Polls come and polls go, most of them resulting in politicos attempting to establish how many seats will be won and lost by the various parties, but do they really tell us anything? Malc and I are halfway through our #sp11 regionwatch series so I thought it was time to check if there is any credibility to the results we are putting out there.

To prove this empirically would probably take a few PhDs and a strong dash of John Curtice so, instead, I’ve pulled together what the 2007 predicted result would be using my model, the 2003 detailed results and the 2007 national voteshare for constituency and region.

The regional detail is below but at a top level, and with actuals in brackets, the model would predict a 2007 election result of:

SNP 47 (47)
Labour 45 (46)
Conservatives 18 (17)
Lib Dems 17 (16)
Greens 1 (2)
Margo 1 (1)

All in all, I say that isn’t too shabby. The main failing is that it predicted that an SNP/Tory coalition is possible when infact it wasn’t (by 1 seat). There were several instances of a hundred or so votes being the difference between who won the 7th regional spot (Patrick Harvie in Glasgow for example) so the margin of error in my model is, of course, quite high.

The detailed reasons for differences are below but the main conclusions I take from this are as follows:

First up, it is worthwhile to predict seat numbers based on national voteshare, safe in the knowledge that regional differences here will be largely mapped out by regional differences there.

Secondly, it is worthwhile to reflect the really obvious local factors but not to get carried away. Green co-convener received a poll bump in Glasgow and Alex Salmond turned a notional Lib Dem hold into an SNP gain comfortably. Popular candidates such as Alex Fergusson, Jim Mather and Bruce Crawford also reversed expectations and this could arguably have been predicted if one was close enough to the ground in these areas.

Thirdly, the competition for the 7th regional spot on the list can be fierce. It is here where the election may well be won and lost for many parties.

My final conclusion, and probably the main one, is that baseline numbers are very important. The prime example is Cumbernauld & Kilsyth where Jamie Hepburn on paper only had to win an extra 500 or so votes to win the constituency, but he was operating from an artificial platform as much of the 2003 SNP vote was due to the impressive Andrew Wilson standing there.

So where might the dodgy baselines exist for 2011? Well, the constituencies that have the largest discrepancy between what did happen in 2007 and what was expected to happen are as follows: Galloway and Upper Nithsdale, Paisley South, Gordon, Ochil, Caithness, Sutherland & Easter Ross, Linlithgow, Carrick, Cumnock & Doon valley, Stirling, Edinburgh East and Airdrie & Shotts. If you are looking for a shock, whatever you perceive a shock to be, I’d look in these constituencies first…

    Detail and methodology

Central
SNP – 7 (7)
Labour – 8 (8)
Conservative – 1 (1)
Lib Dem – 1 (1)
Green – 0 (0)

A good start. No differences between the predictor model and the actual results in total.

NB – The SNP were predicted to win Cumbernauld and Kilsyth. This loss was made up for in the d’hondt allocation.

Glasgow
SNP – 6 (5)
Labour – 9 (9)
Conservative – 1 (1)
Lib Dem – 1 (1)
Green – 0 (1)

Reason for difference: In the model, the Greens were only 95 votes short of winning a seat (from the SNP) so perhaps a personal vote for Patrick Harvie was not reflected. Note that Patrick was the 7th list MSP so the model really wasn’t so far off here at all.

The FPTP element of the model predicted constituency wins accurately with Nicola Sturgeon winning Govan and Labour wins elsewhere.

Highlands & Islands
SNP – 5 (6)
Labour – 3 (3)
Conservative – 2 (2)
Lib Dem – 5 (4)
Green – 0 (0)

Reason for difference: The model predicted that the Lib Dems would win Argyll & Bute which was won by the SNP.

Lothian
SNP – 5 (5)
Labour – 5 (5)
Conservative – 2 (2)
Lib Dem – 2 (2)
Green – 1 (1)

No difference.

All FPTP seats predicted accurately.

Mid Scotland & Fife
SNP – 6 (6)
Labour – 5 (5)
Conservative – 3 (3)
Lib Dem – 2 (2)
Green – 0 (0)

No difference.

NB – Labour was predicted to win Stirling and Dunfermline West but the SNP and Lib Dems took surprise victories there. These variances then balanced out in the d’hondt allocation on the strength of Labour’s second votes.

North East
SNP – 7 (8)
Labour – 3 (3)
Conservative – 3 (2)
Lib Dem – 3 (3)
Green – 0 (0)

Reason for difference: The SNP was 324 votes away from pinching the 7th seat from the Conservatives in my model.

NB – The Lib Dems were predicted to win Gordon if the model was not adjusted for Alex Salmond standing in this constituency. The SNP was predicted to win Aberdeen Central but Labour managed to cling onto it by 382 votes.

South
SNP – 5 (5)
Labour – 6 (5)
Conservative – 4 (4)
Lib Dem – 1 (2)
Green – 0 (0)

Reason for difference: In my model, Labour pinched the last seat from the Lib Dems by 160 votes, once again showing that it is small margins at the tail end of the d’hondt allocation that can make all the difference.

(NB – The SNP upset the horses here, failing to win any FPTP seats despite being predicted to win 2. Alex Fergusson comfortably held Galloway and Upper Nithsdale for the Tories and Jeremy Purvis scraped through in Tweeddale, Ettrick & Lauderdale for the Lib Dems. This impact was corrected through the d’hondt calculator.)

West
SNP – 6 (5)
Labour – 6 (8)
Conservative – 2 (2)
Lib Dem – 2 (1)
Green – 0 (0)

Reason for difference – The SNP were predicted to win Paisley South and West Renfrewshire in my model but, in reality, Labour won them reasonably comfortably, perhaps highlighting the relative impenetrability of the West. These represented two clear gains for Labour between model and actuals as Labour did not win a regional MSP in either set of results.

That leaves the gain for the SNP over the Lib Dems and this seems to be simply due to a poor turnout for the Lib Dems in the West, receiving only 22,515 votes to my modelled 30,217. If this trend has continued, it is a clear reason why Hugh O’Donnell has decided to abandon Tavish’s sinking longboat.

(As a quick overview of my method:

Constituency – I take the individual number of votes for each candidate in each constituency from 2003, divide by the national voteshare from 2003 for the relevant party and then multiply by the 2007 national voteshare to get a new number of votes. In true First Past the Post tradition, whoever receives the highest votes for each constituency is then nominated as the winner)

Regions – The number of votes in each region in 2003 is divided by the national voteshare in the list vote for the relevant party and then multiplied by the actual national voteshare for 2007. The d’hondt allocation method is then applied using these number of votes and the number of FPTP seats predicted as being won in the first stafe of the model)