Opinion Polling: New Methods, Old Problems

1
Opinion Polling New Methods, Old Problems?

• John Curtice
• Strathclyde University

2
Key Questions from 2005

• How accurate are opinion polls now?
• How good is internet polling?
• How successful are recent methodological
  innovations?

3
Key Players and Methods

• Phone quasi-random
• ICM
• NOP
• Populus
• Phone quasi-random and f2f quota
• MORI
• Internet recruited panel
• YouGov

4
The Final Polls
5
A Persistent Bias?
6
During The Campaign
7
Three challenges

• Getting a representative sample
• Getting honest answers
• Forecasting who will actually vote

8
Declining Class Divide
9
Past Vote Weighting

• Past Vote good predictor of current preference
• But people forget systematically
• Align past vote with current preference
• Forget voting Liberal Democrat

10
Changing Memories
11
(Recall) 2001 Vote
12
Target 2001 Vote
13
The Young/Refused/DNV
14
Loyalty Rates
15
Explanations?

• Timing - YouGov 2001 data collected at
  recruitment rather than interview
• So less influenced by current preference
• But more influenced by halo effect?
• So responses still subject to error?
• Differential Lab disloyalty amongst panellists?
• More honest/considered answers?

16
The Wont Says

• May not be a random sub-sample of the population
• Spiral of silence may mean reluctance to
  declare support for an unfashionable party
• In 1992 this appeared to generate shy Tories
• Who wouldnt say how they would vote - but
  perhaps said they voted Tory in past

17
2001 Vote of Wont Says
18
How Wont Says Voted
19
Length of Interview
20
Assessing Turnout

• Ask respondents probability (out of 10) that they
  will vote
• MORI - only take certain to vote
• ICM/NOP/Populus - weight by probability of voting
  (and whether voted last time)

21
All Certain Voters - MORI
22
Average Probability of Voting
23
How well calibrated?
24
Turnout and party support
25
Meeting the challenges

• Pollsters have accepted they do not have
  representative samples
• And that they have to both impute missing data
  and forecast behaviour
• Yet they still have a (diminished?) pro-Labour
  bias

26
In 2005..

• Not clear that attempts to forecast turnout were
  in error in error
• Or that imputation of shy Labour voters was wrong
• But correcting for unrepresentative samples via
  past vote not easy
• As exemplified by internet/phone contrasts