When does a poll sway opinion?

When I was in college, studying statistics, there was a book that was required reading. It was called, "How to Lie With Statistics". I loved that book. It really showed me how polls can really be slanted to "prove" what the pollsters wanted to prove all along.

With regards to that, I remember an old joke, that came out of the days of the Soviet Union. There was a race between the Soviet Union and the U.S. The U.S. won.

The Soviet headline read, "Soviets came in second in race. The United States came in next to last"!

Yes, of course statistics can be lied with. And of course you can slant a poll, if you'd want to. Thats why if you come across a poll commissioned by Target showing that people think Target is da bomb, you'd better be sceptical.

The question here was whether the mainstream media's election polls being undertaken now are examples of that. Specifically, whether that one ABC/WaPo poll A Lone Voice cited was evidence of manipulation, or just an example of the statistical aberrations that are inherently part of polling.

If you'd like to know more about that, feel free to read up on the posts above. Alternatively, of course, you can just throw your arms up and say, "hell I dont care about all that detail, I'm just sure they're fixing stuff anyway, I mean you know you cant trust all dem polls!".

I must have read the same book, because I've come to question many of our government's statistics. LOL

Thanks, nimh, for the in-depth response.

No conspiracy thinking here, other than the clever tags that had been assigned and the way the thread had been thumbed down into oblivion. I'm still learning the reconfigured site, though.

If you noticed, I did post the RealClear politics snapshot polls for that timeframe, and the Post/ABC poll was way off from any of the others, by as much as 6 to 7 points on the average. But the Post poll was the one that received the headlines that day.

Re the over sampling of black voters: Using 163 responses, the Post is using numbers far above not only the black population, but the black voting population. If you read the notes on the poll, the over sampling was 'necessary' because the 'numbers' they (needed, wanted, required?) were not reached in their regular polling.

Don't you think this is flawed methodology? In the bag for Obama? Sure, but I don’t think this was a concerted effort to rig a poll. I just think it was the typical over-effort to be ‘fair.’

I do agree with you about self identification of parties, and more voters this year are leaning dem rather than repub. You bring up a good point for this in the poll that I overlooked, thanks.

It's not like this doesn't cut both ways.

USA-Today/Gallup, 9/5-7/08 823 LV McCain 54 Obama 44

That poll was WAY out of line with any other national poll, yet it's the one that got all the news that week.

Cycloptichorn

That poll by USAToday/Gallup will probably throw their averages way out there in left field. LOL

I havent looked into this poll, and hey I'm also just a layman observer, so I could be wrong. But if they oversampled black voters I'm guessing it is because they wanted enough of them to be able to make statistically reliable conclusions about the views of blacks, specifically. (If you have too few people of a certain group, you cant conclude anything all too specific about that group's preferences with statistical reliability.)

If they deliberately oversampled blacks for such a reason, I'd guess that they weighted down their impact in the end result again to a more proportional share. (I didnt look up whether ABC/WaPo is one of the polls that "weights", but from memory I think it does.)

Yes, sites like Realclearpolitics, which illustrate the overall trend by charting a running average, can be adversely affected by such outliers. Though only for a bit -- there's enough polls out there to smooth it out quickly enough again.

But it is the reason that Pollster.com, for example, doesnt use a running average but a trendline:



No, I dont really know what that stuff in the first bullet point means either ... But from the various explanations in their blog I gather that the specific purpose of the trendlines is to smooth over the centrifugal effects that outlying polls would have on a simple average.

Let me inject a little math here.

The results from well-done polls pretty statistically valid (the voter models to determine "likely voters" are a little bit of art rather than science). The math to determine margin of error is based on real math. If you use them in comparisons, for example how was McCain doing before the debate compared to after the debate... they are very helpful (and mathematically sound).

Regression lines are mathematical hocus-pocus. They should be mostly ignored. A regression line is a valid mathematical technique when the underlying data is linear (i.e. based on a function with a steady rate of change). As voter opinions are not even close to linear, a linear regression line is, mathematically speaking, a little more than useless.

Rolling averages are more valid. They have the affect of turning a one day poll into a three day poll-- and statistically a greater number of samples means less error.

As a former physics teacher I cringe when I see mathematically misleading graphs. Drawing a linear regression on non-linear data is unforgivable.

Ok.. I looked and this is what I found

As for leaners. They asked
901. Generally speaking, do you usually think of yourself as:

then asked of those that said they did not consider themselves a dem or rep.

904. (IF NOT DEMOCRAT OR REPUBLICAN) Do you lean more towards the:

They then combined leaners with party affiliation and called it NET LEANED PARTY:

38% + (46% of of 35%) = 54% (with allowances for rounding)


Pretty simple stuff there. I don't see what your complaint is.


Second -
How do you think of yourself does NOT ask how they are registered. Your complaint is about your understanding not with the poll or its question. People can not be registered at all and consider themselves a party or they can have registered as a party 20 years ago and never changed it but think about themselves differently now then when they registered. You are mixing facts and trying to use something that was NOT asked to dispute what was asked. Nonsensical on your part.

Third
Did you NOT read what it says about the black voters?



Only 12% were Black in the sample. They only added more as an oversample to ensure their sample of blacks was large enough to report on blacks as a whole. They didn't use that in the registered voter results. They quite clearly state it in their explanation and it is shown in the later data. Maybe if you bothered to look you would have seen that.

Fourth
Are you being manipulated? Only by the voices inside your head it seems. The poll isn't doing at all what you claimed it was. If you want to demand intellectual honesty, you should start with your own lack of intellectual honesty in representing what the poll did.

Hm. OK. Well, since they only apply a linear regression trend line if they have between 4 and 7 polls, and this year pretty much any state (not to mention the national polls) had over 7 polls already many months ago, it's maybe not so relevant. I guess the two measures they specified for when there are less than 8 polls available were just stop-gap measures awaiting more polls, and those have become available a long time ago.


Hmm - well I sure as hell aint no expert, so I dont have anything to add here. But on a total sidenote -- irrelevant as this might be since this part of the methodology doesnt seem to be used in any of their current graphs anymore anyway -- I must admit to responding sceptically to this 'I am a former physics teacher and let me tell you, the system this guy is using is ****' stuff. I have this total flashback to Dr. Arzt, the high school science teacher in Lost.

I guess I just mean, Charles Franklin, the guy who made it, is a Professor of political science at the University of Wisconsin, where he teaches statistical analysis of polls, public opinion and election results. He teaches graduate level statistics courses there and at Oxford, as well as at a University Summer School in Social Science Data Analysis. He's on the Special Committee on 2008 Presidential Primary Polling of the American Association for Public Opinion Research. Well, bla bla bla et cetera. All I know is he can go into excruciating detail about statistics and methodology when asked.

None of which means that everything he ever did must be foolproof or anything. Just that, well I guess if something strikes you straight away, I dunno, he's probably given it some thought too, and made the least worst choice available for a temporary status or something. Basically, since I have close to zero education in statistics myself, I just have to cast my lot with one or the other authority when there's a simple opposition of assertions that I dont have the expertise myself to check, and I feel pretty comfortable with him. But again, that just on an aside.

I followed the thread you provided. Please note the statement:
" The added interviews (commonly referred to as an "oversample") were completed to ensure there were enough African American respondents for separate analysis; the group was not over-represented in the reported results from the full sample."
As indicated in the thread, 163 African Americans were interviewed in order to allow a separate analysis of this group's voting trend. However, in considering the poll as an entirety, this group's representation was adjusted to realistically reflect African American percentages in the the wider sample.
I don't trust polls either. Tendentious sampling is always a possibility. Hope no one makes a decision based on any single, or even several, polls