Robo calls with GOTV messages have a small, positive effect on turnout; election protection script robo calls are not effective in mobilizing voters.

Green, Donald and Dean Karlan. 2006. "Effects of Robotic Calls on Voter Mobilization." Unpublished Manuscript. Institution for Social and Policy Studies, Yale University.

Donald Green and Dean Karlan report the results of an experiment conducted in 2004 to gauge the effects of robotic phone calls designed to encourage voter participation. Two scripts, each read by Vanessa Williams, were used as treatments. The scripts were adapted from radio recordings that were used as public service announcements. One script conveyed an election protection message; the other, a message stressing voting as a civic duty. These messages were conveyed to registered voters (with known phone numbers) in North Carolina and Missouri during the day prior to Election Day in November 2004. These voters were randomly sampled from voter lists. The authors assess the effects of the treatments by comparing voting rates of those assigned to receive each message to the control group, which received no treatment. The authors find no evidence that the Election Protection script increases turnout

Electoral Context: This experiment was conducted in the context of the 2004 presidential election in North Carolina and Missouri.

Subject Population: Green and Karlan created three sample frames for the target population of the experiment using nine-digit zip codes in which there were at least five registered voters. The first consisted of areas with high densities of registered voters (at least three) half of which or more are African American. The second stratum consisted of areas of low registered voter density where there were at least 1 or 2 registered individuals in the nine-digit zip code, and at least one was African American. The third sample stratum consisted of areas of high densities of registered voters, fewer than 5% of which are African Americans. Race information is included in North Carolina registration records; race in Missouri was imputed from Census data.

Randomization procedure: Individual voters within each stratum were randomly assigned to treatment and control groups. The treatment group of the first stratum consisted of 181,793 voters and the control group included 221,996 voters. The treatment group of the second stratum consisted of 104,798 voters and the control group included 107,606 voters. The treatment group of the third stratum consisted of 45,655 voters and the control group included 2,348,797voters.

Treatment: Those individuals assigned to the treatment groups heard one of two scripts, prerecorded messages of which were read by Vanessa Williams in robotic calls. The first script, the "Know Your Rights" script was an election protection message encouraging African American voters to know their voting rights and be aware of illegal tactics that had denied African Americans access to voting in recent elections. The second, "GOTV script" cited issues of fair wages, health care and education and encouraged voters to make their voices heard through the practice of voting. The calls were delivered on the Monday prior to Election Day.

Findings: Turnout was measured using voter records. In the first stratum, Green and Karlan find a 64.3% voting rate in the control group. The voting rate is almost identical in the Election Protection group and 0.2 percentage points higher in the GOTV group. In the second stratum, the results are similar. Again, turnout was slightly (and insignificantly) lower in the Election Protection group and slightly (and insignificantly) higher in the GOTV group. In the third stratum, which is largely comprised of whites, the results are similar: turnout was slightly (and insignificantly) lower in the Election Protection group and slightly (and insignificantly) higher in the GOTV group. Green and Karlan find no evidence that the Election Protection script increases turnout. The GOTV script appears to have a very small positive effect, but the results for the GOTV script are not statistically significant at the .05 level.