Research · When the polls missed

Shy Tories and a late swing: how the 1992 polls handed Kinnock a lead he never had

Abstract. The final polls put Labour about a point ahead and the Conservatives won Great Britain by 7.6 points, the worst miss on the lead since polling began. The Market Research Society inquiry blamed late swing, quota samples that did not match the electorate and Conservatives who would not say how they would vote.

Polling Forecast · 2 November 2026 · 7 min read · United Kingdom, 1992 general election, polling error, final polls, inquiry, pollster methods

Summary

On the eve of the general election of 9 April 1992, the last poll from each of the five main companies put Labour narrowly ahead: on average Labour 39.4%, Conservatives 38.3%. John Major's Conservatives won Great Britain by 7.6 points, 42.8% to 35.2%, and kept a majority of 21. The polls missed the lead by 8.7 points, the worst miss since British election polling began in 1945. The Market Research Society inquiry of 1994 found three causes of roughly equal weight: a late swing to the Conservatives, quota samples that did not match the electorate, and Conservative voters who would not say how they were going to vote.

What the polls said

The campaign polls showed a dead heat or a small Labour lead for weeks, and the final round did not change that. These are the last polls from the five companies that published on the final days, as listed in the 2016 polling inquiry's appendix of final polls (Gallup's figure is the unrounded 38.5 it published):

Pollster (client)	Fieldwork end	Con	Lab	LD
Gallup (Daily Telegraph)	8 April	38.5	38	20
ICM (The Guardian)	8 April	38	38	20
MORI (The Times)	8 April	38	39	20
NOP (The Independent)	8 April	39	42	17
Harris (ITN)	7 April	38	40	18
Average		38.3	39.4	19.0

That is a Labour lead of 1.1 points. Of the eight polls the inquiry lists for the final week, only Gallup had the Conservatives ahead, by half a point; ICM had a tie and the other six had Labour ahead by 1 to 3 points. Peter Hutton of MORI, writing in the MRS newsletter in May 1992, put the eve-of-poll average at Conservatives 38, Labour 40.

What happened

In Great Britain the Conservatives took 42.8%, Labour 35.2% and the Liberal Democrats 18.3%. Across the UK the Conservatives won 336 seats, Labour 271 and the Liberal Democrats 20, on a turnout of 77.7%.

Against the five-poll average, the Conservatives were understated by 4.5 points, Labour overstated by 4.2 and the Liberal Democrats overstated by 0.7. The error on the lead, 8.7 points, came almost entirely from the two big parties moving in opposite directions. The polls had set up a hung parliament and a possible Labour-led government; the country got a fourth Conservative term.

Why it happened: what the pollsters found

The Market Research Society set up an inquiry within days. The panel was made up mainly of people from the polling companies, unlike the academic panel of 2015. Its initial report, issued about three months after polling day, accounted for more than half of the gap:

a late change in voting intention, measured in recall polls after the election, worth about 3 points;
refusals to give a voting intention of around 5% rather than the usual 3%, coming disproportionately from Conservatives, worth up to 2 points;
de-registration to avoid the poll tax and overseas voters, worth a little over 1 point;
a residual bias that may have existed since 1959, overstating Labour by 1.5 to 2 points and understating the Conservatives by about the same.

The full report, The Opinion Polls and the 1992 General Election (MRS, 1994), settled on three main causes that contributed "more or less equally":

Late swing. Some voters changed their minds after interviewing ended, and Conservative supporters were more likely to turn out on the day. The report judged that late swing, the easiest factor to isolate, accounted for between a fifth and a third of the total gap.
Inadequate quotas and weights. The quota and weighting targets did not reflect the social profile of the electorate closely enough, and the variables used (sex, age, social class) were not closely enough tied to vote choice to make the samples politically representative. The 2016 inquiry adds that the 1992 polls mostly weighted to the 1990 National Readership Survey, which turned out to differ from the 1991 census in some respects.
Differential refusal. Conservatives were less likely to reveal their vote, both by refusing the question and by saying "don't know". The report thought refusal to be interviewed at all probably worked the same way, though it had no firm evidence for that. Because quotas exist partly to absorb that kind of refusal, the weak quotas let it do more damage.

Smaller factors: sampling points that were slightly biased towards Labour at all but one company, and interviews with people not on the register. The inquiry rejected deliberate lying, postal votes (under 2% of voters, split fairly evenly), overseas voters, street versus in-home interviewing, and sample size.

Ivor Crewe's 1993 paper in the Journal of the Market Research Society, "A nation of liars? Opinion polls and the 1992 election", drew on the MRS report. Nick Sparrow of ICM and John Turner of Oxford Brookes University developed the refusal finding into the "spiral of silence" argument in the same journal in 1995: Conservative voters were going quiet, and polls that did not adjust for it would overstate Labour.

What our data shows

Our long-view table holds eight GB polls from the last two weeks of the 1992 campaign. Every one of them understated the Conservative lead, by between 7.1 and 10.6 points, with an average error of 8.8 points. On average they had the Conservatives 4.8 points too low and Labour 4.0 too high. No company came close, which is what made 1992 a shock rather than an embarrassment for one or two firms.

Across the 22 general elections since 1945 in that table, 1992 has the largest average error on the Conservative–Labour lead. The next largest are 2024 (7.5 points) and 1970 (6.5). The direction was not new. Labour's share was overstated at seven of the nine elections from 1959 to 1987, usually by 1 to 4 points, which fits the MRS finding of a long-running residual bias. In 1992 that standing bias met a close race and a late shift, and the errors added up.

The time series argues against late swing as the main story. Polls ending on 24 and 28 March and on 3 April showed the same picture as the final round: level, or Labour ahead by 1 to 6 points. There was no Conservative recovery building in the published numbers during the last fortnight. Any late movement happened after the final fieldwork or was hidden by the samples, which is why the inquiry could only assign it a fifth to a third of the miss.

What the pollsters changed

The 2016 inquiry summarises the response: after 1992 the pollsters introduced procedures to stop understating the Conservatives, chiefly weighting by past vote and reallocating don't knows and refusals. It judges that these were mostly successful at the next four elections.

ICM went furthest. Its "spiral of silence" adjustment added people who would not give a voting intention back to the party they said they had voted for last time, at a weight of 0.5. ICM also weighted to past vote, using a target 80% based on the last election result and 20% on the recalled vote in its own recent polls, to allow for people misremembering. The British Polling Council's 2017 summary shows ICM still using both methods in 2015. By 2018 YouGov told the House of Lords that past-vote weighting was "standard across almost the whole industry".

Some companies also began weighting by stated likelihood to vote. MORI, later Ipsos, never adopted past-vote weighting. Its 2025 review of the 2024 election notes that it still did not use it, and tested whether it would have helped then (it would have helped only a little).

What it means for our forecast

Our UK correction is fitted on the four elections for which pollsters' raw tables exist: 2015, 2017, 2019 and 2024. It is measured in England on the raw answers, before each pollster's turnout and don't-know steps, and we take off half the average past miss, party by party. We keep that only because it would have improved past forecasts made for each election without knowing its result. 1992 is too early for that test, because the raw tables are not available.

What 1992 adds is the long view. The tendency to overstate Labour runs back decades, so a correction in that direction is not a fluke of recent years. And a miss of almost 9 points on the lead has happened, with every poll on the same side. The ranges around our forecast use the full spread of past misses rather than the average, because a 1992-sized error is rare but real. The method and its tests are in Measuring polling error, and each company's record is in our pollster ratings.

Sources

How to cite

Polling Forecast (2026). Shy Tories and a late swing: how the 1992 polls handed Kinnock a lead he never had. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-uk-1992. Live figures change as new polls arrive; the date above is when the text was last revised.