The War of Two Banners

Chapters of the quest

Two banners, one question

In the autumns of 2000 and 2004, two great houses fought for the realm's highest seat. In every town they courted they hung two kinds of banners: bright banners that praised their own champion, and black banners that attacked the rival.

Brother Knowledge

Pundits swear the black banners win. Others say they make people stay home. Decades of studies can't agree. Our task: find out what each kind of banner really does, both to whether people vote and to whom they vote for.

In the paper's terms: positive (promote) versus negative (attack or contrast) TV ads, every ad coded by tone, and their effects on turnout and candidate choice.

Two choices, not one

A voter makes two decisions: whether to vote at all, and whom to vote for.

The standard model gives each banner a single lever. Pull it and the champion looks better, which raises his share and draws more people to the polls, always in the same direction. Morgana insists that is the only way it can work.

But many suspect black banners win share while keeping people home. A one-lever model can't even express that.

Two levers. Each tone now gets its own effect on a champion's appeal and its own effect on the urge to stay home, so share and turnout can move in different directions.

Both banners fly in the same towns

Battleground towns get flooded with both kinds of banner, and safe towns get few of either. Across markets, bright and black banners moved together with a correlation of 0.88.

When two things always rise together, the raw numbers can't say which one did the work.

Table 1, thousands of gross rating points per market from September 1 to Election Day. Every candidate aired more black banners than bright ones, and the incumbent party aired fewer black banners than the challenger.

The campaigns knew things we don't

Campaigns don't hang banners at random. They study each town: its mood, its polls, its enthusiasm. Then they aim their banners accordingly.

That hidden knowledge moves both the banners and the votes. Compare towns with many banners to towns with few, and you partly measure the towns, not the banners.

Brother Knowledge

To beat Lurk we need forces that change how many banners a town gets, but have no direct say in how its people vote. Hold that thought for two pages.

Blurry counts shrink real effects

How many banners did a household actually see? Nobody counts that directly. It's estimated from what campaigns spent and what airtime cost, and each ad's tone is coded by hand.

Blurry measures drag every estimate toward zero, so real effects look small or vanish. Grumbo then announces there was nothing there.

This helps explain decades of mixed findings: some studies found black banners cut turnout, some found they raised it, and recent work found no turnout effect at all.

The same tool that beats Lurk also beats Grumbo: a force that moves the true number of banners is unaffected by the noise in how we measured them.

Three signals that move banners, not votes

1. Airtime prices. What TV time cost in each market. Cheap airtime means more banners, whatever the town thinks.
2. The neighbors. Campaigns buy a whole TV market at once, so counties next door shape how many banners your county gets.
3. The challenger's ledger. A challenger can attack the incumbent's record. As local unemployment rises, challengers go black more often than incumbents do. Interacting every signal with who holds the seat is what pulls the two tones apart.

Lurk and Grumbo fall. Because these signals move banners without moving votes, they cut through both the campaigns' hidden knowledge and the blurry counts.

Keep the strong few

Mixing and matching the signals produced 1,254 candidate instruments. Use them all and they start fitting noise, and the answer drifts back toward the biased raw comparison.

58 kept. The LASSO keeps only the signals that genuinely predict banner counts. With all the base signals instead, the black-banner estimate swings to −3.08 with a standard error of 18: no answer at all.

Section 4.2 and Table 5. Post-LASSO instrument selection following Belloni et al. (2012).

Black banners persuade, bright banners mobilize

What 1% more of each banner does, once Lurk and Grumbo are beaten:

The trade-off. Black banners win more share, mostly by making the rival less appealing, and may nudge turnout down slightly. Bright banners bring more people to the polls, but some of those new voters choose the rival.

Table 3 elasticities. The black-banner turnout effect (−0.007) could be zero.

Four shortcuts, four wrong answers

Each shortcut skips one piece of the full method. Pick one to see what it gets wrong.

Table 2. Bars show each method's estimated effect of bright and black banners on a champion's own appeal. Faded bars could be zero.

Same budgets, different banners

Keep every campaign's budget exactly as it was, but change the color of its banners. What happens in 2000?

Across both elections, switching from all-black to all-bright banners raises national turnout by about 3.3 percentage points, roughly 10 million voters. All-bright is what a ban on negative ads would look like; some countries outside the U.S. have such bans.

Section 6.1, Figures 4 and 5. In 2004 the margins were wide enough that no state flips.

Aim your banners, and so will they

The smart rule: hang bright banners where you're strong, to bring out your own base, and black banners in your rival's strongholds, to keep theirs home. In Florida in 2000 that meant the north for the Republican and the southeast for the Democrat.

Choose who re-aims their banners in 2000:

Section 6, best-response and competitive-equilibrium scenarios with budgets held fixed.

Relics claimed

Open access under CC BY 4.0. Characters and art from Dungeons & Datasets. The villains are fictional; the findings are not.