Since this post was first published, two new pollsters have new numbers. First, EKOS, which hasn't given numbers all campaign long, comes out with:
PQ 36.0%, CAQ 24.5%, LIB 23.2%, QS 10.7% (sample size 1,749, 8/31-9/3)
Then, Forum gives us:
PQ 36%, LIB 29%, CAQ 25%, QS 6% (sample size 2,781, 9/3)
Furthermore, a fourth projection website, RidingByRiding, was brought to my attention.
This has resulted in the following updates:
1. I moved my own analysis to another post.
2. I added numbers from RidingByRiding, which take all four most recent polls (Forum, EKOS, Léger and CROP) into account.
3. The numbers from Too Close To Call have been updated to take the Forum (but not EKOS) numbers into account.
4. The other two websites have not yet updated for either the latest Forum or the latest EKOS.
Here are the calls of the four projection websites:
Too Close To Call: PQ 66 (33.1%), LIB 33 (27.1%), CAQ 24 (27.4%), QS 2 (7.6%)
ThreeHundredEight: PQ 63 (34.1%), LIB 33 (27.9%), CAQ 27 (26.3%), QS 2 (7.5%)
ElectionsQC: PQ 69 (32.7%), LIB 30 (26.8%), CAQ 23 (27.8%), QS 2 (7.8%), ON 1
RidingbyRiding: PQ 70 (35.0%), LIB 30 (27.8%), CAQ 22 (26.65%), QS 2 (6.5%), ON 1
All four websites call for a thin PQ majority. We note that 308 has different popular vote estimates because it adds an adjustment favouring the PQ and the Liberals at the expense of the CAQ. We also see that the TCTC model to transform popular vote into seats is kinder to the PQ than the 308 model (more seats with lower popular vote), but not as kind as the ElectionsQC model (fewer seats with higher popular vote). RidingByRiding gets results close to ElectionsQC, but probably uses a higher popular vote for the PQ because it includes the Forum poll.
Although I'm out of the projections game this time around, I've still played around the the numbers a little. I believe that the ElectionsQC model is too optimistic for the PQ: if the PQ gets 69 seats, which is certainly possible especially given the latest polls, it would likely need more than 32.7%.
I also disagree with the 308 popular vote adjustment (although if 308 doesn't update for the latest Forum poll, which goes in the same direction, then the adjustment effectively averages that poll in). It could end up being right, but I would not use it as a "best guess", especially since the CAQ's predecessor (ADQ) outperformed the last CROP and Léger polls by 5-6% in 2007, an election with somewhat similar dynamics.
Therefore, of these models, I personally prefer the TCTC and the RxR ones. The difference between their projections can be attributed to the inclusion of the EKOS numbers by RxR, but not by TCTC. However, there is so much uncertainty this time around that I would urge you to be even more cautious than usual when interpreting these models.