24 September 2018

The Windsor Lake By-Election and polling #nlpoli

Market research firm MQO released a poll on September 10 about the Windsor Lake by-election.

They'd surveyed 300 people about how they would likely vote and about who they thought would make the best Premier from among the provincial party leaders.

On their choice among the candidates, MQO reported that 41% of decided and leaning voters favoured Paul Antle,  37% favoured Ches Crosbie, and 22% favoured Kerri Claire Neil.

The result on September 20 saw Crosbie win with 43% of votes cast.  Antle got 38% and Neil garnered 19% of votes.

Pretty much the opposite of what MQO said.

So MQO got it wrong,  some will say.

Others will think that Antle's campaign collapsed and Crosbie surged ahead based on whatever factors they might like to pick.

Well, there are a few problems with either of those scenarios.

If you think the poll was right and think that voters changed their minds in two weeks,  you need a reason for the change.  None comes easily to mind.  Neither of the leading candidates did spectacularly well in the three public debates that followed the poll.  Each of the leading candidates had a few hiccoughs.  Neither tanked but there were no great changes in the campaign that would have caused a major flip that mirrored the poll results so precisely but with the first and second place candidates in the poll changing places.

As for the other choice, MQO might have gotten it wrong. Well, yes.  It might have been a simple case of getting the data labels wrong.  There's no way of knowing that outside MQO but it seems like a rather simple and hence unlikely mistake for a major firm to make.

Major polling firms made substantive technical mistakes in forecasts of British Columbia and Alberta general elections.  Might be the same thing here but at this point let's discount that theory.

Instead let's take a look at something else.

Reworking the MQO results

Let's assume that the high level view offered by MQO is right.  If you stand back and look at all the responses,  Antle was in the lead.

MQO also broke down the data to show how men and women answered the question and how people of different ages answered he question. If you look at the age breakdown,  you actually get a different answer than the total view offered by MQO in their release fro that high-end perspective.

Right at the start,  let's acknowledge MQO's openness and generosity in releasing their tabulations of results.  Your humble e-scribbler asked MQO for the original tables compiled from their raw poll data and they sent them out without condition or hesitation knowing that they'd likely be the subject of a blog post.  There was no commitment to write about them nor any understanding about the content.  No one from MQO has seen this information before publications.  What your humble e-scribbler got weren't the tabulated responses in each cell of the table but percentages. 

While there was nothing apparently wrong with MQO's overall result, there were a couple of anomalies that leaped out from the original news release.  One was the fact that respondents thought Ches Crosbie would be a better Premier than Dwight Ball while at the same time preferring Paul Antle over Ches Crosbie in this election.  The other was that Ches did much better with older voters.  MQO didn't explain that in the release so the numbers might show something interesting.

The MQO sample was already skewed for age.  That is, they undersampled by 18 percentage points those under 34  and oversampled by the same amount those above 55 compared to the 2016 census.  They also have slightly more people in the middle age category than the 2016 census shows. Whether that matches the district is a question but what MQO has done is give us a sampling that is already adjusted to account for the error made in the polls in British Columbia and Alberta.

In the percentages MQO showed,  Crosbie had a significantly larger share of the largest age cohort, which was also the people more likely to vote.  That was true for the original tables and for the one that included "leaners",  that is the people who were originally undecided but who were will to say how they were thinking of voting.

What followed after getting the data tables was a bit of math.   That involved translating the percentages by age group (18-34,  35-54,  and 55+) back to numbers and then working out the resulting total number of votes each received as shares of the total votes theoretically cast. What MQO did was to take each percentage  in an age cohort for each candidate and average them over the three cohorts.  That relatively small difference produced the dramatically different result.

Dropping out the undecideds,  no answers, and refusals cut the sample to 191 (from 300)  for the decideds alone and to 225 when you add decideds and leaners.  MQO got 223 for decideds and leaners so the de-engineering produced a roughly comparable result allowing for rounding.

This doesn't mean that all the others matched up to the non-voters.  That leads into a deep discussion about what these polls actually capture but lets just say that the non-responses in a poll seldom come close to matching very display of actual voting behaviour.  The undecideds and refusals do not match the actual share of non-voters in the district. That's another discussion and for now, let's just leave that sleeping dog snoring in the corner.

Here are the results of the reworking of the MQO numbers. and the actual result.  It doesn't take any explanation to show how closely it matched the final result.


  + Leaners  












With this result in hand, there's no longer a discrepancy between voters' preference in the district for best Premier and best candidate.  The whole thing lines up quite neatly and consistently.  In fact, if you use the same technique to figure out the leader preference,  you get a slightly different result than the one MQO found.  It's approach boosted Rogers' result and shaved some of Crosbie's support compared the Bond technique.  Ball's stayed about the same.

In the bigger picture that minor shift in emphasis doesn't really matter since the key point is that voters preferred Crosbie to Ball.  It's logical that they would prefer Crosbie to Ball's candidate.  Incidentally, if you look at Crosbie's messaging,  he spent as much or more time matching himself with Ball, especially in the last week of the campaign, than he did talking about Antle. In that sense the PC messaging played precisely on that strength and reinforced it.  Expect to see more of it.

Here are results in percent presented as MQO/SRBP:

Ball:                38/37
Crosbie:          41/47
Rogers:           19/16