Track Record: 2005 Errors

This page first posted 6 May 2005

This page shows the seats which were incorrectly predicted at the 2005 general election. In total 52 seats were mis-predicted, and the predicted majority (132) was wrong by 66 seats. Although 52 bad seats was about the expected number, the error in the majority is less satisfactory. However, the reasons for the errors are instructive.

This year, we divided prediction errors into four different types of effect:

Opinion poll error
Regional swing effects
Tactical voting effects
Local factors and residual model error

giving the Total error.

In summary, 21 seats were incorrect due to opinion poll bias, 14 seats due to regional effects, 7 due to tactical voting and 32 seats due to local factors and residual model error. But in total only 52 seats were incorrect because in some seats two or more errors cancelled each other out.

1. Opinion poll error

Opinion poll error causes the final opinion polls of the campaign to show different levels of party support than actually happens at the polls in the election. This is in turn will create errors in the prediction.

The causes of opinion poll error include:

Random sampling error
Over- or under-sampling of various classes and groups
Errors in weighting samples to match quotas
Differential turnout
Refusal of respondents to indicate party preference
Misleading answers by respondents

In 2005, the opinion poll error overstated Labour's lead by 2.9%, which is better than previous years, but still significant. As it happens, the new internet pollster YouGov had a very good election. They consistently showed a Labour lead of around 3%-4%, and their methodology has been vindicated by the actual result. Other pollsters were more variable, but errors of 5% were not uncommon.

The opinion poll error of 2.9% caused 21 seats to be incorrectly predicted. This number is calculated by comparing the prediction with two different sets of inputs. The first uses the average of the final opinion polls of the campaign. Those were: Con 31.7%, Lab 37.5%, Lib 22.8%. The second uses the actual support levels in the election of Con 33.2%, Lab 36.2%, Lib 22.6%. Seats which have different predictions under the two sets of inputs are seats predicted incorrectly due to opinion poll error.

In total, opinion poll error understated the number of Conservative seats by 17, overstated Labour by 15, and overstated the nationalist parties (SNP and PC) by 2.

Seat	Poll Pred Win	Poll Pred Maj	Vote Pred Win	Vote Pred Maj
Bexleyheath and Crayford	LAB	0.2%	CON	2.9%
Bristol West	LAB	0.5%	LIB	0.3%
Clwyd West	LAB	0.2%	CON	2.8%
Dorset West	LIB	0.6%	CON	1.9%
Dundee East	NAT	0.9%	LAB	2.1%
Enfield North	LAB	2.2%	CON	1.1%
Forest of Dean	LAB	1.1%	CON	2.0%
Gillingham	LAB	1.8%	CON	1.3%
Hammersmith and Fulham	LAB	1.0%	CON	2.2%
Harwich	LAB	1.7%	CON	1.5%
Hornchurch	LAB	0.5%	CON	2.8%
Ilford North	LAB	1.6%	CON	1.7%
Inverness Nairn Badenoch and Strathspey	LAB	0.7%	LIB	3.5%
Milton Keynes North East	LAB	0.6%	CON	2.4%
Redditch	LAB	3.0%	CON	0.2%
Rugby and Kenilworth	LAB	1.7%	CON	1.4%
Selby	LAB	0.7%	CON	2.5%
Surrey South West	LIB	1.4%	CON	1.1%
Thanet South	LAB	0.9%	CON	2.3%
Wellingborough	LAB	0.9%	CON	2.4%
Ynys Mon	NAT	1.4%	LAB	2.1%

Since the Conservatives were underestimated by the opinion polls, these seats are mostly those the Conservatives should have been predicted to win, but were not.

2. Regional swing

The model assumes that the whole country (ex-Scotland) moves in a similar direction. However regional factors can cause the swing throughout a geographical area to be different from the national average.

We can calculate how many seats were incorrectly predicted due to regional effects. We do this by taking the vote-based prediction above, and applying the actual regional swings to each seat, depending on its region. The seats which change hands are those for which regional swing was crucial.

Seat	Vote Pred Win	Vote Pred Maj	Reg Pred Win	Reg Pred Maj	Region
Bristol West	LIB	0.3%	LAB	2.1%	West
Ceredigion	NAT	1.4%	LIB	0.7%	Wales
Croydon Central	LAB	1.6%	CON	0.4%	South London
Finchley and Golders Green	LAB	1.5%	CON	3.5%	North London
Guildford	LIB	2.1%	CON	2.3%	South
Hemel Hempstead	LAB	1.1%	CON	2.0%	Essex
Orpington	LIB	0.3%	CON	1.7%	South London
Peterborough	LAB	0.4%	CON	0.5%	East Anglia
Putney	LAB	1.1%	CON	0.9%	South London
Selby	CON	2.5%	LAB	1.0%	Yorkshire
Shipley	CON	3.5%	LAB	0.0%	Yorkshire
Shrewsbury and Atcham	LAB	0.4%	CON	0.0%	Severn
Taunton	LIB	0.7%	CON	1.2%	South West
Weston-super-Mare	LIB	2.2%	CON	0.4%	West

The Conservatives polled relatively better than Labour, allowing for national trends, in London and southern England, but worse in the midlands and northern England. This explains the Conservative victories around London, and their losses in Yorkshire.

The Liberal Democrats gained comparatively in parts of northern England, but lost ground in southern England, particularly the South West and South regions. This is probably due to their realignment as a more left-wing party, and it cost them a few seats in the south and west.

Two distinct trends are visible here. Firstly the Conservative vote is polarising. It is becoming even stronger in the places where it is already strong, and even weaker where it is already weak. Secondly, the LibDem vote is shifting. It is moving away from its previous heartland in the "celtic fringes" of the south west and growing in the "old Labour" north.

3. Tactical voting

Despite advance predictions, there does not seem to have been a large amount of either tactical voting or "tactical unwind" at the 2005 election.

We see this by comparing the prediction above (which makes allowance for poll error and regional swing), and applying the tactical voting parameters. The parameters are estimated using our standard analysis (details), and are:

TVLIB: unwind 0.5%, tactical voting against Labour by LibDem supporters
TVLAB: zero, tactical voting to LibDem by Labour supporters
TVCON: +2.5%, tactical voting to LibDem by Conservative supporters

The first of these is the most important but the size of the effect is small, although it shows some tactical unwind. The last of these is least important as there are few seats where it applies, so the suggested level is speculative.

Making the comparison with the section 2 results, we see that seven seats are predicted to change hands due to tactical voting. They are:

Seat	Reg Pred Win	Reg Pred Maj	TV Pred Win	TV Pred Maj
Bristol West	LAB	2.1%	LIB	0.4%
Dartford	LAB	0.1%	CON	0.4%
Enfield Southgate	LAB	0.3%	CON	0.2%
Rochdale	LAB	1.4%	LIB	1.1%
Shipley	LAB	0.0%	CON	0.5%
St Albans	LAB	0.2%	CON	0.3%
Wimbledon	LAB	0.2%	CON	0.3%

Note that these are still predictions and not the actual result. As it happens, Dartford did not change hands, but all the other six predictions were correct.

Local factors and residual error

We have now made allowance for polling error, regional swing, and tactical voting. But there are still 32 seats which are mis-predicted. The possible explanations for each seat are:

Local factors - such as particularly strong (or weak) candidates or campaigns, or local issues which swing voters in a particular direction.
Residual model errors - features of reality which are not captured by the model, such as demographic-based analysis, anti-war hotspots, and so on.
Random factors - for very marginal seats, the final result can have a degree of chance about it, depending on exactly which electors actually vote.

By comparing the results of the prediction model, as adjusted above, with the actual outcome we can see the seats which are subject to these local factors and residual errors.

The table below shows the 32 seats, with the predicted winner and majority against the actual results. Important local or other factors are included as a comment, if applicable.

Seat	MP elected	Pred Win	Pred Maj	Act Win	Act Maj	Comment
Bethnal Green and Bow	George Galloway	LAB	13.5%	MIN	1.9%	Strong ex-Labour candidate
Blaenau Gwent	Peter Law	LAB	44.8%	MIN	25.9%	Labour in-fighting
Brent East	Sarah Teather	LAB	30.4%	LIB	8.7%	By-election hold
Cambridge	David Howarth	LAB	5.4%	LIB	10.0%	University seat
Dartford	Howard Stoate	CON	0.4%	LAB	1.5%	Very marginal
Devon West and Torridge	Geoffrey Cox	LIB	1.0%	CON	5.5%
Dorset South	Jim Knight	CON	1.9%	LAB	3.7%
Dumfries and Galloway	Russell Brown	CON	1.7%	LAB	5.7%
Dumfriesshire, Clydesdale and Tweeddale	David Mundell	LAB	4.8%	CON	3.9%
Dunbartonshire East	Jo Swinson	LAB	2.2%	LIB	8.7%
Dundee East	Stewart Hosie	LAB	1.4%	NAT	1.0%	Very marginal
Enfield North	Joan Ryan	CON	6.6%	LAB	4.7%
Falmouth and Camborne	Julia Goldsworthy	LAB	7.8%	LIB	3.9%
Finchley and Golders Green	Rudi Vis	CON	4.0%	LAB	1.7%
Gillingham	Paul Clark	CON	1.9%	LAB	0.6%	Very marginal
Gravesham	Adam Holloway	LAB	3.1%	CON	1.4%	Marginal
Hornsey and Wood Green	Lynne Featherstone	LAB	3.9%	LIB	5.1%
Leeds North West	Greg Mulholland	LAB	4.7%	LIB	4.2%	Anti-war hotspot
Ludlow	Philip Dunne	LIB	4.0%	CON	4.4%
Manchester Withington	John Leech	LAB	18.7%	LIB	1.8%	Anti-war hotspot
Na h-Eileanan An Iar (Western Isles)	Angus MacNeil	LAB	7.3%	NAT	10.4%
Newbury	Richard Benyon	LIB	0.8%	CON	6.3%
Oldham East and Saddleworth	Phil Woolas	LIB	7.0%	LAB	8.3%
Preseli Pembrokeshire	Stephen Crabb	LAB	1.2%	CON	1.6%	Very marginal
Reading East	Rob Wilson	LAB	4.6%	CON	1.1%
Redditch	Jacqui Smith	CON	1.2%	LAB	6.7%
Scarborough and Whitby	Robert Goodwill	LAB	3.5%	CON	2.7%
Solihull	Lorely Burt	CON	14.8%	LIB	0.5%	Big swing
Taunton	Jeremy Browne	CON	1.2%	LIB	1.0%	Very marginal
Thanet South	Stephen Ladyman	CON	2.9%	LAB	1.6%	Marginal
Westmorland and Lonsdale	Tim Farron	CON	1.0%	LIB	0.5%	Very marginal, decapitation
Wrekin, The	Mark Pritchard	LAB	0.6%	CON	2.1%	Marginal

We can categorise these seats into four types:

Marginals. Nine seats are either very marginal (swing error less than 1.25%) or marginal (swing error less than 2.5%). These seats are intrinsically difficult to predict as they can literally go either way.
LibDem triumphs. Eight seats saw the LibDems snatch victory against the electoral odds, with residual swings of 5% to 10% (or a mighty 20% for the by-election hold of Brent East). Some of these are anti-war hotspots with large student populations (Cambridge, Leeds NW, Manchester Withington). Others seem due to local factors and the Liberal Democrat strength of targetting local votes independent of national trends.
Labour in-fighting. In both Bethnal Green and Blaenau Gwent, Labour was beaten by a candidate who had been expelled from the party.
Other. I have no suggestions for the remaining 13 seats. Comments and suggestions are welcome as to the local factors which influenced the result.

Total error

Here is a summary of the errors described above:

Error type	Seats Incorrect	CON	LAB	LIB	NAT	MIN
Polls	21	17	-15	0	-2	0
Regional	14	8	-3	-4	-1	0
Tactical	7	5	-7	2	0	0
Local/Model	32	-2	-8	6	2	2
Total	52	28	-33	4	-1	2

Because eleven seats suffered two mis-predictions in opposite directions, the total number of seats mis-predicted (52) is less than the sum of all the seat errors (74).

In summary, the opinipon poll error was less than previous years but still contributed significantly to the total error in Labour's majority. Model error was a more important factor this year, with regional effects, tactical voting, and local factors all working in the same direction as poll error.

As it happens, the most accurate prediction of the result beforehand came from Electoral Calculus users who predicted a Labour majority of 66.

It should be assumed that about 50 seats will be mis-predicted at the next election.

Return to top of page, or track record summary, or home page.