Rich Ulrich
2024-11-02 17:32:01 UTC
Reply
Permalinkthese experts say. Robert Tait
https://www.theguardian.com/us-news/2024/nov/02/what-polls-mean-so-far-trump-harris-election-voters
Of the last 321 polls in the battlegrounds, 124 - nearly 40% -
showed margins of a single point or less, the pair wrote.
Pennsylvania was the most “troubling” case, with 20 out of 59 polls
showing an exact tie, while another 26 showed margins of less than
1%.
This indicated “not just an astonishingly tight race, but also an
improbably tight race”, according to Clinton and Lapinski.
Large numbers of surveys would be expected to show a wider variety
of opinion, even in a close election, due to the randomness
inherent in polling. The absence of such variation suggests that
either pollsters are adjusting “weird” margins of 5% or more,
Clinton and Lapinski argued – or the following second possibility,
which they deemed more likely.
“Some of the tools pollsters are using in 2024 to address the
polling problems of 2020, such as weighting by partisanship, past
vote or other factors, may be flattening out the differences and
reducing the variation in reported poll results,” they write.
When the 'margin of error' is three points, then you expect the
observed results to VARY around the middle. Ignoring all
differences in method -- another survey, done one exactly the
same days by the same people is APT to differ by (say) three
points from the other. People with different methods and
sampling procedures would differ more.
If you flip a coin 1000 times, 500 heads is "most likely" but
if you get exactlly that number a bunch of times in a row,
someone is cheating. A range of results should center on
500, but 3% either way is EXPECTed to come up, fairly often.
The bell this rings for statisticians is the re-analyses done on
Gregor Mendel's data on Dominant/recessive genes. Mendel
wrote before modern statistics; he was a monk doing private
experiments; his work was buried in a minor publication, and
re-descovered years later. He published convincing data
showing ratios like 1-1, 3-1, 1-2-1 -- with TOO MUCH precision
for the number of plants he was growing.
Looking back, it seems that he very likely "fudged" his data in
one way or another. Reporting averages? Throwing out odd
results? His tables seem "unlikely" to have represented the
experiments as he described them.
I don't know how they selected their "321 polls" but I do know
that agencies with HIGH reputations are relatively few; and
the reputable ones do a pretty good job of "showing their
work", so I don't know what is going on here.
I can see how they may get more consistent results if they
do such things as "weighting by partisanship" -- this amounts
to a version of what is called "stratified sampling". Thus, they
are predicting changes in the overall outcome by looking at
changes in subgroups: Are suburban white college-educated
females changing their minds? - some of the surveys these
days are probably internet-based, and ask their volunteers
to give periodic responses, so there is a person-by-person
change measured, which could be precise.
What that LACKs is the randomness of selecting samples; and
where it potentially FAILs is that the generated outcome
makes iron-clad assumptions about turnout: This is our result
IF our percentages of assumed voters by category match
who votes. Unfortunately, WHO VOTEs seems to be the
big explanatory variable for WHO WINs in recent elections.
Nate Silver, of FiveThrtyEight, is the most prestigious of US
poll interpreters. Last week, he gave his statistical opinion,
that, despite all the close calls in polls, there is about a 60%
chance (IIRC) that one candidate or the other will win at
least 6 of the 7 swing states. Sounds about right to me.
Oh, he said that the data (last week) gave Trump about
a 53% chance of winning.
--
Rich Ulrich
Rich Ulrich