Description of the IwsmF Transition Model

Part 1: Global Voter Transitions

Individual Data

In case of split ticket vote (party / candidate) cumulated splitting tables (over all available elections and areas) from official statistics are used, whereupon the 1992 developed so-called "Markov" model is applied. With this model a net transition table without nonvoters can be reconstructed from two splitting tables, definitely being better than any recall, even in case of a party changing the camp. In the one vote case symmetrizised polling tables (combination between present and past voting decision, without nonvoters) are cumulated over as many elections and areas as possible. The original recall is definitely not appropriate because of partly massive bias.

Reconstruction of Individual Transition Behavior with Aggregate Data

Starting point is the factual electoral behavior of all areas in % of the voting-age citizens for the present and past election. A nonvoter plus / minus initially is distributed proportionately to the parties according to their past / present election result for every area. Afterwards reduced marginals without nonvoters are calculated, with calculation of provisional (balanced) interior cells following. According to the split ticket or one vote case either the Markov model with splitting data is used, or the polling data are adjusted to the area marginals (with Iterative Proportional Fitting). Then the net transitions between camps are reduced by one fourth and "substituted" by transitions between parties and nonvoters, so that the original marginals add up. Finally the obtained area tables are balanced, the next step makes use of their marginals reduced by the stable voters.

Recapitulation: The procedure outlined above substantially means, that moderately unsatisfied voters vote and defect to some party of the same camp, whereas heavily unsatisfied voters either don't vote at all (approx. 25%) or vote and defect to some party of the opposite camp (approx. 75%). The probabilities of going to the polls in no way are defined arbitrarily, but result from comprehensive empirical tests with the spectrum of all possible nonvoting probabilities.

The correlation model (Thomsen model without logit transformation and with crude binary choice) is used to estimate the interior cells, based on reduced marginals (by the stable voters) as outlined at the end of the first passage above. For every area an independent estimation is made, using individual weights for all areas. The weights are optimized in such a manner to minimize the difference between the statistical and individual data estimation. The estimates of crude binary choice aren't consistent and therefore have to be adjusted to the should-be marginals. With these "final" interior cells the calculation of individual net transitions is done as outlined above (again without nonvoters, different for split ticket or one vote). As a quality criterion the relative error is defined, i.e. the ratio sum of all deviations (absolute values, for floating voters only) divided by all floating voters.

Part 2: Voter Transitions for Social Structure and Demography

Individual Data

In principle similar to part 1 symmetrizised cumulated polling data are used, subdivided both into all three-dimensional combinations of characteristic values of religion / race with education with employment / activity (social structure) respectively of sex with marital status with age (demography) and into present and past voting decision (incl. nonvoters).

Reconstruction of Individual Transition Behavior with Aggregate Data

Similar to part 1 the correlation model is used to estimate the interior cells. Basis are, on the one hand, the final interior cell area estimations (removed stable voters added again) as marginal 1 and on the other hand subdivided (cumulated micro-)census data (as with individual data) as marginal 2 for all areas, always in % of the voting-age citizens. Further proceeding is formally identical to part 1 up to adjusting the crude binary choice to the marginals 1 and 2. These "final" interior cells now are compared with the adjusted polling data (to marginals 1 and 2 of the particular area). As a quality criterion the sum of all deviations (absolute values) is defined. Voting behavior for social structure respectively demography alone results from summing up the related categories of stable and floating voters.