Why We Weight by Past Vote
The central challenge in modern polling is not simply measuring opinion. It is ensuring that the people we hear from reflect the people we do not.
We have weighted by past vote for 15 years. We haven't changed our methodology. We don't plan to.
Some people assume that when pollsters debate methodology, there's a partisan agenda hiding underneath. I want to be direct: that's not what's happening here, at least not from where I sit.
I care about how we weight surveys because accuracy is the whole point. If our samples systematically over-represent certain kinds of voters and under-represent others, we aren't measuring public opinion.
We're measuring a distorted version of it.
And a distorted read, regardless of which direction it tilts, leads everyone astray: governments, journalists, advocates, and citizens trying to understand where their country actually stands.
Getting the method right isn't about helping one party or hurting another. It's about earning the right to say: this is what Canadians actually think.
A structural challenge every pollster faces
"Are the people answering my surveys like the people who are not?"
— The question that always sits in the back of my mind as a pollster
Polling today is hard. Not because people are doing anything wrong, but because the environment we are all operating in has changed in ways that I think introduces persistent, structural bias into our samples.
The bigger issue is how people are recruited and who ends up in those panels. That process is not random, and it introduces systematic differences that are not evenly distributed across the electorate.
Telephone surveys, which were once considered the gold standard, now face response rates so low that they introduce their own form of bias. We are no longer reaching a broad cross-section of the public. We are reaching a subset of people willing to answer unknown calls and stay on the line.
Most pollsters, including us, weight by demographics such as age, gender, region, and education. That helps. But over time, we have come to believe that it is not enough. That is why we adopted past vote weighting 15 years ago, and it is why we have never stopped using it.
Looking at behaviour, not just characteristics
If demographic weighting cannot fully correct for imbalance, what can? One answer is to look at behaviour, not just characteristics.
Asking people how they voted in the last election gives us a benchmark that reflects the actual electorate. It allows us to compare the partisan composition of our sample to a known outcome and adjust accordingly.
If a group of voters is systematically underrepresented in your sample, and that underrepresentation is not fully explained by demographics, then weighting to match the actual distribution of past vote can help correct for that gap.
43 surveys. Five election cycles. One consistent pattern.
We have been tracking recalled past vote since 2011. I took a sample of our surveys from May 2014 and April 2026 to conduct this analysis. Each survey asks respondents how they voted in the most recent federal election. Before we look at those answers, we weight the data by region, age, gender, and education to reflect the Canadian population. But importantly, we do not apply any past-vote weighting at this stage of our analysis.
That is deliberate. It allows us to see what a standard, demographically weighted sample looks like on its own, without any correction for partisanship.
In every single wave, across every single election cycle, Conservative voters are underrepresented in our demographically weighted sample relative to their actual share of the vote. Not in most waves. Not in some elections. In every case we can observe.
Two things stand out. First, this pattern is consistent. It is not driven by one election, one leader, or one moment in time. It holds across very different political contexts.
Second, the gap appears immediately after each election and persists throughout the cycle. If this were primarily a problem of memory, we would expect the gap to grow over time as recall becomes less reliable. But that is not what we see. The bias is present early and remains relatively stable within each parliament.
That suggests the issue is not primarily about what people remember or report. It is about who is in the sample in the first place.
The empirical answer is yes
Looking at our final polls in the last three federal elections, applying past vote weighting increased Conservative vote estimates, reduced Liberal vote estimates, and brought our numbers closer to the actual result.
| Election | Party | Without Past Vote Weight | With Past Vote Weight | Actual Result |
|---|---|---|---|---|
| 2019 | CPC | 30% | 33% | 34.3% |
| 2019 | LPC | 36% | 34% | 33.1% |
| 2021 | CPC | 26% | 33% | 33.7% |
| 2021 | LPC | 37% | 33% | 32.6% |
| 2025 | CPC | 35% | 39% | 41.2% |
| 2025 | LPC | 45% | 41% | 43.7% |
In every election, past vote weighting moved our Conservative estimates upward and our Liberal estimates downward — consistently in the direction of the actual result. The 2021 election shows the most dramatic correction: a 7-point improvement in our Conservative estimate.
When the Polls Disagreed: The 2025 Canadian Federal Election
The 2025 election produced one of the most dramatic polling stories in Canada's political history — not because the outcome was surprising, but because different firms, measuring the same electorate at the same moment, arrived at conclusions sometimes twelve points apart.
Remarkable Agreement
In early January, on the eve of Justin Trudeau's resignation, every major polling firm was measuring a Conservative lead of between 25 and 29 percentage points. Whatever methodological differences existed, they were not producing meaningfully different pictures.
A Compressed Political Shift
As Trudeau stepped aside, Mark Carney entered the Liberal leadership race, and Trump escalated tariff threats, the Conservative lead eroded across every tracker. By March 23, the average showed a Liberal lead of approximately five points. The direction was clear. The magnitude was not.
Pollster 1 was the clearest outlier on the Liberal-friendly side, averaging approximately seven percentage points above the consensus. In early April, while the average suggested a Liberal lead of around eight points, Pollster 1 was measuring fourteen.
At the other end, we and Pollster 2 were consistently measuring five to six points below the consensus. I raised what might be explaining this discrepancy in the lead up to the election.
The remaining firms — Pollster 3, Pollster 4, Pollster 5, and Pollster 6 — clustered closer to the average, though with their own tendencies.
Did pollsters converge as election day neared?
Liberal–Conservative margin (LPC% − CPC%) by pollster, January–April 2025. Actual result: Liberal +2.5. Dates are approximate for pre-campaign polls.
The actual result — a Liberal margin of 2.5 points — sits below where even the most polling firms were measuring in the final weeks of the campaign. Did the electorate genuinely move toward the Conservatives in the final stretch? Or were even the tightest-reading firms still overcounting Liberal support due to persistent panel composition issues?
The pattern of the data favours a version of both. Some real movement likely occurred. But the size of the inter-firm spread, and the consistency with which methodologically similar firms clustered together, points strongly toward structural measurement differences as the primary driver of divergence.
Limitations, honesty, and what this means
It is important to be clear about what past vote weighting can and cannot do. It does not eliminate all sources of error. If there has been real, large-scale vote switching since the last election, past vote weighting can overcorrect by pulling people back toward their previous choice.
We saw this in 2025 with the NDP, where both weighted and unweighted estimates overstated support because of rapid vote movement during the campaign. Past vote recall is also imperfect. Academic research does find that some respondents misremember or reinterpret their past vote.
"I do not believe that other pollsters are purposively doing anything wrong. Every pollster should be trying to solve the same problem with the tools and evidence available to them."
Our analysis suggests that, in the current cycle, a demographically weighted sample could be underrepresenting Conservative voters by about 5 to 7 points. If that structural gap is not corrected, polls will tend to show a larger Liberal advantage than may actually exist in the electorate.
This helps explain why our polling has often shown a tighter race than others. The difference is not what people tell us about their current vote. It is how we adjust the composition of the sample.
We have been weighting by past vote for 15 years. We have not changed our methodology, and we do not plan to. It remains imperfect. But it is, in our view, a better option than relying only on region, demographics, and education. Until we find a better way to correct for a persistent issue we have seen in our data for years, this is how we will continue to do our work.
What the evidence actually shows
The case for weighting by past vote rests on a substantial and growing body of evidence, spread across multiple countries and electoral systems. Taken together, it points in a clear direction: panels that weight to past vote produce less biased estimates, particularly when that past vote data is collected close to the election in question. But the evidence is not uniformly positive, and the conditions under which the technique works matter enormously. Those conditions, examined honestly, actually make a stronger case for the practice in Canada today, not a weaker one.
The United Kingdom: Where the Practice Was Born
The origin story is the 1992 British general election. Every major polling firm had Labour ahead or neck-and-neck heading into election day. The actual result was a Conservative victory by 7.6 points. The Market Research Society inquiry found the polls had been off by roughly 8.5 percentage points in aggregate. About 2% of the error could be explained by Conservative supporters refusing to disclose their voting intentions.
After 1992, most polling companies changed their methodology. Some asked interviewees how they had voted at the previous election, then assumed that those who had voted Conservative before but were now "unsure" would again vote Tory. Others weighted their panel so that past vote distribution matched the actual election result.
The effect on accuracy was visible almost immediately. Post-1992 adjustments yielded vote share predictions within 1-2 points for major parties across 47 polls in the 1997 election. The 2015 election offered a partial counter-lesson: despite widespread past vote weighting, most firms still missed the Conservative majority. But that miss was due to unrepresentative samples rather than the kind of late swing that affected 1992. Past vote weighting does not fix every source of error. It targets a specific problem: the systematic over- or under-representation of a party's supporters in the raw sample.
Australia: The Clearest Experimental Evidence
The most rigorous direct test comes from Pennay, Misson, Neiger, and Lavrakas (2023) in Survey Practice, using data from a probability-based online panel recruited by random digit dialing rather than opt-in. This isolates the effect of weighting methodology from the confounding problem of a self-selected panel.
Their design tested three versions of past vote recall against a baseline weight using only age, education, sex, and geography. The results were decisive. Adding a short-term recall measure of past vote (collected just three months after the 2016 election) cut the weighted average absolute error on the primary vote nearly in half, from 2.58 pp to 1.41 pp, and reduced the average absolute two-party-preferred error from 4.08 pp to 2.41 pp.
A critical secondary finding: the accuracy of recalled past vote diminishes over time. Short-term recall had an average absolute error of 2.3 pp relative to the actual result, increasing to 3.0 pp for long-term recall collected almost three years later. Long-term recall performed no better than the no-past-vote baseline. Short-term recall was the clear winner.
YouGov's experimental work reinforced this. When they reweighted 2017 UK election data using past vote collected immediately after that election, 41% of respondents reported voting Labour. When collected from the same respondents two years later, only 33% recalled voting Labour. The difference in estimated Labour vote share between those two scenarios was three percentage points. The lesson is not that past vote weighting fails. It is that stale recall data undermines it.
The United States: A Late and Reluctant Convert
Past vote weighting was uncommon among American pollsters before 2020. In 2016, just one in four national surveys used either past vote or party affiliation in their weighting scheme. By 2024, about three-quarters of polls weighted on one of these variables.
Pew Research Center's adoption of past vote weighting in mid-2025 is significant precisely because Pew had long resisted it. Their explanation pinpoints why the technique works better now than in the telephone era. Internet panels reduce social desirability bias, meaning respondents are more willing to admit they voted for the losing candidate or did not vote at all. Panel infrastructure also means past vote can be recorded right after an election, dramatically reducing memory error.
Pew went further and validated self-reported turnout against administrative voter file data. The net effect of adding past vote to their weighting scheme was less than 0.5 percentage points on most questions, but consistent in direction: political questions moved slightly toward Republican positions, and personal finance questions shifted slightly toward lower-income respondents. Both are plausible corrections given who was previously underrepresented.
Pew also noted that Republican voters have become less likely to respond to surveys. Before 2016, there was no evidence that polling systematically underrepresented Republican presidential support. A weighting adjustment is most effective when it addresses a consistent nonresponse pattern rather than one that varies from election to election. That consistency is what makes past vote weighting newly defensible in the American context: it corrects for a structural bias, not a one-time anomaly.
Spain and Canada: Standard Practice
Past vote weighting is not a novelty in many countries. In Spain, traditional weighting by vote recall has been shown to help compensate for the logical biases of working with a sample, and is considered particularly valuable in multi-party scenarios. Durand and Johnson's 2021 study in Public Opinion Quarterly, reviewing electoral polling across four countries including Canada, cited past vote as a standard adjustment method in routine use outside the United States.
When it has not worked
The evidence against past vote weighting falls into two distinct categories, and it is worth being precise about what each actually shows.
Stale recall data. As shown by the Pennay et al. findings and confirmed by YouGov's experiments, weighting on long-term recalled past vote can be worse than not weighting at all. This is not an argument against the method. It is an argument for collecting past vote data shortly after each election and maintaining it as a panel variable, which is exactly what modern online panels are positioned to do.
The pre-2016 US environment. Before 2016, there was no systematic partisan nonresponse bias in American surveys. Republicans and Democrats were roughly equally likely to answer polls. When nonresponse is roughly random across parties, weighting to past vote adds variance without reducing bias. You are solving a problem that does not exist. The technique became beneficial precisely when a stable, directional nonresponse pattern emerged.
As YouGov has noted, the argument for weighting to past vote is the same as weighting to demographics like age, race, gender, or education: if your sample is unrepresentative in terms of some characteristic and the people with this characteristic vote differently than the people without it, the sample will give biased estimates.
The pre-2016 US experience is therefore not evidence against past vote weighting as a principle. It is evidence that the method only earns its keep when there is a persistent compositional problem in who is responding to surveys.
The Canadian volatility argument: why it does not hold up
A version of the argument circulates that the current Canadian political moment makes past vote weighting unreliable or even counterproductive. The reasoning goes roughly like this: Canada is experiencing a large, rapid shift in voter orientation, so anchoring the sample to the 2021 result will pull estimates backward and understate the scale of the change.
This argument misunderstands what past vote weighting does. It adjusts the composition of the sample: specifically, the share of respondents who voted CPC, Liberal, NDP, and so on in 2021. It does not assume those people will vote the same way again. Their current vote intention is measured freely after weighting. If a former Conservative voter in your weighted sample now says they intend to vote Liberal, that is captured in the data. What weighting corrects for is whether that former Conservative voter showed up in your sample in the right proportion in the first place.
This distinction matters enormously in a volatile election. Consider what happens if former CPC voters are underrepresented in your panel, a well-documented problem for Canadian online surveys including our own multi-wave panel analysis. Those are precisely the people who are switching. If they are missing from your sample, you do not just misrepresent the CPC's residual support. You also miss the magnitude of the Liberal gain, because you are under-sampling the very people who are doing the switching.
The volatility argument, if anything, makes the case for past vote weighting more urgent, not less. A sample without enough former CPC voters cannot accurately measure the Liberal surge among former CPC voters, because those voters are not there to be measured.
There is a version of the critique that has more merit: in a period of large swings, the composition of the electorate itself changes. New voters enter, infrequent voters are mobilized, and the 2021 electorate may not map perfectly onto the 2025 electorate. But this is a challenge common to all polling, not a specific problem with past vote weighting. And it points toward using past vote as one variable among several, alongside education, age, geography, and turnout likelihood, rather than abandoning it.
The conditions that make past vote weighting most valuable are precisely: a stable directional nonresponse problem (CPC supporters underrepresented in online panels), a known hard benchmark to weight to (the 2025 election result), and short-term recall data collected close to that election. All three conditions apply in Canada today. The volatility of the moment is a reason to get the sample composition right, not a reason to leave it uncontrolled.
References
Pennay, D., Misson, S., Neiger, D., & Lavrakas, P. J. (2023). "How Weighting by Past Vote Can Improve Estimates of Voting Intentions." Survey Practice. Read the paper
Pew Research Center (2025). "Why and How We're Weighting Surveys for Past Presidential Vote." Decoded. Read the article
YouGov (2019). "False Recall, and How It Affects Polling." YouGov Articles. Read the article
YouGov (2025). "Reliability of Retrospective Vote Choice Questions." YouGov Articles. Read the article
Durand, C., & Johnson, T. (2021). "What About Modes? Differences Between Modes in the 21st Century's Electoral Polls Across Four Countries." Public Opinion Quarterly, 85(1), 183-216. Read the paper
Market Research Society (1994). "The Opinion Polls and the 1992 General Election." MRS Inquiry Report. Read the review
Rigour. Accuracy. Trusted interpretation.
Our mission is to understand people and turn that understanding into insight that leaders trust, use, and act on.
Methodological Rigour
We don't just collect data, we interrogate it. Our commitment to debate and discussion reflects a research culture that prioritizes accuracy over convenience.
Proven Accuracy
In the 2025 federal election, our final poll showed a Liberal lead of +2, one of the closest to the actual result of +2.5 among major polling firms. This was not luck. It was the product of a decade of methodological refinement.
Trusted Interpretation
Numbers without context are just noise. We combine rigorous research with clear judgment and influential storytelling to help leaders understand not just what Canadians think, but what it means for their decisions.
Our Research Toolkit
Good decisions require good data
Whether you need public opinion research, market insights, or strategic counsel, Abacus Data delivers the rigour and clarity your decisions demand.