Contextual Determinants of Electoral System Choice - Ã bo Akademi

etween independent and dependent variables. Particularly large samples canproduce significant values for otherwise inconsiderable effects.The Baysian information criterion (BIC), constructed by A. E. Raftery (1995), istherein a more satisfying test of significance. The BIC value is obtained bysubtracting the logarithm of the population from the Wald value. First of all, theBIC value should exceed zero to reach some level of significance. For coefficientsabove zero, Raftery specifies rules of thumb to evaluate different grades ofevidence in the following way: a BIC difference of 0-2 is weak, 2-6 is positive, 6-10 is strong, and larger than 10 is considered very strong. However, when variablesin a model are compared with each other, the BIC value does not explain morethan other coefficients, because it is, after all, dependent on the square of thelogged odds relative to the standard error. In the following analyses, consequently,I shall present B-coefficients, standard errors, and values of significance. Allindependent variables are coded on a measure scale ranging from 0 to 1. At a firstglance, the demonstration of the odds would be a more appropriate solution thanthe logged odds, since the odds represent the factor by which the probability ofhaving an event is multiplied as a consequence of a one-unit change in theindependent variable. However, a preliminary analysis reveals that the standarderrors vary to a great extent, which in turn implies that the odds have little concretemeaning.The model chi-square value, the –2 log likelihood value, and the Nagelkerke Rsquare value for the model as a whole are also presented. The model chi-squarevalue tests the null hypothesis that all coefficients other than the constant equal 0.The larger the chi-square value, the greater the model improvement above thebaseline. A test of significance on the basis of this value is provided. The –2 loglikelihood value reflects the likelihood that the data would be observed given theparameter estimates. It can be thought of as the deviation from a perfect model inwhich the log likelihood equals 0. The closer the –2 log likelihood value to zero,the better the parameters do in producing the observed data. The Nagelkerke Rsquare is a measure referred to as the pseudo-variance explained, and has aminimum of 0 and maximum of 1. The value cannot decrease when anothervariable is added to the model.207