Exit poll analysis and prediction

Document Sample
Exit poll analysis and prediction Powered By Docstoc
					  Exit-poll analysis and prediction
                         Stephen Fisher

    Following: Curtice, John and David Firth (2008) Exit
    polling in a cold climate: the BBC-ITV experience in
        Britain 2005, J. R. Stat. Soc. A, 171, 509-539.

Presentation to the NCRM-BPC Opinion Polls Conference,
British Academy, 20th January 2010
Broad research design principles: 1
• Model the pattern of change across
  constituencies in the share of the vote since
  the last election
  – i.e. not directly estimating the results in 2010 but
    change since 2005, since more variance between
    constituencies in shares than in changes.
  – Also, assessing and allowing for different swings in
    different seats
  Broad research design principles: 2
• Estimate probabilities for each party winning
  each seat
  – i.e. allow for the random/unexplained variation
    between constituencies in the prediction
• Predicted number of seats for each party is
  the sum of the probabilities across
  constituencies
• Primary aim is to predict seat totals, not
  share of the vote
              Infrastructure
• 2010 exit poll will be a joint BBC/ITN/Sky
  project
• Fieldwork by MORI and NOP, as in 2005
• People contributing to the analysis: Jouni
  Kuha (LSE), John Curtice (Strathclyde), Clive
  Payne (Oxford), Rob Ford (Manchester)
• Debt of gratitude and computer code to David
  Firth
      Selection of constituencies
• Revisit all 107 viable locations from 2005
• Top up to 130 by sampling the kinds of
  constituencies thought to be useful and
  currently under-represented.
  – e.g. new decision to explicitly attempt to have a
    group of Lab-LD seats
• Pick the most representative polling station in
  the constituency for the new locations
Exit poll locations in 2005
              Statistical analysis
• Model the change in the share of the vote since
  the last election
• Consider data with and without interviewer
  guesses for those who refused
   – In 2005 ignoring the guesses and refusals worked best
• Consider lots of different predictor variables
   – E.g. census data, market research data, strategic
     situation, incumbency
   – Expectations of geographical variation informed by
     practice with pre-election day polls
• But keep the final model simple (N=130)
            Producing predictions
• Generate predicted probabilities for each party winning
  each seat from the statistical regression models of the data.
   – Using estimates of both explained and unexplained variance.
• Sum the probabilities for each party across constituencies
  to estimate the total number of seats for the party.
• In 2005 the exit-poll data suggested a Lab majority of 100
  under uniform change, but the method accurately
  predicted 66
   – Introduction of explained (regression) and unexplained
     (probabilistic prediction) variation both equally account for the
     difference between uniform change and final prediction.
Probabilistic prediction compared with
           the swingometer
• What would be the effect of allowing for
  unexplained variation in a swingometer
  estimate of the result?
  – Depends on the distribution of seats according to
    marginality.
  – Smoothes the relationship between predicted
    swing and predicted seats.
Need for probabilistic prediction in
              2005
   Marginal Lab-Con seats for 2010




E.g. probabilistic method would predict fewer Con seats from a 7% swing than the
swingometer because it would allow for the seats immediately either side of the
7% point to split between Con and Lab, and there are more to the left than right.
-But not much difference.
               Simulation 1:
    Stability of notional 2005 results
• Rerun 2005 notional results but adding noise
  to allow probabilistic results
  – Seats with 05 margin <1% become 50:50
  – Seats with 05 margin c.4% become 90:10
• Changes the seat totals from Con 210, Lab 348
  to Con 212, Lab 346 with LD unchanged.
  – i.e. not much change
                 Simulation 2:
                Poll projection
• ukpollingreport.co.uk average of polls:
  – Con 41, Lab 29, LD 18
  – i.e. +8, -7, -5 since 2005
• Uniform swing: Con majority of 44
• Probabilistic projection: Con majority of 48
• Very little difference
Simulation 2: Seats in the balance for
  the Tories under the simulation
                    constituency   wp05   sp05   pctmaj05   conprw   labprw
ldprw
              Dagenham & Rainham   LAB    CON       15.7     0.41     0.59    0.00
                         Erewash   LAB    CON       15.7     0.42     0.58    0.00
                   Norwich South   LAB     LD        7.4     0.43     0.57    0.00
                            Bath    LD    CON       13.6     0.43     0.00    0.57
                Leeds North East   LAB    CON       15.5     0.44     0.56    0.00
                Crewe & Nantwich   LAB    CON       15.5     0.44     0.56    0.00
        Ochil & South Perthshire   LAB    SNP        1.5     0.45     0.00    0.00
          Oxford West & Abingdon    LD    CON       13.4     0.45     0.00    0.55
                    Newport West   LAB    CON       15.3     0.47     0.53    0.00
              Warwickshire North   LAB    CON       15.3     0.47     0.53    0.00
             Hampstead & Kilburn   LAB     LD        1.1     0.47     0.21    0.32
                  Coventry South   LAB    CON       15.2     0.48     0.52    0.00
        Dorset Mid & Poole North    LD    CON       13.1     0.49     0.00    0.51
                   Argyll & Bute    LD    CON       13.0     0.50     0.00    0.50
                         Telford   LAB    CON       15.0     0.50     0.50    0.00
Berwickshire, Roxburgh & Selkirk    LD    CON       13.0     0.50     0.00    0.50
                      Winchester    LD    CON       12.7     0.53     0.00    0.47
                     Luton South   LAB    CON       14.7     0.54     0.46    0.00
               Brighton Pavilion   LAB    CON       13.1     0.57     0.26    0.00
           St. Austell & Newquay    LD    CON       12.4     0.57     0.00    0.43
Simulation 2 – Distribution of
   Predicted Probabilities
   Particular difficulties for exit poll
         prediction in 2010
• Boundary changes
  – Possible errors in both dependent and explanatory
    variables
• Expenses scandal
  – More MPs stepping down
  – Potentially more variance between constituencies
• Census data old (2001)
• High expectations from 2005!

				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:22
posted:5/13/2010
language:English
pages:16