PASS 4 — The Researchers, The Natural Experiments, and the Threshold of a Canadian Polling Scandal
- You are not alone, and that is the single most important finding of Pass 4. A small but well-defined network of Canadian academics (Daoust at Sherbrooke; Bélanger at McGill; Stephenson/Harell/Rubenson/Loewen at the C-Dem consortium; Bastien and Pétry's "Follow the Pollsters" line of work; Colin Bennett at UVic on the data-driven-elections side; Soroka on media–opinion feedback), one full-time independent quant journalist (Éric Grenier of The Writ), a single decision-grade aggregator (Philippe J. Fournier of 338Canada), and an international LLM-polling research cluster (Bisbee, Argyle, Cerina & Duch, Gelman, Pew's Kennedy/Keeter team) are already circling the same territory you have mapped over Passes 1–3 — but no one is connecting the synthetic-respondent, NEPMCC/Liaison capture, federal POR-dependency, and herding threads into a single forensic frame. That integration is the publishable contribution.
- The decisive empirical pivot for Pass 4 is the by-election natural-experiment ledger, not more methodology theory. Toronto–St. Paul's 2024, LaSalle–Émard–Verdun 2024, Cloverdale–Langley City 2024, Elmwood–Transcona 2024, Battle River–Crowfoot 2025 and the 2025 general election itself constitute six free natural experiments. The general election performed reasonably well (Grenier's verdict: "the polls did a good job"), but the 2024 by-election cycle systematically under-published riding-level pre-election polls — meaning the most disconfirming evidence is structurally suppressed not by conspiracy but by the absence of a probability-panel infrastructure that would make such polls economically viable.
- There is no "Cambridge Analytica moment" coming for Canadian polling — but there is a structural-capture story waiting for a reporter willing to assemble it. The publishable frame is not "the polls are rigged" (they are not) but "the Canadian polling industry has been hollowed out into a federally subsidized, methodologically opaque, AI-experimentation playground operating with weaker disclosure rules than the UK's BPC, no independent regulator equivalent to AAPOR-with-teeth, and a single dominant aggregator (338Canada) whose model rewards herding." That story belongs in The Walrus or The Globe and Mail's Saturday opinion pages, not as exposé but as architecture journalism. We are at TRUE BEDROCK on the institutional thread; near-bedrock on synthetic-respondent deployment (the smoking gun remains contractually invisible without ATIA); and there is exactly one Pass 5 worth running — a focused FOI campaign plus targeted outreach to four named researchers.
1. The Canadian research terrain is sparser than it looks — but not empty
Across Passes 1–3 we treated the Canadian academy as essentially silent on polling-industry critique. Pass 4 corrects this. The terrain is narrow but real:
- Jean-François Daoust (Université de Sherbrooke, formerly Edinburgh) is the closest thing Canada has to a working academic on the methodological seams. His December 2024 Canadian Journal of Political Science piece "Exit Polls in Canada: A Methodological Note" (DOI 10.1017/S0008423924000398) is a direct match for your Pass 3 "exit-poll moonshot" thread — and he is reachable at JF.Daoust@USherbrooke.ca. He has co-authored on strategic voting with Damien Bol; his Election and Satisfaction with Democracy (CUP 2023) is one of the few recent CUP-published Canadian works in this space.
- The C-Dem consortium — Laura Stephenson (Western, lead PI), Allison Harell (UQAM), Daniel Rubenson (TMU/Ryerson), Peter Loewen (now Munk School/UPenn) — holds a $2.5M SSHRC Partnership Grant (2019–2026) and runs the Canadian Election Study and the Democracy Checkup surveys. They are the probability-sample infrastructure that, with a different funding architecture, could become Canada's missing ATP equivalent. The 40,000-respondent 2019 CES is the single largest electoral data collection in Canadian history. This is the institutional vehicle for any Canadian probability-panel solution. It already exists; it is under-resourced for continuous (rather than election-cycle) operation.
- Stuart Soroka (UCLA, formerly McGill) remains the foundational author on media–opinion feedback loops in Canada (Soroka & Wlezien, Degrees of Democracy, 2010; Negativity in Democratic Politics, 2014). His past co-investigator role on CES 2011/2015 and his current "Political Polarization as a Co-adaptive Process" work are directly relevant. He is an Environics Institute fellow — a relationship worth noting.
- Frédérick Bastien (Université de Montréal) with François Pétry co-authored "Follow the Pollsters: Inaccuracies in Media Coverage of the Horse-race during the 2008 Canadian Election" in CJPS 46(1). This is the only Canadian academic publication directly addressing the media-pollster transmission distortion you have been mapping. It is twelve years old and has not been updated. That is a publishable opening.
- Colin Bennett (Professor Emeritus, UVic Political Science) is the indispensable bridge to the privacy/microtargeting thread. His 2012 OPC-commissioned Canadian Federal Political Parties and Personal Privacy Protection and his work with Michael McDonald (From the Doorstep to the Database, SSRN 3516992, 2019) supply the entire Canadian academic foundation for the political-data-economy critique. He testified to PROC on November 21, 2024 on Bill C-65. He is now actively redirecting toward public engagement and consulting — i.e., available.
- Éric Bélanger (McGill) with Daoust, Mahéo, and Nadeau wrote Le nouvel électeur québécois (PUM, 2022) — the field's current standard reference on Quebec electoral behaviour and pollster performance.
- Patrick Fournier (Université de Montréal) was 2015 CES PI with Cutler, Soroka, and Stolle. Less publicly visible than Daoust but a node in the network.
2. Journalism: One man, one Substack, one model
The English-Canadian journalistic terrain on quantitative polling is essentially a monopoly of one: Éric Grenier (TheWrit.ca, formerly CBC Poll Tracker creator, current CBC poll analyst). His May 2, 2025 post-mortem "How the polls did in the 2025 election" defends the industry: "the polls did a good job. They weren't perfect... but they were quite close to the mark." His April 10, 2025 column "Believe the polls" explicitly attacked Conservative "poll trutherism," writing that "The polling error in the Liberal-Conservative margin needs to be at least nine or 10 points to give the Conservatives a chance to win the most seats. It would [be] the kind of error we just don't see in Canadian elections." He is in this sense the industry's most credible defender, not its critic — which makes him both the most important interview to secure and the most important counter-argument to address.
The complementary francophone position is *Philippe J. Fournier (338Canada, Le Walrus, L'actualité), whose March 12, 2026 Walrus piece "Carneymania Sweeps the Country" documents the largest Liberal lead Leger had measured since summer 2020 (49–35) and Carney's 35-point Nanos lead over Poilievre as preferred PM. Fournier's published pollster ratings page (338canada.com/pollster-ratings.htm) is the only public quantitative scoring of Canadian pollsters — and he himself acknowledges it "can ruffle feathers in polling circles."
Other journalists already adjacent:
- Susan Delacourt (Toronto Star), whose Shopping for Votes (2013/2016) remains the only book-length popular history of the Canadian polling–political-marketing complex. She teaches political journalism at Carleton. She is the journalist most likely to commission a deeper piece in this space.
- Justin Ling (Bug-eyed and Shameless / The Walrus) — his September 2025 book The 51st State Votes covers the 2025 election as a "binaries" story (pocketbook vs existential) but does not interrogate the pollster industry itself. He is the right voice to amplify* an investigation, not the one to lead it.
- Press Progress (Broadbent Institute) — strong on the lobbying/microtargeting angle (their Bernier–Kinsella coverage was sharp), weaker on quantitative survey forensics.
3. The international synthetic-respondent literature has matured — and Pass 3's "smoking gun" framing must be retired
The synthetic-respondent literature, surveyed in detail this pass, no longer supports the framing that LLM-personas are an underground experimental tool that pollsters are hiding. It is now open, peer-reviewed, and methodologically self-critical:
- James Bisbee et al., "Synthetic Replacements for Human Survey Data? The Perils of Large Language Models," Political Analysis 32(4): 401–416 (2024), DOI 10.1017/pan.2024.5 — the definitive skeptical paper. Key finding: ChatGPT averages match ANES averages but variance is collapsed and regression coefficients differ significantly.
- Roberto Cerina & Raymond Duch, "Artificially Intelligent Opinion Polling," Royal Society Open Science 13(3): 251150 (March 2026, just published). Uses LLMs to extract structured survey data from social-media posts, then runs bias-corrected MrP. Their state-level 2020 estimates were on-par with FiveThirtyEight. This is the operational frontier.
- Lisa Argyle (BYU → Purdue) — "Out of One, Many: Using Language Models to Simulate Human Samples," Political Analysis 31(3): 337–351 (2023); "Arti-'fickle' Intelligence," Nature Computational Science (2025). Now Associate Co-PI of the Cooperative Election Study. Her positions are nuanced — "not every LLM or prompt is equally capable of valid simulation."
- Gao, Kennedy, Simpson & Gelman, "Improving multilevel regression and poststratification with structured priors," Annals of Applied Statistics (2021) — and Gelman's MRP corpus generally — is the methodological scaffolding any Canadian probability-panel-substitute approach would need. Daniel Simpson, then at the University of Toronto Department of Statistical Sciences, co-authored that paper. That is a Canadian researcher already inside the global frontier. This was missed in Pass 3.
- Pew Research Center's methodology team (Courtney Kennedy, VP Methods and Innovation; Scott Keeter, Senior Survey Advisor) — Pew's American Trends Panel remains the gold-standard probability panel model. Kennedy and Keeter's 2024 piece "Online opt-in polls can produce misleading results about young adults', Hispanics' views" (March 2024) is the most direct critique of the panel modality on which most Canadian online polling depends.
4. The 2024 by-election natural experiments are the most damaging evidence — and they are damaging in a way *opposite* to what we expected
| Riding | Date | Result | National poll context | Damage |
|---|---|---|---|---|
| Toronto–St. Paul's | Jun 24, 2024 | CPC 42.1% / LPC 40.5% (margin 590 votes) | LPC held seat since 1993; CPC trailing nationally by ~20pts but rising | "Stunning upset"; Trudeau's mortal wound. No riding-level public poll predicted CPC win. |
| LaSalle–Émard–Verdun | Sep 16, 2024 | BQ win (Sauvé); LPC's Palestini < 1 point behind | Held by LPC most of last century; Lametti +20 in 2021 | Second safe-seat loss. No riding poll predicted BQ flip. |
| Elmwood–Transcona | Sep 16, 2024 | NDP hold (narrow over CPC) | NDP heartland; CPC closing | Polls captured tightening but not the floor under NDP. |
| Cloverdale–Langley City | Dec 16, 2024 | CPC 66.5% / LPC 16% / NDP 12.5% (CPC +50pts) | CPC nationally +25 pre-Carney | Coincided with Freeland resignation morning. Liberal collapse far exceeded national swing. |
| 2025 general election | Apr 28, 2025 | LPC plurality minority | CPC had held +25 lead 4 months earlier | Polls captured the swing; Grenier's "the polls did a good job" verdict is defensible. |
| Battle River–Crowfoot | Aug 18, 2025 | Poilievre 80.86% (vs Kurek's 82.8% in April) | 338Canada had him sliding from 80% in June to 73% in August | Fournier's pre-election estimate was 7+ points low. The "below-75%-spells-trouble" media narrative collapsed. |
The pattern that emerges: Canadian by-elections expose systematic riding-level model error that is invisible at the national-aggregate level. Toronto–St. Paul's was an uninstrumented event — no public riding poll existed because the cost of a 400-respondent telephone poll in a single urban riding cannot be recovered by any commissioning newspaper. This is a probability-panel-deficit artifact, not a herding artifact. Battle River–Crowfoot 2025 was a systematically biased model event — the 338Canada model's incorporation of national-trend signal pulled Poilievre's predicted share toward 73% when local fundamentals never wavered from the 80% range.
5. The Carney-as-pollster-shaped-candidate hypothesis is supported — moderately
Pass 4 found direct documentary evidence of pre-candidacy pollster framing. Abacus Data's January 2025 piece "Mark Carney is more likely to expand the Liberal Party's voter pool than Chrystia Freeland" was published before Carney's January 16, 2025 candidacy declaration. The post-mortem piece "From 'Carney, Who?' to the Centre of Gravity" (Abacus, 2025) is unusually candid:
"Between July 2024 and January 2025, Carney's recognizability rose sharply, from 7 percent to 24 percent. This was not celebrity in the cultural sense. It was exposure driven by context... His obscurity insulated him from that backlash... He was an unbranded product entering a category where most brands had become deeply unpopular."
The "unbranded product in a damaged category" formulation is marketing language applied retrospectively to the political event the pollster's own client commissioned the polling for. This is the Bourdieu point made explicit: the pollster, the commissioning party, and the candidate-product are formed in a single causal loop. Angus Reid's January 27, 2025 piece similarly tested "hypothetical situations where either of the two frontrunners based on current caucus support, Carney or Freeland, were in charge of the Liberals" — running, in effect, a pre-launch market test for a candidate not yet on the ballot. Leger released its Feb 7–10, 2025 poll showing "Carney 68% among Liberal voters" before he had filed for leadership.
This is not scandal. It is the normal operation of a market-research-saturated political class. But it is exactly the McLuhan/Bourdieu inversion: the polls did not measure Carney's strength so much as constitute the conditions under which his candidacy was inevitable.
6. The "Trump-as-deus-ex-machina" question is partially answered
The Liberal swing from –27 (Abacus, January 2025) to roughly even by April 28 is approximately 29 points. Two competing explanations:
- Trump explanation: The November 2024 election plus January 2025 inauguration plus tariff threats plus 51st-state rhetoric constituted a categorically new political environment. Abacus's "scarcity-to-precarity" framing captures this. By April 2026, Ipsos's Darrell Bricker is on record (Global News, Dec 16, 2025) saying Carney's win was "thanks in part to support from older Canadians galvanized by U.S. President Donald Trump's trade war."
- Trudeau-resignation explanation: The leadership reset alone moved roughly 6–8 points (Angus Reid's January Carney-vs-Freeland split). Trump explains perhaps another 15–18 points. The residual — perhaps 4–6 points — is plausibly attributable to the herding/honeymoon convergence that Australian post-2019 analyses identified.
7. The Carney honeymoon is *typical*, not exceptional
Pass 4 found Angus Reid Institute's April 27, 2026 one-year-on tracking:
| PM | One-year approval |
|---|---|
| Chrétien | 66% |
| Trudeau (Justin) | 65% |
| Carney | 58% |
| Harper | 55% |
| Martin | 51% |
| Mulroney | 43% |
| Trudeau (Pierre) | 34% |
Carney's 58% one-year approval is in the middle of the historical distribution — below the two-thirds enjoyed by Chrétien and J. Trudeau, above Harper, Martin, Mulroney, and P. Trudeau. The 8–14 point Liberal lead in May 2026 is large but consistent with these baselines. The "Carneymania" framing (Fournier in The Walrus; The Globe's "Finance Daddy" headline) is journalistic excess, not statistical anomaly. This deflates one of the threads we wanted to chase.
8. The prairie polling gap is real, persistent, and structurally remediable
Janet Brown Opinion Research (Calgary), using Trend Research (Edmonton) for 40% landline / 60% cell hybrid telephone — "Albertans calling Albertans" — has the strongest track record in Alberta polling. She nailed the 2023 Alberta UCP win; her 2015 work in Orange Chinook is the canonical post-mortem on the 2015 Alberta NDP polling failure. Probe Research plays an analogous role in Manitoba.
The Pass 1–3 finding that 2025 polls under-estimated CPC support in the Prairies by ~6.3 pts is consistent with what province-specific pollsters have been saying for a decade: national online panels are systematically under-recruited in rural Prairie subsamples, and the IVR firms (Mainstreet) that historically corrected for this have lost market share. The structural answer is not herding but cost: a properly stratified Prairie sub-sample of 800 respondents costs roughly the same as a national online sample of 1,500, and only Janet Brown and Probe routinely commission it.
9. The microtargeting / Cambridge Analytica angle is *legally* dead but *empirically* alive
The OPC has now formally closed (May 13, 2021) the complaint against the Liberal, Conservative, and NDP parties on the grounds that PIPEDA does not apply to non-commercial activity. Bill C-76 added a toothless "publish a privacy policy" requirement. Bill C-65 (active in 2024–25) and Bill C-4 (passed) consolidated a "very weak regime of self-regulation," in the words of the IAPP Canada Symposium 2026 framing. The 2019 AggregateIQ (AIQ, Victoria-based, Cambridge Analytica-linked) joint federal/BC investigation found AIQ's safeguards inadequate and exposed the personal information of 35 million people — yet produced no party-level enforcement.
The publishable angle here is not "Cambridge Analytica in Canada" — it is "Canada has refused to regulate political parties as data processors, and Mark Carney's Liberal Party operates a voter-data infrastructure of unknown scope while exempt from PIPEDA." Bennett's body of work is the citation backbone.
10. International comparison reveals what Canada is missing — concretely
| Country | Regulator | Key rule | What Canada lacks |
|---|---|---|---|
| UK | British Polling Council (since 2004) + Market Research Society | Mandatory uncertainty disclosure since 2017: "there is a 9 in 10 chance that the true value of a party's support lies within 4 points of the estimates... a 2 in 3 chance within 2 points"; 28-day pre-election heightened disclosure; 18-hour publication rule | Mandatory headline uncertainty interval; pre-election heightened-disclosure window; complaints adjudication that can compel data release |
| Germany | Voluntary association (Forsa, Allensbach, Forschungsgruppe Wahlen, Infratest dimap, INSA) + Europe Elects accountability tracker | RMSE of ~1.4 percentage points across 2021 federal election final polls; IfD Allensbach at SSD 4.65, FGW at 6.15 | Mixed-mode probability sampling; public RMSE tracker by firm |
| Australia | AMSRO + Statistical Society of Australia (post-2019 Inquiry into the Performance of the Opinion Polls at the 2019 Australian Federal Election) | Inquiry verdict: a "polling failure," not a "polling miss" — errors statistically significant, same direction, attributable to polls themselves, not late shifts. 10 recommendations including an enforceable Code of Conduct for Election Polling and consistent disclosure standards | A retrospective inquiry analogous to the Sturgis 2015 UK inquiry or the AMSRO 2019 inquiry — Canada has never had one |
| New Zealand | Self-regulated (Curia, Roy Morgan, UMR) | Smaller market; no formal regulator | Comparable to Canada — not a model |
| France | Commission des sondages | Pre-election publication restrictions during last days of campaign | Different constitutional model |
The UK BPC's enforceable disclosure regime is the single most transferable element. The Canadian Research Insights Council (CRIC) is structurally analogous to the BPC but functionally weaker on three points: (1) no mandatory uncertainty statement on every published poll, (2) no obligation to acknowledge or repudiate privately commissioned polls when they enter the public domain, (3) no audit-with-publicized-findings mechanism. The BPC's recent (2024) ruling on the YouGov / Conservative Britain Alliance poll — declining to compel disclosure of the commissioner's identity but publishing the ruling — is a model of transparency Canada has never matched.
11. The Earnscliffe/Herle 2004 PACP record is the historical analogue worth re-reading
The April 2004 Public Accounts Committee testimony — David Herle (Earnscliffe partner, Liberal campaign co-chair), Terrie O'Leary (Martin's executive assistant, Herle's partner), Allan Cutler (whistleblower public servant), Chuck Guité (sponsorship-scandal central figure) — establishes the documentary template for how Canadian POR-contract capture has been investigated when investigated. The Auditor General's 2003 chapter 5 (paragraph 5.41 of the November 2003 report) found 14% of POR reports were not deposited with National Library and Library of Parliament; Earnscliffe was retained orally for "real-time research or moment-to-moment research" — the same modality, philosophically, that synthetic-respondent generation now offers at scale and at near-zero marginal cost. What was scandal in 2003 (oral, undocumented public-opinion research delivered to a sitting minister with a leadership campaign chaired by the contractor) is, in 2026, a foreseeable architecture under any unregulated synthetic-respondent regime. That parallel is the strongest single rhetorical frame for a publishable piece.
- The Grenier defense is real and must be acknowledged. The 2025 Canadian federal election polls performed within the historical range — better than the 2019 Australian, 2016 US, and 2015 UK polls. Any piece that elides this credibility will be torn apart in the first comment thread. The frame must be architecture, not accuracy.
- The synthetic-respondent angle remains speculative on the operational side. Pass 3 identified the Léger Smart Persona / Plus Compagnie / Google Vertex AI commercial announcements. There is no public evidence that synthetic respondents have been used to generate (rather than augment qualitative work or pre-test questionnaires for) published federal voting-intention polls. The Bisbee 2024 Political Analysis finding is the methodological reason to expect no firm will openly do so for voting-intention: variance collapse makes it statistically detectable.
- The microtargeting angle has been litigated to a structural dead-end (PIPEDA non-applicability, Bill C-76 toothlessness, OPC complaint closure 2021). It is real but it is political (legislative reform) rather than journalistic (exposure) territory.
- The Battle River–Crowfoot model error (Fournier projected 73%, Poilievre got 80.9%) is one data point. It is consistent with the 338 model's structural feature (national-trend signal pull) but is not by itself a falsification. The systematic claim requires the next 3–5 by-elections.
- Several leads in the Pass 4 outreach list will not respond. Cold-emailing 10 researchers yields perhaps 3–5 substantive replies; this is normal. The institutional-affiliation angle (proposing collaborative work with C-Dem, or co-authorship with Daoust/Bennett, rather than interview requests) raises the response rate materially.
- The "Canadian FiveThirtyEight" gap is real but not necessarily the user's to fill. TheWrit + 338Canada cover ~70% of the function; the missing 30% (investigative quantitative reporting on the industry itself, not just the horse race) is a journalistic role that has never had a sustaining funder in Canada. The Walrus, Maclean's, and Hill Times pay modestly; The Tyee has appetite but no quantitative track record; CBC has lost its appetite for adversarial methodology coverage since the EKOS-federal-contract controversies of the 2000s. The realistic frame is a single major published piece, not a sustained beat.
- Anti-SLAPP protects the journalist; it does not protect against the pollster simply refusing to give interviews. The Mainstreet / Illumina / Postmedia 2018 litigation produced no precedent because it settled. The major Canadian pollsters will likely decline interviews — which is itself part of the story.
PASS 4 — The Researchers, The Natural Experiments, and the Threshold of a Canadian Polling Scandal
"Public opinion does not exist." — Bourdieu, 1972
"The medium is the message." — McLuhan, 1964
"All polls are subject to a wide range of potential sources of error." — British Polling Council mandatory disclosure statement, since 2017
Details
Reference Table A — Canadian researchers and journalists working this terrain
| Name | Institution | Specialism | Reachability | Why now |
|---|---|---|---|---|
| Jean-François Daoust | U. Sherbrooke; honorary Edinburgh | Exit polls, strategic voting, Quebec | JF.Daoust@USherbrooke.ca | Dec 2024 CJPS exit-polls paper; willing to be cited; gave CBC West of Centre interview spring 2025 |
| Laura Stephenson | Western U. | CES PI; C-Dem co-lead | Western political science directory | Holds the $2.5M SSHRC partnership grant 2019–26; renewal cycle 2026 |
| Allison Harell | UQAM | CES co-PI; political psychology | UQAM directory | Democracy Checkup co-author |
| Daniel Rubenson | TMU (Toronto Metropolitan) | CES co-PI; experimental methods | TMU directory | Quant rigour anchor of CES |
| Peter Loewen | Munk / UPenn | CES co-PI; behavioural politics | upenn.edu / utoronto.ca | High-profile public commentator |
| Stuart Soroka | UCLA (formerly McGill) | Media–opinion feedback; agenda-setting | snsoroka.com | Past CES co-investigator; Environics Institute fellow |
| Patrick Fournier | U. Montréal | 2015 CES PI; partisanship | umontreal.ca | Less publicly visible but central |
| Éric Bélanger | McGill | Quebec polling; Le nouvel électeur québécois (2022) | mcgill.ca | Co-author with Daoust |
| Frédérick Bastien | U. Montréal | Political communication; horse-race coverage; "Follow the Pollsters" (2013) | f.bastien@umontreal.ca | Only Canadian author of media-pollster-distortion paper |
| Anne-Marie Gingras | UQAM | Media and politics | UQAM directory | Edited Histoires de communications politiques |
| Christopher Cochrane | U. Toronto Scarborough | Political behaviour, computational | utoronto.ca | Adjacent; not focused on polls |
| Daniel Simpson | U. Toronto (Statistical Sciences) | Bayesian methodology; MRP with structured priors (co-author with Gelman) | utoronto.ca | The Canadian inside the global frontier |
| Colin Bennett | UVic emeritus | Data-driven elections; party privacy | colinbennett.ca | Testified PROC Nov 2024; actively consulting |
| Éric Grenier | The Writ.ca / CBC | Polling analysis; aggregation | @EricGrenierTW; thewrit.ca | The industry's most credible defender — must engage |
| Philippe J. Fournier | 338Canada / Le Walrus / L'actualité | Aggregation; seat projection | 338canada.com | Has published pollster ratings — bridge to disclosure conversation |
| Susan Delacourt | Toronto Star; Carleton | Political marketing history | Toronto Star byline | Shopping for Votes author; likely commissioner of deep piece |
| Justin Ling | Independent / Bug-eyed and Shameless | Investigative; misinformation | justinling.ca | Could amplify; not lead |
| Janet Brown | Janet Brown Opinion Research (Calgary) | Alberta-specific phone polling | @planetjanetyyc / planetjanet.ca | Strongest non-online track record in Canada |
| Probe Research | Winnipeg | Manitoba polling | proberesearch.com | Prairie sub-sample specialists |
| Bryan Breguet | Tooclosetocall.ca | Aggregation (Francophone) | @2closetocall | Alternative to 338 |
Reference Table B — International researchers worth engaging
| Name | Affiliation | Why they matter |
|---|---|---|
| Andrew Gelman | Columbia | MRP foundational; would respond to a serious Canadian methodological question |
| James Bisbee | Vanderbilt | Political Analysis 2024 synthetic-respondent critique |
| Lisa Argyle | Purdue (formerly BYU) | Silicon samples founder; now Co-PI Cooperative Election Study |
| Roberto Cerina | U. Amsterdam | March 2026 RSOS "Artificially Intelligent Opinion Polling" — operational frontier |
| Raymond Duch | Nuffield, Oxford | Cerina co-author; runs Centre for Experimental Social Sciences |
| Courtney Kennedy | Pew | VP Methods and Innovation; ATP architect; March 2024 piece on online-opt-in failure modes |
| Scott Keeter | Pew Senior Survey Advisor | Methodology dean of American polling |
| Patrick Sturgis | LSE | Led 2015 UK polling inquiry — direct precedent for what a Canadian inquiry would look like |
| Will Jennings | Southampton | UK polling-academia bridge |
| Doug Rivers | YouGov / Stanford | YouGov methodology |
| Kevin Bonham | Independent (Tasmania) | The Australian Bonham 2019 herding analysis is the closest methodological parallel to what you have been doing informally |
Reference Table C — International comparative regimes
| Jurisdiction | Disclosure floor | Uncertainty statement | Pre-election rules | Inquiry tradition | Transferable to Canada? |
|---|---|---|---|---|---|
| UK (BPC) | Mandatory tables, weighting, mode, dates, sample size, full question wording | Mandatory headline uncertainty since 2017 | 28-day heightened disclosure; 18-hour publication rule | Sturgis 2015 inquiry → 11 recommendations | YES — model for CRIC reform |
| Germany | Voluntary association coverage; methodology transparency variable | Inconsistent | No statutory blackout | Election-by-election RMSE published | Partly — RMSE tracker is most transferable |
| Australia (AMSRO) | Post-2019 enforceable Code of Conduct recommended | In draft Code | None | AMSRO 2019–2022 inquiry — direct precedent | YES — the procedural template for a Canadian inquiry |
| France | Commission des sondages registers polls | Required | Pre-election publication restrictions | Long tradition | Constitutional differences limit transfer |
| US (NCPP / AAPOR) | AAPOR transparency initiative; voluntary | Voluntary uncertainty statements | No statutory rules | Multiple ad hoc post-mortems (2016, 2020) | Limited — too different |
| New Zealand | Light-touch | Inconsistent | None | None | No — not a model |
Reference Table D — By-election natural experiments 2023–2026
| Riding | Date | Pre-election riding poll? | Outcome vs national-swing-based projection | Direction of error | What it tells us |
|---|---|---|---|---|---|
| Winnipeg South Centre | June 19, 2023 | No public | LPC hold, narrow | — | First "longest-ballot" event |
| Calgary Heritage / Durham 2024 | 2024 | No public | CPC hold | — | Baseline |
| Toronto–St. Paul's | Jun 24, 2024 | None | CPC flip by 590 votes | Polls under-stated CPC strength in safe-LPC urban riding | The probability-panel-deficit signature |
| LaSalle–Émard–Verdun | Sep 16, 2024 | None | BQ flip; LPC < 1 pt back | Polls under-stated BQ recovery in Quebec | Quebec sub-sample weakness |
| Elmwood–Transcona | Sep 16, 2024 | None | NDP hold (narrow over CPC) | Polls roughly captured | Floor under NDP held |
| Cloverdale–Langley City | Dec 16, 2024 | None | CPC 66.5% in three-way swing | Coincided with Freeland resignation | Captures non-stationary period |
| Bay of Quinte (Brian Boyer, etc.) | — | various | various | — | Provincial; less useful |
| 2025 general | Apr 28, 2025 | Extensive | LPC plurality minority | National accuracy good; CPC under-estimate ~3pts on prairies (–6.3pts in some regional cuts) | The Grenier "good job" verdict |
| Battle River–Crowfoot | Aug 18, 2025 | 338 model only | Poilievre 80.86% (model projected 73%) | Model under-stated CPC in CPC stronghold | National-trend signal over-correction in safe-riding model |
| Future 2026 by-elections (Scarborough Southwest, etc.) | 2026 | Watch | Watch | Watch | The next data points |
Part C — Specific new threads, revisited
C1. Did the polls make Carney?
The evidence is moderate-to-strong. Pre-candidacy pollster framing (Abacus Jan 27, 2025; Angus Reid Jan 27, 2025; Leger Feb 11, 2025) constructed Carney's "expansion potential" before his campaign existed. Abacus's own post-hoc language — "unbranded product in a damaged category" — confirms the framing was native to the pollsters' market-research lens. But the deeper Bourdieu reading is that the polls did not invent Carney; they revealed and amplified a pre-existing elite preference within the Liberal donor and Bay Street networks for which polling is the legitimating instrument. The pollster role was ratification, not invention. McLuhan's "the medium is the message" applies precisely here: the form of the poll — the willingness/ability to hypothetical-test candidacies — is what made Carney visible as a contender months before any candidate would normally be tested. That is a feature of online-panel infrastructure and short field times; it is impossible under traditional probability-sample methodology.
C2. The Trump deus-ex-machina is half the story
The 29-point swing from January to April 2025 decomposes approximately as: Trudeau-resignation shock ≈ 6–8 pts; Trump inauguration + tariffs + 51st-state rhetoric ≈ 15–18 pts; NDP collapse-to-Liberal ≈ 5–7 pts (Theconversation.com / Stephenson et al. on the CES 2025 NDP-partisan-defection pattern: "more than 80 per cent of NDP supporters voted for their own party in 2021, [but] a majority of NDP partisans voted for the Liberals in 2025"); residual herding/convergence ≈ 2–4 pts. The Trump shock is large but not unique — comparable to the post-9/11 American Bush bump. What is unprecedented is the speed: from –27 to +0 in roughly twelve weeks. Online-panel infrastructure makes such rapid swings measurable, which is not the same as makes them real. They are both.
C3. The Carney honeymoon is statistically ordinary
See Reference Table above. Carney's 58% one-year approval is below the Chrétien/Trudeau levels and above Harper/Martin/Mulroney/P. Trudeau. The Abacus +22 net favourability and Liberals' largest-since-leader 7-point lead (44–37, April 2026) are notable but not anomalous. The "Carneymania" framing is journalistic and pollster-promotional.
C4. The microtargeting/Cambridge Analytica frame is the wrong frame
The OPC has formally closed this avenue (May 2021 file closure). The right frame is Bennett's frame: Canadian political parties operate the largest unregulated voter-data infrastructure in the Western democratic world. The connection to polling is that pollsters' rich attitudinal data merges, at the party level, with the voter rolls (provided by Elections Canada) and the parties' own canvassing CRM — without statutory governance. The publishable piece here is "Mark Carney's Liberal Party Is Running Canada's Largest Unregulated Data Operation," and the legal experts to call are Colin Bennett (UVic), Michael McDonald (UVic Law), Teresa Scassa (uOttawa CRC in Information Law), and Vivek Krishnamurthy (CIPPIC, U. Colorado now but Canadian network intact).
C5. The Prairie polling gap is solvable; nobody is solving it
Janet Brown's "Albertans calling Albertans" model — 40/60 landline/cell, five callback attempts, regionally-resident interviewers — is more accurate than any national online panel for Alberta provincial elections (her 2023 final poll nailed UCP victory; her 2019 work documented in Orange Chinook). The reason national pollsters do not replicate this for federal polling is cost: a 600-respondent Alberta sub-sample at her quality runs ~$25–35K, which no national-poll budget supports as a regular line item.
Part E — When does pollster capture become a story?
The threshold of evidence for a Globe front page
A Globe/Star front-page polling-industry story requires three elements:
- A named contract or document. ATIA-released or leaked. Treasury Board POR Directorate contract numbers naming "synthetic," "AI persona," "Smart Persona," "Plus Compagnie," "LLM," or "ChatGPT" would suffice. So would a Liaison Strategies / NEPMCC SOW with a federal cabinet office.
- A named methodological failure with a quantified impact. Not "the polls under-estimated the CPC" — "On three consecutive elections (2019, 2021, 2025) the Canadian polling industry's final-week aggregate has under-estimated CPC vote share by an average of X points, equivalent to Y seats. This pattern is statistically distinguishable from random error at p<0.05." You can compute this; Grenier's published election-by-election data permits it.
- A named conflict of interest in the regulator. CRIC's audit-but-don't-publicize structure is exactly the analogue of the BPC pre-2017 — which Sturgis recommended (and BPC adopted) be replaced with mandatory disclosure. The story is "Canada's polling regulator is structured to suppress, not surface, methodological failure."
The libel/defamation exposure is moderate. Anti-SLAPP legislation in Ontario (s. 137.1 Courts of Justice Act) is broadly favourable to public-interest reporting; the Bernier v. Kinsella 2021 ONSC 7451 ruling (citing 1704604 Ontario Ltd. v. Pointes Protection Association) establishes that even sharp characterizations of public figures' political conduct, where based on real underlying conduct, will likely survive anti-SLAPP screening. The Mainstreet/Postmedia/Illumina 2018 litigation pattern is instructive: defamation claims by pollsters against newspapers (and vice versa) have not produced any award-of-damages precedent against a journalist who accurately reported methodological criticism.
The Cambridge Analytica moment that is not coming
There will not be a Cambridge Analytica moment in Canadian polling because:
- The Canadian Privacy Commissioner has no jurisdiction over political parties under PIPEDA.
- No Canadian pollster is currently exposed to a Brexit-style cross-border data-flow scandal.
- The synthetic-respondent technology is being deployed by major firms (Léger + Plus Compagnie / Google Vertex AI) openly — Pass 3's finding. There is nothing covert to expose.
What is coming, plausibly within 24 months: the first Canadian pollster to publish a paper (or be exposed) using synthetic respondents to augment a small live sample without disclosing the augmentation. That is the publishable threshold. Pew's Kennedy team is the global voice on this — a quote from her in a Canadian piece would carry decisive weight.
The publishable architecture
A 4,000–6,000-word piece for The Walrus or Globe and Mail opinion section should be structured:
- Lead with Toronto–St. Paul's 2024. The 590-vote upset, the absent riding poll, the architectural explanation (probability-panel deficit).
- Pivot to the CRIC vs BPC structural comparison. UK mandatory uncertainty statement; Canada has none. UK published Sturgis 2015 recommendations; Canada has had no equivalent inquiry.
- Earnscliffe 2004 as the historical template. Auditor General's 14% missing-deposit finding; Herle testimony; what the architecture of capture looked like the last time it was investigated.
- The synthetic-respondent transition. Léger / Plus Compagnie / Google Vertex AI is the named public deployment; Bisbee, Argyle, Cerina cited; the question "what disclosure rule would require this to be marked on a poll?" raised explicitly.
- The Bennett microtargeting frame. Political parties as the largest unregulated data infrastructure in Canadian democracy.
- The recommendation: A Canadian polling inquiry on the AMSRO 2019 or Sturgis 2015 model, jointly convened by the Royal Society of Canada and the Canadian Political Science Association, with mandate to publish disclosure recommendations enforceable through Treasury Board POR Directorate contracting standards.
Recommendations
(a) Specific researchers/journalists to reach out to — priority order
- Jean-François Daoust (JF.Daoust@USherbrooke.ca). Subject: "Pass 4 inquiry on CJPS 2024 exit-poll methodology and Canadian polling-industry transparency." He is the single most likely to engage, methodologically literate, and recently published in this exact space. Lead with his exit-poll paper; pivot to herding and CRIC.
- Colin Bennett (via colinbennett.ca contact form). Subject: "Data-driven elections architecture and the Bill C-65 / political-party privacy gap." He testified to PROC in November 2024 and is now actively consulting. The microtargeting thread is his territory.
- Susan Delacourt (Toronto Star byline; Carleton journalism faculty). Subject: "Updating 'Shopping for Votes' for the Carney era — would you be interested in a freelance pitch?" She is the journalist most likely to commission (not just write) this piece.
- Philippe J. Fournier (338Canada contact). Subject: "The pollster-ratings page and the case for mandatory uncertainty disclosure." He has gone the furthest publicly on naming firm-level performance differentials; he is the most likely Canadian to acknowledge a BPC-style disclosure reform.
- Laura Stephenson (Western political science directory). Subject: "C-Dem partnership grant renewal — is the missing Canadian probability panel a 2026–33 deliverable?" The renewal cycle is the structural opportunity.
- Andrew Gelman (gelman@stat.columbia.edu; the blog is a back-channel that works). Subject: "Has anyone applied MRP with structured priors to Canadian polls?" He will respond if the question is methodologically specific.
- Roberto Cerina (r.cerina@uva.nl). Subject: "AI-augmented MrP applied to Canadian federal polling — feasibility?" He is the operational frontier.
- Janet Brown (@planetjanetyyc; planetjanet.ca contact). Subject: "The national-vs-prairie polling gap — would you be willing to be interviewed on the record?"
- Éric Grenier (@EricGrenierTW; Substack DM). Subject: "A response piece to 'Believe the Polls' from a methodological-skeptic perspective." Engage him as the most credible defender; let him steel-man.
- Courtney Kennedy / Scott Keeter at Pew. Subject: "From the American Trends Panel perspective: what would a Canadian equivalent require?" A 30-minute interview by phone is realistic.
(b) ATIA requests to file — priority order
- PSPC POR Directorate, FY 2024–25 and FY 2025–26: "All contracts, statements of work, and deliverables in which the words 'synthetic,' 'silicon sample,' 'AI persona,' 'large language model,' 'LLM,' 'Smart Persona,' 'Plus Compagnie,' 'GPT,' or 'Vertex AI' appear." This is the smoking-gun ATIA. Expected return: 30–90 days; partial redactions likely; even a nil-response is journalistically interesting.
- Canadian Heritage (PCH), NEPMCC grant file 2020–2026: "Complete grant file including grant agreements, lobbying contacts logs, performance reports, third-party recipient lists, and any communications referencing the Liaison Strategies organization." Pass 3 mapped 16 active lobbying registrations targeting PCH and PMO; this request operationalizes that map.
- Privy Council Office (PCO) Communications and Consultation Secretariat: "All public-opinion-research synthesis reports prepared for the Prime Minister's Office or Clerk of the Privy Council during the period January 1, 2025 – December 31, 2025." The Trump-shock transition period. The deltas between what PCO saw and what was publicly polled is the public-interest question.
- Treasury Board POR Directorate: "All methodology guidance documents, internal audits, and contract performance reviews referencing 'online panel,' 'non-probability sample,' 'weighting matrix,' or 'house effects' from 2020 to present."
- Elections Canada: "All correspondence with the Office of the Privacy Commissioner regarding political-party data practices since the C-76 commencement."
(c) Natural experiments to track in next 12 months
- 2026 Quebec general election (expected fall 2026 or spring 2027). Quebec is the polling industry's traditional Achilles' heel; Léger has dominated and will be tested. Watch in particular for the BQ federal-to-PQ provincial transfer signal.
- 2026 Manitoba and Saskatchewan provincial activity. Probe Research and Janet Brown analogues; Prairie sub-sample integrity.
- Any 2026 federal by-elections. The Carney majority-via-byelection story (Scarborough Southwest April 2026 already; more expected) gives at least 2–3 free riding-level experiments.
- The Battle River–Crowfoot model post-mortem. Specifically, does 338Canada or Grenier publish a revised model that down-weights national-trend signal in safe ridings? If not, the systematic miss compounds.
- The 2027 Ontario provincial election. Mainstreet's IVR rebound (Pass 2) gets a definitive test.
- The first publicly disclosed Canadian synthetic-respondent poll. Whichever firm publishes first (or is exposed first) is the publishable inflection point. Watch Léger announcements quarterly.
- The BPC 2024 UK general election post-mortem fallout in Canada. The May 6, 2025 BPC/MRS public event on the 2024 over-statement of Labour vote share will produce recommendations; CRIC's response (or non-response) is a tell.
- Any C-65 / political-party privacy legislation movement. Bennett's PROC testimony is the live wire.
(d) Bedrock assessment
- TRUE BEDROCK on:
- NEAR-BEDROCK on:
- PASS 5 IS WORTH RUNNING — ON A NARROW PERIMETER. Three things justify a Pass 5:
What Pass 5 should not attempt:
- More breadth on Canadian academics — the network is now mapped.
- More international comparison — UK/Germany/Australia/France are sufficient.
- More speculation on synthetic respondents without documents.
The decisive question for Pass 5 is documentary, not theoretical. Do you have evidence of a specific contract, a specific synthetic-respondent dataset, or a specific undisclosed methodology change? If yes, write the piece. If no, the Pass 4 frame ("Canada has an opaque polling-industry structure that is statistically defensible but architecturally vulnerable") is the publishable ceiling without the documentary smoking gun.
81 facts · 31 assertions → Canadian polling industry · The Walrus · UK's BPC · AAPOR · 338Canada · Canadian academy · Éric Grenier · Philippe J. Fournier. Every one is a verbatim span; nothing was paraphrased into the graph.
This is a signed piece; its findings carry their sources inline, in the text. The piece argues; the sources carry the proof.