Translated from Swedish by Clare James
Click here for Swedish original
Svensk text

Contents

Poststratification
Critique of sources
Monitoring electoral participation
Repeat interviews with people polled in periods without memory displacement
Use of secret ballot
Monitoring the memory-shift trend
Joint poststratification according to previous voting and other variables
What does it mean in practice?

In electoral forecasts and public-opinion polls, interviews can be weighted –– i.e. retroactively stratified –– according to figures on how the respondents voted in previous elections, to make the sample congruent with the voting outcome of this past election. This method is known as “poststratification.” To permit poststratification, an electoral forecast contains not only the question “What party do you think is the best today?” (or the equivalent) but also “Which party did you vote for in the last parliamentary election?” The latter is posed only to respondents who were entitled to vote in the previous election.

One advantage of poststratification by party in the last election is that it yields stable time series of changes in party sympathies, even with fairly small samples of interviewees. The correlation that exists between individuals’ party preferences at various times is used to reduce sampling errors. With the close correspondence that, in Sweden, usually prevails between voting for a party in general elections and preference for the party one, two or three years after the election, poststratification permits an appreciable reduction in margins of uncertainty. One formula ¹⁾ indicates that the margin of uncertainty for a sample of 1,000 voters is reduced to roughly two-thirds of its size in a sample of every nth individual. Poststratification of other variables –– life cycle, income and social class, for example –– yields much smaller gains.

Poststratification is a statistically unassailable method. Nevertheless, the polling of party sympathies is, a priori, a subject that belongs to political science and political sociology; and the statistical methods used in polling must be assessed on the basis of these disciplines. In this assessment, poststratification is not unexceptionable. No one experienced in these sciences would venture to rest an adjustment of party-sympathy figures on the assumption that voters give an entirely correct account of their voting one, two or three years ago. Various researchers are unanimous in pointing out that people’s way of describing their past is affected by their current attitude and situation.

None of us admit with particular pleasure and willingness that we have changed our opinion. It is more attractive to come across as someone who, basically, has always been as sensible we are today. Opinion pollsters are familiar with this phenomenon. By year-end 1972 practically 100 percent of Swedes thought the U.S. should withdraw from Vietnam. Almost everyone said they had always held this opinion. In actual fact, only two-thirds had thought so two years before.

Historians are well aware of such “memory shifts.” In their criticism of sources, they never take an assessment of a controversial event written down years afterwards as an acceptable eyewitness report. Instead, they merely regard such an assessment as a sign that the event was perceived or evaluated in a particular way by the informant, or in the informant’s circles, on the day that assessment was made.

For nearly 70 years, analytical psychologists have known and studied various shifts and identified the “defense mechanisms” that underlie them. In recent years, various attitude theories –– “theories of balance,” as they are termed –– have yielded the wording of legislation and formulation of equations in a special kind of mathematics for the subject. The problem of “biographical consonance,” i.e. our propensity to –– consciously or unconsciously –– alter data about ourselves to give them a slightly better fit with our current views, is a case in point for these theories.

The incidence of psychological memory shifts concerning previous voting has made most experienced opinion pollsters skeptical about uncritical, mechanical and routine poststratification by party in the last election. In the U.S., Dr George Gallup calls the method “very tricky.” In the U.K., Dr Henry Durant rejects it entirely for election forecasts on the basis of experience with the British elections of the 1950s and ’60s. In Norway Björn Balstad has developed a method of, in his electoral forecast, using information on previous electoral behavior only if there are minor disparities between the previous election results and how the voters subsequently state that they have voted. If this difference is large, Balstad relies on the unweighted figures. Professor Jean Stoetzel in France thinks that the truth about the distribution of party sympathies, in the long run, lies somewhere between unweighted figures and figures poststratified according to previous electoral behavior. The proportion depends on the mode of the public during the election campaign and must, in his view, be estimated separately by political experts. The experts then assess whether there is any reason to believe that a party’s sympathizers are more ashamed of their previous electoral behavior than others, or are drawn so strongly to a new party that they also wish to maintain that they voted for it previously.

The same skepticism with which uncritical poststratification should be regarded ought also to apply to other uncritical use, in electoral polls, of data on party support in the previous election. This applies, for example, to the ratio estimate applied by the audience-ratings department (PUB) of the Swedish Broadcasting Corporation (Sveriges Radio) –– and the imputation and calculations carried out by Statistics Sweden (SCB) to estimate the results of hypothetical elections in periods between one election and the next.

What form, then, can critical use of information about party voting in the previous election take? Several methods have been tried by Sifo. The most rewarding have been found to be the following:

1. Controlling the respondents’ electoral participation in the official voter register.
2. Conducting repeat interviews from periods without memory shift.
3. Using a secret ballot, i.e. secret to the interviewer, for interviewees’ responses.
4. Investigating the trend of memory shift concerning party voting in the last election.

Voters’ exaggeration of their electoral participation is a well-known phenomenon. In the 1966 election, participation was low, especially among Social Democrats. But many of the stay-at-home voters stated, before the 1968 election, that they had voted in 1966. This was not counted as increased support for the Social Democratic Party (SDP) in the 1968 pre-election polls since, in the uncritical poststratification then used by Sifo, it was weighted down in the low share of SDP votes in 1966. Sifo therefore underestimated the SDP upturn between 1966 and 1968. The following table shows the poststratified and non-poststratified figures for the SDP during the 1967–68 election year. 

There is surely no serious doubt that the non-poststratified series in this case is the most reasonable: at any rate, it approaches the election result more closely.

Some errors of this type can be avoided by checking, in the official voter register, the selected respondents’ participation in previous elections. Sifo has performed checks of this kind for three elections. One essential result from these is that between a third and half of stay-at-home voters are nonrespondents in surveys involving home visits. Dropouts in surveys and stay-at-home voters are thus recruited, to some extent, from the same categories. If poststratification of previous electoral behavior is to be carried out, the category of stay-at-home voters in previous elections according to these checks should be represented by 56 percent of the number included in the official election figures. In addition, it may be mentioned that checking in the voter register yields, as a by-product, a validation of a “turnout scale.” This scale comprises interview questions, the answers to which can be compiled to predict electoral participation.

In the 1968 election, the Social Democrats achieved their best result of all time –– an absolute majority (50.1 percent of the votes) –– and thereby consolidated their position as the party of government, as they had already been for 36 years. Some non-socialists and communists were then embarrassed or afraid to admit that they themselves were or had formerly been anything but Social Democrats. The proportions who told Sifo’s interviewers that they had voted for the SDP in September 1968 were as follows: 

This distribution confirmed and reinforced Sifo’s scepticism regarding all uncritical use of information about previous electoral behavior. A growing share –– amounting, finally, to some 6 percent –– had undergone a “memory shift.”

In the poll before the 1970 election, Sifo selected its sample from the interviewees of autumn 1968. This sample was then stratified in terms of the data on electoral behavior that were collected in autumn 1968. Accordingly, we were able to utilize party-support data from the 1968 election without any risk of a major memory shift. Moreover, for the first, time, Sifo employed a secret ballot using papers with party names that were laid in a ballot box so that the interviewer never discovered how the interviewee voted. The result was highly satisfactory: in its poll immediately before the election, Sifo obtained the figure of 45.5 percent for the Social Democrats, against an outcome of 45.3 percent in the election. SCB, which uncritically used the data on party support in the previous election, gave the Social Democrats 49.2 percent.

In the pre-election poll of 1970, as we have seen, Sifo asked interviewees to report their party support by means of a ballot paper placed in a box. Accordingly, the interviewers never found out how the interviewee was actually going to vote. To study the functioning of this method in greater detail, during the spring of 1973, Sifo carried out a methodological experiment. Two equivalent samples of some 1,150 voters in the 1970 election were asked which party was the best, both openly by choosing a ballot-paper that they then showed to the interviewer and secretly, by placing a ballot-paper in an envelope and then in a box. The respondents were then asked, in a similar manner, to repeat how they had voted in 1970. The envelopes were sent to Sifo, unopened, with the questionnaire.

 The results were compared, both poststratified by region and party support in the last election and in their unweighted form. The table below summarizes the results. All the figures are percentages. The figures in brackets are unweighted. 

 From the table, we can draw several conclusions:

1. Fewer people answer the questions if the ballot is open ²⁾.
2. The open ballot yields a poll result that reproduces the previous election outcome to a higher degree than the secret ballot.
3. Fringe parties, such as the Left Party Communists and the Moderate Party, receive a lower share with the open ballot.
4. The socially most acceptable parties — the SDP, which was the party of government at the time of the experiment, and the Liberal Party, which enjoyed the strongest support in the press — are voted “the best party” by a higher proportion if the ballot is open.

We were, of course, aware that these conclusions are not necessarily valid in climates of opinion different from (“kinder” than) that prevailing in spring 1973. But the results of the secret ballot were consistently perceived as more correct. On the basis of these results, Sifo decided henceforth to use the secret ballot for all its electoral polls.

It should be pointed out that the experiment does not show an unequivocal improvement in accurate perception of the 1970 election thanks to the secret ballot. For the Liberal Party, which showed a decline between the dates of the election and the experiment, the election result was underestimated in the open ballot and proved fairly correct in the secret one. For the Center Party, which advanced during the same period, the opposite is true. The secret and open ballot yielded, for the Social Democrats, the same estimate of electoral behavior in 1970: a major overestimation. The secret ballot brought clear improvements, but the problems of input data for poststratification were by no means solved.

One crucial advantage of monthly opinion polls is that trends can be ascertained and analysed. The actual purpose of Sifo’s electoral forecast is to describe trends of party sympathies, as measured by the question “Which party do you think is the best today?” 

The question “Which party did you vote for in the last parliamentary election?” should display no trend if no memory shift takes place. Through trend analysis of this question, we can survey the actual extent of the memory shift and take it into account in a poststratification. 

We establish the trend as a centered, sliding mean. For the period between the elections of 1973 and 1976, the trends of memory shifts were as shown by the graph below.

This graph was drawn using the X-11 computer program (Table D12). The dotted line represents the 1973 election outcome and the unbroken line the respondents’ replies to the question of how they voted in 1973, in the election polls conducted at the time of the evenly balanced Parliament.

The perception of the previous election differs from the reality of the election. The Liberal Party, which showed the greatest rise according to the poll, also produces higher proportions of voters who state that they supported the party back in 1973. The Center Party appears to be slightly demoralized except during election periods. For the Social Democrats, support in the election is overestimated –– a common situation during the latter half of their long period in power. For the Left Party Communists (LPC), there is an equally consistent underestimation. We see that, in large measure, the overestimation of the Social Democratic voters’ share corresponds to the underestimation of LPC voters.

One solution to the problem of poststratification in electoral polls is now, in principle, clear. Instead of uncritically poststratifying a given outcome of the previous election, the solution is to perform a critical poststratification that uses, as input data, the trend values for the memory patterns of electoral behavior instead of the election outcome. 

Despite the elegance of this solution, Sifo was unwilling to swallow this method hook, line, and sinker. There are three reasons for this. The trend figures for a given month cannot be established with complete accuracy until we have an additional few months’ measurements. At that point in time, we have only an approximation of the trend. Our experience in econometrics, with Henderson-type graphs, shows that the final figures obtained in the series are easily overdramatized, and that these figures should therefore be treated with great caution. Using final figures as the basis for poststratification would sometimes involve inserting, rather than reducing, errors. The other reason is that political realism calls for responsiveness to sudden changes in political climate, and not only slow changes indicated by gentle trends. A political scandal, major betrayal of the voters’ confidence or sudden crisis can change perceptions of electoral behavior more rapidly than our trend graph can register. Thirdly, the method does not do justice to the underreporting of communist votes. This bias in reporting of LPC voting has a systematic component that appears to be entirely independent of variations in the political climate. Nor is it correlated with the size of the non-response rate. During the 1970s, the average underreporting in Sifo’s polls was 1.4 percent. 

Trend computations of memory shift are, however, valuable in other ways. On the one hand, they provide indications of the mood in each party. On the other, they have given Sifo a clear rule that indicates when it is reasonable to perform poststratification based on party support in a previous election:

1. Only if the memory-shift trend does not deviate substantially from the election result for a parliamentary party can conventional (uncritical) poststratification be used. We count as a substantial deviation 2 percent for parties supported by more than 20 percent of voters, and 1.5 percent for the other parliamentary parties ³⁾. By following this rule, we ensure that poststratification does not perpetuate an election outcome in the figures of the electoral forecast.
   During the election period 1973–76, ordinary poststratification was acceptable according to this rule the whole time, except for a couple of months in spring 1974 when the Social Democrats’ perception of the previous election deviated more than 2 percent from reality and a couple of months in spring 1976 when the same thing happened and the Left Party Communists also showed a substantial deviation.

What, then, shall we do if the memory-shift trend deviates substantially from the election outcome? We apply a graduated procedure, as follows.

2. Regarding the trend of between one and one and a half times the margins specified above, we give the mean of two figures: one poststratified by party and one not poststratified by party. We thus include only part of the stabilizing effect of the poststratification, so as not to risk incurring its whole distorting effect.
3. If the memory-shift trend deviates more than one and a half times the above-mentioned margins, poststratification by party support in the previous election is not usable. We can then apply poststratification solely to variables other than party support in the previous election, such as the respondents’ age, sex and home region. This involves a higher margin of uncertainty in the electoral poll. Should this situation arise, the poll’s readers will of course be informed.

These and the other decision rules for the calculation of Sifo’s electoral polling are exhaustive, i.e. they leave no scope whatsoever for assessments influenced by the personal opinions of the officials who carry out the work.

Regional election results can be derived from the election statistics, thus affording a key for poststratification in terms of both electoral behavior and region. If, on the other hand, the aim is to obtain the weighting targets for poststratification by electoral behavior, age and sex, no official figures are available since voting secrecy does not permit election results to be distributed according to age and sex.

To make tables of this kind (which are, of course, never entirely exact), there are several iterative methods available in the form of computer programs. These are based on a large sample, distributed as follows:

The figures inside the table are approximations from sample polls, while the figures in the margin (outside the double dashed line) relate to population. The programs revise the figures inside with the minimal disruption of their intercellular correlations, bringing them into line with the correct margins. The result is a matrix for poststratification.

Since 1967 Sifo has had a computer program, MITER, for this operation. ECTA, the well-known analysis program, also has an option for proportional iterative adjustments.

When Sifo carries out poststratification according to party support in the previous election, poststratification by age and sex is simultaneously derived. Age is taken from the population register; thus, no memory shift can arise here.

Of the various Swedish electoral polls, that of Statistics Sweden (SCB) has the highest level of ambition (showing what the election outcome would be), the most complicated poll design (with a built-in panel), the largest samples (many thousands of interviews each time) and, of course, the most massive resources (parliamentary grants for millions of kronor). Nonetheless, in its pre-election polls in the 1970s, Sifo achieved results closer to the election outcomes than SCB’s corresponding polls. Sifo published its figures before the elections and SCB after them. The figures and outcomes were as follows:

It is our belief that the difference is due mainly to Sifo’s more critical and sophisticated approach to the use of information on party support in the last election than SCB has shown to date. This has more than offset SCB’s superior resources in other respects.

¹⁾ Leslie Kish, Survey Sampling, John Wiley & Sons, New York, 1967, p. 90. See also p. 92, last paragraph of section 3.4C.

²⁾ It may be added that the share of responses like “don’t know” and “can’t say” on party-political issues is, according to our experience, even larger when open questioning takes place on the telephone.

³⁾ These figures are roughly double the standard deviation of the electoral forecast’s deviations from its long-term trend. The figures correspond to the rules of thumb we have given the mass media and public for assessing the poll results. For a more exact calculation method, see the publication   "Sifos väljarbarometer och dess avvikelser"  (“Sifo’s Electoral Poll and its Deviations”).