Why Do Communities Mobilize against Growth: Growth Pressures, Community Status, Metropolitan Hierarchy, or Strategic Interaction?

ABSTRACT:

Findings from this study challenge the conventional wisdom about the motivations for local growth control. Using data of California ballot box growth controls merged with city level demographic and housing data from the U.S. Census Bureau, logit models are estimated to test four hypotheses for why communities mobilize against growth. Of the four hypotheses, growth pressures, community status, metropolitan hierarchy, and strategic interaction, the only hypothesis that was strongly supported by the logistic regression analyses was strategic interaction. Support for the strategic interaction hypothesis reveals that jurisdictions located in regions where growth control policies are more abundant have a higher probability of mobilizing against growth. In other words, jurisdictions’ growth control policies influence the growth decisions made by neighboring jurisdictions within the same region. One of the most surprising findings in the logistic regression analyses is that low-income suburbs are significantly more likely to mobilize against growth than high-income suburbs. These results refute the commonly held belief that growth control is strictly a concern of elite communities and suggest that residents of low-income suburbs may be turning to the ballot box to control growth because their communities are the locations of choice for noxious land uses.

Notes

1 This is often referred to as the “fiscalization of land use.” For more information on this topic, see Misczynski (1986), Fulton (1999), and Lewis (2001).

2 Upon requesting the sample ballots, it was realized that there were five duplicate measures that were on the list, and therefore these were omitted. Other reasons for the lack of retrieval of ballot measures involved city and county clerks’ staff not being able to find copies of sample ballots or not having sufficient staff members to locate these documents. In general, city and county clerks’ staff attempted to be helpful.

3 We also coded the dependent variable as an ordinal variable with four categories: 0, 1, 2, and 3. The categories 0, 1, and 2 represent the exact number of BBGCs a city qualified throughout the study period. If a city qualified 3 or more BBGCs, the dependent variable was coded 3. There were 314 cities that had 0 BBGCs, 56 cities had 1, 22 cities had 2, and 30 cities had 3 or more. We ran ordinal logistic regression using this alternative dependent variable regressed on all of the same independent variable as found in the logistic models. The results from the ordinal logistic regression models were qualitatively the same as the logistic regression models. We chose to report the results from the logistic regression analyses because there is greater familiarity among readers with logistic regression as compared to ordinal logistic regression.

4 California had 422 incorporated cities in 1980 and 456 in 1990. Since some of the variables used in the logistic regression analyses measure change in sociodemographics between 1980 and 1990, cities that were established after 1980 were not included in the analyses (i.e., 422 cities are included in the analyses).

5 Bivariate correlations were conducted for all explanatory variables to assess possible collinearity. No correlation coefficients exceed 0.70, which is the commonly accepted threshold for variables that may be collinear. The highest correlation coefficient was −0.607, which is well below the level that would be cause for concern.

6 The original full model included three other community status variables: median housing value, percentage of college educated, and percentage of persons in professional or managerial occupations. These variables were highly correlated with the variable median household income, with bivariate correlations of 0.70 and above, and therefore created problems of multicollinearity. The decision to retain median household income and omit these three variables in the analysis was based on income being the most frequently used indicator of social status in the political participation and growth control literatures.

7 When the distribution of independent variables was evaluated, median household income was identified as being positively skewed. Instead of transforming the variable, such as by taking the log value, three categories were created to classify median household income. In a prior analysis, the log value of median household income was included in the model, but it was found to be highly correlated with the interaction variable in the model. Developing three categories for median household income did not pose problems with multicollinearity. Not only do the three income categories make sense for quantitative purposes, but conceptually we think of income in terms of these three categories.

8 The original full model included variables measuring change in Asian population, change in Black population, and change in Hispanic population, but none of these variables were significant in any of the models. For the sake of parsimony, they are not included here.

9 The distribution of population city size was positively skewed, and therefore the variable was transformed by taking the natural log. When a variable is positively skewed, this is a strong indicator that it is exponentially distributed. When variables that are exponentially distributed are included in linear models, such as logit models, it is common practice to transform them (Kay & Little, 1987).

10 Percentage of registered voters who are Democrat was included as a control for political ideology in the original full model, but it was not significant in any of the models and was therefore omitted.

11 For more information about Proposition 13, see Fulton (1999).

12 To determine whether the full model is a better fitting model than the nested model, a likelihood ratio chi-square test is conducted. This test involves subtracting the −2 log likelihood ratio of the restricted model from the −2 log likelihood ratio of the full model. The difference is distributed as a chi-square (χ²). Significance values can be determined by examining the χ² distribution table and by calculating the difference in degrees of freedom between the full model and the restricted model. This test is also called a “goodness of fit” test. The calculations are not shown here, but can be obtained upon request from the author.