The precision of political polls is not what you think it is

The precision of electoral polls is a classical media topic at the end of electoral campaigns. A very quickly forgotten one: a few days ago, i-Télé – a French news channel – devoted several minutes of their broadcast to a jump of François Hollande in the latest release of one of their polls. All told, the jump in question was a 2% one. With only 1000 interviewees, this is noise, not signal.

Quickly forgotten, and anyway in general missing the point. Our article shows that political polls, on the eve of a vote, are in general more precise than the standard mathematical formula would suggest.

If pollsters believed in that formula, they would refrain to comment on a poll giving a 52% vote share to a presidential candidate. With 1000 interviewees, the difference between 48% and 52% is not significant. However, nobody would have said, at the beginning of May 2012, that the French presidential election outcome was uncertain. With reason: that was indeed signal and not just noise.

Where does that unexpected precision come from? It is probably due to the calibration modelling used to get rid of bias in the raw data. At least, this is what we conjecture. It would not be such a surprise that incorporating more information in that raw data would stabilise the results: more – up to the point – information in general means greater precision.