Political Socialization in Context: The Effect of Political Competition on Youth Voter Turnout

Abstract

Adolescence is an important time for political development. Researchers have concentrated on the family as the sole socializing agent of youths; however, as Campbell, Gimpel, and others have shown, political contexts also matter for young citizens. Using the National Education Longitudinal Study of 1988, the Record of American Democracy, and election outcomes data, I find that adolescents who resided in politically competitive locales or states have higher turnout years later compared to those who lived in uncompetitive contexts. These effects are not mediated by the home political environment and act through political socialization. This research adds to a growing literature on the influence of political contexts on political behavior and is the first to explore how political competition during adolescence influences voter turnout in young adulthood.

Appendix

Transforming the ROAD into Zip Code Level Data

Most of the electoral data in ROAD were aggregated from the precinct level into Minor Civil Division Groups (MCDGs), which correspond roughly to towns or cities. There exists some overlap between zip codes and MCDGs; more than one MCDG can be in the same zip code and each MCDG can contain multiple zip codes. In order to match the electoral data with the NELS zip codes, each MCDG was given a political measure based on the raw number provided by ROAD and the corresponding population of the MCDG in relation to the zip code’s population. The proportion of votes for each MCDG, weighed according to size, was then summed to create a new political measure for each zip code. For instance, imagine that three MCDGs overlap with zip code A: MCDG1, MCDG2, and MCDG3. MCDG1 encompasses 40% of the population of zip code A, while MCDG2 has 55%, and MCDG3 has 5%. For each MCDG, we not only know the percentage of people from that MCDG who make up zip code A, but also the total number of votes cast for each election. The total votes cast for a particular party in a particular election in zip code A can then be estimated by the following logic: (1) In MCDG1, 50 votes were cast and 30% of the MCDG lies in zip code A, (2) so it is assumed that 30% of 50 (15 votes) were cast in the portion of the zip overlapping with A and, similarly, (3) 50% of the MCDG2’s votes were in the zip (30 votes) and, (4) 10% of MCDG3’s votes (20 votes). So, it is estimated that 15 + 30 + 20 = 65 votes were cast in zip code A. All valid count variables were summed in this manner, although the ultimate level of aggregation differed according to states. For instance, 1 state (CA) was merged using the Census blockgroup level, 13 states (CT, IL, IN, MA, MN, NH, NJ, NY, OH, PA, RI, VT, and WI) were merged at the MCD level, 26 states (AL, AR, AZ, CO, FL, GA, IA, ID, KS, LA, MD, MI, MN, ME, NC, ND, NE, NV, OK, SC, SD, TN, UT, VA, WA, WV) were merged at the MCDG level, 9 states (DE, HI, KY, MS, MT, NM, OR, TX, WY) were merged at the county level, and 1 state (AK and DC) was merged at the state level. Since the state level political competition measure correlated perfectly with the local political competition measure for those states that were merged at the state level, respondents from AK and DC were deleted from the analyses (N = 50).

California

The California data are organized differently from the rest of the country and at a lower level of aggregation. Unfortunately, information about the 1984 and 1988 elections in California were unattainable; instead, ROAD includes information about the 1992 presidential election. The political competition measures were created using only the 1992 election for the 1,080 respondents who lived in California during the 1988 base survey. Note that although the political competition measure for California is based on the 1992 presidential election, it was still match merged with respondent’s 1988 residences. Consequently, I make the assumption that localities in California did not change significantly in their political competition scores from the 1988 election to the 1992 election so that respondents from California can be included in the analyses. Indeed there is evidence that political competition is highly correlated from one presidential election to the next; the correlation coefficient between political competition in the 1984 presidential election with political competition in the 1988 presidential election is a strong .77 across all levels of aggregation in the ROAD dataset.

Using OLS Instead of HLM

Methodologically, the data structure is of a multilevel nature; respondents are clustered within schools and locales, which are clustered within states. Using Ordinary Least Squares (OLS) to analyze clustered data structures produces unbiased, but inefficient estimators warranting methods that take account of within cluster homogeneity. Robust standard errors and the suite of methods called Hierarchical Linear Models (HLM) both seek to correct this. OLS is comparable to HLM when the majority of the variability of the dependent variable is at the lowest level and when individual level effects are fixed (do not vary randomly across clusters). If individual level variables vary systematically across clusters, these can be modeled as interactions in the robust OLS case or as fixed “level-2” cross-level interactions within HLM. OLS may be preferred to HLM, however, when the complexity of the data causes problems of convergence.

Due to the variability of the dependent variable and the complexity of the data I decide to use OLS regression with robust standard errors. The complex nature of the data occurs because of the intersection of the sampling design (most schools have 6–12 students in the analysis) and residential zip codes used to calculate local political competition (students from multiple contexts of competition may attend the same school). Hence, the data structure is actually cross classified with students clustered within schools and local contexts clustered within states. As a result, the second level was recreated to account for both schools and residence. Students were thus clustered within “school-zipcode” combinations, which were clustered within states. I attempted to estimate a three-level model (with students clustered within school-zips clustered within states) but this three level model failed to converge suggesting a lack of variation at each level in the turnout scale.

I ran a variety of intercept-only models (also called “empty models”) to assess the ICC, which partitions the variability of the dependent variable across the various levels of analyses. First, I estimated an empty model of students clustered within school-zipcodes, ignoring the states. Results indicated that 5% of the variation in the turnout scale was at the school-zipcode level. Next, I estimated an empty model of students clustered within states, ignoring the school-zipcode. Results indicated that 2% of the variation in the turnout scale was at the state level. The combination of these results suggest that at most 3% of the variation of the turnout scale is at the school-zip level while 2% occurs at the state level.

Because the ICC is small, it is assumed that the degree of bias in the standard errors produced by OLS is also small since bias decreases as the ICC decreases. Even in the presence of a small ICC, however, OLS still produces inefficient estimates. To correct the standard errors, I use Huber-White standard errors for clustering within school-zips while ignoring the clustering within states. States are ignored because little variation in the turnout scale occurs at this level. Simulation studies show Huber-White standard errors are similar to those estimated by HLM in the two- and three-level cases (Cheong et al. 2001).

Nonetheless, all OLS models reported were estimated with a two-level HLM of students clustered within school-zips ignoring states. These could not be estimated using survey weights so unweighted models were compared with identically specified OLS models with Huber-White standard errors. Results from the unweighted OLS model with robust standard errors and the unweighted two-level HLM model are nearly identical. Any discrepancies in the coefficients or standard errors are typically at the 10th or 100th decimal place. Most importantly, inferences drawn from both methods are identical. I conclude that the within cluster homogeneity is adequately corrected for by using OLS and robust standard errors.

Table A1 Mean values and standard deviations of independent variables