The Social Ecology of Voting in New York City

Transcription

1 The Social Ecology of Voting in New York City A Multi-Method Approach to Voting Behavior in New York City 2013 Annette Jacoby Abstract Ideally, a functioning democratic society should be characterized by the political participation of the entire adult society, as a well grounded and equally distributed participation in the political decision-making process expresses the state s recognized legitimacy and acceptance. In reality, however, we know that political participation and civic engagement are unevenly distributed and closely related to other underlying ecological factors like age, poverty, education, and racial/ethnic composition and consequentially by no means equally spread across cities. As research has shown, disadvantaged and marginalized people vote less and therefore do not have the same potential to represent their interests and state their political opinions. In addition to the individual characteristics of non-voters, literature has shown that there is less political participation in disadvantaged areas that suffer particularly from a lack of the qualities and elements, which produce and sustain a neighborhood s overall political engagement, which will then in turn negatively impinge on the individual voting behavior. While scholars have focused on the individual characteristics that tend to produce an active voter, voting is also heavily dependent on the context the individual votes in. My paper addresses the ecology of voting in New York City by paying attention to both the individual characteristics of the voters and the contextual setting the voter lives in both in regards to the personal surrounding space and tracts. The reason I chose to use contextual models is that I believe that the nature of the relationship between how people vote and why they vote the way they do is inherently hierarchical. Moreover, research has shown that people who live close to each other tend to vote similarly. In addition to this rather traditional approach, I will also study neighboring activity spaces i.e., the actual spatial and social exposures people experience in their daily routines and their influence on behavioral voting patterns. Considering the intersection of individuals and actual behavior settings through use of so-called activity spaces may provide us with more insight into the processes by which exposure to neighborhood contexts shape the way people participate politically instead of simply relying on pre-assembled neighborhood boundaries. All of this is possible through access to 2013 New York City voter registration data as well as the voter history files that contain individual level, geocoded voter information with microlevel information on age, gender and party affiliation. It is this multifaceted relationship between individual characteristics and the neighborhood and general surrounding, which I want to analyze in this paper. My question is: What is the relationship between personal voting behavior, activity space and neighborhood characteristics? And more specifically: How do contextual characteristics of living i.e., the social exposures people experience in their daily lives influence behavioral voting patterns differently across age? 1

2 Introduction Electoral participation, especially in regard to voter turnout, is one of the most widely studied topics in the social and political science. While the right to vote is universal in advanced democracies not everyone makes use of this suffrage and the vast body of literature on voter turnout attempts to understand both why people vote or do not vote, what type of people vote and what they vote. Although voter turnout and voting frequency may be simple outcomes, the causes are complex, with several factors operating simultaneously at different levels. Thus, it would be ideal to use data that both captures information about every single individual, as well as the conditions or rather contextual effects that apply to all members of a region or country in a roughly similar manner. Basically, there have been two ways to analyze voting behavior in the literature. On the theoretical level one can use system (for instance country or constituency) or individual (micro) level explanations in order to explain why people vote the way they vote. The same distinction is reflected in the choice of empirical data. Again, one can use data from the individual level (usually administered through surveys or, as in my case, through data on registered voters) or aggregated system level data (constituency or country level turnout rates attained through the Census for instance). So far, there has been only very little research that attempts to analyze voting behavior through a combination of these levels, which has various reasons amongst which the limited data on individual level voting outcomes and the lack of geocoded voter information are the most important restrictions. On the individual level, voting has often been related to socio-demographic factors like age, gender, education or income (Topf, 1995). For example, we know that older and well-educated people are more likely to vote than young and less educated people. The problem with individual level data is that they are likely to be very unreliable. Many respondents tend to say that they voted (or intended to vote), despite the fact that in reality they did not. Aggregated data do not suffer from this problem, but obviously tells us less about the individual nuances and interactions I am interested in. Trying to deal with this deficits, this paper covers a wide range of topics, including the personal characteristics that determine an individuals frequency of votes, but also the neighborhood context, thus the ecology of voting, that impacts the way a person is politically involved. I try to 2

3 tie into the debate of the neighborhood effect and its relation to civic participation, which has been done in electoral studies for some time, but with an emphasis on qualitative approaches. There are a few exceptions, though, that base on the contextual neighborhood approach, which is currently reemerging as an important assertion for many of the processes that supposedly shape social behavior and life chances. It is in this context of a renewal of interest in local social relations and particularly the deployment of notions of social capital that this paper picks up on the connection between political behavior and the community, which is particularly interesting in a city like New York that boasts a highly diverse population and is marked by high levels of segregation. A thorough analysis of New York City Voters has been done in 2012 called Who votes? - Voter Turnout in New York City, which was published by the New York City Campaign Finance Board. However, this analysis focuses on the contextual level only by analyzing voter turnout by census tract in the 2008 and 2009 elections using Board of Elections voter history rolls and demographic data from the 2010 American Community Survey. They find that it is especially younger males, people with low educational background, U.S. citizens who were born abroad, and married households that need to be targeted in order to increase voter turnout. What is innovative about my paper is that I also take into account so-called activity spaces, which try to capture the social interaction within residential communities that affect people s political conduct. Activity space represent the spatial movements of an individual's day-to-day lived experience of place and marks an attempt to create a zone of living around the individual s place of residence. While the idea of activity spaces has existed for a while, attempts to use it have been few, which is mainly due to limited availability of the data, expense of collecting spatially referenced data and the computational burden involved. Miller (1977) argued that rather than an individual s social class being the major predictor of how he/she voted, the best predictor was where he/she lived and which people the individual was surrounded by, because those they talk to influence people. Thus, if the majority of a person s social contacts favor one political position and/or party that person is more likely than otherwise to favor it also, even if her/his personal characteristics suggest a predisposition to favor another position/party according to Miller, people who talk together vote together (Miller 1977: 65). The same is thought to be true for voting turnout in general: People living in politically and socially active 3

4 neighborhoods tend to vote more, while people in inactive neighborhoods vote less. Our idea behind this is that conventional theories of neighborhood effects on voting behavior have largely neglected actual routine exposures to local settings i.e., specific locations, organizations, and institutions. Our hypothesis is that considering the intersection of individuals and actual behavior settings through use of so-called activity spaces (Browning/Soller: 2014) may provide us with more insight into the processes by which exposure to neighborhood contexts shape the way people participate politically. Several authors have begun to use innovative context methods instead of simply relying on pre-assembled neighborhood boundaries. Schlichting, Tuckel and Meisel s paper Neighborhood and Community Context Effects on Voter Turnout- A Case Study in Baltimore, MD and Bridgeport (1996) developed a special GIS-related computer algorithm, which they used to construct geographical units, which more closely align with the study's research specifications. What is it that I am hoping to contribute to the discussion on voting? Using New York City as an example I want to a) Combine both the individual level information that I have access to through the voter registration files containing 4.2 million cases of geo-coded registered voters with neighborhood data to reach a more thorough, multifaceted understanding of NYC s voting behavior, b) And introduce an innovative use of voting activity space around every individual to account for influences on the individual that go beyond understanding the living context in terms of tract boundaries. c) Instead of just focusing on one overall model, I intend to analyze voting frequency by age sub groups, which I will construct through supervised discretization techniques that divide a continuous feature into groups (bins) mapped to a target variable. The central idea is to find those cut-points that maximize the difference between the groups in regards to their voting behavior. 4

5 Data sources, description and method The data I used is from both the New York City voter registration as well as the voter history files and contains individual level, geocoded information. My dependent variable is the number of times a person voted in a general election between 2000 and By general election I am referring to ballots that are held to elect candidates to public offices. For partisan offices, nominees from the preceding party primary elections appear on the ballot, along with independent candidates. The individual-level information includes the registered population s names, addresses, ages, party affiliations, family sizes and voting histories (length of registration). While some of the data was provided, other measures had to be configured. Through the geocoded information I was able assess how many people with the same last name live at a given address, which proved tricky as many people report their addresses differently, even if they live at the same address and belong to the same family. Obviously, this measure will have an inherent bias, which I tried to limit through extensive data cleaning. Furthermore, Professor John Mollenkopf of the CUNY Graduate Center, who has collected and constructed the political data, has conducted a surname analysis to identify ethnic and religious backgrounds. As the voter registration forms do not require citizens to identify their countries of origin, he compiled volumes of surnames associated with specific ethnic groups and then categorized voters names by ethnicity in a project that took about eight years. While this tool might not be the most accurate, it nonetheless offers a revealing snapshot of an increasingly diverse electorate, has served as very good predictor of voting patterns in New York City, has helped target minorities and provide voting assistance. Several studies suggest surname analyses produces reasonable estimates of whether an individual is Hispanic or Asian/Pacific Islander. HIver, surname analysis is not accurate for identifying African Americans, which is why I do not have any information on this group. Put shortly, after careful consideration I decided to include to measures, namely family size and before-mentioned ethnic proxies, that I hope will enhance the pool of meaningful level one variables by adding to the relatively sparse set of available data through the New York City voter registration files. 5

6 Originally, the dataset comprised roughly 4,200,000 (all of New York s registered) voters, but I decided to use a smaller sample as too large sample size can artificially inflate the statistical results and make it possible to detect smaller and smaller effects, even though they are not practically relevant. Furthermore, I ran into computational difficulties when trying to work with 4.2 million individual cases. After excluding those under the age of 30, since they do not have the same chance of voting the maximum number of times between 2000 and 2013 as the other voters, and excluding those that have been registered for less than 12 years, my random sample contains people, who theoretically all had the same prospects of voting frequency. Although I will not explicitly take it apart in this paper, individual level determinants will be analyzed in two subcategories: Demographic characteristics (including the ethnic proxies through surname analysis) and political affiliation. I will study these factors to answer questions about how personal attributes such as gender, age, household size, party affiliation and length of registration affect turnout. In general, there are several papers that try to assess the meaning of individual characteristics for voting behavior, but they all base on information given by the respondents. The general findings from this literature are that men vote more than women, the elderly more than the young, the better educated more than the less educated, and the rich more than the poor, though these results vary in their significance, and on rare occasion, direction (Wolfinger and Rosentstone 1980; Powell 1986; Leighly and Nagler 1992; Verba, Schlozman, and Brady 1995). In addition to using the individual level information, I have also aggregated the data to the NYC tract level boundaries and merged in the required ecological data from the Census through geocoding the addresses of each voter. The contextual data drawn from the Census (2010) that I am most interested in is the percentage of foreign-born people per tract, homeownership rates and housing tenure (crude measure of residential stability/ less than 5 years ago), voting results (attributed to the tract level through geocoding) and deprivation (factor analysis of neighborhood measures of educational attainment and poverty/unemployment percentages). As I have pointed out in the introduction, I will enhance my territorial hierarchical analysis based on shared neighborhoods (census tract, other administrative boundaries) by an egocentric definition: Each individual is assigned to a unique neighborhood based on his or her location in 6

7 space and, as implemented here, a fixed radius of 200 feet around that point and calculating the average number of votes within this given radius. For future analysis it would be great to include other information concerning the characteristics of these activity spaces into the analysis. From a methodological standpoint, however, measuring activity space is more data and computationally intensive than distance, and its complexity has resulted in its underutilization. The limited availability and expense of collecting spatially referenced data and the computational burden involved in generating SDEs restricted such studies to small samples. With advances in GIS and increasing availability of spatially referenced data, activity space has become a more viable tool for studying accessibility. Assessing the needs for a multilevel model The reason I chose to use contextual models is that I believe that the link between why/how people vote and where they live is immanently hierarchical. When trying to provide evidence on why it makes sense to analyze a problem through multiple levels it is usually recommended to have three types of justification for the model: Empirical, statistical and theoretical explanations. In order to understand how the frequency of votes is distributed across the city it makes sense to start of by looking at the maps below that show considerable variability from tract to tract regarding voting behavior, especially with the second set of maps showing the strong correlation between clusters of low and high voting frequency and the underlying pattern of deprivation. As we can see there is a lot of variation in voting frequency within but also between the different boroughs indicating that the different values are not randomly distributed, but there seems to be clusters of high and low voting frequency. North Manhattan, East Queens and Staten Island seem to have the highest numbers of times voted between 2000 and The measure of deprivation used in the maps below is a factor-analyzed score I constructed from education, poverty and employment rates, which serves as an example here. I might as well have used contextual measures of residential stability or the percentage foreign-born to show that by glancing at the maps we can see a pattern of coincidence between the different levels of analysis. 7

8 Deprivation turned out to be an excellent example, because we see that clusters of high-high voting patterns tend to be very privileged especially in neighborhoods like Park Slope, Upper East and West Side and Lower Manhattan. We see that areas with high voting population tend to have very low deprivation scores, whereas those neighborhoods, where people have low rates of voting frequency, range amongst those with the highest deprivation scores citywide. In addition to this empirical explanation of why a hierarchical model is preferable, we can argue statistically that regular OLS regressions assume that observations (and hence the error terms) are independent from each other. However, in this paper it is better to assume that there is a nested structure and the independence assumption is violated. What this means is that tracts in Manhattan are probably more similar to each other than to other tracts in Brooklyn. Expressed in a colloquialism, politics are local, and it is reasonable to assume that tract characteristics can shape political behavior as much as other individual factors, such as political affiliation or age. Individuals are nested in Congressional districts and receive campaign information from numerous sources in their district and outside of their district. By using a multilevel model, I hypothesize that many of the contextual variables influenced voter turnout through individual characteristics. Although the results of single-level models and multilevel models are often not extremely different, multilevel models are both theoretically and methodologically more appropriate when examining hierarchical processes. The final multilevel poisson count model my analysis bases on is: E(Votes βj) = λ(ij), log[λ(ij)] = η(ij) η(ij) = β0j + β1j(age(ij)) + β2j(family Size(ij)) + β3j(gender(ij)) + β4j(republican(ij)) + β5j(did not Specify(ij)) + β6j(independent(ij)) + β7j(conservative(ij)) + β8j(working family(ij)) + β9j(green(ij)) + β10j(length of Registration(ij)) + β11j(jewish(ij)) + β12j (Russian(ij)) + β13j(italian(ij)) + β14j(irish(ij)) + β15j(hispanic(ij)) + β16j(chinese(ij)) + β17j(korean(ij)) + β18j(muslim(ij)) + β19j(number Neighbors(ij)) + β20j(average Votes of Neighbors in Activity Space(ij)) β0j = γ00 + γ01(percent Foreign-born j) + γ02(percent Owner j) + γ03(residential Stability j) + γ04(average Votes by Tract j) + γ05(deprivation Score j) + u0j 8

9 Map 1. Number of times People voted between 2000 and 2013, aggregated by NYC 2010 Census Tract 9

10 Map 4 and 5. NYC Deprivation and Inverse Clusters of High-High (3) and Low-Low (4) Voting Participation 10

11 Contextual Variables Ethnic Proxies through Surname Analysis Party Affiliation and Other Voting Information Basic Demographics Results a) Descriptive Statistics Level 1 N MEAN SD MINIMUM MAXIMUM Age Female Family Size Democrat Republican Did not Specify Independent Conservative Working Families Green Years Registered NUMBER OF VOTES (DV) Number of Neighbors within ft Average Votes of Neighbors Jewish Russian Italian Irish Hispanic Chinese Korean Muslim Other Level 2- Tract Level Context N MEAN SD MINIMUM MAXIMUM Percentage Foreign-Born Percentage Owners Residential Stability Average Votes Deprivation

12 b) Multi-Level Regression Analysis (Part 1) Fixed Effects Model Level 1 Variables Only Fixed effects including Level 2 (Population-average model with robust standard errors) (Population-average model with robust standard errors) Fixed Effect Coefficient Standard Error P-value Coefficient Standard Error P-value Constant < <0.001 Percentage <0.001 Foreign-Born Percentage Owners Residential <0.001 Stability Average Votes Deprivation <0.001 Age < <0.001 Family Size < <0.001 Female < <0.001 Republican < <0.001 Did not Specify < <0.001 Independent < <0.001 Conservative Working Family Green Years Registered < <0.001 Number Neighbors Average Vote < <0.001 Activity Space Jewish < <0.001 Russian Italian Irish Hispanic < <0.001 Chinese < <0.001 Korean Muslim (Age Squared) < <0.001 Dependent Variable: Number of Times voted in General Election between 2000 and

13 c) Multi-Level Regression Analysis (Part 2) Fixed effects including Level 2 by Age-Groups Age Group Fixed Effect Coeff SE P Coeff SE P Coeff SE P Constant <0.001 Percentage <0.001 Foreign-Born Percentage Owners Residential < Stability Average Votes Deprivation < <0.001 Age <0.001 Family Size < <0.001 Female < < Republican < <0.001 Did not < < <0.001 Specify Independent < < <0.001 Conservative Working Family Green Years < <0.001 Registered Number Neighbors Average Vote < < <0.001 Activity Space Jewish Russian Italian Irish Hispanic < < <0.001 Chinese < <0.001 Korean Muslim (Age squared) <

14 Fifty-eight percent of adult citizens in this sample were females, the average age was 58 years (ranging from 30 to 93 years) and the mean family size was As I have explained before I constructed this family size measure through aggregating those people with the same last name living at the same addresses. While this is just a proxy and only pertains to those that are registered voters and leaves out children for instance, this measure still gives us a feeling for the family situation of the individual. 45 percent were Democrats, 33 percent were Republicans, another 13 percent did not specify their party affiliation, and 8 percent were conservatives. Only one percent of NYC s voters are green and 3% affiliate with the Working Families. As explained in the introduction, I constructed buffers of 200ft around every voter and calculated the average times voted within these buffers and counted the number of neighbors as a measure of density. On average, a New Yorker has 262 neighbors within a 200ft radius. It is interesting to point out that there is a great variability within this variable, as some people only have one neighbor and some have as much as 2712 neighbors in their close vicinity. I regards to the ethnic surname analysis, 64 percent of the sample was not identified by their last names and served as the control group. 17 percent of surnames in New York City had a Hispanic origin, 8 percent had Jewish last names, while only 3 percent were Chinese, Italian and Irish, and one percent was Korean, Muslim and Russian. These simple tabulations are presented in Table a) in the results section. New York City s Voters are moderately active. Having been given the possibility to vote 11 times between 2000 and 2013, the average New Yorker chose to vote 5 times. In general we can say that the Multi-Level Regression Analysis (Part 1) does not show any substantial differences between Model 1 and Model 2, meaning that individual level outcomes tend to be only mildly influenced by the context they are situated in. However, although the effects tend to be rather small, level 2 variables should not be neglected and play an interesting role in part 2 of the Multi-Level Regression, in which I analyze the frequency of voting through an age-specific lens. In general, we can say that an increase in the tract percentage of people, who were born in a foreign country, tends to influence the voter positively. The positive impact of the foreign-born population is no surprise given that John H. Mollenkopf, director of the 14

15 Center for Urban Research at the Graduate Center of the City University of New York, has found in his analysis that New York City has more foreign-born active voters than any other big city in the US, both in raw numbers and as a share of the total electorate. Furthermore, higher percentages of homeowners tend to depress individual voting outcomes, which stands in stark contrast to the common-held belief that homeowners vote more because they tend to be less mobile and are impacted more by the quality of government decision for the community. However, as residential stability (here measured by the percentage of the population that has resided in the same neighborhood for more than five years) increases so do votes. This shows that people who have lived in a neighborhood for a longer time face less barriers of registration and voting than those who have recently moved may. As the Finance Campaign Board has pointed out in their paper Who Votes? - Voter Turnout in New York City (2012) efforts to increase voter turnout need to focus on those areas that face a more transient, less stable population, especially younger people. As the deprivation score increases for a neighborhood, average voting frequency falls. This finding supports prior research that claims that low levels of education and poverty are major factors in a person s propensity to vote less. Finally, I also looked at the average numbers of votes in a given tract. As you would expect, people who live in higher voting frequency tracts tend to vote more, either by choosing to live in politically active neighborhoods or by being influenced by their active neighbors. Similar to what we find in the literature, our overall model shows that the younger and male population tends to vote less. These figures stay very robust as I introduce the second level. Taking Democrats as my reference group, we see that all groups vote less frequent than the average Democratic voter. Those that did not specify their party affiliation had the lowest frequency of voting when compared to the other groups. When compared to people with not identifiable surnames, people with Jewish, Irish and Italian surnames tend to have higher votes, whereas the other ethnical groups tend to vote less. While this measure is by no means perfect, I think it shows that of those people with ethnic surnames it is those with white assimilated last names that seem to be the most politically integrated. A major focus of this paper is on the voting conditions that enhance or deteriorate a person s propensity to be actively engaged in the election process. First of all, the number of years a 15

16 person has been registered increases the number of times voted between 2000 and 2013, which shows that people do not register and become active voters right away but that it is a life long development. Furthermore, having controlled for the number of neighbors within a 200ft radius of every single individual, I find that the personal activity space is the most relevant indicator of voting frequency. Controlling for the tract voting average, we only see a slight decrease in the importance of the personal activity space. When I started this analysis, I used simple techniques of data mining to determine which were the most important variables I determining how often a person votes. Age and the number of years registered turned out to be among the most influential factors, which is why I decided to look at the different groups separately to see if the overall model looks different when analyzed sub group. As I have pointed out before, these groups boundaries were chosen to optimally discretize age by distributing the values of each variable into bins. Bin formation is optimal with respect to a categorical guide variable (number of times voted) that "supervises" the binning process. What I find is astonishing. Looking at the second level coefficients, we see that a higher percentage of people born in a foreign country is only positive for people aged 55 and older, whereas higher rates of foreign-born people among the two younger age groups tend to reduce individuals voting agility (although these results are not statistically significant). The number of average votes no longer is significant for age groups 1 and 2, when I control for personal activity space (which is highly significant for all three age groups), but the oldest age group, those above 55, still seem to be impacted by the average tract voting frequency in addition to their personal voting space. Looking at the level 1 coefficients yields even more interesting results. While family size is still beneficial to all the age groups activeness, being female seems to be less important as the age groups increase. For those aged 55 and above there seems to be no distinguishable difference between men and women in regard to their voting behavior, while women aged tend to vote 11 percent more than their male counterparts. Women aged vote 6% more than men, when controlling for all the other variables. 16

17 In regards to the party affiliation, we can see that in general democrats still tend to be the most active voters, topped only by Green voters aged What stroke me as interesting in regards to the ethnic surnames is that, while the most active voters still tend to have assimilated white family names, the picture becomes less unidirectional and clear when we take into account the different age groups. For Jewish family names, it is mainly those above 38 that tend to vote more than the reference group. The opposite is true for the Irish: While the older groups were insignificant, people aged with Irish surnames tend to vote substantially more than the control group. CONCLUSION In a multilevel analysis that goes beyond the conventional hierarchical models, I looked at the relationship between voting frequency in New York City between 2000 and 2013 and neighborhood characteristics such as contextual socio-economic status and the percentage of foreign-born people, as an addition to the individual attributes of residents such as race and age, which are provided in the New York City voter registration file. Higher frequency of votes tends to indicate greater civic engagement, which is a good quality for any neighborhood in terms of citizenship participation, but also with other indicators of neighborhood quality and child well being, for instance. Furthermore, the number of people voting shows, who feels engaged as a citizen and who does not. In this paper, I wanted to analyze these questions while going beyond the tract-related understanding of context and trying to understand the direct impact of neighbors on a person s voting behavior net of the tract s impact. Therefore, I used the geocoded addresses of each single individual and calculated the average values for the surrounding half-mile radius. Having access to 4.3 million registered voters in NYC gave me the rare opportunity to make use of these activity spaces. The paper develops the existing literature in the following three ways. First of all, voting seems to be highly clustered and therefore we know that there is a tendency of people living close to 17

18 each other to vote similarly, which I have addressed in this paper through a multilevel hierarchical analysis of voting behavior that takes into account both the personal activity space surrounding a single individual and the contextual conditions a person lives in. My most evident finding is confirmation of the higher voting frequency of female, older and more assimilated citizens. Another finding is the inclination toward lower participation by people who are not Democrats and live in deprived instable neighborhoods. Contrary to previous research, I find that these results are highly dependent on age and cannot simply be analyzed in an overall model. By looking at three optimally binned age groups separately, I was able to detect differences and even contraries that the aggregate model did not reveal. For instance, we now know that the smaller personal activity space of an individual tends to be much more important in determining how often a person votes when compared to the times voted by tract. In fact, the tract effects for age groups and were insignificant once I controlled for personal activity space, and remained significant only for those aged 56 and above. What does this mean? Does this mean that older people are also influenced outside of their personal zone? I believe that the research on personal activity spaces needs to be greatly enhances, as it tells us much about where people live and move and how they are influenced politically. Therefore I suggest that political participation of the electorate cannot be fully understood without taking into account the variations between different age groups. We need to be more careful about drawing oversimplified conclusions, especially when it comes to targeting voters with low voting frequency and turnout. 18