What is the REAL Tory poll lead?
In my thread "are ALL the polls wrong?" I identified sources of error that inflated the Con lead.
In this thread, I go further.
I have re-modelled/weighted the data to try to correct these errors.
The real Con lead? Read on to find out.
1/17
In my thread "are ALL the polls wrong?" I identified sources of error that inflated the Con lead.
In this thread, I go further.
I have re-modelled/weighted the data to try to correct these errors.
The real Con lead? Read on to find out.
1/17
The sources of error I will be attempting to correct are;
1. Turnout
2. Over-weighting of leavers
3. Under-weighting of GE17 non-voters and newly registered voters
I will be clear exactly how I am doing it, so anyone is welcome to tell me if I'm wrong, and how.
2/17
1. Turnout
2. Over-weighting of leavers
3. Under-weighting of GE17 non-voters and newly registered voters
I will be clear exactly how I am doing it, so anyone is welcome to tell me if I'm wrong, and how.
2/17
Turnout is easy to correct.
A few polls have crazy turnout figures (90% 65+, 31% 18-24 etc), and it skews the average Con lead.
Taking the 10 most recent polls and applying 2017 GE turnout, the average Con lead is not 10%, it's 8.4%.
Our starting point is an 8.4% lead.
3/17
A few polls have crazy turnout figures (90% 65+, 31% 18-24 etc), and it skews the average Con lead.
Taking the 10 most recent polls and applying 2017 GE turnout, the average Con lead is not 10%, it's 8.4%.
Our starting point is an 8.4% lead.
3/17
Now to the over-weighting of leavers.
In my previous thread, I identified the demographic changes which meant that weighting to the 2016 EU Ref no longer made sense.
Since the EU ref there has been a net loss of 667,494 leavers and a net gain of 882,735 remainers.
4/17
In my previous thread, I identified the demographic changes which meant that weighting to the 2016 EU Ref no longer made sense.
Since the EU ref there has been a net loss of 667,494 leavers and a net gain of 882,735 remainers.
4/17
In simple terms, it seems there are too many leavers in the weighted data of the polls.
For every 1% the gap between Leave and Remain closes, Con lose 0.35% of their VI, and Lab gain 0.25%.
On the basis of demographic change alone, the Con lead is reduced by 2.4%.
5/17
For every 1% the gap between Leave and Remain closes, Con lose 0.35% of their VI, and Lab gain 0.25%.
On the basis of demographic change alone, the Con lead is reduced by 2.4%.
5/17
We could be more aggressive with EU Ref weighting, and say that the underlying picture is 52/48 remain so the Con lead is -4.8%
But I'm being cautious and going with a demographic adjustment with imputed values for those who have turned 18, adjusted for age-based turnout.
6/17
But I'm being cautious and going with a demographic adjustment with imputed values for those who have turned 18, adjusted for age-based turnout.
6/17
At this point, we have applied 2017 turnout adjustment to correct for the anomalies in some polls (-1.6% Con lead)
And, with a simple demographic adjustment to EU ref weighting, we reduce that lead by another -2.4%
On these two adjustments, the average Con lead is now 6%.
7/17
And, with a simple demographic adjustment to EU ref weighting, we reduce that lead by another -2.4%
On these two adjustments, the average Con lead is now 6%.
7/17
So now it's time to deal with the most challenging issue - people who didn't vote in 2017 (DNV's), and new voter registrations.
This gets fairly complex.
I'm going to use YouGov weighted data as my base for this. It looks about right, although has too few DNV''s.
8/17
This gets fairly complex.
I'm going to use YouGov weighted data as my base for this. It looks about right, although has too few DNV''s.
8/17
We need to answer three questions about newly registered and 2017 GE DNV's:
1) Who will they vote for?
2) How many of them are there?
3) How likely are they to vote?
If we can answer these questions, we can model the impact on the polls with some accuracy.
9/17
1) Who will they vote for?
2) How many of them are there?
3) How likely are they to vote?
If we can answer these questions, we can model the impact on the polls with some accuracy.
9/17
Starting with who they are going to vote for.
Data from IPSOS Mori shows that DNV's and newly registered voters break for Labour over Con by over 2-1.
The exact figures are:
27%Lab
12%Con
21%Don't Know
19%Won't vote
10%LD
So that's who they're voting for.
10/17
Data from IPSOS Mori shows that DNV's and newly registered voters break for Labour over Con by over 2-1.
The exact figures are:
27%Lab
12%Con
21%Don't Know
19%Won't vote
10%LD
So that's who they're voting for.
10/17
In terms of how many of them there are, as at December 2018 there were 47m people on the electoral roll.
In 2019 there were 9m new registrations.
We know 37% are duplicates.
Based on averages, 2m are house movers. That leaves 4m genuine new voters, mostly young.
11/17
In 2019 there were 9m new registrations.
We know 37% are duplicates.
Based on averages, 2m are house movers. That leaves 4m genuine new voters, mostly young.
11/17
This means the size of the 2017 DNV electorate has expanded considerably.
From 15.3m to 19.3m in a year (+26%)
On this basis, let's make an adjustment to the YouGov DNV pool. Let's increase the weighting by 26%, so it moves from 25% to 31.5% of the weighted voter pool.
12/17
From 15.3m to 19.3m in a year (+26%)
On this basis, let's make an adjustment to the YouGov DNV pool. Let's increase the weighting by 26%, so it moves from 25% to 31.5% of the weighted voter pool.
12/17
In terms of likelihood to vote, most pollsters heavily downweight 2017 DNV's, including first time voters.
Given the record number of registrations and record high levels of recorded political engagement, this seems foolish.
Instead, let's apply 2017 GE turnout by age.
13/17
Given the record number of registrations and record high levels of recorded political engagement, this seems foolish.
Instead, let's apply 2017 GE turnout by age.
13/17
Applying 2017 GE turnout by age makes sense, because we already factor in the likelihood to vote of a demographic comprised of many new voters.
There is no logical need to down-weight further.
Now we know 2017 DNV VI, numbers, and likely turnout, we can model the impact
14/17
There is no logical need to down-weight further.
Now we know 2017 DNV VI, numbers, and likely turnout, we can model the impact
14/17
Many pollsters apply, over and above age-based turnout, an additional downweighting of 50% to 2017 DNV's
In the 2019 election, the data suggests that this is not a credible approach.
If we remove this downweighting, Con lead drops another 2.36%
But there's more...
15/17
In the 2019 election, the data suggests that this is not a credible approach.
If we remove this downweighting, Con lead drops another 2.36%
But there's more...
15/17
What about the the 2017 DNV's who are currently 'don't know'?
Let's impute voting intention as for other DNV's, but on the basis that 50% of them don't bother voting at all.
This reduces the Con lead not by 2.36%, but 3.16%.
16/17
Let's impute voting intention as for other DNV's, but on the basis that 50% of them don't bother voting at all.
This reduces the Con lead not by 2.36%, but 3.16%.
16/17
The impact of attempting to correct these sources of error is as follows:
Polling average
Con +10%
Turnout modelled to GE 2017 across all polls
-1.6%
EU Ref demographic changes
-2.4%
Re-weighting 2017 DNV's/New voters:
-2.84%
Actual Con lead: 2.86%
END
Polling average
Con +10%
Turnout modelled to GE 2017 across all polls
-1.6%
EU Ref demographic changes
-2.4%
Re-weighting 2017 DNV's/New voters:
-2.84%
Actual Con lead: 2.86%
END
CORRECTION
Polling average
Con +10%
Turnout modelled to GE 2017 across all polls
-1.6%
EU Ref demographic changes
-2.4%
Re-weighting 2017 DNV's/New voters:
-3.16%
Actual Con lead: 2.84%
THE REAL END
Polling average
Con +10%
Turnout modelled to GE 2017 across all polls
-1.6%
EU Ref demographic changes
-2.4%
Re-weighting 2017 DNV's/New voters:
-3.16%
Actual Con lead: 2.84%
THE REAL END
