One problem with these polls is that they can not see the micro effects smaller parties are having.
Ruth Helm
WP106
Boundary House
Walsall Campus
Gorway Road
Walsall
WS1 3BD
01902 518879
-----Original Message-----
From: A UK-based worldwide e-mail broadcast system mailing list
[mailto:[log in to unmask]]On Behalf Of E.E.Bassett
Sent: Monday, April 25, 2005 2:26 PM
To: [log in to unmask]
Subject: QUERY: Accuracy of UK political opinion polls
With the general election due in 10 days, several voting
intention polls are published each day. (See, for example, the
BBC poll tracker on
http://news.bbc.co.uk/1/shared/vote2005/polltracker/html/polltracker.stm
which shows results for all the recent polls from 6 polling firms.
The subpages give "margin of error" figures, all of +-3% (for
polls with sample size about 1000) or +-2$ (for YouGov, with
sample size about 1800). These are consistent with the usual
casual "plus or minus 2 standard errors" calculations for simple
random samples (srs) of about those sizes. (Of course, the
firms do *not* use srs.)
But I find the apparent stability of the polling results quite
inconsistent with these margin of error figures. For example,
the last 3 ICM results have all shown Labour on 39% and
Conservatives on 33%. The other firms have a little more
variation than that, but not much. Even if the electorate is
totally static in its view, mere sampling variation would seem
likely to make the time series for the firms much more variable
than they are.
This suggests that either the levels of accuracy for individual
polls are much greater than claimed, or that pollsters are doing
something to make their figures more stable. One possibility is
that some sort of smoothing (e.g. using moving averages) is
going on. Does anybody know what sort of tweaking, if any, is
happening?
On a related point, the YouGov poll depends on internet users
only. To a sceptical statistical mind, this seems incredibly
risky. (In terms of the sort of bias involved, it's very
reminiscent of the infamous "Dewey Wins" poll in 1948, based on
telephone subscribers.) I'm sure that YouGov make use of all
the tools of stratified random sample estimation (perhaps
post-hoc) to adjust for demographics. But this can't deal with
the bias caused by restricting the sample to internet users
only, let alone the self-selection involved in applying to join
their panel. Does anybody know of any methods YouGov may have
to overcome these objections to their methodology?
As usual, I suggest you reply to me rather than flooding
allstat, and I'll try to summarise responses if there seems to
be interest in this.
Eryl Bassett
|