Friday, May 23, 2008 at 6:33 pm by Darryl
Poll Analysis: Obama Makes Gains Against McCain
| Obama | McCain |
| 32.7% probability of winning | 65.4% probability of winning |
| Mean of 262 electoral votes | Mean of 276 electoral votes |
Yesterday, Sen. Barack Obama had a 19.3% probability of defeating Sen. John McCain in a general election. Today, there were four new polls today that contribute to this match-up. Obama regains gains some lost ground.
Now, after 10,000 simulated elections, Obama wins 3,266 times (plus he takes the 196 ties), and McCain wins 6,538 times. If a general election were held today, Obama would have a 34.6% (32.66% plus 1.96% for ties) probability of winning the election, and McCain has a 65.4% probability of taking the election.
The long term trends in this race can be seen from a series of elections simulated every 7 days using polls from 23 Sep 2007 to 23 May 2008, and including polls from the preceding 1 month (FAQ).
Here is the distribution of electoral votes [FAQ] for the current time period:
- 10000 simulations: Obama wins 32.7%, McCain wins 65.4%.
- Average ( SE) EC votes for Obama: 262.5 ( 14.9)
- Average (SE) EC votes for McCain: 275.5 ( 14.9)
- Median (95% CI) EC votes for Obama: 262 (238, 293)
- Median (95% CI) EC votes for McCain: 276 (245, 300)
| State | EC Votes | # polls | Total Votes | % Obama | % McCain | Obama %wins | McCain %wins |
|---|---|---|---|---|---|---|---|
| Alabama | 9 | 1* | 812 | 39.5 | 60.5 | 0.0 | 100.0 |
| Alaska | 3 | 2 | 1001 | 45.7 | 54.3 | 0.1 | 99.9 |
| Arizona | 10 | 1 | 490 | 44.7 | 55.3 | 1.4 | 98.6 |
| Arkansas | 6 | 1 | 480 | 40.6 | 59.4 | 0.0 | 100.0 |
| California | 55 | 3 | 1978 | 57.6 | 42.4 | 100.0 | 0.0 |
| Colorado | 9 | 1 | 450 | 53.3 | 46.7 | 92.5 | 7.5 |
| Connecticut | 7 | 1* | 1476 | 59.8 | 40.2 | 100.0 | 0.0 |
| Delaware | 3 | 1* | 553 | 55.0 | 45.0 | 98.6 | 1.4 |
| D.C. | 3 | 0 | (100) | (0) | |||
| Florida | 27 | 3 | 2899 | 47.9 | 52.1 | 0.6 | 99.4 |
| Georgia | 15 | 3 | 1733 | 42.9 | 57.1 | 0.0 | 100.0 |
| Hawaii | 4 | 1* | 546 | 66.3 | 33.7 | 100.0 | 0.0 |
| Idaho | 4 | 1* | 553 | 42.9 | 57.1 | 0.0 | 99.9 |
| Illinois | 21 | 1* | 546 | 65.9 | 34.1 | 100.0 | 0.0 |
| Indiana | 11 | 2 | 1779 | 49.0 | 51.0 | 15.9 | 84.1 |
| Iowa | 7 | 1 | 430 | 51.2 | 48.8 | 70.2 | 29.8 |
| Kansas | 6 | 1 | 445 | 38.2 | 61.8 | 0.0 | 100.0 |
| Kentucky | 8 | 1 | 546 | 36.3 | 63.7 | 0.0 | 100.0 |
| Louisiana | 9 | 1* | 465 | 44.1 | 55.9 | 0.9 | 99.1 |
| Maine | 4 | 1 | 445 | 57.3 | 42.7 | 99.7 | 0.3 |
| Maryland | 10 | 1* | 577 | 57.0 | 43.0 | 99.9 | 0.1 |
| Massachusetts | 12 | 1 | 450 | 56.7 | 43.3 | 99.4 | 0.6 |
| Michigan | 17 | 1 | 445 | 49.4 | 50.6 | 39.6 | 60.4 |
| Minnesota | 10 | 1 | 994 | 55.0 | 45.0 | 100.0 | 0.0 |
| Mississippi | 6 | 1 | 558 | 41.9 | 58.1 | 0.0 | 100.0 |
| Missouri | 11 | 2 | 1856 | 48.0 | 52.0 | 4.5 | 95.5 |
| Montana | 3 | 1* | 455 | 47.3 | 52.7 | 11.0 | 89.0 |
| Nebraska | 5 | 1 | 445 | 43.8 | 56.2 | 0.8 | 99.2 |
| Nevada | 5 | 1 | 430 | 46.5 | 53.5 | 6.9 | 93.1 |
| New Hampshire | 4 | 3 | 1314 | 48.8 | 51.2 | 17.1 | 82.9 |
| New Jersey | 15 | 1 | 707 | 63.6 | 36.4 | 100.0 | 0.0 |
| New Mexico | 5 | 2 | 983 | 52.3 | 47.7 | 92.0 | 8.0 |
| New York | 31 | 2 | 976 | 57.9 | 42.1 | 100.0 | 0.0 |
| North Carolina | 15 | 4 | 2243 | 46.2 | 53.8 | 0.0 | 100.0 |
| North Dakota | 3 | 1* | 218 | 46.3 | 53.7 | 13.8 | 86.2 |
| Ohio | 20 | 4 | 2970 | 49.8 | 50.2 | 40.9 | 59.1 |
| Oklahoma | 7 | 1* | 569 | 40.1 | 59.9 | 0.0 | 100.0 |
| Oregon | 7 | 1 | 455 | 57.1 | 42.9 | 99.7 | 0.3 |
| Pennsylvania | 21 | 6 | 4784 | 53.6 | 46.4 | 100.0 | 0.0 |
| Rhode Island | 4 | 1* | 572 | 58.2 | 41.8 | 100.0 | 0.0 |
| South Carolina | 8 | 1* | 554 | 48.4 | 51.6 | 20.0 | 80.1 |
| South Dakota | 3 | 1* | 221 | 39.8 | 60.2 | 0.2 | 99.8 |
| Tennessee | 11 | 1* | 450 | 42.2 | 57.8 | 0.1 | 99.9 |
| Texas | 34 | 2 | 1001 | 44.9 | 55.1 | 0.0 | 100.0 |
| Utah | 5 | 1 | 537 | 30.4 | 69.6 | 0.0 | 100.0 |
| Vermont | 3 | 1* | 576 | 68.4 | 31.6 | 100.0 | 0.0 |
| Virginia | 13 | 3 | 1725 | 48.6 | 51.4 | 9.5 | 90.5 |
| Washington | 11 | 2 | 1088 | 56.2 | 43.8 | 100.0 | 0.0 |
| West Virginia | 5 | 1* | 540 | 39.8 | 60.2 | 0.0 | 100.0 |
| Wisconsin | 10 | 1 | 450 | 47.8 | 52.2 | 15.7 | 84.3 |
| Wyoming | 3 | 1* | 508 | 39.4 | 60.6 | 0.0 | 100.0 |
* denotes that an older poll was used
Details of the methods are given in the FAQ.
The most recent analysis in this and other match-ups can be found from this page.
Friday, May 23rd, 2008 at 8:18 pm
Did you include the Virginia poll from yesterday? Survey USA, Obama +8
Friday, May 23rd, 2008 at 8:51 pm
I did include the Virginia poll. I wrote about it yesterday in the poll summary…here.
BTW: you can always see the polls included in an analysis. Click on the state of interest, then in the summary table, click on the number of polls in the state.
Friday, May 23rd, 2008 at 11:47 pm
Darryl – Personally I think these polls are meaningless right now, and the variability of them is proof enough.
While Clinton is in this race Obama’s polling numbers will suffer, and McCain will benefit from that. Your approach of holding onto only the most recent numbers doesn’t eliminate the variability.
Saturday, May 24th, 2008 at 6:57 am
I’m a bit confused by your methodology. You cited the reason VA poll was not showing better for Obama but the same logic would seem to place Indiana more in his favor. i.e. A larger more recent poll. Also the SUSA OH poll is more recent though it is smaller in sample.
The other thing I find quite strange is the fact that polls as old as February are used if there are no other polls avavailbe yet polls as recent as early april seem tossed out if there are more recent polls.
looks like you toss 30 maybe 45 days IF there is a more recent poll. But you’ll keep an 4 month old poll around otherwise.
This amplifies variability on the one hand while rewarding the influence of dated polls in other infrequently polled states. This gives the overall map an apples and oranges element akin to super tuesday polling that showed Obama winning just IL and GA.
e.g.
state 1: 1 old poll
state 2: 1 new poll but discarding 1 poll that is more recent than state 1’s old poll
stae 3: 3 recent polls
etc.
There are advantages to what you are doing but in reality when looking at a whole map you have a combination of decent data and poor (or should I say old) data. In other words if a feb poll is of use in one state an april poll should not be discarded in another else the “big picture” is a distortion of sorts. In essence the overall map is operating with 50 differnet rule sets because Feb is fine for Maryland but unsuitable for Ohio.
Saturday, May 24th, 2008 at 8:25 am
Daniel,
“Personally I think these polls are meaningless right now”
In the sense that a half-time score is “meaningless” as to the final score of a football game, sure. But…I like following the score over the entire game, not just after the two minute warning. Given that many, many millions of dollars have been spent on polls for this election (state head-to-head polls as well as national polls), it would be hard to argue that the information is “meaningless.” It is, apparently, just not the information you want (the final score instead of a running score).
“and the variability of them is proof enough.”
I am not sure how variability (presumably you mean variability over time?) “proves” this. The methods explicitly try to quantify the sampling variability in a meaningful way, in addition to finding a central tendency. The result is a relatively smooth time-trend plot. If the analysis were meaningless, I would expect a pattern in the time-trend plot that looks “noisy.”
“While Clinton is in this race Obama’s polling numbers will suffer”
Yep…but isn’t that like arguing a first quarter football score is “invalid” because the first string quarter back is temporarily out of the game for your team?
“and McCain will benefit from that.”
Probably. But the benefit is real…I mean, people are being asked who they would vote for in a match-up and they give a response.
“Your approach of holding onto only the most recent numbers doesn’t eliminate the variability””
The goal isn’t to “eliminate the variability.” The goal is to quantify that variability by generating, at each time point, the entire distribution of Electoral College outcomes.
Saturday, May 24th, 2008 at 8:39 am
in light of your response on the graph I reiterate my previous observation.
data points on the graph get impacted by variable rules with respect to time and age of the poll. I understand the method and it is what it is I guess.
an older Feb Maryland poll impacts the graph while a March Ohio poll does not becasue of a may OH poll. Seems wierd to me put obvioulsy some compromises are required when all polls arrive at different times anyway.
using your football game analogy plays from the first quarter impact the score but some plays from the second quarter do not if there is a more representaive “play” from the third quarter that erases action in the second quarter.
good effort …I’ll punt
Saturday, May 24th, 2008 at 8:59 am
Tone,
“I’m a bit confused by your methodology. You cited the reason VA poll was not showing better for Obama but the same logic would seem to place Indiana more in his favor. i.e. A larger more recent poll. Also the SUSA OH poll is more recent though it is smaller in sample.”
The algorithm is to use polls taken in the last month (unless there are none, in which case the single most recent poll is used). In Indiana, there are two “recent” polls:
– A SurveyUSA poll taken 28 Apr through 30 Apr on 1,279 individuals giving Obama 48% and McCain 47%.
– A Research 2000 poll taken 23 Apr through 24 Apr on 600 individuals giving Obama 43% and McCain 51%
– For the SurveyUSA poll there were 1,279×0.48=614 “votes” for Obama and 1,279×0.47=601 “votes” for McCain.
– For the Research 2000 poll there were 600×0.43=258 Obama “votes” and 600×0.51=306 “votes” for McCain.
– There were a total of 614 + 601 + 258 + 306 = 1779 total “votes” for either Obama or McCain.
– In aggregate, Obama’s percentage is (614 + 258)/1779 = 49%.
– In aggregate, McCain’s percentage is (601 + 306)/1779 = 51%.
Hence, McCain has a slight advantage in Indiana currently.
“The other thing I find quite strange is the fact that polls as old as February are used if there are no other polls avavailbe yet polls as recent as early april seem tossed out if there are more recent polls.”
The states that are polled less frequently tend to be those states in which the outcome of the election is believed to be fairly certain. Organizations pay a lot of money for polls to be conducted, so it would be considered a big waste of money to repeatedly poll states like Wyoming, Vermont, Rhode Island, and Oklahoma. (And D.C. hasn’t yet been polled, so I am using the 2004 election results.)
States that are nearly tied (e.g. Ohio, Pennsylvania), previously gave a surprising result (e.g. Virginia, North Carolina), or are potentially changing (i.e. there is active campaigning before a primary, like Indiana in late April) tend to be polled most frequently. That is…some organization is curious enough to pay to get a poll done.
“looks like you toss 30 maybe 45 days IF there is a more recent poll. But you’ll keep an 4 month old poll around otherwise.”
The algorithm is given in the FAQ.
“This amplifies variability on the one hand while rewarding the influence of dated polls in other infrequently polled states.”
Nope…your criticism would have more merit if polling were done in states randomly. But they aren’t.
“There are advantages to what you are doing but in reality when looking at a whole map you have a combination of decent data and poor (or should I say old) data.”
But…the “poor data” are in states whose outcome is not likely to change. Therefore the data are not likely “poor.”
On top of that, the winner-take-all nature of the electoral college (in all but 2 states) means that the precise percentage isn’t needed for states whose outcome is relatively certain.
“In other words if a feb poll is of use in one state an april poll should not be discarded in another else the “big picture” is a distortion of sorts. In essence the overall map is operating with 50 differnet rule sets because Feb is fine for Maryland but unsuitable for Ohio.”
This would only be true if the frequently polled states were just like the infrequently polled states. They’re not.
Saturday, May 24th, 2008 at 10:22 am
I see Daryl thanks.
That means assuming there are no new polls by 6/9 in this method Virignia would switch to pink about 6/9 when the 5/8 Rasmussen poll expires.
I guess looking back I see that is basically what recently happened in WI and IN when the indianapolis star expired for example.
That star poll is only 2 days older than the recent poll you use so i imagine the oldest 4/23 poll would fall away tomorrow and you will have only the most recent IN poll. I’ll check to see if Indiana turns blue tomorrow with no new polls. lol
I get it it just seems strange. WHy not average slightly longer and weight the most recent polls more? That would appear to lessen what we are getting with IN flipping around on poll expiration versus actual new polls.
Of course i thought IN was like wyoming anyway.
Saturday, May 24th, 2008 at 12:21 pm
Tone,
“I get it it just seems strange. WHy not average slightly longer and weight the most recent polls more? That would appear to lessen what we are getting with IN flipping around on poll expiration versus actual new polls.”
I could. But that would require a more complex set of assumptions than just using a cut-off (i.e. a decay function). BTW: my cut-off will be reduced below one month as the frequency of polling picks up.
“Of course i thought IN was like wyoming anyway.”
Nope…not in late April. There was a bunch of campaigning going on. The folks who pay thought the situation was dynamic enough to be worth paying for the polls.
Saturday, May 24th, 2008 at 12:30 pm
Tone,
“data points on the graph get impacted by variable rules with respect to time and age of the poll. I understand the method and it is what it is I guess.”
The rule-set is quite small…it is the arrival times of polls that vary.
“an older Feb Maryland poll impacts the graph while a March Ohio poll does not becasue of a may OH poll. Seems wierd to me put obvioulsy some compromises are required when all polls arrive at different times anyway.”
That is because the “polling market” has determined that there is relatively little going on in MD right now.
“using your football game analogy plays from the first quarter impact the score but some plays from the second quarter do not if there is a more representaive “play” from the third quarter that erases action in the second quarter.”
In fact, there are plays with low probability of scoring and plays with a high probability of scoring. If we only got instant feedback on the latter and delayed feedback on the former, we would usually have a good estimate of the score.
Obviously, it would be idea to have a fresh set of polls every day, but nobody will pay for that. So I use the data that are available.
The alternatives are: (1) using national polling, which does a better job of getting the score out quickly, but has the disadvantage that they don’t actually use the same scoring system (i.e. they estimate a popular vote, not an electoral college vote). In the football analogy, it is like being given only the number of times each team has scored, without mention of whether the points arose via a touchdown, 2 pt conversion, 1 pt conversion, field goal or safety. And (2) using the “markets” which is an interesting, albeit experimental, way of generating an approximate score.
Saturday, May 24th, 2008 at 12:51 pm
Daryl not really meaning to be too critical I was just curious about what seemed like a few strange results. I think the falling away or “expiring” polls which are sometimes only days apart is actually the bulk of what I didn’t realize was happening. But now I expect to see one of those delayed “field goal” results which is IN turning light blue as soon as a previous “safety” poll expires.
Thanks for your responses to my obvious misunderstandings.
I am surprised MI has not had more recent polling given aborted primary and the seeming importance on the electability arguments of Clinton.
Sunday, May 25th, 2008 at 3:07 pm
Tone,
You’re welcome, and no offense taken.
“I am surprised MI has not had more recent polling given aborted primary and the seeming importance on the electability arguments of Clinton.”
I agree…Michigan’s results have been rather surprising, yet only two March, one April and one May polls (so far). I suspect this will change.