 2019).

Conclusion
Focusing on the current cutting edge of media substitution, namely streaming on CTV, we analyzed a large and granular log dataset.Despite the limitations noted above, we made considerable progress toward comprehending the replacement of linear TV with streaming, which is currently underway worldwide.We demonstrated that this replacement is not uniform but involves multiple internal relationships resulting from functional similarities and complementary effects.This structural explanation supported media substitution theory and foreshadowed that ongoing functional expansion of streaming would further contribute toward linear TV's replacement.
Previously, streaming and linear TV were functionally different.However, the prevalence of broadband and CTV has allowed streaming services to offer content broadly and linearly, even on TV devices.That is, streaming now encompasses the functionality of linear TV.In light of this study, media industry players should utilize the broad functionality of streaming to fulfill the audience's purpose and gratification efficiently, even in the domain in which linear TV has excelled.Broadcasters also should not hesitate to functionally expand their streaming service even when cannibalization occurs because cannibalization means that streaming efficiently fulfills the same function as linear TV.
CTV's functionality is no longer limited to video.Cloud gaming is one example.Since time allocation does not arise only within video media (Jang & Park, 2016), we propose more comprehensive research as one future direction, including non-video activities.Our methodology could be applicable in other areas beyond media where there is resource allocation between new and traditional ways of life.
excluded the explanatory power of the fixed effects.

Figure 1 .
Figure 1.95% Confidence intervals for differences in the coefficients for each program genre of live viewing.Note.A confidence interval not containing zero indicates statistically significant differences in regression coefficients.Confidence intervals are wider for streaming services and program genres with lower viewing hours.
excluded the explanatory power of the fixed effects.

Table 1 .
Daily viewing hours by formats and services.SD includes variance between devices and excludes variance between quarters.Overall, 96.8% of the "Others" consists of using HDMI-connected devices, including recorders and gaming consoles.Streaming viewing through PCs and other devices is included here, although we assume that streaming viewing through HDMI-connected devices is small, given that smart TVs inherently possess streaming capabilities.Other services in the streaming category consist largely of video streaming services but include non-video services such as music streaming.

Table 2 .
Daily viewing hours of live viewing by genre(N = 197,273).

Table 3 .
Quarterly aggregation of daily viewing hours(N = 197,273).Total TV viewing time increased in 2020 when the COVID-19 pandemic began.The composition percentage of Others decreased from 20.4% to 17.2%.While we cannot break down data on this change, replacing gaming consoles and optical disc players with streaming might be one reason.

Table 4
(Xu et al., 2014)lts for live viewing between ALL and HDD groups should be mentioned.The HDD group exhibited more negative coefficients than the ALL group for all streaming services [columns (1) and (2) in Table4].We presume the reason was the longer TV viewing time in the HDD group, as shown in Table1because complementarity is weaker for more time-constrained households(Xu et al., 2014).

Table 4 .
Results of the estimation model for live and recorded viewing.

Table 5 .
95% Confidence intervals for differences in the coefficients for live and recorded viewing.
Device Group and Dependent Variable

Table 6 .
Results of the estimation model for each program genre of live viewing.