Methodology

Confidence Intervals

Medians

The median is the value separating the higher half of a probability distribution from the lower half. One could therefore think of it as the “middle” value.

For a continuous probability distribution, the median is the value such that a number is equally likely to fall above or below it. Since the number of voters is usually rather large in the areas considered on these pages, the probability mass functions for the voting intentions behave a lot like continuous probability distributions. One can therefore say that for the medians of the voting intentions, the actual number of voting intentions for a particular party is equally likely to fall above or below it.

For the number of seats, this doesn’t hold true though. The number of seats is usually much too small for the probability mass functions to behave more or less like a continuous probability distribution. As a consequence, each of the values in the 99 percent confidence intervals will have a substantial probability, including the median itself. In these cases, the median is the value such that the probability that the value is equal to it or above it is at least fifty percent, and that the value is equal to it or below it at least fifty percent too. One could say that median is the value that pushes the accumulated probability above the fifty percent mark.

Notice that often, the median will also be the most likely value. However, that doesn’t need to be so, and in many cases, it won’t be either.

Shoudn’t the Medians Add Up to the Total Number of Seats?

In most cases, the medians for the number of seats per party will add up to a number close to the total number of seats. But in many cases, the sum of the medians won’t be exactly equal to the total number of seats. Here’s an example illustrating why this isn’t an error:

Image three parties competing for a single seat in a constituency. All three parties have the same chance of winning seat, so the probability mass function is 33⅓ percent of getting 1 seat for each of them, and 66⅔ percent of getting no seat. This means that for all three parties, the median value for the number of seats is 0, with a 95 percent confidence interval of 0—1. The sum of the median values is 3 × 0 = 0, which is less than 1.

Weighted Monte Carlo Simulations

Projecting National Results on Electoral Districts

Opinion polls usually only give data on a national level. Unless there is only one, national electoral district, these data have to be translated to the local electoral districts one way or the other. The model used to produce the numbers on these pages is a variant of the so-called uniform national swing, which means that it projects the national tendencies onto the local electorale districts without any adjustments.

Proportional National Swing

Assume that a party had a voting share of 20 percent at the last election, but that an opinion poll shows that the voting intentions for the party have risen to 22 percent. One way to project this change onto the local electoral districts, is to take the last election result in every electoral district, and simply add the national change to each of them. This means that if the party scored 25 percent in a particular district, the new score used to calculate the number of seats in that district will be 25 + 2 = 27. If the party had a voting share of only 15 percent, the new score would be 15 + 2 = 17 percent.

This model works fine for parties increasing their result, and parties with a modest to large vote share across all electoral district. However, for small parties losing votes, the model can result into negative numbers in the electoral districts where it performed poorely during the last election. Just imagine a party going from 8 to 5 percent nationally, but with a local score of only 2 percent in one of the electoral districts. The uniform national swing model described above would project the party at 2 - 3 = -1 percent in that district.

As an alternative, the model used on these pages projects the national results onto the local districts in a proportional way. That means that when a party increases from 20 to 22 percent, the model doesn’t add 2 percent to the party’s voting share in every district, but multiplies the local results with a factor of 22 / 20 = 1.1. In the example mentioned above, this would put the party in the district where it score 25 percent last time at 27.5 percent, and in the district where it scored only 15 percent at 16.5 percent. For the party losing votes, the factor would be 5 / 8 = 0.625, thus projecting the party in the district where it score 2 percent last time at 1.25 percent.

Is proportional national swing more correct than uniform national swing? Hard to say. Proportional national swing will make parties swing more in electoral districts where they had a high score at the last election. For parties gaining support, that may be incorrect, if they are gaining more national support because they’re mainly getting more support in those districts where they performed less during the last election. But it’s easy to produce counterexamples where a multiplicative approach will be more correct than the additive.

As a general rule, using proportional or uniform national swing shouldn’t make much of a difference for seat projections on the national level. Overestimations in one electoral districts will often be compensated by underestimations in other electoral districts, such that the overall result will be more or less correct. Remember that in the end, a party gaining support on a national level may very well be losing votes in some parts of the country. The projection of the national results onto the local electoral districts should therefore not be used to predict local results, but only as a way to calculate national seat projections. If you want to know how a party is doing locally, the best way to find out about this is to run a local opinion poll.

Regional Data

From time to time, pollsters will run local opinion polls, or publish regional tendencies in their national polls. Advanced models could use this information in order to project national polling results more accurately onto local electoral districs. In some occassions, e.g. when there are parties with a particular local stronghold, this may give more accurate results on a national level. On the other hand, if these local results give better results on a national, it would probably be wiser not to project the national results onto those local electoral districts, and instead just use the regional data directly.

Historic Data

Some models try and use more of the historic data in every electoral district. But in Western countries, with election cycles of four or five years, this means that models use data that is eight to ten years old if last two elections are used, and twelve to fifteen years old if three elections are used. It is questionable how relevant such historic data can be with new parties emerging, old parties disappearing, in addition to changing demographics. Sometimes, even using data from the last election, which may not be older than three or four years, seems hopeless because of deep changes in national politics.

Incumbency

Some studies have shown that incumbent candidates have a slight advantage over their opponents. However, when trying to calculate national results, incumbency should make a difference, since any significant change would already have been included in the national polling results. Therefore, incumbency has not been modeled in for the calculation of the numbers presented on these pages.

House Factors and Other Corrective Measures

Some modelers try to improve the raw polling results using “house factors”. Basically, these house factors try to compensate for the historic errors pollsters have made, e.g. if they have consistently underestimated a party with a certain percentage. Others also include economic or demographic data, e.g. giving a slight advantage to the parties in government if the economy is doing well, or vice versa.

House factors may work to a certain degree in countries with a bipartisan political landscape, but they quickly break down when there are many parties involved, or when there are large or even disruptive changes. Furthermore, for obvious reasons, house factors won’t work for new parties.

The numbers presented on these pages try to explain in more detail the true meaning of opinion polling results. Therefore, the model doesn’t try to “improve” any opinion polling result with house factors or other parameters.

As a personal note, it is also my belief that pollsters have a large interest in trying to get their opinion polling results as correct as possible, and therefore will have a huge incentive to learn from theire mistakes. House factors seem therefore a bit patronizing.

New Parties

Since the method described above relies on regional data in order to calculate the number of seats in parliament, new parties that haven’t participated in the previous election pose a probleme. However, there are number of strategies that can be used to estimate the regional distribution of the new party’s voters:

Note however that in the case where more information is known about the newly formed party, it is always wise to include it. If it’s e.g. known that the new party will only participate in the election in a part of the country, the regional data for the new party should be adjusted in that way.

Results

The following sections show the results for past elections. In particular, for each of the tables below, the final results in terms of number of votes were feeded to the model as a set of polling results. The sample sizes for these fictive polls were 600, 800, 1000, 1200, 1500, 2000, 3000 and 5000, sizes often used by polling firms. For each of the sample sizes, the 95% confidence intervals for the number of seats for each parties are compared to the actual number of seats the parties received.

The results show, quite naturally, that the seat projections are more accurate for electoral systems that are more proportional. But even for non-proportional electoral systems, the seat projections stay relatively close to the actual seat distribution of the election. Larger parties, for which non-proportional electoral systems often will behave in a fairly linear way, get more accurate results than smaller parties, which will often strugle with e.g. local or national electoral thresholds.

United Kingdom, 8 June 2017

Sample Size CON LAB UKIP LIBDEM SNP GREEN PC
Election Result (8 June 2017) 42.4%
318		 40.0%
262		 1.8%
0		 7.4%
12		 3.0%
35		 1.6%
1		 0.5%
4
600 38–46%
278–348		 36–44%
241–331		 1–3%
0		 6–10%
1–19		 2–5%
0–54		 1–3%
0		 0–1%
0–7
800 39–46%
283–343		 37–43%
247–327		 1–3%
0		 6–9%
2–17		 2–4%
0–52		 1–3%
0		 0–1%
0–5
1000 39–45%
287–340		 37–43%
251–322		 1–3%
0		 6–9%
2–16		 2–4%
0–51		 1–3%
0		 0–1%
0–5
1200 40–45%
291–338		 37–43%
254–319		 1–3%
0		 6–9%
2–15		 2–4%
1–50		 1–2%
0		 0–1%
0–5
1500 40–45%
294–335		 38–43%
257–316		 1–3%
0		 6–9%
3–15		 2–4%
1–48		 1–2%
0		 0–1%
0–5
2000 40–45%
300–332		 38–42%
263–312		 1–2%
0		 6–9%
3–15		 2–4%
1–41		 1–2%
0		 0–1%
0–5
3000 41–44%
304–328		 38–42%
269–307		 1–2%
0		 7–8%
4–14		 2–4%
2–35		 1–2%
0		 0–1%
0–5
5000 41–44%
308–326		 39–41%
275–303		 1–2%
0		 7–8%
5–13		 3–4%
2–29		 1–2%
0		 0–1%
0–4
Previous Election Result (7 May 2015) 36.9%
331		 30.4%
232		 12.6%
1		 7.9%
8		 4.7%
56		 3.8%
1		 0.6%
3

Only polls for which at least the sample size has been published are included in the table above.

Legend:

Norway, 11 September 2017

Sample Size Ap H FrP KrF Sp V SV MDG R
Election Result (11 September 2017) 27.4%
49		 25.0%
45		 15.2%
27		 4.2%
8		 10.3%
19		 4.4%
8		 6.0%
11		 3.2%
1		 2.4%
1
600 24–31%
44–57		 22–29%
38–52		 13–18%
23–33		 3–6%
1–11		 8–13%
14–24		 3–6%
1–11		 4–8%
8–15		 2–5%
1–9		 1–4%
1–2
800 24–31%
45–56		 22–28%
39–51		 13–18%
23–33		 3–6%
1–10		 8–13%
15–23		 3–6%
1–11		 5–8%
8–14		 2–5%
1–8		 2–4%
1–2
1,000 25–30%
45–56		 22–28%
40–51		 13–18%
23–32		 3–6%
2–10		 9–12%
15–22		 3–6%
1–10		 5–8%
8–14		 2–4%
1–8		 2–4%
1–2
1,200 25–30%
45–55		 23–28%
40–51		 13–17%
24–32		 3–5%
2–10		 9–12%
16–22		 3–6%
2–10		 5–8%
9–14		 2–4%
1–8		 2–3%
1–2
1,500 25–30%
46–55		 23–27%
40–50		 13–17%
24–31		 3–5%
2–9		 9–12%
16–22		 3–6%
2–10		 5–7%
9–13		 2–4%
1–8		 2–3%
1–2
2,000 25–29%
46–54		 23–27%
41–50		 14–17%
25–31		 3–5%
2–9		 9–12%
16–22		 4–5%
2–10		 5–7%
9–13		 3–4%
1–7		 2–3%
1–2
3,000 26–29%
47–54		 24–27%
42–49		 14–17%
25–30		 4–5%
2–9		 9–11%
17–21		 4–5%
2–9		 5–7%
9–13		 3–4%
1–4		 2–3%
1–2
5,000 26–29%
48–53		 24–26%
42–48		 14–16%
26–30		 4–5%
2–9		 9–11%
17–20		 4–5%
3–9		 5–7%
10–12		 3–4%
1–2		 2–3%
1–2
Previous Election Result (8–9 September 2013) 30.8%
55		 26.8%
48		 16.3%
29		 5.6%
10		 5.5%
10		 5.2%
9		 4.1%
7		 2.8%
1		 1.1%
0

Legend:

Iceland, 28 October 2017

Sample Size D V P B C A S F T R M
Election Result (28 October 2017) 25.2%
16		 16.9%
11		 9.2%
6		 10.7%
8		 6.7%
4		 1.2%
0		 12.1%
7		 6.9%
4		 0.1%
0		 0.2%
0		 10.9%
7
600 22–29%
14–20		 14–20%
9–14		 7–12%
4–7		 8–13%
5–9		 5–9%
0–5		 1–2%
0		 10–15%
6–9		 5–9%
3–5		 0–1%
0		 0–1%
0		 9–14%
5–8
800 22–28%
15–19		 14–20%
10–13		 7–11%
5–7		 9–13%
5–8		 5–9%
3–5		 1–2%
0		 10–14%
6–9		 5–9%
3–6		 0%
0		 0–1%
0		 9–13%
5–8
1,000 23–28%
15–19		 15–19%
10–13		 8–11%
5–8		 9–13%
6–8		 5–8%
3–5		 1–2%
0		 10–14%
6–9		 5–9%
3–5		 0–1%
0		 0–1%
0		 9–13%
6–9
1,200 23–28%
15–19		 15–19%
10–13		 8–11%
5–7		 9–13%
5–8		 5–8%
3–5		 1–2%
0		 10–14%
7–9		 6–9%
3–5		 0%
0		 0–1%
0		 9–13%
6–8
1,500 23–28%
16–18		 15–19%
10–12		 8–11%
5–6		 9–12%
6–8		 6–8%
3–5		 1–2%
0		 11–14%
7–8		 6–8%
4–5		 0%
0		 0–1%
0		 9–13%
6–8
2,000 23–27%
15–18		 15–19%
10–12		 8–11%
5–7		 9–12%
6–8		 6–8%
3–5		 1–2%
0		 11–14%
7–9		 6–8%
3–5		 0%
0		 0–1%
0		 10–12%
6–8
3,000 24–27%
15–18		 16–18%
10–12		 8–10%
5–6		 10–12%
6–8		 6–8%
3–5		 1–2%
0		 11–13%
7–8		 6–8%
4–5		 0%
0		 0%
0		 10–12%
6–8
5,000 24–26%
16–18		 16–18%
11–12		 8–10%
5–6		 10–12%
6–8		 6–7%
4–5		 1–2%
0		 11–13%
7–8		 6–8%
4–5		 0%
0		 0%
0		 10–12%
6–8
Previous Election Result (29 October 2016) 29.0%
21		 15.9%
10		 14.5%
10		 11.5%
8		 10.5%
7		 7.2%
4		 5.7%
3		 3.5%
0		 1.7%
0		 0.3%
0		 0.0%
0

Legend:

Holes and Spikes in Probability Mass Functions

Calculation of the Error Estimate