The Forecasting Model

These visual patterns are put to the test in Table 1, which displays the results of a regression analysis in which the dependent variable is the percent of the two-party vote going to the incumbent presidential party and the independent variable is interacted with a dichotomous variable scored 0 for incumbent contests and 1 for open-seat contests. Lamentably, with only 14 observations, collinearity is a major concern and it is not possible to enter approval, personal finances, and their separate interactions with incumbency in the same model.Footnote ² Instead, I present in Table 1 the results of the national conditions model, which combines both approval and personal finances into a single variable, as described above, and requires only one interaction term. If there is really less accountability during open-seat contests, the interaction slope should be negative and significant; and, indeed, the results in Table 1 do indicate less accountability in years in which no incumbent is running. The impact of personal finances and presidential approval is much greater in incumbent races than in open-seat contests. And despite the high level of collinearity (tolerance = .026), the interaction term is significant at the 0.05 level (one-tailed).

Table 1.

Incumbency-based Forecasting Model for Presidential Elections, 1952–2004

One important question is whether these findings are robust, or if they are due to just one or two influential data points. One way of assessing this is to focus on the out-of-sample predictions, derived for each year by dropping that observation and predicting the outcome based on data from the other 13 years. Doing so for this model generates an average absolute out-of-sample error of 2.68 percentage points. Another way of evaluating the expected accuracy of the model is by looking at pseudo a priori predictions from the last several elections based on estimates that would have been generated at the time from previous elections. So, for instance, the predicted outcome for 1992 is based on slopes generated from data from 1952 to 1988. The accuracy of these “predictions” is displayed in Table 2, where we see that the model estimates are, on average, within just less than 2 percentage points of the actual outcome.Footnote ³

Table 2.

Pseudo Predictions from 1992 to 2004

Note: Cell entries are actual and predicted percent of the two-party vote for the incumbent presidential party.

Also important, however, is whether the model overall, and the interaction term in particular, change very much when single years are dropped. Table 3 provides some evidence on this point. Here we see that for the most part the magnitude and significance of the interaction slope do not change much as cases are dropped. Two cases stand out as having the greatest impact on the interaction effect, 1952 and 2000. The interaction coefficient becomes much stronger if 1952 is dropped and slightly weaker if the 2000 election is dropped. The 2000 election also represents the case whose deletion has the greatest impact on the significance of the slope. It is worth pointing out, though, that even with the 2000 election excluded, the model predicts an outcome for the 2008 election that is within 1 percentage point of the prediction generated by the full model.

Table 3.

The Sensitivity of the Incumbency Interaction Term to Case Deletion

The 2008 Forecast

With the 2008 values of presidential approval (30.8%) and consumer satisfaction (71.3), the national conditions index stands at 47.6, which generates a predicted 2008 outcome of 44.3% of the two-party vote for John McCain and 55.7% for Barack Obama. Using the out-of-sample standard error and the t-distribution, the estimated probability of a McCain victory stands at 0.08. These are long odds and do not auger well for the McCain camp.

Caveats

While the model has a high level of predictive accuracy and there is every reason to expect the actual outcome will be close to the predicted outcome, there are also reasons to exercise caution. First, there is no way for the model to incorporate factors such as candidate characteristics that are truly unique and important to individual elections. The assumption I make is that these factors are usually given relatively little weight by voters, in part because candidates are usually drawn from the same pool of relatively well-off, White, male politicians. The 2008 election, of course, is the first in which an African American candidate is representing one of the two major political parties. Will this matter in the end? Will it help or hurt McCain? It's hard to tell what role race will play in the end but, given that there is evidence (Kinder and Sanders Reference Kinder and Sanders1996) that racial attitudes played an important role if the 1988 election (Willie Horton), when two white men opposed each other, it would be naïve to think race will not play a role in 2008. The problem for forecasters is that there are no data points (no mixed-race presidential contests) and, hence, no real way to model the expected impact of race.

Another possibility is that Obama will “Gore” the forecasting model by ignoring the George W. Bush/Dick Cheney Record and give McCain a free pass. Based on the savvy demonstrated by his campaign staff during the primary season, I don't see this as a real possibility, but it could have an impact if he does. If there is a lesson for the Obama campaign to take from past elections it is that presidential performance matters, but it may require greater effort to frame the contest as a referendum when the president is not on the ticket. And, of course, this means that the McCain campaign's best strategy is to distance their candidate from President Bush and, first and foremost, avoid having the election turn into a referendum on the Bush presidency. The fact that campaigns generally act strategically usually means that voters get the referendum cue (more strongly I would suggest when the president is on the ticket) and election outcomes are fairly predictable. In this way, a “campaigns matter” perspective is entirely consistent with predictable elections. If either McCain's or Obama's assumed strategies dominate this fall, then the actual outcome could differ substantially from the predicted outcome.

Finally, this forecast is based on a single model from a relatively small number of years, and the findings might not hold with other retrospective variables. In fact, when percentage change in GDP over the first two quarters is substituted for aggregated personal finances in the model presented here, the incumbency interaction is not significant. This may indicate that the conditional effects from incumbency are more likely to emerge when using survey-based economic measures, such as economic perceptions, which may capture some political as well as economic content. Even with this important caveat, however, the bottom line is that forecasters should check to see if the reward-punishment aspects of their models are sensitive to the conditioning influence of incumbency.