One of the (many!) important questions raised by the robocall scandal is whether or not the deceptive calls did in fact achieve their presumed goal of inducing supporters of opposition parties to stay home and not vote.
If you look at riding-level data, there's not much to see. But Simon Fraser University's Anke Kessler has dug deeper into Elections Canada's poll-level databaseand uses information that is available at the poll level. Outcomes at polling stations differ in turnout and in vote shares for particular candidates; this makes each riding look like a smaller copy of a country-wide election. In a first step, she finds that polling stations with predominantly non-Conservative voters generally experienced a decline in voter turnout from 2008 to 2011. In a second step, she asks how the extent of this decline varies with reported robocalls. She finds that it was larger in the former, meaning that in ridings where robocalling was reported, polling stations that voted predomininantly non-Conservative in the 2008 election saw a greater-then-average decline in voter turnout. The paper "Does misinformation demobilize the electorate? Measuring the impact of alleged 'robocalls' in the 2011 Canadian election" is available here.
The paper presents evidence on the effect of voter demobilization in the context of the Canadian 2011 federal election. Voters in 27 ridings (as of February 26, 2012) allegedly received automated phone calls (‘robocalls’) that either contained misleading information about the location of their polling station, or were harassing in nature, claiming to originate from a particular candidate in the contest for local Member of Parliament. I use within-riding variation in turnout and vote–share for each party to study how turnout changed from the 2008 to the 2011 election as a function of the predominant party affiliation of voters at a particular polling station. I show that those polling stations with predominantly nonconservative voters experienced a decline in voter turnout from 2008 to 2011, and that this effect was larger in ridings that were allegedly targeted by the fraudulent phone calls. The results thus indicate a statistically significant effect of the alleged demobilization efforts: in those ridings where allegations of robocalls emerged, turnout was an estimated 3 percentage points lower on average. This reduction in turnout translates into roughly 2,500 eligible (registered) voters that did not go to the polls. The 95%-confidence interval gives a lower bound estimate of 1,000 fewer votes cast in robocall ridings, which is still a sizable effect.
And here is the passage in the introduction where she describes how she was able to extract her results:
The problem with identifying any possible causal effect of the robocalls is that a crucial variable that likely affects both turnout and the selection of ridings that were targeted is the expected margin of victory, which is unobserved. Instead of proxying for the perceived margin of victory by the actual margin (or the margin in the previous election), I employ a different approach that allows for arbitrary unobservables at the level of the riding and instead uses within riding variation for identification. Specifically, the starting point of my identification strategy is that there is considerable variation of outcomes (turnout, vote shares) on the level of the polling station within a riding. On average, an electoral district has roughly 250 polling stations, some predominantly supported the Conservatives in the past election, while others predominantly supported the main opposition parties (Liberals, NDP, Bloc) in 2008. In the 2011 election, turnout at the latter polling stations fell, while turnout at stations with more Conservative leaning voters increased, relative to the district average. This effect can likely be attributed to the failure of the main opposition parties to mobilize their supporters as effectively as the governing Conservatives. Now assume that the instigator behind the robocalls targeted opposition voters evenly within the riding. Because some polling stations have more eligible voters favoring the opposition than others, we would expect those polling stations to have a differentially lower turnout compared to similar polling stations in ridings that were unaffected, assuming the misinformation worked and supporters of the main opposition parties were in fact demobilized. In other words, comparing ridings with robocall complaints to ridings without, the decline in turnout at those polls with more Liberal NDP or Bloc support should be more pronounced in the former. The result suggest that this is indeed the case: those electoral districts that were allegedly targeted by robocalls experienced a (relative) drop in voter turnout. On average, voter turnout was 3 percentage points lower in those ridings from which complaints had been received as opposed to ridings from which no such complaints had been received. Using the average such riding as a benchmark, this translates into roughly 2,500 fewer voters at the polls.
Here is her graph showing the model's predicted turnout rates:
Section 4 of the paper is devoted to robustness checks.
And there are important caveats to be kept in mind when in interpreting these results:
It is important to note that my findings in no way can ‘prove’ whether misconduct or an illegal act has occurred. First, since I can only rely on self-reported incidences of misleading or harassing calls, which are made available through the Canadian media as my data source, there might be a considerable amount of noise present in the data as the actual occurrence of misconduct is obviously not observed. That is, I can only estimate the effect of (the average level of) robocalling in a riding, conditional on the calls being reported, relative to the effect of (the average level of) robocalling in districts where the calls, if any, have not been reported and consequently do not appear on the robocall list. Second, the findings only apply to the artificial construct of an “average” riding, i.e., the interpretation of the results necessitates an electoral district with average characteristics (voter turnout, margin of victory, etc.) which does not actually exist. For this reason, I wish to emphasize that the analysis and the corresponding results are not suited to bring the outcome in a particular riding into question.
This is an important insight into the robocall scandal, and people who are following this file would do well to read Professor Kessler's paper.