BludgerTrack 2013 methodology
BludgerTrack 2013 is an aggregate of all published national opinion polls: Newspoll, Nielsen, Galaxy, Essential Research and Morgan. Each poll is adjusted to account for observed biases and weighted according to sample size and past reliability. Local regression analyis (LOESS) is used to plot trendlines through the available data and determine current results based on the modelled results for the most recent available point in time. State-level polling is used to measure each state’s deviation from the national trend, which is in turn used to produce vote and seat projections for each state.
The preferred method for determining bias is to determine an average error based on pollsters’ final pre-election polls against the ensuing election results. However, because federal elections are infrequent and only a few pollsters conduct state election polling, in most cases the number of observations available within an acceptably recent time frame is prohibitively small. The two exceptions are Newspoll and Galaxy, for which the seven most recent case studies have been used, respectively going back to 2009 and 2007. A further exception in this respect is Morgan phone polls, for which I am content to use a performance average going back to 2007 despite the small number of case studies owing to its record of accuracy over the longer term. For other pollsters, bias measures are derived from a trend measure of their deviations from the BludgerTrack model, to be re-calculated monthly.
Pollsters are also weighted according to their historical accuracy, so that polls with better records have a greater influence on the total result. In most cases this has been determined by measuring the historical accuracy of their final pre-election polls. In these cases, the observations used are either the most recent ten results or, where fewer then ten are available, all results going back 10 years. For two newer poll series, Essential Research and Morgan’s multi-mode polling, poll results over the current term have been compared with the aggregated BludgerTrack results at the time of the poll’s publication.
For each pre-election poll, the accuracy of the primary vote results for Labor, the Coalition and the Greens are measured using a normal distribution function, resulting in a percentage figure representing the proportion of poll results that would be expected to be less accurate when taking into account the poll’s sample size and theoretical margin of error. A composite of the three results is then produced with the result for each party weighted according to its share of the vote. The final accuracy measure is derived by indexing each pollster’s composite result against Newspoll’s. This means Newspoll’s accuracy measure is always 1, with other pollsters rated either higher or lower depending on whether they are found to be more or less accurate.
It should be noted that polls with larger samples tend to have lower accuracy ratings, as their lower theoretical margins of error mean they are held to a higher standard in the accuracy weighting calculations. This is especially notable in the case of the Morgan multi-mode poll, for which sample sizes are around 3000.
National and state result aggregation
A total weighting score for a given poll is determined by multiplying its sample size by its accuracy measure. Pollsters which conduct surveys on a weekly basis, namely Essential and Morgan multi-mode, have their sample sizes halved to prevent them from dominating the overall result. The trend charts are then generated from the bias-adjusted poll results using LOESS, with the modelled results for the most recent available date used to determine the headline BludgerTrack results. For the states and territories, deviations from the national result are determined using all available state-level data, with LOESS used to calculate a trend measure for the most recent available point in time. This includes both published and, from those pollsters kind enough to provide me with it, unpublished data. The deviation measurements are then applied to the national BludgerTrack totals to determine state voting intention results.
Electorate two-party projections
Two-party results are projected for each electorate based on the state swing figures, subject for adjustment where a seat is being vacated by a sitting member, or where a member first elected at the previous election can expect to enjoy a “sophomore surge”. The effect of these factors has been determined using regression modelling based on historical election results going back to 1993, which takes into account their tendency to have a greater impact in non-metropolitan electorates.
Seat totals for each state are determined by adding together win probabilities for each electorate and rounding to the nearest whole number, as distinct from the blunter “Mackerras” technique of assuming uniform swings and marking off seats with lower margins (though in practice it will rarely make much of a difference). Win probabilities are determined by assuming normally distributed swings and determining the likelihood of a result below 50% based on a seat’s two-party projection. Historical observation has been used to determine that a highly significant relationship of y=0.131x+0.019 exists between the standard deviation (y) of swings within a given state and the state’s overall swing (x). This calculates as a 1.9% standard deviation in the event of no swing, increasing to 3.2% for a swing of 10%.
Independent and minor party seats
Since BludgerTrack is essentially a two-party model, complications are presented by the five electorates which were not won by the major parties in 2010 — New England (Tony Windsor, Independent, NSW), Lyne (Rob Oakeshott, Independent, NSW), Kennedy (Bob Katter, Independent, Queensland), Denison (Andrew Wilkie, Independent, Tasmania) and Melbourne (Adam Bandt, Greens, Victoria). With the retirements of Windsor and Oakeshott, New England and Lyne are being treated as two-party contests in which the Labor-versus-Coalition two-party preferred count which the AEC conducted after the 2010 election are used as the starting point for the determinations. Since these measures respectively showed Nationals majorities of 16.8% and 12.4%, they are effectively being allocated to the Nationals. Kennedy, Denison and Melbourne will be allocated to the incumbents until electorate-level polling evidence emerges to suggest otherwise.