![Sample our Politics & International Relations journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5947/TOC-Subject-Sample-2021-Politics-International-Relations.jpg"})
ABSTRACT
This article describes a statistical method for aggregating the information from Irish opinion polls. Such aggregate estimates provide academic researchers with a time series of support for political parties, and inform the public better about opinion polls by focusing on trends and uncertainty in these estimates. The article discusses the challenge of aggregating opinion polls in a multi-party setting with a comparatively limited number of polls available and presents daily estimates of party support for the 1987–2016 period. The article develops a method to better model major sudden political and societal events, which have been common in Ireland since 2007. Finally, it discusses how polling aggregation estimates can enhance opinion poll reporting in the media.
KEYWORDS:
Acknowledgments
An earlier draft of this paper has been presented at the Annual Conference of the Political Studies Association of Ireland (PSAI), Cork, 16–18 October 2015. I am thankful for the research assistance in collecting the polling data that has been provided by David Beatty and Stefan Müller. I thank Stefan Müller for his comments on an earlier draft of this paper.
Disclosure statement
No potential conflict of interest was reported by the author.
ORCiD
T. Louwerse http://orcid.org/0000-0003-4131-2724
Notes
1 Pollster Red C makes use of Random Digit Dialling with a 50–50 split between mobile phones and landlines.
2 Ideally, this parameter would be allowed to vary between pollsters. In practice, however, if this is allowed, the model tends to estimate a tiny design effect for the most popular pollster, resulting in an estimate that very closely follows the polls of that particular pollster. Pickup et al. (Citation2011) report similar issues. Therefore, I estimate a single industry-level design effect.
3 In the mathematical notation, I will specify the variance of the normal distributions. The JAGS script in the appendix uses the precision, which is the inverse variance.
4 As Ipsos MRBI changed its adjustment method during the 1997–2002 and the 2007–2010 term, we split these terms into multiple parts (three and two, respectively) (Marsh and McElroy, Citation2003: 163). For each part, we estimate the house effects separately (each subject to the zero-sum constraint discussed below).
5 Moreover, the implementation of the model is more difficult as the software used to estimate this model, JAGS, does not allow for missing values in the vector of counts.
6 I also experimented with non-logged ratios, but this requires the implementation of a truncated normal distribution in the random walk model, which means that the algorithm runs less quickly. The results, however, were rather similar to the log-ratio approach.
7 An alternative approach is to set to the prior election result and assume that differences between that election and subsequent polls are due to house effects (Fisher et al., Citation2011). This strategy has the advantage that it allows us to correct for ‘industry bias’, that is under- or overestimation of a party by all pollsters. The problem is, however, that relatively few polls are done after an election. In the Irish case this is particularly true, with sometimes a few months passing between elections and the first post-electoral poll. One solution would be to fix the parameter
, where n is the last day observed, to the subsequent election result (Jackman, Citation2005). This option is, of course, only feasible for historical data, not the current term. Moreover, one has to assume that the day-to-day change is similar in the final stages of the election campaign compared to the entire period; this assumption might not necessarily hold.
8 Strictly speaking, the house effect prior is over the unstandardized house effect. Also note that the prior for the variance of the random walk is over rather than
.
9 If a party's support was not included in the breakdown of only a few polls, for these polls the party support (as well as support for ‘others’) was defined as missing. One advantage of the aggregation model above, estimated using Markov Chain Monte Carlo methods is that it allows for flexibility regarding missing values.
10 Omitting parties that did not participate in these elections, but were polled on in the previous term (WP 1992–1997 and PD 2007–2011).
11 If we limit the MSE to the three traditional large parties in Irish politics, the mean squared error was highest in 2002 (8.19), just before 2016 (7.12) and 2007 (6.98). The lowest mean squared errors for those three parties occurred in 2011 (1.52).
12 The analysis of Marsh and Mikhaylov starts in September 2005; therefore, for the period between September 2005 and 2007 I use the estimates from the regular polling indicator for that time period.
13 This is also true () when we limit the model to one observation per month resulting in 66 monthly observations.
14 I also replicated the Marsh & Mikhaylov model using the original Irish Polling Indicator series. Findings are similar except for Withdrawal from the bond market, which loses significance. This seems to support the case for taking into account major events when modelling opinion polls.
15 Note that confidence intervals associated with frequentist statistics, such as error margins associated with individual opinion polls, cannot generally be interpreted this way. Frequentist 95 per cent confidence intervals should contain the population mean in 95 per cent of times across an infinite number of replications.