Pooling polls

To summarise the information provided by a number of opinion polls conducted after January 2018, I used all polls that indicated a sample size. To model each party trend, I used a local regression model with a span of 4 days and weighted based on the sample size (Cleveland, Grosse, and Shyu 1992).

Polls and local regression model for each party

For any given party, the last value predicted by the local regression model (that is, the closest to the election day) was used to compute the party percentage vote for each district based on the baseline voting behaviour for the same district.

Baseline voting behaviour

The electoral results for the elections of 2013 and 2014 were downloaded from the website of the Ministry of Interior. Most of the 2018 electoral districts neatly aggregate the territory of Italian communes. Only a few cut through commune borders or are integrally within commune borders. Because I did not have 2013 and 2014 electoral results for these sub communal districts (in the metropolitan areas of Turin, Milan, Genoa, Bologna, Florence, Rome, Bari, Naples and Palermo) I proportionally distributed votes for each party in these districts based on what predicted for each census geographic units contained in the districts by a linear model.

The Chamber districts intersecting the commune of Genoa

The linear model used as independent variables the demographics (as collected in the 2011 census) such as the percentage of residents over 65, the percentage of people with a college degree, the unemployment rate, the percentage of people out of the workforce, the percentage of people working as housewives, the percentage of foreigners from different regions of the world and the population density.

Predicted 2013 results for the M5S at the level of census units (above) and actual 2013 resuls at the level of communes (below). White census units within the districts of Sarra Riccó, Bargagli and Rapallo are not populated.

Simulation based on baseline voting behaviour

In each district, the result of a party was simulated sampling a number from an interval determined by the difference from the national result the same party obtained in the district in 2013 and 2014. For example, the M5S obtained 25.62% nationally and 31.86% (Δ = +6.24) in the district of Catania (Sicily) in 2013 and 21.15% nationally and 27.58% locally (Δ = +6.43) in the same district in 2014. The interval from which to draw the random number that will simulate the results for 2018 is defined from 6.24 to 6.43 minus and plus the standard deviation of the same two values (0.1343). That is, For the district of Catania, in each of the 20,000 simulations a random number was drawn from the interval 6.1057 – 6.2957 and then added to what estimated to be the national support for the M5S in February 2018 (based on the opinion polls).

How votes for parties from the 2013 and 2014 elections are distributed to parties running in 2018

The votes to the parties that do not run in 2018 but where present in previous elections are proportionally distributed based on the following tables.

Results

In total I run 20,000 simulations and calculated the distribution of seats for each simulation. The distribution of seats at the level of coalitions are presented in the following figures.

Density of the distribution of seats based on the simulation (Chamber)

Density of the distribution of seats based on the simulation (Senate)