This article was downloaded by: [University of West Florida]

On: 09 October 2014, At: 12:30

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Australian Journal of Political Science

Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/cajp20

Pooling the polls over an election

campaign
a
Simon Jackman
a
Stanford University
Published online: 15 Aug 2006.

To cite this article: Simon Jackman (2005) Pooling the polls over an election campaign, Australian
Journal of Political Science, 40:4, 499-517, DOI: 10.1080/10361140500302472

To link to this article: http://dx.doi.org/10.1080/10361140500302472

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
 Australian Journal of Political Science,
Vol. 40, No. 4, December, pp. 499– 517

Pooling the Polls Over an Election Campaign

SIMON JACKMAN
Stanford University

Poll results vary over the course of a campaign election and across polling
organisations, making it difficult to track genuine changes in voter support. I
Downloaded by [University of West Florida] at 12:30 09 October 2014

present a statistical model that tracks changes in voter support over time by
pooling the polls, and corrects for variation across polling organisations due to
biases known as ‘house effects’. The result is a less biased and more precise
estimate of vote intentions than is possible from any one poll alone. I use five
series of polls fielded over the 2004 Australian federal election campaign
(ACNielsen, the ANU/ninemsn online poll, Galaxy, Newspoll, and Roy
Morgan) to generate daily estimates of the Coalition’s share of two-party
preferred (2PP) and first preference vote intentions. Over the course of the
campaign there is about a 4 percentage point swing to the Coalition in first
preference vote share (and a smaller swing in 2PP terms), that begins prior to
the formal announcement of the election, but is complete shortly after the
leader debates. The ANU/ninemsn online poll and Morgan are found to have
large and statistically significant biases, while, generally, the three phone polls
have small and/or statistically insignificant biases, with ACNielsen and (in
particular) Galaxy performing quite well in 2004.

Polls are the lifeblood of media coverage and punditry during an election campaign.
Each major newspaper has a contractual arrangement with a polling organisation,
providing them with poll numbers—grist for the journalistic mill, as it were—a
critical component of the campaign coverage in the media. As an election draws
closer, the tempo of polling increases, as does the number of polling organisations
in the field, and the intensity with which the poll numbers are studied. In sum,
media polls are perhaps the most readily available information on how the parties
are faring over the campaign.1

Simon Jackman is an Associate Professor in the Department of Political at Stanford University and
Director of the Political Science Computational Laboratory. An earlier version of this paper was
prepared for the annual meeting of the Australasian Political Studies Association, University of
Adelaide, 29 September– 1 October 2004. I thank David Gow, Andrew Leigh, Kevin Quinn and
Justin Wolfers for useful discussion, and Randall Thomas for some useful references.
1
Other sources of information include the prices offered in election betting markets (eg Wolfers and
Leigh 2002) but my focus here is strictly on the polls.

ISSN 1036-1146 print; ISSN 1363-030X online/05/040499-19 # 2005 Australasian Political Studies Association
DOI: 10.1080/10361140500302472
 500 S. JACKMAN

Although media-commissioned polls are central to the media’s coverage of the

election campaign, they are actually of limited use for political scientists. As I
show below, media-commissioned polls employ sample sizes that are too small to
reliably detect the relatively small day-to-day or week-to-week movements in
voter sentiment we would expect to occur over an election campaign. Thus, as
they currently stand, media-commissioned polls are essentially of ‘news value’
rather than social-scientific value. As Murray Goot explains:
The problem in the reporting of the polls does not arise, fundamentally, from the
fact that journalists are ill-trained to deal with such data—though that to some
extent is true. Fundamentally, the press plays up differences which are otherwise
insignificant because it has to. Its only alternative is to say that what a poll
found today is not significantly different from what it found yesterday; and under
Downloaded by [University of West Florida] at 12:30 09 October 2014

most (though not all) circumstances, that sort of news is no news at all. (Goot
2000, 46)
Likewise, in commenting on the polls in the 2004 election, Peter Brent (2004)
observes that
[q]uantitative opinion polls aren’t that precise. But the process that pays for them
pretends they are. They [polls] cost a bundle and so are given pride of place. Once
they’re there, everyone involved goes along with the charade.
Beyond endorsing these observations, my goal is not to dwell on or further criticise
the media’s reporting of poll results. Rather, my goal here is more constructive, to
present tools that make better use of published polls, so as to more reliably track
changes in voter sentiment over the course of an election campaign. In short,
while the precision of any one poll is quite limited, I show that we can systematically
combine the information in the published polls, leveraging them against one another,
so as to obtain a clearer picture of what might be going on in the electorate over the
campaign.
In the next section I discuss the limits of polls, highlighting (1) imprecision due
to sampling error; and (2) the possibility of bias, due to the procedures and
methods employed by the particular polling organisation. Pooling the polls helps
deal with the first problem (more data are better than less), but is a valid strategy
only if polls are all measuring the same thing (ie each poll produces an unbiased
estimate of the current state of voter support). In addition, (3), voter support is
likely to be changing over the course of a campaign, while polls provide snapshots,
and imprecise and possibly biased snapshots. I then present a statistical model that
addresses these three issues simultaneously: ie (1) the model pools the polls, over-
coming the limits to precision inherent in any one poll; (2) the model smoothes
over time, consistent with the notion that although support for either party fluctuates
or trends over the course of a campaign, each campaign day need not be considered
de novo—yesterday’s level of Coalition support will generally be an excellent
predictor of today’s level of support; (3) the model estimates and corrects for the
possibility that any single poll is subject to bias or ‘house effects’ (bias induced by
methodological procedures specific to each polling organisation). I apply this statisti-
cal model to polling data generated in the lead-up to the 2004 Australian federal
election, drawing on data from five polling organisations, spanning traditional tele-
phone interviewing through to face-to-face and self-selected samples completing
surveys via the Internet.
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 501

The chief benefit of pooling poll results (after correcting for house effects) is that
we are much better positioned to ascertain movements in levels of voter support in
response to campaign events. As it turns out, there is simply not a tremendous
amount of volatility over the course of the 2004 campaign, at least as far as we
can detect it with the available polling data and the model I use here: to foreshadow
one of my findings, in terms of aggregate, two-party preferred (2PP) voting inten-
tions, there was a steady shift towards the Coalition over the opening weeks of the
campaign, and little movement thereafter. Larger movements are apparent in the
Coalition’s share of first preferences over the course of the campaign. But a striking
feature of the 2004 polling data is that a good portion of the differences across polls
is not due to movements in voter sentiment but to differences in the polls’ method-
ologies. The model and analysis I present here lets us distinguish between these
different sources of variation in polling data: movement due to fluctuations over
Downloaded by [University of West Florida] at 12:30 09 October 2014

the campaign, sampling error, and non-sampling error (bias) specific to each
polling organisation (‘house effects’).

The Limits of Polls: Margins of Error and Sample Size

A poll’s reported ‘margin of error’ is almost always a 95% confidence interval around
the poll’s estimate of the 2PP vote split, provided by the approximation (valid for
large samples)
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a^ (1
a^ )
a^ + 1:96 , (1)
n

where a^ [ ½0,1 is the poll estimate of a, some proportion of interest (eg the
proportion of 2PP vote intentions for a particular party) and n is the sample size.
The 1.96 in equation (1) comes from the fact that (1) via the central limit theorem,
uncertainty about a statistic such as a^ computed with a large sample follows a
normal distribution, and (2) if a random variable follows a normal distribution,
then, with probability 0.95, it lies within 1.96 standard deviations of the mean of
that normal distribution.
Equation (1) highlights the fact that statistical precision increases in the square root
of the sample size, so that increases in sample size produce diminishing marginal
gains in statistical precision. But the marginal cost of an additional survey respondent
remains more or less fixed, or may even be increasing in sample size. Since time and
money are not infinite, there are limits to the precision we can reasonably expect from
any given poll. Most media polls are conducted with sample sizes in the neighbour-
hood of 1000 –1500, with larger samples of above 2500 being fielded in the days
immediately prior to the election. These sample sizes are simply too small to reliably
detect small fluctuations in support for the parties over the course of an election cam-
paign. Routine statistical calculations2 reveal that sample sizes of the order of 65,000
per sample are required to give a researcher a 95% chance of detecting a true 1 per-
centage point change with a conventional 95% level of confidence; if the researcher
is more risk acceptant and is willing to accept a lower level of statistical significance
2
See, for example, Fleiss, Levin and Paik (2003, ch. 4). For the specific context of sample sizes required
to assess campaign effects, see Gow (2001).
 502 S. JACKMAN

the sample size requirements become slightly less onerous, but still well beyond
the resources deployed in media-commissioned polls. For instance, to give a
researcher an even-money chance of detecting a 1 percentage point change using a
90% confidence level, the researcher requires roughly 13,500 respondents per sample.
Larger changes in voter support are easier to detect, in the sense of requiring fewer
respondents. Figure 1 summarises the sample size requirements over a range of
assumed differences between polls. For instance, around 4000 respondents per poll
are sufficient to detect changes of 4 percentage points (again, in the sense that a
researcher has a 95% chance of rejecting the null hypothesis of no change at the
traditional benchmark of 95% confidence levels); for changes this large, a sample
size of 1200 per week is sufficient to give the researcher a 50– 50 chance of rejecting
the null hypothesis of no change at the 95% confidence level.
To summarise, the sample sizes used by the published media polls (eg the 1400
Downloaded by [University of West Florida] at 12:30 09 October 2014

respondents typically seen in the ACNielsen polls for the Fairfax papers) have a
reasonable chance of detecting moderate to large swings in voter support. But the
probability of detecting a small swing, say, of the order of a percentage point, is
less than 1 in 10, given the sample sizes used by most commercial polls. To the
extent that campaigns generate these small to moderate changes in voter support,
we clearly need more data than those provided by a single media-commissioned
poll if we are to detect it.

Figure 1. Sample size requirements. Notes: Each curve shows the sample size (vertical axis, log scale)
required to detect the indicated change in support (horizontal axis, assuming a baseline level of 50%),
with probability given by the label next to each line. In each instance it is assumed that the researcher’s
decision problem is whether to reject the null hypothesis of no change in favour of a two-sided,
alternative hypothesis, using a 95% confidence level or better (ie a p-value of 0.05).
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 503

Pooling the Polls

One way to get more data is to pool the polls. If two or more polls are in the field at
more or less the same time, then we have the potential to combine the information
in each poll, to arrive at an estimate that is more precise than any single poll.
Suppose that two polling organisations, A and B, generate unbiased estimates of a,
the Coalition’s 2PP vote share. Organisation A conducts a poll with sample size nA
and produces the estimate a^ A , and so, recalling equation (1), our beliefs about a
given A’s poll can be p represented with a normal distribution with mean a^ A and
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
standard deviation sA ¼ a^ A (1
a^ A )=nA . Likewise, we obtain a^ B and sB from organi-
sation B. Since A and B are trying to estimate the same quantity, we can combine the
information in their polls; in particular, it is a fact about normal distributions that the
pooled estimate is a precision-weighted average of the two separate estimates: ie
Downloaded by [University of West Florida] at 12:30 09 October 2014

pA a^ A þ pB a^ B
a^ AB ¼ , (2)
pA þ pB

where pA and pB are the precisions of the two polls, defined as 1=s2A andp1=s 2
B , respect-
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
ively. The standard deviation we obtain using the pooled estimate is 1=(pA þ pB ),
and so the pooled estimate is more precise (has smaller standard deviation) than any
one of the polls individually.

Pooling: An Example
Figure 2 presents a graphical illustration of the potential gains from pooling. Suppose
polling company A estimates the Coalition’s 2PP vote share at 53%, based on a
sample size of 1400 respondents, and so via equation (1) the poll has a ‘margin of
error’ of 2.6 percentage points. Suppose further that polling company B is in the
field at roughly the same time and estimates the Coalition’s 2PP vote share at 50%
based on a sample size of 2500 respondents. Because of the larger sample size,
company B’s estimate has a smaller margin of error than the estimate of company
A, of 2.0 percentage points. When pooling the two estimates, the greater precision
of company B’s estimate means we give more weight to it. Using equation (2), the
pooled estimate is the precision-weighted average, 51.1%, slightly closer to B’s
estimate of 50% than the simple unweighted average of 51.5%. The pooled estimate
has a margin of error of 1.5 percentage points, smaller than the margins of
error associated with either A or B’s results taken separately; after pooling the
results, we conclude that with 95% probability, the Coalition’s 2PP vote share lies
between 49.5% and 52.6%.

‘House Effects’: Bias in the Polls

Pooling polls will always enhance precision, but the validity of the pooled estimate
rests on a critical assumption: that the polls are unbiased. This is often not the case.
Polls are subject to bias, and, in particular, biases specific to particular polling
organisations, known as ‘house effects’ (the term ‘house’ here refers to a polling
company, not a sampled household). Variations in mode of interview (telephone,
face-to-face, or Internet), sampling and weighting procedures, the day of the week
 504 S. JACKMAN
Downloaded by [University of West Florida] at 12:30 09 October 2014

Figure 2. Pooling the polls: an example. Notes: The uncertainty attaching to each poll’s estimate can
be summarised with a normal distribution; see equation (1). Pooling the information in polls A and B
produces a normal distribution that has (a) its mean equal to a precision-weighted average of the two
polls, via equation (2), and (b) a smaller standard deviation than the normal distributions
summarising polls A and B, reflecting the fact that pooling is equivalent to combining information.

or the time of day a company interviews, the age, ethnic and gender composition of a
given company’s interviewers, question wording and question ordering are all poten-
tial sources of bias in survey research. A critical facet of polling in the Australian
context is how or whether to solicit second preferences from respondents who
report a first preference for a minor party; as we shall see below, differences on
this score can generate quite different estimates of 2PP vote shares (see also Brent,
2004). All survey companies make choices about these aspects of surveying, balan-
cing considerations of cost and timeliness against beliefs about the magnitude of the
biases likely to be introduced from the sources listed above; Groves (1989) provides a
book-length survey and McDermott and Frankovic (2003) examine the magnitudes
of specific sources of house effects in pre-election polling for the 2000 US presiden-
tial election. Again, for commercial reasons, survey companies tend to settle on a set
of procedures (mode of interview, sampling, instrumentation, weighting); thus, any
biases in one poll from a given organisation are often present in their other polls,
and a more or less fixed aspect of their estimates. But the consequence of house
effects is that sampling variation is not the only source of error. Put differently,
the biases in poll A and poll B may be such that they can’t validly be considered
to be estimating the same population quantity, and pooling them would be invalid.
On the other hand, there are commercial pressures to eradicate bias, and survey
research firms will review and occasionally alter their procedures to this end.
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 505

Election polling is somewhat unusual in this regard, since the election outcome itself
provides a highly visible benchmark against which to assess a poll’s estimates and,
by extension, the validity of the polling organisation’s methodology; see, for
example, the essays by Goot (2000, 2002, 2005) reviewing the performance of the
polls over recent Australian election campaigns. It is rare (but not unheard of) for
polling companies to change their procedures in the middle of a high-profile event
like a national election campaign;3 acting on the advice of an anonymous referee,
I contacted each of the polling organisations whose data I draw on here to verify
that their procedures remained constant over the course of the campaign (and I
received assurances that this was the case).
In summary, naively pooling the polls addresses one issue (the relative lack of
precision in any one poll), while assuming house effects are non-existent. In
general, this is a dangerous strategy: pooling two biased polls does not necessarily
Downloaded by [University of West Florida] at 12:30 09 October 2014

remove bias, since the pooled estimate will inherit some of the biases of each poll
(in proportion to the precision of each poll). If the biases offset (say, poll A has a
pro-Coalition bias and poll B has an ALP bias of roughly the same magnitude)
then pooling will tend to alleviate the bias of any one poll. But if the biases run in
the same direction, then the gain in precision from pooling results in higher levels
of confidence about a biased estimate.
This indicates that we require estimates of the bias in each poll as we attempt to
pool the polls. Of course, presuming that we have knowledge of each poll’s biases
is equivalent to knowing the population quantity being estimated. This is rarely
ever the case, at least before an election: indeed, if we knew the population quantity
of interest (eg the Coalition’s share of 2PP vote intentions), we wouldn’t be conduct-
ing polls! After the election, the actual election outcome provides a critical reference
point, from which we can calibrate the various polls and, in turn, estimate house
effects. I exploit this strategy in the analysis below.4

The Polls in 2004

I consider the following major national polls:
(1) ACNielsen, which provided polls for the Fairfax papers (eg the Sydney Morning
Herald and The Age);
(2) Newspoll, which provided polls for News Ltd, primarily its national daily The
Australian;
(3) Galaxy, which provided polls for other News Ltd metropolitan daily newspapers,
such as Sydney’s Daily Telegraph and the Brisbane Courier Mail;
(4) Roy Morgan, apparently without a media client in 2004, but formerly providing
polls to The Bulletin;
3
In the 2000 US presidential election campaign, some large fluctuations in poll estimates were due to
changes in weighting procedures. Another complication and likely source of bias in the United States is
the definition of a likely voter; these definitions vary across organisations and sometimes change within
organisations over the course of a campaign. See Erikson, Panagopoulos and Wlezien (2004).
4
Of course, this is not possible before the election: possible strategies for dealing with house effects
over the course of the campaign include (a) using past assessments of the performance of the polls
(although this isn’t feasible for new polling companies, with no history of election polling); (b) impos-
ing an assumption of some sort, such that the polls are collectively unbiased, such that, on average, the
polls get it right (as I did in a ‘pre-election’ version of this paper; as I show below, this assumption is not
supported by the data).
 506 S. JACKMAN

(5) the Australian National University, which provided daily polls from 12
September up until 8 October (election eve) for Consolidated Press Holdings
outlets such as ninemsn.com, Channel Nine, and The Bulletin.

Three of the five organisations use telephone polling, the exceptions being Morgan
(almost exclusively face to face) and the ANU/ninemsn.com collaboration (polling
via the Internet).
Between mid-June 2004 and the election on 9 October, these five organisations
published the results of 66 polls, comprising over 72,000 interviews via phone
(the dominant interview mode), face to face, and Internet. The tempo of polling
increased markedly after the formal announcement of the election. The data are sum-
marised graphically in Figure 3. In the figure, each poll’s estimate of the Coalition’s
share of the 2PP vote (vertical axis) is plotted over time with the mid-point of a poll’s
Downloaded by [University of West Florida] at 12:30 09 October 2014

field period taken as the effective date of the poll. The vertical lines in the figure cover
95% confidence intervals around each poll’s estimate of the Coalition’s share of 2PP
vote intentions, using the formula in equation (1). In the analysis, below, I also
examine the polls’ estimates of the Coalition’s first preference vote share.

Figure 3. Polls in the 2004 campaign and historical election outcomes. Notes: Each poll is represented by
a dot, with the vertical lines extending to cover a 95% confidence interval. The vertical axis is identically
scaled in each panel; note that the ANU/ninemsn series of daily polls starts during the campaign. The
vertical lines labelled 1 to 5 correspond to the following campaign events: (1) election announced,
29/8; (2) Jakarta Embassy bombing, 9/9; (3) Leader debate, 9/12; (4) Liberal Party campaign launch,
9/26; (5) ALP launch, 9/29.
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 507

The Evidence for House Effects in 2004

Figure 3 makes several features of the 2004 polls quite obvious. First, most of the
polls do not provide unambiguous evidence that the Coalition is leading the ALP
or vice versa. Telling exceptions are: (1) most of the Morgan polls, especially
those from July, August and early September, with relatively large sample sizes
(and hence smaller confidence intervals) that strongly suggest a sizeable
Labor lead; (2) a mid-August Nielsen poll that found Labor in the lead, and the
23– 24 September and the election eve Nielsen polls that estimated the Coalition’s
2PP vote share at 54%; (3) two Newspolls that again point to a Labor lead; (4) 12
of the ANU/ninemsn daily Internet polls strongly indicate a Labor lead; and (5)
the early August Galaxy polls showing the Coalition unambiguously leading the
ALP. Tellingly, the polls that do give an unambiguous lead to one party or the
Downloaded by [University of West Florida] at 12:30 09 October 2014

other point in contradictory directions. In short, the polls are all over the map:
some polls point to a Coalition win with high probability; some point to a Labor
win with high probability; and most polls fall somewhere in between.
To give some sense of the variability across the polls, the lower right panel of
Figure 3 shows the distribution of actual results in the 22 House of Representatives
elections from 1949 to 2001. The Coalition’s 2PP vote share has never been below
the 46.8% result recorded in 1983 and has never exceeded the 56.9% result recorded
in 1966, a range of 10.1 percentage points. Half of the Coalition’s 2PP results lie in a
range spanning 49.1 – 53.4%, or 4.2 percentage points. In contrast, the polls in the
second half of 2004 vary between a low of 42% (recorded in the ANU/ninemsn
online poll of 29 September) and a high of 54% (recorded in a Galaxy poll from
early August, and two Nielsen polls, including the Nielsen poll immediately
before the election). That is, the range of poll results is larger than the range observed
in 22 actual elections covering some 55 years of Australian political history.
Pursuing this point, note that the standard deviation of the 22 elections
(1949–2001) is 2.8 percentage points. This degree of variation roughly corresponds
to the uncertainty that accompanies a poll with a sample size of about 315 respon-
dents.5 How is it that polls using sample sizes typically in the neighbourhood of
1000 respondents display essentially just as much variability as the (nominally less
precise) historical data?
The data presented in Figure 3 provide a strong hint that ‘house effects’ are at work
in the 2004 polling data. The polls’ point estimates of Coalition vote share are widely
dispersed, to be sure; but the variation across polling organisations in the point
estimates is substantial. That is, large fluctuations within the data from any single
polling organisation are rare; eg the ANU/ninemsn estimates display the greatest
variation, in no small measure because they are based on relatively small daily
samples (an average daily sample size of 685). The Morgan polls seem to be
consistently on the low side of the estimates of the Coalition’s 2PP, while Galaxy
seems to be consistently on the higher side of the estimates. In fact, simple analysis
of variance (ANOVA) reveals that 41% of the variation in the polls’ point estimates
is house-to-house variation; an F-test leads to an overwhelming rejection of the null
hypothesis that there are no differences among the polling organisations
(F4,61 ¼ 10.61, p ¼ 0.001). Augmenting the analysis with various trend terms
5
That is, if a poll of 315 respondents generated an estimate of the Coalition’spvote share of 50%, then
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
via equation (1), a 95% confidence interval would be plus or minus 1.96 times 0:25=315 or about 1.96
times 0.028.
 508 S. JACKMAN

(to address the hypothesis that the house effects are confounded with possibly real
fluctuations in vote support) generates identical results.6 The extent of this house-
to-house variation is even more impressive when we recall that sampling error
alone is large and an unavoidable source of variation across the polls; ie even
without house effects and/or change in voter sentiment over the campaign, sampling
variation alone would generate considerable dispersion in the polls’ point estimates.
The model presented below provides a more rigorous assessment of house effects, by
explicitly dealing with both sampling error and the possibility that voter sentiment is
moving over the campaign period.

A Statistical Model for Pooling the Polls

Downloaded by [University of West Florida] at 12:30 09 October 2014

Here I present a statistical model that simultaneously tackles the three problems
discussed above: (1) pooling polls so as to increase precision; (2) estimating and
adjusting for the bias of any one poll; (3) tracking the trends and fluctuations in
voter sentiment over the course of the election campaign. Some notation will help
to clarify matters. Let at be the Coalition 2PP intended vote share at time t, with t
indexing days, where t ¼ 1 on 18 June 2004 (corresponding to the field date of the
first poll in my data set); below, I also consider the polls’ estimates of the Coalition’s
share of first preference votes. Let i ¼ 1, . . . , n index the polls available for analysis.
Each poll result is assumed to be generated as follows:

yi  N(mi , s2i ), (3)

where yi is the result of poll i. Each of the n polls is generated by organisation ji on

field date ti . si is the standard error of the poll (a function of yi and the poll’s sample
size; again, see equation (1)) and

mi ¼ ati þ d ji , (4)

where dj is the bias of polling organisation j, an unknown parameter to be estimated.

To model change in vote intentions, I use the following simple random-walk
model:

at  N(at
1 , v2 ), t ¼ 2, . . . , T (5)

with the distribution

a1  Uniform(0:4, 0:6) (6)

initialising the random walk (ie before we see any polling, I assume that Coalition
support is anywhere between 40% and 60%, bracketing the historical range of
election results reported above). In adopting this model I assume that vote shares
are locally constant, ie on average, today’s level of Coalition support is the same
6
For instance, adding a cubic polynomial in time to the analysis picks up an extra 0.03 in r2 ; the r 2 from
the regression with the cubic trend term alone is just 0.01.
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 509

as yesterday’s, save for random shocks that come from a normal distribution with
mean zero and standard deviation v.
The model is fit subject to the constraint implied by the actual election outcome;
that is, on 9 October 2004, at ¼ 0:5274 (ie the Coalition received 52.74% of the 2PP
House of Representatives vote, according to the Australian Electoral Commission).
This constrains the trajectory of the estimated at to culminate in the actual election
outcome, and, in turn, lets us estimate the dj parameters tapping house-specific
biases. This is an important constraint; without being able to anchor the estimated
levels of Coalition support to the actual election outcome, the model unravels, it
being impossible to simultaneously estimate underlying levels of support for the
Coalition and house effects. Accordingly, the model cannot be used for ‘real-time’ track-
ing over the course of the campaign unless we impose a priori restrictions on the
house effects; in fact, one goal of the current analysis is to provide estimates of house
Downloaded by [University of West Florida] at 12:30 09 October 2014

effects that then might be used to calibrate poll results in the next election campaign.
This model is essentially a Kalman filter, used in signal processing in engineering
applications, for tracking a moving target with noisy (and possibly biased) obser-
vations. Kalman filters are also used widely in time-series econometrics
(eg Harvey, 1989) and in political science, where they have been used to track presi-
dential approval (eg Beck 1990; Baum and Kernell 2001) and public opinion
(eg Stimson 1991; Green, Gerber and De Boef 1999). Equations (3) and (4) define
the measurement or observational part of the model, relating the observed poll esti-
mate to the latent target at , while equation (5) specifies the way in which the hidden
target moves over time. The unknown model parameters are: (a) 114 at parameters,
the daily levels of intended 2PP vote share for the Coalition over the 114 ‘campaign
days’ in my analysis; (b) the five bias parameters dj specific to each polling company;
and (c) v, the standard deviation of the normal distribution characterising day-to-day
volatility in the at . The data used to estimate these parameters are the survey results
yi and their standard errors si . Note that not all polling organisations report data on
every day; in fact, I have data from just 66 polls. At first glance, the paucity of
data relative to the number of parameters would seem to prohibit meaningful statisti-
cal analysis, but this is misleading: in the absence of polling data, equation (5) pro-
vides a predictive model for at , but constrained by the influence of previous polling
data on estimates of at
1 , at
2 , . . ., and of future polling data on estimates of
atþ1 , atþ2 , . . .. Estimation of the model is via Bayesian simulation methods discussed
in the Appendix. All computer programs and data are available upon request.

Results
Daily estimates of the Coalition’s share of 2PP vote intentions are shown in Figure 4.
The solid line connects the 114 daily estimates of the at parameters, and the dotted
lines indicate a 95% confidence interval around each daily estimate. As in Figure 3,
campaign events are indicated with vertical lines, and actual poll results are overlaid
with plotted points.
There is some evidence of trending over the campaign in the data, with a small rise
in the Coalition’s share of the 2PP vote after the election was formally announced on
29 August 2004. For this date, the estimate of the Coalition’s share of the intended
2PP vote is 50.6%, with a 95% confidence interval ranging from 48.2% to 52.6%.
The final election result of 52.7% lies just at the upper limit of the 95% confidence
interval of Coalition support at the formal start of the campaign; put differently, the
 510 S. JACKMAN
Downloaded by [University of West Florida] at 12:30 09 October 2014

Figure 4. Estimated Coalition share of two-party preferred vote intentions, and pointwise 95%
confidence intervals. Notes: The shaded area covers the 95% confidence intervals around the
estimated levels of Coalition support, given the model and the polls. Individual polls are represented
with a plotted point at their respective point estimates. See Figure 3 for campaign events.

probability that the Coalition gained support over the campaign is 0.977 (ie we would
reject the null hypothesis of ‘no change over the campaign’ with greater than the
conventional standard of 95% confidence).
My estimates also indicate that Coalition 2PP support peaked on 16 September,
some three weeks before the election (but just after the leader debates) at 52.8%
of the intended 2PP. Of course, there is considerable uncertainty accompanying
this point estimate; a 95% confidence interval ranges from 51.0% to 54.5%, indicat-
ing that this ‘peak’ estimate is indistinguishable from the actual election outcome
(52.7%). In sum, the estimates reported here indicate that the Coalition’s gains
over the campaign amounted to about 2 percentage points, almost the size of the
overall swing to the government in the 2004 election, and that they came relatively
early in the campaign, as events such as the Jakarta embassy bombing and the
leaders’ debate took place. The formal policy launches of both major parties seem
to have made no impact on aggregate, 2PP vote shares.
The benefits of pooling data are also apparent from Figure 4. The width of the
shaded area—a 95% confidence interval around the estimated at , the Coalition
share of 2PP vote intentions—is substantially smaller than the confidence intervals
attaching to any one poll. Moreover, as Figure 5 highlights, the width of the 95%
interval is falling over time as more polls are published in the closing weeks of
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 511
Downloaded by [University of West Florida] at 12:30 09 October 2014

Figure 5. Width of 95% confidence interval around daily estimates of intended Coalition 2PP vote
share. Notes: The axis on the right-hand side of the graph shows the implied sample size of a
hypothetical poll if it were to generate a confidence interval as precise as that obtained by the
pooling model.

the campaign. In particular, the constraint that estimated series of at culminate with
the actual election result on 9 October sees the confidence interval shrink towards
zero (and the implied sample size tend to infinity) in the last days of the
campaign. A more realistic appraisal of the benefits of polling is the pattern, say,
over the second last week of the campaign, where polling is quite frequent and the
benefits of pooling and smoothing are quite substantial: in this time period the
95% confidence intervals are never more than 3.5 percentage points in width, corre-
sponding to daily polls with sample sizes of the order of 3000 respondents.
A similar picture emerges when we turn to consider the polls’ estimate of first
preferences. Figure 6 presents the estimated daily track of first preference vote
shares, generated by applying the model presented above to the poll data on
Coalition first preferences. The Coalition’s actual share of the first preference
vote, 46.7%, is statistically distinguishable from the Coalition share of first prefer-
ence vote intentions as estimated at the start of the series (43.5%, on 18 June,
p , 0.01), at the estimated low point in Coalition support on 14 August (42.3%,
p , 0.01), and on the day the election was formally announced (44.0% on 29
August, p , 0.01; all one-sided p-values). On the other hand, the Coalition’s
maximum level of estimated support (47.2%) is not statistically distinguishable
from the actual election result of 46.7% ( p ¼ 0.29, one-tailed). Thus, the overall
pattern is extremely similar to that which we observe for the 2PP estimates (see
Figure 4): ie a shift towards the Coalition that appears to pre-date the formal
announcement of the election, levelling off after the leader debate, and with no
 512 S. JACKMAN
Downloaded by [University of West Florida] at 12:30 09 October 2014

Figure 6. Estimated Coalition share of first preference vote intentions and pointwise 95% confidence
intervals. Notes: The shaded area covers the 95% confidence intervals around the estimated levels of
Coalition support, given the model and the polls. Individual polls are represented with a plotted
point at their respective point estimates. See Figure 3 for campaign events.

statistically discernible movement thereafter. This movement towards the Coalition

in terms of first preferences is over 4 percentage points, a large and politically
consequential shift by any reasonable standard,7 and is considerably larger than
the movement in 2PP vote intentions.

Estimates of House Effects

Table 1 summarises the estimates of house effects, presenting point estimates and the
lower and upper limits of a 95% confidence interval around the point estimates
(ie point estimates with 95% confidence intervals overlapping zero are indistinguish-
able from zero at the traditional 95% confidence level). Results are provided for both
the 2PP data and the polls’ estimates of first preference vote shares.
For the 2PP estimates, three out of five organisations have estimated bias
parameters that are distinguishable from zero at conventional levels of statistical
7
Consider that the 4.4 percentage point movement from the Coalition’s lowest level of estimated first
preference support (42.3% on 14 August) to the actual result of 46.7% corresponds to the difference
between a clear victory (ie the 2004 result) and a defeat (1972, 41.5%) or the slimmest of wins
(1961, 42.0%).
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 513

Table 1. Estimates of house effects parameters, and limits of 95% confidence intervals. Positive house
effects imply a systematic overestimate of Coalition support. All terms are expressed as percentage points
Two-party preferred First preferences

Estimate 2.5% 97.5% Estimate 2.5% 97.5%

AC Nielsen 21.1 22.8 0.6 1.1 20.8 3.0

ANU/ninemsn 25.7 27.3 24.2 26.0 27.7 24.2
Galaxy 0.1 22.0 1.8 0.4 21.6 2.5
Newspoll 22.9 24.6 21.2 21.4 23.3 0.4
Roy Morgan 24.9 26.6 23.2 23.7 25.5 21.8
Average 22.8 24.3 21.2 22.0 23.7 20.3
Average of Nielsen/ 21.3 22.9 0.2 20.1 21.9 1.8
Galaxy/Newspoll
Downloaded by [University of West Florida] at 12:30 09 October 2014

significance, and all tending to underestimate Coalition support. The ANU/ninemsn

online poll has the largest bias, underestimating Coalition support by an average of
5.7 percentage points; to better appreciate the magnitude of this bias, recall that
between 1949 and 2004 actual Coalition 2PP vote shares have varied in a range of
10 percentage points, with an inter-quartile range of 4 percentage points. Morgan
has the next largest bias in these data, resulting in an average underestimate of
Coalition support of 4.9 percentage points; note that this estimate is based on
treating the Morgan house effect as constant across the face-to-face polls and
the one, final phone poll it fielded. Finally, Newspoll is estimated to have a 2.7
percentage point bias towards Labor that is statistically significant at the
conventional 95% level. Recall that the pre-election Newspoll estimated Coalition
vote share at 50%, based on a large sample of 2500 respondents; the disparity
between this poll and the actual election result almost exactly mirrors the bias
estimate, as it does for the other polling organisations. Newspoll fares better in
estimating first preferences, with the bias parameter indistinguishable from zero at
the conventional level of statistical significance. Both the ANU/ninemsn Internet
poll and Morgan have large and statistically significant biases in their estimates of
Coalition first preference vote shares, again systematically underestimating Coalition
support.
Galaxy and Nielsen are estimated to have negligible biases in both their estimates
of 2PP and first preference vote intentions (in the sense that the corresponding bias
parameters cannot be distinguished from zero at conventional levels of statistical
significance). Again, these negligible bias estimates reflect the fact that Galaxy
and Nielsen’s final poll estimates were quite close to the actual election outcome.
Although Nielsen’s last poll estimated the Coalition’s 2PP vote share at 54%
(an overestimate), over the course of the campaign Nielsen (like most polling
organisations) tended to underestimate Coalition support, resulting in a negative
bias estimate, albeit not statistically significant at conventional levels. Galaxy
appears to have had an excellent campaign, with its estimated bias parameters
almost exactly zero for both first preferences (0.4 percentage points, plus or minus
about 2 percentage points) and 2PP (0.1 percentage points, again plus or minus
about 2 percentage points).
It is also telling that the average bias of Nielsen, Galaxy and Newspoll is virtually
zero on first preferences, but the 2PP average bias is over a percentage point
 514 S. JACKMAN

(underestimating Coalition 2PP support) and on the cusp of statistical significance.

Differences in the ways the polling houses arrive at 2PP estimates are certainly
implicated here: eg both Morgan and Newspoll have larger 2PP bias estimates
than they do on first preferences, while Galaxy used minor-party preference
flows observed at the 2001 election to estimate 2PP; see Brent (2004) for additional
detail.
The house effects estimated here arise from numerous sources (as discussed
above). But it is interesting to note that the largest house effects are associated
with the mode of interview. The ANU/ninemsn Internet poll produces the largest
biases, followed by Morgan (face to face), and the three telephone polls. The prob-
ability that the ANU/ninemsn Internet poll has a larger bias than the Morgan poll is
0.92 (in estimating 2PP vote shares), while the probability that the Morgan poll has a
larger bias than the telephone polls is greater than 0.99. The average bias for the
Downloaded by [University of West Florida] at 12:30 09 October 2014

phone-based polling organisations (ACNielsen, Galaxy, Newspoll) is small—just

over a percentage point underestimate of the Coalition’s 2PP vote share—and on
the threshold of conventional levels of statistical significance;8 these three phone-
based polls collectively did quite well in estimating first preference vote share,
with the average bias almost exactly zero (0.1 of a percentage point, plus or minus
about 1.8 percentage points).
Although the magnitudes of the biases do map onto mode of interview rather
neatly, there is much more to house effects than interview mode. For instance, it
may be that sampling and weighting procedures rather than mode underlie the
poor performance of the self-selected Internet poll (ANU/ninemsn) and face-to-
face polls (all but the final Morgan poll). This distinction seems especially important
when considering the ANU/ninemsn poll, where mode (self-completion via the
Internet) is confounded with sampling (self-selection). Although the ANU/
ninemsn poll has the largest bias of the five polls I consider here, it would seem
hasty to write off the Internet as a polling medium: Internet polls offer the promise
of precision via large sample sizes (note that the marginal cost of an additional Inter-
net respondent is close to zero) and timeliness, and may well be the wave of the future
(eg McAllister 2004), provided the thorny issue of the non-representativeness of a
self-selected sample of Internet users can be solved. Clearly, the large bias estimate
attaching to the ANU/ninemsn Internet poll suggests that considerable work needs to
be done on this score. The problem confronting the use of the Internet as a tool for
election polling in Australia, at least in the near term, is that (1) compulsory voting
means that the relevant population is the entire adult citizenry; (2) characteristics
which predict Internet usage in Australia appear to be correlated with politically rel-
evant characteristics, in particular age and the urban/rural divide (see, for example,
Australian Bureau of Statistics 2003; Reid 2002). Deriving weighting schemes so
as to make self-selected Internet samples representative of broadly defined popu-
lations is a topic of considerable interest in both academic and commercial circles
(eg Berrens et al 2003; Schonlau et al 2004), and the future does look promising.

8
We would reject the null hypothesis that the average bias of the three phone polls is zero, in favour of
the one-sided alternative at conventional 95% levels of statistical significance; the probability that the
average bias of the phone polls leads to an underestimate of Coalition 2PP is 0.93, just short of the
conventional 95% standard.
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 515

Conclusion
I have presented a statistical model for dealing with two problems with election polls.
By pooling and smoothing the polls, we obtain a more precise estimate of underlying
vote intentions than can be formed from any single poll. By constraining the
estimated trajectory of daily vote intentions to culminate with the known election
result, it is possible to estimate bias parameters for each polling organisation, and,
in turn, to recover day-by-day estimates of vote intentions that are purged of the
biases afflicting any one polling organisation.
The estimates of daily vote shares suggest that there was a 2 percentage point shift
towards the government over the course of the campaign, almost all of it coming in
the opening weeks of the campaign. Substantial biases specific to polling companies
are also apparent: Morgan and the ANU/ninemsn Internet polls systematically
Downloaded by [University of West Florida] at 12:30 09 October 2014

underestimate Coalition support by large magnitudes. On the other hand, the

average bias of the telephone-based polls is small, and on the threshold of statistical
significance at conventional levels.
The analysis reported here exploits the luxury of hindsight. The fact that the elec-
tion outcome is known plays a powerful role in my analysis; the known election
outcome provides a fixed point from which the daily stream of 2PP estimates and
house effects are anchored. This is not to say that the model I present here is of no
utility to researchers tracking shifts in vote support over the course of a campaign.
Indeed, the model presented here can be easily deployed in ‘real time’, but the
researcher needs to first make a commitment on the question of house effects. One
strategy might be to ignore the house effects, naively pooling the polls with the
bias parameters all (implicitly) set to zero; the results presented here strongly
suggest that the house-specific biases are not ignorable, certainly when we admit
Morgan and the ANU/ninemsn Internet poll to the pool. Ignoring house effects
might work if one focused solely on phone polls and the house effects estimated
here are in fact more or less fixed features of the various polling organisations. A
better strategy, to be implemented in the next federal election campaign, is to use
the house-specific bias estimates reported here to calibrate and then pool the
estimates produced by each polling organisation.

References
Australian Bureau of Statistics. 2003. ‘Household Use of Computers and the Internet.’ Australia Now
series. , http://www.abs.gov.au/ausstats/[email protected]/46d1bc47ac9d0c7bca256c470025ff8%7/
c1d866341d03d9e9ca256d39001bc362!OpenDocument..
Baum, M.A. and S. Kernell. 2001. ‘Economic Class and Popular Support for Franklin Roosevelt in
War and Peace.’ Public Opinion Quarterly 65: 198– 229.
Beck, N. 1990. ‘Estimating Dynamic Models Using Kalman Filtering.’ In Political Analysis, ed. James
A. Stimson, vol. 1. Ann Arbor: University of Michigan Press: 121– 56.
Berrens, R.P., A.K. Bohara, H. Jenkins-Smith, C. Silva and D.L. Weimer. 2003. ‘The Advent of Internet
Surveys for Political Research: A Comparison of Telephone and Internet Samples.’ Political Analysis
11: 1 – 22.
Brent, P. 2004. ‘For Whom the Polls Tell.’ Walkley Magazine 30: 17.
Carlin, B.P., N.G. Polson and D.S. Stoffer. 1992. ‘A Monte Carlo Approach to Nonnormal and
Nonlinear State-Space Modelling.’ Journal of the American Statistical Association 87: 493– 500.
Carter, C.K. and R. Kohn. 1994. ‘On Gibbs Sampling for State Space Models.’ Biometrika 81: 541– 53.
Erikson, R., C. Panagopoulos and C. Wlezien. 2004. ‘Likely (and Unlikely) Voters and the Assessment
of Campaign Dynamics.’ Public Opinion Quarterly 68: 588.
Fleiss, J.L., B. Levin and M.C. Paik. 2003. Statistical Methods for Rates and Proportions, 3rd ed.
Chichester: Wiley.
 516 S. JACKMAN

Goot, M. 2000. ‘The Performance of the Polls.’ In Howard’s Agenda, eds Marian Simms and
John Warhurst. St Lucia: University of Queensland Press: 37 – 47.
Goot, M. 2002. ‘Turning Points: For Whom the Polls Told.’ In 2001: The Centenary Election, eds John
Warhurst and Marian Simms. St Lucia: University of Queensland Press: 63 – 92.
Goot, M. 2005. ‘The Polls: Liberal, Labor, or Too Close to Call.’ In Mortgage Nation, eds Marian
Simms and John Warhurst. Perth: API Network/Edith Cowan University Press.
Gow, D. 2001. ‘Sampling in the Australian Election Studies: Past Practices and Future Prospects.’
Presented to the Australian Election Study Workshop, Australian National University, Canberra.
Green, D.P., A. Gerber and S.L. De Boef. 1999. ‘Tracking Opinion Over Time: A Method for Reducing
Sampling Error.’ Public Opinion Quarterly 63: 178– 92.
Groves, R.M. 1989. Survey Errors and Survey Costs. New York: Wiley.
Harvey, A.C. 1989. Forecasting, Structural Time Series Models and the Kalman Filter. New York:
Cambridge University Press.
Jackman, S. 2000a. ‘Estimation and Inference are Missing Data Problems: Unifying Social Science
Statistics via Bayesian Simulation.’ Political Analysis 8: 307– 32.
Jackman, S. 2000b. ‘Estimation and Inference via Bayesian Simulation: An Introduction to Markov
Chain Monte Carlo.’ American Journal of Political Science 44: 375– 404.
Downloaded by [University of West Florida] at 12:30 09 October 2014

McAllister, I. 2004. ‘A Margin for Error.’ The Bulletin 19 October.

McDermott, M. and K.A. Frankovic. 2003. ‘Horserace Polling and Survey Method Effects: An Analysis
of the 2000 Campaign.’ Public Opinion Quarterly 67: 244–64.
Reid, A. 2002. Town and Country and the Digital Divide. Sydney: ACNielsen/NetRatings.
Schonlau, M., K. Zapert, L.P. Simon, K.H. Sanstad, S.M. Marcus, J. Adams, M. Spranca, H. Kan, R.
Turner and S.H. Berry. 2004. ‘A Comparison between Responses from a Propensity-weighted
Web Survey and an Identical RDD Survey.’ Social Science Computer Review 22: 128– 38.
Stimson, J.A. 1991. Public Opinion in America: Moods, Cycles, and Swings. Boulder: Westview.
West, M. and J. Harrison. 1997. Bayesian Forecasting and Dynamic Models. New York: Springer.
Wolfers, J. and A. Leigh. 2002. ‘Three Tools for Forecasting Federal Elections: Lessons from 2001.’
Australian Journal of Political Science 37: 223– 40.

Appendix: Bayesian Estimation and Inference via Markov Chain Monte Carlo
Equations (3), (4) and (5) define the statistical model used for pooling and smoothing
the polls, as well as estimating house effects. The unknown parameters are (1) at , the
Coalition’s vote share on day t, t ¼ 1, . . ., T 2 1; (2) dji , the house effect of polling
organisation ji , where j indexes the set of five polling organisations analysed here;
and (3) v2 , a variance parameter tapping the magnitude of day-to-day variability
in the at .
Inference for these parameters is via Bayesian methods, meaning that we require a
characterisation of the posterior density of the model parameters, ‘posterior’ literally
in the sense of after having observed the polling data. I rely on computationally inten-
sive, simulation-based methods to provide this characterisation; specifically, I use
Markov chain Monte Carlo (MCMC) techniques, and in particular Gibbs sampling,
to provide a large number of approximately independent samples from the posterior
density of the model parameters. In this case the Gibbs sampler was run for 26
million iterations, discarding the first one million iterations as burn-in, and each
1000th iteration of the remaining 25 million iterations retained for inference, yielding
25,000 approximately independent samples from the posterior density of the model
parameters. Point estimates for parameters reported in the body of the paper are the
average of the sampled values for the respective parameters; 95% confidence inter-
vals are formed by computing the 2.5 and 97.5 percentiles of the sampled values.
Reviews of MCMC methods tailored for a political science audience can be found
in Jackman (2000a, b). I adopt the Bayesian approach to dynamic, latent-variable
modelling as discussed in West and Harrison (1997), Carter and Kohn (1994) and
Carlin, Polson and Stoffer (1992). Complete details on the implementation of the
 POOLING THE POLLS OVER AN ELECTION CAMPAIGN 517

Gibbs sampler for this problem, and the prior distributions assumed, are available in
the longer, working paper version of this article on the author’s Website.
Prior distributions are a critical component of Bayesian modelling, a formal state-
ment of the researcher’s a priori beliefs about the model parameters. Equations (5)
and (6) supply priors for the at parameters. For the house-effects parameters dj
I use a vague normal prior centred at zero dj  N(0, d 2 ), where j indexes survey
organisations, and with d an arbitrary large constant; I use d 2 ¼ (0:15=2)2 ¼
0:005625.
A prior distribution is also required for the variance of the day-to-day changes in
voter sentiment, v. I use a uniform prior:

100 if 0 , v , 0:01
f (v) ¼ (A1)
0 otherwise
Downloaded by [University of West Florida] at 12:30 09 October 2014

That is, I presume that day-to-day changes are not massive; even at this maximum
prior value of v ¼ 0.01 (1 percentage point), 95% of the daily changes in the at
are no larger than plus or minus 2 percentage points.
Downloaded by [University of West Florida] at 12:30 09 October 2014