Possums Pollytics

Politics, elections and piffle plinking

Polling Analysis and Carr-ying on

Posted by Possum Comitatus on December 12, 2007

A long time ago in a poll far, far away – the September quarterly Newspoll breakdown to be precise, some people got a bee in their bonnet about such outrageous overanalysing of the polling data.

The key problem, despite many a clarification to the contrary both here, Poll Bludger and just about everywhere else in the known pollyjunkie universe, was a simple one where critics refused to listen to what was actually being said, preferring to make up their own interpretations of what the key figures produced actually meant. Explanations became pointless and the only way to address their particular problem was to simply wait for the election results and demonstrate the point with real world data.

So today we can use actual electoral data to repeat the exercise to achieve two things, firstly to test how this method stacks up against using the usual national pendulum approach when it comes to estimating the number of seats to fall from a given swing, and secondly to highlight using real world data why some critics completely missed the point.

The Newspoll quarterly breakdowns give us 2 sets of figures as ammunition for polling analysis, firstly they give us the State swings for NSW, Vic, Qld, WA and SA. Secondly they give us the swings in safe Coalition held seats, safe ALP held seats and marginal seats – where safe seats are defined as being held on a margin greater than 6%. So what we will use here is what we used last time, 139 seats in the 5 states that Newspoll measures (we’ll remove the two Independent seats from the mix).

So if go over to the AEC and extract just that data for the election result (simulating Newspoll quarterly data), we end up with the following:

  NSW Vic Qld SA WA Swing
Marginal           5.1
Safe Coalition           6.08
Safe ALP           4.79
Total Swing 5.98 5.26 7.81 6.76 2.13 5.6

Next we need to take the ratio of the Marginal Seat swing to the National Swing, which in this case is 5.1/5.6 = 0.91, then do the same for Safe Coalition Seats 6.08/5.6=1.09 and again for Safe ALP seats 4.79/5.6= 0.86

What we will do here is make the assumption that the ratio of the swing types will hold between States – meaning that in every state the average swing in the marginal seats for that state will be 0.91 multiplied by the State swing for that State. So the swing in NSW marginal seats will be 0.91*5.98 = 5.45. We then do that for every State and we end up with a populated swing matrix of:

swings NSW Vic Qld SA WA Swing ratio
marginal 5.446071 4.790357 7.112679 6.156429 1.939821 5.10 0.910714
safecoal 6.492571 5.710857 8.479429 7.339429 2.312571 6.08 1.085714
safe alp 5.115036 4.499179 6.680339 5.782214 1.821911 4.79 0.855357
Swing 5.98 5.26 7.81 6.76 2.13 5.60 1

This assumption may not hold exactly – but that’s OK, the differences should come out in the wash at the end, the point here is to estimate the number of seats to fall given the data that Newspoll quarterly breakdowns provide us with. It’s a pendulum within pendulums approach.

Next up we need to remove any over or under cooked feedback effects within states, between the movements in their 3 seat categories and their total state swing – so let me introduce to you a thing called a swing unit. A swing unit is simply the number of seats multiplied by a swing. If we have 10 seats, and applied a 5% swing, we would have 50 swing units.

So the number of seats and their type can be represented as:

seats NSW Vic Qld SA WA Total
marginal 11 13 7 6 5 42
safecoal 20 14 19 4 8 65
safe alp 17 10 2 1 2 32
Total 48 37 28 11 15 139

To get the swing units for each seat type, we simply multiple, for example, the marginal NSW seat number (11) by the marginal NSW seat swing (5.45) to get 59.9 swing units for that type. If we do that for all seat types (and where we also multiple the totals of the State seats by their respective State swing) we get:

swingunits NSW Vic Qld SA WA
marginal 59.90679 62.27464 49.78875 36.93857 9.699107
safecoal 129.8514 79.952 161.1091 29.35771 18.50057
safe alp 86.95561 44.99179 13.36068 5.782214 3.643821
Total swing units 276.7138 187.2184 224.2586 72.0785 31.8435
State swing units 287.04 194.62 218.68 74.36 31.95

As we can see, the total swing units for NSW using the sum of the marginal and safe seats is 276.7, but the total number of NSW seats multiplied by the NSW swing produced 287 swing units. So what we want to do is adjust these swings by the ratio of those two numbers for all estimated swings.

So for NSW marginal seats, the swing becomes the original estimated marginal seat swing in NSW (5.45) multiplied by the ratio of total swing units for NSW (276.7) to State swing units for NSW (287.4).

Hence adjusted NSW Marginal Seat Swing becomes 5.45*(287.04/276.7)= 5.65

Doing that for all seat types gets us the following swing matrix:

Adjusted swings NSW Vic Qld SA WA
Marginal 5.649303 4.979741 6.935746 6.351298 1.946309
Safe coal 6.734856 5.936633 8.268498 7.571743 2.320306
safe alp 5.305914 4.677051 6.514162 5.965239 1.828004

Now it’s simply a matter of applying these swings to the relevant seats. The easiest way to do it is to simply list all 139 seats we are talking about and their pre-election margins – where positive margins represent ALP seats and negative margins represent Coalition seats. Then we just add these swings to the seat margins according to the type of seat e.g. NSW marginal seats all have 5.64 added to their margin, QLD safe Coalition seats all have 8.27 added to their margins and safe WA ALP seats all have 1.83 added to their margin.

The purpose of the end result is to try and get an accurate estimate of how many seats would fall given the data that Newspoll quarterly breakdowns provide us with. What isn’t important is the actual projected margin on any of the seats – that is entirely unimportant – and it’s where earlier critics lost the plot despite having it told to them repeatedly.

What is important is how many seats would be projected to fall using those numbers, not any given number itself.

This methodology is effectively a large number of pendulums all put together, pendulums within pendulums, so seats with a given projected margin will more than likely end up having either a greater or lesser actual margin than what was projected – but for every seat that ends up with a higher margin, another seat will end up with a smaller margin simply because swings tend to be normally distributed around a given mean swing. To give an example of this, if we look at the 148 seats where major parties were the victor, and show the size of the swings to the ALP as a histogram – we get a very normal looking distribution, a bell curve:

normswing1.jpg

So armed with all that, and applying those swings to the relevant seat types we end up with the following projected number of ALP seats:

ALP seats Division Proj margin Coal seats Coalition Seats Proj margin
1 Grayndler 26.61 1 Mallee -18.86
2 Batman 26.08 2 Murray -18.16
3 Melbourne 25.88 3 O’Connor -18.08
4 Sydney 22.61 4 Mitchell -13.97
5 Wills 21.68 5 Riverina -13.97
6 Blaxland 20.61 6 Maranoa -12.73
7 Watson 19.91 7 Curtin -12.38
8 Gellibrand 19.68 8 Barker -12.33
9 Gorton 19.58 9 Parkes -12.07
10 Scullin 19.48 10 Moncrieff -11.63
11 Throsby 19.21 11 Bradfield -10.77
12 PortAdelaide 18.97 12 Groom -10.73
13 Fowler 18.81 13 Pearce -10.68
14 Chifley 17.41 14 Indi -10.36
15 Reid 17.31 15 Tangney -9.48
16 Cunningham 17.01 16 Mackellar -8.77
17 Hunter 16.51 17 Farrer -8.67
18 Griffith 15.01 18 Moore -8.58
19 Shortland 14.61 19 Forrest -8.18
20 Maribyrnong 14.18 20 Lyne -7.37
21 Newcastle 14.01 21 Canning -7.28
22 KingsfordSmith 13.91 22 Aston -7.26
23 Oxley 13.71 23 Fadden -7.03
24 Charlton 13.71 24 Cook -6.97
25 Lalor 13.48 25 Wannon -6.46
26 Barton 12.91 26 Berowra -6.37
27 Calwell 12.88 27 Grey -6.33
28 Werriwa 12.41 28 Hume -6.17
29 Lilley 12.34 29 Mayo -6.03
30 Prospect 12.21 30 McPherson -5.73
31 Hotham 12.18 31 Casey -5.46
32 Brisbane 10.94 32 Flinders -5.26
33 Capricornia 10.74 33 Fairfax -5.03
34 Corio 10.68 34 Menzies -4.76
35 Rankin 9.94 35 Forde -4.73
36 Fremantle 9.63 36 Fisher -4.73
37 Jagajaga 9.48 37 Warringah -4.57
38 Banks 8.95 38 Macarthur -4.37
39 Lowe 8.75 39 Greenway -4.27
40 MelbournePorts 8.68 40 Goldstein -4.16
41 Perth 8.63 41 Kalgoorlie -4.08
42 Bruce 8.48 42 WideBay -3.93
43 Adelaide 7.75 43 Kooyong -3.66
44 Chisholm 7.68 44 Dunkley -3.46
45 Ballarat 7.28 45 NorthSydney -3.37
46 Richmond 7.15 46 Higgins -2.86
47 Brand 6.65 47 Gilmore -2.77
48 Holt 6.58 48 Ryan -2.23
49 Isaacs 6.48 49 Hughes -2.07
50 Hindmarsh 6.45 50 Leichhardt -2.03
51 Bonner 6.34 51 Dawson -1.93
52 Kingston 6.25 52 Gippsland -1.86
53 Macquarie 6.15 53 Calare -1.17
54 Bendigo 5.98 54 LaTrobe -0.92
55 Wakefield 5.65 55 Dickson -0.83
56 Makin 5.35 56 Bowman -0.63
57 Parramatta 4.55 57 McEwen -0.56
58 Moreton 4.14 58 Hinkler -0.53
59 Wentworth 3.05 59 Corangamite -0.42
60 Lindsay 2.75 60 Robertson -0.17
61 Cowan 2.75 61 Stirling -0.15
62 Eden-Monaro 2.35 62 Paterson -0.07
63 Herbert 2.17 63 McMillan -0.02
64 Swan 2.05 64 Deakin -0.02
65 Longman 1.67      
66 Bennelong 1.65      
67 Blair 1.24      
68 Boothby 0.95      
69 Dobell 0.85      
70 Sturt 0.77      
71 Flynn 0.47      
72 Petrie 0.37      
73 Page 0.15      
74 Cowper 0.13      
75 Hasluck 0.05      

Out of the 139 seats analysed, we have the ALP winning 75 of them. Then using the national swing to project to the ALP the seats in the states and territories that Newspoll doesn’t use in the quarterly breakdown we get: Tasmanian seats (5), ACT seats (2) and NT seats (2).

The total estimated number of seats using just the data of the type that the Newspoll quarterly provides is 75+5+2+2= 84

Which just so happens to be the actual number of seats that the ALP won.

If we used the national pendulum approach instead, and projected a 5.62% swing – we end up with only 81 seats being projected to fall.

This is why I use this methodology for the Newspoll Quarterly breakdown. I’ll say it again, it’s not about any given projected margin – for it’s simply a set of pendulums, it’s about the total number of seats that those projected margins estimate will fall.

So those that criticised the methodology on the basis of not understanding it in the first instance, refusing to allow it to be explained to them in the second instance, and simply making shit up about it in the third by projecting onto it meaning it does not contain (which tends to happen when one doesn’t understand something and refuses to listen to explanations of it) – well the proof is in the pudding. 84 seats projected to fall using this methodology (and only the type of data that Newspoll provides in its quarterly breakdown) vs. 84 seats actually falling.

Over to you Dr Adam Carr.

UPDATE:

Adam gave a reply you can see over here.
AddThis Social Bookmark Button

add to kwoff

45 Responses to “Polling Analysis and Carr-ying on”

  1. kwoff.com said

    Polling Analysis and Carr-ying on « Possums Pollytics

    A long time ago in a poll far, far away – the September quarterly Newspoll breakdown to be precise, some people got a bee in their bonnet about such outrageous overanalysing of the polling data.

    The key problem, despite many a clarification to the con…

  2. slartybardfast said

    Oh goodie, I was hoping this would start soon😉, “runs to kitchen to get beer and popcorn”

    Ok, I’m ready…lets go🙂

  3. Rain said

    Don’t any of you folks have day-jobs?

  4. Cat said

    Now do not take this the wrong way Poss but promise me you will not make a speech at your wedding! Could I reasonably summarise this with suck shit Carr, I was right and you were not. ?

  5. George said

    Very nice Poss. Although to be fair to poor Carr, he does write on his site:

    “Let no-one call me a sore loser. Since Labor did win Dawson (sending an astonished James Bidgood on his first ever trip to Canberra), I have emailed the King of Sweden and nominated Possum for the Nobel Prize in Psephology.”

    Although, there’s always a “however”:

    “Correspondingly, however, I think Possum might now concede that his table was wrong about all the other seats I mentioned earlier (in fact there were 28 seats on Possum’s list of predicted ALP gains which the Coalition retained). Greensborough Growler might like to reflect on his allegation that I was a prize prat for daring to apply some political commonsense to the sacred findings of the psephometricians. HarryH might be more wary in future about blind reliance on “data” without critical analysis.”

    Oh well, can’t help himself.

  6. Harmless Cud Chewer said

    Hey Possum, can you link those graphics so I can see the right half?🙂

  7. Andos the Great said

    Clearly the fault lies with neither Possum nor Adam, but with Newspoll. Since Possum’s earlier analysis/forecast was based solely on Newspoll’s data, then retrospectively we can see that there was clearly a problem with the data (I think the aim of the previous exercise was to highlight the extreme nature of the opinion poll data).

    Adam was 100% (or close enough) right to be sceptical about Labor making a gain of 52 seats, and I’m also sure that Possum was not for one minute suggesting he believed that this result would come to pass, but simply (or complicatedly) using Newspoll’s quarterly breakdown to project what this data would reflect if the electorate were to cast their votes accordingly.

    Clearly, the conclusion to be reached from this investigation runs thus: breaking down polling data by state and safe or marginal seat can give an accurate pendulum of electorates, and that Newspoll should in no way be interpreted as an accurate predictor of an election outcome.

  8. The Finnigans said

    Poss, correct me if I am wrong. What you are saying is that your methodology is correct. It’s the input data, namely the Newspoll quarterly breakdowns, that is wrong. So the old story of “gargabe in, garbage out”.

  9. lurker said

    This is my explanation for the narrowing – the undecided/refused voters who wouldn’t commit to pollsters eventually made up their minds in the last couple of weeks – and they tended to favour the Coalition. That Four Corners episode in the last couple of weeks of the campaign following a few of them is an example.

    http://www.ozforums.com.au/viewtopic.php?id=330

    Somewhere on Morgan’s website he says that people who refuse to speak to pollsters tend to be more conservative than the rest of the population. Though how he knows I have no idea – by definition these people are the “unknown unknowns” (to paraphrase Donald Rumsfeld of all people)!

  10. Ron said

    CAN YOU REPEAT THE EXERCISE BASED ON the 2004 last qtly Newspoll
    to forecast the 2004 seats vs the 2004 actual seats ?

    What would the projected 2007 seats been without the use of swing units to adjust the swing ?

    Why do not the matrix’s and pendulums and pendulums within pendulums
    not cater statistically at all for the Howard baseball bat

  11. Never did take any heed of Carr. I always knew – Possum Rules!

  12. Ron said

    Possum , another question ,
    IF you applied your formula to the September qtly Newspoll data,
    how many seats would labor be forcast to win please ?

  13. Country Kid said

    Hmmm…

    The analysis debate is heating up.

    Luckily for the rest of us – it isn’t too hot in the kitchen – or across our fair land, because the Rudd rains keep falling.

    http://www.bom.gov.au/cgi-bin/climate/rainmaps.cgi?page=map&variable=totals&period=cmonth&area=aus

  14. Ferny Grover said

    Yep CK – rainfall is always higher under Labor. So who do you trust to keep rainfall high????

  15. Possum Comitatus said

    Cud Chewer – I can, which ones are looking a bit out of whack at your end?

    Ron, over the next few weeks I’ll go back and do it for the last 5 or 6 elections.. depending on when I get the time. That September quarterly would have represented about 112 seats falling.

    Finnigins – yep, the methodology seems to work, so it’s only as good as the data then goes into it.

    The moving 4 pollster average seemed to be the best measurement, so I might play around with that between now and the next election to figure out a way to adjust the quarterly newspolls to reflect that 4 pollster average.

  16. Possum Comitatus said

    Cat back at 4, it’s actually nothing that bad. I quite like Adam and respect his views enormously.

    I’m just poking him with a stick as a bit of harmless fun in the same way he did to me:mrgreen:

  17. gusface said

    “your majesty is like Possum piss,when all else is dark you shine out like a golden light” (apologies to Monty Python)

    and long may you Comitatus

  18. ViggoP said

    Possum,

    I’m intrigued – do you take your nom de guerre from the
    Posse Comitatus Act?

    http://www.reference.com/search?q=Comitatus

    or from the more conventional meaning of comitatus?

  19. Possum Comitatus said

    Viggo, it’s a play on the Posse Comitatus Act

  20. exilemerc said

    This is so much more entertaining than tea-leaves.

    Ok, so after the election you can produce a chart to show that you predicted the result. Hmmm.

    Quick question Poss – please remind the more forgetful of us…what was your actual Labor-seat count prediction *before* the vote took place? Can you please point us to your pre-poll prediction once again…. ‘cos claims to predictive authority are so much more convincing when the predictions are made, you know, *before* things happen.

    Just sayin’, is all.

    Love your work, BTW.

    I’m just astonished that Mungo actually got something right, being the first and way earliest to predict the whole Howard-losing-his-seat thing. But then I guess a stopped clock…

  21. Harmless Cud Chewer said

    Possum @15, it’s everything graphical in the middle column. Just simply doesn’t have the width. I’m on 800×600 in case you want to do a test.

  22. fred said

    Disclaimer[s].[1] I’ve disagreed strongly with Adam at PB over the ALP Right preferencing FF in the past and his continuing justification of that.[2]I’m a big fan of Possum. [3]I’m an ‘old lefty’ disillusioned by the ALP [Right in particular] who accordingly has moved to the Greens. [4] I campaigned [non-stop for about 2 months] for the ALP this time for 2 reasons – a relative was an ALP candidate, the realistic situation was and is that the only mob who could replace the ‘deshpikkable’ [Daffy Duck speak] Lib/Nat coalition is the ALP. One small, tiny slight step forward is better than a huge step backwards as a Coalition victory would have meant.

    Both Adam and Possum were right.

    The numbers Possum showed, his analysis ‘Pollycide’, was quite simply telling us what Newspoll was saying…..at the time. No more, no less. They were electrifying. And bloody close to the mark as it turned out in many ways. My campaigning was in a seat that my touring showed was strongly reflecting the trends Possum detected via Newspoll then and on election day. And close contact [at official and rank and file levels] with other seats was also confirming of Possum’s analysis.
    But there was no way the ALP did win the hundred plus seats that Newspoll showed in September.
    Nor, in Adam’s opinion, was it ever going to be thus. And he was right, was he not?
    I think his essential point that if the polls showed what Newspoll said, as revealed by Possum, then they were wrong proved and that turned out to be accurate.
    Possum’s analysis above is based on AEC votes on the day, not Newspoll taken months before.

    The problem is that polls are snapshots of where some people claim they are at some time before the election as recorded by an organisation with credible but imperfect methodology.
    They are certainly worthy of a lot of notice.
    the problem is that Adam

  23. fred said

    Damn hit the submit thingy by mistake!

  24. fred said

    Continuing….
    Adam doesn’t place the importance on polls that Possum does.
    Possum prefers his wealth of experience…which is not to be underestimated.
    In the words of the Beatles…”Come together” guys, 2 valuable insights that should be complementing each other. I know there is a mutual respect [well, it seems that way] so add to each others expertise and maybe when the next election comes around the double-barrelled approach will get it more precisely right.

    FWIW I told my candidate relly what the result would be nationally and in our electorate virtually exactly. Within half a percent.
    Based on Possum [and Mumble and Geoff Lambert and other].

    Oh and if I have mis-perceived the positions of our 2 combatants I apologise, just saying what seems to me to be the case, I coild be wrong, have been before.
    Next election i will be reading both Adam and possum for guidance.

  25. Steve_E said

    Guys there is a lot to say about the election and what happened.

    However, today was a day to be remembered (12/12/2007) as the AEC declared the result for Bennelong. Long Live MAX. Not only has this proven a profitable call, it is a three times win: no more JWH, no more LIBS in Government and now there is Hope for Change when there was none (cause without hope there aint much left).

  26. Possum Comitatus said

    Exilemerc, this is actually about the best way to turn the type of data we receive from a quarterly newspoll into the most accurate representation of the number of seats that poll suggests would end up on each side of the house, ceteris paribus.

    It’s not about a prediction, it has absolutely nothing at all to do with prediction and for the life of me I’m bamboozled where people keep coming up with this idea. It’s about how to turn a set of data (marginal, safe Coalition and safe Alp seat swings, plus 5 state swings – what the Newspoll quarterly seat breakdowns provide) into the most accurate estimate of the number of seats that would change on that polling outcome.

    The process was misunderstood by some earlier on back in September, so to demonstrate the *process* I’ve simply substituted the type of data that we get from the big Newspolls with the same type of data from the actual election result, to show that the methodology does what it’s supposed to – estimate the number of seats that would fall given that type of data.

    I only made two predictions for the election. The forecast into November using Newspoll data up until October giving 55.15 but only 89 seats, and one way back in May where it used satisfaction variables and things like interest rates to disposable income to forecast 6 months ahead to arrive at a result of 53.7% TPP for Labor which would be 87 seats on a uniform swing.

    Cud Chewer, is that any better now?

  27. Harmless Cud Chewer said

    Hmm, works in Internal Exploder, not in FlamingFox:/

  28. kiwipundit said

    Adam’s response (to Possum) on his blog does highlight one key point – that it wasn’t the upper-income, socially liberal “doctor’s wives” but the lower-income to middle-income socially conservative “battlers” that put Rudd and Labor in power. So it was Work Choices and interest rates not climate change and the war in Iraq that swung the election for Labor.

  29. Harmless Cud Chewer said

    Rather interesting is the comparison between say Robertson and Paterson. The projection puts them at -0.17 and -0.07. Yet the declared result is 0.11 and -1.51 respectively. Could this be a clue about the difference between actively campaigning in a seat (Robertson) and running dead (Paterson) ?

    Also interesting is the difference between the Higgins projection of -2.86 and the near final result -7.01. Definitely a demographic bump in that one.

    -moo-

  30. Chris said

    Well I picked ALP on 83 seats using nothing more than my than my finger and my arse.

    Take that Carr!! Take that multivariate analysis!!

  31. Ron said

    I should have added I did understand this was a forecast method

    I accept the formula is not an election forecast BUT a calculation of how many seats would be won IF the election is held at that time

    So the formula translates the poll at any time into seats won at that time

    Therefore we are back to assessing if the Poll itself represents reality given circumstances then may affect the poll and campaigns can change voting dynamics

    Maybe the forecast answer lies in how to weight the respective polls or weighting questions like preferred PM etc.

  32. JP said

    HCC @ 29

    Same for Cowper. Projections +0.13. Ran dead. Result -1.23. Bugger!

  33. caf said

    “Then we just add these swings to the seat margins according to the type of seat e.g. NSW marginal seats all have 5.64 added to their margin…”

    Shouldn’t the swing be adjusted based on the actual margin in each seat? Eg the swing represents a certain number of Coalition voters changing their vote to ALP, so even if the average swing was uniformly distributed across the country, it would be proportionally higher in seats with a higher Coalition vote.

    An example – the national average swing was 5.6%, on a nationwide Coalition TPP of 52.7 – so we expect 0.11% swing for every % of Coalition TPP, on average. Applying this to say Fraser (Coalition TPP of 36.7% before the election) gives an expected swing of 4.0%, if the 5.6% national swing was uniform.

    You could obviously apply this with finer granularity in your pendulums-within-pendulums approach. Does it make much difference to the result?

  34. David Richards said

    Yes JP – if the ALP had not run dead in those seats.. they might have won even more. You can be sure that with the benefits of incumbency they won’t run dead in any seats next time.. and there are a lot more seats within reach now. If they run a strategy just to hold the seats they won this time, that might prove to be a mistake. They should go for all those seats that are under 2%, as well as holding the seats they hold by under 2%. That is the likely swing either way for a first term government, given past results.

  35. Possum Comitatus said

    Caf, the swings that are applied actually have whatever differing proportion of higher (or lower) Coalition votes changing built in, as the three swing types (the safe Coalition seats, safe ALP seats and marginal seats) represent those differing proportions – even if it’s only been split into three sizes.

    So those seats with higher proportions of Coalition voters have the swing for the safe Coalition seats applied, where the swing for those safe seats measured in the polling partially captures any proportional effect, for example.

    I have played around with applying swings on a more continuous proportional basis and it didnt seem to make a great difference, but I might try it again with the election data to see if it works better now we have a set of results to verify it against.

  36. KeepingALidOnIt said

    OOOOhhhhhh (small AHA moment)

    This is the key: “It’s not about a prediction, it has absolutely nothing at all to do with prediction and for the life of me I’m bamboozled where people keep coming up with this idea. It’s about how to turn a set of data (marginal, safe Coalition and safe Alp seat swings, plus 5 state swings – what the Newspoll quarterly seat breakdowns provide) into the most accurate estimate of the number of seats that would change on that polling outcome.

    The process was misunderstood by some earlier on back in September, so to demonstrate the *process* I’ve simply substituted the type of data that we get from the big Newspolls with the same type of data from the actual election result, to show that the methodology does what it’s supposed to – estimate the number of seats that would fall given that type of data.”

    I understand now. I geddit. I think most people reading these blogs do think that it’s all about prediction. And Adam is right – for most of us these blogs are about “expert” predictions confirming our own prejudices hopes and dreams.

    We should read things more closely. So much for informed readers.

  37. psephoblog said

    I have made a (final) comment on this controversy at my webolg here:
    http://psephoblog.wordpress.com/2007/12/13/stirring-the-possum/

  38. psephoblog said

    By “webolg” I of course mean “weblog.”

  39. JP said

    David Richards @ 34

    Yes, I think you’re right. As long as they don’t screw up too badly, seats like Cowper and Paterson should be ripe for the plucking next time around. Now that Lyne no longer have the Deputy PM to vote for, or even the Nationals leader, Labor could sweep the NSW North Coast. But they’ll need to try, which I’m sure they will.

    Interesting that the “runnng dead” result in Cowper and Paterson was the same – about 1.3% less than the projection. Could that be the value of a lopsided campaign? Are there any other electorates where the result was out by a similar amount after the ALP ran dead?

  40. Gippslander said

    There’s no more enjoyable sport than “let’s you and him fight”, specially if the rapiers are all buttoned. the beer and popcorn are alone worth the admission price.
    In the heady days of September I made a comment on Simon Jackman’s site that I would be unavailable to a phone pollster because I automatically hang up on unsolicited phone calls. I think that if I didn’t know for whose benefit a face to face poll was being made I would also refuse to participate. I’m to the right of many posters here (but not Dr Carr). So there might be some truth in Morgan’s hypothesis about refusers being conservative. However it must be said the in the absence of an oxymoronic poll of people who refuse to be polled, it’s just something he pulled out of his a*se (like so many other of his “statistics”).
    As Possum says, polls are not predictive.
    Dr Carrs’s seat of the pants methods are also not predictive.
    Nothing is predictive.
    So why bother? certainly post facto analysis by folks like Possum provide valuable insights for practical operatives like Adam to shape future tactics.
    I’m not so sure that polls taken during a campaign are much help, due to large MOE. Even secular movement in polls is highly doubtful for tactical purposes. (I would have thought, in my biased way, that Rudd comprehensively won nearly every day of the campaign, but there was precious little evidence in the polls).
    I think the real stars in the psephology field were Crosby Textor, who fairly accurately mapped the unfavourable electoral terrain for the Coalition.
    Thanks, Possum, for an invariably interesting site (but 13 parameters to describe garbage like the Morgan polls?). And thanks, Adam, for making an old man very happy by not prefencing FFP (you also avoided the terrible risk of carpal tunnel syndrome inherent in below the line voting)

  41. CL de Footscray said

    Poss, you still de man as far as I’m concerned (or still the possum, I suppose, to be pedantic).

    Forecasting of anything is a complete pain in the a#*e and ties one in knots, as I know from having tried to do it in diverse applications from time to time, but I do know that anyone who can get as close as you did, and make perfect sense in explaining how you got there, is doing well. And it seems quite clear to me that there were swag of seats where the final outcome was decided by less than 1% – say less than 1,000 votes -which had they fallen the right way could have pushed the ALP up to 87 or even 89 (my own prediction at Pollbludger). Anyway, Adam makes a comment about not ‘trusting’ statistical analysis that no-one believes. Unfortunately if we’re interested in knowing what’s happening out there we have few options. As they say in flying school, trust the instruments. Nothing else will tell you which way is up!

    Anyway, I’m just enjoying Julia PM (acting), the tall bald eniviromnment Minister and the gay Climate change Minister of Chinese antecedents. Suddenly it’s the 21st century!

    Thanks again for all the fun during the elction and before. It was a hoot!

  42. Ron said

    14/12/07
    Hello Possum ,

    I have looked at your ‘swings formula’ in a purely clinical manner
    and have since read & re-read your post on Adam’s blog carefully

    My conclusion is that I may be talking at cross purposes to your intent.
    So perhaps I would like to be a little more precise in my
    statistical understanding to you so that where that differs you may wish to correct

    My understanding after 6hr is that your ‘pendulum swings formula’:

    1/ does not predict anything at all about a future election result

    2/ produces a seat count (later defined by me) for each Party
    as at the quarterly Newspoll Poll date ONLY

    3/(a) uses the Quarterly Newspoll data of swings by State and
    National swings by 3 seat type categories (of safe Liberal ,
    safe labor and Marginal).

    The National 3 seat categories swings are converted
    (by ratio to the State swings) into State swings
    by State seat category (of safe Liberal , safe Labor &
    Marginal) by state
    ..safe seat being over 6% swing..then the correction adjustment

    (b)Then the adjusted swing per State seat per category is
    applied to each respective State category seat per State
    to produce for each individual seat, a seat won result

    (c) aggregating all individual seat won results then provides
    a total seat count by Party IF an election had been held
    on the date of the qtr. Newspoll & based on that Newspoll data

    ie. it is a translation of Newspoll’s %’s into a “seat count”
    as an ‘alternative’ to the McKerras % Pendulum of seat numbers

    4/ “seat count’ is my term because I am unsure ;
    (a)whether your ‘pendulum swings formula’ SOLE PURPOSE is
    to provide a National total seat count by Party only , or

    (b)whether the your ‘pendulum swings formula’ PURPOSE is also
    to provide a State seat count by Party as well, or

    (c)whether your ‘pendulum swings formula’ PURPOSE is also to
    provide (a) and (b) AND a state seat count by category by Party

    5/ Notwithstanding your advice re 4/ (a) ,(b) or (c) ,
    the National seat count for each Party is an aggregation
    of the number of seats within each state AND within each of
    the three seat categories per State calculated as per 3/

    that would be won by each Party IF an election had been held
    at the Newspoll Poll date

    6/(a) your ‘pendulum swings formula’ DOES nominate specific seats
    per Party within each State and within each of the three seat
    categories per State

    that would be won by each Party IF an election had been held
    at the Newspoll Poll date
    (the aggregation of which provides the National total count
    for each Party)

    (b) BUT your ‘pendulum swings formula’ ASSUMES EITHER:

    (b)(i) that if ANY nominated specific seat ‘to be won’ by a Party
    in any of the 3 categories in any State are not won by a Party

    THEN another seat in that specific category in THAT State will
    be ‘won’ by that Party to offset the nominated specific seat
    that was not won

    EG. IF your ‘pendulum swings formula’ did nominate North Sydney
    to be ‘won’ by Labor in the ‘safe Liberal seat’ category of
    NSW State (being a plus 6% swing safe Liberal seat.)

    THEN your ‘pendulum swings formula’ ASSUMES that
    if North Sydney is not won by Labor ,
    then another ‘safe Liberal seat’ category seat in NSW will be
    won in its place by Labor

    OR

    (b)(ii) that if ANY nominated specific seat ‘to be won’ by a Party
    in any of the 3 categories in any State are not won by a Party

    THEN another seat in that specific category in ANY State
    (BUT NOT NECESSARILY IN THE SAME STATE as the seat NOT ‘won’)
    will be ‘won’ by that party to offset the nominated specific
    seat in that category that was not won by the Party

    OR

    (b)(iii) that if ANY nominated specific seat ‘to be won’ by a
    Party in any of the 3 categories in any State are not won
    by a Party

    THEN another seat in ANY category (but still in that State)
    will be ‘won’ by that party to offset the nominated specific
    seat in one of the 3 categories in that State that was not won

    7/ the naming of nominated specific seats to be won IS relevant only in identifying which of the 3 categories in each State
    will produce a calculated number of seats won by each Party

    but is not relevant in saying that a specific seat in a category in a State will be won by a Party
    (only that if that seat is not won , then another will be won by that Party to replace it and further that the replacement seat will come from whatever answer you give to above 6/(b)(i) or 6/(b)(ii) or 6/(b)(iii))

    PS/
    I have put my understanding of the formula in point form to allow you easily to quote your agreement or disagreement by point number
    rather than you needing to repeat my words

    I do have some possible suggestions for the formula for you to consider but they are academic until my understanding is on the same page as you

    regards

  43. caf said

    After some further thought, I think the most interesting thing in all of this is actually the swing histogram, showing that the distribution of swings across seats approximates a normal distribution with a standard deviation of 3.0.

    Have you done this analysis for older Federal elections? I would be interested to know if the standard deviation is generally constant. This would allow us to convert a particular approximation to the mean swing, say from a poll, into the probability of a particular seat falling. If you do this for every seat you can create an overall probability distribution showing the probability of each possible “total seats” outcome – this is like a pendulum analysis where instead of getting a black and white answer (“ALP will get X seats”), we get an (hopefully more realistic) answer with shades of grey (“ALP is X% likely to get at least Y seats”). This essentially removes the underlying approximation “swings are uniform” and replaces it with “swing are normally distributed with known variance”.

    Thoughts?

  44. David Richards said

    Ron, you don’t happen to work writing bureaucratic regulations, do you?

  45. Ron said

    David Richards

    the first sign of ignorance is non understanding of converging variables

Sorry, the comment form is closed at this time.

 
%d bloggers like this: