Disentangling the direct and indirect effects of the initiative process

For a description and overview of direct democracy practices across the globe, see International Institute for Democracy and Electoral Assistance (2008) and Kaufmann et al. (2010).

For surveys, see Lupia and Matsusaka (2004), Matsusaka (2005), and Garrett (2010). For recent evidence on same-sex marriage, see Lupia et al. (2010).

For example, when public opinion is 80 % or more on one side, 82 % of initiative states choose the majority policy while only 46 % of noninitiative states choose the majority policy.

The literature that uses a dummy variable or a simple comparison between initiative and noninitiative states is extensive, for example, Boehmke (2005) on death penalty and Indian gaming; Feld and Kirchgassner (2001) on debt; Feld and Savioz (1997) on economic performance; Gerber (1999) on abortion and death penalty; Matsusaka (1995, 2004) on fiscal policy and fiscal institutions; Matsusaka (2008) on the executive branch; Matsusaka (2009) on public employment; and Persily and Anderson (2005), Matsusaka (2006), and Smith (2008) on election law.

Boehmke and Witmer (2004) find a significant direct effect but an insignificant indirect effect of the initiative on state adoption of Indian gaming compacts. This finding may be closer to an example than evidence of a general pattern because it is based on a specific and somewhat narrow issue and because the direct effect is identified based on only two cases in their sample of 576 observations.

All such models build on the agenda-setter framework of Romer and Rosenthal (1978, 1979).

For a discussion and evidence that legislators do change their positions in light of information revealed in elections, see Kousser et al. (2007).

The article reports results from linear probability models (1), but every regression was also estimated in a logistic specification, which has theoretical advantages given the dichotomous dependent variable. I have chosen to report coefficients from the linear specification because they have a direct interpretation as marginal probabilities, and the linear specifications do not produce findings that differ in material ways from the logistic specification.

A simple dummy variable formulation implicitly assumes that the initiative process is “equally effective” in every state that allows it. While this is a reasonable starting point, there is some evidence that initiative effects depend on how easy it is to use the process, as determined, for example, by signature requirements or petition periods. I discuss this issue below.

More precisely, b captures differences between initiative and noninitiative states. To the extent that initiative and noninitiative states are otherwise identical conditional on X, one can interpret b as a causal parameter. Matsusaka (2004) contains a longer discussion, and presents evidence suggesting that initiative and noninitiative states are not likely to differ in terms of unobserved ideology or culture.

In principle, the communication effect should depend on vote totals. For example, a one-sided rejection of a measure might convey more information than a 49–51 rejection; similarly, the results from a high turnout election may communicate more than results from a low turnout election. Some estimates in this vein are discussed below.

One particular limitation is worth noting here: for the most part the ANES is not designed to be representative at the state level (the exception is the 1988–1992 pooled Senate study). To the extent that state-level public opinion is measured incorrectly, the regression coefficients will be estimated with error, and biased toward zero. Thus, the limitations of the ANES data bias against finding significant results, and strengthen our confidence in findings when they occur.

There is a large literature concerned with estimating policy congruence and policy responsiveness. The present article does not attempt to break new ground on the core research design questions in that literature but rather uses “off the shelf” variables that were developed previously. The measure in (4) is essentially the concept stated in Gerber (1999) as implemented in Matsusaka (2010).

As is usual in the direct democracy literature, I classify Illinois as a noninitiative state. Illinois’ initiative process can be used only to amend structural and procedural subjects contained in Article IV of the state constitution, which has to do with the legislature. The state does not permit initiatives concerning any of the issues considered in this study, e.g. in 1994 the Illinois Supreme Court prevented an initiative concerning term limits from appearing on the ballot.

Initiatives requiring parental notification were included in the “parental consent” category. The results are similar if they are excluded. The “term limits” category does not include initiatives that allowed a representative to take a non-binding term limits pledge; only initiatives that limited terms by law. For term limits, I counted laws placing term limits on state legislators even though the ANES question asks about term limits on congressmen; see discussion in the appendix.

In states with an indirect initiative process, after citizens collect enough signatures to place a measure on the ballot, the legislature has the option to adopt the proposal without sending it to the voters. I was able to identify three initiatives that became law in this way, without going to the voters, and include those in the category of initiatives that were approved: Alaska approved capital punishment in 1986, Michigan prohibited public funding of abortion in 1987, and Michigan required parental consent for abortions by minors in 1990.

Additional information on data sources is contained in the appendix to Matsusaka (2010).

Note that the direct effects are not identified entirely from cross-state differences; the identification also exploits variation across issues.

I estimated numerous exploratory regressions using other control variables, none of which changed the findings in a material way or were reliably statistically significant. Among the demographic variables considered were different measures of education, population, urbanization, race, and ethnicity. I also explored alternative variable specifications, such as log of income instead of its level, and allowed the initiative effects to interact with the opinion variables. I experimented with other measures of political culture, but none had explanatory power. In a working paper version of the study, I included a variable for the age of the state, which does have explanatory power in some specifications; I excluded this variable in the present version because its interpretation is unclear and its inclusion does not alter the main findings related to initiative status. I also ran exploratory regressions with control variables representing legislative structure and professionalism, which did not alter the main results.

If there was more than one failed initiative in a state on an issue, the average approval rate was used.

Another robustness concern is whether any one particular issue out of the nine is driving the results. Based on regressions that delete issues one by one, it can be determined that the term limits issue contributes more to the results than any other single issue. If term limit observations are deleted from the sample, the regression continues to show an overall statistically significant indirect effect and a positive direct effect, but the standard error increases so that the coefficient on the direct effect is not different from zero at conventional levels of significance. The lack of significance could be due to having many fewer observations that use the direct channel once the term limits observations are removed. A similar pattern holds for the policy regressions in Table 4, although the direct effect remains significant in some specifications. A plausible conclusion is that the sample displays robust evidence of an overall initiative effect, but the evidence for the direct channel depends to a large extent on the term limits observations.

For a detailed comparison of citizen ideology in initiative and noninitiative states, see Matsusaka (2004).

I do not consider approved liberal initiatives because the sample contains only two of them.

The conservative tilt emerges from studies that cover the last several decades. Research from the early twentieth century finds a liberal tilt associated with the initiative process (Matsusaka 2000).