Do Ads Harm News Consumption?
DDo Ads Harm News Consumption? * Shunyao Yan, Klaus M. Miller, Bernd Skiera † Goethe University Frankfurt May, 2020
Abstract:
Many online news publishers finance their websites by displaying ads alongside content. Yet, remarkably little is known about how exposure to such ads impacts users’ news consumption. We examine this question using 3.1 million anonymized browsing sessions from 79,856 users on a news website and the quasi-random variation created by ad blocker adoption. We find that seeing ads has a robust negative effect on the quantity and variety of news consumption: Users who adopt ad blockers subsequently consume 20% more news articles corresponding to 10% more categories. The effect persists over time and is largely driven by consumption of “hard” news. The effect is primarily attributable to a learning mechanism, wherein users gain positive experience with the ad-free site; a cognitive mechanism, wherein ads impede processing of content, also plays a role. Our findings open an important discussion on the suitability of advertising as a monetization model for valuable digital content.
Keywords: online advertising, news consumption, ad blocking, monetization of content, difference-in-differences * The authors thank participants of the 2019 EMAC Conference, 2019 EMAC Doctoral Colloquium, 2019 SCECR Conference, 2019 JAMS Thought Leaders' Conference on Innovating in the Digital Economy, 2019 Marketing Science Conference, 2019 ZEW Conference on the Economics of Information and Communication Technologies, AMA CBSIG 2019, 2019 China Marketing International Conference, WISE 2019, VHB 2020, and seminar participants at Goethe University Frankfurt, LMU Munich, Queensland University of Technology, National Central University, National Tsing Hua University, Adblock Plus / Eyeo, Tel Aviv University, and Hebrew University of Jerusalem for helpful comments. All remaining errors are our own. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement No. 833714). † Department of Marketing, Faculty of Business and Economics, Theodor-W.-Adorno-Platz 4, 60323 Frankfurt am Main. Email: [email protected]; [email protected]; [email protected].
News is essential to our daily life. It contributes to shaping voters’ political attitudes (Strömberg 2015), investors’ financial expectations (Tetlock 2015), and individual’s health-related awareness (DellaVigna and La Ferrara 2015). And it is well established that well-informed individuals constitute a functioning society (Feddersen 2004). Recent evidence shows, however, that individuals’ tendency to consume and share low-quality news (e.g., fake news) may far exceed their tendency to consume and share high-quality news (Vosoughi et al. 2018), a phenomenon that may have negative social consequences. Accordingly, it is of interest to identify factors that might encourage individuals to consume high-quality news. This paper addresses this question by focusing on a factor that, to our knowledge, has received little consideration as a potential determinant of news consumption: exposure to ads that are featured alongside news content. Given that bundling ads with content is a common means of financing news sites, it is perhaps surprising that very little is known about whether and how ads impact news consumption online. Prior research points to three plausible relationships between exposure to ads and news consumption: First, since ads can provide information that benefits consumers (e.g., informing consumers about the existence of a particular product and its price), it is possible that the presence of ads makes news sites more attractive and thus enhances news consumption (Kaiser and Song 2009). Second, as users can easily bypass ads on websites (e.g.,
Drèze and Hussherr (2003))—a well-documented phenomenon known as “banner blindness”—ads might not affect the consumption of online news at all. Indeed, prior studies of online and print news consumption have tended to assume that ads do not affect the reading experience (Aribarg and Schwartz 2019; Pattabhiramaiah et al. 2018). The third possibility is that ads, particularly those that incorporate animation or are otherwise intrusive, might annoy users and distract them from news reading, thereby affecting news consumption negatively. Goldstein et al. (2014) used lab experiments on MTurk to find support for this effect. Yet, the capacity to generalize the results of controlled lab experiments to real-world news consumption behavior is highly limited—not least because of discrepancies in the nature of the ads to which consumers are exposed. For example, in the real world, ads are typically targeted to consumers, potentially providing them with more benefits but at the same time raising their privacy concerns compared with ads that have been generated for an experiment. Clearly, it is necessary to obtain real-world evidence regarding the relationship between ad exposure and news consumption. However, obtaining such evidence is a highly challenging empirical task. For example, a randomized field experiment in which a website exposes only some users to ads is likely to be highly costly, particularly when the experiment is implemented for a reasonable amount of time, and it may also be limited in terms of realism (e.g., by providing some users with the mistaken impression that the website itself does not incorporate ads). Analysis of actual browsing behavior, in turn, presents substantial identification challenges. For example, simply comparing the behavior of users who view ads on a site with that of users who do not (e.g., because of use of an ad blocker) raises self-selection concerns, making it difficult to determine how users who do view ads would have behaved in their absence, and vice versa.
Herein, we introduce the first real-world empirical study that examines how exposure to ads on a news site affects individuals’ news consumption on that site. To overcome the identification challenges outlined above, we exploit a unique individual-level panel dataset, which contains 3.1 million browsing sessions of 79,856 users on a news website. Our dataset provides information about users’ news reading behavior, their usage of ad blockers, and the timing of their adoption (or dis-adoption) of ad blockers. In effect, these data enable us to construct a “counterfactual” world without ads. Specifically, we compare similar users who adopt ad blockers at different points in time and analyze differences in their news consumption before and after their adoption. We suggest that the decision to adopt an ad blocker and the timing of doing so are likely to be independent of a user’s prior news consumption behavior. This assumption is grounded in the following reasoning: First, the decision to adopt an ad blocker has been shown to be driven by factors such as advertising annoyance, page loading speed, and privacy concerns regarding targeted ads, none of which is directly related to a user’s news consumption behavior on a specific news website (Vratonjic et al. 2013). Second, when a user installs an ad blocker in his or her browser, ads are turned off automatically on all websites that the user visits after adoption and not just on one particular website. A rare exception of this rule is whitelisting, which we can observe thus control for in our analysis. Thus, adoption of an ad blocker can be assumed to be independent of a user’s activity on a particular website as long as it is not that website that triggers a user’s decision to adopt an ad blocker. As our news website and its competitors did not change the way they displayed ads or reported news during our observation period, it is unlikely that our news website triggered users’ decisions to adopt ad blockers at specific points in time. Given that adoption of an ad blocker is likely to be independent of a user’s prior consumption behavior, we suggest that such adoption can be considered as an exogenous shock in terms of its effect on subsequent news consumption. Thus, we obtain a quasi-experimental setting to isolate the impact of seeing versus not seeing ads on news consumption. Nevertheless, it is necessary to account for the possibility that news consumption might be correlated with other individual-level characteristics that affect users’ ad blocker adoption decisions (e.g., age, education, or unobserved taste towards news). To address this concern, we follow previous work by
Datta et al. (2018) and use an analysis approach that combines matching and difference-in-differences (DiD), taking advantage of the granular information on users of our large panel dataset to remove all observed and time-invariant unobserved confounders. In our main analysis, we focus on users who adopt an ad blocker after a period of being exposed to ads, where users who are continuously exposed to ads serve as the control group. To address remaining concerns regarding unobserved confounders, we carry out two additional analyses using the same identification strategy, but with alternative constructions of treatment and control groups. In the first of these analyses, we include only users who adopted ad blockers, exploiting the variation created by differences in the (plausibly exogenous) timing of adoption. In the second analysis, we consider ad blocker dis-adoption (i.e., abandonment), rather than adoption, as treatment. The latter analysis enables us to support the causality of the effect we investigate.
Our analyses, which control for self-selection, reveal that adoption of an ad blocker (i.e., not seeing ads) leads to a 20% increase in the quantity of news consumption (i.e., number of article views) and a 10% increase in the variety of news consumption (i.e., number of news categories to which the viewed articles correspond). The effects are driven by an increase in consumption of hard news (defined as political news, economic news and opinion news; see
Angelucci and Cagé (2019)). We further seek to shed light on the mechanism of the effect, testing two theoretical explanations. First, we examine a “cognitive effect”, in which the diminished ad consumption of users who view ads is attributed to the fact that users consciously process ads or are subconsciously distracted by them (Kahneman 1973; Vakratsas and Ambler 1999). Second, we look for a “learning effect” in which users who do not view ads enjoy and benefit from the positive website experience (Johnson et al. 2003) . We distinguish these two mechanisms and find evidence for both of them, with the cognitive effect being small in magnitude, whereas the learning effect is substantial and persistent. Our results also highlight substantial heterogeneity in the effect of ad exposure across different users: First, users with a stronger tendency to read news on their mobile phones (as opposed to on desktop devices) exhibit a stronger treatment effect. This result may indicate that annoyance caused by ads, which is likely to be more intense on a smaller screen, may play a role in the effects observed. Second, users with an older Java version exhibit a stronger treatment effect, suggesting that page loading speed, which tends to be slower on older Java versions, may also contribute to the ad effect. Third, light users exhibit a stronger treatment effect, which provides implications for news publishers who wish to engage users and to turn them into heavy and even subscribed (i.e., paying) users. The remainder of the manuscript proceeds as follows.
In the next section, we review the related literature. Then, we introduce our empirical setting and dataset and explains how we construct our main variables. Next, we present our identification strategy, followed by the empirical results. In the end, we conclude the manuscript.
LITERATURE REVIEW
Our study draws from and contributes to three main streams of literature. The first is the broad literature on the effects of advertising. Although advertising has been studied extensively in economics and marketing, prior studies have tended to focus on the success of advertising for advertisers, as captured by measures such as recall and recognition of ads (Bagozzi and Silk 1983), click-through rate (Dinner et al. 2013), willingness-to-pay (Goldfarb and Tucker 2011), sales (Blake et al. 2015), and brand awareness (Joo et al. 2014). Much less research has examined how ads affect the platforms that publish them (“publishers”) and users’ engagement with those platforms. Much of the empirical work in this vein has focused on non-digital markets (e.g., traditional TV, magazines, and yellow page books), documenting both positive and negative effects of ads on media consumption (Kaiser and Song 2009; Rysman 2004; Wilbur et al. 2013). However, the formats and delivery of ads in non-digital markets differ substantially from those in digital markets. Studies on digital advertising, in turn, have primarily taken place in highly controlled lab settings and thus may be of limited relevance to real-world consumption (e.g., the work of
Goldstein et al. (2014) discussed above). Our paper complements previous studies by examining the effect of banner ads on the behavior of users on a real news website. Notably, unlike lab experiments, our work is able to capture long-term effects. A notable exception is the work of Sahni and Zhang (2020), who carried out a field experiment to examine how the prominence of ads on a search engine affected usage of that search engine. The authors found that users responded positively to higher levels of search advertising, yet also noted that search advertising on a search engine is typically less annoying and more informative than banner advertising on a news website. Marketing researchers generally recognize that ads are designed to attract people’s attention and thus will have a cognitive impact (Hong et al. 2004; Vakratsas and Ambler 1999). In this vein, cognitive studies provide a theoretical foundation to understand the relationship between exposure to ads and news consumption. The central capacity theory proposes that people’s working memory has limited cognitive capacity and simultaneous tasks compete for the limited cognitive resources (Kahneman 1973). When people view ads and news at the same time, part of the cognitive resources will process the information of ads or suppress the distraction of ads, thereby leaving less cognitive resources for processing the news. The effect of the limited cognitive capacity is especially strong when the simultaneous tasks are irrelevant(Hong et al. 2004), as is often the case with online news and banner ads, resulting in a potential negative relationship between exposure to ads and news consumption. A complementary theory lies in experiential learning, the process that people learn from their usage experience (Lin et al. 2015). The learning literature suggests that consumers have uncertainty about the utility they obtain from a product (i.e., news consumption in our case). Consumers update their belief of the expected utility based on their product usage experience (Lovett and Staelin 2016; Roos et al. 2020) and the future consumption decision is based on the expectation of the updated belief. A positive and undistracted usage experience can thus help consumers in the process to form the belief (e.g., on how well the news website matches their preference). In this case, the exposure to banner ads on the news website can have a momentum effect that not only negatively impacts the current consumption but also the future consumption. To the best of our knowledge, all studies on learning and advertising analyze how advertising helps consumer learning of the advertised product, leaving the negative learning effect of advertising on the media that carries the ads unexplored (for a review, see (Ching et al. 2017)). Second, we contribute to the understanding of ad blockers, a new technology that is generally considered as a threat to the publishing industry and even the internet as a whole (Shiller et al. 2018). Early studies in this stream theoretically modeled publishers’ and advertisers’ strategic responses to ad blocking, such as changing the levels of advertising or ad targeting rules. These studies concluded that when some consumers use ad blockers, ad cluttering increases for ad blocker non-adopters (Anderson and Gans 2011; Johnson 2013). Shiller et al. (2018) offered the first empirical investigation into how ad blockers affect publishers. Using website-level data, the authors showed that an increase in the use of ad blocking among a website’s users is associated with a reduction of the website’s traffic over 35 months. They proposed that this effect was a result of the fact that the lost revenue 0 attributable to ad blockers led the publisher to reduce its investments in website content. Additional studies, relying on surveys, have explored the potential reasons for users to adopt ad blockers (Vratonjic et al. 2013). Herein, we use individual-level data to document how ad blockers affect user behavior, and point to ways in which publishers might draw from this knowledge to improve their own outcomes. Third, our research joins a growing literature and an ongoing debate about news supply and demand. In the domains of economics and marketing, studies on the supply of news focus primarily on evaluating pricing models for the content itself (e.g., paywalls and subscription fees; Lambrecht and Misra (2017); Pattabhiramaiah et al. (2019)). Discussions and research on the demand for news, in turn, explore questions related to the nature of the content that people seek out. Studies in this vein investigate why people demand low-quality news (e.g., fake news) or news that only covers limited viewpoints (e.g., echo chambers). These studies have attributed such demand patterns to influences related to social media (Scharkow et al. 2020; Schmidt et al. 2017) among other factors. This paper, which focuses on advertising rather than on monetary payment for content, bridges between the supply and demand facets of this literature by showing that a publisher’s reliance on ads can also affect users’ patterns of engagement with the content—and specifically, the quantity and variety of news that they consume. Thus, our work can contribute to the discussion on whether and how to use advertising to finance news or other socially valuable digital content.
EMPIRICAL SETTING, DATASET, AND VARIABLE CONSTRUCTION We rely on a unique internal and proprietary dataset from a highly reputable European news publisher who prefers to remain anonymous. Our news publisher runs an online news website that publishes daily news, with a focus on politics and business while reporting a variety of other topics. The news website ranks among the top 10 in its country in terms of weekly usage, with over 3.6 million weekly page impressions (total clicks). At the time of our study, around 78% of the traffic to the news website came from its own country. In the industry, our news publisher has long been regarded as a national “newspaper of record” and its reputation in its linguistic area is comparable to that of the
New York Times , the
Financial Times , or the
Guardian . During the period of our analysis, all content on the news website was offered free of charge to all users. Users were required to register with the website (i.e., to enter their email addresses) to access archival content and content newsletters, but were not required to pay for this content. Approximately 20% of visits on the website come from registered users. Our dataset was composed of clickstream data for all registered users who visited the news website from the second week of June, 2015 (week 1) to the last week of September, 2015 (week 16). We focus on registered users for both econometric and socio-economic reasons. We can only track registered users on the individual-level over time, which provides us with a unique panel setting that we use for our empirical analysis. In addition, data on registered users effectively provide information on individuals who are likely to value high-quality news sources and thus to be more politically engaged (Prior 2007). Thus, their beliefs based on news consumption are more likely to shape the political and social discourse. Users 2 were anonymized. The clickstream data for each registered user included a full record of that user’s browsing activities (including, among others, the time stamp, the page views, and whether an ad blocker is used) on the news publisher’s website over the course of the data collection period. We further combined the clickstream data with self-reported user demographics from the publisher’s CRM database. In total, we analyzed 79,856 unique users with 3.1 million visit sessions.
Information About the Ads on the News Website
Our news website runs display advertising according to the standard advertising formats outlined by the Interactive Advertising Bureau (IAB) (IAB 2017). More precisely, our website runs leaderboard ads (728 × 90 pixels) on top of the page and rectangle ads (300 × 250 to 336 × 280 pixels) in the middle of the page on both desktop and mobile devices. In addition, our website runs skyscraper ads (120 × 600 pixels) on the side of the page on desktops. On average, there are five display ads on the homepage and three display ads on an article page. These levels of advertising are comparable to, and in some cases even lower than, the levels of advertising on other similar premium news websites such as the
New York Times , the
Washington Post and the
Guardian . Moreover, the ads displayed on our news website are less annoying (i.e., less animated, with better aesthetics, and provided by more reputable advertisers; see (Goldstein et al. 2014)) compared with those featured on tabloid news websites such as the
Daily Mail , the
Sun , or the
National Enquirer . We further note that our news website does not run half-page ads (300 × 600 pixels) or large mobile banner ads (320 × 100 pixels), whose removal by an ad blocker could lead to substantial changes in the 3 display of the content. In addition, our website did not run native advertising during the observation period of our study.
Information on Ad Blockers
In the European countries, 20%-38% of internet users used an ad blocker during our observation period (Newman et al. 2016). An ad blocker is usually an extension that a user downloads on her browser and that, in most cases, automatically removes all ads on every webpage the user visits—with the exception of websites that the user has whitelisted. Some websites require users to disable their ad blockers in order to view content; neither our website nor any of its main competitors did so during the observation period. Since ads are used extensively online, installing an ad blocker can lead to a noticeable difference in the display of a website, as shown in Figure 1, and the respective user experience. [Insert Figure 1 about here] Crucial to our identification strategy are two features of the ad blocker. First, using an ad blocker removes all ads on the website and thus creates a “counterfactual” of the original version of the website that displays ads. Second, an ad blocker is adopted on the browser and thus is not targeted to remove ads for a specific website. Specifically, a user’s decision to adopt an ad blocker is likely to be motivated by factors that are unrelated to her prior news consumption behavior. For example, the user may learn accidentally about the existence of ad blockers at different points in time, or could be triggered by seeing (annoying) ads on some random other website.
Thus, adoption of an ad blocker can generate exogenous variation in 4 news consumption on our focal news website, which we exploit for our identification strategy. Next, we introduce how we construct our key variables—corresponding to ad blocker adoption and news consumption—which we use to empirically test how exposure (or non-exposure) to ads impacts news consumption.
Independent Variables – Ad Blocker Adoption
Our dataset indicates, for each user daily browsing session, the number of page impressions that were blocked using an ad blocker. We use this information to derive an indicator of ad blocker usage. Specifically, a non-zero number of blocked page impressions indicates that the user is implementing an ad blocker, whereas zero impressions blocked indicates non-ad blocker usage (i.e., exposure to ads). Whereas our dataset contains records of news consumption from week 1, ad blocker usage was recorded only from week 10 onwards. Of the 79,856 users whom we observed, 19,088 users used an ad blocker during this period (as indicated by a non-zero number of page impressions blocked), and 60,768 users did not use an ad blocker during this period. Thus, 24% of users in our dataset used an ad blocker; this percentage is comparable to the ad blocking adoption rates across European countries at the same time, ranging from 20% in Italy to 38% in Poland (Newman et al. 2016). As discussed above, we suggest that, because adoption of an ad blocker is unlikely to be driven by news consumption behavior per se, such adoption can serve as source of exogenous variation in exposure to ads on a specific news website. We construct our main treatment 5 measure—ad blocker adoption—according to whether a user experienced a zero to non-zero change in the number of ads blocked. The fact that ad blocker usage is only recorded from week 10 onwards creates a potential left-censoring problem: A user who we observe using an ad blocker in week 10 could already use it back in week 9 or earlier, which we do not observe. Thus, such a user would not experience a change in her exposure to ads. Following an approach to address a similar problem in the context of adopting Spotify (Datta et al. 2018), we designate a two-week cut-off period (week 10 and week 11) for defining our group of ad blocker adopters. Specifically, users who have zero ad blocker usage in week 10 and week 11 (i.e., the first two weeks of observed ad blocker usage) and then have non-zero ad blocker usage during week 12 or later are classified as ad blocker adopters. These ad blocker adopters constitute the treatment group for our main analysis. In other words, we classify users into the treatment group only if their first ever observed ad blocker usage occurred during week 12 or later. We classify a user into the control group, in turn, if the user had no ad blocker usage throughout weeks 10–16. According to this construction, our treatment group comprised 6,366 users and our control group comprised 38,270 users (see the top part of Figure 2 for an illustration of our construction of the treatment and control groups). For robustness, we carry out two additional analyses using alternative definitions of treatment and control groups. Our second analysis compares early adopters of an ad blocker (treatment group) to late adopters (control group). An early adopter is defined as a user who adopts an ad blocker in week 12 ( n = 1,124). A late adopter is defined as a user who adopts an ad blocker in week 14 ( n = 1,167). This approach enables us to control for bias related to 6 the decision to adopt an ad blocker at all, given that both treatment and control groups adopted an ad blocker and only differed in the timing at which they did so (that is, early adopters adopted two weeks earlier than late adopters). Our third analysis leverages the fact that, in our sample, 9,055 of the total 79,856 users had already adopted an ad blocker during our cut-off period (weeks 10 and 11). Thus, they are censored users for whom we cannot identify the timing of adopting an ad blocker. However, for these users, we can identify the timing at which they dis-adopted an ad blocker—based on their non-zero to zero change in ad blocker usage. We use these censored adopters to identify the effect of abandonment of ad blockers. Specifically, we classify these censored adopters into the treatment group if they undergo a change from non-zero to zero ad blocker usage during week 12 or later. In other words, the users in this treatment group ( n = 2,882) do not see ads in weeks 10 and 11 but start to see ads during week 12 or later. The control group ( n = 6,173) consists of users in this censored sample who have non-zero ad blocker usage throughout weeks 10–16. The bottom part of Figure 2 depicts the construction of these treatment and control groups. [Insert Figure 2 about here] Dependent Variables – News Consumption
Our analysis considers a wide range of news consumption measures, which we summarize in Table 1. We report all measures and the corresponding analyses at the user-week level. Of primary interest are two main measures: The first is article views, i.e., a count of the number of news articles a user clicks on the news website; this measure captures the 7 quantity of news consumption. The second is breadth, a count of the number of unique news categories of article views; this measure captures the variety of news consumption. In addition, we analyze the mechanism underlying the effect of ad exposure. To this end, we decompose the effect on article views by visit. A visit is an entry to our news website and ends with user inactivity for 30 minutes, consistent activity for 12 hours, or other activities that indicate robot behavior. We decompose the number of article views for each user into two components: the number of article views per visit and the number of visits. The product of these two components yields again the original measure: the number of article views. An effect of treatment on the number of article views per visit is likely to indicate a cognitive effect, wherein seeing (vs. not seeing) ads influences the extent to which the user is able to cognitively process website content. An effect on the number of visits, in turn, may indicate a learning effect, in which ad exposure influences the user’s experience with the website and desire to return. To further investigate the cognitive effect, we scrape all the news articles that users click on and check the number of words per article (i.e., length of the article) and the number of words per sentence, a readability measure suggested by Loughran and McDonald (2014). To further understand the learning effect, we classify visits into direct visits (i.e., users navigate to the website directly) or referral visits (i.e., users are referred from social media, search engine, or newsletter). To further understand the main effect of ad exposure on article views, we also count the number of article views separately for each news category (e.g., political news and economic news). We report all news categories of our website in Panel 2 of Table 1. In addition, we 8 count the number of page views on the home page. The page views from news articles and the home page account for more than 90% of the browsing activities during our observation period. Other browsing activities on our news website include browsing account-related pages and the weather forecast. [Insert Table 1 about here] Recall that our main analysis is based on two groups of users: ad blocker adopters (treatment group) and non-ad blocker adopters (control group). We report in Table 2 the group means of our main dependent variables: article views and breadth, a before-and-after difference within groups, and a DiD between groups. We start with the before-and-after analysis. For treated users (ad blocker adopters), all news consumption measures increase after treatment, but these measures decrease for the control group (non-ad blocker adopters), leading to a positive value when we compute a simple DiD estimator (e.g., +1.89 for article views and +0.62 for breadth). These results provide preliminary evidence of the positive effect of not seeing ads (by adopting an ad blocker) on news consumption, in terms of both quantity and variety. However, the DiD reported in Table 2 is not meant to represent a causal effect. In particular, we observe that ad blocker adopters differ from non-adopters in the quantity of news consumption in the pre-treatment period—specifically, adopters read more than non-adopters, suggesting that self-selection occurs. [Insert Table 2 about here] In the next section, we describe our approach, inspired by (Datta et al. 2018) , to disentangle any potential bias from self-selection. To identify the causal effect, we combine 9 matching with DiD and establish the robustness of the result by repeating the analysis with the alternative treatment definitions defined previously.
IDENTIFICATION STRATEGY
Selection bias into treatment can come from both observable and unobservable confounders. In our analysis, we first non-parametrically control for observable confounders using coarsened exact matching (CEM). Then, to remove any time-invariant unobserved confounders, we use DiD with individual-level fixed effects. As for time-varying confounders, we use a placebo treatment test to show that they do not bias our results. In what follows we describe this identification strategy for our main analysis, in which ad blocker adoption at week 12 and onward (as defined above) serves as a treatment. We use a similar strategy in our two robustness analyses, in which, respectively, early adoption and dis-adoption of an ad blocker serve as treatment (see above). Recall that our sample covers 16 weeks, running from June 8, 2015 to September 27, 2015. Our treatment starts from week 12. We use the first 11 weeks, i.e., the entire pre-treatment period, for matching. For our estimation, we use weeks 7 to 11 as the pre-treatment period and the remaining weeks, weeks 12 to 16, as the post-treatment period to obtain balanced numbers of pre- and post-treatment periods. We also use weeks 1 to 11 as a pre-treatment period in an additional robustness check, reported in Web Appendix Table S9, which confirms the robustness of our results. 0
Coarsened Exact Matching
To remove observable confounders, we use matching to refine our treatment and control groups (Heckman et al. 1998). Matching methods are commonly combined with DiD in the statistical treatment effects literature (e.g., see Datta et al. (2018)). Among the different matching techniques, we chose CEM for its advantages over other methods (such as propensity score matching; see King and Nielsen (2019)) in terms of balancing the covariates. We also use propensity score matching as a robustness check (Web Appendix Table S5) and the results remain similar. Specifically, we match the treatment and control groups on the basis of user demographics, pre-treatment browsing activities, and the beginning and end of each user’s observation window (Datta et al. 2018). We include three controls for demographics—age, gender, and income—as prior empirical studies show that these factors are important determinants of news consumption (Fan 2013).
We note the tradeoff in using demographic data for matching (introducing a selection problem, as those who report their demographics in CRM data can be different from those who do not report). In the Web Appendix Table S4, we check that missing value of demographics is random. We also report the estimation using unmatched data as robustness checks. The results remain similar. To capture a user’s pre-treatment browsing activities, we use page views, breadth, and time per visit. To make sure users are active throughout the same observation window, we include a user’s first and last observed week in our sample. [Insert Table 3 about here] 1
Table 3 compares the matched and unmatched samples of the treatment group (ad blocker adopters) and the control group (non-ad blocker adopters) in terms of the observed characteristics that we used for matching. The first three variables (gender, income, age) heading down the left part of Table 3 show that the treatment group and control group are similar in terms of the demographics before matching. The next five variables show that treatment group is more active online than the control group: ad blocker users tend to stay longer with the website, spend more time on each visit, and view more pages than non-ad blocker users. The right part of Table 3 shows that, after matching, the treatment group and the control group are balanced in terms of all their demographics and browsing activities. CEM removes the differences in all of the observed controls, and thus any potential remaining selection bias can only come from unobservable confounders, which we discuss in the next section. [Insert Figure 3 about here] Figure 3 depicts the distribution of the propensity scores, the probability of being treated predicted by the matching variables, of the treatment group and the control group before and after CEM. The distribution shows that CEM further removes from the sample the users who are the least likely to adopt an ad blocker. Thus, in effect, we mimic an experimental setting in which users in the treatment and control group are equally likely to adopt an ad blocker but decide randomly whether to adopt it or not. 2
Difference-in-Differences (DiD)
Having produced our matched control and treated samples, we subsequently apply DiD with individual-level fixed effects, which removes time-invariant unobserved confounders by taking the temporal differences within each user. We eliminate all variation in news consumption caused by time-invariant unobserved heterogeneity between individuals (e.g., education and preference towards news or ads). In addition, DiD removes any bias due to time trends that are common to both groups (e.g., resulting from seasonality or news shocks) by taking the difference once again across groups. Specifically, we estimate the following DiD model: (1) 𝑌𝑌 𝑖𝑖𝑖𝑖 = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 , where 𝑌𝑌 𝑖𝑖𝑖𝑖 is the dependent variable, one of the news consumption measures listed in Table 1, for user 𝑤𝑤 in week 𝑤𝑤 ; 𝛼𝛼 𝑖𝑖 is a user fixed effect, controlling for time-constant differences across users, such as education or tastes towards news; 𝛿𝛿 𝑖𝑖 is a week fixed effect, controlling for common trends or changes over time that affect all users equally, such as breaking news; 𝐼𝐼 𝑖𝑖𝑖𝑖1 is an indicator variable that is equal to one if the observation for individual 𝑤𝑤 in week 𝑤𝑤 is within 1 week after the treatment (so that this binary variable is one in the treated week and the following week); 𝐼𝐼 𝑖𝑖𝑖𝑖2 is an indicator variable that is equal to one if the observation of individual 𝑤𝑤 in week 𝑤𝑤 is after the treatment (so that this binary variable is 1 in the treated week and all subsequent weeks); 𝛽𝛽 and 𝛽𝛽 distinguish an immediate effect and a “permanent” effect of the treatment (i.e., an effect that persists throughout the entire post-treatment period); and 𝜀𝜀 𝑖𝑖𝑖𝑖 is the error term, clustered at the user level. 3 Our identification strategy builds upon the changes in news consumption that occur after treatment (in our main analysis, adoption of an ad blocker, after which users no longer see ads). DiD compares these changes to those that occur for comparable users who are not treated (in our main analysis, users who never adopt an ad blocker). Thus, this identification strategy uses the trend in news consumption of the control group as a counterfactual for the trend in news consumption of the treatment group. Crucial for the identification is that all confounders are either controlled for or quasi-random; that is, any unobserved time-varying confounders follow parallel trends in the pre-treatment periods. We visually inspect this assumption by observing pre-treatment and post-treatment trends (Figure 4 presents a graph for our main analysis, with article views as the dependent variable). In the Web Appendix, we formally test this identification condition for all dependent variables using a placebo treatment test (Angrist and Pischke 2008). [Insert Figure 4 about here] Recall that the treatment starts from week 12 in our sample. In Figure 4, our treatment and control group experience very similar trends before week 11 (red dotted line). However, news consumption in the treatment group increases more strongly than that in the control group after week 12.
Heterogeneous Treatment Effects
We next explore whether treatment effects differ across individual users. One goal in doing so is to validate our argument that the decision to adopt an ad blocker is not driven by news consumption behavior per se but rather by other factors related to the effects of ads on 4 the web browsing experience. Accordingly, in this step of the analysis, in addition to user demographics, we focus on user characteristics that might serve to capture factors that prior literature has identified as drivers of the decision to adopt an ad blocker. Specifically, drawing from the results of previous surveys (Newman et al. 2016; Vratonjic et al. 2013), we identify three main reasons that users often report for adopting an ad blocker: the annoyance of ads, page loading speed, and privacy concerns. We test whether these three factors influence the treatment effects we investigate. To test the role of annoyance, we examine whether the treatment effect is stronger for users with a stronger tendency to use mobile (rather than desktop) devices; we assume that ads are likely to be more annoying on mobile devices, which tend to have smaller screen sizes than desktops. For page loading speed, we take into account users’ Java versions, and examine whether users with older Java versions, which are assumed to load pages more slowly than newer versions, are more strongly affected by treatment. Finally, to test the role of privacy concerns, we use an indicator of whether a user has ever rejected a cookie, under the assumption that users who have rejected cookies have greater privacy concerns and thus may be more sensitive to treatment. To examine the heterogeneous treatment effect, we interact these three user characteristics (mobile usage, java version, and cookie setting in the pre-treatment period) with the treatment indicator. In addition, we interact the treatment indicator with a variable indicating the extent to which, during the pre-treatment period, the participant was a light or 5 heavy user of the website (i.e., visited the website frequently). Thus, we estimate the following fully saturated model: (2) 𝑌𝑌 𝑖𝑖𝑖𝑖 = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝐴𝐴𝑖𝑖𝑖𝑖𝑜𝑜𝐴𝐴 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝑃𝑃𝑇𝑇𝑤𝑤𝑟𝑟𝑇𝑇𝑠𝑠𝐴𝐴 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝑆𝑆𝑆𝑆𝑤𝑤𝑤𝑤𝑆𝑆 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝐿𝐿𝑤𝑤𝐿𝐿ℎ𝑤𝑤 𝑖𝑖 + 𝜀𝜀 𝑖𝑖𝑖𝑖 , where 𝑌𝑌 𝑖𝑖𝑖𝑖 is the news consumption measure for user i in week t ; 𝛼𝛼 𝑖𝑖 is the individual-level fixed effect; 𝛿𝛿 𝑖𝑖 is a time fixed effect at the week level; 𝐼𝐼 𝑖𝑖𝑖𝑖2 is an indicator variable that is equal to one if the observation of individual 𝑤𝑤 in week t is after the treatment (similar to 𝐼𝐼 𝑖𝑖𝑖𝑖2 in equation (1)); 𝐴𝐴𝑖𝑖𝑖𝑖𝑜𝑜𝐴𝐴 𝑖𝑖 is coded as 1 if the mobile usage of a user i is above the average mobile phone usage (12% of the page impressions taken place on a mobile phone); 𝑃𝑃𝑇𝑇𝑤𝑤𝑟𝑟𝑇𝑇𝑠𝑠𝐴𝐴 𝑖𝑖 is coded as 1 if user i has not always accepted a cookie; 𝑆𝑆𝑆𝑆𝑤𝑤𝑤𝑤𝑆𝑆 𝑖𝑖 is coded as 1 if the Java version of user i is below the median Java version (Java 1.6); 𝐿𝐿𝑤𝑤𝐿𝐿ℎ𝑤𝑤 𝑖𝑖 is coded as 1 if user i ’s number of visits in the pre-treatment period is below the median (4 weekly visits); We use dummy coding for all these variables; and 𝜀𝜀 𝑖𝑖𝑖𝑖 is the standard error clustered at the user level. RESULTS
Main Effect on Quantity and Variety of News Consumption
We are primarily interested in two measures of news consumption: the number of article views (i.e., the quantity of news consumption) and the number of news categories to which viewed articles correspond (i.e., the variety of news consumption). Table 4 reports these results for our main analysis as well as for our two robustness analyses with alternative treatment and control group designs (as described in section 3.3). The dependent variables are the natural logarithms of news consumption measures; thus, a simple transformation of β i in 6 regression model (1) can be directly interpreted as percentage change of news consumption: exp( 𝜷𝜷 ) − reports an immediate effect and exp ( 𝜷𝜷 ) − reports a permanent effect. [Insert Table 4 about here] In our main analysis, we used ad blocker adopters as a treatment group and ad blocker non-adopters as a control group. We find a significant and consistent positive effect of ad blocker adoption (corresponding to a negative effect of exposure to ads) on both the quantity and variety of news consumption. After ad blocker adoption, the number of article views permanently increases by 15% (=exp(0.140)-1), with an additional immediate increase of 13.4% (=exp(0.126)-1) within one week of the treatment. The variety of news consumption, in turn, permanently increases by 8.8% (=exp(0.084)-1), with an additional immediate increase of 9.3% (=exp(0.089)-1). As reported in Table 2, ad blocker adopters on average read 11.46 news articles and 4.41 news categories per week in the pre-treatment period. Taken together, our estimates indicate that users read 1.7 (=11.46 * 15%) to 3.3 (=11.46 * (15% +13.4%)) more articles per week and 0.8 (=4.41 * (8.8%+9.3%)) more news categories in total without ads. Our second analysis, in which we compared early adopters (treatment group) with late adopters (control group), produced results consistent with those of our main analysis. Specifically, early adopters increased the quantity of their news consumption by 27.6% (=exp(0.244)-1) and increased the variety of their consumption by 16.6% (=exp(0.154)-1) within 1 week of adoption. In this analysis, we are comparing the difference of news consumption between early and late adopters when early adopters adopt but late adopters not adopt an ad blocker (i.e., week 12 and week 13) with the difference of their news consumption 7 when neither early adopters nor late adopters adopt an ad blocker (i.e., before week 12). In this approach, however, we cannot estimate a permanent effect because the observation period between early and late adoption is too short. Our third, complementary analysis, in which ad blocker abandoners served as the treatment group and users who used ad blockers throughout the entire observation period (weeks 10–16) served as a control group, provides further robustness to our findings. Specifically, we find that users who abandon ad blockers decrease their quantity of news consumption by 10.2% (=exp(0.097)-1) and the variety of their news consumption by 7.9% (=exp(0.076)-1); both effects are permanent. In absolute terms, the size of this effect is similar to the effect sizes obtained in our previous approaches. These findings support the causality of the effect of ad exposure on news consumption behavior. Decomposition of the Main Effect: Article Views by News Category
Having established the robustness of the effect of exposure to advertising on the quantity and variety of news consumption, we decompose in Table 5 the effect on article views into different news and non-news categories. For clarity of presentation, in what follows we only report the results of our main analysis, with ad blocker adopters as the treatment group and non-adopters as the control group. The results obtained with our other approaches (using ad blocker early adoption or ad blocker abandonment as treatment) are largely similar and are reported in the Web Appendix. [Insert Table 5 about here] 8
We find that our treatment effect on the consumption of hard news (i.e., political, economic, and opinion news) persists over time. For most soft news categories (e.g., lifestyle and art & culture), the effects vanish in the long run though there are immediate effects. In addition, ad blocker adoption does not affect user consumption of non-news article pages (e.g., account settings and play pages, which are games such as Sudoku or Mahjong). This finding indicates that what we picked up is not an activity effect (i.e., users being more active online after not seeing ads) but an effect on news consumption. One potential explanation for the increase in consumption of hard news is that our news website displays hard news content more prominently than soft news on its home page. To test this possibility, we reran the analysis while controlling for home page views. The results, shown in the Panel 2 of Table 5, indicate that the effect on hard news (i.e., regional political news and opinion news) is persistent.
Cognitive Effect: Number of Article Views by Visit
Previous studies have shown that ads have a cognitive impact on consumers, regardless of whether consumers pay attention to them (Vakratsas and Ambler 1999). The reason is that our brain processes information both consciously and subconsciously (Kahneman 1973). Thus, one explanation for an increase in news consumption after adoption of an ad blocker is the increased availability of cognitive resources attributable to not seeing ads. If this explanation holds, then we should expect increases in news consumption to happen within visit sessions (i.e., consumers should view more articles per visit and from more categories), because the cognitive system for processing the information is the working memory that 9 functions in the short-term (Baddeley 1992). We suggest that users are unlikely to be able to save these extra cognitive resources over prolonged periods of time, such that alternative realizations of heightened news consumption—e.g., increases in the number of visits to the website—are unlikely to be attributable to the cognitive mechanism. Panel 1 of Table 6 shows that the number of article views per visit increases only by 2.5% (=exp(0.025)-1) immediately and 2.4% (=exp(0.024)-1) permanently, which is much less than the immediate increase of 13.4% and the permanent increase of 15% of total article views (Table 4). The number of home page views per visit does not increase significantly as well. Taken together, these results suggest that the cognitive effect only explains a small part of the increase of news consumption. [Insert Table 6 about here] Another means by which the availability of additional cognitive resources (attributable to ad blocker adoption) might influence users’ news consumption behavior is by enabling them to read longer or more complex articles. To explore this possibility, we scraped all the articles that users viewed and analyzed their length (number of words per article) as well as their readability (number of words per sentence). Ad blocker adoption had only a short-term and small effect on the length and readability of the news article. We observed, however, that the time per visit increased rather large after ad blocker adoption, suggesting that not viewing ads might have enabled users to devote closer attention to the articles they read. 0
Learning Effect: Number of Visits
The fact that the number of article views per visit did not increase permanently after ad blocker adoption indicates that the number of visits to the website should be the driver behind the increase in the number of article views. Indeed, we find both an immediate and a permanent increase of website visits (see column “Visits” in Panel 2 of Table 6). These effect sizes are comparable to those corresponding to the effect of treatment on article views (Table 4), which further establishes the robustness of our main result. Learning, the process of acquiring knowledge and experience about a product, provides an intuitive explanation for repeated visiting behavior (Johnson et al. 2003). Learning helps users to experience more aspects of the news website, which can also explain the main effect on breadth: users read more news categories. Hoch and Deighton (1989) suggest that learning involves actively seeking experience with a product. To further examine this learning effect, we separately analyzed visits to the news website according to the referring website. Specifically, users could visit the news website directly (e.g., by using a bookmark or directly typing in the URL) or could be referred by a social media website (e.g., Facebook), a search engine (primarily Google) or an email with a newsletter of the newspaper. We find that the increase in article views is driven by users directly visiting the news website, which coincides with the active seeking process and indicates that users enjoy the website experience without ads and start to develop routines and habits associated with the website. Such routines and habits might last even longer than our observation period and represent a truly permanent 1 effect, particularly given the inherently recurring nature of news consumption (DeFleur and Ball-Rokeach 1989).
Heterogeneous Treatment Effects Across Users with Different Characteristics
Table 7 presents the results of the regression that uses user characteristics to derive heterogeneous treatment effects, as defined in equation (2). The treatment effect is significantly stronger for users with a stronger tendency to visit the website via a mobile phone ( β = 0.089) and for users with older Java versions ( β = 0.131), indicating that ad annoyance and page loading speed play a role in driving the effect of ad exposure on news consumption. In addition, we find that individuals who were light users in the pre-treatment period (i.e., frequency of visiting the website was below the median) were more strongly affected than heavy users by not seeing ads ( β = 0.163), which coincides with a larger learning effect: users who have less prior experience with the website are more strongly affected. Notably, the stronger effect for light users also indicates the potential effect for unregistered users, who are presumably light users compared to registered users. [Insert Table 7 about here] SUMMARY AND CONCLUSION
We used 3.1 million anonymized browsing sessions from 79,856 users on a news website and the quasi-random variation created by ad blocker adoption to show that exposure to ads has a robust negative effect on news consumption. Our analysis, which controlled for self-selection and observable and unobservable trends, revealed that users read 20% fewer news articles and 10% fewer news categories when they were exposed to ads than they did 2 when they were not exposed to ads (as reflected in ad blocker usage). These percentages translate into 2 fewer news articles per week and 1 less news category in total. Our effects were robust to three alternative treatment definitions applied to different subsamples of users (ad blocker adoption vs. non-adoption; early adoption vs. late adoption; ad blocker abandonment vs. continuous usage). These effects persisted over time and were primarily attributable to a decrease in the consumption of hard news, defined as political news, economic news and opinion news. Stated differently, not seeing ads increases news consumption, both in quantity and variety. We find that the effect is driven primarily by a learning mechanism, as reflected in the fact that users who adopt an ad blocker subsequently visit the news site more frequently. We also identified another, less prominent, and less persistent driver of the effect: a cognitive mechanism, wherein users who adopt an ad blocker subsequently read more articles per visit. This finding suggests that, in the presence of ads, consumers consciously or subconsciously devote cognitive resources to processing such ads. Taken together, these findings suggest that adopters of an ad blocker benefit twice: the ease of consuming news content improves in the absence of ads, and users further benefit by consuming more (ad-free) news than they did previously. We also observed that the magnitude of the main effect differs across individual users, where light users are more strongly affected by ad blocker adoption. The effect is also stronger for users with a stronger tendency to visit the site on mobile (as opposed to desktop) devices, as well as for users with older Java versions, suggesting that the annoyance of ads and page loading speed had roles in the effect. 3
Our findings counter the common assumption in news consumption studies that exposure to advertising does not affect users’ engagement with the news content itself (Aribarg and Schwartz 2019; Pattabhiramaiah et al. 2018). Rather, they provide support for previous lab findings (e.g.,
Goldstein et al. (2014)) that suggest that ads are annoying and distract users. Our findings are also in line with previous research that identified negative effects of ads on media consumption in non-digital markets (Wilbur et al. 2013). Our observation that adoption of an ad blocker led to an increase in the frequency of visits to the website diverges somewhat from the findings of Shiller et al. (2018), who showed that increasing use of ad blocking on a website leads to a reduction of the website’s traffic over 35 months. Yet, we suggest that the results are complementary, rather than contradictory: Shiller et al. (2018) attributed their findings to the fact that ad blockers result in a loss of revenue for publishers, thus leading them to reduce their investment in the website’s content, ultimately making the website less attractive. Our results show that, in the short term, ad blocker usage might actually increase user traffic, owing to the immediate benefits that users experience from not seeing ads. This finding does not preclude the possibility that, in the long term, that effect might be compensated by a reduction in the quality of the content. Our results suggest that the use of ads to finance online news content—a highly common business model—might, in fact, reduce news consumption. This finding has clear implications for news publishers, most of whom struggle to find a sustainable business model. Many publishers rely on advertising, and they suffer greatly from ad blockers, which 4 diminish their revenues. In some cases, publishers even pay heavy consulting fees to ad blocking providers to benefit from their knowledge on how to create “acceptable ads”. Yet, our findings show that eliminating ads can enhance users’ consumption behavior, resulting in more repeat visits to the website, one of the most common loyalty metrics. This finding suggests that alternative business models that do not rely on advertising might offer benefits to publishers and users alike. For example, publishers might want to consider exploring ad-free subscription plans. Future research could leverage our findings to weigh the advantages and disadvantages of the advertising-based publishing model against those of alternative mechanisms that do not rely on ads. Notably, the total cost of other monetization methods of news is not clear, in particular as not all news might be considered to be equally valuable. Our findings that ad exposure can serve as a means of influencing the quantity of news that individuals consume, as well as the number of news categories that they access, might also have societal implications. A question that arises, one that goes beyond the scope of this manuscript, is how much news consumption, and what kinds of news, are desirable from a societal perspective—and whether it is worthwhile for policy makers to intervene in this regard. In general, it seems that it could be beneficial to discourage consumers’ consumption of low-quality or fake news or to encourage more diverse news consumption, towards mitigating problems such as echo chambers (i.e., consumption of content corresponding to narrow viewpoints). Hypothetically, if policy makers were able to classify specific news items as being more or less “valuable” from a societal perspective, then it is possible that not showing ads for valuable content and showing ads for less valuable content could encourage 5 consumption of the former rather than the latter. Taxation of advertising revenue—and specifically, differential taxes across different types of content—might help to move in such a direction. Clearly, such a tax should be considered cautiously, given the prevalence of the advertising business model in publishing and the potential to severely harm the publishing industry by further increasing its costs. For some recent discussion in that area, see Stourm and Bax (2017), Kerkhof and Münster (2015) and Nobel laureate Paul Romer (Romer 2019). Finally, it would be of interest to explore whether our findings hold in contexts other than online news, e.g., online videos or online education, and even offline education. It seems likely that the relationships between ad viewing and the content consumption experience are not specific to news content per se. If this is the case, our findings reveal that consumers’ exposure to advertising might serve as a pathway towards influencing their consumption, and thus potentially improving societal benefit. References
Anderson, Simon P. and Joshua S. Gans (2011), "Platform siphoning: Ad-avoidance and media content,"
American Economic Journal: Microeconomics , 3 (4), 1-34. Angelucci, Charles and Julia Cagé (2019), "Newspapers in times of low advertising revenues,"
American Economic Journal: Microeconomics , 11 (3), 319-64. Angrist, Joshua D and Jörn-Steffen Pischke (2008), Mostly Harmless Econometrics: An Empiricist's Companion. Princeton: Princeton university press. Aribarg, Anocha and Eric M Schwartz (2019), "Native advertising in online news: Trade-offs among clicks, brand recognition, and website trustworthiness,"
Journal of Marketing Research , 57 (1), 20-34. Baddeley, Alan (1992), "Working memory,"
Science , 255 (5044), 556-59. Bagozzi, Richard P and Alvin J Silk (1983), "Recall, recognition, and the measurement of memory for print advertisements,"
Marketing Science , 2 (2), 95-134. Blake, Thomas, Chris Nosko, and Steven Tadelis (2015), "Consumer heterogeneity and paid search effectiveness: A large ‐ scale field experiment," Econometrica , 83 (1), 155-74. Ching, Andrew T, Tülin Erdem, and Michael P Keane (2017), "Empirical models of learning dynamics: A survey of recent developments," in
Handbook of Marketing Decision Models , Berend Wierenga and Ralf Van der Lans, eds.: Springer. Datta, Hannes, George Knox, and Bart J Bronnenberg (2018), "Changing their tune: How consumers’ adoption of online streaming affects music consumption and discovery,"
Marketing Science , 37 (1), 5-21. DeFleur, Melvin L. and Sandra Ball-Rokeach (1989), Theories of Mass Communication. New York: Longman. DellaVigna, Stefano and Eliana La Ferrara (2015), "Economic and social impacts of the media," in
Handbook of Media Economics , Simon P Anderson and Joel Waldfogel and David Strömberg, eds. Vol. 1. North Holland: Elsevier. Dinner, Isaac M, Harald J Van Heerde, and Scott A Neslin (2013), "Driving online and offline sales: The cross-channel effects of traditional, online display, and paid search advertising,"
Journal of Marketing Research , 50 (5), 527-45. 7
Drèze, Xavier and François-Xavier Hussherr (2003), "Internet advertising: Is anybody watching?,"
Journal of Interactive Marketing , 17 (4), 8-23. Fan, Ying (2013), "Ownership consolidation and product characteristics: A study of the US daily newspaper market,"
American Economic Review , 103 (5), 1598-628. Feddersen, Timothy J. (2004), "Rational choice theory and the paradox of not voting,"
Journal of Economic Perspectives , 18 (1), 99-112. Goldfarb, Avi and Catherine Tucker (2011), "Online display advertising: Targeting and obtrusiveness,"
Marketing Science , 30 (3), 389-404. Goldstein, Daniel G, Siddharth Suri, R Preston McAfee, Matthew Ekstrand-Abueg, and Fernando Diaz (2014), "The economic and cognitive costs of annoying display advertisements,"
Journal of Marketing Research , 51 (6), 742-52. Heckman, James, Hidehiko Ichimura, Jeffrey Smith, and Petra Todd (1998), "Characterizing selection bias using experimental data,"
Econometrica , 1017-98. Hoch, Stephen J and John Deighton (1989), "Managing what consumers learn from experience,"
Journal of marketing , 53 (2), 1-20. Hong, Weiyin, James YL Thong, and Kar Yan Tam (2004), "Does animation attract online users’ attention? The effects of flash on information search performance and perceptions,"
Information Systems Research , 15 (1), 60-86. IAB (2017), "IAB New Standard Ad Unit Portfolio," Interactive Advertising Bureau. Johnson, Eric J, Steven Bellman, and Gerald L Lohse (2003), "Cognitive lock-in and the power law of practice,"
Journal of Marketing , 67 (2), 62-75. Johnson, Justin P. (2013), "Targeted advertising and advertising avoidance,"
The RAND Journal of Economics , 44 (1), 128-44. Joo, Mingyu, Kenneth C Wilbur, Bo Cowgill, and Yi Zhu (2014), "Television advertising and online search,"
Management Science , 60 (1), 56-73. Kahneman, Daniel (1973), Attention and Effort. Englewood Cliffs, N.J.: Prentice-Hall. Kaiser, Ulrich and Minjae Song (2009), "Do media consumers really dislike advertising? An empirical assessment of the role of advertising in print media markets,"
International Journal of Industrial Organization , 27 (2), 292-301. 8
Kerkhof, Anna and Johannes Münster (2015), "Quantity restrictions on advertising, commercial media bias, and welfare,"
Journal of Public Economics , 131, 124-41. King, Gary and Richard Nielsen (2019), "Why propensity scores should not be used for matching,"
Political Analysis , 27 (4), 435-54. Lambrecht, Anja and Kanishka Misra (2017), "Fee or free: When should firms charge for online content?,"
Management Science , 63 (4), 1150-65. Lin, Song, Juanjuan Zhang, and John R Hauser (2015), "Learning from experience, simply,"
Marketing Science , 34 (1), 1-19. Loughran, Tim and Bill McDonald (2014), "Measuring readability in financial disclosures,"
The Journal of Finance , 69 (4), 1643-71. Lovett, Mitchell J and Richard Staelin (2016), "The role of paid, earned, and owned media in building entertainment brands: Reminding, informing, and enhancing enjoyment,"
Marketing Science , 35 (1), 142-57. Newman, Nic, Richard Fletcher, Antonis Kalogeropoulos, David AL Levy, and Rasmus-Kleis Nielsen (2016), "Digital news report 2016," in Reuters Institute for the Study of Journalism. Pattabhiramaiah, Adithya, S Sriram, and Puneet Manchanda (2019), "Paywalls: Monetizing online content,"
Journal of Marketing , 83 (2), 19-36. Pattabhiramaiah, Adithya, S Sriram, and Shrihari Sridhar (2018), "Rising prices under declining preferences: The case of the US print newspaper industry,"
Marketing Science
Journal of Marketing Research , 57 (3), 395-421. Rysman, Marc (2004), "Competition between networks: A study of the market for yellow pages,"
The Review of Economic Studies , 71 (2), 483-512. 9
Sahni, Navdeep S and Charles Zhang (2020), "Search advertising and information discovery: Are consumers averse to sponsored messages?," in Stanford University Graduate School of Business Research Paper No. 3441786. Scharkow, Michael, Frank Mangold, Sebastian Stier, and Johannes Breuer (2020), "How social network sites and other online intermediaries increase exposure to news,"
Proceedings of the National Academy of Sciences , 117 (6), 2761-63. Schmidt, Ana Lucía, Fabiana Zollo, Michela Del Vicario, Alessandro Bessi, Antonio Scala, Guido Caldarelli, H Eugene Stanley, and Walter Quattrociocchi (2017), "Anatomy of news consumption on Facebook,"
Proceedings of the National Academy of Sciences , 114 (12), 3035-39. Shiller, Benjamin, Joel Waldfogel, and Johnny Ryan (2018), "The effect of ad blocking on website traffic and quality,"
The RAND Journal of Economics , 49 (1), 43-63. Stourm, Valeria and Eric Bax (2017), "Incorporating hidden costs of annoying ads in display auctions,"
International Journal of Research in Marketing , 34 (3), 622-40. Strömberg, David (2015), "Media coverage and political accountability," in
Handbook of Media Economics , Simon P Anderson and Joel Waldfogel and David Strömberg, eds. Vol. 1. North Holland: Elsevier. Tetlock, Paul C (2015), "The role of media in finance," in
Handbook of Media Economics , Simon P Anderson and Joel Waldfogel and David Stromberg, eds. Vol. 1. North Holland: Elsevier. Vakratsas, Demetrios and Tim Ambler (1999), "How advertising works: What do we really know?,"
Journal of Marketing , 63 (1), 26-43. Vosoughi, Soroush, Deb Roy, and Sinan Aral (2018), "The spread of true and false news online,"
Science , 359 (6380), 1146-51. Vratonjic, Nevena, Mohammad Hossein Manshaei, Jens Grossklags, and Jean-Pierre Hubaux (2013), "Ad-Blocking Games: Monetizing Online Content Under the Threat of Ad Avoidance," in
The Economics of Information Security and Privacy , Rainer Böhme, ed. Berlin, Heidelberg: Springer Berlin Heidelberg. Wilbur, Kenneth C, Linli Xu, and David Kempe (2013), "Correcting audience externalities in television advertising,"
Marketing Science , 32 (6), 892-912. 0
Table 1. Description of News Consumption and User Activity on the News Website
Panel 1. Summary Statistics of News Consumption Variables min median mean max sd Main Variables Article Views 0.00 4.00 9.40 1204.00 17.99 Breadth 0.00 3.00 3.57 20.00 3.30 Learning Variables Visits 1.00 4.00 7.72 180.00 9.88 Direct Visits 0.00 3.00 7.14 178.00 9.58 Social Media Visits 0.00 0.00 0.12 149.00 1.37 Search Engine Visit 0.00 0.00 0.55 117.00 2.70 Newsletter Visit 0.00 0.00 0.00 19.00 0.10 Cognitive Variables Article Views Per Visit 0.00 1.00 1.28 213.00 2.09 Homepage Views per Visit 0.00 1.00 1.20 117.12 1.09 Time per Visit (in seconds) 0.00 259.83 408.31 36163.00 568.76 Words per Article 0.00 577.00 606.73 10203.00 302.03 Words per Sentence 1.00 15.68 15.76 44.20 2.28
Panel 2. User Behavior on the News Website
Category of News Articles % of Page Views Category of Non-News Articles % of Page Views
International Political News 8.35 Homepage 44.17
Economy News 6.44 Account Related Page 3.72
Sport News 6.22 Weather Forecast 1.82
Regional Political News 4.79 Search 1.02
Finance News 4.03 Others 0.56
Opinion News 3.81 Play Page 0.22
Outlook News 2.97 Archive 0.06
Local Political News 2.94
Art & Culture News 2.20
NewsTicker News 1.22
Sunday News 1.12
Science News 1.01
Digital News 0.92
Lifestyle News 0.81
Photostream News 0.71
Transportation News 0.31
Brief News 0.30
Video News 0.18
Special News 0.09
Data News 0.02 Table 2. Before-and-after Analysis of Ad Blocker Adopters and Non-Ad Blocker Adopters
Variable Ad Blocker Adopters Group Mean Ad Blocker Non Adopters Group Mean Difference in Differences (DiD)
Pre-Treatment Post-Treatment Difference Pre-Week 12 Post-Week 12 Difference Article Views
Breadth
Pre-Treatment Post-Treatment Difference Pre-Week 13 Post-Week 13 Difference Article Views
Breadth
Pre-Treatment Post-Treatment Difference Pre-Week 14 Post-Week 14 Difference Article Views
Breadth
Pre-Treatment Post-Treatment Difference Pre-Week 15 Post-Week 15 Difference Article Views
Breadth
Number of Users Table 3. Comparison of Non-Ad Blocker Adopters and Ad Blocker Adopters before and after Coarsened Exact Matching (CEM)
Variable Operationalization Unmatched Sample Std. Mean Difference Matched Sample
Std. Mean Difference Control Group Mean Treatment Group Mean Control Group Mean Treatment Group Mean Dummy Variables
Gender Male=1
Income Index2=1 -0.0512
Index3=1
Index4=1
Index5=1
Inde6=1
Age 25-29=1
First Week Week2=1
Week3=1
Week4=1
Week5=1
Week6=1
Week7=1
Week8=1
Week9=1
Week10=1
Week11=1
Last Week Week11=1
Week12=1
Week13=1
Week14=1
Week15=1
Week16=1
Continuous Variables
Page Views
Time per Visit -0.0570
Breadth N Notes: Standardized mean difference is non-significant between control group and treatment group in the matched sample. Table 4. Treatment Effect on Article Views and Breadth
Ad Blocker Adoption Ad Blocker Early Adoption Ad Blocker Abandonment
Article Views Breadth Article Views Breadth Article Views Breadth 𝜷𝜷 *** *** *** *** -0.110 *** -0.067 *** (0.025) (0.017) (0.044) (0.028) (0.030) (0.019) 𝜷𝜷 *** *** -0.097 ** -0.076 *** (0.030) (0.020) (0.035) (0.021) N 26,128 26,128 7,225 7,225 16,945 16,945 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Clustered standard errors appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table 5. Treatment Effect on Article Views Decomposed by News Category Panel 1. Effect on Article Views by News Category
Hard News Soft News Non-News Article Pages
Political Economic Opinion Sports Art & Culture Lifestyle Weather Play Page Account 𝜷𝜷 ** * * ** * -0.002 -0.019 (0.025) (0.022) (0.018) (0.020) (0.015) (0.010) (0.011) (0.004) (0.015) 𝜷𝜷 *** ** *** ** Panel 2. Effect on Article Views by News Category Controlling for Home Page Views
Hard News Soft News
International Political Regional Political Local Political Economic Opinion Sports Art & Culture Lifestyle 𝜷𝜷 * -0.005 0.009 0.035* 0.012 0.036 * * 𝜷𝜷 -0.007 0.037 * * 𝜷𝜷 * * * * * * * * (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.000) N 26,128 26,128 26,128 26,128 26,128 26,128 26,128 26,128 R Notes: In Panel 1, each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. In Panel 2, each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐻𝐻𝑜𝑜𝑇𝑇𝑤𝑤𝑆𝑆𝑇𝑇𝐿𝐿𝑤𝑤𝐻𝐻𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. In both panels, β represents the immediate effect and β represents the permanent effect. R computation includes the explanatory power of the fixed effects. Clustered standard errors appear in parentheses. * p < 0.1. Table 6. Treatment Effect on Cognitive and Learning Variables Panel 1. Effect on Cognitive Variables
Article Views per Visit Home Page Views per Visit Words per Article Words per Sentence Time per Visit 𝜷𝜷 * * * ** (0.011) (0.009) (0.026) (0.004) (0.053) 𝜷𝜷 * *** (0.012) (0.011) (0.027) (0.005) (0.058) N 26,128 26,128 22,542 22,253 26,128 R Panel 2. Effect on Learning Variables
Visits Direct Visits Social Media Visits Search Engine Visits Newsletter Visits 𝜷𝜷 *** *** -0.000 0.012 0.001 (0.017) (0.018) (0.005) (0.013) (0.001) 𝜷𝜷 *** *** * Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Clustered standard errors appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table 7. Heterogeneous Treatment Effects on Article Views and Breadth Across User Characteristics
Article Views Breadth 𝛽𝛽 𝛽𝛽 (Annoy) * (0.041) (0.026) 𝛽𝛽 (Privacy) -0.235 0.062 (0.274) (0.143) 𝛽𝛽 (Speed) *** *** (0.040) (0.027) 𝛽𝛽 (Light) *** *** (0.036) (0.025) N 26,128 26,128 R Notes: Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝐴𝐴𝑖𝑖𝑖𝑖𝑜𝑜𝐴𝐴 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗𝑃𝑃𝑇𝑇𝑤𝑤𝑟𝑟𝑇𝑇𝑠𝑠𝐴𝐴 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝑆𝑆𝑆𝑆𝑤𝑤𝑤𝑤𝑆𝑆 𝑖𝑖 + 𝛽𝛽 𝐼𝐼 𝑖𝑖𝑖𝑖2 ∗ 𝐿𝐿𝑤𝑤𝐿𝐿ℎ𝑤𝑤 𝑖𝑖 + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Clustered standard errors appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Figure 1. Comparison of a News Website without and with an Ad Blocker
Source:
Screenshots from ft.com . Figure 2. Construction of Treatment Groups and Control Groups Figure 3. Distribution of Propensity Score in Matched and Raw Sample Figure 4. Trends in the Pre-and Post-Treatment Period for Ad Blocker Adopters and Non-Ad Blocker Adopters Web Appendix A. Placebo Treatment Test
The identification assumption under difference-in-differences (DiD) is that in the absence of the treatment (i.e., ad blocker adoption), there would have been no differential changes in news consumption between the treatment and control group. To formally test the parallel pre-treatment trends condition for the treatment and control group, we perform a “placebo” treatment test by estimating the week-wise treatment effects before and after treatment. Specifically, we replace the 𝐼𝐼 𝑖𝑖𝑖𝑖1 and 𝐼𝐼 𝑖𝑖𝑖𝑖2 in the equation (1) with a set of week-wise dummy variables 𝐼𝐼 𝑖𝑖𝑖𝑖−𝜏𝜏 which equal one if 𝜏𝜏 weeks before treatment and another set of dummy variables 𝐼𝐼 𝑖𝑖𝑖𝑖+𝜏𝜏 equal to 1 if 𝜏𝜏 weeks after treatment: (3) 𝑌𝑌 𝑖𝑖𝑖𝑖 = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + ∑ 𝛽𝛽 −𝜏𝜏 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖−𝜏𝜏 ( 𝜏𝜏 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑏𝑏𝑤𝑤𝑜𝑜𝑜𝑜𝑇𝑇𝑤𝑤 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜏𝜏=5𝜏𝜏=1 ∑ 𝛽𝛽 𝜏𝜏 ∗ 𝜏𝜏=4𝜏𝜏=0 𝐼𝐼 𝑖𝑖𝑖𝑖+𝜏𝜏 ( 𝜏𝜏 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 , where 𝑌𝑌 𝑖𝑖𝑖𝑖 is the pre-treatment news consumption for user i in week t ; 𝛼𝛼 𝑖𝑖 is a user-level fixed effect; 𝛿𝛿 𝑖𝑖 is the week fixed effect; 𝐼𝐼 𝑖𝑖𝑖𝑖−𝜏𝜏 is a set of interactions of the treated users and 𝜏𝜏 weeks before treatment; 𝐼𝐼 𝑖𝑖𝑖𝑖+𝜏𝜏 is a set of interactions of the treated users and 𝜏𝜏 weeks since treatment; 𝜀𝜀 𝑖𝑖𝑖𝑖 is the standard error clustered at the user-level. We choose the last week before treatment ( 𝐼𝐼 𝑖𝑖𝑖𝑖−1 ) as the omitted default category. If the trends of the treatment and control group are parallel, then 𝛽𝛽 −𝜏𝜏 will be statistically indistinguishable from zero. As reported in Table S1, all the main news consumption measures we used pass this test with a non-significant point estimate of 𝛽𝛽 −𝜏𝜏 in the pre-treatment period. Table S1. Placebo Treatment Test on News Consumption Variables
Article Views Breadth Visits Article Views per Visit International Political News Regional Political News Local Political News Economy News 𝛽𝛽 −2 𝛽𝛽 −3 𝛽𝛽 −4 𝛽𝛽 −5 -0.049 -0.036 -0.047 -0.002 -0.001 -0.045 0.003 -0.045 (0.032) (0.022) (0.023) (0.014) (0.022) (0.021) (0.018) (0.022) R Finance News Opinion News Sport News Art & Culture News Lifestyle News Weather Forecast Play Page Account 𝛽𝛽 −2 -0.020 0.022 0.006 0.004 0.015 0.004 -0.002 0.002 (0.017) (0.017) (0.018) (0.013) (0.010) (0.013) (0.004) (0.015) 𝛽𝛽 −3 -0.011 0.005 0.033 0.004 0.021 -0.012 -0.007 -0.015 (0.018) (0.018) (0.020) (0.014) (0.010) (0.014) (0.004) (0.016) 𝛽𝛽 −4 -0.014 -0.001 0.022 0.010 0.009 -0.003 -0.007 -0.012 (0.018) (0.019) (0.020) (0.014) (0.010) (0.014) (0.005) (0.018) 𝛽𝛽 −5 -0.048 * Home Page Views per Visit Words per Article Words per Sentence Time per Visit Direct Visits Social Media Visits Search Engine Visits Newsletter Visits 𝛽𝛽 −2 -0.016 -0.019 -0.002 -0.055 0.028 -0.001 -0.005 -0.000 (0.010) (0.029) (0.005) (0.060) (0.021) (0.005) (0.013) (0.000) 𝛽𝛽 −3 -0.016 0.027 -0.001 -0.006 0.024 -0.000 0.009 0.000 (0.010) (0.028) (0.005) (0.064) (0.022) (0.005) (0.013) (0.001) 𝛽𝛽 −4 -0.001 -0.018 -0.000 0.010 0.014 0.006 0.009 -0.000 (0.011) (0.031) (0.005) (0.065) (0.023) (0.005) (0.014) (0.000) 𝛽𝛽 −5 -0.005 0.028 -0.008 -0.121 -0.044 -0.003 -0.007 -0.001 (0.011) (0.027) (0.005) (0.067) (0.023) (0.005) (0.013) (0.000) R Notes: Each column refers to a separate regression with the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + ∑ 𝛽𝛽 −𝜏𝜏 ∗ 𝜏𝜏=5𝜏𝜏=1 𝐼𝐼 𝑖𝑖𝑖𝑖−𝜏𝜏 ( 𝜏𝜏 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑏𝑏𝑤𝑤𝑜𝑜𝑜𝑜𝑇𝑇𝑤𝑤 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + ∑ 𝛽𝛽 𝜏𝜏 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖+𝜏𝜏 ( 𝜏𝜏 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜏𝜏=4𝜏𝜏=0 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample. 𝛽𝛽 −𝜏𝜏 are the placebo treatment effects and are reported, with 𝛽𝛽 −1 omitted as the default category. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. * p < 0.01. Web Appendix B. Robustness Check on Adding Time-Varying Controls
The placebo treatment test (reported in Table S1) statistically validates the identification condition (parallel pre-treatment trend) of DiD. Recall that DiD removes all time-invariant confounders. Given that the parallel pre-treatment trend holds, DiD also eliminates any bias from time-varying confounders because a common pre-treatment trend implies that time-varying confounders, if any, impact the treatment and control groups in the same way in the pre-treatment period. Concerns may still remain that a time-varying confounder kicks in at the same time with the treatment and, thus, will bias our result. For example, a user may read news with different browsers, which might change his ad blocker usage and also impact his news reading behavior. To establish the robustness of our main result, we rerun our main estimation by adding the following time-varying control variables: browser switching (i.e., the number of different browsers a user uses during one week), ordering (i.e., the number of orders a user places on the website during the week, such as purchasing access to the news archive), and commenting (i.e., the number of comments a user leaves during one week). Specifically, we estimate the following model (4) 𝑌𝑌 𝑖𝑖𝑖𝑖 = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐵𝐵𝑇𝑇𝑜𝑜𝑤𝑤𝑤𝑤𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝛽𝛽 ∗ 𝑂𝑂𝑇𝑇𝑆𝑆𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝛽𝛽 ∗ 𝐶𝐶𝑜𝑜𝑇𝑇𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝜀𝜀 𝑖𝑖𝑖𝑖 , The results are reported in Table S2. We find that our main treatment effects ( 𝛽𝛽 and 𝛽𝛽 ) stay robust. From now on, we only report results for news categories, classified into the following: hard news (political, economic and opinion news) and soft news (sports, culture & art, lifestyle news). 4 Table S2. Robustness of Main Model after also Controlling for Browser Switching, Ordering, & Commenting
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** * *** *** (0.025) (0.017) (0.016) (0.011) (0.026) (0.023) 𝛽𝛽 *** *** *** *** * (0.029) (0.019) (0.020) (0.012) (0.030) (0.024) 𝛽𝛽 ( 𝐵𝐵𝑇𝑇𝑜𝑜𝑤𝑤𝑤𝑤𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 ) 0.325 *** *** *** ** *** *** (0.014) (0.010) (0.010) (0.005) (0.014) (0.013) 𝛽𝛽 ( 𝑂𝑂𝑇𝑇𝑆𝑆𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 ) -0.220 -0.194 * -0.032 -0.151 -0.185 -0.026 (0.153) (0.086) (0.082) (0.086) (0.151) (0.073) 𝛽𝛽 ( 𝐶𝐶𝑜𝑜𝑇𝑇𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤𝑤𝑤 𝑖𝑖𝑖𝑖 ) -0.007 0.004 0.040 *** -0.028 *** Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 )+ 𝛽𝛽 ∗ 𝐵𝐵𝑇𝑇𝑜𝑜𝑤𝑤𝑤𝑤𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝛽𝛽 ∗ 𝑂𝑂𝑇𝑇𝑆𝑆𝑤𝑤𝑇𝑇𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝛽𝛽 ∗ 𝐶𝐶𝑜𝑜𝑇𝑇𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤𝑤𝑤 𝑖𝑖𝑖𝑖 + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starting on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Web Appendix C. Robustness Checks on Matching
Table S3. Robustness on Unmatched Sample
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** *** *** *** (0.011) (0.007) (0.007) (0.005) (0.011) (0.010) 𝛽𝛽 *** *** *** *** *** (0.012) (0.007) (0.008) (0.005) (0.012) (0.009) N 252,428 252,428 252,428 252,428 252,428 252,428 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on the unmatched sample. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table S4. Robustness on Missing Observations Due to Coarsened Exact Matching (CEM)
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 ( 𝐼𝐼 𝑖𝑖𝑖𝑖2 ) *** *** *** *** *** *** (0.014) (0.009) (0.010) (0.006) (0.014) (0.011) 𝛽𝛽 ( 𝐼𝐼 𝑤𝑤𝑤𝑤 ∗ 𝑀𝑀𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑖𝑖𝐿𝐿 𝑤𝑤 ) Notes: Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 𝑜𝑜𝑜𝑜𝑇𝑇 𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑤𝑤𝑆𝑆 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 𝑜𝑜𝑜𝑜𝑇𝑇 𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑤𝑤𝑆𝑆 ∗ 𝑀𝑀𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑖𝑖𝐿𝐿 𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on the unmatched sample. β represents the permanent effect and β represents the interaction effect of the permanent effect and any missing observations due to users not revealing full information in our CRM data. Insignificant 𝛽𝛽 indicates matching does not induce bias in the estimation. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table S5. Robustness on Propensity Score Matching (PSM)
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** * *** *** (0.020) (0.014) (0.014) (0.009) (0.021) (0.019) 𝛽𝛽 *** *** *** * *** *** (0.025) (0.017) (0.018) (0.011) (0.026) (0.021) N 33,007 33,007 33,007 33,007 33,007 33,007 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a sample matched using propensity score matching (psm). R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Web Appendix D. Robustness Checks on Different Cutoff Periods
Table S6. Robustness on 1 Week as Cutoff Period
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** *** *** *** (0.009) (0.006) (0.006) (0.004) (0.009) (0.008) 𝛽𝛽 *** *** *** * *** *** (0.010) (0.007) (0.007) (0.004) (0.010) (0.008) N 203,852 203,852 203,852 203,852 203,852 203,852 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on unmatched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table S7. Robustness on 3 Weeks as Cutoff Periods
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** ** *** *** (0.013) (0.009) (0.009) (0.006) (0.013) (0.012) 𝛽𝛽 *** *** *** *** *** (0.014) (0.009) (0.010) (0.006) (0.014) (0.011) N 167,668 167,668 167,668 167,668 167,668 167,668 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on unmatched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table S8. Robustness on 4 Weeks as Cutoff Periods
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** *** *** (0.028) (0.020) (0.019) (0.014) (0.030) (0.027) 𝛽𝛽 ** * * * *** *** (0.029) (0.020) (0.020) (0.014) (0.031) (0.027) N 142,074 142,074 142,074 142,074 142,074 142,074 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on unmatched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Web Appendix E. Robustness Checks on Longer Pre-Treatment Period
Table S9. Robustness on Using Week 1 to Week 11 as Pre-Treatment Period
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** *** ** *** *** (0.025) (0.016) (0.017) (0.011) (0.026) (0.023) 𝛽𝛽 *** *** *** *** *** (0.026) (0.016) (0.019) (0.011) (0.027) (0.022) N 39,922 39,922 39,922 39,922 39,922 39,922 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample from week 1 to week 16 (full observation period instead of 5 weeks before and after treatment starts on week 12). R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Web Appendix F. Robustness Checks on Effect Decomposition Using Other Quasi-Experimental Designs
Table S10. Robustness on Effect Decomposition Using Ad Blocker Early Adopters as Treatment Group and Ad Blocker Late Adopters as Control Group
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 *** *** * *** *** *** (0.044) (0.028) (0.018) (0.030) (0.045) (0.037) N 7,225 7,225 7,225 7,225 7,225 7,225 R Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05. Table S11. Robustness on Effect Decomposition Using Ad Blocker Abandoners as Treatment Group and Continuous Ad Blocker Users as Control Group
Article Views Breadth Visits Article Views per Visit Hard News Soft News 𝛽𝛽 -0.110 *** -0.067 *** -0.054 * -0.024 -0.095 ** -0.094 *** (0.030) (0.019) (0.021) (0.013) (0.031) (0.027) 𝛽𝛽 -0.097 ** -0.076 *** -0.121 *** -0.009 -0.119 *** Notes: β represents the immediate effect and β represents the permanent effect. Each column refers to a separate regression of the following model: log( 𝑌𝑌 𝑖𝑖𝑖𝑖 + 1) = 𝛼𝛼 𝑖𝑖 + 𝛿𝛿 𝑖𝑖 + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖1 ( 𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑤𝑤𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝛽𝛽 ∗ 𝐼𝐼 𝑖𝑖𝑖𝑖2 ( 𝑇𝑇𝑎𝑎𝑎𝑎 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤 𝑤𝑤𝑤𝑤𝑖𝑖𝑠𝑠𝑤𝑤
𝑇𝑇𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑇𝑇𝑤𝑤𝑖𝑖𝑤𝑤 𝑖𝑖𝑖𝑖 ) + 𝜀𝜀 𝑖𝑖𝑖𝑖 on a matched sample centered around 5 weeks before and after treatment starts on week 12. R computation includes the explanatory power of the fixed effects. Standard errors clustered at the user level appear in parentheses. *** p < 0.001, ** p < 0.01, * p < 0.05.p < 0.05.