Income (Inc) is defined in our model as the average income per capita in the municipality. The sign of income is a priori ambiguous. On the one hand, we can expect a negative sign as people with high income are associated with higher opportunity costs of their time (Frey, 1971, 102). On the other hand, high income people tend to be better educated (thus having lower information costs) and are more sensitive to social pressure and the associated costs of not-voting (Fraser, 1972, 117). Though the major effects of income can be seen to be cost effects, the latter element in the reasoning indicates there might also be an effect of increases feelings of ‘duty’ (D) associated with higher income.
Age is defined as the percentage of the population over 65. We assume a negative relation with turnout based on the so-called “life cycle model”. This model asserts that there is a non-linear relation between age and turnout. First, turnout increases with age as people become more experienced in voting and acquiring information and as they become more socially integrated (Strate et al, 1989, 447-453). However, when one reaches a certain age, one’s social involvement tends to decrease and one is confronted with the disabilities and health problems often accompanying old age (Hout and Knoke, 1975, 53) increasing the costs of voting. Although the average age (and its square) may well be better to gauge this relationship, these figures were unavailable to us.
Electronic voting is a dummy variable with a value equal to one when electronic voting is possible and zero otherwise. The possibility of voting electronically is likely to reduce the costs of voting and thus increase turnout. However, it may also be the case that people unfamiliar with modern technologies are not at ease with or may distrust computers such that they will prefer not to vote. This could then lower turnout. The effect of this variable thus is a priori ambiguous.
The ‘civic duty’ element of the rational voter model can in aggregate-level models be approximated by measures of social cohesion. Indeed, if a community is more socially integrated, an individual in that community is more likely to feel socially and morally obliged to vote. The reason is that he does not want be known as someone that does not care about the ‘common good’ (Overbye, 1995, 376). In this case, voting can be seen as an act of strategic reputation building. Three measures in our analysis look at the effect of social cohesion.
Mobility (Mob) is defined as the number of immigrants and emigrants at the municipal level divided by the total population of the municipality (lagged by one year). Increased mobility of the population is supposed to depress turnout. Firstly, municipalities with a high turnover in the population tend to show weaker group connections (Hoffman-Martinot, 1994,14). This leads to lower social pressures to turn out reducing the cost of not voting (Schram, 1991, Ch. 8).
Number of pre-1976 municipalities is used as a second (admittedly imperfect) proxy for social cohesion within the municipality (Sub). In 1976, the Belgian municipal landscape was drastically reformed in that on average every four municipalities were merged into a single “new” one. It is assumed that a higher number of pre-1976 communities reduces the level of social cohesion in this “new” municipality. This should then lead to lower levels of social pressure, decreasing turnout.
Finally, Geographical location is a set of dummy variables representing the five Flemish provinces. This may be seen as a test of whether civic duty is stronger or weaker in certain areas of the country, as we hold constant all other elements we believe to influence turnout. Of course, this is certainly not a perfect measure of the possible geographical influence on voting behaviour. There is little reason to assume that the areas where civic duty is significantly larger (or smaller) coincide perfectly with provincial boundaries. However, the use of a more suited variable to measure spatial correlation would greatly complicate our model and time restraints prevented the use of them.
Analysis Before we start with the actual analysis of the data through panel data estimation, we first wish to infer whether there may be problems of multicollinearity in our dataset. The presence of this could be problematic for the preciseness of our estimates and would thus possibly seriously inhibit analysis. As exact multicollinearity is very unlikely to occur, we will only have to check for the degree of multicollinearity between our explanatory variables. We do so via two tests.
The first is simply to look at the bivariate coefficients between the various explanatory variables. Following Davis (1991, 83), we employed a cut-off point of 0.80 to eliminate variables from the model. The results are presented in Table 1. It is clear that none of the variables in the model reaches the cut-off point. In fact, the highest correlation in our dataset is 0.5253 between income and electronic voting. As also mentioned in Tolbert et al (2001, 631), correlations of 0.5 are usually seen as evidence for no more than “moderate correlation” between two variables. Hence, none of the variables proposed earlier needs to be eliminated on the ground of multicollinearity.