How Big a Problem? How Low Response Rates Affect Models of Electoral Turnout

• In Event: Drawing Better Inferences from Biased Surveys

Abstract

Response rates to pre- and post-election sample surveys have been declining almost everywhere. Research indicates that low response rates are not necessarily fatal for modelling vote choice between parties, but there is far less confidence about modelling of turnout. Most studies on the quality of turnout data address the problem of misreporting. But even if samples are validated and the official data is robust, survey non-response correlates with not casting a vote, and increasing non-response makes the problem worse. Non-voter subsamples have even lower response rates than those of voters. Weighting is one way to correct the distribution between the two groups, but will not solve problems of bias if responding non-voters are different from nonrespondents on important variables of interest, as is likely the case.

In some countries, big data is available including vote or non-vote and this makes it possible to model turnout without issues of survey non-response. The New Zealand electoral roll includes age and occupation at the individual level, and ethnicity and gender can be coded from a combination of roll data and names. Census data can also be merged in at low levels of aggregation. However as vote/non-vote data can only be collected manually, a ‘bigger’ data sample of 30,000 was collected at the 2014 election, and followed through into the 2017 and 2020 elections. In this paper we compare baseline turnout models from the New Zealand Election Study (NZES) sample survey datasets of those elections to those using the ‘bigger’ data and explore ways in which the survey analysis may be improved using that information.

Authors

• Jack Vowles, Victoria University of Wellington

• Matthew Gibbons, Victoria University of Wellington