WorksComputer Science
Measuring Organizational Influence on Digital News: An Information-Theoretic Framework for Media Pluralism and Editorial Independence
Abstract
News consolidation is accelerating, yet regulators have no measurement frameworks for determining whether owners shape editorial decisions. Existing evidence comes from single cases, where pluralism and independence are measured separately, if at all. We ask whether ownership impacts editorial decisions across the digital news ecosystem, and if so, how acquisitions affect this. We derive editorial independence and internal pluralism using partial information decomposition, which separates news coverage overlap with sister outlets from coverage overlap with all outlets. Using public datasets, we apply our measures monthly from 2018 to 2026 to 1,364 US news websites on three editorial decisions: which events they cover, which sources they link to, and what they put on their front pages. We find that co-owned outlets share up to 40 times more information about these decisions than expected by chance. Ownership accounts for far more variation in independence than network affiliation does and, upon being acquired, outlets lose independence sooner than matched controls, quickly converging on their new owner. In broadcast news, by comparison, we find that it is network affiliation, not ownership, that accounts for more content overlap. Finally, we discuss policy implications and suggest that media regulators have blocked large deals rather than editorially detrimental ones. We hope that this work can serve as a framework for comparative media analyses going forward.
CCS Concepts
• Information systems → Web mining; • Social and professional topics → Governmental regulations; • Mathematics of computing → Information theory.
Introduction
Paramount Skydance’s $110 billion acquisition of Warner Bros. Discovery, approved by the DOJ and the FCC, places CBS News and CNN under one owner. Together, their outlets now reach 97–99% of US households. Given concerns from media critics, a five-journalist editorial board to protect the two newsrooms’ independence is being established.
Yet it is unclear how such a board would know whether independence has been maintained, or what would erode it. While research on organizational influences over editorial decisions is well established, and there is general agreement that such influence exists, current research falls short of measuring it. Existing studies rely on case studies and adopt heterogeneous methods for analysis [36], treating pluralism and independence as unrelated measures. What’s more, regulators seem to approve acquisitions based on audience reach and local market overlap, disregarding editorial decisions. No measure tells a regulator before a deal, or an editorial board after it, the degree to which editorial decisions may change. In terms of transparency for the information ecosystem, no comparative measures exist to inform the public about which news outlets have the least editorial independence.
However, editorial decisions leave traces that can be measured at Web scale, such as which events an outlet covers (GDELT), which sources it cites (Common Crawl web graphs), and which stories it puts on its front page (GDELT Global Frontpage Graph). Because editorial decisions are documented at Web scale, they can be used to audit owner influence, which is a question of information integrity and media governance. We ask four questions. Does ownership organize outlets’ editorial decisions? If so, how much, relative to network affiliation? Do organizational changes such as acquisitions and layoffs impact editorial independence? And do regulators block the deals that most reduce editorial independence? We define pluralism and independence from the same primitives: an outlet’s monthly editorial decisions, and the information that its sibling outlets, and the rest of the ecosystem, convey about them. A partial information decomposition separates what an outlet shares specifically with its siblings from what it shares with any other outlet; with political bias as the grouping, the measure is internal pluralism, and with ownership, editorial independence. We demonstrate that pre-existing measurement methods cannot account for this and instead mistake the common news agenda, such as content from wire services such as the Associated Press (AP News), as evidence of low editorial independence between sibling stations. We then apply our measures to Web-scale datasets covering 1,364 US news websites, monthly from 2018 to 2026. A generalized additive mixed model sizes the owner effect against network affiliation, and matched survival designs follow outlets after acquisitions and layoffs.
Compared to the shared information between two random outlets, co-owned outlets share ∼ 20× more information for event selection, ∼ 2× more for (unweighted) source selection, and ∼ 40× more for front-page placement. Meanwhile, outlets that share only a network affiliation have reduced, though still significant, overlap. Owners account for far more variation in independence than affiliate networks (owner SD 0.72, affiliate SD 0.15). This effect depends on media type, however, and we find that it reverses when considering content from television broadcasts. On air, stations’ shared coverage follows their network affiliation; online it follows their owners.
Acquired outlets lose independence sooner than matched controls (hazard ratio 2.12 at six months), quickly converging on their new owner. Projecting these trends over past deals, we suggest that regulators have blocked large deals rather than ones that stand to reduce editorial independence. In this way, our measures can track acquired newsrooms over time, giving editorial boards such as the one for CBS News and CNN a benchmark. We contribute a unified framework for editorial independence and pluralism on three editorial decisions, the findings derived from them, and an ownership panel of 1,364 US news websites.
Related Work
Studies of organizational influence on editorial decisions broadly agree that such influence exists. Of the 56 empirical studies reviewed by Theine et al. [36], 33 report strong ownership influence, 13 weak influence, seven no systematic relationship, and three are unclear. However, because the studies are methodologically heterogeneous and mostly examine a few outlets, a comparative analysis gap remains.
Pluralism. Comparative media research distinguishes external pluralism, a system of outlets that each carry a different political tendency, from internal pluralism, balance within each outlet [13, 26]. Content studies measure these with diversity indices such as Shannon entropy over various categories of content like topics, actors, and viewpoints [21, 24].
Whether an index measures internal or external pluralism depends on whether it is applied to a single outlet’s distribution or to the whole market. We apply this decomposition to editorial decisions. At time and for a given type of editorial decision, with indicating that outlet covered item (an event, or a domain it cites), let be outlet ’s coverage distribution, its share of all coverage, and the ecosystem-level coverage distribution. Total diversity decomposes into the diversity of each outlet’s coverage distribution separately and the differences between those distributions [18]:
A system is externally plural when total diversity is high and mostly supplied by the between-outlet term, and internally plural when it is mostly supplied by the within-outlet term. The between-outlet term is the mutual information between two outlets in terms of the items they cover, so external pluralism is high when outlets select different items. However, this formulation alone does not explain why some outlets select alike and others do not. To answer this, our measure introduces a further decomposition where we split the shared coverage information into the part specific to an outlet’s own group (its owner or political bloc) and the part it shares with the rest of the media ecosystem (Section 3).
Ownership and editorial independence. Policy monitors use institutional safeguards [4] and survey studies from expert ratings [14, 30] to score editorial independence at the country level. Studies
that measure an owner’s influence over its outlets’ output do so in two ways. The first is the owner’s share of the variance in an outlet trait, as with the owner random effect on newspaper slant of Gentzkow and Shapiro [12]. The second is the change in coverage after an acquisition, relative to comparison outlets [2, 23, 25]. The review of Theine et al. [36] names online news and cross-national comparison as gaps in this literature, which this work addresses. We estimate the owner’s variance share for each editorial decision with our central model (Section 5).
The closest outlet-level measure of independence is content recycling, the share of an outlet’s article titles that also appear in co-owned outlets [15]. With as in (3), let indicate that some other outlet in ’s owner group did (Section 3.1). Content recycling is then .
Crucially, this measure can present ecosystem-level redundancy as co-ownership overlap. In contrast, our partial information decomposition (PID) [40] separates the two: the information about an outlet’s coverage that is conveyed by its owner is split into what the owner uniquely provides () and what the rest of the ecosystem also provides (redundancy ; Figure 1; Section 3):

Figure 1. The PID of the information that outlet ’s co-owned siblings have about ’s editorial decisions. is unique information, what only ’s group predicts about its coverage, is redundancy, what group and field both predict, and is synergy, what only their combination predicts.
We apply the same construction with political bias in place of ownership, so internal pluralism and editorial independence are computed from a unified framework. While several web-scale studies use political bias to describe news ecosystems from hyperlink and co-coverage networks [6, 29, 31], we know of no work that measures pluralism and independence together.
Measures
We define outlet measures as ratios of the partial information decomposition (PID) atoms of the outlet’s coverage, with pairwise quantities derived with respect to the partial entropy decomposition (PED). Each measure is defined for both the political bias and parent owner outlet groupings. For a mapping between the two sets of measures, see Appendices C and F.
Coverage variables
An outlet is active in window (a GDELT month, a Common Crawl graph, or a Sunday front page) if it covers at least 50 items in it; is the set of active outlets. Items are GDELT events, cited domains or front-page events, each representing a different editorial decision (Section 5). For an item drawn uniformly at random,
where covering means publishing an article about the event, or linking to the domain. To reduce noise from overly granular event coding in GDELT, we filter out the items covered by fewer than active outlets:
Because the filter itself induces dependence between outlets, we test the measures against a null model where they are recomputed after conducting degree-preserving randomization on the underlying data, which preserves each outlet’s number of items and each item’s number of covering outlets (Appendix C).
A grouping assigns outlet to a group : its bias category, or its owner in month . Two sources describe each item from ’s point of view:
how many other outlets of ’s group covered , and how many outlets outside the group did, binned on a log scale. Binning is used to restrict to a six-state variable, so that estimation of the PID’s joint table is tractable. Under , ’s co-owned outlets in the same bias category are dropped from both sources, so bias and ownership are not conflated.
Outlet measures
The information the two sources carry about ’s coverage is split by PID [3, 40] into four non-negative atoms (Figure 1):
Internal pluralism () and editorial independence () are the shares of the explained information that is not specific to the group:
Hence, an outlet that covers exclusively what every outlet covers still has a high . The characteristic that distinguishes from content recycling (§ 2) is that shared-agenda coverage of this nature is assigned to instead of . To demonstrate this distinction, we also report the share of ’s coverage entropy that its full set of co-owned outlets leaves unexplained: (). As in content recycling (§ 2), counts shared-agenda coverage as dependence.
Pairwise measures
For two outlets, the PED of their joint coverage has a shared atom and a synergistic atom , and their mutual information is the difference, [16]. We compare outlet pairs by the normalized mutual information
the share of the smaller coverage entropy that the two outlets share. Weighting the ordinal distance between bias categories, , by this coupling gives ’s cross-bias breadth, . Unlike , uses the ordering of bias categories, where coupling with a distant category counts more than coupling with a neighboring one. Building this ordering into the PID would require a group source, , with one state per category. While not intractable, this would make PID estimates unreliable at our sample sizes (Appendix C.3). We therefore keep the PID measures nominal and only use ordinal variables for political bias in .
Data
Event selection. Event selection comes from the Mentions table of the Global Database of Events, Language, and Tone (GDELT) [20], which records the source domains that published articles about each news event that GDELT extracts. We aggregate these events by calendar month. During periods where chain-owned sites blocked GDELT’s crawlers, we mask the affected outlet-months.
Source selection. Source selection comes from the Common Crawl domain-level web graphs [8], which record links between domains. We parse each of the 55 web graphs published since February 2018 as a separate snapshot.
Front-page placement. Front-page placement comes from the GDELT Global Frontpage Graph (GFG) [35], which snapshots the links on about 50,000 news homepages every hour and records each link’s position on the page. We use one Sunday a month, starting in 2018. We note that, as of October 2025, GDELT appears to have stopped actively maintaining the GFG. Measurements made on this data after that point are therefore unreliable.
Outlets and other outcomes. Outlets are the US news domains of our ownership panel, which labels each domain with its parent owner month by month and with its network affiliations (Appendix A); political-bias labels in five ordinal categories come from Media Bias/Fact Check [39]. We also measure productivity (GDELT article and byline-author counts) and engagement (Ahrefs organic web traffic estimates [1]). Byline authors are the distinct names in the author field GDELT extracts from each article, split and normalized (lower case, letters only) and counted per outlet-month. Wire dependence is the share of an outlet’s bylined articles in the month that credit the Associated Press, by name, as “AP” or as an AP desk.
Organizational events. Treatment groups for our survival analysis are made up of outlets from organizations that experienced layoffs of at least 100 jobs. This data is collected from public trackers of newsroom cuts [37]. For US news media acquisitions and mergers, ranging from 2018 to mid-2023, we use pre-existing datasets [11, 22] (see Appendix § 11 for more details). Failed deals (11 blocked or abandoned, 2017–2026) are compiled by hand from FCC orders, DOJ announcements, SEC filings and trade press.
Model
We fit one generalized additive mixed model per outcome (outlet , month ) with mgcv [41]:
where is the outlet’s AP byline share (defined when , i.e., the outlet published bylined articles that month), a smooth of calendar time, and random intercepts for the owner of in month and its network affiliate, and a per-outlet smooth of time. Because each outlet has its own smooth, the owner effect measures what co-owned outlets have in common rather than differences between individual outlets. Where residuals remain correlated from month to month, we refit with first-order autoregressive errors within outlets. Finally, the ownership effect is the owner random-effect SD in outcome-SD units, where zero variance indicates no owner influence [12].
Outcome measures. Outcomes for outlet are and on each of three types of editorial decisions. We define the items and of (3) as the GDELT events covered (event selection, monthly), domains links to in a Common Crawl graph (source selection), and GDELT events linked from ’s homepage on one Sunday a month (front-page placement).
Results
First, we establish the existence of organizational influence over three editorial decisions. Then, we turn to quantifying the level of influence using the GAMM (9), investigating the impact of organizational changes such as acquisitions and layoffs, and comparing digital with broadcast news. Finally, we explore how platform policies impact digital news outlets with a longitudinal analysis of online engagement data.
Ownership and editorial decisions
Co-owned outlets make the same editorial decisions far more often than other outlets (Figure 2). Same-owner pairs share 20.4 times the information of the shuffled-owner null in event selection, 2.2 times in source selection and 40.2 times in front-page placement. In contrast, pairs that share only a network affiliation sit just above the label-shuffled null, and share a significant proportion of their event-selection information only (Figure 2a).
Figure 2. Shared editorial decisions by co-organizational type: shared owner and affiliation, shared owner only, shared affiliation only, and no relation. Each plot is the distribution of pairwise NMI (8) between two outlets’ monthly decisions. Diamonds show the median of the owner- and affiliation-randomized null.
We also see that median shared information between pairs of outlets that are co-owned and affiliated is relatively small for event selection (12%). In other words, 12% of the information in the smaller outlet in these pairs is redundant given the other outlet. In contrast, source selection (~43%) and front-page placement (~35%) have more shared information between co-owned and affiliated outlets. For source selection, higher rates of shared information are a result of the higher density and connectivity of the web graph in comparison with the co-event coverage graph. Credible news outlets eventually link to one another at least once, so linking to the same sources is common. This is evident in Figure 2b where even random outlets share 13% of their information. We note that the Common Crawl web graph is binarized, so this analysis does not factor in link weights. In addition, NMI scores over GDELT events increase when computed between outlet groups of increasing size, as event coverage overlaps increase (see Figure 8 in the appendix). For front-page placement, the higher rates of shared information likely result from a tradeoff between the limited space on the front page and the need for a news outlet to have major news stories (e.g., the shared agenda) represented there.
Figure 8. Event selection (GDELT): NMI between an outlet’s coverage and the binned summed coverage of other outlets, by group size ; “all” uses every other outlet (for same-owner groups, all co-owned outlets). Lines are medians and bands interquartile ranges over 300 draws per in each of the 84 monthly windows. Dashed: the same draws on Curveball (degree-preserving) randomizations. Dotted: the plug-in bias floor ( shuffled over events). Same-owner groups are shown where at least three owners have outlets.
Across all three editorial decisions, a shared affiliation on its own implies a lack of shared information between outlets. We note, however, that this is primarily the case for digital news, and that media type plays an important role in moderating the role of network affiliations in editorial decision making. We return to this in § 6.4.
Modeling organizational effects
We apply our GAMM (9) to size organizational effects (Figure 3). For every editorial measure the owner SD is several times the affiliate SD (: 0.72 vs 0.15; : 0.68 vs 0.14; : 0.92 vs 0.15), and co-ownership accounts for 0.258 to 0.436 of the total variance. Here, SD is the standard deviation of the owner (or affiliate) random intercepts, in units of the outcome’s SD, so owners differ in their typical level of by 0.72 SDs. Importantly, ownership influences productivity and traffic as much as editorial decisions. The owner SD is 0.61 for article counts, 0.73 for byline authors and 0.88 for web traffic. Because productivity and engagement distributions are long-tailed, article and author counts enter as and web traffic as .
Figure 3. (a) Size of the owner and affiliate effects in the GAMM, by random-effect SDs in outcome-SD units, with 95% intervals; (b) the AP wire-share slope per +1 SD of share. Editorial outcomes, productivity (articles, byline authors; log) and engagement (web traffic; asinh).
AP wire sharing separates our measure from content recycling (§ 2): a higher AP share lowers , which counts shared-agenda coverage as dependence, about three times as much as (-0.210 vs -0.063, Figure 3b). AP wire sharing is also associated with a modest increase in .
The larger influence of ownership relative to affiliation holds on the other two editorial decisions (Figure 10, Appendix § 14): the owner SD of is 1.05 on source selection and 0.69 on front-page placement, compared with affiliate SDs of 0.06 and 0.16. Owners differ as much in which sources they link to and which stories they lead with as in which events they cover. As such, the ownership effect generalizes across multiple data sources and decision types. Given the possibility that model results are impacted by irregularities in the data, we also conduct a sensitivity check where we refit every model after dropping outlet-months whose GDELT volume fell below a quarter of the outlet’s 2017 monthly mean. The estimates are essentially unchanged: the owner SD is 0.76 for (against 0.72), 0.75 for (against 0.76) and 0.59 for (against 0.68), and the affiliate SDs stay small.
Figure 10. Owner (•) and network-affiliate (▲) random-effect SDs in outcome-SD units, 95% intervals, from the central GAMM refitted on (a) source selection (one window per Common Crawl graph; 595 outlets) and (b) front-page placement (monthly Sundays; 656 outlets); analogous to Figure 3a.
Acquisitions and organizational change
We follow every outlet acquired or merged into a new owner between 2018 and 2023, each matched with at least one never-acquired outlet (within the period of observation) of the same chain tier (i.e., part of a multi-outlet group or not), same political bias group (left, center, or right), and same decile of pre-deal size (in terms of mean byline author counts over the six months before the event).
With this design we ask two questions. First, does the acquired outlet lose independence sooner than its matched controls? Here, we define failure as a sustained loss of at least two consecutive months where the treated outlet’s measure is SD below its pre-event level. We estimate Cox hazard ratios with follow-up truncated at 3, 6, 12 and 24 months post-event. Second, does the acquired outlet converge on its new owner? We track the outlet’s NMI to its new sister outlets, and the share of the pre-deal gap to the acquirer’s coverage profile that it closes, against targets of deals that failed. While similar dynamics have been identified by Rappaz et al. [31], their measures are not interpretable and rates of convergence cannot be established with this method. We further extend our matched design to large layoffs of at least 100 jobs, and add a non-editorial outcome: byline author counts.
In a time-to-event (survival) analysis, each outlet is followed from the event until its first sustained loss or the end of the follow-up period. Figure 4 plots Kaplan–Meier estimates [19] of the cumulative share of outlets that have experienced a sustained loss by each month since the event, for treated outlets and their matched controls. Because outcomes fluctuate, controls may fail too, and so evidence for organizational influence is the gap between the treatment and control curves. A solid curve above the dashed one means treated outlets lose the outcome sooner, and the vertical gap at a month is the excess share of treated outlets that have failed by then. The Cox proportional-hazards model [9] summarizes the gap as a hazard ratio, the rate at which treated outlets fail relative to the matched controls, with a ratio of 2 meaning twice the controls’ rate, and 0.5 half of it. Truncating follow-up at 3, 6, 12 and 24 months reveals change over time, as organizational influences on each of the measures may follow different time scales. Figure 5 shows these ratios on a log scale, one row per outcome and horizon. As we will see, these results reinforce the findings from the Kaplan–Meier estimates in Figure 4.
Figure 4. Survival curves for acquisitions and layoffs show the cumulative share of treatment and control groups that experience a sustained loss. Sample sizes and Cox hazard ratios at 6 and 24 months (2 and 8 graphs for Common Crawl) are given.
Figure 5. Cox hazard ratios [95%] of a sustained loss after (a) acquisitions and (b) layoffs of at least 100 jobs, with follow-up truncated at 3, 6, 12 and 24 months. A hazard ratio > 1 means that treated outlets lose the outcome sooner than controls (filled markers exclude 1).
Acquisitions. Acquired outlets lose editorial independence sooner than matched never-acquired outlets (Figure 4, Figure 5). Within six months 54% of acquired outlets have experienced a sustained loss, against 26% of matched controls. The hazard of a sustained fall in is 2.12 [1.47, 3.07] times that of the controls at six months and 1.55 [1.02, 2.34] at 24 months (212 acquired outlets, 297 controls). The loss replicates on source selection, where it is larger (5.81 [2.16, 15.64] two Common Crawl graphs, about six months, after the deal; 88 acquired outlets). Acquisitions do not appear to have a substantial impact on front-page placement, although this measure covers too few acquired outlets (28) to test for significance.
Turning to media type, while all outlets are measured according to their digital footprint, much of their content may come from their medium of origin. However, we observe that the drop in associated with acquisitions is consistent across outlets regardless of whether they are primarily broadcast or print outlets. For outlets that began as television stations, 0.74 of the pre-deal gap to their acquirer’s coverage is closed within six months; for newspapers, 0.83 of the same gap is closed within a year. We come back to the question of organizational influence by media type in § 6.4.
Furthermore, we can show that content convergence is a result of acquisitions, because coverage gaps do not close when deals fail. In these cases (11 blocked or abandoned deals), target outlets close none of the pre-deal gap to the acquirer (-0.07 after one year). Likewise, the internal pluralism of acquired outlets does not change significantly (0.61 [0.23, 1.61] at six months). In terms of productivity and engagement, we find that acquisitions do not negatively impact either. In fact, acquired outlets are less likely than their controls to lose byline authors (0.67 [0.53, 0.85] at 24 months). This may be related to pre-acquisition budget cuts leading to job losses before deals close.
Evaluating acquisition policy. To ask whether regulators stop the deals that matter editorially, we project each deal’s effect on ecosystem independence over 12 months from the convergence we measure: an approved deal is compared with a no-deal path in which its targets drift like targets of failed deals, and a blocked deal with a with-deal path in which its targets close the measured share of the gap to the would-be acquirer. Over 6 blocked and 47 cleared deals, the projected loss is larger for blocked deals (difference +0.0067, over the three independent regulatory decisions) suggesting that regulators block large deals rather than editorially worse ones. The approved deals are small in isolation, yet combined, their projected loss is about 0.05 in ecosystem independence (+2.0 month-to-month standard deviations).
Layoffs. Layoffs act on staff rather than on editorial independence. Outlets that cut at least 100 jobs lose byline authors more often than their controls (44% against 33% by six months). The lower hazard of loss replicates our source selection (0.45 [0.21, 0.95]), but not on front-page placement, where the three-month ratio goes the other way (3.09 [1.22, 7.84], 46 outlets, see Figure 9 in the appendix).

Figure 9. Cox hazard ratios [95%] of a sustained loss of and measured on (a, b) source selection and (c, d) front-page placement, after acquisitions (a, c) and layoffs of at least 100 jobs (b, d), against matched controls; follow-up truncated at 2, 4 and 8 Common Crawl graphs (source selection) or 3, 6, 12 and 24 months (placement); intervals clustered by outlet and by event; filled markers exclude 1. Analogous to Figure 5.
Here, significant author count losses within the control group are indicative of the persistence of small-scale job cuts across the media industry. These types of job cuts are difficult to track in a dataset because organizations are not required to disclose them under the federal WARN Act. As such, matched controls eventually lose staff at the same rate: outlets without a large announced layoff shrank too, only without the initial cut.
Organizational influence on broadcast TV
Our measures observe news on the Web, while television groups also coordinate on air. Comparing the same station pairs on air (closed captions) and on their websites (Figure 6 shows the two layers organized differently). Here, closed captions from broadcast TV come from the Internet Archive’s TV News Archive, spanning 2015–2016 [17, 33], and news articles scraped from news websites come from the Common Crawl News dataset (CC-News), spanning 2016–2024 [27]. Analogous to the pairwise NMI measure, we calculate pairwise linguistic redundancy between two outlets as the drop in a language model’s per-token surprisal of one station’s text when the other’s same-day text is in context.
Figure 6. Owner versus affiliation redundancy on air (broadcast captions [33]) and on the same stations’ websites (CC-News [27]). Trends on air are in opposition to those observed on the web, with affiliation, not ownership, accounting for shared broadcast coverage (as measured by LM-surprisal).
First, on the web, the surprisal measure corroborates the pairwise NMI findings for the ownership effect on digital news (as in Figure 2), where ownership adds redundancy within a network (+0.00788, ) and affiliation adds almost none. On air, however, redundancy follows the network affiliation, where co-affiliated outlets share +0.257 more surprisal reduction than random outlets and co-owned outlets share almost nothing (-0.022). The specific ownership effects we report are therefore a property of digital media. For other media types, affiliation and content syndication may play a more important role.
Longitudinal analysis
We find that platform decisions, not ownership, seem to dominate the ecosystem-level engagement and newsroom employment measures (Figure 7). Organic search traffic rises from late 2020 as outlets turn to search, then falls after Google’s US launch of AI Overviews, which reduce clicks to linked sites [7], from an estimated 2.8 billion monthly visits in May 2024 to 1.6 billion in June 2026. Summed over outlets, byline authors decline throughout, from 157,000 in March 2018 to 38,000 in September 2026. Our byline author count trendline correlates highly with monthly US newspaper employment data () compiled by the Bureau of Labor Statistics [38]. In contrast, and despite the pronounced changes that we describe at the individual outlet level, our ecosystem-level editorial measures of external pluralism () and are stable over the period. varies little (monthly SD for every decision), and the information-weighted mean of is steady throughout, with an SD of 0.05 for event selection.
Figure 7. Productivity and engagement over time: total byline authors (294 outlets, to 2026-09), Ahrefs estimates of organic web traffic (2,601 outlets, to 2026-06), each summed over a fixed outlet set, z-scored over its own window and smoothed with a centered 3-month mean. Markers: start of COVID-19 lockdowns in the US (March 2020) and Google’s US launch of AI Overviews (May 2024).
Discussion
Our results answer the four questions we set out with. First, ownership influences editorial decisions on the Web. Co-owned outlets share 20.4 times the information of the shuffled-owner null in event selection, 2.2 times in source selection and 40.2 times in front-page placement (Figure 2). Second, ownership matters far more than network affiliation: the owner SD of is 0.72 against 0.15 for affiliates (Figure 3). Only on broadcast television does affiliation dominate (Figure 6). Third, acquisitions reduce editorial independence while layoffs do not. Acquired outlets lose sooner than matched controls (Figure 4) and close most of the coverage gap to their acquirer within a year, while outlets involved in 11 failed deals do not move. Fourth, our projections suggest that regulators tend towards blocking large deals rather than editorially harmful ones: per target outlet, our projections suggest that blocked deals would not have been more determinant to independence than approved deals, although this rests on few regulatory decisions and on projections rather than observed outcomes.
Two deals in particular show that regulatory considerations can and should extend beyond size and reach. In 2018 the FCC stopped Sinclair Broadcast Group’s $3.9 billion acquisition of Tribune Media, which would have taken an owner of over 190 local stations [32] to about 70% of US households; its review did not measure editorial independence, although that April a compilation of Sinclair anchors reading the same scripted message had gone viral [28] and Sinclair’s outlets are the least independent of the 61 owners in our model (owner random effect −1.53 outcome SDs). The GateHouse–Gannett merger of 2019, a print deal outside the FCC’s remit, was cleared without any editorial review. The impact of each cleared deal like this is small (about 0.001 in ecosystem independence), but together our projections put their cost at just over 1 SD of our measure.
These findings extend a literature built on single owners and single decisions [36]. Martin and McCrain [23] found that Sinclair-acquired stations shifted their coverage toward national politics. We observe ecosystem-wide convergence on the acquirer across three different editorial decisions, and further show that outlets involved in failed deals are not impacted. Gentzkow and Shapiro [12] found that newspaper slant adapts to readers, not owners. Our measures show that when it comes to coverage decisions, ownership matters. While post-acquisition convergence in news coverage had already been demonstrated by Rappaz et al. [31] for a sample of outlets, their measures are not interpretable, so rates of convergence cannot be established. We extend the idea behind their case study to every acquisition in the panel, with rates that can be compared across owners and decisions.
Our findings have several implications. First, editorial independence measures require formal information-theoretic definitions of redundancy. When multiple Sinclair anchors read from the same script [28], the combination of editorial control with a lack of transparency enables audience manipulation. Measures that count duplicated content, such as recycling [15] or our , cannot separate these instances from benign AP news wire usage. Wire copy lowers (our definition of a content recycling measure) by about three times as much as , and so content duplication measures reads AP wire use as dependence on the owner. What’s more, after an acquisition falls while does not, so the loss lies in choices specific to the new owner, which a duplication measure misses. Separating owner-specific from shared-agenda information is what lets track independence.
Second, organizational influence depends on the medium. On air, stations’ shared content follows their affiliation. On their websites, it follows their owner (§ 6.4). While broadcast regulation counts stations and audience reach, treating network affiliation as the content content syndication mechanism, this is not the case on the web. Here, the owner is the stronger influence. A board for CBS News and CNN would therefore need to watch both the affiliate’s feed on air and the owner’s choices online.
Third, the window for such oversight is short and harms from editorial negligence accumulate. Acquired outlets the coverage gap to their new owner within a year, so remedies applied after a deal closes arrive once editorial changes have already begun. And because each cleared deal is small on its own while together they cost about 0.05 SDs of ecosystem-level independence, reviewing each deal separately hides the ecosystem effect. Editorial measures computed continuously from Web-scale data allow both the screening of individual deals before they close, and the tracking of ecosystem-level effects across multiple concurrent mergers and acquisitions.
Of course, our measures are not without limitations. First, our evidence is mostly observational and not causal. If co-owned outlets decide alike, this does not imply that owners made them do so. In addition, chains may have a preference towards acquisitions that already align with their outlets, or whose audience is perhaps more suited towards the same viewpoints. Although existing FCC regulations guard against a single company owning multiple television stations in the same locale (often called the duopoly rule), if co-owned outlets do share regions and audiences, our measure may mistake their locality for dependence. A locale term would need to be added to our decomposition to prevent this (6). In general, reading PID and PED atoms as independence and pluralism is a modeling choice (Appendix § 15), and many other valid definitions may exist that also address the same issues of separating group level redundancy from ecosystem level redundancy.
There are also several data issues that impacted our study. The filter removes events covered by fewer than three outlets in order to address overly granular GDELT event codings may also filter out articles with local stories or investigative and niche reporting. What’s more, GDELT stopped crawling many chain-owned sites after ownership changes resulted in their crawlers being blocked, and we are forced to mask these gaps. Front-page placement is particularly affected, having coverage over fewer outlets than GDELT events and the CommonCrawl webgraph offer. In addition, starting in October 2025 the front-page data pipeline no longer appears to be actively maintained by GDELT, limiting its future use.
Several extensions follow. The FCC already weighs viewpoint diversity, but through ownership counts and reach; our projections suggest screening mergers on editorial measures as well, and more research should be done to support this. Given that the information-theoretic measures generalize to groupings of any kind, the influence that of chain tier, private-equity ownership, or political alignment has over editorial decisions can be studied using this framework. Finally, given the international scope of GDELT and Common Crawl, the analysis done here may be extended to other countries. To support these efforts, we release the measures, the code and the owner-at-month panel on publication.
Conclusion
Owners influence what news sites cover, cite, and feature, and acquisitions move outlets toward their new owner within months. Through web-scale datasets and our information-theoretic framework, editorial independence can now be measured.
Acknowledgements
We thank Dr. Gregory Martin and for providing an acquisition dataset and Dr. Zachary Metzger at the Medill Institute for providing ownership data for print outlets. This work was supported in part by the Office of Naval Research grant (N000142412568) and the Knight Foundation. The views and conclusions in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Knight Foundation, Office of Naval Research, or the U.S. Government.
News ownership labels
Construction. Owner labels come from crowd-sourced lists of media-company assets on Wikipedia. For each broadcasting or publishing company with a Wikipedia category of the outlets it owns (for example, the category of Nexstar Media Group), we retrieve the member pages with the Petscan service. Each outlet’s official domain is taken from Wikidata through a SPARQL query, with the website field of the Wikipedia infobox as a fallback. Domains are resolved through HTTP redirects to their canonical form and normalized (lower case, without www. or paths). Outlets are mostly newspapers, television stations and radio stations. Network-affiliate labels (for example, ABC or NBC affiliation of a local television station) are built the same way, from Wikipedia’s network-affiliate categories; an outlet can carry several. The same procedure, applied worldwide and filtered with web-traffic estimates and site categories, produces a multi-country news-ownership dataset; this paper uses only its US news outlets, plus the affiliate labels, which that dataset lacks.
Validation on US outlets. The labels were validated against human annotation and against an independent ownership directory. On a sample of 321 US news domains annotated by hand with their owner, the Wikidata owner labels agree with the annotation for 95.8%. The Medill Local News Initiative directory of US newspapers [10, 11] provides a second check, on the 638 domains where its print titles coincide with our news websites. Our owner labels agree with Medill’s for 73.5% of these domains, which receive 91.9% of the overlap’s web traffic. Where the two disagree, Medill is usually right, because Wikipedia lags recent acquisitions (for example, the sale of Gannett titles to CherryRoad Media [5]). Counting either Medill’s current or its previous owner as a match raises agreement to 93.6% of domains and 98.8% of traffic. The labels are therefore accurate up to the timing of recent sales, which are the owner-at-month correction below addresses. These figures are for the US outlets of the multi-country collection, built like our labels.
Owner at month. The Wikipedia labels are a snapshot compiled in 2023–2025 and carry no dates, so they miss the deals made during our window (among them GateHouse–Gannett, Nexstar–Tribune, Gray–Raycom and Lee–BH Media). For every one of the 1,364 owner-labeled outlets we therefore researched the owner on 2015-01-01 and on 2026-09-30 and every change in between, with its closing date and a source. The sources are Wikipedia station and newspaper histories (1,788 of 2,318 owner spells), SEC filings and company press releases (165), and trade press, the Medill directory snapshots for 2014, 2016, 2018, 2020 and 2025, and live site checks (365). Every outlet’s starting owner was checked against at least one source, and 1,755 of the 2,318 spells are rated high confidence. The result is a panel of 2,318 owner spells over 278 dated deals. It corrects 99 snapshot labels (51 outlets never held by the labeled owner in the window, 48 sold or closed before the snapshot), and 667 outlets change owner at least once in 2015–2026. As a cross-check, 123 of the 135 in-window acquisitions recorded for these outlets in FCC ownership-transfer records coincide with a change in the panel within two months. The other 12 are explained by license versus operating-control timing, sales of non-license assets, closings before the window, month offsets, or sub-brands.
Conventions. The owner of month is the owner at the end of the first day of , so a deal switches an outlet’s owner, and its own group , from the first month that starts on or after the closing. Stations run under local marketing or shared-services agreements carry the operating group rather than the licensee. Events at the level of a parent company that do not change the operating owner (for example, the CBS–Viacom merger) leave the label unchanged. Predecessor owners keep their own names, so a merger splits an owner’s history into a pre-deal and a post-deal owner; this is why the central model has more owner levels than there are owners today (§ 5). Closed outlets have no owner after closure.
Data and decision details
Front-page placement. The Global Frontpage Graph records each homepage link’s position as a percentage of the page’s links, from 0 (first in the page) to 100 (last). For one Sunday a month (March 2018 to December 2024) we match each outlet’s homepage links to the article URLs GDELT links to events in the three days up to that Sunday, and keep each event’s best (highest) position on the page. An outlet is active if at least 50 events appear on its front page that Sunday; is the event’s rank among them and their number. The outlet measures use front-page presence, if is on ’s front page, because the PID requires a binary coverage variable. Pairwise comparisons use four levels, absent, 1 lower page (), 2 upper page, 3 lead (the top positions), with NMI computed over events on at least three front pages.
Organizational events. Layoffs: 451 announced job cuts at US news organizations, compiled by hand from public trackers of newsroom cuts and coded from the report or announcement cited for each entry (outlet, date, jobs cut, closure); we use cuts of at least 100 jobs. Acquisitions: changes of owner in the owner-at-month panel (Appendix A), which draws on the television-station ownership records of Martin et al. [22] and the Medill newspaper ownership records [11]; we use acquisitions and mergers from 2018 to mid-2023.
Measure and model details
The shared agenda
The estimand is how outlets divide attention among items that receive attention from several of them, not pluralism over all news. Two consequences are handled explicitly. First, without the filter the joint distribution over thousands of binary variables is so sparse that almost every coverage pattern is unique and plug-in entropy estimates count items rather than structure. Second, conditioning on the column sum in (4) induces positive dependence among the even when outlets choose independently; every measure is therefore reported against a null model that preserves the filter (Section C.4).
The field source is a count rather than an indicator because, under the filter (4), the event that some outside outlet covered is almost always true and would carry no information; the count records the item’s prominence outside ’s group.
Partial information decomposition
We use the two-source measure of Bertschinger et al. [3] (BROJA), for which unique information is defined by a convex optimization over distributions sharing the observed source–target marginals; with two sources the leading PID measures agree on the qualitative behavior of these atoms, and the aggregation of the own group to a single binned count in (5) is what keeps the decomposition at two sources. The four atoms together with the unexplained remainder , the coverage predictable from neither source, partition . The four explained atoms divided by form outlet ’s compositional profile under C: they sum to one, and the explained fraction is retained so that any atom can be re-expressed as a share of .
Estimation
Marginal entropies and the joint tables of are estimated by plug-in over the items in . The outlet-level PID uses the BROJA measure, computed with our own solver of its convex program and validated against the dit implementation to bits; pairwise mutual information is plug-in from tables. Plug-in dependence carries a positive bias of order , largest for outlets with small , which the null models below calibrate.
Null models and reporting
Two null models calibrate the measures. A degree-preserving randomization of the outlet-by-item incidence on (Curveball [34]; 200 draws for the pairwise measures, 30 for the PID measures) keeps each outlet’s coverage volume and each item’s coverage count, so the filter (4) and the dependence it induces are present in the null. A permutation of the group labels with the incidence fixed turns a within-group statistic into evidence about the grouping; the pairwise comparisons of Figure 2 use it. The models and hazard analyses use the raw series.
Central model
In (9), indicates that the outlet-month published at least one authored article and is the share of those articles carrying an Associated Press byline (read as zero when no wire-share record exists), so that is the wire-dependence slope among months with authored output; the information-theoretic outcomes carry no political-bias term, since the bias label enters their construction; is a thin-plate smooth of calendar time with basis dimension 12; is a random intercept for the parent owner , the owner at month (Appendix § 10); is a random intercept for the network affiliate, present only for affiliated outlets; and is a per-outlet factor smooth of time with basis dimension 5, which absorbs each outlet’s own level and trend so that the ownership terms are identified from what co-owned outlets share beyond their individual histories. Counts enter as , engagement as , and shares, distances and entropies on the identity scale. Outlet-months in GDELT coverage gaps are dropped.
Serial dependence. Outlet-months are not independent draws. After the full fit we compute each outlet’s lag-one residual autocorrelation; when the median across outlets exceeds 0.30 the model is refitted with a first-order autoregressive error at the mean residual autocorrelation, and every comparison below is made between models with the same error structure.
Variance share. The variance share of a term is the variance of its fitted contribution as a share of the variance of the full fitted predictor plus the residual variance; it is an effect size, not evidence. Slopes are reported in standard deviations of the outcome.
Shared information with groups of outlets
Pairs of unrelated outlets share little event selection (Figure 2a). To test whether groups share more, we hold out an outlet , draw other outlets, and compute the NMI between and the binned count of those outlets covering each event, built like the field source;
is the pairwise NMI of Eq. 8. Shared information rises steeply with group size: from a median of 0.0008 for one random outlet to 0.011 for seven and 0.11 for all other outlets (Figure 8). This rise reflects event popularity, not coordination. On a degree-preserving randomization of the outlet–event matrix (Curveball), which keeps how many events each outlet covers and how many outlets cover each event, the curve is higher at every (0.009 for pairs, 0.14 for the field): unrelated outlets are more independent than their coverage totals imply, as expected if local outlets cover different markets. Groups of co-owned outlets sit four to seven times above that baseline for to 32; eight co-owned outlets explain about 21% of an outlet’s coverage entropy, eight random outlets about 1%.
Acquisitions and layoffs on source selection and front-page placement
Figure 9 repeats the hazard analysis of Figure 5 with and measured on the other two editorial decisions: the design, matching and failure rule are unchanged, and intervals are clustered by outlet and by event. On source selection, time is counted in Common Crawl graphs (two, four and eight graphs, about 6, 12 and 24 months). Acquired outlets lose editorial independence in their linking sooner than their controls (5.81 [2.16, 15.64] at two graphs, 3.56 [1.67, 7.58] at eight), while their internal pluralism does not change significantly; outlets with large layoffs are less likely to lose internal pluralism (0.45 [0.21, 0.95] at four graphs). On front-page placement the samples are small (28 acquired outlets for , 12 for
and every interval includes 1, except a higher hazard of losing three months after a large layoff (3.09 [1.22, 7.84]).
Mutual-information forms of the measures
Each measure in Section 3 has a form written in conditional entropies, mutual information, co-information or total correlation rather than decomposition atoms. This appendix gives those forms, the exact identities that relate them to the atoms where such identities exist, and the differences where none does. The MI forms are computed on the same source variables and the same items as
the decomposition forms, so the two can be compared on identical inputs. Throughout, denotes the pairwise mutual information between outlets and outlet ’s share of total pairwise coupling.
Identities
Two identities link the atoms of (6) to standard quantities:
where CoI is co-information. Identity (11) states the sense in which the MI forms are less expressive than the decomposition forms: they report redundancy net of synergy.
Family 1
The MI form of internal pluralism, on the same source variable so that the relation is exact, is
by (10); likewise for .
References
- [1]Ahrefs. See tools & resources to grow your traffic. 2023. URL https://ahrefs.com/.
- [2]A. M. N. Archer and J. D. Clinton. Changing owners, changing content. Political Communication, 35(3):353–370, 2018. doi: 10.1080/10584609.2017.1375581.DOI
- [3]Nils Bertschinger, Johannes Rauh, Eckehard Olbrich, Jürgen Jost, and Nihat Ay. Quantifying unique information. Entropy, 16(4):2161–2183, 2014.DOI
- [4]K. Bleyer-Simon, E. Brogi, et al. Monitoring media pluralism in the digital era: MPM2023. Technical report, Centre for Media Pluralism and Media Freedom, European University Institute, 2023.
- [5]Mark Caro. Paving the Cherryroad. Local News Initiative, Northwestern University. https://localnewsinitiative.northwestern.edu/posts/2025/05/12/jeremy-gulban-cherryroad-q-and-a/index.html, May 2025.
- [6]Peter Carragher, Evan M Williams, and Kathleen M Carley. Detection and discovery of misinformation sources using attributed webpages. In Proceedings of the International AAAI Conference on Web and Social Media, volume 18, pages 214–226, 2024. URL https://doi.org/10.1609/icwsm.v18i1.31309.
- [7]Athena Chapekis and Anna Lieb. Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/, July 2025.
- [8]Common Crawl Foundation. Common crawl host- and domain-level web graphs. https://commoncrawl.org/web-graphs, 2024.
- [9]David R. Cox. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2):187–220, 1972.
- [10]Tim Franklin. The state of local news 2024: Expanding deserts, shifts in ownership, and expanded digital coverage. Online report, Northwestern University, 2024. URL https://localnewsinitiative.northwestern.edu/projects/state-of-local-news/2024/. Accessed January 2026.
- [11]Tim Franklin. The state of local news 2025. Online report, Northwestern University, 2025. URL https://localnewsinitiative.northwestern.edu/projects/state-of-local-news/2025/. Accessed January 2026.
- [12]Matthew Gentzkow and Jesse M. Shapiro. What drives media slant? Econometrica, 78(1):35–71, 2010. doi: 10.3982/ECTA7195.
- [13]Daniel C. Hallin and Paolo Mancini. Comparing Media Systems. Cambridge University Press, 2004. doi: 10.1017/CBO9780511790867.DOI
- [14]Chris Harnetty. Media outlets and their moguls. European journal of Communication, 29(3):335–350, 2014. doi: 10.1177/0267323114523150.
- [15]J. Hendrickx and H. Ranaivoson. Why and how higher media concentration equals lower news diversity. Journalism, 22(11):2800–2815, 2021. doi: 10.1177/1464884919894138.
- [16]Robin A. A. Ince. The partial entropy decomposition: Decomposing multivariate entropy and mutual information via pointwise common surprisal. arXiv preprint arXiv:1702.01591, 2017.arxiv.org/abs/1702.01591
- [17]Internet Archive. TV News Archive. https://archive.org/details/tvarchive, 2026. Accessed October 2026.
- [18]Lou Jost. Partitioning diversity into independent alpha and beta components. Ecology, 88(10):2427–2439, 2007.pubmed.ncbi.nlm.nih.gov/18027744
- [19]Edward L. Kaplan and Paul Meier. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282):457–481, 1958.
- [20]Kalev Leetaru and Philip A. Schrodt. GDELT: Global data on events, location, and tone, 1979–2012. In ISA Annual Convention, volume 2, pages 1–49, 2013.
- [21]Felicia Loecherbach, Judith Moeller, Damian Trilling, and Wouter van Atteveldt. The unified framework of media diversity. Digital Journalism, 8(5):605–642, 2020. doi: 10.1080/21670811.2020.1764374.
- [22]Gregory Martin, Nicola Mastorocco, Joshua McCrain, and Arianna Ornaghi. Media consolidation. Kilts Center at Chicago Booth Marketing Data Center Paper, 2026.
- [23]Gregory J. Martin and Joshua McCrain. Local news and national politics. American Political Science Review, 113(3):372–384, 2019. doi: 10.1017/S0003055418000965.DOI
- [24]A. Masini et al. Measuring and explaining the diversity of voices and viewpoints in the news. Journalism Studies, 19(15):2324–2343, 2018. doi: 10.1080/1461670X.2017.1343650.
- [25]Nicola Mastrocoracco and Arianna Ornaghi. Who watches the watchmen? American Economic Journal: Economic Policy, 17(2):285–318, 2025. doi: 10.1257/pol.20230356.
- [26]Denis McQuail. Media performance: Mass communication and the public interest, 1993.
- [27]Sebastian Nagel. News dataset available. Common Crawl blog. https://commoncrawl.org/blog/news-dataset-available, October 2016. CC-NEWS, s3://commoncrawl/crawl-data/CC-NEWS/.
- [28]Erin Nyren. Sinclair broadcast group faces backlash over scripted promos: ‘this is extremely dangerous to our democracy,’ 2018. URL https://variety.com/2018/tv/news/sinclair-promos-backlash-1202741019/.
- [29]Victoria Patricia Aires, Fabiola G. Nakamura, and Eduardo F. Nakamura. A link-based approach to detect media bias in news websites. In Companion Proceedings of The 2019 World Wide Web Conference, pages 742–745, 2019.DOI
- [30]M. Popescu, G. Tóka, T. Gosselin, and J. Santana Pereira. European media systems survey 2010. Technical report, Central European University, 2011.
- [31]Jérémie Rappaz, Dylan Bourgeois, and Karl Aberer. A Dynamic Embedding Model of the Media Landscape. In The World Wide Web Conference, pages 1544–1554, San Francisco CA USA, May 2019. ACM. ISBN 978-1-4503-6674-8. doi: 10.1145/3308558.3313526. URL https://dl.acm.org/doi/10.1145/3308558.3313526.DOI
- [32]Sinclair. Sinclair broadcast group, 2023. URL https://sbg.net/news/.DOI
- [33]Gaurav Sood and Suriyan Laohaprapanon. Closed caption news transcripts from the Internet Archive (2014–2023), 2022.
- [34]Giovanni Strona, Domenico Nappo, Francesco Boccacci, Simone Fattorini, and Jesus San-Miguel-Ayanz. A fast and unbiased procedure to randomize ecological binary matrices with fixed row and column totals. Nature Communications, 5:4114, 2014.DOI
- [35]The GDELT Project. Announcing the GDELT global frontpage graph (GFG). GDELT blog. https://blog.gdeltproject.org/announcing-gdelt-global-frontpage-graph-gfg/, March 2018.
- [36]Hendrik Theine, Julia Bartsch, and Mandy Troeger. Does media ownership matter for journalistic content? a systematic scoping review of empirical studies. International Journal of Communication, 19:865–887, 2025.
- [37]Charlotte Tobitt. More than 3,000 journalism job cuts tracked in UK and US in 2025. Press Gazette. https://pressgazette.co.uk/publishers/journalism-job-cuts-2025-tracked/, 2026. Published January 13, 2026; updated April 28, 2026.
- [38]U.S. Bureau of Labor Statistics. All employees, newspaper publishers (NAICS 511310), not seasonally adjusted (series CEU5051311001), current employment statistics. https://data.bls.gov/timeseries/CEU5051311001, 2026. Accessed 2026-10-25.
- [39]Dave Van Zandt. Media bias/fact check: A comprehensive media bias resource. https://mediabiasfactcheck.com, 2024.
- [40]Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. arXiv preprint arXiv:1004.2515, 2010.arxiv.org/abs/1004.2515
- [41]Simon N. Wood. Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC, 2 edition, 2017.







