Sunday, May 19, 2013

The B.C. election and polling fiasco

In August 1972 a hamburger poll at the Pacific National Exhibition accurately predicted Dave Barrett and the NDP would win the B.C. Election held August 30th of that year. TC has long had great faith in the capacity of polls, even hamburger polls, to capture public sentiment in election campaigns in a way that correctly, if not precisely, anticipates the election result.

The polls on the whole have delivered. Using Pollster.com's capacity to create one's own prediction map  and based on their aggregation of polls TC was able to accurately predict the outcome in 49 of 50 states in last year's U.S. Presidential election. The fiftieth, Florida, I had as too close to call. Most polls got the 2012 U.S. election right.

However, in B.C. in 2013 the polls were wrong and the errors were extreme (something similar happened in Alberta last year and there were significant regional errors in the 2011 Canadian federal election).  For the two leading parties (the minor parties should be ignored for the purpose of this analysis) only three polls out of the final seven had a result inside the margin of error for the NDP and all were outside the margin for the Liberals.



If we rank the pollsters by methodology we see that while all the polling was weak, the online polls were generally weaker than the phone polls (IVR or interactive voice response polls are still phone polls conducted using automated computer equipment rather than human operators). In the table below you can see that on the whole the phone was a bit more accurate.


Telephone polls do have two problems today that were not present in the not too distant past. First, many people either don't answer the phone or use call display to filter out unwanted telemarketing or polling calls. Partly because of this response rates have fallen dramatically. In addition phone polls to be valid need to call cell phones, which have become so widespread especially among the young that you can't capture an adequately representative sample without them but not all do. The low response rates mean the potential for capturing an accurate representation of the population must be declining.

Online polls are surveys conducted among panels recruited by the company to answer surveys on a variety of topics mostly market research (TC was formerly part of an online panel and has answered many such surveys).

The most important point about this methodology is there are no established standards or methods for what is a recently developed technique for polling.  Some good results have been obtained (there have been some bad results too) but considerable uncertainty about the method remains.

But here is something the public doesn't know and something journalists should be demanding from the online companies.  While companies such as Ipsos say their national samples are large, as large as 200,000, we don't know the exact size of the B.C. part of its national panel (is it about 26,000 representing roughly BC's share of Canada's population) nor do we know fully what methods were employed. Almost by definition, members of online panels are more interested participating in survey research than is typical of the population. Is this not likely to be an important factor in determining outcomes?

The UK firm YouGov has a decent reputation (Nate Silver who generally likes phone polls said in evaluating polls in 2010 "A firm that conducts surveys by Internet, YouGov, also performed relatively well.") YouGov provides a detailed description of its methods on its web site. For example, YouGov notes "restrictions are put in place to ensure that only the people contacted are allowed to participate". Online polls are going to be part of our future, but the industry should establish some minimum standards about how such polls are to be conducted, and should be much more transparent about their methods than they have been to date.

BC may be a more difficult place to poll than other provinces. Reflecting its diverse geography BC is a province of micro-climates both meteorological and political. Pollsters should have a tougher time capturing its variety and likely do. Polls also have a hard time with turnout, which was low in this election. For example the weighting of results demographically by age has a critical impact (older voters are much more likely to vote than younger voters). No doubt some of the error was rooted here.

Paradoxically, at the same time that publicly available media polls are getting weaker we are entering an era where, in private, well-funded sophisticated campaigns such as the 2012 Obama campaign spend massive sums of money to give themselves an extensive almost-census like knowledge of political preferences.  This quote describes the Obama's campaign's activity in battleground states:
For each battleground state every week, the campaign’s call centers conducted 5,000 to 10,000 so-called short-form interviews that quickly gauged a voter’s preferences, and 1,000 interviews in a long-form version that was more like a traditional poll. To derive individual-level predictions, algorithms trawled for patterns between these opinions and the data points the campaign had assembled for every voter—as many as one thousand variables each, drawn from voter registration records, consumer data warehouses, and past campaign contacts.
The failures of polling matter  Our knowledge of politics depends on them in a way that was not the case many years ago. It clearly affected the perceptions of the campaign from beginning to end on the part of the parties, the media and the public. Watching election night coverage it was clear that the Liberals were just as surprised as the NDP. I am distrustful of all claims by those who now say they had foreknowledge.

In hindsight it is clear now that the NDP lost in no small part because it was receiving misleading signals from the polls.  It ran an unfocused weak campaign that took victory for granted (somewhat like that of Lyn Macleod and the Ontario Liberals in 1995 when a big lead at the outset became a Mike Harris government on election day) when they should have been looking over their shoulder.

The whole story is complex but let me discuss two factors:

Adrian Dix made a disastrous decision to forgo attack ads and insisted on a positive campaign.  It seems pundits and politicians confuse negative politics with unfair dirty politics. Negative political discourse whether it is question period, a TV debate or an ad critical of an opponent represents the normal content of political debate.  There is a key distinction to be made, however, between negative campaigning and dirty politics.  It is the latter that should be denounced at every opportunity, especially by journalists and political scientists. Those who have difficulty with this should read Kathleen Hall Jamieson's classic work, Dirty Politics, Deception, Distraction and Democracy.  The BC NDP could have run ads critical of the many twists and turns in Christy Clark's style of governing and the Campbell government's handling of the HST fiasco, which could have been tough but fair, without being the equivalent of the Willie Horton ad (for context see analysis here).

Whether deserved or not, the NDP suffers from a longer term reputation as a party that is weak on economic growth and fiscal management. The Clark campaign ads (they strike TC as quite effective) took advantage of the fact that Adrian Dix was associated with an NDP government in the 1990s that has been portrayed as playing into this NDP stereotype. His mid-campaign shift on the pipeline issue fed into this. This was always going to be a potential weakness the NDP needed to address not just through the proposals it made as a party, but through tough attacks on the record of its opponents as well. It is not something that can be ignored.

There is always a next election. This year's Liberal campaign talked about "getting rid of BC's debt".  Perhaps the BC NDP should make careful note of that pledge.

Tuesday, May 14, 2013

BC Election - Final Poll Average

I have averaged the final polls from BC including all those released as of the morning of election day.

The results are:


BC Liberal NDP BC Cons. Green Other Sample Margin of Error
Closing average 35.7 43.2 7.1 10.9 3.0      5,928 1.3

In seats this yields 50 NDP, 34 BC Liberals and one other.  However, although my estimate model doesn't show it, I would not be surprised if the Greens picked up one or two.  The ridings they would be most likely to win would come from the BC Liberals.

My seat projection also shows six very close races including Premier Clark's seat of Vancouver-Point Grey.

Sunday, May 12, 2013

BC Election - Efficiency of the Vote

I heard on the radio this morning yet another reference to efficiency of the vote in the B.C. election.  The generally accepted assumption is that the NDP has the more efficient vote.

This is something that I can use my projection model to measure.  Using a projection based on the 2009 election, TC calculates that the BC Liberal vote is very slightly more efficient but the difference is negligible.  It is too small to be worth thinking about.

Saturday, May 11, 2013

British Columbia's Election

B.C.'s election is drawing to a close and in TC's view it will be a comfortable majority for the NDP and Adrian Dix. Despite all the excitement about a Liberal comeback, the shift in the polls from big NDP leads in the early going to an apparently closer race now looks like a reversion to the mean: in other words we are seeing a return to something closer to the longer run average of political support for the two major parties in BC politics. Most polls except for a Forum Poll conducted on May 8 can be found here on the Tyee's Election Hook website.

There is little doubt the early polls captured something.  However, most of the movement appears to have been disaffected BC Liberal voters returning home, voters who were anti-NDP to begin with but undecided due to antipathy to the Clark government. The Liberal party's attacks on the NDP paid dividends here.  However, almost every poll has had the NDP above 40 per cent and the BC Liberals below (the two exceptions to this are both Forum polls).  The average of the polls conducted this week is:

BC Liberal - 37.2%
NDP - 43%
Green - 10.2%
BC Conservatives - 6.6%

TC expects to see about 50 New Democrats and about 35 BC Liberals win, possibly one Green although that is probably getting less likely.  Perceptions of a close race will drive some voters, who view both the Greens and the NDP favourably, towards the NDP in order to avoid re-electing the BC Liberals to a fourth term.

One or two points about the polling.  Forum research commits a sin that I think needs to be noted: they conduct their polls on a single day.  Their polling is conducted using Interactive Voice Response, a valid methodology where respondents answer a short phone survey where the questions are played back from a computer and the responses digitally recorded by punching phone keys. However, it means they can turn on their computers and just let them run until they have the sample they need (all firms weight their samples so that they resemble the population at large). In TC's view they should spread out the survey over a few days the way other firms do.

Consider Forum's description of its methodology of its April 30th poll:
The Forum Poll™ was conducted by Forum Research with the results based on an interactive voice response telephone survey of 1,055 randomly selected residents of British Columbia 18 years of age and older. The poll was conducted on April 30th, 2013.
Now compare that to an American IVR polling company with a good reputation, Survey USA, as they describe a poll they conducted recently on the Los Angeles Mayoralty race:
Cell-phone and home-phone respondents included in this research: SurveyUSA interviewed 1,000 city of Los Angeles adults 04/08/13 through 04/10/13. Of the adults, 840 were registered to vote. Of the registered, 478 were determined by SurveyUSA to be likely to vote in the 05/21/13 runoff. This research was conducted using blended sample, mixed mode. Respondents reachable on a home telephone (82% of likely voters) were interviewed on their home telephone in the recorded voice of a professional announcer. Respondents not reachable on a home telephone (18% of likely voters) were shown a questionnaire on their smartphone, tablet or other electronic device.
One approach is clearly more sophisticated than the other. In the 2011 federal election the polls did not cover themselves with glory in terms of accuracy.  However, according to my calculations, Forum was eighth out of nine in total error on its national poll numbers.

Most of the other polls used online panels to conduct their surveys, a methodology that keeps getting better and allows better targeting. For example, the last Angus Reid survey was able to ask respondents only about parties actually running candidates in their constituency. Online polls, however, still can't guarantee their sample is fully representative of the population.