What would an Irish General Election forecasting model look like?

13-nate-silver_nocrop_w529_h615_2xThe past few year’s has been an exciting time in field of election statistics. The emergence of Nate Silver and his FiveThirtyEight blog (now a major news site) has changed the dynamic of US elections forever. For those unfamiliar with Nate, he was named one of The World’s 100 Most Influential People by Time magazine in 2009 for his accurate prediction of the 2008 Presidential Election, calling the outcomes of 49 of the 50 US state contests. In 2012 he predicted all 50 states correctly.

So, can we just take his model and use it in Ireland? Sadly, no. However the principles of his approach provide a strong basis for any export to Ireland.

 

OPINION POLLS

At the core of FiveThirtyEight is a collection of opinion polls that are averaged to indicate what the public are thinking at any given moment. Powerful stuff, but there are a few differences between US polls and Irish polls which are worth bearing in mind from the outset:

  1. Frequency – In the course of a US election campaign there are so many opinion polls that it is difficult to track them all. In the UK we find similar frequency where at the time of writing at least one poll is published every day during the 2015 General Election. During the 2011 Irish General Election 14 polls were published in the two months prior to polling day. Outside of a General Election less than half that number are published across the same period.
  2. Multi-Party System – I’m reminded of The Simpsons episode where Kang and Kodos reveal their true (alien) identities to Springfield and declare “It’s a two-party system. You have to vote for one of us”. Meanwhile, Ireland’s Register of Political Parties runs to 19 names, with Renua Ireland soon to make it 20. The ability for opinion polls to accurately capture support for small political parties is a major concern. If a poll places The Green Party on 2%, it is customary for commentators to point out in jest that this could mean a support level of -1%, when one applies the usual +/-3% margin of error.
  3. Constituency level polling – Since the US operates an Electoral College system for its Presidential Elections, opinion polls on a national basis have little relevance. Instead, polling companies operate on a state-by-state basis, the Irish equivalent of polling Roscommon-East Galway or Cork North Central on their own. Polls at this level are costly and for that reason they are rarely done in Ireland, and there is no evidence to support their accuracy.

So we are confined to national level polls with some weighted average calculated. Some of this work is already underway by other observers of Irish Elections, and more “poll of polls” tend to emerge in the build-up to elections.

 

CONSTITUENCY SUPPORT LEVELS

The next component of a model for Irish elections must answer the question of constituency support levels. If Sinn Féin doubles its support nationally (since the last election), will its vote double in my local constituency? The answer to this is of course, no. Sinn Féin may triple its vote in some constituencies, but may stand still in others. The same applies for parties set to lose support. If Labour’s vote is cut in half nationally, not every constituency will mirror this trend.

Adrian Kavanagh is a Lecturer in the Maynooth University Department of Geography and his website on Irish Elections attempts to gauge constituency support levels. His approach assumes that rises/falls in support levels are felt evenly across all constituencies. This is a very rough model, something Dr. Kavanagh highlights in his posts, however it is the first of its kind in Ireland and it provides an excellent basis for an improved model to take shape.

Such changes should attempt to incorporate:

  1. Local Election Results – Local Elections in Ireland take place in the middle of a Dáil term. 21 months separated the 2009 Local Elections and the 2011 General Election with a similar interval set to separate the 2014 Local Elections and the 2016 General Election. In examining “the locals” a model should examine where certain parties made gains/losses. For example, the Sinn Féin vote in Dublin West tripled between 2011 and 2014 with other areas only seeing slight increases. I question whether the level of support at local elections is predictive in its own right (many other factors are at play) however the change in vote share could prove very helpful to any model.
  2. Constituency Volatility – “My father voted for his father, and you’ll vote for his son – he fixed the road!”. The conservative nature of the Irish electorate should never be underestimate and The Savage Eye has captured this phenomenon well in a 14 second skit. Stereotypes would tell us that rural voters are less likely to change their votes at election time, with the reverse true for urban areas, especially Dublin. While most Irish stereotypes are subjective, we CAN measure volatility in political constituencies and incorporate it into a statistical model. It has just never been attempted before.

 

CANDIDATE STRENGTH

While the above changes can help us gauge party support levels for particular constituencies, this tells us nothing about support for particular candidates. Any model for Irish Elections should seek to measure the strength of candidates within their party tickets, within their constituencies and the state of independent candidates. There are a number of factors which should be considered:

  1. Historical Results – Where candidates have run before in their constituency, or in local elections, their past results can give us an indication of future support.
  2. Incumbency – Outgoing TDs have a distinct advantage above newcomers, or those who contested before but lost. This is true in any electoral system and the extent of this phenomenon in Irish elections should be further examined. Outgoing party candidates will have received state funding for the previous Dáil term and Independent TDs a “leader’s allowance”.
  3. Outgoing Ministers – A seat at the cabinet table is highly regarded in many parts of the country, and at the very least it provides a national platform upon which a candidate can increase their chance of re-election. An examination of previous outgoing ministers may give us some indication of the prospects for outgoing Fine Gael and Labour ministers at the next election.  
  4. County Councillors/MEPs/Senators – Name me a new Dublin TD from the 2011 Election who wasn’t in another elected position before their election? If you struggle to answer this question you are not alone. Only ONE TD falls into this category, Peter Matthews who was elected for Fine Gael in Dublin South before leaving the party in 2013. A regular contributor to “Tonight with Vincent Browne” during the crisis, Matthews proves that if you are not already elected to another position, “celebrity” is nearly the only other factor that can aide your election. This is true for Gerry Adams, Mick Wallace, Peter Fitzpatrick and Stephen Donnelly, all of whom joined the Dáil in 2011 with similar advantages. The only exception is Labour’s Michael McNamara who was elected in Clare having never served before as a councillor, MEP or Senator. He did however contest the 2009 European Elections as an independent. The remaining new TDs who joined the Dáil in 2011 did so from the platform of some other elected position, usually Councillor or Senator. Any model of Irish elections must take this into account, and score candidates on that basis.
  5. Strength Within “The Ticket” – Let’s assume that Fianna Fáil has won 10,000 votes in a constituency and two candidates are fielded by the party. The above factors would score candidates individually based on their elected positions and previous results. The next step is to take these strengths and proportion them to the available votes for that party. If “Candidate A” emerges with a score of 0.6, and “Candidate B” with a score of 1.8 we can conclude that “Candidate B” is three times stronger than their running mate. Applying this proportion to the votes available, “Candidate A” wins 2,500 first preference votes, and “Candidate B” wins 7,500.

 

RE-CAP

Before bringing our factors together into a complete model, let us re-cap and summaries the components at play:

  1. Opinion Polls
    • Average of recent party polls, weighted for most recent polls, and those from the most accurate polling companies.
    • Adjust polls for over/understatement of certain political parties (if applicable)
  2. Constituency Support Levels
    • Include local election results to help determine regional spread of support changes for each party.
    • Use Pearson volatility measure or other method to determine volatility of each constituency. Ensure that predictions do not deviate from this volatility.
  3. Score candidates based on:
    • Past election results
    • Elected position currently or previously held (some measure of “celebrity” for others)
    • Relative strength versus candidates from same party or grouping

Having achieved all of the above (and an estimate for turnout in each constituency), we should be left with some prediction for first preference votes for each candidate in each constituency. The model must hold that party support mirrors its poll support nationwide and that constituency proportions show some reflection to historic results.

 

TRANSFERS

Unlike the US and UK electoral systems, our elections are determined not by first preference votes, but by transfers through the elimination of candidates (and surplus votes from those elected). An estimate of transfers is required to model election counts under our PR-STV system along with the rate of non-transferable votes:

  1. Transfer Patterns – Examine recent election results to determine likely transfer patterns between parties. Some trends are already known including a strong level of transfer between the government parties and a level of “transfer toxicity” for Fianna Fáil, meaning they attract very few transfers. Figures can be arrived at for these patterns and incorporated into the model.
  2. Non-transferable votes – As elections counts proceed, the number of remaining candidates diminishes. As each candidate is eliminated, their transfers have fewer places to go, and many do not transfer to any candidate. The rate at which this figure rises can be computed from historical results and a simple function applied to the model.

 

SIMULATION

With first preference votes and transfer patterns calculated, the final step is to simulate each election count under various circumstances. I will outline some steps in this process here, with further reading recommended over at FiveThirtyEight’s Methodology page:

  • Perform the steps outlined above and produce an initial outcome for the various constituencies.
  • Allow for variations in all the variables used (for example, a poll may estimate Fine Gael’s vote at 30% nationwide. Allow for this figure to *hover* anywhere between 27% and 33%. Apply the same to candidate scores. A candidate may be on course to take 5,000 votes. Allow this number to hover between 4,500 and 5,500).
  • Simulate the model at least 1,000 times (FiveThirtyEight runs their model 100,000 times). At each “replication” (one of the 100,000 runs) the variables above will hover at a different combination of results. Essentially, this allows us to reach into 100,000 parallel universes, each one producing a slightly different result, in and around what national polls and candidate scores suggest to us.
  • Record the results at each replication.
  • Produce probability figures for each candidate and for each party.

538

This final point is the fun part (if you weren’t having fun reading this already). If the model works, it should allow us to track the chances for each election candidate as the race progresses. Some examples:

  • alan-shatter-3994-310x415Taoiseach X has a 90% chance of reelection.
  • Cllr. Y has a 59% chance of winning a seat.
  • Minister Z has a 31% chance of reelection. (Illustrated right)
  • Party A is likely to win 56 seats, and is 95% likely to be within the range of 52-59 seats.
  • Party B is likely to win 10 seats, and is 95% likely to be within the range of 8-11 seats.

By 90% chance, we mean that out of 100,000 replications, Taoiseach X was re-elected under 90,000 of those simulations.

 

CONCLUSION

The possibilities are endless with such a model however I realise that all of the above is rooted in sizeable optimism and is shadowed by immeasurable uncertainty. Even if the model draws accurately on all possible sources (and if subjective measures like “celebrity” can be turned into numbers), what’s to say that it will prove no more accurate than a coin toss in a General Election?

However, I shouldn’t end this piece on a sour note. This approach has potential, and at the very least it could prove a bit of fun putting together. The insights from election results are invaluable to our democracy and I genuinely believe that the construction of a forecasting model could radically transform the way we examine Irish General Elections. Having deconstructed the components of this model, the next step is to build it and that’s where I need some help. If you were interested by this piece and want to get involved, please get in touch.

 

David Higgins

higgins.davidw@gmail.com

@ElectionStatsIE

 

 

One thought on “What would an Irish General Election forecasting model look like?

Leave a Reply