House effects: what it costs to read one pollster instead of another

Every polling house has its own sampling design and its own weighting scheme, and that produces what the literature calls a house effect: a systematic tendency to put a party a little above or below where the rest of the market puts it. With the polling archive now complete, we tried to measure it — and to measure it properly, which is not the obvious way.

Why the raw average is not enough

The instinctive comparison is to line up each house's average over the same months. The problem is that houses do not publish on the same days: in a period when a party's support is falling — which is exactly what is happening to the centre-right — a house that mostly publishes early in the window will look "higher" than one that mostly publishes late, with no bias involved at all.

The correction is simple, and we apply it here: for every poll, take the difference from the average of all houses in the same month, then average those differences. What is left is the house's deviation at a fixed point in time.

The spread between the house most generous to Fratelli d'Italia (Tecnè, +1.56) and the strictest (EMG, −1.42) is almost 3 full points. That is not a detail: it is more than the distance currently separating the country's second list from its third.

The effect is not the same across parties

Three things worth reading out of this table:

• On PD and M5S the market is far tighter. Setting BiDiMedia and Eumetra aside, every house sits within three tenths of the average. Almost all of the dispersion is concentrated on the right-wing lists.
• Lega is the second battleground. From −1.24 (YouTrend) to +0.68 (EMG), and +2.15 for BiDiMedia: on a party currently worth just over 6 points, those are large relative differences.
• The BiDiMedia and Eumetra rows deserve caution. With 8 and 10 polls, and deviations all of the same sign, they are more likely measuring a different roster of lists (a different "Other", a different treatment of undecideds) than a genuine estimation preference.

The edge case: when a single poll has to be dropped

The value of an aggregate average is clearest in the pathological cases. On 3 August 2026 the official registry took in an IZI poll, fielded from 30 July, putting Fratelli d'Italia at 8.1% and the Democratic Party at 9.0% — while every other reading in the same period places them between 25% and 29%, and around 21%, respectively.

Our quality pipeline flagged it automatically as suspect, with the flags sample_size_unparsed, margin_error_confidence_unknown and few_parties_q2. Looking at the extracted rows makes it clear why:

• the poll contains several questions fused into a single result block — alongside the party lists there are values like "Yes 34.1", "No 96.5" and a "Five Star Movement 34.3" that plainly belong to different questions;
• the shares are not renormalised: undecided respondents (17.6%) and those saying they would not vote (8.7%) stay in the base, which mechanically depresses every list.

So it is not a collapse in support, and it is not a mistake by the pollster either: it is an extraction problem on the filed document. But a single data point like this, left unfiltered, visibly shifts a monthly average — which is why our averages exclude suspect polls by default.

Source: ElectionLab polling archive — official register of political and electoral polls (database under art. 8 c.3 of Law 28/2000, Presidency of the Council of Ministers) and the onData dataset (CC-BY-4.0). House effects computed on national polls with data_quality other than suspect from January 2026, as the mean deviation from the all-house monthly average. Data as of 29 August 2026. Compare the houses yourself in the Polls section.

— ElectionLab Studio