Public Accountability and Political Participation: Effects of a Face-to-Face Feedback Intervention on Voter Turnout of Public Housing Residents

For example, see research on the effects of higher rates of turnout on election outcomes (DeNardo 1980; Nagel and McNulty 1996; Tucker et al. 1986), or margins of victory (Citrin et al. 2003).

There is a debate about the extent to which participation in public assistance programs galvanizes or depresses political participation among recipients. For a review of theories relating the receipt of public assistance to one’s level of political engagement and participation, see Soss (1999) or Mettler and Soss (2004).

Edna Carrasco of CBPH provided background information about the scope and work of the organization as well as contextual information related to the issues facing tenants of the public housing developments included in the sample.

Information related to the approximate demographic composition of the housing developments and surrounding neighborhoods was provided by Edna Carrasco of CBPH through interviews and personal correspondence. Demographic attributes of the sample were obtained from voter files and the voter information database used by the organization.

A large portion of Complex A was demolished after the study was conducted and residents were relocated to a different public housing development.

Standard errors of the multinomial regressions are clustered by the unit of randomization, apartment entrance.

Pew Research Center for People and the Press. June 2008. “Voter Attitudes Survey,” accessed online at http://​people-press.​org/​reports/​questionnaires/​436.​pdf on December 20, 2008.

Contact rates were statistically the same in both treatment groups as indicated in Table 2. Results of probit regressions of contact on treatment assignment among subjects assigned to either treatment group (N = 540), with and without covariates for demographic attributes and previous turnout, imply no statistically distinguishable relationship between contact and the assignment to one treatment group relative to the other. Furthermore, restricting the sample to individuals who were successfully contacted by canvassers, results from probit analysis in which contact by one treatment is regressed on demographic and turnout covariates imply that covariates cannot predict which script contacted subjects heard (χ²(13) = 12.12; p-value = 0.52).

In general, binary response models are considered to be more appropriate for estimation of effects of explanatory variable on binary dependent variables. However, for models in which the explanatory variables are all or mostly discrete and take on few values, as in the case of this analysis, in which all but one explanatory variable is binary, “the [linear probability model] need not provide very good estimates of partial effects at extreme values of” the explanatory variables (Wooldridge 2001, pp 455). For this reason, OLS and probit estimates do not differ substantially in the majority of the results.

Canvassers recorded the results of their contact attempts on walk lists. I define contact as an encounter in which the canvasser has an interaction with the targeted subject. I do not count messages left with another member of the household as a successful contact.

The basic model of the instrumental variables approach to estimating the difference between the canvassing and treatment effects may be specified as follows:

In the preceding system of equations, Y_ia is a binary variable equal to one if subject i voted in the November 2007 election. X_1ia is a binary variable equal to one if subject i was successfully contacted with either treatment, and X_2ia is a binary indicator equal to one if subject i was contacted with the voter history feedback intervention. T_ia is a binary variable equal to one if subject i was assigned to either treatment group, and H_ia is a binary variable equal to one if the subject was assigned to the treatment group in which canvassers presented subjects individual voter histories, the feedback intervention. In the full model, covariates for past turnout and demographic attributes are added as specified in Eq. 2.

The largest estimated treatment effect in the present analysis, the effect of actually receiving one’s record of previous turnout, modeled with covariates for previous voter turnout and demographic attributes, is 18.7 percentage points. This estimated effect, while it is substantial in magnitude, implies that in order to generate 1,000 additional votes, the organization would have to successfully contact 5,348 voters with voter histories. In the contested race of November 2007, the at-large city council contest, the four candidates who received the most votes won seats. The fourth place winner, John Connolly, received 21,997 votes, while the fifth place candidate (first runner-up), incumbent Felix Arroyo, received 18,579 votes. Felix Arroyo, the incumbent who lost his seat, was the only Latino ever elected to the city council. The next-closest candidate in the race received only 4,008 votes. The most likely scenario in which higher turnout could have influenced the outcome of the election would have been one in which approximately 3,900 more supporters of Felix Arroyo turned out to vote resulting in a margin of victory for fourth place of about 480 votes, which would be sufficient to avoid a recount. Turnout was particularly low in predominately Latino precincts relative to previous municipal elections. Conceivably, a very large mobilization campaign targeting Latino voters could have impacted the outcome of the election. The 18.7 percentage point local average treatment effect implies that a campaign would have to have successfully contacted 20,855 voters using voter histories, all of whom would have to vote for Felix Arroyo (a strong assumption) in order to have altered the outcome of the election. The contact rate of 37.7% among the feedback intervention experimental group implies that such a campaign would have to target 55,320 voters to obtain a sufficient number of contacts.