Monday, November 03, 2008
On Polling
Wow, I sparked a lot of comments in response to me saying that I don't trust most of the polls and that I think McCain has about a 30% chance to win. (For fun, I've taken to calling him "The Arizona DJ, Emcee Cain". I encourage others to do as well).
Since a quick comment did an apparently piss-poor job of it, here's a post so I can use some formatting to help me better convey my thoughts.
==============
Main points:
1) I do not trust radically new and untested weighting methodologies. If you like, view this as part and parcel of having a fundamentally conservative outlook.
2) I do not believe the youth vote will turn out in greater proportion than *ever* before. A massive youth vote has been asserted for every election I have experienced. It has never materialized.
3) I believe that white conservatives and white liberals are about equally energized so their turnouts should be approximately equal. I believe that blacks will turn out in greater proportion than ever before and harder for the Dem candidate, but this will only translate into an extra 2-3 pct bump for Obama (there's a reason blacks are called a minority population). I believe hispanics will end up about the same (there's a Dem wave in the media, but I think this will be offset because of hispanic / black mutual dislike - see Hillary).
4) I believe undecideds will break hard for McCain (see late Dem primary and that Obama has more than enough exposure for people to make up their mind and the election is fundamentally a referendum on Obama)
5) That lots of polls have adopted the same (to me unproven) methodology (and I believe wrong methodology because of points 2-4) does not give me confidence in the methdology.
More importantly, aggregating together many polls together only improves accuracy when the estimates are independent and unbiased. Because most every poll has adopted similar assumptions, it means that if you believe there is an error in the assumptions, averaging will not converge on the true mean; it will instead converge to a biased estimate. Running Monte-Carlo sims ala 538 do not fundamentally fix this if the MC model fundamentally conforms to the same assumptions / models.
Or in real short, averaging reduces variance, but only improves accuracy if there's no systematic bias (methodological, not political).
6) The poll which has not significantly changed their weightings is the IBD poll (which as they are fond of noting was the closest poll in 2004). Their methods largely conform with my implicit model in points 2), 3) and 4). Because their assumptions best agree with my assumptions, that's the poll I trust.
For the past month or so, they've had Obama generally up 2-4 (strictly 1-7 up) with 9-13 undecided.
7) For the sake of simplicity, assume uniform distributions on Obama's reported lead (over 2-4%) being the actual decided lead. Assume 10% undecideds. And assume the break of undecideds is uniformly distributed from 55-45 (+1 pt for McCain) to 75-25 (+5 pt for McCain). That actually yields a 50-50 split.
Now factor in a little humility (i.e., assume I have to be a little bit off because I'm disagreeing with almost everyone) and we'll bump McCain's odds down a bit and call it 30%.
Note that I effectively changed two modeling assumptions (weighting and undecided break) and got something radically different than the results of assuming a "transformational" election.
==============
On the yard signs anecdote, sweet mother guys. When someone prefaces a paragraph with "Also anecdotally," maybe, just maybe they mean it as an anecdote, not as a statistical argument. It's funny that an McCain sign was put in my yard because my neighbors think that *I* am an Obama supporter. It's weird that I live in neighborhood that's 40-0 in yard signs over 109 houses (I have a hard time figuring that out even in a conservative part of a conservative area though the blind squirrel effect may be in play.)
And I am baffled that folks who should know me well don't pick up on this being my normal M.O., i.e., I cannot help myself from making asides if some random connection amuses me (and I don't really care if it doesn't amuse the person I'm talking to under my Joe Pesci rules). When I write something for an actual publication, I have to spend about half of my writing time figuring out if I want to strke my asides or integrate them into the main flow.
=============
On Obama not valuing signs - that's nothing new and something that every national campaign has known for years (Rep and Dem).
And indeed, no vote is changed by a yard sign. However, people want signs (and bumper stickers) as a show of group solidarity and as such yard signage is a good proxy for intensity of support. Thason and I have had this discussion before, though I don't know if I've blogged it.
And since there's an apparent need to burst "the Obama / Axelrod has all sorts of new insights into what is and is not important to a campaign" bubble, take a look at McCain's site. Signs are sold. For money. Specifically, $35 for 5. Campaigns leverage the fact that people emotionally attach to them like a sports team and use campaign paraphenalia to help raise money. Again it's different for local elections where name recognition matters.
============
On state distributions, I am not impressed by the number of polls used to craft an average because I believe they share a methodological flaw.
However, in response to Thason's questions, here's my current beliefs for the states Thason listed assuming my weighting model for declared intentions: (Note - that's not a direct shifting down from the national polls and includes touchy-feely factors about how I feel about the politics / demographics of each state)
FL: McCain + 2
NC: McCain +4
VA: Obama + 2
OH: McCain + 3
CO: Obama + 4
NV: Obama + 2
MO: McCain + 1
Other relevant states
PA: Obama + 4
NH: Obama + 3
MI: Obama + 7
NM: Obama + 5
IA: Obama + 5
Of these, only MI do I put out of reach for McCain under my second assumption of a hard undecided break for McCain (listed for comparison purposes). A very small undecided break for McCain (55-45) and VA goes blue. A medium size break (66-33) and VA stays red. A huge break (75-25) and CO, PA and NH go McCain with NM and IA being really tight.
============
On the disingenuity of only focusing on the "reverse-Bradley" effect: a) only so much time in a day and b) no one expects the Spanish inquisition.
The Bradley effect is well known and discussed (heck, it was Kaus's Halloween costume this year). However, what I don't see discussed is that Obama can't pick up 1 point from the black shift between the last poll and the election that we normally see because he already has the entire black vote (for all intents and purposes). As my 30-50% estimate of a McCain win relies on a hard undecided break for McCain, this is important to note (since I apparently have to be explicit - I don't expect the normal break to be a countervailing force).
That folks may have reported for McCain or Obama with the intention of voting for the other I find unlikely. It's far easier to hide intentions by declaring as undecided. We have twice as many undecideds as 2004. That's where they're hiding, not in each others' declared numbers. Based on the slant of media coverage (*very* pro-Obama, to the point of embarrassing Dan Rather!), I also expect that hidden intentions run much more for McCain than for Obama.
============
On Liberty registering - it's an anecdote I find interesting. Never seen a school make such an explicit effort before.
And to conclude with another quasi-related aside, which tangentially demonstrates my awareness of other registration efforts, in Roanoke I signed a petition to help get McKinney on the ballot in VA this year. My friends (none of which are my neighbors) thought I was nuts, but I figured more choices were better.
Update
Thason and I are on: $25 (me), $100 (him) for 4-1 odds on a McCain victory.
I'm not taking any more bets (I don't have a lot of disposable income now and need to start squirreling away my money in my mattresses for the upcoming Obama administration), but thought we should let everyone else be aware of the bet for vicarious election night fun.
|
Since a quick comment did an apparently piss-poor job of it, here's a post so I can use some formatting to help me better convey my thoughts.
==============
Main points:
1) I do not trust radically new and untested weighting methodologies. If you like, view this as part and parcel of having a fundamentally conservative outlook.
2) I do not believe the youth vote will turn out in greater proportion than *ever* before. A massive youth vote has been asserted for every election I have experienced. It has never materialized.
3) I believe that white conservatives and white liberals are about equally energized so their turnouts should be approximately equal. I believe that blacks will turn out in greater proportion than ever before and harder for the Dem candidate, but this will only translate into an extra 2-3 pct bump for Obama (there's a reason blacks are called a minority population). I believe hispanics will end up about the same (there's a Dem wave in the media, but I think this will be offset because of hispanic / black mutual dislike - see Hillary).
4) I believe undecideds will break hard for McCain (see late Dem primary and that Obama has more than enough exposure for people to make up their mind and the election is fundamentally a referendum on Obama)
5) That lots of polls have adopted the same (to me unproven) methodology (and I believe wrong methodology because of points 2-4) does not give me confidence in the methdology.
More importantly, aggregating together many polls together only improves accuracy when the estimates are independent and unbiased. Because most every poll has adopted similar assumptions, it means that if you believe there is an error in the assumptions, averaging will not converge on the true mean; it will instead converge to a biased estimate. Running Monte-Carlo sims ala 538 do not fundamentally fix this if the MC model fundamentally conforms to the same assumptions / models.
Or in real short, averaging reduces variance, but only improves accuracy if there's no systematic bias (methodological, not political).
6) The poll which has not significantly changed their weightings is the IBD poll (which as they are fond of noting was the closest poll in 2004). Their methods largely conform with my implicit model in points 2), 3) and 4). Because their assumptions best agree with my assumptions, that's the poll I trust.
For the past month or so, they've had Obama generally up 2-4 (strictly 1-7 up) with 9-13 undecided.
7) For the sake of simplicity, assume uniform distributions on Obama's reported lead (over 2-4%) being the actual decided lead. Assume 10% undecideds. And assume the break of undecideds is uniformly distributed from 55-45 (+1 pt for McCain) to 75-25 (+5 pt for McCain). That actually yields a 50-50 split.
Now factor in a little humility (i.e., assume I have to be a little bit off because I'm disagreeing with almost everyone) and we'll bump McCain's odds down a bit and call it 30%.
Note that I effectively changed two modeling assumptions (weighting and undecided break) and got something radically different than the results of assuming a "transformational" election.
==============
On the yard signs anecdote, sweet mother guys. When someone prefaces a paragraph with "Also anecdotally," maybe, just maybe they mean it as an anecdote, not as a statistical argument. It's funny that an McCain sign was put in my yard because my neighbors think that *I* am an Obama supporter. It's weird that I live in neighborhood that's 40-0 in yard signs over 109 houses (I have a hard time figuring that out even in a conservative part of a conservative area though the blind squirrel effect may be in play.)
And I am baffled that folks who should know me well don't pick up on this being my normal M.O., i.e., I cannot help myself from making asides if some random connection amuses me (and I don't really care if it doesn't amuse the person I'm talking to under my Joe Pesci rules). When I write something for an actual publication, I have to spend about half of my writing time figuring out if I want to strke my asides or integrate them into the main flow.
=============
On Obama not valuing signs - that's nothing new and something that every national campaign has known for years (Rep and Dem).
And indeed, no vote is changed by a yard sign. However, people want signs (and bumper stickers) as a show of group solidarity and as such yard signage is a good proxy for intensity of support. Thason and I have had this discussion before, though I don't know if I've blogged it.
And since there's an apparent need to burst "the Obama / Axelrod has all sorts of new insights into what is and is not important to a campaign" bubble, take a look at McCain's site. Signs are sold. For money. Specifically, $35 for 5. Campaigns leverage the fact that people emotionally attach to them like a sports team and use campaign paraphenalia to help raise money. Again it's different for local elections where name recognition matters.
============
On state distributions, I am not impressed by the number of polls used to craft an average because I believe they share a methodological flaw.
However, in response to Thason's questions, here's my current beliefs for the states Thason listed assuming my weighting model for declared intentions: (Note - that's not a direct shifting down from the national polls and includes touchy-feely factors about how I feel about the politics / demographics of each state)
FL: McCain + 2
NC: McCain +4
VA: Obama + 2
OH: McCain + 3
CO: Obama + 4
NV: Obama + 2
MO: McCain + 1
Other relevant states
PA: Obama + 4
NH: Obama + 3
MI: Obama + 7
NM: Obama + 5
IA: Obama + 5
Of these, only MI do I put out of reach for McCain under my second assumption of a hard undecided break for McCain (listed for comparison purposes). A very small undecided break for McCain (55-45) and VA goes blue. A medium size break (66-33) and VA stays red. A huge break (75-25) and CO, PA and NH go McCain with NM and IA being really tight.
============
On the disingenuity of only focusing on the "reverse-Bradley" effect: a) only so much time in a day and b) no one expects the Spanish inquisition.
The Bradley effect is well known and discussed (heck, it was Kaus's Halloween costume this year). However, what I don't see discussed is that Obama can't pick up 1 point from the black shift between the last poll and the election that we normally see because he already has the entire black vote (for all intents and purposes). As my 30-50% estimate of a McCain win relies on a hard undecided break for McCain, this is important to note (since I apparently have to be explicit - I don't expect the normal break to be a countervailing force).
That folks may have reported for McCain or Obama with the intention of voting for the other I find unlikely. It's far easier to hide intentions by declaring as undecided. We have twice as many undecideds as 2004. That's where they're hiding, not in each others' declared numbers. Based on the slant of media coverage (*very* pro-Obama, to the point of embarrassing Dan Rather!), I also expect that hidden intentions run much more for McCain than for Obama.
============
On Liberty registering - it's an anecdote I find interesting. Never seen a school make such an explicit effort before.
And to conclude with another quasi-related aside, which tangentially demonstrates my awareness of other registration efforts, in Roanoke I signed a petition to help get McKinney on the ballot in VA this year. My friends (none of which are my neighbors) thought I was nuts, but I figured more choices were better.
Update
Thason and I are on: $25 (me), $100 (him) for 4-1 odds on a McCain victory.
I'm not taking any more bets (I don't have a lot of disposable income now and need to start squirreling away my money in my mattresses for the upcoming Obama administration), but thought we should let everyone else be aware of the bet for vicarious election night fun.
|
