Yesterday’s post promised boring details about the prediction models posted there. I will try to temper the boring details with brevity.
This guess about outcomes is based on polls, but it is not enough simply to take a poll and apply the mathematical formula for calculating the number of seats (an observation that Slovak papers have been frustratingly slow to incorporate). To make a meaningful prediction based on polls it is necessary to answer several questions:
- What are the actual polling numbers?
In practice this requires predictors to use more than one poll. I have tried to solve this problem here by averaging all polls for each calendar month. This throws away some information about trend, since some polls are taken at different times than others, but I tend to think that it is preferable to miss trends than it is to assume that random, sample-based variations in survey results actually reflect trends. Fortunately, the polling numbers of various firms tend to correspond quite well, with the exception of a few survey firms. The graph and table below give the average of results from multiple firms.
- How do polling numbers translate into votes?
This question is much harder. We can use a variety of data sources that determine the degree to which those who specify a particular party (or a disinclination to vote) will actually go out and vote for that party. Unfortunately, despite my best efforts, I have only been able to obtain some of the information: the degree to which party voters are attracted to other parties as an alternative. An even more relevant question is the degree to which party supporters are willing actually to go out and vote. I’ve been able to glean some of this from news sources (many firms have promised this information to clients on a proprietary basis) and from analogies to extensive 2002 data sources, but these are woefully incomplete.
Without the information necessary to get into voters’ heads, the best alternative–and the mainstay of my analysis here–but simply to assume that voters will both speak and act in the way they have in the past, in other words that polling numbers will translate into votes at the same rates as in previous elections. The limitations of this approach are so obvious that they need little comment, but with little else available this is where my efforts at prediction begin. This way of answering the question requires answering two further questions
- To what degree will party supporters turn actually vote?
My efforts here begin from the baseline that if everyone voted, the results would look like polls (this is a flawed beginning, but not entirely implausible, especially if you believe, as I do, that campaigns do not have a particularly large effect). From there, I look at levels of actual turnout and actual vote for particular parties and use the results to calculate to what degree a drop in overall turnout will correspond to a drop in the turnout of likely voters for a particular party. This, of course, depends on which election is used for a baseline. I have created models based on the 2002 parliamentary election and the 2004 Europarliament election. The former suggests extremely high turnout effects for particular parties (major losses for Smer, gains for SDKU); the latter shows relatively low turnout effects, but the patterns tend to be quite consistent with 2002. In an arbitrary, but I hope not unjustified, choice, I have based my own predictions on an average of these two models. Inserting the most recent polling average into these equations produces the following results, displayed here both as graph and table.
These calculations differ from my earliest model not only in the use of the most recent polling data but in the modification of assumptions for SF. Since the party did not campaign in 2002, I used a zero-slope for that part of the equation but in retrospect it is clear that this is not sufficient. For lack of better data, I have assumed that the behavior of SF voters would resemble that of other new party voters and so have based the assumptions on an average of voter behavior for the new parties of 2002 (ANO, HZD, SDA and Smer). This is highly questionable, I know, but better than leaving the answer blank.
- To what degree will Slovakia’s electorate actually vote?
One advantage of this model is that it allows for some degree of sensitivity to turnout. The tendency of parties to outperform or underperform their polling numbers obviously depends on a number of factors and is not the same from election to election. A turnout-based model allows us to incorporate, for example, Smer’s significant under-performance in the 2004 Europarliament election (28.3% in polls but only 16.9% of votes cast) without assuming that the party would experience such significant losses in a larger turnout election.
Employing these assumptions, however, requires a meaningful guess about the actual turnout. Here I am limited by available data. My only basis for predicting turnout is to look at figures from polls for those who say they will not vote or do not know who they will vote for (often a good proxy for staying home) and to compare these to the previous election. Current data shows the average number of "won’t vote and don’t know" in multiple polls for May to be just over 30%. In 2002 at a corresponding time period, the average was just under 27%. All else being equal–which it never is–this suggests a turnout about 3% lower than last time, or about 66%. This is the baseline that I have so far used in my predictions.
Another, baseline, less secure but nevertheless interesting, can be derived from more nuanced questions about participation itself. Here, however, we have only partial data and approximate comparisons. The most recent results reported by MVK suggest that 38.2% of respondents would certainly vote. In a corresponding poll from 2002 by FOCUS, 46.6% of respondents announced their firm intent to vote, suggesting a difference of 8.4% or a turnout of about 61%. Unfortunately, however, the question asked by FOCUS offers different options, including "probably will vote" and "probably won’t vote" and so the results are not strictly comparable. Still, this comparison suggets that MVK’s estimates of 46% to 54% are probably too low.
A final brief note: this sort of comparative baseline, while always flawed, is absolutely necessary to make any reasonable guess. Stories like Pravda’s "60% would come to the ballot box" make the same mistake as the other analysis of polling numbers discussed above, taking respondents at their word without seeing how their word translated in the past into votes. For this consistent oversight, much of the Slovak press deserves low marks.
What does it all portend?
As the above analysis indicates, there are many sets of assumptions, and error in any of these will produce faulty results. My hope, however, is that best guesses in each based on contextually-grounded comparisons will yield a fairly accurate outcome from the model without too much ad hoc tweaking of guesses for individual parties (though of course as my SF modification shows, ad hoc circumstances may cause me to change the model).
In developing this model, I have been pleased to see that the results correspond to certain presumptions that I share with others who follow these developments:
- that SMK, KDH, and HZDS will gain from low turnout because of their stable electorates but that they will not gain enormous amounts
- that Smer will not do as well as its numbers predict (Interestingly, Pavol Haulik of MVK has been emphatic on this point–which is understandable since he will be judged by the performance of his polls–and on numerous occasions has suggested directly or indirectly that Smer will fall short of its current numbers, including an interesting recent suggestion that late polls showing a Smer victory may lower the party’s margin since "Slovaks do not like it when somebody wins too much")
- that KSS has some hidden strengths that may allow it to cross the 5% threshold regardless of its current weakness, particularly in a low-turnout election.
- that SNS is reaping some benefit from its position as a non-parliamentary party and that these, as with Smer, may be reflected in voting that is lower than the polled numbers.
- that SDKU somehow manages to pull rabbits out of hats (even when the data offers little reason for expecting it).
Obviously this last point–the expectation of magic–is the weakest of
all here, and therefore the most suspect. At a certain point, however
it may not be that relevant since even the best prediction for SDKU will not put it in a position to form a government, and only a limit case analysis of numbers would even allow it to form a two-party coalition with Smer.
But more on that in another post.