The Keys to the White House, developed by Allan Lichtman, is a system for predicting the electoral college winner of American presidential elections.
The model relies on the theory of retrospective voting. That is, the model assumes that the American electorate chooses a president not according to events of the campaign, but according to how well the party in control of the White House has governed the country. If the voters are content with the party in power, it gains four more years in the White House; if not, the challenging party prevails.
Thus, the model implies that the choice of a president does not turn on debates, advertising, speeches, endorsements, rallies, platforms, promises, or campaign tactics. Rather, presidential elections are primarily referenda on the performance of the party holding the White House.
Lichtman first developed the Keys system in 1981, in collaboration with Vladimir Keilis-Borok. The methodology used in the development of the Keys is described in Keilis-Borok and Lichtman (1981). As shown in the table below, each of the thirteen keys is stated as a threshold condition that always favors the re-election of the party holding the White House.
Each key can then be assessed as true or false prior to an upcoming election. The following table shows Lichtman’s coding for making forecasts for the ten elections from 1984 to 2020.
The model then uses a simple decision rule to predict the election winner:
When five or fewer keys are false, the incumbent party wins; when any six or more are false, the challenging party wins.
The Keys model is often cited as having a perfect record in correctly predicting each U.S. presidential election since 1984. This is not correct. In 2000, Lichtman coded five keys as false for the incumbent Democratic party. Given today’s implementation of the model, this would have implied a forecast that Democratic candidate Al Gore should have won the electoral college, which he didn’t (Gore did win the popular vote, however). Hence, the model’s track record in picking the electoral college winner since 1984 is a 9 out of 10. Yet, the value of counting correct predictions for binary (yes/no) forecasts is questionable.
The trouble with such an approach, or any other producing binary predictions, is that landslides such as 1964, 1972, and 1984 are easy to predict, and so supply almost no information relevant to training a model. Tie elections such as 1960, 1968, and 2000 are so close that a model should get no more credit for predicting the winner than it would for predicting a coin flip.Gelman, Hellmann & Wlezien (2020, p. 866)
A better way to judge the quality of the Keys model is thus to translate the binary forecasts into forecasts of the popular vote, as suggested by Armstrong & Cuzán (2006). In order to translate Lichtman’s coding into a forecast of the incumbent’s popular two-party vote (V), Armstrong & Cuzán (2006) used the number of Keys favoring the incumbent (i.e., keys coded as
True) as the single predictor in a simple linear regression model estimated based on historical data back to 1860. The table below shows the resulting vote share forecasts for each of the ten elections from 1984 to 2020.
The model did quite well, with an average error of roughly two points. This is particularly notable given that the forecasts were made with very long lead time, much longer than most other available models, as can be seen in the following table.