Shark attacks and the 1916 election

In July 1916 four people were killed by sharks off the New Jersey coast in twelve days. Sixty years later the episode gave Peter Benchley his plot. In between, it became the single most repeated example of a claim that voters are not really evaluating anything: they punished Woodrow Wilson at the ballot box for the sharks. It is a wonderful story, and as far as the data can tell, it did not happen.

New Jersey, summer 1916

Four fatal shark attacks along the Jersey Shore in twelve days, in the middle of a presidential election year.

Wilson’s 1916 New Jersey share and the reported shortfall in the affected communities, from Achen and Bartels.

Twelve days in July

In 1916 four people were killed by sharks along the New Jersey coast. The attacks emptied the beaches, wrecked the resort season, and eventually gave Hollywood its plot.

And then, allegedly, the vote

Achen and Bartels used the episode to argue that voters punish incumbents for things no government controls. Wilson, they report, ran about three points behind his statewide numbers in the towns that took the hit.

Look at what the evidence is

One state. One election. A few towns, in an era of thin and irregular county records. There is no replication available, because there is only ever one 1916.

Count the choices instead

Fowler and Hall pointed out that the analyst picks which towns are affected, which prior election is the baseline, and what to control for, all after knowing what happened. Here is that tree.

Defensible choices, different answers

Redefine "hit" as all shore counties and the estimate moves toward zero, or past it. Fowler and Hall also found no relationship in coastal counties in other elections, and no effect where you would most expect one.

The dot is not the finding

When a claim rests on a single dot from a cloud this wide, the honest description is that the data cannot settle the question. Read the article for what survives.

The claim

Christopher Achen and Larry Bartels, in a 2002 working paper that became a chapter of their 2016 book Democracy for Realists, argue that voters routinely punish incumbents for events outside anyone’s control: droughts, floods, the local football team, and sharks. The shark case is the showpiece. They report that in the New Jersey beach communities affected by the 1916 attacks, Wilson ran roughly three percentage points behind his statewide performance.

The wider thesis is serious and does not depend on sharks. It is that retrospective voting, the mechanism most defences of democracy rest on, is much cruder than political science assumed. Voters feel that things have gone badly and punish whoever is in office, whether or not the office had anything to do with it.

fatal attacks in 12 days

New Jersey coast, July 1916

~3 pts

claimed shortfall for Wilson in affected towns

election, in one state

no replication is possible

How they found it

County and township returns from the 1916 presidential election, compared against Wilson’s statewide share and against his own 1912 result. The affected towns are the ones near the attacks, principally in Ocean County. The finding is the gap between what Wilson got there and what the comparison implies he should have got.

That is the entire evidentiary base. There is no second 1916, no other coast with a comparable event in a presidential year, and no randomisation of anything. Everything therefore rides on the analyst’s choices: which towns count as affected, which prior election is the baseline, what else to adjust for.

The mechanism

The proposed mechanism is blind retrospection. Beach economies collapsed for the season, people were frightened and out of pocket, and an incumbent president was on the ballot. Voters do not carefully assign responsibility; they register that the year was bad and vote accordingly.

There is a stronger version of this literature that does not rest on sharks, covering droughts and floods, and a well-known related claim about college football results and incumbent vote share. Some of those have their own critiques. The shark case is simply the most quotable and the thinnest.

How much to believe it

The re-analysis is unusually thorough. Anthony Fowler and Andrew Hall (2018, Journal of Politics, “Do Shark Attacks Influence Presidential Elections?”) went back to the returns. They report that the result depends on which coastal places are coded as affected, that it does not appear in coastal counties in other elections where attacks occurred, and that comparable specifications produce estimates on both sides of zero. Their conclusion is that the correlation is a product of analytic choices in a very small sample.

The tree in the scene above is the actual argument. With three binary-ish decisions you already have dozens of defensible analyses. None of them is fraudulent. The researcher does not have to go looking for the significant one, because the analysis that matches the story you already believe simply feels like the correct one. This is Gelman and Loken’s garden of forking paths, and it is the dominant failure mode in observational social science.

The base rate problem. Shark attacks are rare, so any study of them has almost no observations, and rare events attract attention precisely when they coincide with something interesting. The 1916 attacks are famous partly because they happened in an election year.

Achen and Bartels did not fold. They have responded that Fowler and Hall’s coding choices are themselves contestable and that the broader retrospective-voting evidence stands. On the narrow question the fairest reading is that the shark result is not robust, and that the general thesis has to be argued from the drought and flood work instead.

What it is good for. The episode is now more valuable as a teaching case about forking paths than as a fact about voters, which is a decent second career for a finding.

What it means for you

The number of analyses you could have run is part of the result, whether or not you ran them. If a dashboard number only becomes interesting after you picked this cohort, this window and this exclusion rule, it is not a finding yet; it is one draw from a distribution you never looked at.

The practical fix is cheap. Before you cut the data, write down the cut. Then run the four or five other cuts you would have accepted as equally reasonable and put all of them on the same axis. If the answer is stable across them, you have something. If it moves around zero, you have learned that too, and much more cheaply than by shipping it.

Suggested chart

A specification curve. Run every reasonable combination of affected-area definition, baseline election and control set, sort the resulting estimates, and plot them as a single line with a confidence band, with the published specification marked. Underneath, a matrix showing which choices produced which estimate. The 1916 returns are public, so this is genuinely reproducible rather than schematic.

Sources

Achen, C. H. and Bartels, L. M., 2016, Democracy for Realists: Why Elections Do Not Produce Responsive Government, Princeton University Press. press.princeton.edu.
Achen, C. H. and Bartels, L. M., 2002, “Blind Retrospection: Electoral Responses to Drought, Flu, and Shark Attacks,” working paper, later Estudio/Working Paper 199, Juan March Institute. march.es.
Fowler, A. and Hall, A. B., 2018, “Do Shark Attacks Influence Presidential Elections? Reassessing a Prominent Finding on Voter Competence,” Journal of Politics, 80(4), 1423-1437. journals.uchicago.edu.
Gelman, A. and Loken, E., 2013, “The garden of forking paths,” working paper, Columbia University. stat.columbia.edu.