Mobilization Around New Convenience Voting Methods: A Field Experiment to Encourage Voting by Mail with a Downloadable Ballot and Early Voting

We recognize that outcomes beyond the method of voting and turnout also are important but confine this analysis to those issues. Future research should collect data on the requests for absentee ballots to shed light on the reliability of new methods of ballot delivery.

There are several reasons for the divergent findings. For example, McDonald et al. (2015) show the coding of election laws and model specifications can alter recent results regarding early voting. Similarly, Herron and Smith (2012, 2014) demonstrate the importance of coding the length of the early voting period, and any reduction in that period when assessing the effects of early voting by party and race. See also Menger et al. (2015).

The recent implementation of automatic registration in several states is an important exception, though its application will not be universal for all eligible but unregistered citizens. Hanmer (2009) makes an argument similar to the one advanced here and demonstrates the apparent motives behind the adoption of registration reforms influence their implementation and effect on turnout.

As we note below, we test this via an early voting message. The Maryland legislature first adopted early voting in 2005, but Republican Governor Bob Ehrlich vetoed the law. In 2006, the Democrat-dominated legislature overrode the veto and seemingly cleared the path for early voting, but later that year an Anne Arundel County Circuit court judge ruled that Maryland’s early voting law violated the state’s Constitution. In 2008, 72% of Maryland voters supported a ballot question to amend the Constitution to allow early voting. Of course, we do not expect everyone to switch to early voting, but we do expect those predisposed to it and who learn the policy was the result of government responsiveness to its citizenry are more likely to respond by voting early.

Scheibehenne et al. (2010) were unable to identify a set of necessary conditions or a tipping point that would result in negative consequences from multiple choices. They report that some results from the studies showing signs of choice overload have not been successfully replicated. Most of the 50 experiments they reviewed presented more than 10 choices, the range was 2–300, and only 6 offered the same or fewer choices (as the lower bound) than the three new choices presented in this study. This evidence comports with our argument that the presentation of three new voting choices is insufficient to bring about negative consequences from choice overload. We acknowledge that choice also could be seen as the number of days on which one could vote. When thought of in these terms a large number of days could lead to choice overload. Since our treatments focus on the choice of voting method rather than the timing we did not anticipate timing options would cause overload, nor do our results provide evidence for it.

We recognize that we cannot separate out the practical aspects of having more choices (e.g. removing an excuse for nonvoting) from the psychological benefits.

We recognize that there are subtle differences between cost and convenience. In following the literature (e.g. Gronke et al. 2008), we focus on the overlap between these concepts. The theoretical and empirical distinctions between them, and their influence on how voters evaluate new voting methods, are subjects for future research.

We used two criteria to identify overseas voters: if the voter filed a mailing address different from residential address (and that address was not in the United States), or if the voter’s mailing address was an APO address (for voters in uniformed services). We conducted separate experiments relevant to their circumstances (Hanmer et al. 2015; Herrnson et al. 2015).

Maryland is a typical state in many respects. Its population is average in size and, like many other East Coast states, is somewhat more racially diverse, more educated, and more affluent than the national average. The state’s politics have been historically dominated by one-party (the Democrats), but there is considerable variation in the competitiveness of its legislative and local races, and its recent gubernatorial elections have been competitive and resulted in partisan turnover. Maryland’s legislature scores 0.194 on Squire’s (2007) professionalism index, just above the 0.187 average. As discussed in note 4, Maryland voters were directly involved in the decision making over its voting policies. While a one state study has many benefits we recognize that there are also potential limits with regard to generalizability.

In addition to the methodological gains articulated by Nickerson (2005), sample stratification can provide substantive advantages (see Niven 2004; Parry et al. 2008).

We constructed the strata as follows: (1) if the voter is registered and voted three times out of all three elections since 2004, or twice out of two consecutive elections (turnout rate is 100%); (2) if the voter is registered only in one election and voted (turnout rate is 100%, but for only one vote); (3) if registered since 2004 and voted only two out of three elections (turnout rate is 67%); (4) if registered for more than one election but voted only once (turnout rate is 34–50%); (5) if newly registered after the 2008 election, or registered in 2008 but didn’t vote (turnout rate is 0%, but for only one election); and (6) if registered for more than one elections but never voted (turnout rate is 0%). Multinomial logit models for each stratum predicting assignment to treatment based on age, sex, and party identification suggests the randomization worked well, as these factors are not good predictors of treatment assignment. The χ² statistics from the likelihood ratio tests and p values are as follows: stratum 1, χ² = 26.49 p = 0.8489; stratum 2, χ² = 31.79 p = 0.6237; stratum 3, χ² = 29.96 p = 0.7099; stratum 4, χ² = 27.27 p = 0.8214; stratum 5, χ² = 31.22 p = 0.6511; stratum 6, χ² = 33.27 p = 0.5520.

We believe this is a reasonable assumption based on our knowledge of the role of universities in the community and research showing that university experts are viewed as credible sources (Verplanken 1991).

We recognize that we cannot separate out the effect of including the form in the mailer from the effect of the messages about absentee voting themselves. In our discussion with state officials, we agreed that election officials or others who might send mail in envelopes to spur absentee voting would prefer to include the required application form to make it easier for recipients to request the absentee ballot. This, plus the small cost of doing so led us to include the form. As we discuss below, our results should further alleviate concerns with this issue (see “Results” section).

The 2010 election was a typical election for governor, with no significant events that could have interacted with the timing of the treatments and thus led to differential effects.

Our substantive conclusions hold if we adjust for the contact rate.

The analysis of heterogeneous effects is beyond the scope of the present endeavor. That said, we found few differences across voting history. We also tested for potential conditioning effects among groups that traditionally have exhibited low levels of political participation, including voter turnout: young people, women, and registered voters unaffiliated with a major party (groups for which we had data). The results (available from the authors) show that members of these groups did not systematically respond to particular messages in terms of increased use of a specific voting method or turnout. The only exception is that all of the EABDS messages increased its usage among unaffiliated voters.

OLS is more efficient and is the standard analytical approach in the literature. Since all of our right hand side variables are binary variables this is simply an efficient way to get the weighted treatment effects. OLS is equivalent to doing means tests and avoids having to convert output into predicted probabilities (see, e.g. Hanmer and Kalkan 2013). The results hold when we use a probit model or combine all of the analyses into a single multinomial logit model. The full results of the OLS, probit, and multinomial logit models are reported in Online Appendix.

This is a common approach in the literature (see, e.g. Gerber and Green 2000; Gerber et al. 2015; Bennion and Nickerson 2011). This had little effect on the results (see Online Appendix Table A2).

One might argue for the use of a placebo group to compare the impact of any message to no message. We are not aware of designs like this for direct mail. That said, our facts-only messages were sufficiently bland to function as a basis for comparisons. Moreover, the lack of evidence of turnout effects for the facts-only message and other messages promoting a single reform suggests that simply receiving a mailer was not enough to alter turnout.

The effect sizes increase as the denominator (only those who voted) is smaller. These results are available upon request.

See Note 4.

“Judge Voids Md. Laws for Early Voting; Decision, a Blow to Democrats, Is Stayed Pending an Appeal.” Matthew Mosk, August 12, 2006, The Washington Post. “Early Voting Gets Underway.” Liam Farrell, September 4, 2010, The Capital.

Alford et al. (2005) raise an important criticism of contagion effects—selection bias. We recognize that since we cannot randomly assign household membership we cannot rule out the influences that lead people to live together as mechanisms. That said, these forces should exist across all treatment and control conditions and thus we can credibly evaluate whether those who receive the treatment are more likely respond. Although we cannot isolate the mechanisms, it stands to reason that those who choose to live together will be more likely to interact and behave similarly; we leave it to future research to explore the underlying mechanisms.