what are the statistical rules to guarantee the reliability of surveys?

Voting opinion polls are multiplying in the press and are beginning to punctuate the legislative campaign.

For these legislative elections of June 30 and July 7, as for the other elections, the more the days pass, the more observers will have their eyes glued to the voting opinion polls. They rely on a mathematical law, the law of large numbers, which states that an average observed in a population sample has an increasing probability of getting closer to the real value in the entire population, as the sample grows.

If the sample is well representative of the population, that it is sufficiently large in number, in politics we often have samples of 1000 to 2000 people, if the wording of the questions is not biased, then a survey has therefore every chance of reflecting the state of opinion at a given moment. In France, the publication of political polls is regulated by the law of July 1977. The polling commission controls the details of the methodology used, which is calls the notices, and the institutes a legal obligation to publish their margin of error.

This margin of error essentially depends on the size of the sample, the larger it is, the lower the margin of error, but also on the score measured. The closer it is to 50%, the greater the margin of error. On the occasion of this campaign for the legislative elections, the polling commission recalls that seat projections are not strictly speaking polls and that they introduce numerous interpretation biases. To represent the future National Assembly, remaining within the logic of the poll, it would actually be necessary to carry out 577 polls, or one per constituency, each time with a sample of the local population. As this is not possible, seat projections are made on the basis of a national survey, the results of which are transposed into seats. This method cannot take detailed account of local issues and the personality of the candidates.

Technologically speaking, the use of the Internet makes it possible to obtain a representative sample and responses to a questionnaire more quickly than by telephone or face-to-face, therefore the statistical photograph of opinion is more effective, more instantaneous, but as at the same time political opinions are more volatile than 30 years ago, polls are no more predictive today than yesterday. The watchword really remains caution in their interpretation, especially when the election date is still far away.