Archive
Introducing the 112th Congress
Now that the dust from the 2010 Elections has cleared, it is time to start looking at how the 112th Congress will compare to the 111th. The figure below tracks the ideological positions of the median House member, the mean Democrat, and the mean Republican since 1988. The replacement effect of the 2010 Midterm Elections is unlike anything in recent memory. The shift in the House median is two and a half times what was observed after the 1994 Election, wiping out the effect of Democratic gains in the previous two elections and then some. The 111th was the most liberal Congress in the past three decades; the 112th will be the most conservative.
The 2010 Elections had a profound effect on congressional polarization. Not only will the 112th House be the most polarized on record; 2010 will surpass 1994 as the most polarizing election cycle.
The figure below helps place the historic jump in polarization in perspective. It displays the ideological distributions of House members continuing on from the 111th Congress, alongside the distributions of the entering class of Republicans and the sitting Democrats they defeated. Two features stand out. The first is that the mean entering Republican (1.09) is substantially more conservative than the mean continuing Republican (0.82). In total, 77 percent of freshmen Republicans in the 112th Congress will locate to the right of the party median from the 111th. In other words, nearly 8 in 10 incoming House Republicans would have been on the right wing of the party in the 111th Congress.
The second standout feature is that, contrary to many media accounts, the Republican wave was an equal opportunity un-employer for Democrats. Democrats who lost their reelection bids were slightly more moderate than those who retained their seats—0.61 compared to 0.64—but the difference is statistically insignificant. The attrition rate was greater among centrist Democrats, but there were so few Democrats remaining in the political center that the losses had to come from elsewhere.
The polarization resulting from the 2010 Midterms is fundamentally different and more worrisome than what had preceded it. By historical standards, the post-war era stands out as a period of relatively low partisan polarization. This is largely attributable to the coalition between Northern and Southern Democrats. The increase in polarization during the 103rd through 105th Congress corresponds to the tail end of the Southern partisan realignment, a period during which southern districts that had traditionally elected moderate Democrats (a.k.a. Dixiecrats) began electing conservative Republicans. As the Southern Democrats gradually disappeared throughout the 1980s and 1990s, the parties became more clearly defined, thus returning congressional polarization to the historical norm.
The hollowing out of the political center explained the momentous rise in polarization during the Southern realignment. Now that only a handful of moderates remain in the House, polarization can no longer be portrayed as a story of vanishing moderates. It appears the rise of the extremists has stepped up as the driving force behind congressional polarization.
Forecasting House Elections with FEC Records
As election day nears, I thought it might be an interesting exercise to see how accurately I could forecast election outcomes using only information derived from campaign finance records. Campaign finance records represent a rich data source that speaks to many areas of U.S. politics. Elections are no exception. I’m not the first to incorporate information on fundraising into a forecasting model. However, to the best of my knowledge, I am the first to attempt to forecast election outcomes based solely on FEC records. Despite the handicap of excluding all information from polls, InTrade, expert raters, and other data sources used to forecast elections, the model’s predictions are remarkably accurate. In fact, the out-of-sample predictions for House races outperform the polls.
Contribution records contain far more useful information than what can be expressed by fundraising tallies. For instance, they provide a way to estimate a reliable set of candidate positions via CFscores. The CFscores update almost in real time as FEC records are released during the course of an election cycle. In other words, we don’t have to wait until the election is over to get ideological measures for candidates. The CFscores enable my forecasting model to account for factors that other models ignore, such as adjusting for whether the ideological extremity of Tea Party candidates will hinder Republican electoral prospects this November. In addition, I can forecast ideological quantities of interest such as how the location of the median member of Congress will change after the election. This is perhaps less useful in terms of the horse-race but is probably a better overall measure for the type of policy we should expect from the next Congress.
Also informative are the patterns of individual donors across elections. I’ve been working on assigning unique contributor IDs that link contribution records from the same donor across election cycles and across state and federal elections. This may not seem like a big deal, but the ability to track the behavior of individual donors across elections cycles unlocks a wealth of information that had previously been trapped inside the dataset. For instance, linking records across years makes it possible to calculate the proportion of a candidate’s funds that came from first-time donors as opposed to veteran contributors. At the level of campaigns, this can convey information about a candidate’s success in activating supporters. At the national level, the median CFscore across all first time donors serves as a good proxy for the enthusiasm gap.
Much of the model’s predictive power is owed to an idea I borrowed from Sandy Gordon. Sandy has an interesting paper in which he uses past performance of expert election raters to calibrate their predictions in future elections. The paper gave me the idea of treating the hundreds of thousands of donors who had given in previous election cycles as de facto expert raters by looking at the percentage of funds given to winning candidates in previous election cycles. The idea behind this is simple. Some contributors give a greater proportion of their money to candidates that go on to win, while others spend the majority of their money on losing candidates. All else equal, the more money a candidate raises from the type of donors who give to winners, the more likely he is to win. (For additional details on what goes into the model, I include at the bottom of this post a description of the other model predictors, as well as links to download the R script and dataset.)
Tom Holbrook over at Politics by the Numbers nicely overviews the accuracy of election polls. He shows that during the 2006 and 2008 election cycles approximately 85 percent of House candidates who led in the polls 45 days before the election went on to win. As a comparison, the out-of-sample predictions from my forecasting model correctly identify the winner in over 94 percent of combined 2002-2008 House elections.
It is worth noting that my sample includes a number of less competitive House races that lack polling data and hence are excluded from the poll based predictions. The larger sample accounts for some of the increase the prediction rate but not all of it.
The model’s seat share predictions are also close to the mark. The table below shows the number of seats won by Democrats that the model predicts above/below the observed outcome (e.g. a value of 4 indicates the model predicts that Democrats would win four more seats than they actually did; a value of -4 indicates the model predicts Democrats would win four fewer seats than they actually did). I ran 1000 bootstrapped simulations for each election cycle to get uncertainty estimates. The first column reports the median value from bootstrapped runs and the other two columns display the upper and lower 95 percent confidence bounds.
|
Median |
CI Lower Bound |
CI Upper Bound |
|
|
2002 |
4 |
-1 |
8 |
|
2004 |
4 |
1 |
6 |
|
2006 |
-12 |
-18 |
-5 |
|
2008 |
-1 |
-1 |
5 |
The predictions are very close to the actual outcome in all but one election cycle. In 2006 the model under predicts Democratic gains by a considerable margin. This might be a result of the Mark Foley factor but I would need more evidence to support that claim. It is just as likely that it reflects a failure of the model to adjust for the effects of partisan momentum in landslide years. (As a note, I noticed that the confidence bounds only contain the observed value in two of the four elections. This suggests that I should probably up the amount of uncertainty in the bootstrapping scheme above the software’s default.)
A major advantage of using campaign finance data to forecast elections is that it costs next to nothing. There is no need to commission polls or pay expert raters. One needs only to collect freely available data and fit a model. That being said, campaign finance based forecasting could have the greatest impact in state level elections where polling data is sparse but fundraising abounds. Although I’m not convinced that they would be, it might be interesting to seeing if the model predictions are as accurate for state elections as they are for federal elections. The catch is that not every state is quite up to speed with the FEC in releasing contribution records to the public in a timely manner, but this problem is fast solving itself as the disclosure process becomes increasingly digitized.
As much as I would like to have predictions ready for the upcoming election, I haven’t had the time to fully organize the dataset. Those forecasts will have to wait until this weekend. In the meantime, I’ve made the dataset and R script available for download for anyone who might be interested.
Predictors
Candidate Positioning: I use candidate CFscores to measure ideology.
Picking Winners: For each contributor, I calculate the percentage of funds given to winning candidates in previous election cycles. Then for each candidate, I calculate the mean value of his contributors weighted by the dollar amounts received.
First-Time Donors: I calculate the percentage of a campaign’s donors that gave the first time during that election cycle. This is intended as a proxy for the enthusiasm gap.
Donor Mood: I construct an aggregate variable by calculating the CFscore of the median first-time donor to proxy public mood.
Fundraising Success: This is measured by the log total amount raised by each candidate and the log sum of unique contributors.
Source of Funds: For each candidate, I calculate the percentages of funds raised from PACs, individual donors, party committees, and self-funding.
A Visual History of Senate Polarization from 1967 to 2010
(Maximize the video and set the quality to 720p HD for best viewing. The video is also available for download from here.)
Each senator is marked by a two-letter state code and is color coded by party membership (Democrat, Republican, Independent). The top grey bar is the distance between the mean Republican and the mean Democrat (the standard measure of partisan polarization). The size of each party’s coalition is displayed at the end of the bars. The lower grey bar is the gridlock interval, which is measured as the distance between the filibuster pivots (the 34th and 66th most conservative senators before the Senate rule change in 1975 and the 41st and 60th thereafter). The ‘M’ imposed on top of the bar marks the position of the median senator.
The ideological estimates are constructed by scaling senate voting records with dynamic optimal classification. Keith Poole’s paper on optimal classification can be found here, and the paper which explains how to extend the method to smooth legislator ideal points over time can be found here.
Ideological Cartography 
