Presentation headline here

Download Report

Transcript Presentation headline here

The Polls and
The 2015 Election
John Curtice
9 June 2015
Two Halves
 The exit poll
 The pre-polling day polls
1
The Difficulties of Exit Polling
 Very geographically clustered sample.
 No precinct level counts, so difficult to draw a sample
of districts that is known to have been representative
of the result last time.
 Some polling locations are easier to cover than
others!
 Relatively high (and potentially differential) refusal
rate.
 Most voting takes place in the early evening, giving
little time to analyse the data!
2
The Solution
 Geographical variation in change in party support
much less than variation in the level of party support.
 Thus any sample of locations more likely to produce
an accurate estimate of change in vote share than in
level.
 But need an estimate of vote share last time.
 So poll in the same places as last time, and compare
exit poll results this time with last time.
 Invite respondents to complete mock ballot paper
3
The Design in 2015
 Covered 141 polling stations. Interview random 1 in n
sample, where n function of station electorate.
 Four were instances where polling district split into
two; so 139 estimates of change since 2010.
 129 were places covered by 2010 exit poll, though
some had undergone boundary change. Only two
locations replaced.
 Six new locations added in Scotland and four in areas
of UKIP strength. In each case, attempted to select a
polling station representative of the constituency.
4
Thereafter…
 Model the 139 changes (for each party) using
whatever constituency data appears theoretically and
empirically appropriate.
 Use resulting equation to estimate change in vote
shares in each constituency – and thus vote shares.
 From these vote shares estimate probability of each
party to win each seat.
 Forecast seats for each party is sum of probabilities.
(No national vote forecast.)
5
How close were we?
350
316
331
300
239 232
250
200
150
100
58 56
50
10
8
27 23
0
Con
Lab
SNP
Forecast
6
Outcome
Lib Dem
Others
The Previous Record - 2010
350
307 307
300
255 258
250
200
150
100
59
57
50
29
28
0
Con
Lab
Forecast
7
Lib Dem
Result
Others
The Previous Record - 2005
400
356 356
350
300
250
200
209 198
150
100
53
50
62
28
30
0
Con
Lab
Forecast
8
Lib Dem
Seats
Others
Performance of Final Polls
38
40
35
34
34
31
30
25
20
13 13
15
9
10
8
5
5
4
6
6
0
Con
Lab
UKIP
Poll Average
9
Lib Dem
Green
Others
Result
Based on 10 polls whose fieldwork did not end before 5.5; Polls by Opinium; YouGov; Survation; ComRes; Populus;
Ashcroft, Ipsos MORI: BMG Research, Panelbase; ICM Research
A Longer Term Pattern?
5
4
4
Average Error in Polls
3
3
3
2
1
1
0
-1
1992
1997-1
2001
-2
-3
-5
-6
2005
-1
-1
2010
Con
2015
-2
-3
-4
10
4
-4
-5
Lab
Modal Differences
40
35
35
33
34 33
30
25
20
14
15
11
10
9
9
4.5 5
5
6.5 6
0
Con
Lab
UKIP
Phone
11
Lib Dem
Internet
Green
Others
Possible Explanations
 Late Swing
 Shy Tories
 Lazy Labour
 Question Order/Wording
 Poor samples that are inadequately weighted
(perhaps because they cannot be!)
12
A Shortage of Young Voters
40
37
35
% 18-34
30
30
30
29
26
25
24
29
22
20
15
10
5
0
YG (18-39)
Populus
Unweighted
13
ICM
Weighted
Ipsos MORI
An Excess of Middle Class Ones
70
61
60
62
61
56
54
YouGov
Populus
56
59
54
50
40
30
20
10
0
Unweighted
14
ICM
Weighted
Ipsos MORI
The Inquiry
 Sponsored by BPC and MRS
 Chaired by Prof. Patrick Sturgis, Director of NCRM
 Eight other members – 5 academics, 3 commercial
researchers
 None directly involved in polling in 2010
 Asked to report by March 2016
 First open public meeting, RSS, 19 June
 Call for evidence: www.ncrm.ac.uk/polling/
15