Update to Model Parameters and Discussion of Poll Weighting

As explained in the methodological post on poll weighting, the new formula is one of the main improvements that I have made to the model for this election: it derives the weights from a statistical framework rather than basing them on arbitrary discount factors. After a month of data coming in (I include all national polls that include field dates after July 1), it has become apparent that the individual-pollster variance parameter of 0.0001 (1% standard deviation) is too low; I am changing it to 0.0002 to better fit the recent data.

(Update Aug. 11: I have also changed the pollster common variance from 0.0002 to 0.0006 based on the frequency and magnitude of misses in recent provincial elections. This will only affect the confidence intervals, and will be reflected the next time I give such intervals.)

This update should make projections slightly more stable by increasing the model's weariness of concentrating weight on any one pollster. It causes minor changes to the latest projection, which is now:

Seats projected ahead as of the latest national poll (midpoint: August 4)

LIB - 155 (33.3%)
CON - 143 (35.5%)
NDP - 23 (13.9%)
BQ - 11 (4.2%)
GRN - 5 (10.0%)
IND - 1 (Wilson-Raybould)

If you're new to my 2019 projections, view key interpretive information here.

- Rimouski-Neigette--Témiscouata--Les Basques now stays NDP.
- Québec now remains a Conservative pickup (with caveat).
- Richmond Hill now remains a Conservative pickup.
- Hamilton East--Stoney Creek is now an NDP pickup.

What I want to do now is to illustrate how the weighting formula is working. As I explained in the earlier post, negative weights on some polls can arise, and I gave a theoretical explanation. In practice, negative weights are playing an even greater role than I was expecting... and it makes perfect sense.

For a full list of included polls and a link to each of them, see previous Projection posts.

Consider the following:
- 17 national polls combining about 29,000 respondents have been fed into the model.
- The total weight on the three most recent polls, totaling under 20% of the total respondents, by Innovative, Mainstreet and Abacus, is 64%.
- The total weight assigned to each pollster is:
16% Mainstreet
16% Abacus
13% Innovative
11% Forum
9% EKOS
7% Léger
6% Angus Reid
6% Campaign Research
6% Nanos
5% Research Co.
5% Ipsos

As you can see, the formula achieves the dual objective of putting high weight on recent polls without giving any pollster undue influence. If you look more closely, you'll note that the three pollsters with the most recent polls (totaling 64% weight) have a combined weight of just 45%. This is achieved by assigning negative weight to those pollsters' older polls, and the sum of those negative weights is a significant -19%. If those pollsters did not have an older poll, then their newer poll would receive a lower weight, as there would be no way to correct for their pollster-specific bias.