Sunday, November 21, 2010

Election Results vs. Predictions

I have held off counting coup on the Democrats in the House of Representatives since a number of races still hang in the balance.

I have not found any single source tracking all those races other than the liberal local New York paper's (LLNYP) House race map.  It changes occasionally, which indicates that it is still being updated.

Their partial count shows that Democrats have been declared winners in 190 races and Republicans in 240, with five races remaining to be determined.  Republicans lead in four of those five, with the average lead being about 500-600 votes, though in one case it is only 27 votes.  Those races, and their status as this is posted are here:


Dem  Rep R - D
CA-11 McNerney  Harner -628
CA-20 Costa  Vidak 27
NY-01 Bishop  Altschuler 383
NY-25 Maffei  Buerkle 659
TX-27 Ortiz  Farenthold 799

If these races were all to go the way they now lean, Republicans would hold a 244 to 191 seat majority, for a swing of 66 seats.

My November 1 prediction, corrected from October 31, and based on the final Gallup generic congressional ballot question and a Monte Carlo model, was 243-192, a 65 vote swing.

I can't take much credit for that, since the model was not mine, but I can say that Nate Silver's final prediction was for a 45-ish vote swing to about a 233-202 Republican majority.  The difference isn't quite as large as it seems, since it requires a change in only ten results, so, if the five seats in doubt go Dem, Silver's prediction looks about as good as mine.

Silver chose explicitly to ignore the Gallup GCB.  His analysis used 100,000 Monte Carlo simulations of the nationwide election doing independent random draws of each district's result from a separate probability distribution for that district determined by his analysis of all the polling data available.  The model I used did a similar thing, but used the Generic Congressional ballot to shift the 2008 results in each district.

The approach I favored tends to allow for correlations between races in separate districts, where Silver's approach tends to give uncorrelated results, though correlations do get in through the polling results.