Tuesday, April 26, 2011

"The Polls Are Bunk"

In 2007, when Obama was losing terribly to Hillary in the polls during that summer, his campaign manager David FLUFF PLOUFFE sent out  mass emails to supporters saying “the polls are bunk” and do not believe them. Hillary was seen as the one who would easily win the nomination and Obama was still playing the role of the underdog.
http://tpmelectioncentral.talkingpointsmemo.com/2007/08/obama_campaign_national_polls.php
Of course, Obama and POOF—like all dirty D’s the done dirty cheap lying Democrat D’s Do—were losing in the game, so they had to make some excuse up so that their base of pot smoking, patchouli oil smelling, college kids wouldn’t get demoralized over the fact that “the One” might get smoked.
While POOFE was being a Democrat deceiver,  I’m a believer.  The polls are not only bunk—THEY’RE BULLSHIT!
I swear people, when I see media outlets like Fox News and others citing WAPO/ABC News Polls and WSJ/NBC polls as credible polls like they did today, I just want kick my TV set in.  See the WaPo article with the poll here:
http://www.washingtonpost.com/politics/economic-anxiety-threatens-obama-in-2012-but-in-poll-he-edges-gop-rivals/2011/04/18/AFUFQN2D_story.html?hpid=z2
Barry’s in trouble but, of course, he’s certainly going to win handily against ANY of the Republican potentials.  Naturally.  Hey he’s even got a good lead over Mitt Romney, and no surprise, a double digit lead over the “unelectable” Sarah Palin.
BULLSHIT!  These polls are skewed heavily to Democrats and liberals.  And the independents that make up 30% of those polled, are they right or left leaning? Who knows!
See links covering this here:


http://hotair.com/archives/2011/04/19/obama-down-to-47-in-seriously-skewed-wapoabc-poll/
http://www.businessword.com/index.php/weblog/comments/2645
http://newsbusters.org/node/8221
Below is a copy and paste of a study that was originally in a PDF so forgive me if there are some spacing errors as it may not have transferred into Word properly.  The study was done in 2004 by AAPOR to explain why polls are faulty.   What’s especially troubling about this study is, in 2004, pollsters were just getting used to a new phenomenon in communication called the cell phone.  Pollsters are producing inaccurate results because they’re mostly calling landlines.  More people today rely on cell phones rather than land phones so that would mean the cellphone user bias might be bigger in 2012 than it was in 2008 and in 2004.
http://pewresearch.org/pubs/1761/cell-phones-and-election-polls-2010-midterm-elections
Leading up to 2012 there will be many erroneous polls, although any respectable pollster will try to correct the lie in the last days before voting so they don’t come off as complicit in some political lie.
Another interesting read is this study done just might put a smile on your face.  This team uses algorithms to conclude that there are at least SEVEN Republican candidate will be able to be Barack Hussein Obama in 2012.  Have at it:
http://members.verizon.net/~vze3fs8i/air/pres2012.html
Don’t believe BULLSHIT polls.  Keep your heads up, chins up and DON’T GIVE UP!
Mr.L
Mr.L's Tavern
www.mrltavern.blogspot.com
www.youtube.com/mrltavern

SOURCES OF VARIATION IN PUBLISHED
ELECTION POLLING: A PRIMER
By Cliff Zukin, Professor of Public Policy, Rutgers University
October 2004
The ideas and views contained in this document are those of the author.
Preface
During any campaign season, a multitude of polls about the candidates’ relative
positions appears in the broadcast and print media. AAPOR representatives are
called on repeatedly to explain the how and why of these surveys – especially why
their results in the horserace might differ one from another. Cliff Zukin, AAPOR
President Elect/Vice President, graciously agreed to prepare the following primer on
sources of variations to help us answer inquiries about these polls, and I thank him for
his excellent attention to this task. Many thanks are also extended to AAPOR
members Larry Hugick, David Moore, Betsy Martin, Michael Traugott, and Jon
Krosnick for their considerable assistance and comments in drafting this document.
The result of this work is a short primer which goes a long way in explaining election
polling to journalists and others. While I am sure some readers will differ with one or
another element of this essay -- which is not meant to represent a formal AAPOR
position -- it gives journalists and others the background to make sense of what can
be perplexing differences in a complex and evolving field.
The primer addresses the surveys conducted by and for the media – not surveys
conducted by political pollsters to assist their candidates in conducting winning
campaigns. Those campaign-sponsored polls may use many but not precisely the
same practices and methods as the published polls described here.
-- Nancy Belden, President, AAPOR
A Primer
Election polling is a special breed among public opinion surveys. It calls for more
judgments—the art rather than science of the craft—on the part of the pollster than other
types of polls. And this brings into play a host of other reasons why the estimates of well
established and well done polls may differ from one another, even when surveys are
conducted at a similar point in time.
Sources of Variation in Published Election Polling: A Primer
2
Real Sampling Error
Most seasoned political observers are familiar with the notion of sampling error in public
opinion polling. That is, because we select a sample to represent a population we are
making an estimate, of candidate preference for example, in inferring back to that
population. This margin of error, expressed as plus or minus a number of percentage
points, is the most commonly known source of variation or why polls may differ. In this
election year we often hear statements, such as Bush leads Kerry by five percent, 48% to
43%, with a sampling error of plus or minus four percentage points.
What is less commonly known is that the margin of error does not apply to the spread
between the two candidates, but to the percentage point estimates themselves. If applied
to the five point spread the four point margin of error would seem to say that Bush’s lead
might be as large as nine (5 + 4), or as little as one (5 – 4). But when correctly applied to
the percentage point estimates for the candidates Bush’s support could be between 52 and
44% (48 ± 4), and Kerry’s between 39 and 47% (43 ± 4). Thus the range between the
candidates could be from Bush having a 13 point lead (52 – 39) to Kerry having a 3 point
advantage (44 – 47). So, sampling error is generally much larger than it may seem, and is
one of the major reasons why polls may differ, even when conducted around the same
time.
Sampling and Coverage Concerns: Household Selection
Sampling is the foundation of survey research, and is based on the branch of mathematics
having to do with probability theory. In short, a probability sample is necessary for its
numbers to be legitimately generalized from the sample back to the population from
which it was drawn. This assumption is not warranted is the case of non-probability (or
convenience) samples. Any poll in which respondents were able to select themselves,
including call- in polls, and Internet or Web surveys where people volunteer to participate, are non-probability samples, and thus have no basis in science. It is probably a disservice to the public to report the results of such pseudo surveys.
Most pre-election surveys are conducted by telephone, using one of two types of
sampling frames, or sources of eligible sampling units (a unit may be a person, a
household, etc.). The most common approach in the US is what is called a RDD sample,
short for random digit dialing. In this case, samples of phone area codes and exchanges
are taken, and then random digits added to the end to create 10 digit phone numbers. The
first step ensures proper distribution of phone numbers by geography; the second step,
adding the random numbers, makes sure that even unlisted numbers are included. This is
the standard practiced by almost all public pollsters.
An alternative is called registration based sampling, or RBS. This begins with a sample
of individuals drawn from lists of registered voters, to which phone numbers are then
matched. This is less costly and more efficient, as almost all calls result in reaching a
working phone number and a registered voter, which is not true of an RDD sample. The
primary disadvantage of an RBS survey is that it misses people who have recently moved
or have unlisted telephone numbers, which may be a significant portion of the electorate.
In New Jersey, for example, an RBS sample would miss approximately 30% of those
with landline telephones, who tend to be younger, more urban and more Democratic in
their voting behavior. Also, the purging and updating of voter registration lists varies
from state to state, so the accuracy of RBS sampling will vary.
Finally, it is important to note that all telephone polls are reaching a somewhat smaller
portion of the electorate than in the past because of the growth of cell phones. Since cell
phone owners are charged for calls that come into them, these exchanges cannot be
included among those available in an RDD sample. A paper presented at the 2004
AAPOR conference estimated that perhaps six percent of households were “cell phone
only.” As one might expect, this group is made up disproportionately of younger
citizens. Another five percent of households had no cell phone or landline and would not
be able to participate in a telephone survey. There is no question that telephone polls
miss those people without landline telephones, and we are unsure as to how they may be
different from others in terms of their voting behavior. How many in these groups come
out on Election Day--and how they are dividing their votes --will be missed by all pre election telephone polls in 2004. Although weighting may be used to try to make up for this shortfall (see “Weighting” below), “cell phone only” households are a relatively new phenomenon and we do not yet fully understand the consequences of this bias.
Respondent Selection
Sampling a household is only the first stage of RDD surveys, and in itself is insufficient
to ensure a representative sample. If  interviewers simply spoke with whoever answered
the phone the resulting samples would be older and more female than the population as a
whole. To produce a representative sample, survey organizations must also go through a
respondent selection procedure among potential eligible respondents in the household.
Some organizations use a random technique, such as the “last birthday” technique, where
the interviewer asks to speak to whoever had the last birthday in the household. There
are other techniques of randomization, but the idea is to ensure that everyone has an equal
chance for inclusion. Other organizations use a systematic technique, such as asking for
the youngest male/oldest female at home, that have produced representative samples in
the past.
Some surveys, such as those using automatic or recorded interviewers (touch 1 if you
plan to vote for Kerry, 2 if voting for Bush) use no selection technique when contacting a
household, but instead try to compensate for this by weighting at the end. Like nonprobability samples, these types of surveys have little claim of scientific validity and
probably should not be reported.

Timing and Field Procedures
Timing of course refers to when a poll is done. As all pollsters are fond of saying, even
the most well done poll is a no more than a snapshot in time. Polls do not predict; they
describe the situation of the moment. Obviously, polls with different field dates may
yield different results, as voter preferences may change with time. However, a largely
invisible reason for differences is that that polling organizations have different field
procedures. Field procedures refer to the ground rules under which the interviewing is
done. And there are some tradeoffs to be made. For example, a field period of seven
days would allow for a number of callback attempts to reach the designated respondent
before allowing substitution with a new household and respondent. But campaign events
may happen in those seven days, making that poll harder to interpret. A three-day poll
may focus more narrowly on a particular point in time, but perhaps at the sacrifice of
callback attempts. Callbacks matter because one respondent may not be the same as the
next; extra field time may be necessary to reach younger voters for example, who may be
more Democratic in orientation than others. So factors like the number of callbacks, days
of interviewing, and response rates may also be reasons why polls purporting to measure
the same thing give different results.

Question Ordering and Wording
It has long been known that the ordering and wording of questions in a survey can affect
the results. That is, responses to questions asked early in the interview schedule may
affect later ones, as frames of reference are set, and respondents strain to be consistent in
their responses to interviewers. For example, a survey that asked respondents a set of
questions on the economy before asking for whom they planned to vote could lead to a
bias in favor of Kerry, who is stronger on domestic issues. And a line of questioning that
asked people about terrorism could lead to a bias in favor of Bush when the “vote”
question was finally asked.
In order to minimize this problem, most researchers will ask the horserace question (If the
election were held today, for whom would you vote…) before any other substantive
election question on the survey. This does not include neutral questions about whether
people are registered and how interested they are. After all, when people go into the
voting booth on November 2 they will have had no warm up questioning. However,
perhaps in hopes of simulating the campaign, some polling organizations begin their
surveys with substantive policy or election-related questions before asking about vote
intentions. When interpreting poll results it is always useful to know the context in which
a question was asked. While two polls may have asked the horserace question in the
same form, one may have done so after unconsciously pushing some respondents in one
direction or the other by earlier questioning. Thus, question ordering also becomes a
source of possible variation in the results among published polls.
The wording of questions -- even the horserace question -- may also vary from one poll to
the next. Some polls will ask a two-way vote intention question, naming only the major
party candidates but recording all answers, while others will explicitly add Ralph Nader’s
name, or ask about the Green party, or add a response choice of voting for an independent
candidate. Some polls ask the horserace question twice in the same survey, once with the
two major party candidates and once with a more expansive list of candidates. But one of
these must be asked before the other, and the order may influence responses. While most
polling organizations asking about the candidates add their party labels as a cue, some
may just name the candidates. And trial heat questions that also name the Vice
Presidential candidates may produce somewhat different results than when only Bush and
Kerry are mentioned. So differences in question wording also may be a reason why polls
have small differences in their reported findings.

Weighting
Weighting is an important and common practice in survey research. Even the best polls
cannot interview a perfect sample, due to non-response and non-coverage among other
reasons. (Non-response occurs when people refuse to take part in the survey; non coverage occurs when people cannot be included in the survey– for example, an Internet
survey would miss households without computers and a telephone survey would miss
those without landlines example.) Thanks to the U.S. Census, we know how many
people with fixed characteristics of age, race and education are in the population and we
should have in our samples. When we look at whom we actually interview in our
samplings, we can adjust -- or weight -- for these characteristics to make sure they are
correctly represented.
Weighting likely voters : Because of the Census, we know the characteristics of
the general adult population -- but of course, we do not know what the voting public will
look like on Election Day. We do not know how many of any racial, educational or
generational group will be going to the polls until after the fact. In the 2000 election, the
national exit poll estimated that African Americans made up about 10% of the electorate,
and they gave about 90% of these votes went to Al Gore. What if seven percent of a
polling organization’s sample is made up of African Americans in 2004? What if it is
13%? It will obviously make a difference in the horserace estimate, but we will not know
which is correct until Election Day. And, of course, the past is no guarantee of the future.
So the pollster’s dilemma is “What do we weight to?” Most pollsters of published
surveys first ask a sample of the general population about their race, age, etc., weight
their data to what a random sample of the population should look like, and then go on to
pull likely and non- likely voters out of that big (already weighted) general population
sample. Some, however, weight to a picture of what they believe turnout will be, based
on past experience and elections -- and not everyone doing so is painting the same
portrait. (This practice is more common among campaign-sponsored polls than published
polls.)
Weighting party identification: Another source of differences in pre-election
polls is whether the pollster uses party identification as a weighting factor, that is,
weighting to reflect an assumed distribution of electorate by the political party of
respondents. A party identification question, generally placed near the end of the survey,
asks people to state whether they consider themselves a Democrat, Republican,
independent or something else. Some pollsters, including this author, do not feel it is
appropriate to weight by party. They note that party is not a fixed attribute, like race or
gender or age, and that peoples’ responses to this question change based on
circumstances and events. And indeed, the American public does show fluctuation in
partisanship over time, as well as individual changes. So these pollsters do not consider
party ID as a variable for weighting.  Some others weight using an aggregate of their recent phone polls, in an effort to smooth out the ups and downs of party ID that they find in the polls with short field times. Still other pollsters take party ID into consideration either weighting by what they believe party distribution is or weighting their data using party ID estimates from exit polls. This has been a subject of on- going debate among pollsters in recent years and will no doubt continue to be.

Likely Voters
One problem pollsters face here is that not all the respondents who tell us they plan to
vote will actually do so. When respondents’ self-report of intentions in pre-election polls
have been compared to actual turnout (again, known only after the election) we have
historically found a large over-report of voting intentions. So the pollsters’ dilemma here
is to separate the wheat from the chaff: Of all those saying they will vote on Election
Day, which ones will really do it, and which ones will stay home? And, of course, people
change in their commitment to voting as the campaign unfolds. Respondents are
probably better able to tell if they really are going to vote as it gets closer to Election
Day. This means that the definition of likely voters is somewhat of a moving target,
compared to the definition of registered voters, for example.
Research finds no magic bullet question or set of questions that can reliably determine
likely voters with 100% accuracy. Thus, different organizations have different ways of
estimating who are probable voters. Most polls ask a combination of questions that cover
self-reported vote intention, measures of engagement (following the election closely,
interest) and past behavior (voted in prior elections). They then combine responses to
create an index that gives each respondent a total score. Most then use a cutoff point so
that only the candidate preferences of the “most” likely voters are used, and the choices
of others are discarded. But even while most use such a scale, the component questions
that go into the scale differ, and so this too is a source of variation among polling
organizations. There are other approaches as well. At least one national poll relies on a
single question of reported intention; some may give voters weights based on the
probability of voting to all in the sample rather than using a cutoff; some use a fixed set
of screening questions that have worked well for them in the past.
A second issue in determining likely voters is estimating how many there will be, which
may affect the division of the vote. In New Jersey, for example, the percentage of
registered voters turning out in the last three presidential elections was 82% in 1992, 72%
in 1996, and 70% in 2000. Based on this, a pollster might expect turnout to be about 70
% in 2004. But, by comparison, surveys have noted that interest in the election is higher
in 2004 than in prior years. So, what should be used as the expected turnout figure --
70% or 80%, or 85%? Suppose a choice of a cutoff point of 70 % gives an estimate that
Kerry leads Bush in the state by four percentage points. But when expanding the
expected electorate to 80 %, it may be that the data show Kerry leading by six percentage
points. So, another source of possible differences is what percentage of voters is let in
during the likely voters scoring process. Moreover, some may start with a base of
registered voters (72% in New Jersey in 2000) while others may work with a percentage
of the voting age population (52%) as their base. There are also differences in the voting
age and voting eligible populations in each of the states. Thus while all polling
organizations will release figures for who they believe are likely voters, no two
organization will define them in exactly the same way.
In Summary
There are a number of choices to be made in the course of conducting election polling
beyond sampling error. We call these “house differences” where different organizations
have different ways of doing this type of research. To look for trends it is probably safest
to compare polls done by the same organization at different times, rather than to try to
compare polls with different methodologies done at different times. Given the unique
nature of election polling, it is likely that outsiders may look at them with puzzlement and
ask, “What’s going on”? We hope this essay is helpful to our journalistic and other
colleagues in understanding some of the sources of variation in election polling. From
the inside, those of us conducting election polls see a fair amount of consistency in
findings amid the complexity of a science-based-art.

No comments: