Tuesday, October 19, 2010 at 4:46 pm by Darryl
A little while ago at Horsesass I briefly wrote about today’s Public Policy Polling (PPP) poll in the senate race between Sen. Patty Murray (D-WA) and Dino Rossi (R). The poll, taken from 14th to the 16th of October on a sample of 1,873 likely voters, has Murray leading Rossi 49% to 47% (2.3% margin of error). PPP conducts robopolls.
Here I add to that post in several ways:
- Do the usual Monte Carlo analysis of this poll to determine the probability that the leader wins in a hypothetical election held now.
- Do a combined analysis of all the October polls to determine who would win in an election held now.
- Evaluate the idea (from Stuart Elway, and further developed by me) that robopolls have to race too close
Part 1: From a million simulated elections of 1,873 “voters” voting probabilistically for Murray 49% of the time and Rossi 47% of the time, Murray won 730,066 times and Rossi won 264,429 times. Thus, this poll give evidence that, in an election held now, Murray would beat Rossi with a 73.4% probability. (Since this value is less than 95%, a statistician would say that the result falls within the margin of error.) Here is the picture of how the simulated elections came out:
Part 2: Here are the eight October polls
Pooling these eight give us a whopping 6,779 “votes” of which 6,400 go to Rossi or Murray. Murray gets 3,316 (48.9%) and Rossi gets 3,084 (45.5%), with 379 (5.6%) going elsewhere. In the simulated elections, Murray won 979,617 times and Rossi won 19,978 times. In an election held now, these polls give evidence that Murray would have a 98% probability of beating Rossi. Translating words into pictures…
Part 3: Based on Elway’s suggestion, I hypothesized last Tuesday that the enthusiasm gap might explain why robopolls seem to systematically have the race very close while the live-interview polls give Murray a solid lead. Since then, there have been five new polls in the race.
The robopolls are represented by PPP (+2 for Murray), Rasmussen (+3) and SurveyUSA (+3). The live-interview polls are represented by the KCTS/KPLU/Washington Poll (+8) and CNN/Time/Opinion Research (+8).
Clearly, all of the poll released after Elway’s suggestion have been entirely consistent with the idea. The live-interview polls have a more than doubled spread between the candidates. So, lets look at what each type of poll says about the probability of each candidate winning.
Using the three October live-interview polls (add Elway to the list), there were a total of 1,800 “votes” of which 1,660 go to Rossi or Murray. Murray received 913 (50.7%) to Rossi’s 746 (41.4%) votes; 141 (7.8%) votes went to neither candidate. The simulated elections went to Murray 998,204 times and Rossi 1,707 times.
The evidence from live polls is that Murray would win an October election with a probability of 99.8% to Rossi’s 0.2%.
There were four robopolls released in October—add the older Rasmussen poll to the above list. There were a total of 4,229 “votes” of which 4,028 went to the candidates. Murray received 2,036 (48.1%) to Rossi’s 1,993 (47.1%) votes—quite close! But with such a large sample size, even this close result weighs heavily toward the leader. In the simulated elections, Murray won 682,825 times to Rossi’s 313,350 wins.
So based on the robopolls, Murray would have a 68.5% probability and Rossi would have a 31.5% probability of winning an election held now.
So…99.8% versus 68.5%—the difference between the two methods is quite striking. If I am correct and this difference reflects an enthusiasm gap—that is, Murray supporters are less engaged or excited about the election than the Rossi supporters, leading to more hang-ups on robopolls—then Murray’s lead is being systematically underestimated in the robopolls. The phenomenon might even be more widespread, leading to underestimation of the lead, or overestimation of the loss percentage, by Democrats in other races across the nation.
What I expect as we approach election day is that some Democrats will become more engaged and excited by the race. This should reduce the gap between robopolls and live-interview polls. Specifically, I think robopolls will start looking more and more like the live-interview polls as election day approaches.