A New Methods of Predicting Voting Behaviour

International Journal of Market Research

A New Methods of Predicting Voting Behaviour Janet A Hoek and Philip J Gendall Massey University, Palmerston North, New Zealand

INTRODUCTION The prediction of voting behaviour interests many political and social scientists and preoccupies most political parties. However, despite several decades of research, the success of attempts to forecast election outcomes has been variable, and on some occasions political poll estimates have borne little relationship to the actual results.1 One reason for these discrepancies may be the influence of ‘undecided’ voters; that is, respondents who are unable or unwilling to state a clear preference for a particular party or candidate when questioned in a pre-election poll. The proportion of these undecided respondents frequently ranges from 20% to 45% of the sample; clearly a group of this size is likely to have a significant effect on the predictive ability of such polls. Fenwick et al (1982) summarise the problem: 'the voting behaviour of the undecided group can have a significant impact on the final results and.... pre-election knowledge of their likely voting behaviour can have important political strategy implications' (p 383). However, while researchers acknowledge the problem, they have made comparatively little progress in eliminating the undecided group or resolving the difficulties they create. This article describes a study which provides a possible solution.

REVIEW OF LITERATURE A number of researchers have examined methodological problems associated with political opinion polling (see Miller 1952; Freeman 1953; Kraut & McConahay 1973; Perry 1960; 1973; 1979; Traugott & Katosh 1979; Day & Becker 1984; Shamir1986; Bolstein 1991; Petrocik 1991). However, most investigations have concentrated on problems such as data validity and sampling. Thus researchers have sought to assess whether discrepancies in registration status and intended behaviour exist (Traugott & Katosh 1979), compared different interview methods (Perry 1979) and have analysed the estimates obtained using different sample selection procedures (Mitofsky 1981). Nevertheless, there is a small body of research devoted specifically to the issue of undecided respondents. One approach to reducing the problem of undecided respondents, developed by the Gallup Organisation, involved the use of a secret ballot. They found this considerably reduced the size 1 of 14

29 Sep 2017 13:57:10

of the undecided group, so much so that it decreased to only 25% of the proportion recorded using direct questioning (Perry 1979). However, while Gallup’s innovation reduced the size of the undecided group, it only decreased the extent of the problem this group presents. Thus researchers still had to consider whether they would attempt to eliminate the remaining undecided respondents from their analysis or whether they would attempt to allocate them to candidates or parties. Those opting to reallocate undecided respondents had then to decide what basis they would use for such an allocation. Researchers considering allocation criteria have devised attitude scales to indicate the party or candidate respondents appeared to lean towards, as well as their voting intentions. Lampert (1978), for example, described a 'Choice Pollimeter', a device similar to a slide rule, and based on the same principle as a semantic differential scale, which respondents used to indicate the extent to which they favoured, or did not favour, two election candidates. When respondents had performed this task, interviewers, looking at the reverse side of the rule, were able to read a percentage support figure. Lampert claimed the Pollimeter combined voters attitudes and intentions and allowed researchers to measure both respondents’ probability of voting and their vote distribution. However, his approach fails to account for the fact that attitudes are often poor predictors of behaviour (a problem Day & Becker, 1984, have discussed specifically in the context of polling). Nor does Lampert suggest how his device would aid forecasting when more than two major candidates or parties are represented. Even his subsequent study (Lampert & Tziner 1985), which attempts to deal with the latter problem, explicitly excluded from the analysis the 20% who classified themselves as undecided. Some researchers have not sought to quantify respondents’ attitudes, but have attempted to elicit other information using more subtle questions. Instead of asking respondents directly who they intended to vote for, they have used indirect measures to try to ascertain the direction in which undecided voters leaned (Fenwick et al 1982). The information obtained from decided voters enabled the development of a discriminant analysis model which Fenwick et al used to predict the party affiliation of undecided voters. Post-election surveys showed the model had correctly assigned 61% of this undecided group, however, Fenwick et a l suggest researchers first use leaning questions as an allocation basis before using their model to classify those respondents who had not provided ‘leaning’ information. Gallup have used a similar methodology, not to develop an allocation model, but to classify voters according to their likelihood of voting, in the belief that a sizable proportion of the undecided group are nonvoters (Perry 1973 & 1979). The Gallup turnout scale uses nine key questions related to voting intention and enables the identification of a group of ‘likely’ voters. Studies comparing estimates derived from this group with those obtained from the total sample suggest the likely voters produce more accurate estimates (Perry 1973 & 1979). Gallup’s method differs from those discussed earlier as it does not attempt to allocate undecided voters, but rather focuses on voters with clear preferences and a high voting intention. Traugott & Tucker (1984) continued to develop this method by proposing a simplified version of Gallup’s turnout scale and comparing estimates obtained from it with those generated by a logit regression model. The index model attempted to identify a likely voter group, while the regression model assigned a probability to voters which was then used to weight the final estimates. Both models produced very similar estimates, although Traugott & Tucker concluded that the index model was preferable because of its simplicity and because the regression coefficients may not apply exactly to subsequent elections. Building on this work, Petrocik (1991) suggested using characteristics associated with non –voting to discount respondents’ projected voting intention and thus identify likely nonvoters. 2 of 14

29 Sep 2017 13:57:10

While this improves the accuracy of predictions from Traugott & Tucker’s model, it continues to allocate undecided voters according to Converse’s (1966) normal vote distribution. That is, undecided voters are distributed in the same proportions as those voters who had clear party preferences. While this is an explicit attempt to deal with undecided voters, it assumes that their mean probability of voting for the various parties or candidates is the same as that of decided respondents, an assumption which Perry (1973) has already queried. In other words, there seems no logical reason to suppose that undecided voters’ preferences should resemble those of decided voters. In summary, the problem of ‘undecided’ voters is generally dealt with in one of two ways: researchers either allocate the undecided respondents in the same proportions as those who have made up their minds, or, alternatively, they allocate undecided respondents according to other criteria, such as who they feel they are leaning towards, or on the basis of their past voting behaviour. However, both approaches rest on assumptions which are at best questionable. First, they assume that undecided respondents will vote, or at least be as likely to vote as those who have made up their minds. Yet if this group is unlikely to vote, reallocating them may actually decrease the accuracy of voting estimates rather than improve them. Secondly, these approaches assume that it is possible to establish a clear preference for each voter. This ignores the possibility that, before an election, people feel some affinity for several parties or candidates – they are genuinely ‘undecided'. Allocation methods do not capture this equivocation. A further problem with the allocation methods described above is that they tend to be complex, making them less accessible to many researchers and last resorts rather than first-to-hand techniques. What is required is a method which addresses the problem of undecided voters without the problematic assumptions or methodological complexities of existing methods. The remainder of this article describes a new method of predicting voting behaviour which meets these criteria, and discusses the results of a study designed to compare its predictive ability with that of the usual method employed by New Zealand pollsters.

A NEW METHOD OF PREDICTING VOTING BEHAVIOUR The proposed method for predicting voting behaviour uses the Juster Scale, an eleven point probability scale employed successfully by marketers to predict aggregate consumer behaviour, to estimate both voter turnout and the level of support for each party or candidate.

THE JUSTER SCALE Juster developed the probability scale which bears his name in response to the poor predictive performance of buying intention scales. He observed that many respondents who stated no buying intention accounted for a large proportion of purchases, while only a proportion of those who said they intended to buy actually did so (Juster 1966). Juster surmised that verbal intentions were really disguised probability statements and suggested that these probabilities could be collected directly using a probability scale. After some experimentation with format and wording, he developed the scale shown in Table 1, which combines verbal probability descriptions and numeric probabilities. Multiplying the number of responses for each probability by that probability and dividing the result by the total number of responses gives an estimate of the mean population purchase rate.

TABLE 1: THE JUSTER SCALE 3 of 14

29 Sep 2017 13:57:10

10 9 8 7 6 5 4 3 2 1 0

Certain, practically certain Almost sure Very probable Probable Good possibility Fairly good possibility Fair possibility Some possibility Slight possibility Very slight possibility No chance, almost no chance

(99 in 100 chance) (9 in 10 chance) (8 in 10 chance) (7 in 10 chance) (6 in 10 chance) (5 in 10 chance) (4 in 10 chance) (3 in 10 chance) (2 in 10 chance) (1 in 10 chance) (1 in 100 chance)

Evidence to date suggests the Juster Scale produces more accurate estimates of consumers’ behaviour than purchase intention scales. Juster (1966) found that probability data explained twice as much of the variance in actual purchase rates as buying intentions data, and subsequent research has confirmed the Juster Scale’s superior predictive ability and crosscultural applicability (Pickering & Isherwood 1974; Day et al 1991; Gendall et al 1991). The strength of the Juster Scale lies in its recognition that virtually all self-predictions of future behaviour are conditional; they depend on what happens over the period concerned. By allowing respondents to take this uncertainty into account and express it in their purchase probability, the Juster Scale acknowledges that some individual ‘non–intenders’ will buy and that some individual ‘intenders’ will not. These individual variations are reflected in the aggregate mean purchase probability which is used as a predictor of the population purchase rate. Although the Juster Scale was developed as a means of predicting purchases of consumer durables, it has also been applied to purchases of other items, including services and fastmoving consumer goods (Gendall et al 1991), and there is no reason why it should not be applied to any future behaviour with a conditional outcome. The proposed procedure applies the Juster Scale to the prediction of voting behaviour. The method specifically involves asking respondents their probability of voting for each party or candidate and their likelihood of voting at all. This allows respondents to indicate their relative level of support for all parties or candidates, rather than requiring them to make an absolute choice. It also explicitly acknowledges that respondents may have different probabilities of voting. In addition, the method recognises that undecided respondents may not be homogeneous, but are likely to comprise at least two groups: those who are undecided because they are uninterested and unlikely to vote, and those who are likely to vote, but who are genuinely undecided about who they will vote for. Although on election day voters will either vote or not and, if they do vote, they will have to support one candidate or party, the Juster Scale provides estimates of aggregate behaviour which take into account any discrepancies between individuals’ projected and actual behaviour.

METHOD To test the effectiveness of using the Juster Scale to predict voting behaviour, we selected one electorate in which we conducted a face-to-face survey prior to the 1990 New Zealand General Election. The electorate concerned, Palmerston North, is an urban electorate which, prior to 1990, was a safe Labour seat.

SAMPLE AND PROCEDURE

4 of 14

29 Sep 2017 13:57:10

The survey sample was randomly selected from the June 1990 Palmerston North Electoral Roll, and sample members were randomly assigned to one of two groups, one of which was interviewed using the traditional polling method while the other was interviewed using the Juster Scale method. Respondents were interviewed over a two-day period two weeks prior to the General Election in October 1990. Two call backs were made to those respondents who were either not at home or who were unavailable at the time the interviewer called. Respondents who were not contacted after two callbacks, or who refused to participate in the study, were deleted from the sample and replaced by randomly selected respondents who lived in the same geographic area. In total, 645 contacts were made, resulting in 438 successful interviews, 16 refusals and 190 instances where the designated respondent was unavailable. This represents an overall response rate of 67.9%. Samples for face-to-face polls are usually generated by a random walk around random starting points. However, such clustering decreases the precision of the estimates obtained. By selecting our sample randomly from the electoral roll, we hoped to maximise the likelihood that any differences in the estimates arose from the instruments tested, not the sampling mechanism. In addition, Palmerston North has a large number of students (over 10% of the population) who remain registered in their home electorates, thus sampling the general population would have resulted in a larger proportion of ineligible respondents. Voting registration is mandatory in New Zealand and, as the roll we used had been updated only one month prior to our sample selection, we assumed it contained an accurate listing of potential voters in this electorate. However, a supplementary roll published immediately prior to the election contained 23,444 names, an additional 2250 potential voters who were not part of our original population. Some implications of this are discussed later in the paper.

INSTRUMENTS Three different questionnaires were used in the face–to–face interviews. One employed the traditional voting question: ‘If a general election had been held yesterday, which party would you have voted for?’, while the others required respondents to use the Juster Scale, to indicate their probability of voting for each of the parties represented in the Palmerston North electorate. The second and third versions of the questionnaire differed only in the order in which the competing parties were presented, to counter any order bias that may otherwise have arisen. All respondents then used the Juster Scale to indicate their probability of voting in the forthcoming election. The two main questionnaire formats are reproduced in the appendix at the foot of this document.

ANALYSIS In order to analyse the Juster Scale data, it was first necessary to adjust the probabilities given by respondents. Logically, the sum of the party probabilities allocated should have equalled each respondent’s overall probability of voting. However, in many cases this did not occur, thus each respondent’s party probability was recalculated using the following formula: (Allocated party probability/sum of party probabilities)* probability of voting. The two sets of data were also weighted so that their age–sex distribution corresponded with that of the Palmerston North electorate. 5 of 14

29 Sep 2017 13:57:10

RESULTS The results of the study are reported in two sections. The first examines the probability that respondents would cast a vote arid explores the extent to which this varied among decided, undecided and declared nonvoters. The second section examines the voting predictions based on the traditional voting intention approach and the proposed Juster Scale method and compares these predictions with the actual election outcome.

VOTER TURNOUT The projected voter turnout, estimated by calculating the mean probability of voting, was 82.7% which compares favourably with the electorate turnout of 84.6%. Just over 20% of the voters asked the traditional voting intention question were undecided about who they would vote for. Their probability of voting was lower than that of the ‘decideds’, but not significantly so, shown in Table 2, thus reinforcing the importance of including this group when predicting voting behaviour.

TABLE 2: PROBABILITY OF VOTING Voting intention category Decided voters Undecided voters Non voters Total

Number in sample 154 49 15 218

Percentage of sample 70.8 22.4 6.9 100.0

Mean voting probability 0.93 0.86 0.29 0.87

7% of the sample said they would not vote, but their mean probability of voting was nearly 30%. This demonstrates the difference between a stated intention and the probability of performing that behaviour. However, the small size of this group and their low probability of voting suggests their effect on the actual election outcome would have been minimal.

PREDICTING VOTING BEHAVIOUR In this study, the collection of voting probabilities allowed us to calculate two sets of estimates. The first set used the traditional method of allocating undecided respondents in the same proportions as those who expressed a preference, while the second took these estimates and then weighted them by respondents’ likelihood of casting a vote. These weighted voting estimates and the estimates derived from the traditional approach to predicting voting behaviour are compared in Table 3.

TABLE 3: COMPARISON OF TRADITIONAL AND PROBABILITY WEIGHTED ESTIMATES Party

Green Labour National 6 of 14

Predicted voting behaviour Traditional Survey Result1 Probability of voting method2 % 10.1 21.3 36.1

% 0.87 0.93 0.97

n = 218 % 13.0 27.5 46.5

Probability weighted method3 12.1 27.5 48.3 3.92 29 Sep 2017 13:57:10

Other Undecided Non voters

3.3 22.4 6.8 100.0

0.86 0.86 0.29 0.87

– 8.84 – 100.0

– 8.2 – 100.0

NOTES

1. The question asked was 'if an election had been held yesterday, which party would you have voted for?' 2. Undecided voters allocated proportionally to other categories. 3. Traditional estimates weighted by probability of voting. 4. This figure was calculated by weighting the survey result by the probability of voting and adding to this a relative proportion of the undecided groups. As no party had a 100% probability of voting among its supporters, the difference between the estimated probability for each party and 1 was reallocated to the non voters. The traditional and probability weighted estimates were both obtained after the undecided respondents had been allocated to parties in the same proportions as the decided voters. Because ‘decided’ voters had high probabilities of voting, these two sets of estimates consequently do not vary markedly. The new method of predicting voting behaviour which we propose does not result in a group of undecided respondents and produces estimates weighted by the likelihood of voting as a matter of course. These estimates, obtained using the Juster Scale, are compared with those derived from the traditional voting intentions question and with the actual election outcome in Table 4.

TABLE 4: PREDICTED AND ACTUAL RESULTS Party

Green Labour National Other4 Total

Traditional1 (n = 218) % 13.0 27.5 46.5 13.0 100.0

Actual

weighted2

Juster3 (n = 220)

(n = 218) % 12.1 27.5 48.3 12.1 100.0

% 12.8 32.5 41.8 12.9 100.0

% 7.4 41.8 40.0 10.8 100.0

Probability

NOTES

1. 2. 3. 4.

Undecided respondents allocated proportionately to other categories. Traditional estimates weighted by probability of voting. Probability of voting for each party weighted by overall probability of voting. Other includes non voters as well as minor parties.

The Juster Scale estimate for Labour was closest to the actual election outcome, but was not substantially different from the estimates produced by the other methods. However, this method provided a more accurate estimate of the National Party’s share of the vote and was clearly a better predictor in this instance than either of the other methods. The Juster Scale method allocated more votes to the minority parties than they actually obtained; but this was characteristic of all three methods.

7 of 14

29 Sep 2017 13:57:10

characteristic of all three methods. In order to assess the overall accuracy of the different methods, we calculated the difference between the actual outcome and the outcome predicted by each method; Table 5 contains the results of these calculations.

TABLE 5: RELATIVE ACCURACY OF ESTIMATES Party

Green Labour National Other Absolute average Difference Mean square Difference

Difference between predicted and actual outcome Traditional method Probability Juster method weighted method +5.6 +4.7 +5.4 –14.3 –14.3 –9.3 +6.5 +8.3 +1.8 –2.2 +1.3 –2.1 7.2 7.2 4.7 8.4

8.6

5.8

On the basis of both the absolute average difference and the mean square difference between the predicted and actual outcome, the Juster Scale method produced more accurate voting estimates than either of the methods based on the traditional intention question. For all practical purposes, there was no difference in the relative accuracy of these latter methods. Attempts were made to interview ‘undecided’ respondents after the election to determine if and how they had voted. However, 25 of the 49 respondents in this group either refused to be interviewed or could not be contacted, so the conclusions drawn from this exercise can only be tentative. Nevertheless, 63% of the undecided respondents had voted for the Labour Party, 25% for National, 4% for the Green Party and 8% for other parties. This supports the argument that proportional allocation of undecided voters may actually bias rather than improve the accuracy of the final estimates.

DISCUSSION The projected turnout figure, 87.2%, was very similar to the actual voter turnout of 84.6%. However, instead of involving a number of questions in order to develop a turnout scale, the Juster Scale method provides a simpler method of assessing voter turnout that appears just as accurate as the methods pollsters currently employ (Perry 1979; Mitofsky 1981). Furthermore, our results suggest that to ignore undecided voters and base estimates only on the preferences of likely voters may introduce bias, since undecided voters have a high probability of voting and show a markedly different vote distribution to decided voters (see Perry 1973; Converse 1966). Overall, the Juster Scale method was a more accurate predictor of voting behaviour than the two alternative methods based on the traditional voting intention question. However, none of the three methods tested correctly predicted the outcome of the election. The Palmerston North electorate was won by Labour against a general trend which saw a landslide victory for the National Party. This contradiction may be partly explained by the fact that about 10% of the voters in the electorate (2250 people) registered late and consequently were not included on the Electoral 8 of 14

29 Sep 2017 13:57:10

Roll from which we selected our sample. Although no formal studies of the allegiances of these people were conducted, political scientists were of the view that they were mainly Labour supporters who registered late in a last minute attempt to stem the tide of waning support for their party. If this is so, it would help to explain why all of our predictors underestimated the level of support for Labour. In retrospect, it may have been better to have selected our samples from the whole eligible population, rather than just from the latest Electoral Roll. As well as having a more straightforward methodology than the models present by Lampert & Tziner (1978); Perry (1979); Fenwick et al (1982) and Traugott & Tucker (1984), the Juster Scale method also eliminates the undecided group from polling results. Its theoretical basis appears more rigorous and logical than the attitude and intention models discussed earlier as several researchers (Juster 1966; Pickering & Isherwood 1974; Day et al , 1991) have demonstrated the superior predictive ability of probability scales over intention scales. However, while the method we have proposed appears more accurate and less complex than other methods, its accuracy could be improved, and further research is needed to compare its performance with that of other methods which take a more sophisticated approach to dealing with the undecided voter problem than proportional allocation. In addition, while the Juster Scale is easily administered face-to-face, it is more difficult to administer by phone, especially in surveys employing random digit dialling where respondents cannot be mailed a copy of the scale before the interview. Although a substantial proportion of polling studies are still conducted face-to-face, the speed and cost efficiency of CATI surveys suggest that future research could explore methods of using the Juster Scale method in telephone interviews.

CONCLUSIONS Undecided voters cannot be ignored by researchers attempting to predict voting behaviour through pre–election opinion polls. Furthermore, the common practice of proportional allocation of this group on the basis of preferences of respondents who have made up their minds is not sustainable theoretically or empirically. The method of predicting voting behaviour which we have proposed provides a practical and effective solution to this problem. It also addresses the fact that, despite their avowed intention, many professed non voters actually have a probability of voting which is greater than zero. Our research suggests that the Juster Scale can accurately predict voter turnout and that allowing voters to express support for more than one party provides more accurate estimates of their subsequent voting behaviour than forcing them to make an unequivocal choice of one party. However, our research has also identified a problem, the over-estimation of support for minor parties, which requires further study, as does the administration of the Juster Scale over the telephone. Despite these unresolved issues, this article has outlined a new and promising polling methodology which we believe merits further scrutiny and investigation. ENDNOTE

1 In the most recent example, the 1992 UK General Election, British polls consistently predicted either victory for the Labour party or a 'hung' parliament. In fact, the Conservative party won the election with a small majority. This discrepancy between poll estimates and the election results prompted the Market Research Society to institute an enquiry into the discrepancy. ACKNOWLEDGEMENT

9 of 14

29 Sep 2017 13:57:10

The authors wish to thank the Massey University Research Fund for providing a grant-in-aid which supported this project, and Don Esslemont for his perceptive comments on an earlier draft of this paper.

APPENDIX ID______________ OUTCOME______________ VERSION 1 MASSEY UNIVERSITY DEPARTMENT OF MARKETING VOTING BEHAVIOUR QUESTIONNAIRE AUGUST 1990 Hello, may I speak to----------------please. I am a student from Massey University and am conducting a research project to estimate the election results as part of my studies. Would you mind if I asked you a few questions? It will only take about 2 minutes of your time. 1. Which party did you vote for in the last general election? DO NOT PROMPT

Democrat Green Party Labour National New Labour Social Credit Don't know None (didn't vote) Other Wasn't eligible to vote

CODE IN COLUMN A 5–6 COLUMN A Last election 1 2 3 4 5 6 7 8 9 10

7 COLUMN B Yesterday 1 2 3 4 5 6 7 8 9 10

2. If a General Election had been held yesterday, which party would you have voted for? DO NOT PROMPT

CODE IN COLUMN B

3. Now please look at this card and tell me how likely it is that you will vote in this year's election? PRESENT SHOWCARD Certain, practically certain Almost sure Very probable Probable Good possibility Fairly good possibility Fair possibility Some possibility 10 of 14

(99/100) (9/10) (8/10) (7/10) (6/10) (5/10) (4/10) (3/10)

10 9 8 7 6 5 4 3 29 Sep 2017 13:57:10

Slight possibility Very slight possibility No chance, almost no chance

(2/10) (1/10) (1/100)

2 1 0

4. Just to ensure that my sample is representative, can you please tell me which year you were born? 19-5. Please record the sex of the respondent Male 1 Female 2 Thank you very much for your help. VERSION____1___ SURVEY NO.____3___

-----------------------------------------------------------------------------------------------ID______________ OUTCOME______________ VERSION 1 MASSEY UNIVERSITY DEPARTMENT OF MARKETING VOTING BEHAVIOUR QUESTIONNAIRE AUGUST 1990 Hello, may I speak to----------------please. I am a student from Massey University and am conducting a research project to estimate the election results as part of my studies. Would you mind if I asked you a few questions? It will only take about 2 minutes of your time. 1. Which party did you vote for in the last general election? DO NOT PROMPT Democrat Green Party Labour National New labour

1 2 3 4 5

Social Credit Don't know None (didn't vote) Other Wasn't eligible to vote

6 7 8 9 10

2. Please look at this card. If a General Election had been held yesterday, how likely is it that you would have voted for the Social Credit Party? CIRCLE APPROPRIATE NUMBER IN FIRST COLUMN

Certain, practically 11 of 14

7–8 SocCre (99/100) 10

9–10 11–12 13–14 15–16 17–18 19–20 21–22 NewL Nat Lab Green Demo Other Vote 10 10 10 10 10 10 10 29 Sep 2017 13:57:10

certain Almost sure (9/10) Very probable (8/10) Probable (7/10) Good possibility (6/10) Fairly good (5/10) possibility Fair possibility (4/10) Some possibility(3/10) Slight possibility (2/10) Very slight (1/10) possibility No chance, (1/100) almost no chance

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

9 8 7 6 5

4 3 2 1

4 3 2 1

4 3 2 1

4 3 2 1

4 3 2 1

4 3 2 1

4 3 2 1

4 3 2 1

0

0

0

0

0

0

0

0

Repeat question for: New Labour party National Party Labour Party Green Party Democrat Party, and A Party I haven't mentioned. 3. Still using this card, please tell me how likely it is that you will vote in this year's election? CIRCLE APPROPRIATE NUMBER IN THE LAST COLUMN 4. Just to ensure that my sample is representative, can you please tell me which year you were born? 19-5. Please record the sex of the respondent Male 1 Female 2 Thank you very much for your help. VERSION_____3__ SURVEY NO.___3____

REFERENCES

1. HOLSTEIN, R (1991). Comparison of the likelihood to vote among pre-election poll respondents and non respondents. Public Opinion Quarterly, 55, 648–650. 2. CONVERSE, P (1966). The concept of a normal vote. In Elections and the Political Order. Campbell, A (Ed), New York: Wiley. 3. DAY, D; GAN, B; GENDALL, P & ESSLEMONT, D (1991). Predicting purchase behaviour. Marketing Bulletin, 2, 18–30. 12 of 14

29 Sep 2017 13:57:10

4. DAY, R & BECKER, K (1984). Pre-election polling in the 1982 Illinois gubernatorial contest. Public Opinion Quarterly , 48, 606—614. 5. FENWICK, I; WISEMAN, F; BECKER, J & HERMAN, J (1982). Classifying undecided voters in pre-election polls. Public Opinion Quarterly , 46, 383—391. 6. FREEMAN, H (1953). A note on the prediction of who votes. Public Opinion Quarterly , 17, 288–292. 7. GENDALL, P; ESSLEMONT, II & DAY, D (1991). A comparison of two versions of the Juster Scale using self-completion questionnaires. Journal of the Market Research Society , 33, 257–263. 8. JUSTER, F (1966). Consumer Buying Intentions and Purchase Probability. Occasional Paper 99, National Bureau of Economic Research , Columbia University Press. 9. KRAUT, H & MCCONAHAY, J (1973). How being interviewed affects voting: an experiment. Public Opinion Quarterly , 37, 398–406. 10. LAMPERT, S (1978). A new approach to pre-election polling. Public Opinion Quarterly , 42, 259–264. 11. LAMPERT, S & TZINER, A (1985). A predictive study of voting behaviour using Lampert’s Pollimeter. Social Behaviour and Personality , 13, 1–9. 12. MILLER, M (1952). The Waukegan Study of voter turnout prediction. Public Opinion Quarterly , 16, 381–398, 13. MITOFsKY, W (1981). The 1980 pre-election polls: a review of disparate methods and results. American Statistical Association Proceedings of the Section on Survey Research, 47-52. 14. PERRY, P (1960). Election survey procedures of the Gallup Poll. Public Opinion Quarterly , 24, 581–542. 15. PERRY, P (1973). Comparison of the voting preferences of likely voters and likely nonvoters. Public Opinion Quarterly , 37, 99–109. 16. PERRY, P (1979). Certain problems in election survey methodology. Public Opinion Quarterly 43, 312–325. 17. PETROCIK, J (1988). An algorithm for estimating turnout as a guide to predicting elections. Public Opinion Quarterly , 55, 643–647. 18. PICKERING, J & ISHERWOOD, B (1974). Purchase probabilities and consumer durable buying behaviour. Journal of the Market Research Society 16, 203 – 226. 19. SHAMIR, J (1986). Pre-election polls in Israel: structural constraints on accuracy. Public Opinion Quarterly , 50, 62–75. 20. TRAUGOTT, M & KATOSH, J (1979). Response validity in surveys of voting behaviour. Public Opinion Quarterly , 43, 359–378. 21. TRAUCOTT, M & TUCKER, C (1984). Strategies for predicting whether a citizen will vote and estimation of electoral outcomes. Public Opinion Quarterly , 48, 330–343.

© Copyright Warc 2017 Warc Ltd. 85 Newman Street, London, United Kingdom, W1T 3EX Tel: +44 (0)20 7467 8100, Fax: +(0)20 7467 8101 All rights reserved including database rights. This electronic file is for the personal use of authorised users based at the subscribing company’s office location. It may not be reproduced, posted on intranets, extranets or the internet, 13 of 14

29 Sep 2017 13:57:10

e-mailed, archived or shared electronically either within the purchaser’s organisation or externally without express written permission from Warc.

14 of 14

29 Sep 2017 13:57:10