Public opinion polls and elections 2012 Review by Public Issue

Transcription

1 Public opinion polls and elections 2012 Review by Public Issue Yiannis Mavris a Yiorgos Symeonidis b a: Political scientist (PhD), President & CEO of Public Issue b: Statistician (MSc), Statistical modeling analyst for Public Issue INTRODUCTION The parliamentary election in May 2012 and the repeat election in June provided a challenge for Greek public opinion polls. Apart from the impressions created, the problems that arose are numerous. It is more than likely that these problems will continue and perhaps even intensify. The phenomenon of the growing and largely unforeseen rate of abstention, along with spiral of silence effect, in various forms will in the future limit the predictive power of Greek public opinion polls. Most importantly however, in conditions of generalized electoral volatility, coupled with ideological-political polarization, the ban on the publication of opinion poll results during the last two weeks prior to an election has developed into a crucial issue. Apart from the abolition of citizens democratic right to information and the blatant propaganda it nurtured, by abolishing in practice the relevant legislative provision in force, this time the ban also seriously compromised the reliability of opinion polls

2 Unlike the experience of previous electoral contests, in both May and June the strong trends that gained momentum during the final two weeks and were recorded in pre-election polls, because of the ban could not be published in time. If this problem is not immediately resolved, then serious questions will be raised about whether the democratic prerequisites actually exist for the conducting and publication of opinion polls, while at the same time it will become increasingly likely that their use will inevitably be curtailed in the private domain. The following analysis has a twofold objective. On the one hand, to reply to the (well-intentioned or malicious) criticism leveled at Public Issue regarding the June estimate and, on the other, to investigate the real errors that arose in the estimate of the trends that were taking shape among the electorate in the period between the two electoral contests. Continuing with consistency the effort that began with the political Barometer in 2004, we shall attempt a scientific evaluation of the pre-election surveys (as in any case took place with the previous elections) and, in addition, we shall make available the raw data of all the opinion polls conducted by the company during the two pre-election periods. We shall clarify exactly what Public Issue did and did not do with regard to the June election and shall set out in detail the methodology used. Furthermore, in the belief that dialogue and the exchange of arguments, not only among polling organizations but also between polling industry and the academic community, are obvious tools for the furthering of scientific discussion, we shall present the company s view concerning political weighting and endeavor to compare the application of weighting with Public Issue s own methodology, in the case of the two parliamentary elections in We would like to hope that other companies, academic researchers and whoever else may be genuinely interested in the future of Greek opinion polls will take part in this dialogue. 1. The election result of 17 June and the momentum of the pre-election period The result of the 17 June election was in no way certain from the outset. On the contrary, the outcome of the contest remained open until the last moment. The impact of the pre-election period on the election result is not always the same. There have been electoral contests, e.g. in 2004 and 2007, in which it turned out that the election result had to a great extent already been determined for some time and, as a consequence, the impact of the final run-up campaigns of the political parties did not play a substantial role (Mavris & Symeonidis 2005, 2007). In the case of the 17 June election, exactly the opposite occurred. The momentum of the most recent pre-election period proved to be the greatest ever observed historically, clearly stronger than the corresponding impetus seen in the May election. Developments during the final two weeks were rapid and continued right up to the day of voting, having a decisive impact on the result. Quite contrary to the 2

3 rhetoric that was employed after the election, in order to cover up the manipulative use of opinion polls, the final election outcome, i.e. the defeat of SYRIZA, emanated from the momentum that had been shaped during that crucial last fortnight. The total losses of SYRIZA, which held a significant lead after the 6 May election, are estimated at 4.5 percentage points (Figure 1). Any well-intentioned or non-fanatical reader will accept that the fact New Democracy (ND) eventually won the election in no way proves automatically that it was ahead and moreover constantly throughout the preceding period. In any case, this is precisely what emerges from the marked variations among the last prior to the ban opinion polls announced by all the polling companies variations, of course, which should not be overlooked or suppressed in a scientific discussion of their assessment (Table 2). Of the eleven (11) in total final pre-election opinion polls made public in the media, prior to the commencement of the ban, two (Public Issue and VPRC) put SYRIZA ahead, one (Pulse) showed SYRIZA and ND tied, one (Metron Analysis) a difference of less than one percentage point in favor of ND, three (GPO, MRB, MARC) a difference of less than two percentage points and only four (DATA RC, RASS, Alco, Kapa Research) a difference of more than two percentage points. According to the surveys conducted by Public Issue, it emerges that after the May election and as a continuation of the momentum of the poll result, a very strong and broadly diverse electoral surge appeared in favor of SYRIZA, chiefly at the expense of the Communist Party of Greece (KKE), but also of right-wing antimemorandum parties (Independent Greeks and Golden Dawn). This was the classic phenomenon of an election tide: first the flow (31.5%), then the ebb (26.9%). 1 The current was on the rise during the first part of the pre-election period, and more specifically from the day after the 6 May election up until the announcement of SYRIZA s program (1/6/12), before falling sharply, for 2½ weeks, possibly even right up to election day (Figure 1). The flip side of SYRIZA s decline in voter support was the rallying achieved by the smaller anti-memorandum party formations (especially Independent Greeks and Golden Dawn Figures 2 & 3, Table 3). The reasons for the reversal of SYRIZA s electoral momentum should be sought: a) in the unprecedented fear campaign launched against the party when it appeared that it might win the election. The coordinated scaremongering focused on the supposed consequences of such a victory for bank deposits, salaries and pensions, along with shortages of medicines and fuel, etc. The pressure exerted from both within the country and abroad, based on the euro or drachma dilemma, proved to be effective, b) in the grey communication assault launched by ND and c) in the party s program, especially those points relating to its policies on policing and immigrants, which alienated a segment of conservative anti-memorandum voters who, after the initial euphoria in May, had been switching to SYRIZA. Taking also into account the impact of the higher abstention rate (young people, out-of- 1 See the analysis by Yiannis Mavris: Elections of 17 June the right and the wrong, in KATHIMERINI on SUNDAY (24/6/12). Available online at:

4 constituency voters), which appears to have worked against it, SYRIZA lost a total of 4.5 percentage points and its electoral support finally shrank to 26.9%. 2 At the same time, also according to Public Issue s surveys, it can be seen that in the last fortnight there was a rise in ND s voter support (about 2%), which up until the last 10 days of May had remained unchanged. From 25.5%, ND increased its voter support on the last Thursday before the election to 27.5% (Figure 1). To a great extent, ND s rise during the final two weeks was also due to the fear that prevailed among the electorate and took the form of tactical voting. It is characteristic that 2 in 10 ND voters (approximately 5% of the electorate) opted for the conservatives in order to remain in the eurozone, while roughly 2.5% of the electorate to avert chaos and anarchy (8% of ND voters). It is also indicative that 10% of its voters gave as a reason for their choice of vote that they would be voting for ND to prevent SYRIZA from winning (Figure 4). However, the size of the shift from Democratic Left (DIMAR) and PASOK to ND and consequently the magnitude of ND s rise was underestimated, due quite simply to the problem of spiral of silence (that is, people who do not like to admit they support a certain party but who vote for them nonetheless), both in the preelection opinion polls and in the exit poll. A segment of voters who eventually voted for ND did not state this in the surveys. (See in detail below the ex post calculation of the bias in the pre-election opinion polls). Unfortunately, however, hidden vote in opinion polls, a common phenomenon observed internationally from as early as the 1930s, can be ascertained only retrospectively, i.e. solely on the basis of the election result. Whoever maintains that ND had a clear lead from the outset, in effect is asserting that the events of the last fortnight prior to the election did not play a substantial role in the shaping of the final result, in other words, that the preelection period did not have a significant impact on the final election result. However, such a standpoint not only clearly runs counter to common sense but also contradicts the majority estimates of most analysts, who (correctly) ascertained the unprecedented volatility of the vote, the (actual) erosion of party identification, etc. Moreover, the fluidity of the vote is also plain to see in the exit poll data, both in May and in June. It is revealing of the unprecedented electoral uncertainty that compared to the 17% of voters in the 2009 election who decided which party to vote for during the final two weeks or even on election day, the corresponding percentage in May 2012 was more than double, 41%. While in June, 14% (against 10.9% in 2009) decided on election day, 1 in 8 voters (12%) during the last week and almost 4 in 10 (37%) during the final days in the period between the two contests (Table 9). Most importantly though, the view that events during the pre-election period had no real influence masks the magnitude and decisive role of the fear campaign that took place. 2 An important indication of the impact of abstention at the expense of SYRIZA is the voting difference between in-constituency and out-of-constituency voters. Among in-constituency voters, ND received 31% and among out-of-constituency voters 25%. In contrast, SYRIZA polled 26% among inconstituency voters and 30% among out-of-constituency voters, i.e. 5 percentage points more than ND (Source: Public Issue, aggregate record of the Pre-Election Barometer June 2012, N = 5,862 individuals)

5 2. Public Issue s estimate for the elections of June and May Public Issue s final estimate, of the three published before commencement of the ban, was presented on SKAI television and KATHIMERINI newspaper on 1/6/ In this survey, Public Issue estimated SYRIZA s voter support at 31.5% and ND s at 25.5%, reflecting the trends that had been shaped among the electorate up to that moment. This was an estimate of voter support at the given moment that is a snapshot and in no way constituted a forecast of the final election result. As has been stressed on numerous occasions, these are two entirely different things 4. On election day, after polling stations had closed and before the election result was known (19:00), Public Issue presented also the final two (unpublished) pre-election surveys which it had conducted during the period when the publication ban was in force. And these revealed the way in which the preferences of the electorate had changed, as recorded in the company s survey data, after the prohibition had come into force. 5 The first secret survey (1-7/6/2012) showed a significant narrowing of the gap (to 2.5 percentage points), still with SYRIZA ahead (29.5%-27%), while the next one (11-14/6/2012), on the Thursday before the election, that the gap (0.5 percentage points) had almost been closed (28%, against 27.5% Figure 1). As subsequently became clear, this particular trend continued in the same direction, a fact that led to the final result See the explanatory note on the webpage [in Greek]. 5 For the relevant substantiation, see the video of the election program on SKAÏ (17/6/2012), available online at: Also, the company s relevant webpage: [in Greek]

6 6

7

8 TABLE 1: Comparison of pre-election opinion polls for parliamentary election of 6 May 2012 COMPANY CONDUCTED FOR PUBLICATION DATE PASOK ND KKE SYRIZA LAOS ECOLOGISTS DIMAR DEMOCRATIC ALLIANCE ANEL GD ND+PASOK PULSE PONTIKI 19/4/ MEAN ABSOLUTE ERROR VPRC EPIKAIRA 19/4/ MRB REAL FM 20/4/ MARC ETHNOS 20/4/ K- Research ΝΕΑ 20/4/ RASS EL. TYPOS 20/4/ METRON ANALYSIS CRISIS MONITOR 20/4/ ALCO NET 20/4/ PUBLIC ISSUE SKAI 20/4/ GPO MEGA 20/4/ ELECTION 6/5/

9 TABLE 2: Comparison of pre-election opinion polls for parliamentary election of 17 June 2012 COMPANY CONDUCTED FOR PUBLICATION DATE ND SYRIZA PASOK ANEL ΚΚΕ GD DIMAR RECREATE GREECE Kapa Research ΝΕΑ 1/6/ OTHER ND+ SYRIZA ND- SYRIZA PARTIES IN PARLIA MENT MEAN ABSOLUTE ERROR ALCO Newsit.gr 30/5/ RASS EL. TYPOS 1/6/ MARC ALPHA 1/6/ MRB Real.gr 30/5/ GPO MEGA 30/5/ METRON ANALYSIS ANT1 1/6/ PULSE PONTIKI 30/5/ VPRC Epikaira 30/5/ Public Issue SKAI/KATH 1/6/ DATA RC pelop.gr 31/5/ ELECTION 17/6/

10 Consequently, on the basis of the final wave, SYRIZA s estimated voter support (28%) diverged from the final result by just 1%, while ND s (27.5%) by 2.5%. These are the real figures and these are the actual divergences from the election result, i.e. the true error. This ascertainment renders unacceptable and willful moreover with the invocation of mathematics and statistics! the criticism leveled against the company, mainly in social media, regarding the failure to forecast the gap between the first and the second party, from +6 percentage points (in favor of SYRIZA) to -3 percentage points (in favor of ND)! The objective of public opinion polls is first of all to estimate voter support for parties, not the gap between them which, in any case derives precisely from the estimate. Furthermore, the gap between the two leading parties constitutes a measure of polling accuracy primarily in quintessentially two-party systems (as for example in the USA, where the aggregate of the two big political parties usually surpasses 95%) and it can certainly not be considered a suitable measure in the case of the June election, where the object of the estimate was the percentages of at least seven (7) parties. The problem regarding the strong impressions created by the last opinion poll to be published by Public Issue is due exclusively to the adverse effects of the ban in force, which were greater than at any other time. It is quite clear that if there had been no publication ban and the surveys of the final two weeks had been presented in the normal course of events, no such problem would have arisen. No impressions would have been created to that extent and reality would not have been manipulated and distorted so much, nor would responsibility have been unjustly ascribed to the company, when the problem lies solely and entirely in a drastic legislative provision. Why the pre-election surveys were published on election day Regrettably, the fact that the two secret surveys, which revealed the rapid change that had taken place, were presented on SKAI television on election day, immediately after polling stations had closed and before the election result was known 6, was unable to moderate the negative impressions; on the contrary, it probably reinforced them. This was because the presentation of the final preelection surveys was misinterpreted. Although they were pre-election surveys that reflected the situation up to and including the Thursday before election day, the percentages of the parties which they recorded were mistakenly considered to be the result of an exit poll, and consequently a forecast of the election result. The unpublished pre-election surveys were presented for scientific and ethical reasons as well as out of a sense of responsibility toward the public. It was considered (mistakenly, as it turned out) self-evident that the surveys which had been conducted during the period of the ban should be presented, albeit necessarily after the event, so as to inform public opinion how the electoral contest had taken shape. This choice was made on the basis of the reasoning that the company considered it proper to publicize any data it had. In any case, Public Issue had 6 For the purpose of substantiation, see the link in the preceding footnote

11 adopted the same practice in the 2009 election and the May election, just one month previously, i.e. in all cases where the company did not itself conduct or participate in an exit poll. Lastly, a brief reference to Public Issue s estimate for the May election is considered necessary for reasons of completeness and appraisal of the company s overall effort during the elections in In the last survey, which was presented on SKAI television and in KATHIMERINI newspaper on 20/4/2012 7, ND s voter support was estimated at 21.5%, PASOK s at 14% and SYRIZA s at 13%. It is particularly important that Public Issue was the only polling company (along with VPRC) which estimated, 15 days prior to the election, the gap between PASOK and SYRIZA to be so narrow. On Election Day (6/5/2012), precisely as in the case of the June election, Public Issue presented its final ( secret ) pre-election survey, which was conducted during the prohibition period (27/4-2/5/2012). The company s estimate, which was based on this survey, was that SYRIZA (16%) had now moved ahead of PASOK (15%), while ND (22.5%) remained firmly in first place. Despite any divergences from the final election result, the company s surveys successfully recorded the trends that developed during the pre-election period and the estimates published by Public Issue proved to be particularly accurate (see Table 1). 3. Comparison of pre-election estimate with exit poll On Sunday 17 June, at exactly 7 p.m., immediately after voting had ended, two different estimates were broadcast on television in Greece. The first, aired by the TV channels MEGA, ANT1, ALPHA and NET, was based on the joint exit poll. The second, by Public Issue, was presented on SKAI television 8 and was based on an unpublished pre-election survey (constituting a wave of the Political Barometer with an increased sample of 1,200 individuals) and not on an exit poll, which in any case the company never conducted. 7 See [in Greek] 8 See footnote no

12 TABLE 3: Voting intentions: valid responses (excluding undecideds, abstainers & refusals), Public Issue s Political Barometer, January-June 2012, unweighted data DATE N ND SYRIZA PASOK KKE DIMAR ANEL GD 1/ a/ b/ a/ b/ a/ b/ c/ Election 6/5/ a b c a b Election 17/6/ TABLE 4: Voting intentions: valid responses (excluding undecideds, abstainers & refusals), Public Issue s Political Barometer, January-June 2012, data smoothed estimates using Kalman smoothing, including elections results DATE N ND SYRIZA PASOK KKE DIMAR ANEL GD 1/ a/ b/ a/ b/ a/ b/ c/ Election 6/5/ a b c a b Election 17/6/

13 TABLE 5: Ex post calculation of bias in Public Issue s pre-election surveys, for the elections of May & June 2012 June 2012 May 2012 Party Smoothing forecast (17/6) Election result (17/6) Bias (17/6) Smoothing forecast (6/5) Election result (6/5) Bias (6/5) ND SYRIZA PASOK ANEL GD DIMAR KKE LAOS ECOLOGISTS OTHER TABLE 6: Voting intentions: valid responses (excluding undecideds, abstainers & refusals), Public Issue s Political Barometer, pre-election surveys June Adjusted by past vote DATE N ND SYRIZA Election 6/5/ a b c a b Election 17/6/ TABLE 7: Comparison of Public Issue s estimate and syndicated exit poll, May-June 2012 (7 p.m.) June May Party PI Estimate Exit Poll Election result PI Estimate Exit Poll Election result ND SYRIZA PASOK KKE ANEL DIMAR GD LAOS ECOLOGISTS

14 TABLE 8: Election Day Forecast, based on the first announced actual election results), May & June 2012 Party June (20:40, 16% of Polling Stations) PI forecast based on election results Standard error Election result May (21:10, 6% of Polling Stations) PI forecast Standard Election based on error result election results ND SYRIZA PASOK KKE ANEL DIMAR GD ECOLOGISTS LAOS RECREATE GREECE OTHER ABSTENTION Note: The percentages for Independent Greeks and DIMAR are those observed in the sample of polling stations which had been gathered up to 21:10 and do not derive from regression models, as in the case of the other parties, since neither of the two parties had stood in the 2009 election. TABLE 9: Time of voting decision, Exit Poll Data 6/2012** 5/2012* 10/2009* Today In the past 15 days*/in the past week** Some time ago*/ The week before last / When elections were called** I always vote the same No Answer Source: (*) MRB, (**) Metron Analysis, Exit Poll Parliamentary Election 17/06/2012, [in Greek]

15 Public Issue stopped using exit polls in 2007, on account of the serious drawbacks ascertained with this method and the high degree of unnecessary risk it entails. The joint exit poll of the TV channels became established in the European elections of 2009, though without the participation of PI, for reasons it had explained from the outset. So, despite the fact that in the last four electoral contests (European 2009, parliamentary 2009, parliamentary 2012 May, parliamentary 2012 June) it did not conduct exit polls, nevertheless, even today i.e. after almost four years this has not been understood, not only by a large segment of TV viewers, but also by many of the country s political figures and a significant number of journalists. On the contrary, in fact, the conviction is so strong among television viewers that in the evening of election day all figures are exit polls that it was not possible for the difference between estimate and exit poll to be perceived, nor of course that the former was based on a pre-election survey. This confusion ultimately proved to be a communications disaster. However, even though Public Issue s estimate was based on a pre-election survey, in terms of statistical accuracy it did not differ substantially from the forecast of the joint exit poll conducted by the TV channels, since it both cases it was estimated that the result would be extremely close and basically unforeseeable within the framework of the science of statistics. It should be noted at this point that the pre-election telephone survey was conducted with a sample of 1,200 individuals with a margin of error of +/-2.9%, while the exit poll had a (more than six times larger) sample of 7,709 individuals/voters 9 and for this sample size the margin of error is +/-1.2%. One of the reasons for preferring a telephone survey to an exit poll was most certainly the financial aspect: the cost of an exit poll is approximately times higher than a pre-election telephone survey. The simple comparison of the two estimates (i.e. of the final pre-election survey conducted by Public Issue and the joint exit poll of the TV stations) is presented in Table 7 and it is quite revealing. To begin with, one can see that the actual result was within the confidence interval estimated by Public Issue for all parties. In contrast, the exit poll forecast failed with respect to both PASOK and KKE, while, albeit marginally, it also failed to forecast SYRIZA s percentage. Setting aside some of the odd views expressed, the claim that Public Issue had estimated a win for SYRIZA is quite startling to say the least. Anyone with an elementary knowledge of statistics knows full well that the essential conclusion of the estimate was a tie between ND and SYRIZA, in other words the inability to predict the election winner, with the data that were available up to and including the Thursday before election day. This fact, i.e. that on the basis of the available estimate the two leading parties were neck and neck, was repeatedly stressed with emphasis during the course of the TV program on election night Visit [in Greek} 10 See video in footnote no

16 If ND s lead was so steady and consolidated during the last two weeks of the pre-election period, as some assert, why did the exit poll which was conducted ON THE VERY DAY of voting fail to forecast ND s comfortable victory, by giving a lead of just 0.5% instead of the actual 2.8%? In essence, that is, the result of the exit poll did not differ from a pre-election telephone survey. ND s lead of 0.5% in no way pointed to victory for ND, just as correspondingly, SYRIZA s lead of 0.5% in no way pointed to victory for SYRIZA. And yet, the attacks charging failure were directed solely at Public Issue. The failures of the exit poll were hardly noticed or were concealed, while in many cases it was even presented as a success. It is very disappointing to ascertain that, while in any other country of the world an estimate that was based exclusively on pre-election surveys, cost much less than the exit poll and yielded at least equally accurate results as the latter would be viewed as a success, in Greece it was considered a failure. This peculiar clash between pre-election survey and exit poll was also seen, in precisely the same circumstances, in the election of 6 May. Then too, the statistical accuracy of both estimates was absolutely comparable. Public Issue s pre-election estimate failed only in the case of ND and Independent Greeks (marginally), while the exit poll forecast failed with regard to PASOK. Despite the relative overestimation of ND, Public Issue s estimates (in its final published survey and on election day) are deemed to be very successful, taking into account also the timely recording of SYRIZA s uptrend, in contrast with the final pre-election estimates of the companies which conducted the exit poll, which considerably overestimated the aggregate percentage for two-partyism (ND-PASOK), at the same time underestimating SYRIZA s percentage (Tables 1 and 7). 4. Forecast based on election results, as opposed to exit poll On Election Day, instead of the precarious method of the exit poll, in both the May and June elections Public Issue applied a new method for forecasting the election result and abstention rate. In reality, this was the only forecast made by Public Issue on Election Day. In May, the forecast was announced at 21:00, while in June at 20:40, i.e. less than two hours after polling stations had closed. Bearing in mind the speed at which official election results are these days processed and made available, as well as the now well established official estimate of the Interior Ministry, one might reasonably ask whether it is worth not only the polling companies but also the TV stations risking their reliability so openly by making a forecast that can quite easily be presented with absolute safety just 1-2 hours later

17 Public Issue s forecast was not based on the results of public opinion surveys or an exit poll; it used exclusively the actual election results. Due to the abovedescribed confusion that was created, this innovation too unfortunately went completely unnoticed, or was considered to be an exit poll or, even worse, was regarded as a correction of the pre-election estimate! This specific forecasting method is based on the first election results, as these are announced by each polling station (rather than a pre-selected sample, as in the case of the official estimate), by applying the following methodology: The entire country is divided into 1,034 clusters, corresponding to the number of municipalities resulting from the Capodistrias local government reform in By matching the polling stations in the municipalities to which they belong, a sample is created of all the Capodistrian municipalities in the country. Knowing the results of the aforesaid municipalities in the previous elections, it is possible to forecast the results of the new election using linear regression models. The precise technique used is two-stage cluster sampling combined with stratification. The clusters are the 1,034 Capodistrian municipalities and the polling stations in each municipality. All the municipalities are divided into 3 strata on the basis of their population size. Forecasting on the basis of election results is the safest method. It yields very accurate results even when figures from only 5% of all polling stations have been announced. The forecasts made both in May and June are shown in Table 8 and proved to be a complete success The methodology used by Public Issue for vote estimate in the elections of May and June Public Issue s methodology The methodology used to estimate voter support for political parties, which Public Issue has been systematically developing since 2004, is based on time series analysis and is carried out by combining the analysis of short- and long-term voting intentions trends. 12 On the basis of this methodology, in the first election of 6 May 13, the company announced its final estimate prior to the publication ban, on Friday 20 April 11 See the presentation of the June forecast in the video of the election program aired on SKAÏ (17/6/2012), available online at: For the corresponding forecast in May, go to 12 For a detailed presentation of Public Issue s methodology for estimating the electoral support of the parties, see: and Also, Mavris & Symeonidis 2005, 2007 [in Greek]. 13 For the May election, Public Issue conducted a total of six (6) waves of the Political Barometer. The company s published pre-election surveys for the May 2012 election are the following: Political Barometer 100, 1 st extraordinary wave 2 nd half of February 2012 ( ), Political Barometer 101, March 2012 ( ) Political Barometer 102, 2 nd extraordinary wave 2 nd half of March

18 , which proved in general to be accurate: Two-partyism (ND and PASOK) would not exceed 35%, ND would just surpass 20% and there was a possibility that SYRIZA would come second (Table 1). 15 For the repeat election of 17 June, Public Issue conducted five (5) weekly waves of the Political Barometer. The first three were published before the ban (1/6/2012) 16, while the last two, which were conducted while the prohibition was in force, were made public on election day, after polling stations had closed, on SKAI television (Figure 1). From necessity, the vote estimate was this time based exclusively on shortterm trends 17, mainly due to the extraordinary nature of this particular electoral contest, but also because the repeat election was being held in the same quarter as the May election. The adjustment of bias in the June surveys was based chiefly on a statistical model (linear regression), which correlates the deviation in voting intention for the leading party from the actual election results, with the percentage of undecided voters and respondents who refused to answer the question on voting intention. To estimate the above model, data were used from all the electoral contests (parliamentary and European) which had taken place between 2004 and the May 2012 election. In addition, account was also taken of the divergences observed between the pre-election surveys of May and the actual election result. The adjustments of the raw data resulting from the above methodology were relatively minor, mainly because of the low percentage of undecided voters and nonrespondents (refusals) which was recorded in the surveys. Public Issue s estimates, as calculated prior to the election, are shown in Figures 1 and Ex post calculation of bias in the pre-election surveys and adjustment of vote estimate The real scientific and methodological question that needs to be answered does not concern the impressions created by the considerable divergence of Public Issue s last published opinion poll from the election result. The real question concerns the actual bias of the opinion polls, not only in the June election but also in ( ), Political Barometer 103, April 2012 ( ), Political Barometer 104, 3 rd extraordinary wave 2 nd half of April 2012 ( o-ektakto-kyma/ ). 14 Political Barometer 104, 3 rd extraordinary wave 2 nd half of April 2012, ( ). 15 Indeed, among 10 published opinion polls in May, it showed the lowest mean error. On the other hand, the last survey published by Public Issue in June showed the highest mean error. 16 Public Issue s published pre-election surveys for the June 2012 election: Political Barometer 106 for the June election, 1 st wave ( ), Political Barometer 107, 2 nd wave ( ), Political Barometer 108, 3 rd wave May 2012 ( ). 17 In greater detail, regarding the concept of short-term trends, see: [in Greek] Also, Norpoth & Gschwend (2003)

19 May, which was ascertained on the basis of the two election results. This part of the article will be dealing with this bias. By the term bias, we mean the systematic deviation (overestimation or underestimation) of a party s percentage in the voting intention, from its actual voter support, which is not reduced by any increase in sample size or as election day approaches. 18 With this sense, the actual bias of public opinion surveys can be accurately measured (unfortunately) only retrospectively, i.e. after each electoral contest has been held. On the basis of the retrospective calculation of the bias, it is possible to obtain an adjusted ex post vote estimate. The adjustment of the estimate of voter support is achieved by eliminating the survey bias and becomes possible after having integrated the actual election results in the time series of voting intention for each party. The ex post adjustment of the vote estimate for ND and SYRIZA, particularly during the June pre-election period, is extremely useful, not only for evaluating Public Issue s published estimates, but also and more importantly for examining the momentum of the pre-election period and the way in which the election result was finally shaped. The ex post calculation of the bias in the company s pre-election surveys for the May and June elections was performed using the Kalman smoothing technique. Kalman smoothing is a technique used in a series of tracking polls in order to distinguish the real change in a party s percentage in voting intention, between two moments in time, from the changes that are due to sampling error. This distinction is made by considering that the underlying stochastic process that generates the real change in public opinion is an autoregressive AR(1) model (Green, Gerber and De Boef 1999 Mavris and Symeonidis 2007). The two election results (of 6 May and 17 June) were integrated in the time series of voting intention for each party. The election results of the two contests were considered opinion polls with an extremely large sample size and consequently with an extremely low (almost zero) sampling error. Then, on the basis of the smoothed value of the last wave of the preelection Barometer (for the May election the last wave was completed on Wednesday, 2 May 2012 and for the June election, correspondingly, on Thursday, 14 June 2012), the projection was calculated for each party for the next three (3) days, and therefore also the forecast of the time series value for the election day (6/5/2012 and 17/6/2012 respectively), according to the corresponding estimated model (out of sample forecast). The deviation between the actual election result and the aforementioned projection that emerges from the proposed method constitutes a measure of the survey bias, for each party. The bias calculation method used is set out in detail by Fisher, Ford, Jennings, Pickup and Wlezien (2010). 18 There are many possible causes of bias in surveys including, for example, sampling errors (sampling bias), concealment of true opinion by a segment of respondents (response bias), etc

20 Table 3 presents the raw data (i.e. without any adjustment whatsoever) of both the published and unpublished surveys (during the publication prohibition periods of May-June) conducted by Public Issue, since early 2012 to the Election Day (17/6/12). Table 4 and Figure 5 show the corresponding smoothed values, following integration of the results of the two electoral contests (May, June). Lastly, Table 5 presents the bias of the pre-election surveys conducted, not only for the election of 6 May but also the election of 17 June, which derives from the above-described method for each party. The most important conclusions to be drawn with regard to the estimate of voter support, based on the ex post calculation of the bias in the pre-election surveys of June, are the following: A) The voting intention for ND was underestimated by approximately 2% (1.9%), whilst conversely, the voting intention for PASOK was overestimated by about 1% (0.8%) and for DIMAR by around 1.5% (1.6% - Table 5). This phenomenon of underestimation/overestimation was due to the hidden vote of a segment of center-left voters, mainly of DIMAR and PASOK, who in reality, without stating it, voted for ND, to remain in the euro area or to prevent SYRIZA from winning (tactical voting, see also Figure 4). It should be noted that vote concealment of this kind, i.e. opinion poll underestimation of the party that was leading in respondents perception of which party will win (election winner prediction) 19, which is not a usual phenomenon in the country s electoral history, was most probably impossible to 19 In the last wave of the Pre-Election Barometer, which was completed on 14/6/2012, 57% of respondents predicted that ND would win the election, against 35% who believed SYRIZA would win

21 detect in the pre-election surveys (nor, of course, in the exit poll), let alone be adjusted prior to the elections by means of statistical techniques. Quite possibly, part of the underestimation of ND may also have been due to the unanticipated increase in abstention (young people, out-of-constituency voters). At this point it should be noted that the available data prior to the elections (overestimation of ND in the May election, marked bipolarization and comfortable lead for ND in respondents election winner prediction), pointed more to the possibility of an overestimation of the intention to vote for ND. In the end, this assumption proved to be mistaken. B) In the June election, voting intention for Golden Dawn was also underestimated highly (1.7%, Table 5), while something similar had not been observed to such an extent in the first election in May, a fact that clearly has political importance, since it concerns the significant degree of social legitimization for the neo-nazi party. Concealment of the far right vote is a common phenomenon in many countries of Europe (e.g. France). So it was not really a surprise, especially given the more general negative climate for Golden Dawn which had prevailed in the mass media as a whole during the second pre-election period. C) In contrast, no significant bias appeared in the two major poles of the anti-memorandum vote, namely SYRIZA and Independent Greeks (0.1% - Table 5). On the basis of the ex post calculation of bias in the pre-election surveys of May, it is clear that its impact was very different to that in June. The voting intention for ND was overestimated by about 2% (2.2%), while the voting intention for PASOK was also overestimated by approximately 0.5% (0.6% - Table 5). There was a significant overestimation of voting intention for DIMAR (2%), while the percentages for the Ecologists (0.6%) and marginally SYRIZA (0.3%) were also overestimated. On the other hand, the voting intention for Independent Greeks was underestimated by around 1.5% (1.7%), for Golden Dawn by approximately 0.5% (0.6%) and the smaller parties by about 3%. It emerges from the above results that in the May election there was a significant underestimation mainly of the newlyformed parties (Independent Greeks, Recreate Greece and Democratic Alliance) and of the groupings at the extremes of the party spectrum (Golden Dawn and ANTARSYA), while in contrast, the traditional parties were overestimated. However, the most important drawback of the smoothing methodology is that it cannot determine the starting point for the appearance of the biases calculated. Given the uptrend for ND and the downtrend of PASOK and DIMAR during the last two weeks prior to the election, it is more than likely that the concealment of votes for ND appeared during this period of time and gathered pace especially during the final week, as opposed to remaining steady throughout the entire pre-election period. Consequently, it appears that because of this concealment, the shift of voters from DIMAR and PASOK to ND was underestimated in both of Public Issue s secret surveys. 20 The retrospective adjustment of the vote estimate is possible only on the assumption that the bias in the opinion polls remained constant throughout the preelection period, since the method used to calculate the bias is unable to detect any 20 We do not know if this was observed also in the corresponding raw data in the possession of the other polling companies, since the relevant data have not been disclosed

22 change in bias over time. Therefore, the estimate of voter support for ND and SYRIZA may be adjusted if, from the value calculated with Kalman smoothing (Table 4), we remove the bias in the opinion polls which we assume to be constant (Table 5). Applying the above method, voter support for ND and SYRIZA, on Friday 1/6/12, the last day on which opinion poll results could be published, is estimated ex post at 27.8% and 30.6% respectively (see Tables 4 and 5 in combination). That is, the lead of SYRIZA, instead of 6 percentage points as given prior to the election, is adjusted to approximately 3 percentage points. Using the same method, the estimated voter support for ND and SYRIZA on Thursday 14/6/12, i.e. three days before the election, is adjusted to 29.8% and 27.7% respectively. In other words, if calculated ex post and the survey bias is removed, the lead changes to 2 percentage points in favor of ND, instead of half a percentage point in favor of SYRIZA as estimated before the election. Put differently, the two main conclusions of Public Issue s published preelection estimates continue to hold even after integration of the actual result in the time series of survey data: 1) Fifteen days before the election, Public Issue had estimated that SYRIZA had a clear lead over ND. According to the ex post adjusted calculation of the estimated voter support of the two parties, it emerges that two weeks before the election SYRIZA indeed had a clear lead over ND (approximately 3 percentage points, against 6 as initially estimated), a fact that does not refute the company s pre-election estimate. 2) From the data published by the company on election day, Sunday, according to the situation indicated by its surveys up to and including Thursday, it emerged that there had been a reversal of the climate, it was estimated that the electoral contest was a tie and the outcome largely unforeseeable. The ex post estimation of the two parties voter support indeed calculates the lead of ND at about 2 percentage points. However, the relatively small size of the lead, in conjunction with the margin of error of the two estimates (± 2% for ND and ± for SYRIZA) confirm also retrospectively the company s preelection conclusion regarding the uncertainty of the outcome of the electoral contest. That is, according to the available opinion poll data three days before the election, there could have been no reliable forecast of the election outcome, even if the bias in the surveys was known from the outset. In conclusion, Public Issue s last unpublished pre-election estimate of the election result, SYRIZA 28% and ND 27.5%, overestimated SYRIZA by about 1% and underestimated ND by around 2.5%. SYRIZA s downturn, which continued after Thursday 14/6/12, the day on which the final wave of the pre-election Barometer was completed, explains the divergence in the estimate of SYRIZA s voter support. In contrast, the overestimation of ND was due to systematic bias, which could not have been detected from the outset