The interplay between media-for-monitoring and media-for-searching: How news media trigger searches and edits in Wikipedia

Abstract

This study investigates how traditional news media and Internet services have become entangled in recipients’ habits of gathering information on current topics. Push media enable citizens to scan the issue environment while pull media enable them to seek out in-depth information if information needs have been elicited. Furthermore, content quality in many pull media may increase when more users generate content, removing flaws and adding information. We expected that TV and newspaper coverage of an issue will lead to increases in (a) searches for and (b) user edits in related articles in the online encyclopedia Wikipedia. Our findings reliably support the hypotheses, but the extent to which the count of page views increases is highly dependent on the topic at hand and how the search keyword relates to the issue. This matches the predictions of information-seeking theories and the dynamic transactional model of media effects.

Keywords

Dynamic transactional model information seeking news media user-generated content Wikipedia

Introduction

Initially, the emergence of the Internet gave rise to predictions that it would cause a revolution in media use, rendering the use of traditional media obsolete (Brown, 1999; Nguyen and Western, 2006). Despite some evidence for displacement of, for example, TV use by Internet use (e.g. Kayany and Yelsma, 2000), it has become clear that any such processes occur gradually rather than abruptly and dispersedly rather than comprehensively.

As different media have been integrated into everyday life to serve different purposes (Jung et al., 2012), new media often complement the function of older media rather than displacing them (Dutta-Bergman, 2004; Nguyen and Western, 2006). This study investigates how complementarity between functionally different media works in practice. We expect that traditional news media and newly established Internet-based knowledge stores have become entangled in recipients’ habits of gathering information on current issues, leading us to the following research questions: (a) How does news media coverage influence whether users search online for information on the issue? (b) How does it influence what kind of information they search for? (c) How long do these effects on information seeking last? Furthermore, we are interested in (d) whether and how news media coverage influences how much issue-related user-generated content (UGC) users contribute. In particular, we assume that increasing coverage of newspapers and TV newscasts about an issue will lead to an increase in searches and edits of issue-related articles in the Internet encyclopedia Wikipedia (Roessing, 2009).

To answer these questions, we first distinguish two different media types: media-for-monitoring and media-for-searching. After this differentiation, we introduce the dynamic transactional model (DTM) from which we derive hypotheses and research questions. We then examine the salience of several issues and keywords in media-for-monitoring (TV news, newspapers) and use them to predict the frequency of page views (searches) and edits of the corresponding Wikipedia articles.

Our analysis shows how media users combine media services. Media-for-monitoring provide general information on current-affairs issues that can be easily assessed in a passive mode, satisfying the human need for surveillance of the environment (Shoemaker, 1996). Some users will be motivated to turn to media-for-searching to acquire in-depth information. In the aggregate, small and conditional individual effects cumulate to waves of searches for and edits of issue-related Wikipedia articles. The interaction between different media types contributes to our understanding of processes of public opinion formation and knowledge acquisition.

Media-for-monitoring and media-for-searching

The Internet has enriched the general population’s opportunities to retrieve information about current issues. People tend to use traditional media to monitor the information environment and then use the Internet to look up additional information (Jung et al., 2012; Scharkow and Vogelgesang, 2011). Negroponte’s (1995) terminology of push versus pull media (see also Chaffee and Metzger, 2001) can be applied to the monitoring function of traditional (mass) news media (to scan the environment) and services for in-depth search on subjectively relevant aspects of the issue: news media push an issue to the top of mind of their mass audience (media-for-monitoring), and their audience pulls subjectively relevant in-depth information from web-based knowledge stores (media-for-searching). However, push media (media-for-monitoring) are not identical with traditional media and pull media (media-for-searching) are not always Internet-based. Online newspapers would, for example, fit our concept of media-for-monitoring, and a printed encyclopedia would fit our concept of media-for-searching.

To classify information services as media-for-monitoring or media-for-searching, we use three criteria: (a) Content richness: how much information is available? (b) Content access and accessibility: can the information be passively received or is active retrieval of information necessary? (c) Integration of new information: does new information replace older content or is new information added to modify, supplement, or amend existing content? Media-for-monitoring (e.g. TV newscasts, online newspapers, data journalism sites) pick up an issue as current events unfold. Recipients turn to these media regularly and habitually (passively) to get an overview over current issues. Media-for-monitoring typically hold a small and superficial selection of information on a broad variety of issues which is continuously replaced by new information. Media-for-searching provide much and in-depth information about various issues. Existing information may be enriched, corrected, or updated as the events unfold (also by UGC), rather than deleted. Accessing the information necessitates active effort and purposeful search.

Information seeking as dynamic transactions

The few precursory studies with a similar research focus employed the agenda-setting approach to account for changes in aggregate information-seeking behavior (Granka, 2010; Holbach and Maurer, 2014; Scharkow and Vogelgesang, 2011). They show that media-for-monitoring may influence recipients’ perceptions of an issue’s importance and salience. However, agenda setting merely accounts for the increase of page views in general. It is very limited in explaining which pages are viewed and when, and whether users edit content.

To extend this theorizing and overcome its shortcomings, we assume that online searches do not primarily express the public salience of issues but rather indicate subjective information gaps. Therefore, we turn to a more flexible theoretical framework: the DTM of media effects (see Früh, 1991: 31; Früh and Schönbach, 1982). It predicts that search behavior is driven by both content features (e.g. frequency of repetition of a keyword), individuals’ features (e.g. ability to extract a searchable word from media texts), and interactions between individuals’ and content features.

Individual-level mechanisms

We apply the DTM to the interplay between newspapers and TV news outlets (media-for-monitoring) and Wikipedia (media-for-searching). Our research interest is focused on the collective-level upswing of information seeking and UGC generation as a consequence of increasing mass media coverage. Therefore, predictions from the DTM—on the individual level—need to be aggregated to obtain testable hypotheses on the collective level. The progression of dynamic transactions (an idealized reception process) proceeds in five steps: (a) Users routinely/habitually use media-for-monitoring to stay up-to-date with current affairs. (b) Recipients encounter a news story capturing their interest; possibly media-for-monitoring do not gratify all information needs. (c) In this case, recipients refer to media-for-searching, for example, Wikipedia, to gratify their interest. To do so, it is necessary for them to “lemmatize” the news input, that is, identify issue-related keywords they can look up via a search mask (Rose and Levinson, 2004; Van Deursen and Van Dijk, 2011). To successfully do so, users need preconceptions as to which kinds of information are available from which source and what kinds of keywords are most likely to lead to the desired information. (d) Recipients extract additional knowledge from media-for-searching. Interest recedes as knowledge approximates the desired level (Case, 2012: 72–73). The most active recipients may contribute UGC if they feel knowledgeable and motivated to do so. (e) Interest and media use normalize.

However, this progression of steps can be terminated at any stage for each individual, for example, if a recipient is not sensitive toward the issue, fails to make sense of the information provided, cannot access the Internet, or lacks the time, motivation, or skills to engage in searching for information (see Howard and Massanari, 2007; Wirth et al., 2007).

Collective-level predictions

Search activity

Individual characteristics affect whether and how intensively one seeks out information (Dervin, 1998), which makes it hard to predict individuals’ information-seeking behavior. However, collective information behavior is more readily predictable (see, for example, Rössler, 1999). Given the conditional nature of recipient reactions proposed by the DTM, the likelihood that a person will use particular media-for-searching (here: Wikipedia) to gratify information needs is generally low. However, out of the millions of media users, even a low likelihood of triggering search behavior will incite hundreds, thousands, or hundreds of thousands of people to retrieve information they would never have encountered had the issue not been raised in media-for-monitoring as the likelihood and frequency of contact with issue-related information increases.

H1. The more media-for-monitoring cover an issue, the more recipients will access media-for-searching content related to the issue.

Media-for-monitoring may suggest keywords to search for, which makes searching less cognitively demanding. The users will be more likely to visit content connected to keywords often mentioned in media-for-monitoring coverage (push-cued searches). Those who already read content in media-for-searching are likely to also encounter issue-related hyperlinks. This offers another low-cost information opportunity (Case, 2012: 100–102) which they are also more likely to follow if they are interested (pull-cued searches).

H2. (a) The more prevalent a keyword is in media-for-monitoring coverage, the more users will search for corresponding media-for-searching content (push-cueing). (b) The more hyperlinks in relevant media-for-searching content point to some content, the more users will access the media content pointed to (pull-cueing).

Individuals will differ in their actual levels of knowledge and in the level of knowledge they aspire. Therefore, they will search for different issues and different keywords (lemmas). However, due to shared cultural background norms and knowledge, in a given situation, particular issues and keywords will be more resonant in society as a whole, attracting much more searches than other issues and keywords. On the individual level, the most probable mechanism is that some issues and keywords appear more important, more urgent, more interesting, or more appealing to a larger part of the audience than others.

RQ1. How do (a) issues and (b) keywords moderate how the amount of coverage in media-for-monitoring influences search activity?

Editing activity

According to a similar mechanism, a (small) share of the audience will feel motivated to edit content in media-for-searching they encounter during their search procedure. While 73% of German Internet users read Wikipedia at least once in a while, only 2% of the German Internet users have ever contributed content to Wikipedia (Busemann and Gscheidle, 2011: 362). Wikipedia users edit out of (a) social motivations as they enhance the feeling to belong to the Wikipedia community, (b) intrinsic motivations as they enjoy using and refining skills needed for writing in Wikipedia, and (c) epistemic motivations as they learn new perspectives, information, and arguments (Kuznetsov, 2006; Nov, 2007; Schroer and Hertel, 2009). They may even actively search for content that needs editing: when an issue is in the news, Wikipedians may assume that thematically related articles need updating and revising, leading them to access these articles. But also contact with content (according to H1, H2, RQ1) may incite or activate the social, intrinsic, and epistemic motivations of potential editors. Overall, the same mechanisms leading to more page views should also lead to more page edits, although with much lower absolute numbers. Therefore, H3, H4, and RQ2 mirror H1, H2, and RQ1.

H3. The more media-for-monitoring cover an issue, the more recipients will edit content of media-for-searching related to the issue.

H4. (a) The more prevalent a keyword is in media-for-monitoring, the more users will edit corresponding media-for-searching content (push-cueing). (b) The more hyperlinks in relevant media-for-searching content point to some content, the more users will edit the media-for-searching content pointed to (pull-cueing).

RQ2. How do (a) issues and (b) keywords moderate how the amount of coverage in media-for-monitoring influences editing activity?

Temporal patterns

A successful search for information decreases the drive to search for (even more) information. So, there is a bias toward a short-term effect of media stimulation on information seeking: even with continuing stimulation, search and editing activities are likely to normalize quickly. This is in line with previous findings: Granka (2010) found that search activity on Google typically wore out more quickly than mass media coverage (see also Holbach and Maurer, 2014).

H5. Search activities and editing activities decline more quickly than mass media coverage.

Research design and method

We conducted two similarly set studies on search and editing activities on Wikipedia. The first study investigated 12 issues for between-issue comparison (broadband study [BBS]). The second study investigated one salient issue (2010 Haiti Earthquake) in-depth (Haiti Case Study [HCS] with three sub-issues). We tracked the frequency of page views (searches) and edits of the corresponding Wikipedia articles (dependent variables) and recorded how salient the issues were in (a) TV news and in (b) newspaper coverage, (c) which keywords were mentioned how often in newspaper coverage (push-cueing), and (d) how strongly hyperlinks in Wikipedia pointed to particular Wikipedia articles (pull-cueing).

Choice of media

Our study focuses on newspapers and TV news (media-for-monitoring) and Wikipedia as an example of media-for-searching. The population in Germany mostly relies on TV news and newspapers to get information about current affairs. Taken together, newspapers and TV newscasts reach 90% of the German population (Busemann and Engel, 2012). With an average audience of over 9 million viewers (3 years and over) every day (Zubayr and Gerhard, 2011), Tagesschau is by far the single most important TV news provider in Germany. As the salience of issues is rather similar across different TV newscasts in Germany (Geiß, 2013), Tagesschau can serve as a proxy for TV newscasts in Germany. To meaningfully capture the very diverse newspaper landscape in Germany, we conducted keyword searches within the content of all newspapers included in the Genios database (with 166 newspapers of different types).

Wikipedia is an online encyclopedia (Roessing, 2009) that can be viewed and edited by every Internet user, free of charge. Within Wikipedia, articles are usually accessed via a search mask, but they may also be found via topically arranged portals and indexes. However, most traffic to Wikipedia flows from general search engines such as Google. Wikipedia has a high information load (there are over 1.6 million articles in the German-language version), its information is actively retrieved (rather than passively received), and older information is amended rather than replaced, fulfilling all criteria for media-for-searching.

Choice of issues and lemmas

Broadband study (BBS)

To identify salient issues for which news coverage and search activities could be tracked over time, we listed all headline issues in Tagesschau in the first half of January 2010 (January 1–15) and found 12 different headline issues. We assumed that these issues were broadly discussed in other media as well.

Transcripts of all news items in the Tagesschau dealing with any of the 12 headline issues (January 1–31, 2010) were searched for keywords (lemmas) that are clearly related to the issue and for which articles in the German-language version of Wikipedia exist. To ensure comparability between the Haiti issue (with many relevant keywords) and the other issues, the number of lemmas per issue was limited to five (this was the highest number of lemmas for any of the issues apart from the Haiti issue). In the Haiti issue, we chose those keywords that were mentioned most frequently in news coverage. According to the procedure, 37 lemmas were identified (3 × 1, 1 × 2, 2 × 3, 4 × 4, 2 × 5 lemmas per issue).

Haiti case study

Lemmas for the Haiti case study (HCS) were collected in a three-step procedure: First, word counts were run over transcripts of Tagesschau news items about the earthquake in Haiti (January 1–31) to identify all meaningful vocabulary that was associated with the event and for which a Wikipedia article in the German-language version existed. 66 potential key terms were identified. Second, searches for these key terms were run in all newspaper items under investigation (n = 10,613). Ten terms were mentioned in at least 10% of the newspaper articles (big circles in Figure 1). Nineteen more keywords were mentioned in at least 1% of newspaper articles (small circles in Figure 1). The remaining 37 keywords were discarded. Third, content of the 10 Wikipedia articles on the 10 often-mentioned lemmas were scanned for hyperlinks that would point to topically relevant other lemmas/articles (version of record on 31 January 2010, 11:59 p.m.); only paragraphs above the table of content were scanned, yielding 16 additional purely pull-cued lemmas (rectangles in Figure 1), resulting in a list of 45 lemmas. We clustered lemmas into sub-issues of the Haiti Earthquake issue. We used three categories: (a) Politics, geography, culture; (b) disaster, event and geology; (c) international aid and relief.

Figure 1.

Map of Haiti earthquake-related lemmas in Wikipedia.

Data collection

TV salience

We recorded the number, length, and serial position of issue-related news items for all issues. Twelve TV news time series (one per issue) were constructed in the BBS, whereas we used only one time series (Haiti issue) in the HCS. Journalists order news items in TV newscasts according to the news value they attribute to the items, with the most valuable item up front (Tagesschau.de, 2013). As recipients infer the newsworthiness of events from the serial order (Ruhrmann, 1989), we weighted news items according to their serial position in the newscast.

Newspaper salience

The coverage of newspapers was assessed using the Genios database, searching for the selected issue-specific keywords day-by-day (with some specifications to exclude off-topic hits). The resulting intensities of coverage (number of items) were arranged as 12 newspaper time series (one per issue) in the BBS. Again the HCS uses only the Haiti issue time series.

Push-cueing

The number of mentions of each of the 37 plus 45 lemmas in newspaper coverage (again with specifications to exclude off-topic hits) was used as a proxy to the relative prevalence of the keywords in media coverage (push-cueing).

Pull-cueing

The degree of pull-cueing (HCS only) was assessed by analyzing the 10 Wikipedia articles about keywords mentioned in at least 10% of newspaper articles (pointer articles). For every hyperlink, the number of views of the pointer article is added to the pull-cueing score of the article the link points to (target article). The pull-cueing score represents the popularity-weighted cross-linkage of an article with the other topically relevant Wikipedia articles: a link between two articles should increase the likelihood of viewing the target article. This increase should be the higher, the more popular the pointer article is.

Search data

The number of page views of the 37 (BBS) plus 45 (HCS) articles on Wikipedia was retrieved from the website https://stats.grok.se, which visualizes technically collected data on http://dumps.wikimedia.org. These data were recorded as time series of the raw number of page views per lemma in January 2010 (37 plus 45 time series). Data for January 24–25 were missing in the original database due to technical error. They were replaced using linear interpolation.

Editing data

The number of edits per day was recorded using the Wikipedia article history. We counted changes per lemma and recorded them as 37 (BBS) plus 45 (HCS) editing time series.

Data preparation

An autoregressive integrated moving average (ARIMA) (1,1,0) (searches in the HCS) or ARIMA(1,0,0) (all other time series) model was applied to all time series and removed problematic data patterns. We ran all computations (using the ARIMA residuals) with both linear (dependent variable: raw data) and log-linear regression models (dependent variable: natural logarithm of raw data). Only the log-linear models generated normally distributed residuals. However, we present both the linear and the log-linear models as the linear models react more strongly to major changes in search and editing activities whereas the log-linear models primarily account for small and moderate changes. We account for the differences between models and results in the discussion section in more detail.

Dynamic regression analyses indicated no time lags of one or more days. To make a case for news being responsible for increased viewing of Wikipedia articles, we inspected the hour-by-hour changes of page-viewing activity on Wikipedia directly after the news on the Haiti Earthquake broke (13 January). The share of Haiti-related page views (“Haiti,” “Port-au-Prince,” and “Dominikanische Republik”) on Wikipedia (compared to all page views in Wikipedia) increased strongly from .1% to 1.3% in the early morning (7–10 a.m.). It fell to .5% at 5 p.m. and again increased in the early evening (7–9 p.m.) to 1.0%; Presumably, these were the effects of morning news (mostly radio news and newspapers) and evening news (mostly TV news), respectively.

Findings

The BBS is based on 2,197,105 searches and 2712 edits of 37 lemmas on 12 issues. The HCS is based on 2,360,203 searches and 1712 edits of 45 lemmas in three sub-issues of the Haiti Earthquake issue. On average, a lemma received M = 2728 newspaper stories (SD = 4077; Median = 1025) and the equivalent of M = 5.37 (SD = 6.79; Median = 2.36) Tagesschau headline items in the BBS. The HCS identified 12,730 newspaper stories and the equivalent of 21.7 Tagesschau headline items dealing with the Haiti Earthquake. After ARIMA procedures, both dependent variables (number of searches and edits) were significantly and strongly correlated (r = +.46 in the BBS and r = +.41 in the HCS). TV news was weakly correlated with newspaper coverage (r = +.15; +.14), search (r = +.21; +.13), and editing activity (r = +.24; +.15). Newspaper coverage was only very weakly correlated with search (r = +.07; +.01) and editing activity (r = +.09; +.04).

We assume that, aggregate search and editing behaviors will be more prevalent (a) the more the media-for-monitoring cover the issue (H1, H3), (b) the more often the media-for-monitoring or popular Wikipedia articles mention the respective lemma (H2, H4), and (c) the degree of impact of media-for-monitoring coverage will vary between issues and lemmas (RQ1, RQ2). Technically, these are effects of media-for-monitoring coverage (H1, H3, models 1 and 2), effects of push-/pull-cueing and interactions of media-for-monitoring coverage with push-/pull- cueing of lemmas (H2, H4, models 3 and 4), and interactions of media-for-monitoring coverage and issues or lemmas (RQ1, RQ2, model 5 and saturated model). The performance of models should be compared to the “saturated” model in which the gradient of news coverage’s impact on search activities is estimated for each lemma separately. This is the optimal but also the least parsimonious model against which the performance of more parsimonious models can be benchmarked.

Search behavior

In both BBS and HCS, the volume of searches increases with increasing TV news coverage. Including TV news significantly improves model fit in four out of four cases (model 2). In contrast, including newspaper coverage yields no model improvement (model 1). H1 is therefore supported for TV news coverage but must be rejected for newspaper coverage. However, this crude model explains only between 1% and 3% of variations in search activity (Table 1).

Table 1.

Predicting aggregate search activity.

Predictors	Broadband study (BBS)				Haiti case study (HCS)
	Search volume				Absolute change in search volume
	Log-linear		Linear		Log-linear		Linear
	ΔR ²	B	ΔR ²	B	ΔR ²	B	ΔR ²	B
Model 1: newspapers	.000		.000		.000		.000
(Intercept)		+4.28***		+723^#		+3.44***		+542^#
Newspaper stories		+.03		+8.71		+.02		+52.2
Model 2: TV news	.010***		.029***		.005**		.016***
TV news stories		+.21***		+950		+.16***		+692
Model 3: Push/pull	.316***		.136***		.228***		.095***
Push-cueing (prevalence in news)		+1.15***		+2020**		+.98***		+1669**
Pull-cueing (Wikip. cross-linkage)		–		–		+.35*		+31.9
Model 4: News × push/pull	.003 ^#		.235***		.003		.213***
Push-cueing × TV		+.02		+2199**		+.03*		+2523*
Push- cueing × newspapers		−.07**		−430.3*		+.00		−216
Pull-cueing × TV		–				+.12***		+115
Pull-cueing × newspapers		–				+.00		−38.0
Model 5: Issues × TV	.206***		.116***		.004		.011^#
Total R²	.535***		.516***		.240***		.335***
Satur. model (Lemma × TV) R²	.746***		.556***		.591***		.515***

N_BBS = 1110 (37 lemmas á 30 days; ARIMA(1,0,0)); N_HCS = 1305 (45 lemmas á 29 days; ARIMA(1,1,0)). Linear and log-linear models with cluster-robust standard errors (lemmas as clusters). Unstandardized regression weights. All non-dummy predictors were z-standardized. Estimates for issue dummies (model 5), lemma dummies (saturated model), issue × TV interactions (model 5), and lemma × TV interactions (saturated model) are not displayed. Mixed models (with lemmas as level 2 units) yielded comparable results.

p < .10; *p < .05; **p < .01; ***p < .001;

When including the push-cueing of lemmas (model 3) and its interactions with newspaper and TV coverage (model 4), the explanatory power increases significantly and substantially in both the BBS and the HCS (to between 24% and 40% of explained variance). The findings support H2a. There are differences between linear and log-linear models, however, the linear model fits improve when the interaction terms are included (model 4) whereas the log-linear model fits improve when the main effect of push-cueing is added (model 3). H2b receives only partial support in the HCS and cannot be tested in the BBS. Hyperlinks from pointer Wikipedia articles (pull-cueing) seem to attract additional views of the target article if a log-linear relationship is assumed, but not if a linear relationship is assumed. The same applies to the interaction between pull-cueing and TV news stories.

The magnitude of increase in Wikipedia article views due to TV news coverage is contingent on the respective issue in the BBS but not in the HCS (model 5; RQ1a). The issues investigated in the BBS were more distinctive compared to the sub-issues in the HCS, and hence have more explanatory power. Searches for lemmas are best explained by TV news coverage if regression gradients are estimated for each lemma separately. These “saturated” models explain between 52% and 75% of variation in search activity, respectively. However, it is questionable whether the additional explanatory power of the “saturated” models justifies the higher complexity compared to models 4 and 5 (RQ1b).

Which issues and which articles were more popular or less popular among users than predicted based on their prevalence in media-for-monitoring coverage? In the HCS, issues neither influenced the number of searches nor moderated the impact of salience of the issue in the news; only the lemma “2010 Haiti Earthquake” was searched for more frequently than predicted. The likely reason is that this was the article explicitly treating the earthquake event covered in the news whereas other articles were only thematically related to that event. In the BBS, users reacted less responsive than expected to coverage of the issues “Left Party,” “Winter storms,” and “State visit to Turkey,” whereas they reacted more responsive than expected to the “Haiti” and “Yemen” issues. The variations in responsiveness may reflect that Wikipedia—as an encyclopedia—does not feature current information about political conflicts, the weather, or state visits and hence was not the source of choice for most potential users.

Editing behavior

In the BBS and the HCS, the number of edits were not predicted by the amount of media-for-monitoring coverage when log-linear models were used. The linear models, however, reveal that the number of Tagesschau items and newspaper stories affect editing behavior, as indicated by significant improvements in model fit (models 1 and 2). This partially supports H3 (for linear, but not log-linear models). Even in linear models, news coverage alone accounts for only 1% to 3% of variation of editing activities (Table 2).

Table 2.

Predicting aggregate editing activity.

Predictors	Broadband study (BBS)				Haiti case study (HCS)
	Number of edits				Number of edits
	Log-linear		Linear		Log-linear		Linear
	ΔR ²	B	ΔR ²	B	ΔR ²	B	ΔR ²	B
Model 1: Newspapers	.000		.000		.000		.000
(Intercept)		−1.11**		+1.65***		−2.02***		+1.02**
Newspaper stories		+.03		+0.11		+.05^#		+.07*
Model 2: TV news	.000		.032***		.000		.014***
TV news stories		+.02		+.90		−.01		+.53
Model 3: Push / pull	.185***		.120***		.230***		.147***
Push-cueing (prevalence in news)		+.99**		+1.74***		+1.07***		+1.72***
Pull-cueing (Wikip. cross-linkage)		–		–		+.30		−.06
Model 4: News × push/pull	.001		.107***		.001		.110***
Push-cueing × TV		+.06**		+1.36***		+.05*		+1.50**
Push- cueing × newspapers		+.01		−.11 ^#		−.02		−.23*
Pull-cueing × TV		–				+.05		−.03
Pull-cueing × newspapers		–				+.03		+.07
Model 5: Issues × TV	.339***		.055***		.031***		.019 ^#
Total R²	.525***		.314***		.262***		.290***
Satur. model (Lemma × TV) R²	.866***		.537***		.861***		.540***

N_BBS = 1110 (37 lemmas á 30 days; ARIMA(1,0,0)); N_HCS = 1305 (45 lemmas á 29 days; ARIMA(1,0,0)). See Table 1 for details on calculation and tabulation.

p < .10; *p < .05; **p < .01; ***p < .001;

The more frequently the media-for-monitoring mention a lemma-related keyword (push-cueing), the more often do users edit the corresponding Wikipedia articles (model 3); furthermore, push-cueing becomes more effective if TV news emphasize the issue (model 4). The findings support H4a. H4b is rejected in the HCS and cannot be tested in the BBS: articles that are better linked to other relevant Wikipedia articles (pull-cueing) are not edited significantly more often than other articles (model 3), and neither is there an interaction between pull-cueing and TV or newspaper coverage (model 4).

TV coverage more strongly impacts editing activities in particular issue categories. In the HCS, this effect is very weak, however. Within the BBS, the moderating effect of issues is stronger when a log-linear (rather than a linear) model is used (RQ2a). The impact of TV coverage on editing activities is also contingent on the particular lemma. Both in the BBS and in the HCS, adding the lemma/TV news interaction improves the explanatory power of the models by over 20 percentage points (RQ2b).

Again, what issues and lemmas were users particularly responsive or irresponsive to when it comes to editing? In the HCS, the article on “Earthquake in Haiti 2010” was edited more often than expected. Also, articles on “Port-au-Prince” and “Haiti” were edited more often than predicted. The continuing stream of news about Haiti obviously necessitated more updating of these articles compared to other articles. The BBS suggests the same conclusion: Haiti-related articles were edited significantly more often than predicted.

Temporal patterns

Visual inspection

The development of media-for-monitoring coverage, collective search activity in Wikipedia, and editing activity in Wikipedia is displayed in Figure 2 (HCS only). It is apparent that search activities (half-life: 8 days) and editing activities (half-life: 7 days) fade more quickly compared to TV news (half-life: 13 days) and newspaper coverage (half-life: 18 or more days) of the issues. This visual inspection supports H5 for the raw time series data.

Figure 2.

Temporal dynamics of Wikipedia search and editing activities (HCS; aggregated across 45 lemmas), Tagesschau coverage, and newspaper coverage. All values standardized to a range from 0 (minimum) to 1 (maximum). Raw peak values are given in brackets.

Statistical analysis

The relatively long time during which search and editing activities remain above-average may be explained via autocorrelation as pure reverberations of an initial stimulus. Therefore, using techniques from time series analysis can more adequately trace the duration of the effects exerted by media-for-monitoring coverage, wherefore we use the ARIMA-treated time series for a statistical analysis. We tried to predict search activities and editing activities by (a) locating the “take-off” point in search and editing activities (i.e. the day on which activities rose most pronouncedly for the article at hand) and (b) using idealized effects progressions (impulse models) with different half-lives: (1) The effect occurs only on the take-off day and instantly declines to zero the next day (no reverberations model, NRM); The effect halves (2) every day (HL1); (3) every four days: (HL4); (4) every eight days (HL8); (5) every 16 days (HL16). Apart from using a different main regressor, models resembled those used to analyze H1–H4 and RQ1–RQ2. As the findings in the BBS and the HCS were similar, we present only the HCS findings.

According to the log-linear models, the HL8 and the HL16 models have the highest explanatory power to account for search activity and all models perform similarly in predicting editing behavior. These findings support the notion that the effect of TV news is relatively long-lived. The linear model suggests just the opposite: media-for-monitoring coverage has an immediate yet short-lived effect on search behavior (Table 3) and editing behavior (Table 4). The NRM model outperforms every other model, indicating that the effect completely vanishes after only one day and all subsequent above-average searches can be interpreted as reverberations of the initial “shock,” that is, the initial increase in searches. This predicts search behavior much better than the model based on actual media-for-monitoring coverage. The findings from the linear model strongly support H5, while the findings from the log-linear models contradict H5.

Table 3.

Impulse models of search activity (HCS).

Change in search volume	Input	Impulse models: half-life
	TV news	0 days	1 day	4 days	8 days	16 days
Predictors	R ²	R ²	R ²	R ²	R ²	R ²
Log-linear models
Model 1: input/impulse ΔR²	.005***	.018***	.029***	.057***	.074***	.072***
Model 2: push/pull ΔR²	.228***	.228***	.227***	.228***	.227***	.227***
Model 3: push/pull × input ΔR²	.003*	.010***	.014***	.018***	.018***	.021***
Saturated model R²	.591***	.615***	.639***	.683***	.711***	.711***
Linear models
Model 1: input/impulse ΔR²	.016***	.033***	.030***	.024***	.017***	.011***
Model 2: push/pull ΔR²	.095***	.094***	.095***	.094***	.094***	.095***
Model 3: push/pull × input ΔR²	.213***	.439***	.393***	.291***	.291***	.133***
Saturated model R²	.515***	.911***	.819***	.636***	.480***	.365***

N_HCS = 1305 (45 lemmas á 29 days; ARIMA(1,1,0)). Linear and log-linear models with cluster-robust standard errors (lemmas as clusters). Unstandardized regression weights. All non-dummy predictors were z-standardized. Estimates for lemma dummies and lemma × input/impulse interactions (saturated model) are not displayed. Mixed models (with lemmas as level 2 units) yielded comparable results.

p < .10; * p < .05; ** p < .01; ***p < .001

Table 4.

Impulse models of editing activity (HCS).

Number of edits	Input	Impulse models: half-life
	TV news	0 days	1 day	4 days	8 days	16 days
Predictors	R ²	R ²	R ²	R ²	R ²	R ²
Log-linear models
Model 1: input/impulse ΔR²	.000	.000	.000	.001*	.001*	.000 ^#
Model 2: push/pull ΔR²	.229***	.230***	.230***	.230***	.230***	.229***
Model 3: push/pull × input ΔR²	.231*	.231***	.231***	.232***	.231**	.231*
Saturated model R²	.861***	.867***	.863***	.863***	.862***	.861***
Linear models
Model 1: input/impulse ΔR²	.014	.030***	.027***	.020***	.013***	.008***
Model 2: push/pull ΔR²	.161***	.177***	.174***	.167***	.160***	.155***
Model 3: push/pull × input ΔR²	.271***	.416***	.364***	.295***	.242***	.206***
Saturated model R²	.540***	.876***	.730***	.539***	.428***	.362***

N_HCS = 1305 (45 lemmas á 29 days; ARIMA(1,0,0)). See Table 3 for details on calculation and tabulation.

p < .10; * p < .05; ** p < .01; ***p < .001

At the side, models 2 and 3 provide additional support for the partially supported H2b: At the time of the “shock,” those articles with a stronger cross-linkage in Wikipedia (pull-cueing) are searched more often than articles with less cross-linkage in Wikipedia, and the cross-linkage significantly interacts with the impulses.

Discussion

Interpretations

Linear versus log-linear models

The results of our calculations are strongly contingent on the choice of statistical models. In contrast to linear models, the meaning of coefficients in log-linear models depends on the magnitude of the initially predicted value: for example, in a log-linear model, a coefficient of B = +.50 means that the number of page views of a lemma increases by the factor 1.65 or 65% (e^0.50 = 1.65) if media-for-monitoring coverage increases by one SD unit. If our initial expected value is 100 page views, this corresponds to an absolute increase of 65 page views, whereas on an already popular page with an expected 100,000 page views, the log-linear model with the same coefficient of B = +.50 would predict an absolute increase of 65,000 page views. Compared to the linear models, the log-linear models will decrease the leverage of the few cases with very high numbers of page views. These outliers are re-interpreted as mild rather than as extreme outliers. We argue that both models are valid but primarily account for different things: the linear model accounts for the cases with extremely high search and edit counts—for example, the enormous increase of search activities on Haiti or on earthquakes after the earthquake on 12 January. The log-linear model primarily explains the differences within the bulk of cases with moderate and low search and edit counts.

Contingency on mass media

All findings point in the direction that media-for-monitoring coverage plays two important parts in shaping search and editing behaviors: first, major increases in media-for-monitoring coverage of an issue lead to an increase in searches for H1 and edits of H3 issue-related articles in Wikipedia. Second, media-for-monitoring coverage suggests keywords to the media users which they could search for H2a. Searches for and edits of articles whose lemmas corresponded to keywords often mentioned in media-for-monitoring coverage were more strongly affected by the media’s initiating stimulus. That newspaper coverage had no effect on search activities can be attributed to the fact that the earthquake (12 January, 10:53 p.m. CET) was not included in most morning newspapers on 13 January 2010, leading to a statistical null finding. The differences between linear and log-linear models indicate that extreme increases in search and editing activities are explained by increased TV news salience and interactions between TV news salience and push-cueing (linear model, Models 2 and 4), whereas the degree of push-cueing by itself explains the differences within the bulk of cases with moderate and low search and edit counts (log-linear model, Model 3).

Wikipedia cross-linkage

Only the log-linear models show an effect of cross-linkage within Wikipedia (pull-cueing, H2b), which suggests that the relevance of this mechanism is limited to explaining low and moderate variations in search activity. That the effects of Wikipedia cross-linkage on search activities are weaker (H2b) and that no effects on editing are detectable (H4b) is in line with the fact that only few individuals encounter these cues within Wikipedia, hence more than weak effects would not be expected.

Contingency on issues

Our findings suggest that search and editing activities are somewhat issue-dependent (RQ1 and RQ2), particularly in the log-linear BBS model (Model 5). This indicates that issues matter in explaining differences within the bulk of low and moderate search and edit counts (log-linear models), particularly if fundamentally different issues are under study (BBS). Issues were less important in predicting extreme outliers (linear models) and if issues are mere sub-issues (HCS). Issue contingency—even when controlling for push-cueing intensity—may indicate that news users do not turn to Wikipedia if they believe other sources are better suited for their current information need.

Time frame

When removing autocorrelation from the time series, the independent effect of media-for-monitoring coverage on search and editing activities vanished within one day’s time in the linear models. Actually, the time-spans are not as low as the findings seem to indicate: search and editing activities remained above-average for several more days. The conservative procedures of time series analysis attribute this to reverberations of the initial stimulus, but it could also be attributed to continuing media-for-monitoring coverage. These findings primarily explain major outliers, that is, the major increases in search and editing activities. The log-linear models suggest that in the majority of cases where a low or moderate level of search and editing activities is to be predicted, time frames are generally less important and initial impulses lead to longer reverberations than the linear models suggest.

Search and editing behaviors decline to normal levels when media-for-monitoring coverage still emphasizes the corresponding issue (Figure 2). An optimistic interpretation is that recipients quickly close their subjective information gaps. A pessimistic interpretation is that recipients quickly get bored or frustrated by the issue or habitualize their attention. Most probably, many recipients discover a gap between their aspired and their current levels of knowledge when news breaks. Many recipients get the information they feel they need from the media-for-monitoring coverage they receive—this is one reason why the daily views of the relevant Wikipedia articles is in the 100,000s and not in the millions. Others seek more or other information and draw on a variety of sources, Wikipedia being an important one among them. However, the information in Wikipedia articles is encyclopedic and therefore rather static. Therefore, most users will view Wikipedia pages quickly after the news breaks and will not return in the next days or weeks, which explains why searches and edits in Wikipedia decline more quickly than does coverage in media-for-monitoring.

In sum, the reconstruction of individual-level mechanisms is in line with the hypotheses derived from the DTM, while other models of media use and media effects cannot account for our findings in a convincing manner. Whether our reconstructions actually capture the individual-level mechanisms needs to be checked using a different methodology with individual-level data. An obvious limitation of the DTM is that it was designed for more traditional media environments—a thorough theoretical reworking would be in place.

Limitations of the study

The two studies reported here are limited in several ways. First, we used only one TV newscast to represent TV news coverage. Second, we focused on search and edits in Wikipedia, ignoring other relevant media-for-searching. Third, our analysis looks at collective information seeking while we argue on an individual level. However, the collective-level patterns are in line with our individual-level arguing from which we derived our hypotheses which we designed to bridge individual and collective levels.

Implications and conclusion

Our analysis clarified how media users combine information services with different strengths and weaknesses. Media-for-monitoring suggest what issues recipients can seek information about and what keywords they use (push) to retrieve relevant knowledge from media-for-searching (pull), in our case from Wikipedia. This complementary use of both media outlets may reinforce the formation of public opinion.

Of course not every user of media-for-monitoring searches for information in Wikipedia: some do not search for any information at all, others fail to retrieve any information, or they retrieve information from other sources. However, the overall volume of search increase in Wikipedia due to media-for-monitoring coverage is tremendous: according to our calculations—which use search activity for the same lemmas in January 2009 and January 2011 as a benchmark—more than 1,700,000 additional searches were triggered by the media-for-monitoring coverage of the Haiti earthquake alone. A substantial number of people received a substantial amount of additional information because TV and newspaper coverage motivated people to search for issue-related information in Wikipedia. Particularly regarding the Haiti earthquake, a substantial number of people extended their knowledge about earthquakes, the underlying seismological and geological mechanisms, or the geographical, political, and social context of Haiti. Furthermore, many people were motivated to read about emergency aid agencies, organizations, and programs. They may have made up their minds as to whether and how much to donate to which organization. Such information-seeking practices conform to the ideal of the monitorial citizen (Schudson, 1998), which posits that citizens engage in in-depth information seeking once their need for sense-making is evoked (Zaller, 2003). Media-for-monitoring coverage of these issues also increased the scrutiny of the thematically related Wikipedia articles, stimulating collective efforts to improve quality of Wikipedia articles.

Footnotes

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Author biographies

Stefan Geiß is a post-doctoral researcher at the Department of Communication at the University of Mainz. His research focuses on agenda and frame building, agenda and frame setting, political scandals, and audience–journalist relations.

Melanie Leidecker is a post-doctoral researcher at the Department of Communication Psychology and Media Education at the University of Koblenz-Landau. Her areas of interest include political communication, election campaigns, journalism studies, and media content analyses.

Thomas Roessing was a post-doctoral researcher professor at the Department of Communication at the University of Mainz and is now project manager at mct, a private media research firm in Dortmund, Germany. His research focuses on public opinion, political communication, and Wikipedia.

References

Brown

(1999) FEAR.COM. State of the American newspaper. American Journalism Review 21(5): 50–71.

Busemann

Engel

(2012) Wandel der Mediennutzungsprofile im Zeitalter des Internets [Change of media use in the Internet age]. Media Perspektiven 2012(3): 133–146.

Busemann

Gscheidle

(2011) Web 2.0: Aktive Mitwirkung verbleibt auf niedrigem Niveau [Web 2.0: active participation remains low]. Media Perspektiven 2011(7–8): 360–369.

Case

(2012) Looking for Information: A Survey of Research on Information Seeking, Needs and Behavior. London: Emerald.

Chaffee

Metzger

(2001) The end of mass communication? Mass Communication and Society 4(4): 365–379.

Dervin

(1998) Sense-making theory and practice: an overview of user interests in knowledge seeking and use. Journal of Knowledge Management 2(2): 36–46.

Dutta-Bergman

(2004) Complementarity in consumption of news types across traditional and new media. Journal of Broadcasting and Electronic Media 48(1): 41–60.

Früh

(1991) Medienwirkungen: Das dynamisch-transaktionale Modell—Theorie und empirische Forschung [Media Effects: The Dynamic-Transactional Model—Theory and Empirical Research]. Opladen: Westdeutscher Verlag.

Früh

Schönbach

(1982) Der dynamisch-transaktionale Ansatz: Ein neues Paradigma der Medienwirkungen. [Dynamic-transactional approach: a new paradigm of media effects]. Publizistik 27: 74–88.

10.

Geiß

(2013) Zwischenmediale Einflüsse: Koorientierung und Verselbständigungstendenzen der Medien in Orientierungs- und Routinephasen [Inter-media influence: co-orientation and disengagement of mass media in orienting and routine phases]. In: Quiring

Kepplinger

Weber

Geiß

(eds) Lehman Brothers und die Folgen: Berichterstattung zu wirtschaftlichen Interventionen des Staates [Lehman Brothers and the Aftermath: Coverage of Government Interventions into Economy]. Wiesbaden: Springer VS, pp. 91–110.

11.

Granka

(2010) Measuring agenda setting with online search traffic: influences of online and traditional media. Paper presented at the 2010 annual meeting of the American Political Science Association, 5 September. Available at: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1658172

12.

Holbach

Maurer

(2014) Wissenswerte Nachrichten: Agenda-Setting-Effekte zwischen Medienberichterstattung und Online-Informationsverhalten am Beispiel der EHEC-Epidemie. [News worth knowing: agenda setting effects between media coverage and online information gathering—the case of EHEC]. Publizistik 59: 65–81.

13.

Howard

Massanari

(2007) Learning to search and searching to learn: income, education, and experience online. Journal of Computer-Mediated Communication 12: 846–865.

14.

Jung

Lin

Kim

(2012) The dynamic relationship between East Asian adolescents’ use of the internet and their use of other media. New Media & Society 14(6): 969–986.

15.

Kayany

Yelsma

(2000) Displacement effects of online media in the socio-technical contexts of households. Journal of Broadcasting and Electronic Media 44(2): 215–229.

16.

Kuznetsov

(2006) Motivations of contributors to Wikipedia. Computers and Society 36(2): 1–7.

17.

Negroponte

(1995) Being Digital. New York: Alfred A. Knopf.

18.

Nguyen

Western

(2006) The complementary relationship between the Internet and traditional mass media: the case of online news and information. Information Research 11(3): Article 259.

19.

Nov

(2007) What motivates Wikipedians? Communications of the ACM 50(11): 60–64.

20.

Roessing

(2009) Wikipedia and public opinion: the online encyclopedia as a transmission belt for political communication. Paper presented at the second Cairo international conference on public opinion polls “Public Opinion Polls in a Changing Society,” Cairo, Egypt, 8–9 November.

21.

Rose

Levinson

(2004) Understanding user goals in web search. In: Feldman

Uretsky

(eds) Proceedings of the 13th international conference of the World Wide Web. New York: ACM, pp. 13–19.

22.

Rössler

(1999) The individual agenda designing process: how interpersonal communication, egocentric networks, and mass media shape the perception of political issues by individuals. Communication Research 26(6): 666–700.

23.

Ruhrmann

(1989) Rezipient und Nachricht: Struktur und Prozess der Nachrichtenrekonstruktion [Media User and News: The Structure and Process of News Re-Construction]. Opladen: Westdeutscher Verlag.

24.

Scharkow

Vogelgesang

(2011) Measuring the public agenda using search engine queries. International Journal of Public Opinion Research 23: 104–113.

25.

Schroer

Hertel

(2009) Voluntary engagement in an open web-based encyclopedia: Wikipedians and why they do it. Media Psychology 12(1): 96–120.

26.

Schudson

(1998) The Good Citizen: A History of American Civic Life. New York: The Free Press.

27.

Shoemaker

(1996) Hardwired for news: using biological and cultural evolution to explain the surveillance function. Journal of Communication 46(3): 32–47.

28.

Tagesschau.de (2013) Wir über uns [About us]. Available at: http://intern.tagesschau.de

29.

Van Deursen

Van Dijk

(2011) Internet skills and the digital divide. New Media & Society 13(6): 893–911.

30.

Wirth

Böcking

Karnowski

. (2007) Heuristic and systematic use of search engines. Journal of Computer-Mediated Communication 12: 778–800.

31.

Zaller

(2003) A new standard of news quality: burglar alarms for the monitorial citizen. Political Communication 20(2): 109–130.

32.

Zubayr

Gerhard

(2011) Tendenzen im Zuschauerverhalten: Fernsehgewohnheiten und Fernsehreichweiten im Jahr 2010. [Trends in audience behavior: habits of TV viewing and audience ratings in 2010]. Media Perspektiven 2011(3): 126–138.