I think it’s rather clever how these researchers, Tom Cornwall & Anke Kessler, realized that the individual polling stations within a riding could be used to determine the effect of robocalls and fraudulent live-calls targeted at non-Conservative voters! The various neighbourhoods tend to vote in a particular manner, possibly related to income, religion, or other consistent factors. By comparing polls within ridings with mostly constant variables, it’s easier to expose the voter suppression campaign’s effectiveness. Conservative voters in the riding would have had fewer robocalls misdirecting them, so their turnout comparison ratio to the polls in 2008 would be more identical than the non-Conservative leaning polls.
The data support this presumption: the average winning margin for districts with no robocall allegations was 10,903 votes or 22.8 percentage points. Ridings where allegations of impropriety have emerged, in contrast, had a margin of victory that was almost 28% lower: 8,719 votes or 16.3 percentage points.
Identifying a causal effect of demobilisation efforts on turnout
For this reason, we employ a slightly different strategy: instead of using between-district variation to identify the effect of alleged misconduct, we use within district variation at the level of the polling station. This allows us to control for district-level effects such changes in candidate quality, district level time trends, local TV ad buys, etc. On average, an electoral district has roughly 250 polling stations, which display a considerable within-district variation in turnout and voting patterns: while some polling stations voted predominantly Conservative, others leaned mostly towards opposition parties. Moreover, the latter experienced a drop in voter turnout from the 2008 to the 2011 election, whereas turnout at the former rose between the 2008 and the 2011 election, relative to the district average. Our identifying strategy is then the following: Assume the instigator of the robocalls randomly called all non-Conservative voters in the 27 treated electoral districts. If the calls had the intended effect, then polls with a higher-fraction of non-Conservative voters should experience a larger drop in turnout than comparable stations in non-robocalled districts. Expressed differently, the difference between how Conservative voters and non-Conservative voters turned out at the polls should be more pronounced in those ridings that were allegedly targeted by calls directed to suppress the (presumably non-Conservative) vote, controlling for poll level turnout in the previous election.
“To reduce this level of statistical detail for researchers in the future, expect there to be a single polling place per riding (or just no vote totals reported) in the future,” I add sarcastically.