|
Post by bigfatron on Dec 27, 2021 12:01:07 GMT
Well, if Labour get ~30% in East Surrey and the Lib Dems get ~10% then I will eat Paddy Ashdown's hat.... Remembering that last time out was LDems 19.5%, labour 13.5%, even taking the minimum/maximum range points (Labour minimum 22%, LDem maximum 18%) seems highly unlikely!
Labour's highest ever vote in this constituency since the Liberals/LDems started standing in 1959 was the 21% achieved in 1997 - according to this MRP they are bound to surpass that?
There are equally silly results giving the Lib Dems quite reasonable results (6-7-8%) in inner city seats where they scarcely had a voter in 2019...
|
|
|
Post by froome on Dec 27, 2021 12:15:10 GMT
This one shows the Tories winning Oxford West & Abingdon on just under 30% of the vote and only very marginally ahead of Labour with the Lib Dems in third. And it actually has Labour winning Bath... Bizarrely it has Labour winning Edinburgh East but losing Edinburgh South to the Nats. Re Bath, I've already commented on the Focaldata thread. But elements in the media, and seemingly some polling companies, do seem to be utterly enamoured with the idea that Labour could win Bath. It is a fantasy based on the type of demographic we have here, and that the same media can only see Labour as the repository for their votes. At the last two elections Labour trotted out leaflets which 'showed' that they could win here. They weren't believed then but I'm sure they will follow the handbook saying that if you say something enough times, someone is eventually going to believe you. We will see, but I don't know anyone here, except maybe one or two core Labour activists, who actually believe it could be possible, and not that many amongst people who might consider voting Labour who even want it.
|
|
iain
Lib Dem
Posts: 10,709
|
Post by iain on Dec 27, 2021 12:20:27 GMT
Cheers, just had a quick scroll through.
Seems like the same "Squashing" as the focaldata one - every party squeezed significantly towards their national vote share in every seat. Labour predicted >20% in every seat in England apart from Beaconsfield, East Devon, Lewes and St Albans. Various bizarre individual seat results including the ones you mention earlier.
What is the point of modelling that sells itself on giving accurate, precise seat by seat predictions, and yet produces clear nonsense?
It isn't complete nonsense, but it has limitations and needs to be interpreted. It can't take account of tactical considerations in individual seats, even if it works on second order changes between parties. You can look at where you think tactical voting will occur and adjust - that's how you treat models in any sphere. You can ask as part of your polling whether people would consider voting for another party and if so which, but you won't get anything like the actual tactical considerations that squeeze people in a real campaign. Also don't knows are higher in polls than in elections and more importantly they are different people. But there are very many seats where tactical voting is small, and here the model can throw up real features that you can't get with universal swing. You can be pretty sure that all the parties look at these models in deciding where to concentrate their efforts. Oh, and as 1997 showed, the Labour Party can sometimes come from a poor third - if they campaign and the Liberal Democrats focus elsewhere. See Hastings or Leeds NW, or even St Albans. If an MRP doesn’t take account of LD v SNP seats, or Brighton Pavilion, then fair enough. It’s not supposed to get into that level of granularity. But if it can’t be sensible about large groups of seats such as, say, Conservative-Lib Dem marginals or the whole of Scotland, then I’d argue the model isn’t much use at all.
|
|
graham
Non-Aligned
Posts: 953
Member is Online
|
Post by graham on Dec 27, 2021 13:01:38 GMT
This one shows the Tories winning Oxford West & Abingdon on just under 30% of the vote and only very marginally ahead of Labour with the Lib Dems in third. And it actually has Labour winning Bath... Bizarrely it has Labour winning Edinburgh East but losing Edinburgh South to the Nats. Re Bath, I've already commented on the Focaldata thread. But elements in the media, and seemingly some polling companies, do seem to be utterly enamoured with the idea that Labour could win Bath. It is a fantasy based on the type of demographic we have here, and that the same media can only see Labour as the repository for their votes. At the last two elections Labour trotted out leaflets which 'showed' that they could win here. They weren't believed then but I'm sure they will follow the handbook saying that if you say something enough times, someone is eventually going to believe you. We will see, but I don't know anyone here, except maybe one or two core Labour activists, who actually believe it could be possible, and not that many amongst people who might consider voting Labour who even want it. I know this was discussed at length on the FocalData thread, but these MRP findings do make me wonder whether they provide some hint of the likely outcome in certain seats were all tactical voting to be unwound. In Bath the underlying Labour vote is a good deal higher than recent election results imply - and the LDs quite a bit weaker.To what extent the overall result in the seat would change were ALL tactical voting to unwind, is very difficult indeed to say. Labour was only 800 votes from unseating the Tories there in 1966 - though of course 75% of those voters are no longer alive.
|
|
|
Post by John Chanin on Dec 27, 2021 13:03:30 GMT
Implied swings from MRP data (Labour gains from Conservative) Bassetlaw | 18.4% | Gloucester | 10.8% | Peterborough | 8.9% | Mansfield | 17.5% | Warrington S | 10.8% | Durham NW | 8.8% | Harlow | 16.8% | Derby North | 10.7% | Morecambe & Lunesdale
| 8.8% | Stoke South | 16.3% | Clwyd South | 10.7% | Bury South | 8.6% | Telford | 16.0% | Carmarthen W etc | 10.6% | Southampton Itchen | 8.6% | Grimsby | 14.9% | Rossendale & Darwen | 10.6% | Dewsbury | 8.5% | Bishop Auckland | 14.6% | Don Valley | 10.6% | Rushcliffe | 8.5% | Wolverhampton NE | 14.4% | Crawley | 10.5% | Loughborough | 8.4% | Leigh | 14.1% | Workington | 10.4% | Worthing E & Shoreham | 8.2% | Stockton South | 14.1% | Bolton NE | 10.4% | Delyn | 8.2% | Corby | 13.8% | Uxbridge & S Ruislip | 10.4% | Southport | 8.2% | Southend E | 13.5% | Bolton West | 10.3% | Bury North | 8.1% | Stoke North | 13.3% | Barrow & Furness | 10.3% | Milton Keynes N | 8.0% | West Bromwich W | 13.3% | Northampton N | 10.2% | Reading West | 8.0% | Wakefield | 12.8% | Ashfield | 10.2% | Milton Keynes S | 7.9% | Scunthorpe | 12.8% | Colchester | 10.2% | Preseli Pembrokeshire | 7.8% | Blackpool S | 12.8% | Bournemouth E | 10.1% | Norwich North | 7.7% | Bournemouth W | 12.6% | Rother Valley | 10.1% | High Peak | 7.7% | Bolsover | 12.6% | Darlington | 10.1% | Chingford & Woodford Gn | 7.6% | Carlisle | 12.6% | Hastings & Rye | 10.0% | Colne Valley | 7.5% | Penistone & Stocksbridge | 12.4% | Watford | 9.9% | Calder Valley | 7.5%
| Thurrock | 12.3% | Blyth Valley | 9.8% | Wrexham | 7.3% | Newcastle (Staffs) | 12.2% | Northampton S | 9.8% | Aberconwy | 7.2% | Blackpool North | 12.0% | Keighley | 9.8% | Filton & Bradley Stoke | 7.1% | Scarborough & Whitby | 11.8% | Birmingham Northfield | 9.7% | Shipley | 6.8% | West Bromwich E | 11.8% | Ipswich | 9.6% | Finchley & Golders Green | 6.2% | Lincoln | 11.5% | Pendle | 9.6% | Hendon | 6.1% | Thanet South | 11.4% | Harrow East | 9.5% | Altrincham & Sale E | 6.1% | Heywood & Middleton | 11.4% | Sedgefield | 9.4% | Wycombe | 5.9% | Burnley | 11.4% | Vale of Clwyd | 9.3% | Bridgend | 5.8% | Crewe & Nantwich | 11.3% | Gedling | 9.2% | Pudsey | 5.6% | Wolverhampton SW | 11.2% | Clwyd West | 9.1% | Stroud | 5.0% | Swindon South | 11.1% | Camborne & Redruth | 9.0% | Truro & Falmouth | 4.9% | Redcar | 11.0% | Copeland | 8.9% | Kensington | 4.3% | Stoke Central | 11.0% | Ynys Mon | 8.9% | Chipping Barnet | 3.3% | Hyndburn | 10.9% | Worcester | 8.9% |
|
|
This enables you to see how this MRP differs from universal swing, and what sorts of seats it expects high and low swings
|
|
graham
Non-Aligned
Posts: 953
Member is Online
|
Post by graham on Dec 27, 2021 13:47:15 GMT
Again, a notably low LibDem number. It's difficult to believe Labour would win Bath which sort of makes me sceptical of any MRP study that includes such a finding. I suppose it is not impossible that 12,000 Labour voters have regularly been switching to the LDs there for tactical reasons.Adding such a figure to the Labour total - and deducting it from the LDs - changes past election results rather dramatically. Moreover, Labour polled over 18,000 there in 1966 when the voting age was still 21!
|
|
|
Post by No Offence Alan on Dec 27, 2021 13:55:15 GMT
Again, a notably low LibDem number. It's difficult to believe Labour would win Bath which sort of makes me sceptical of any MRP study that includes such a finding. My understanding is that MRP measures "the kind of people likely to vote for Party X" and doesn't really take levels of local activity into account.
|
|
|
Post by woollyliberal on Dec 27, 2021 14:05:35 GMT
It's difficult to believe Labour would win Bath which sort of makes me sceptical of any MRP study that includes such a finding. My understanding is that MRP measures "the kind of people likely to vote for Party X" and doesn't really take levels of local activity into account. That's pretty much it. For each constituency, they know how many people there are in each of the 70 odd Experian groups. They then poll each Experian group and construct a voting intention for each constituency. They don't do is take account of tactical voting. So places where it is SNP vs whichever of Con / Lab / LD, they predict an SNP gain a few percentage points before they should. They also don't account for local activity, places where one opposition party tries and another doesn't. What we need is an "overlay" of the typical difference between MRP and actual election results for each constituency to tune out the tactical situation.
|
|
cogload
Lib Dem
I jumped in the river and what did I see...
Posts: 8,072
|
Survation
Dec 27, 2021 14:12:20 GMT
via mobile
Post by cogload on Dec 27, 2021 14:12:20 GMT
|
|
|
Post by andrew111 on Dec 27, 2021 14:23:13 GMT
Again, a notably low LibDem number. Once again, notably ridiculous. Includes such seat by seat highlights as us falling to a poor third in Bath, but closing to within only 1% of winning in Cities. I am always quite suspicious of people who publish spreadsheet data to 9 decimal places. They also give a finite SNP vote in every English constituency, which signifies a badly designed machine learning model. If we assume the SNP% in all the seats then has to add up to a fixed UK %, that might lead to an underestimate of the SNP vote in Scotland.. I don't know what the sample size is in this poll, but Focal Data had 24k. That does give the chance to group data from Lib Dem seats and target seats and perhaps get a significant sample to use separately from Lab-Con seats. Right now these MRPs are a black box and as far as I can find out the algorithms are not published.
|
|
|
Post by carolus on Dec 27, 2021 14:28:41 GMT
My understanding is that MRP measures "the kind of people likely to vote for Party X" and doesn't really take levels of local activity into account. That's pretty much it. For each constituency, they know how many people there are in each of the 70 odd Experian groups. They then poll each Experian group and construct a voting intention for each constituency. They don't do is take account of tactical voting. So places where it is SNP vs whichever of Con / Lab / LD, they predict an SNP gain a few percentage points before they should. They also don't account for local activity, places where one opposition party tries and another doesn't. What we need is an "overlay" of the typical difference between MRP and actual election results for each constituency to tune out the tactical situation. This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
|
|
|
Post by bjornhattan on Dec 27, 2021 14:38:22 GMT
That's pretty much it. For each constituency, they know how many people there are in each of the 70 odd Experian groups. They then poll each Experian group and construct a voting intention for each constituency. They don't do is take account of tactical voting. So places where it is SNP vs whichever of Con / Lab / LD, they predict an SNP gain a few percentage points before they should. They also don't account for local activity, places where one opposition party tries and another doesn't. What we need is an "overlay" of the typical difference between MRP and actual election results for each constituency to tune out the tactical situation. This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
I can't help but feel like using MRPs for general elections is a lost cause, at least when done outside a campaign. There just seem to be too many issues which need to be corrected for - everything from incumbency to tactical voting to local issues. They might be more useful if we had proportional representation. There would be less need to adjust for tactical voting (possibly not at all depending on the system used), and if local MPs no longer exist then incumbency effects would be less important. But it seems to me that they're much more suited to issues based polling. Survation appear to have attempted to do this by asking respondents whether they think the government abides by the Nolan principles, and modelling how different constituencies feel about this. While they do have some odd results (such as neighbouring constituencies with similar demographics and politics diverging sharply), these seem to be much more plausible than the "headline" MRP where they attempt to model the election.
|
|
|
Post by John Chanin on Dec 27, 2021 15:03:41 GMT
Another try. “All models are wrong but some are useful” is a standard aphorism. These models are not predictions. They are models based on a regression of relevant factors. You can’t adjust a model “by hand” because it doesn’t look right. The model gives the results it gives. Personally I find these MRP models interesting and useful.
I understand why Liberal Democrats get exercised as well represented here. But the underlying fact is that given their standing in the polls and the diffuse nature of their support, they are not going to come out well from demographic models. You could construct an entirely different sort of model which would show Liberal Democratic prospects, but it would have very different features. There are probably some statisticians doing exactly that at Liberal HQ (do they have an HQ?). But it is essentially a sideshow. The Survation and Focaldats MRPs are designed to provide national models. They don’t do a great job in Scotland either, although that’s easier to incorporate in a national model.
|
|
|
Post by carolus on Dec 27, 2021 15:05:19 GMT
This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
I can't help but feel like using MRPs for general elections is a lost cause, at least when done outside a campaign. There just seem to be too many issues which need to be corrected for - everything from incumbency to tactical voting to local issues. They might be more useful if we had proportional representation. There would be less need to adjust for tactical voting (possibly not at all depending on the system used), and if local MPs no longer exist then incumbency effects would be less important. But it seems to me that they're much more suited to issues based polling. Survation appear to have attempted to do this by asking respondents whether they think the government abides by the Nolan principles, and modelling how different constituencies feel about this. While they do have some odd results (such as neighbouring constituencies with similar demographics and politics diverging sharply), these seem to be much more plausible than the "headline" MRP where they attempt to model the election. I completely agree. Even setting aside the voting dynamics issue, the philosophical question of "What is this actually modelling?" doesn't seem to have a good answer. It's not a projection of the next election, since as we keep getting reminded, they're snapshots not forecasts. But if they're not predicting an actual event, then they're essentially unverifiable. If we got another one from another agency in a week that had a 50 seat difference for Lab or Con, or a 10 seat difference for LD or SNP or whatever, would we be able to make any meaningful conclusions about what it means? I'm not sure we would. And at that point the error bounds on these become so big, and the entire thing so shorn of context that it's pointless. I suppose if we were getting them frequently from the same agency then we might at least see trends - but if the model is nonsense then the trend may well be too I think your point about the issues polling is reasonable - I admit I'd glossed over it when I looked at the results. But it does seem like a coherent task, and avoids the voting dynamics problem.
|
|
|
Survation
Dec 27, 2021 15:14:47 GMT
via mobile
Post by carolus on Dec 27, 2021 15:14:47 GMT
Another try. “All models are wrong but some are useful” is a standard aphorism. These models are not predictions. They are models based on a regression of relevant factors. You can’t adjust a model “by hand” because it doesn’t look right. The model gives the results it gives. Personally I find these MRP models interesting and useful. I understand why Liberal Democrats get exercised as well represented here. But the underlying fact is that given their standing in the polls and the diffuse nature of their support, they are not going to come out well from demographic models. You could construct an entirely different sort of model which would show Liberal Democratic prospects, but it would have very different features. There are probably some statisticians doing exactly that at Liberal HQ (do they have an HQ?). But it is essentially a sideshow. The Survation and Focaldats MRPs are designed to provide national models. They don’t do a great job in Scotland either, although that’s easier to incorporate in a national model. I am not, and have not, proposed adjusting anything "by hand". When you design and fit your model you must decide: - What data you will fit the model on - What model architecture you will use - Above all, what you are trying to predict, and how you will evaluate that prediction. Your assertion, as I understand it, is that the only "relevant factors" they have included are demographic ones, and that voting dynamics are irrelevant for their modelling objective. My assertion is that they are not irrelevant and that you certainly could include relevant data as features in your model if your objective was the advertised one of providing predictions of the outcomes of a GE on a seat by seat basis. You could also incorporate modelling of voting dynamics on an architectural level if you want. These aren't particularly abnormal modelling ideas - I find it very strange to think that the organisations have not successfully implemented them. But at least on the surface it seems thst they haven't.
|
|
batman
Labour
Posts: 8,726
Member is Online
|
Post by batman on Dec 27, 2021 15:26:55 GMT
Implied swings from MRP data (Labour gains from Conservative) Bassetlaw | 18.4% | Gloucester | 10.8% | Peterborough | 8.9% | Mansfield | 17.5% | Warrington S | 10.8% | Durham NW | 8.8% | Harlow | 16.8% | Derby North | 10.7% | Morecambe & Lunesdale
| 8.8% | Stoke South | 16.3% | Clwyd South | 10.7% | Bury South | 8.6% | Telford | 16.0% | Carmarthen W etc | 10.6% | Southampton Itchen | 8.6% | Grimsby | 14.9% | Rossendale & Darwen | 10.6% | Dewsbury | 8.5% | Bishop Auckland | 14.6% | Don Valley | 10.6% | Rushcliffe | 8.5% | Wolverhampton NE | 14.4% | Crawley | 10.5% | Loughborough | 8.4% | Leigh | 14.1% | Workington | 10.4% | Worthing E & Shoreham | 8.2% | Stockton South | 14.1% | Bolton NE | 10.4% | Delyn | 8.2% | Corby | 13.8% | Uxbridge & S Ruislip | 10.4% | Southport | 8.2% | Southend E | 13.5% | Bolton West | 10.3% | Bury North | 8.1% | Stoke North | 13.3% | Barrow & Furness | 10.3% | Milton Keynes N | 8.0% | West Bromwich W | 13.3% | Northampton N | 10.2% | Reading West | 8.0% | Wakefield | 12.8% | Ashfield | 10.2% | Milton Keynes S | 7.9% | Scunthorpe | 12.8% | Colchester | 10.2% | Preseli Pembrokeshire | 7.8% | Blackpool S | 12.8% | Bournemouth E | 10.1% | Norwich North | 7.7% | Bournemouth W | 12.6% | Rother Valley | 10.1% | High Peak | 7.7% | Bolsover | 12.6% | Darlington | 10.1% | Chingford & Woodford Gn | 7.6% | Carlisle | 12.6% | Hastings & Rye | 10.0% | Colne Valley | 7.5% | Penistone & Stocksbridge | 12.4% | Watford | 9.9% | Calder Valley | 7.5%
| Thurrock | 12.3% | Blyth Valley | 9.8% | Wrexham | 7.3% | Newcastle (Staffs) | 12.2% | Northampton S | 9.8% | Aberconwy | 7.2% | Blackpool North | 12.0% | Keighley | 9.8% | Filton & Bradley Stoke | 7.1% | Scarborough & Whitby | 11.8% | Birmingham Northfield | 9.7% | Shipley | 6.8% | West Bromwich E | 11.8% | Ipswich | 9.6% | Finchley & Golders Green | 6.2% | Lincoln | 11.5% | Pendle | 9.6% | Hendon | 6.1% | Thanet South | 11.4% | Harrow East | 9.5% | Altrincham & Sale E | 6.1% | Heywood & Middleton | 11.4% | Sedgefield | 9.4% | Wycombe | 5.9% | Burnley | 11.4% | Vale of Clwyd | 9.3% | Bridgend | 5.8% | Crewe & Nantwich | 11.3% | Gedling | 9.2% | Pudsey | 5.6% | Wolverhampton SW | 11.2% | Clwyd West | 9.1% | Stroud | 5.0% | Swindon South | 11.1% | Camborne & Redruth | 9.0% | Truro & Falmouth | 4.9% | Redcar | 11.0% | Copeland | 8.9% | Kensington | 4.3% | Stoke Central | 11.0% | Ynys Mon | 8.9% | Chipping Barnet | 3.3% | Hyndburn | 10.9% | Worcester | 8.9% |
|
|
This enables you to see how this MRP differs from universal swing, and what sorts of seats it expects high and low swings Labour requires a swing well in excess of 6.2% to gain Finchley & Golders Green
|
|
|
Survation
Dec 27, 2021 19:47:05 GMT
via mobile
Post by andrew111 on Dec 27, 2021 19:47:05 GMT
This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
I can't help but feel like using MRPs for general elections is a lost cause, at least when done outside a campaign. There just seem to be too many issues which need to be corrected for - everything from incumbency to tactical voting to local issues. They might be more useful if we had proportional representation. There would be less need to adjust for tactical voting (possibly not at all depending on the system used), and if local MPs no longer exist then incumbency effects would be less important. But it seems to me that they're much more suited to issues based polling. Survation appear to have attempted to do this by asking respondents whether they think the government abides by the Nolan principles, and modelling how different constituencies feel about this. While they do have some odd results (such as neighbouring constituencies with similar demographics and politics diverging sharply), these seem to be much more plausible than the "headline" MRP where they attempt to model the election. I suspect 99% of those polled have no idea what the Nolan Principles are.. Presumably they explain. Here are the Principal Nolans though (introduced in Sweden..)
|
|
|
Post by greenchristian on Dec 29, 2021 17:34:02 GMT
That's pretty much it. For each constituency, they know how many people there are in each of the 70 odd Experian groups. They then poll each Experian group and construct a voting intention for each constituency. They don't do is take account of tactical voting. So places where it is SNP vs whichever of Con / Lab / LD, they predict an SNP gain a few percentage points before they should. They also don't account for local activity, places where one opposition party tries and another doesn't. What we need is an "overlay" of the typical difference between MRP and actual election results for each constituency to tune out the tactical situation. This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
You still have the problem that your sample per constituency is very small, so you'll only pick up national tactical voting trends. MRP gives a better read on which seats are actually in play than UNS, but neither approach is likely to pick up tactical voting very well. To do that you'd need to base your model on data that includes a large sample from every constituency where tactical voting could plausibly make a difference to the result (as well as enough data from elsewhere to calibrate it against). It is very difficult to design a model which doesn't underestimate parties which are only competitive in a relatively small proportion of seats unless you have high quality constituency-level polls from a very wide spread of constituencies.
|
|
|
Post by carolus on Dec 29, 2021 19:03:56 GMT
This may, in practice, be the case for some of these MRPs. But it certainly isn't how they are advertised.
It would not be a tremendously difficult modelling task to take these things into account - you might want an additional squeeze question on your respondents - "How would you vote in your constituency, if there were an election tomorrow?" or similar, you would make sure your model is taking into account things like previous vote shares and so forth. Of course you couldn't teach your model about the particular importance of the A1234 bypass in West Southshire, but you don't need to.
In fact it is almost beyond belief to me that they would not attempting to do this - the obvious expectation of your model to predict elections on a constituency by constituency basis is that it should... predict what will happen. In fact, if you weren't trying to do this, it's not clear that you could even validate your model - you can't check it against real GE results, because they do have these voting dynamics!
You still have the problem that your sample per constituency is very small, so you'll only pick up national tactical voting trends. MRP gives a better read on which seats are actually in play than UNS, but neither approach is likely to pick up tactical voting very well. To do that you'd need to base your model on data that includes a large sample from every constituency where tactical voting could plausibly make a difference to the result (as well as enough data from elsewhere to calibrate it against). It is very difficult to design a model which doesn't underestimate parties which are only competitive in a relatively small proportion of seats unless you have high quality constituency-level polls from a very wide spread of constituencies. That's a fair point about the squeeze question (although probably still helps a bit). But if you decide you want your model to take into account tactical voting I don't think you actually need to do it constituency by constituency in that way.
I can think of a few different ways to proceed with it: - bin your constituencies e.g. LD vs Con seats, Lab vs SNP targets, and so on and so forth. That probably gets you a vaguely useable amount of responses per category
- use historic data and so on to build a tactical/squeeze model - investigate how vote shares typically change in two party marginals, and so on - if you have enough historic national data then you could probably construct some demographic "ease of squeeze", although I'd imagine you'll find by far the most important feature is which party was second last time. - the whole hog, give up on the actual "MRP" part, and just throw everything into an MLP or some other more powerful model
But the fundamental point is just that none of this is implausible, and that there are ways to do it - if your stated objective is to provide seat by seat predictions you have to be doing something to deal with tactical voting, otherwise you really aren't doing anything of merit at all.
I should note - it's entirely possible that either the modellers are trying to do this and it has gone wrong somehow, or there is something major I am missing.
|
|
timmullen1
Labour
Closing account as BossMan declines to respond to messages seeking support.
Posts: 11,823
|
Post by timmullen1 on Dec 29, 2021 19:11:38 GMT
You still have the problem that your sample per constituency is very small, so you'll only pick up national tactical voting trends. MRP gives a better read on which seats are actually in play than UNS, but neither approach is likely to pick up tactical voting very well. To do that you'd need to base your model on data that includes a large sample from every constituency where tactical voting could plausibly make a difference to the result (as well as enough data from elsewhere to calibrate it against). It is very difficult to design a model which doesn't underestimate parties which are only competitive in a relatively small proportion of seats unless you have high quality constituency-level polls from a very wide spread of constituencies. That's a fair point about the squeeze question (although probably still helps a bit). But if you decide you want your model to take into account tactical voting I don't think you actually need to do it constituency by constituency in that way.
I can think of a few different ways to proceed with it: - bin your constituencies e.g. LD vs Con seats, Lab vs SNP targets, and so on and so forth. That probably gets you a vaguely useable amount of responses per category
- use historic data and so on to build a tactical/squeeze model - investigate how vote shares typically change in two party marginals, and so on - if you have enough historic national data then you could probably construct some demographic "ease of squeeze", although I'd imagine you'll find by far the most important feature is which party was second last time. - the whole hog, give up on the actual "MRP" part, and just throw everything into an MLP or some other more powerful model
But the fundamental point is just that none of this is implausible, and that there are ways to do it - if your stated objective is to provide seat by seat predictions you have to be doing something to deal with tactical voting, otherwise you really aren't doing anything of merit at all.
I should note - it's entirely possible that either the modellers are trying to do this and it has gone wrong somehow, or there is something major I am missing.
Everybody seems to be overlooking this quote from the Opinium survey, which I imagine holds just as true for Survation: “The seat predictions are based on a uniform swing, that is, they take no account of local factors, tactical voting or boundary changes and should be regarded as a broad guide, not a precise projection.”
|
|