by Stephen Fisher
Publishing polls in the last five days before an election is illegal in Israel, so the final pre-election polls were published on 13th March. They suggested that Binyamin Netanyahu’s Likud party would get around 22 seats to 25 for the Zionist Union (the alliance between Labor and Haunuah, led by Issac Herzog and Tzipi Livni).
The exit polls yesterday suggested 27 each.
The actual result was 30 for Likud and 24 for the Zionist Union.
These are big discrepancies given the Knesset has just 120 members. At least the exit poll figures were on average closer to the final result than the five-day-old polls. But even the exit polls were three seats out for each of the two main parties. This would be equivalent of being 16 seats out for a UK House of Commons prediction.
Put another way, both electionsetc.com and electionforecast.co.uk are predicting the Tories to get 285 seats at the moment as their central forecast. If this was under by 16 seats then the Conservatives would get 301, which would produce quite a different political situation for government formation.
Moreover, the discrepancy in the Likud figures with the pre-election polls just five days out from the election is equivalent to 43 seats in the House of Commons. This is more than the difference between the current central forecast of a seriously hung parliament and a Tory majority in one direction and almost a Labour majority in the other. Such events are only just within the 95% prediction intervals for electionsetc.com and electionforecast.co.uk, but they are perhaps more likely than we can tell from past British polling data.
Israel has a closed list proportional representation electoral system with a single nationwide constituency so, apart from the 3.25% threshold, changes in seats are pretty much perfectly sensitive to the changes in the share of the vote. This is not so under the first past the post system for the British House of Commons. But on average over the post-war period in Britain the two main parties have gained (or lost) 7.5 seats for a percentage point change in their vote. This is actually slightly but not much more responsive than pure proportionality, suggesting that polling errors on the scale Israel has just suffered could have still more dramatic consequences in Britain.
Huffpollster did warn that there could be surprises, not least because of strategic coalition voting, and indeed the polls and exit polls look much better if you aggregate Likud with the religious right parties. But this does not really not necessarily explain the discrepancies, and certainly not for the exit poll. It will be interesting and helpful to see what the post-mortems conclude.
The point of this post is not to criticise Israeli pollsters. Polls everywhere have the potential and frequency to make big errors that mean that they could be pointing to a very different political situation from the one that will actually transpire in the results. For example, in Britain the Liberal Democrats did not increase their seat tally by the 60% expected from the final polls in 2010. In the 1992 debacle the final polls had Labour fractionally ahead, but the Tories won with a 7.6 point lead.
So while it is very helpful (and much appreciated) to see the central predictions of various election forecasts summarised at may2015.com and elsewhere, observers and commentators would do well to take seriously the prediction intervals from those forecasts which are attempting to estimate the uncertainty based on the historical variability and accuracy of the polls. These include the forecasts at electionsetc.com, electionforecast.co.uk and from the Polling Observatory team. Even we might be under-estimating the true level of uncertainty so it would be wise not to completely rule out either a Tory or a Labour majority in May.