Voter Turnout Sensitivity Analysis: Towards a More Parsimonious Combinatorial Likely Voter Model Presentation to 2015 AAPOR Conference May, 2015 © 2015 Ipsos. All rights reserved. Contains Ipsos' Confidential and Proprietary information and may not be disclosed or reproduced without the prior written consent of Ipsos. Typically ‘Likely Voter Models’ include a 5 or 6 item summated index from which “cut points” are defined, which typically correspond to the historical turnout rates for a given type of election (Presidential or Midterm). Recent work has explored the challenges of this model, specifically the ‘coarse’ nature of the index, which does not allow for as much precision as needed to approximate expected turnouts. The problem is especially acute in Primary or low-turnout Midterm elections like 2014, where only 25% - 35% of the population cast a vote. The problem has been in part addressed via the use of logistic regression models that allow for more discrimination between intervals. 2 However, the need for a simple question battery allowing for sufficient granularity at low-turnout elections persists. Our paper will undertake a sensitivity analysis for a broad range of turnout cutpoints from over 30,000 interviews conducted in the run-up to the 2014 Midterm elections, for generic congressional ballot voting intention measures. We will re-analyze the data in order to assess both question sensitivity and the full Likely Voter battery, in an attempt to create a combinatorial model that is more parsimonious. To measure performance, we will employ the Average Absolute Difference between the survey estimates and election results. 3 Ipsos Likely Voters Question Battery – 2014 Question Q1. Are you currently registered to vote, or not? Responses Yes No Don’t know / Refused Q2. Sometimes things come up and people are not able to vote. In the [last] election, did you happen to vote? Yes No Don’t know / Refused ASK IF ‘NO’ Q2A. Why not? Did not live in state/district at the time Too young to vote at the time Not registered, too busy, something came up Other (SPECIFY) Don’t know / Refused Q3. Did you happen to vote in any of these other elections? 2006 midterm congressional election 2008 presidential election 2010 midterm congressional election None of these [EXCLUSIVE RESPONSE] Q4. In November [year], the next [election type] election will be held. Using a 1-to-10 scale, where 10 means you are completely certain you will vote and 1 means you are completely certain you will NOT vote, how likely are you to vote in the upcoming [election type] election? 10 – Completely certain 9-6 5-2 1-Completely certain will not vote Don’t know Q5. How much interest do you have in following news about the upcoming [election type] election? A great deal Quite a bit Only some Very little No interest at all Don’t know / Refused Score Voter Non-voter Voter Non-voter Voter Non-voter Voter Non-voter Voter Non-voter Voter Non-voter 4 Likely Voter Index Formulation • Ipsos General Election Likely Voter Formulation • Summated index: ln(vote/not vote = Past voting(Voted 2006 + Voted 2008 + Voted 2010 + Voted 2012) + Prob.of Voting 2014 + Interest in election • Easy to implement. • Provides resolution for turnout levels from 40-70%. • Ipsos Primary Election Likely Voter Formulation • Logistic regression index: • More complicated implementation. • Relies on partisan sentiment – acceptable for primaries, less so for general. • Provides resolution for turnout levels from 10-30%. 5 • Problem: • Ipsos has no unified likely voter model approach for both primaries and general elections. • Lack of resolution when turnout is between 30% and 40% ( a la the 2014 midterm election). • Solution: • Unified likely voter approach providing resolution across entire spectrum of scenarios. 6 Likely Voter Battery Components Voter Registration Likelihood to Vote Likelihood to vote Base LV Model Interest in election Past Election Behavior 10% 20% 30% 40% 50% 60% Partisan ID intensity 70% 80% 90% 100% Percentile of Vote Eligible Public 7 Ipsos Standard – Summated Likely Voter Index – 2014 • Constructed by combining Q1+Q2+Q3+Q4+Q5 • Each variable has equivalent weight Likelihood to Vote Always votes Primary turnout range Never votes Midterm turnout range Presidential turnout range 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Percentile of Vote Eligible Public 8 Alternative One – “Kitchen Sink” Additive • Constructed by combining Q1 + Q2/3 + Q4 + Q5 + States Preference + Intensity of Partisan ID Likelihood to Vote Base LV Model Alt-1 LV Full Model 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Percentile of Vote Eligible Public 9 Alternative Two – Stated Interest Only • Constructed by combining Q4 + Q5 Likelihood to Vote Base LV Model Alt-2 LV Interest Only 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Percentile of Vote Eligible Public 10 Alternative Three – Historic Only • Constructed by combining Q1+Q2/3 Likelihood to Vote Base LV Model Alt-3 LV Historic Only 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Percentile of Vote Eligible Public 11 Alternative Four – Historic + Stated Intention • Constructed by combining Q1+Q2/3 + Q4 Likelihood to Vote Base LV Model Alt-4 LV Historic + Stated intention 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Percentile of Vote Eligible Public 12 • Accuracy = proximity to actual 2014 election results. • Measure accuracy by comparing absolute average error of findings vs. actual results. • Use standard practice of dropping respondents who say “don’t know” or “refuse” and rebase to 100%. 13 Impact on Poll Results Absolute average error 50% 6 6 3 3 4 3 2 46 46 50 50 47 51 49 47 46 Democrats 3.3 6 8 8 6 3 6 6 6 51 48 42 42 47 46 45 47 48 Republicans Other Ipsos Data Reuters / Ipsos Poll: Oct-Nov 2014 14 • Likely voter model significantly improves accuracy of election polling. • Historic election behavior is the better measure of likely voter. • However, still underreport Republican vote share in 2014. 15 • Hypothesis 2: Dropping respondents who show up as highly likely to vote but are “don’t know” or “refuse” loses actual voters. • Challenge: How to impute vote preference on voters who have no stated preference? • Potential solution: code “don’t know” or “refuse” likely voters as voting for the incumbent party for their district. 16 Thinking about the elections in 2014, if the election for U.S. Congress were held today, would you vote for the Democratic candidate or the Republican candidate in your district where you live? Democratic Republican Other Not voting DK/ Refused Party ID of Democrat 53% 36% 44% 46% 43% the office holder of Republican 47% 64% 56% 54% 57% the 113th Congress 113th Congress "US House 2012" by Mr. Matté File:113th U.S. Congress House districts. 17 Impact on Poll Results – Don’t Know Other to Incumbent Absolute average error 50% 5 5 2 3 3 2 2 46 46 49 49 46 51 46 47 46 Democrats 3.3 6 8 8 6 3 6 6 6 51 48 46 45 48 46 48 49 49 Republicans Other Ipsos Data Reuters / Ipsos Poll: Oct-Nov 2014 18 Implications 1. Interest in election news provides no significant addition to our likely voter model and can be removed. 2. A likely voter model consisting of voter registration, past voting behavior and stated intent to vote in the upcoming election is as accurate as the “full” Ipsos model. 3. All Ipsos likely voter models continue to understate Republican vote share. 4. Coding likely voters with no stated preference to support the incumbent for their district improves estimates. 19 Thank you Chris Jackson Ipsos Public Affairs [email protected] 20
© Copyright 2026 Paperzz