How Initiatives Don’t Always Make Citizens: Ballot Initiatives in the American States, 1978–2004

The initiative states are Alaska, Arizona, Arkansas, California, Colorado, Florida, Idaho, Illinois, Maine, Massachusetts, Michigan, Mississippi, Missouri, Montana, Nebraska, Nevada, North Dakota, Ohio, Oklahoma, Oregon, South Dakota, Utah, Washington, and Wyoming. See Waters (2003) for complete rules by state. Initiatives are distinct from referendums in that citizens may place the former on the ballot themselves.

See also Gilens et al. (2001).

Closing dates coded from The Book of the States (1978–2004).

Because it treats the differences among points as constant while the distances among items on our scales are unknown, ordinary least squares regression is inappropriate for scales such as these (Greene 2000, 875).

Although respondents over-report whether they voted, it seems highly unlikely that the presence or absence of voter initiatives would affect whether they lie to pollsters. The best predictor of falsely reporting having voted is education, and controls should negate any such effects (Silver et al. 1986).

Where possible, the scale includes the Vice President, Speaker of the House, and Chief Justice. The 2000 ANES asked about substantially more obscure figures, however, so the names are Trent Lott, William Rehnquist, and Janet Reno.

Specifically, we use the “sandwich” estimator, commonly referred to as the Huber–White estimator (Huber 1967; White 1980), adjusted for clustering (Rogers 1993). When using CPS data, this clustering procedure substantially raises the standard errors on state-level variables as it effectively reduces degrees of freedom from 80,000 (or so) respondents to 48 states. Since the ANES has a smaller overall sample size, and fewer observations from each state, the standard errors change little. Regardless, the substantive pattern of results remains the same in models with and without robust standard errors.

The number of simulations for both the vector of parameters and each quantity of interest is set at 1,000. All analyses were conducted in the R statistical environment, version 2.6.2, using code modified from the Zelig package. Results on replication may differ slightly due to estimation uncertainty in the simulation procedure.

More specifically, for each respondent in each model, 1,000 simulations estimate first differences for changes in the initiative variable, while all other variables are set to their observed values. Next, we average these draws. From these individual respondent estimates, we compute a mean and variance for the overall treatment effect.

While scholars have added various statewide controls to models of initiatives’ effect on citizenship, we find their effects to be strikingly modest. In order to keep the focus on initiatives, we eschew statewide controls in the figures shown here. Nevertheless, these results remain robust to adding a series of state-level variables. We consistently find zero effect in presidential elections, although the effects at midterm look substantially weaker with controls for state political culture and for the percentage of residents 25 and over with a college education than with controls for the state’s Black and Hispanic population, or no statewide controls. Without controls, and with controls for Black and Hispanic populations, the mean effect of initiatives from simulating the full change in initiatives, averaged across the seven midterm elections, is 0.08. For political culture, however, the figure is 0.05 and for share of residents with a high-school education 0.07. Similarly, when statewide controls are added to the models below for knowledge and efficacy, results change only minimally. Political culture coded from Sharkansky (1969). State-level education and race/ethnicity coded from U.S. Bureau of the Census (1982, 2006a, b).

The results are nearly identical without controls for party ID; in all but three elections, the average treatment effects differ by less than 0.01.

Appropriately designed experiments could gain traction on this notion.

To follow the distinction introduced in Fig. 3, the null effects hold for voters and non-voters alike; the average treatment effect is positive in 26 of the 52 questions for voters, and in 25 questions for non-voters.