When Ballot Issues Matter: Social Issue Ballot Measures and Their Impact on Turnout

It is not clear whether mobilization (of new voters), retention (of presidential election voters, i.e. Traugott 2004), or some combination of the two is taking place (but see Dyck and Seabrook 2010). This ambiguity in the literature, however, does not hinder our ability to investigate the hypothesized effect of an increase in turnout due to direct democracy.

The argument that the presence of the process is the mechanism increasing turnout has largely been abandoned by scholars due to concerns about its nonrandom assignment across states, which raises serious endogeneity concerns about determining the relationship between the institutional factor and its effect (March and Olsen (1984); see Hanmer (2009) for analysis in the context of registration laws). Smith and Fridkin (2008) illustrate the problem of endogeneity by demonstrating that the adoption of the initiative process was strongly influenced by varying considerations across the states, and Keele (2009), using a synthetic case control, finds little evidence that the introduction of the initiative process boosts a state’s turnout.

Some scholars have objected to this measure (Tolbert et al. 2001), questioning whether front-page newspaper coverage is a measure of salience or of the opinions of editors who decide what is placed on the front page. Lacey (2005) counters this point, arguing that market constraints on editors (i.e. the requirement that newspapers run stories that reflect their market’s interests if they wish to stay in business) add validity to this measure of salience.

This definition omits a number of issues that are sometimes considered social issues, such as other crime issues. They are excluded, however, because the values and attitudes driving opinion on these issues are significantly different; they do not tap into morality politics, and/or attitudes on these issues are not derived from primary identity or religion. In addition, one might question the inclusion of all drug legalization issues (decriminalization and the legalization of medical marijuana). Drug use policy is a key tenant of morality politics (Mooney and Lee 1995), and medical marijuana appears to relate to morality politics and tap into attitudes grounded in religious values as well. The existing literature has yet to differentiate between the two, and doing so is beyond the scope of the present study. As such, all marijuana issues are treated as social issues.

These issues include same-sex marriage, homosexual rights, stem cell research, abortion, and euthanasia.

Hanushek and Jackson (1977) explain that it is a “fact that individual level data are usually richer than aggregate data, permit estimation of more elaborate models, and thus are to be preferred when available” (p. 180). In addition, aggregating across individuals can obscure the relationship one is investigating by suppressing variation in the independent variables across individuals (Hanushek and Jackson 1977).

Initiatives differ from other measures in the fact that they allow citizens to actually draft the legislation, which likely affects both what issues are placed on the ballot and the legislative parameters of these issues. Both types of propositions, however, permit direct participation by citizens in the legislation-making process, and it is unclear why a measure would be more likely to mobilize citizens simply based on who placed it on the ballot. This is especially true if we accept that, as discussed above, it is not the actual initiative process itself that increases turnout, as well as the likelihood that the average citizen does not know the difference between the two types of measures.

As previously stated, individual-level data is employed because that is the level of the relationship in which we are interested.

Prior to 1984, social issues propositions were sporadic (they were on ballots in CA, MI, and ND in 1972; CA and OR in 1978; and AK in 1982). From 1984 to 1990, social issue propositions appeared on the ballots in three states in each election (CO, OR, and WA in 1984; AR, ME, and OR in 1986; AR, CO, and OR in 1988; and AK, NV, and OR in 1990). The small number of states with such issues on the ballot in each election over the period preceding the start point of this study, as well as the fact that these ballot measures were limited to a few specific states, raises concerns about the ability to generalize any results across the country. Although the 1992 election only has four states with a social issue on the ballot, it is included to expand the number of presidential elections in the study, while the 1996 election, which had only two states with social issues on the ballot, is included for continuity (though I recognize the results may not generalize across the country).

All noncitizens and those under the age of 18 surveyed in the CPS and CCES are excluded from the analyses.

Clustering arises because the attributes associated with the states in which individuals reside do not vary across individuals within each state. The failure to account for the multilevel nature of the data leads to misestimated standard errors, as we fail to take into account the dependence among individual responses within the same state, which violates the independent and identically distributed assumption of OLS regression (Primo et al. 2007). In most cases, the correlation between the errors is positive, causing the estimated standard errors to be too low, t-statistics to be too high, and leading to predictors that appear to have a significant effect on the dependent variable when in reality they do not. The clustered nature of the data necessary to investigate the relationship in question has routinely been ignored by scholars (but see Schlozman and Yohai 2008; Tolbert et al. 2009).

It is well known that self-reported turnout rates are consistently higher than actual turnout rates. Self-reported turnout rates from the CPS are as follows: 69.5% in 1992, as opposed to the actual voter eligible population turnout rate of 58.1%; 51.6% in 1994, as opposed to the actual rate of 41.1%; 63.0% in 1996, as opposed to the actual rate of 51.7%; 49.4% in 1998, as opposed to the actual rate of 39.3%; 66.4% in 2000, as opposed to the actual rate of 55.3%; 51.4% in 2002, as opposed to the actual rate of 40.5%; 71.8% in 2004, as opposed to the actual rate of 60.7%; 54.3% in 2006, as opposed to the actual rate of 41.3%. While the overall reported rate is certainly inflated, it is less than the over-reported rates associated with the NES of 14–18% over the same period (Duff et al. 2007 and own analysis). Self-reported turnout rates in the CCES are significantly higher, with 85.9 and 72.5% of respondents reporting they voted in 2006 and 2008, respectively (actual turnout rate in 2008 was 62.3%).

See Appendix 1 for a complete description of the coding of each variable.

The nature of the 2006 CCES data does not allow for the inclusion of residential mobility in the model. See Appendix 1 for a complete description of the coding of each variable in the CCES models.

See Appendix 2 for a complete list of all social issue propositions.

The employment of a count of the number of social issues that appear on the ballot is designed to control for when two measures addressing the same issue appear on the ballot at the same time, as the measures may mobilize the same individuals (i.e. there may be no cumulative effect for multiple measures dealing with the same social issue). While such issue overlap is rare, it does occur twice in 1998: the Colorado ballot had two measures that dealt with abortion (the banning of partial-birth abortion and the requiring of parental notification prior to a minor’s abortion), while the Oregon ballot had two measures that dealt with marijuana (the legalization of medical marijuana and the criminal punishment for possession of small amounts of marijuana). Though these measures address different aspects of the same social issue, it is unclear whether the second proposition should theoretically increase turnout beyond the effect of the first, so the more conservative measure is employed. Results from a model run using a count of the number of social issue measures that appear on the ballot in 1998 are roughly the same.

Using a count of the number of measures on the ballot may seem counterintuitive, given the argument above that it is the content of the measure, not the number on the ballot, which is likely to determine any mobilization effect. Just as all ballot measures cannot be considered equal, all social issue measures are likely not equal as well. A count is employed, however, to allow for the cumulative effect of multiple social issues. To check if simply the presence of a social issue on the ballot stimulates turnout, a dummy for states with a social issue on the ballot was employed as a variable. The dummy only exhibits a positive impact on turnout in 1994.

To compile these lists, I relied on the subject description provided by the National Conference of State Legislatures and the classification of each measure by issue concern carried out by the Initiative and Referendum Institute.

Specifically, questions can be raised as to whether diminished returns for each additional measure are expected, which would require some transformation of the variable (such as employing the log of the variable). Analyses of cross tabulations reveal a relatively constant, linear effect across the range of the social issue variable in years that it exerts a statistically significant effect on turnout. Dummy variables could be employed for each number of social measures, but, because of the relatively few states with more than one or two social issues on a state’s ballot, in many cases doing so essentially means that we would be dummying out one state. While it is likely that at some point an additional measure produces a significantly smaller effect than the previous one, the social issue measure likely does not reach this point due to the small number of social issues that ever appear on the same ballot. The possible nonlinear effect of the total number of measures is not considered here, and the effect is treated as linear because this is how previous literature has treated the variable.

In 2002, environmental issue measures received, on average, a slightly larger percentage of votes than social issue measures (almost half a percentage point more). This appears to be an outlier year for environmental issues (their next highest total, 1994, is roughly one and a half percentage points less), and it may be a product of the relatively few number of environmental issues on state ballots that year (six). In addition, it highlights the ability of other issues, given the right circumstances, to grab voters’ attention.

This presentation format is not without its weaknesses, as it does not allow for the presentation of other estimated coefficients, model specifications, or additional information about control variables. This information can be found in Tables 1–8 of the electronic supplementary material. All variables in every model, when significant, exhibit the expected behavior.

Due to the large number of same-sex marriage bans on the ballot in 2004, and, more generally, in all of the elections under investigation, it may be that the results are driven exclusively by such measures. Models for all years were run with a dummy for states that had a same-sex marriage ban on the ballot. The relationship between turnout and same-sex marriage bans only holds for the 1998 election, while the relationship in 2000 is actually negative.

These two elections provide an interesting contrast, in that 1996 was not particularly salient in comparison to other presidential elections, while 2004 was characterized by an active campaign to increase conservative turnout through the presence of social issues on the ballot. While the results for the 1996 election suggest that social issue propositions can increase turnout in low salient presidential elections (either when the outcome is not in doubt or the election is not particularly interesting), the 2004 results highlight their ability to increase turnout in the most salient of presidential elections, especially when candidates and parties make them a focal part of their campaign.

The number of tax and environmental propositions were also employed as a direct democracy measure. The results (not shown) demonstrate a lack of uniformity. The number of environmental measures is associated with an increase in voting in 1996, 1998, 2002, and 2004, though their small number and restriction to a few states suggest we should exercise caution in generalizing any effect across all states. An additional environmental measure on the ballot increased the likelihood of turning out by 0.98% in 1996, 2.6% in 1998, 3.6% in 2002, and 3.5% in 2004. Tax measures only have a positive impact on turnout in 1994 and 2008. An additional tax measure on the ballot increased the likelihood of turning out by 2% in 1994 and 1.3% in 2008. These findings further suggest that the issue content of the ballot measure is paramount, while revealing that social issue measures have a positive impact on the most elections and are alone in increasing turnout uniformly in midterm elections.

Confidence intervals are generated through a simulation process that draws one thousand sets of coefficients from the multivariate normal distribution, based on a mean vector created for the coefficient and covariance matrices. For each individual, the change in the predicted probability is calculated from each set of simulated coefficients, generating a thousand simulated probabilities from which the confidence intervals are computed (see Hanmer and Kalkan (2009) and Herron (2000) for further explanation of the simulation procedure).

The exception is the increase in the likelihood of voting by 1.2 percentage points for 2006 model run with the CCES data, which, while it may be due to the inclusion of variables unavailable for the CPS models (strength of partisanship and interest in politics), may also be due to the high self-reported turnout in the sample (85.9%), which leaves little room for any increase in the probability of voting to occur.

Because social issues only appeared on the ballot in two states in 1996, we should be cautious in generalizing this effect across the entire population.