A Note on Robocalls in the 2011 Federal Election
Summary Recently Prof. Anke S. Kessler of Simon Fraser University published a working paper on a regression model that suggests alleged robocalling had “a statistically significant impact on voter turnout and election results” by suppressing anti-Conservative turnout in 27 ridings in 2011.1 I estimated her regression model using 27 other, similar ridings, where there were no robocall allegations (though the PMO may have used secret mind control devices to suppress antiConservative turnout). Not surprisingly (to me), Prof. Kessler’s model suggests that the effect of the entirely fictional mind control devices was greater than the effect of the alleged robocalls. Background Prof. Kessler’s study suggests that robocalls may have suppressed non-Conservative turnout in 27 ridings, allowing the Conservative Party of Canada (CPC) to win several additional ridings. As might be expected, media interest in the study has been considerable: • • • • • • • • • • Study supports vote suppression allegations (David Akin, Sun News) Did robocalls affect the last election? Apparently so, says a new study (Montreal Gazette) ELECTION FRAUD! Robo-Calling May Have Significantly Impacted Voting, Says SFU Study (The Link) Study says misleading 'robocalls' may have affected federal election results (StarPhoenix) Misleading calls may have affected election: study (Victoria Times Colonist) Calls may have affected election results: professor (Vancouver Sun) Alleged robo-calling may have significantly impacted voting (E Science News) SFU prof: Robocalls did suppress voter turnout (The Tyee) The effect of robocalls (Ottawa Citizen) Robocalls could have reduced turnout (Burnaby Now)
Skeptical commentators like Gerry Nicholls, who dismissed the study as “statistical nonsense,” were pilloried as anti-science philistines on Twitter. For my part, I noted that any apparent correlation between robocall allegations and lower turnout in any particular riding should not be presented as evidence that any alleged robocalls caused a decline in turnout. Prof. Stephen Gordon of Laval disagreed,2 and suggested that if there is evidence of correlation, there should be a presumption in favour of causation unless there is another “causal story” that would generate the same correlation. But as I tried to explain via Twitter, the causes of variation in
“Using the fact that the average targeted riding had 83,268 registered voters, this translates into an estimated absolute number of roughly fewer 2,500 voters showing up at the polls. This is substantial. Among those districts allegedly affected by robocalls that are on the list I employ, a total of 6 ridings had winning margins smaller than that. The lower bound of the 95% confidence interval is 1,032 voters that did not vote in robocall ridings on election day, which is still larger than the winning margin in three ridings (two of which were won by a Conservative candidate).” (Page 11) 2 “‘Correlation is not causation’ does NOT mean ‘.. so we can ignore the correlation.’ If you’re using this line properly, you have thought of another, alternative causal story that gives the *same* correlation. If you can’t, you’re using a slogan you don’t understand.”
turnout are varied and complex; there may be no alternative “causal story” – at least, not one that can be measured and tested in an econometric model. In any case, my call for a bit of humility from the “pro” econometricians was unpopular on Twitter. Prof. Gordon advised me, hopefully not entirely sarcastically, to “Write [my objections] up and show us how it’s done.” This note is my response, though as I said I’m more interested in showing how it’s not (or shouldn’t be) done. Thus, I estimated Prof. Kessler’s model once again, replacing the 27 ridings where there were robocall allegations with 27 similar ridings, where there were none (according to Prof. Kessler’s paper). Etobicoke Centre was replaced by Etobicoke-Lakeshore, Mississauga East-Cooksville by Mississauga South, and so on. I chose the new ridings based on: • Proximity. Ideally, the new riding is adjacent to the robocall riding. For example, I replaced Winnipeg South Centre, MB with Winnipeg South, MB, and Egmont, PEI with Malpeque, PEI. If this was not possible, I tried to choose a new riding from the same region; for example, Sudbury was replaced by Thunder Bay-Rainy River in Northern Ontario. Demographics. Ideally, the new riding has a similar urban/suburban/rural mix as the robocall riding. For example, Willowdale is replaced by York Centre in predominantly urban Toronto, Oakville is replaced by Burlington in suburban Halton, and Haldimand-Norfolk is replaced by Elgin-Middlesex-London in rural Southwestern Ontario. Political History. Ideally, election results in the new riding, in both 2008 and 2011, were similar to those of the robocall riding, in terms of the winning party, margin of victory, incumbency, etc. Partly for this reason, I replaced some ridings in Eastern Ontario (Ottawa Orleans and Ottawa West-Napean), Southwestern Ontario (Guelph, Kitchener-Conestoga, Kitchener-Waterloo and London North Centre), and Niagara (St. Catharines) with ridings in the GTA.
As I noted above, none of the new ridings are included in either Prof. Kessler’s original list of 27 ridings where robocalls were alleged, or her “extended” list of 72 ridings where reports of robocalls later emerged. Thus it is safe to assume that if the model’s interaction term is still statistically significant, it is not because of the effect of alleged robocalls. Instead, there were obviously other factors at work in these ridings that we cannot fully understand (let alone measure). However, since this is an exercise in econometrics, and for fun, we may postulate that I have obtained evidence that, instead of robocalls, the PMO directed a secret, evil mind control device at anti-Conservative voters in the 27 new ridings in order to suppress turnout. I label this dummy variable “mindctrl,” and instead of Prof. Kessler’s model,
Not surprisingly, the new interaction term mindctrli × lagoppvoteshareij is statistically significant. Perhaps surprisingly, though, the effect of the PMO’s secret mind control devices actually exceeds that of the alleged robocalls.3
Including both interaction terms in the same model, we can estimate the following:
Table 7 is reproduced from page 19 of Prof. Kessler’s working paper. Note that the standard error for the interaction term robocalli × lagoppvoteshareij appears to be incorrect; 0.008 is the standard error for the constant term, while 0.0137 is the standard error for the interaction term. Also note that I estimated the non-weighted regression model, though the weighted results are similar.
All variables are significant at both the 95% and 99% levels, but in a contest between the effect of alleged robocalls and the effect of fantasy mind control rays from the PMO, the mind control rays win. Additional calculations could provide estimates of the precise number of anti-Conservative voters (in the thousands) who did not vote, according to this model, because of the PMO’s mind control devices, and of the number of ridings the CPC won because of this evil technology. I will leave that exercise to others.
Michael R. Smith firstname.lastname@example.org Twitter: michael_r_smith