| Previous: Format: | Up: IMPUTE | Next: Recommended Use: |
As a researcher, in evaluating elections, you face a difficult choice in dealing with uncontested districts. You could use the actual reported election results, but this clearly is undesirable; candidates who received more than 95 percent were very unlikely to have done so given a credible opponent. Therefore following this method is almost certain to distort your predictions and analysis, for example by making the system seem less competitive than it really is.
On the other hand, you could choose to eliminate these districts from the analysis, but in most cases doing so is equally undesirable. Presumably, uncontested districts tend to be among the least competitive districts, so using only contested districts would distort results to make the electoral system seem more competitive than it actually is. Furthermore, eliminating observations reduces the information available in developing your model estimates, at the cost, in general, of sacrificing the accuracy of predictions.
Probably a more desirable means of coping with this difficulty is to estimate the underlying distribution of party preferences in a district at the time of the uncontested election; that is, to estimate what the results of a hypothetical contested election in the district probably would have been. The CHG! operator provides one blunt means of adjusting these district votes by setting hypothetical election results in all uncontested districts as 75 percent for the uncontested victor and 25 percent for the hypothetical opponent. While this rather crude method of substitution might be adequate for certain purposes, in general we probably need a more sophisticated method to estimate these district level results. Arbitrarily setting these numbers suggests that we have more certainty regarding underlying preferences in the district than we actually do, thereby deflating the standard errors reported in the model.
The IMPUTE command offers one method of gauging the hypothetical vote values in these uncontested districts. IMPUTE transforms the YVOTE variable by creating a new dependent variable for use in JudgeIt analysis, as follows. For each of the uncontested districts (by default, those in which a candidate wins more than 95 percent of the reported vote), JudgeIt will use the dependent and explanatory variables to formulate predictions of the underlying vote in each of these districts.For the purposes of this estimation, JudgeIt treats the YVOTE variable as though a CHG! operator has been applied to it. That is, it sets uncontested district votes to 0.75 and 0.25, or whatever replacement values have been indicated with the UNC command. Then it adjusts them up or down randomly to simulate the uncertainty in these predictions.Error in vote predictions (or any other social science prediction) comes from two sources--imprecision in our model's coefficient estimates, and uncertainty caused by the nondeterministic nature of political behavior. Therefore the IMPUTE command actually adjusts the predictions twice, once randomly according to a posterior distribution representing theoretical expectations of coefficient error, and once randomly according to a posterior distribution representing the stochastic nature of political events. Since the predictions are adjusted randomly, they will contain variance that will be reflected in higher estimates of the standard errors. That is, the program will allow you to use likely underlying vote estimates for uncontested districts, but also will adjust your standard errors upward to reflect the lower level of certainty in these results.