Contribution
It documents two-way favors between a government and connected media in a democracy, with a control that the earlier evidence lacked: private firms’ advertising in the same outlets at the same time. It then separates two mechanisms that produce the same observed slant, an owner who agrees with the government and an owner who is paid by it. It does so under explicit structural assumptions, and shows that targeted advertising can also change who owns the media.
Key insight
The same pro-government coverage can come from ideology or from exchange, and the two are told apart only by a change that alters one and not the other. Hold the owner fixed, so ideology is fixed, and change the reward for slant: by a fallout that ends the connection, or by a rise in the government’s advertising budget. Any change in relative slant is then the exchange motive. Conversely, ideology is not exogenous either. Advertising favors make outlets worth more to investors who already lean toward the government.
Assumptions
- Classification of connections from Mertek Media Monitor reports, checked against ownership records. The analysis uses connections to the right; left connections under left governments are not settled.
- Slant is measured as minus the share of an outlet’s articles covering corruption scandals, from a hand-built keyword list of scandals with a hand-checked subset. It is unbounded in the model and bounded in the data.
- Model (Section 4.1). Two outlets choose slant s̃ᵢ to maximize β·E[Σⱼ pᵢqᵢⱼ] − (NT/2)(s̃ᵢ − ŝᵢ)², where ŝᵢ is the owner’s ideal point. Readers choose one outlet by logit with slant weight γ. Advertisers have Cobb–Douglas weights proportional to readership and, for state-owned advertisers, exp[θᵢ + δᵢ(sᵢ − s_r)], with δᵢ > 0 only for a connected outlet. Budgets Wⱼ are exogenous and prices predetermined. Measured slant is sᵢ = s̃ᵢ + ζ + ξᵢ, with ζ an unobserved neutral level and ξᵢ noise, and a reference outlet r (hvg.hu, a critical outlet) nets out ζ.
- Identification. An owner’s ideal point is constant across periods: Telekom’s around Origo’s change of editor, and Simicska’s around the 2015 fallout, which the authors call “arguably strong”. The government’s online advertising budget is not endogenous to the outlets’ slant.
- Episodes are separate one-shot equilibria. Future changes are unanticipated.
Key results
- Advertising favors (Section 3.1; Tables A.II–A.IV).
- Changes in government: the right-connected daily’s share of state-owned firms’ advertising (against the left daily) rose from 16% to 72% in 1998–2002 and from 47% to 95% after 2010, while private firms’ share tracked circulation. In regressions, state-owned firms shifted 26 percentage points (unweighted) to 34 (weighted, advertiser fixed effects) toward the right daily under right governments.
- Metropol’s purchase by Simicska’s partner in 2011: its share of state-owned print advertising went from below 20% to above 50% at once, with private advertising unchanged (weighted event estimates about +0.28 to +0.32).
- The Orbán–Simicska fallout of February 2015: the share of state-owned advertising in Simicska’s two dailies fell from above 60% to below 20%, starting months before the fallout was public.
- Origo: state-owned firms’ share rose by 18–22 percentage points after the change of editor in 2014, and by 57–61 after the purchase by a connected owner in 2016.
- Prices from annual reports: state-owned firms paid connected media about three times what private firms paid.
- Content favors (Section 3.2; Tables A.V–A.VI). After the fallout, Magyar Nemzet’s share of articles on corruption rose by 1.27 percentage points (s.e. 0.23), from the level of an unaffected right daily toward that of the left daily. Relative to Index, Origo’s share fell by 0.44 points after the new editor (s.e. 0.26 with period effects, 0.105 with month effects) and by 1.36 after the new owner (s.e. 0.30 or 0.21).
- Against audience explanations. Private advertising did not move. The online audiences’ demographics and the political views of 2013 readers’ demographic profiles evolved in parallel. In the fallout, advertising moved before content, the reverse of the order an audience story predicts.
- Proposition 1 (demand for slant). log(a₁ⱼ/a₂ⱼ) = α + γ(s₁ − s₂) + θGⱼ + δ(s₁ − s_r)Gⱼ + νⱼ.
- Proposition 2 (supply of slant) and (8). With masⱼ = E[a₁ⱼ(1 − a₁ⱼ)], s₁ − s₂ = (ŝ₁ − ŝ₂) + (βδ/NT)·Σ over government advertisers of masⱼWⱼ + ξ. Readers’ preference cancels. Both propositions are re-derived in THEORY-860.
- Table IV. δ is 421 online, 513 in print and 441 pooled. γ is −53, −149 and −69, so the marginal reader dislikes slant. β is 1.38, 1.63 and 1.50, all with bootstrapped intervals excluding zero. Origo’s new owner’s ideal point is 1.3 points more pro-government than Index’s; Telekom’s was 0.3 less. Letting Simicska’s ideal point change at the fallout gives a small, insignificant change.
- Table V, decomposition for Origo. Against no targeted advertising and Telekom ownership throughout: favor exchange cut corruption coverage by 29% (period 1), 44% (period 2) and 42% (period 3). The new owner’s ideology alone cut it by 63%, and both together by 74%, so the two are substitutes. A ban on targeted advertising (δ = 0) would have raised coverage by 41%, 79% and 42%.
- Table VI, endogenous ownership. With favors, the pro-government owner’s net-profit advantage rose from −7% to 0% of revenue. Without favors it stays between −20% and −17%.
- Tables VII–VIII. A 60% fall in private advertising, or a doubling of government advertising, cuts Origo’s corruption coverage by 67% or 75% and raises the return to a connected takeover. Capturing Index as well can raise or lower Origo’s slant, depending on how close the two are in the government’s market.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Hungarian governments of the right steered state-owned firms’ advertising to connected outlets, beyond what audiences explain | strong | five events, private advertisers as control, audience data online, price evidence |
| C2 | Connected outlets reduced corruption coverage when connected and raised it when the connection ended | strong for online, moderate for print | Tables A.V–A.VI; timing in the fallout |
| C3 | In the model, the slant gap between connected and unconnected outlets is ideal-point gap plus a favor-exchange term, with readers’ preference cancelling | strong (derivation) | Propositions 1–2; re-derived in THEORY-860 |
| C4 | Both owner ideology and favor exchange drove Origo’s coverage, with favor exchange alone cutting it by 29–44% | moderate | Table V, conditional on the identifying assumptions |
| C5 | Targeted advertising can make media ownership more attractive to pro-government investors | moderate in the model; suggestive in fact | Table VI; consistent with Origo’s sale |
| C6 | Government advertising affects media content worldwide | weak | extrapolated from Hungary and the qualitative reviews of Appendix A.1 |
Concepts
- Favor: “an allocation not driven by a conventional market force like audience preferences”.
- Owner ideology ŝᵢ: the owner’s ideal slant, which costs her quadratically to depart from.
- Favor exchange: content favors given in return for advertising favors, the βδ term. Shared ideology: content favors given because the owner agrees, the ŝ₁ − ŝ₂ term.
- Connected outlet: one with δᵢ > 0, whose slant the government’s advertisers reward, standing in for Besley and Prat’s low τ.
- Marginal advertising share masⱼ = E[a₁ⱼ(1 − a₁ⱼ)]: how responsive advertiser j’s split is to the outlet’s slant.
Connections
It builds on Besley and Prat (LIT-1333) for the government’s demand for slant and on Gentzkow and Shapiro (2010) for the structural approach to slant. It advances on Di Tella and Franceschelli (LIT-1319) by ruling out the market-based explanations they left open. It relates to DellaVigna et al. (2015) on favors from businesses to connected media, and to Petrova (2011) on private advertising and press independence.
Bearing on the record
- CLAIM-407. A25’s point, “institutions may produce similar observable outputs through different causal organizations”, has a documented instance here, in the shape of the claim’s hidden governance bit. In equation (8) the same slant gap s₁ − s₂ is produced by different splits between owner ideology and favor exchange, so a description that keeps only the outlets’ output is the same under both. Under an intervention on the mechanism the two come apart: banning targeted advertising (δ = 0) removes the exchange term and leaves the ideology term, so coverage at Origo would rise by 41–79% under the estimated split and not at all under pure ideology (Table V). The paper tells the mechanisms apart only from episodes in which the reward changed with the owner fixed, that is, from variation in the very term the coarse description drops. The paper’s numbers are a model counterfactual, not an observed intervention.
- CLAIM-404. A26 §4 used the paper for how control over evidence and control over the evaluative objective reinforce one another. The paper supports a narrower, documented form. A government’s discretion over an evaluator’s resources (advertising) changed the evaluator’s output, and it also changed who owns the evaluator, and so the owner’s ideal point, which is the evaluator’s own objective. That second channel, Table VI’s endogenous ownership, is the closest thing in the record’s readings to influence on the resources becoming influence on the criteria. It is a model counterfactual, not an observation.
- CASE-123. A second documented variant beside Di Tella and Franceschelli, stronger in design: the press variant with the connection switched on and off. The fallout is the case’s one-time perturbation in reverse. When the reward stopped, the content favors stopped too, after a lag. Coverage did not stay where the captured period had put it, which is evidence against self-sustaining capture in this instance.
- TERM-066. The paper’s “media capture” covers both mechanisms, and its decomposition shows why the term’s “driven by the party being evaluated more than by evidence independent of it” needs the outside standard. The paper’s standard is a critical reference outlet (hvg.hu, or 444.hu for robustness), which nets out the unobserved neutral level of coverage.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Identification of β in print rests on Simicska’s ideology being unchanged by the fallout. The authors call this “arguably strong”. The online identification rests on the exogeneity of the government’s budget, supported by parallel trends across media.
- Unobserved favors. Only advertising is measured as a reward, and only corruption coverage as slant. The authors show that the results survive special cases (a linear rescaling, proportional rewards) and say the general case is open.
- Episodes are one-shot equilibria. No dynamics of reputation or anticipation. The ban counterfactual assumes no other channel and a permanent policy.
- Measurement. Corruption coverage is a keyword count. Slant is bounded in the data, and some counterfactuals push it past the bound, where the authors read negative coverage as “praise”.
- Left-wing favoritism is neither shown nor ruled out.
- The proposition text has a typo. It defines ξ₂r = ξ₁ − ξ_r, which must be ξ₂ − ξ_r.
Open questions
- Is favor exchange self-reinforcing over time, for example through ownership changes that raise slant, which raises favors, which attracts further connected owners? The model is static per episode and cannot say.
- Do the results hold where government advertising is a small share of revenue? The authors suggest that is why Gentzkow and Shapiro found no political effect in the US.